JP2024065881A - Prediction system, prediction method, and system management system - Google Patents

Abstract

A prediction model for a prediction target whose response characteristics have changed is determined with a low load, and the determined prediction model is updated.
[Solution] A prediction system generates update data. The update data includes sample data for each observation period for one or more prediction targets. For each prediction target, for each observation period, the sample data includes one or more observation values obtained for the prediction target during the observation period and observation values obtained for each of one or more factors during the observation period. The prediction system calculates a degree of discrepancy between a first sample group and a second sample group for each prediction target. For each prediction target, the first sample group is one or more sample data used to identify a prediction model corresponding to the prediction target, and the second sample group is one or more sample data that have not yet been used to identify a prediction model corresponding to the prediction target. The prediction system updates a prediction model corresponding to a prediction target whose degree of discrepancy is equal to or greater than a threshold.
[Selected figure] Figure 2

Description

The present invention generally relates to data prediction.

Predictions of future data are made in a variety of fields, including weather forecasts, energy demand forecasts, and forecasts of various economic indicators. For example, in the power sector, there are cases where transmission losses are reduced by controlling grid equipment based on predictions of power demand for tens to thousands of substations every few tens of minutes to every few hours.

The object to be predicted (prediction object) has the property (response characteristics) of showing a specific response when influenced by one or more factors. For example, the power demand of a substation (one example of a prediction object) changes in response to fluctuations in temperature (one example of a factor). When the temperature is high, the operation rate of air conditioning equipment increases, and there is a tendency for power demand to also increase. In prediction, a prediction model is identified that shows the response characteristics of the prediction object to factors estimated from past observation data.

However, response characteristics change over time. For example, the response characteristics of electricity demand change over time due to seasonal changes, changes in people's lifestyles, or changes in factory operating conditions. In order for the response characteristics of the prediction target to track changes over time, it is desirable to continue updating the prediction model. Furthermore, when repeating prediction and control at short intervals (for example, when performing control based on demand forecasts at the substation level), it is desirable to continue updating the prediction model at short intervals.

Technologies for updating a prediction model are known from JP 2000-276460 A (Patent Document 1), JP 2004-86896 A (Patent Document 2), and JP 2022-28338 A (Patent Document 3).

Patent Document 1 discloses a technology for learning a prediction model, and predicting power demand using the learned prediction model as the best prediction model only when the error in the output data of the learned prediction model has decreased. Patent Document 2 discloses a technology for updating a prediction model when the error in the predicted value output from the prediction model is equal to or exceeds a set tolerance. Patent Document 3 discloses a technology for updating a prediction model when an outlier is detected in the observed value of the state value of a learning dataset.

JP 2000-276460 A JP 2004-86896 A JP 2022-28338 A

In Patent Document 1, model learning is required to determine whether a prediction model needs to be updated, so the processing load for determining whether to update all prediction models is large, and it is difficult to complete the updates within a specified time. In Patent Document 2, the prediction model is not updated until the error in the predicted value reaches or exceeds an allowable value, and the timing of the model update is delayed relative to changes in the response characteristics of the prediction target, resulting in long-term errors. In Patent Document 3, the state values of the learning data are factor data used to identify the prediction model, and even if an outlier occurs in the observed value, it does not necessarily mean that the response characteristics have also changed, so there is a risk of unnecessary model updates (model updates that result in reduced prediction accuracy).

The present invention has been made in consideration of the above points, and aims to determine, with low load, the prediction model of a prediction target whose response characteristics have changed, and to update the determined prediction model.

The prediction system generates update data from the prediction target observation data and the factor observation data. The prediction target observation data includes, for each of one or more prediction targets, time series data of past observation values obtained for the prediction target. The factor observation data includes, for each of one or more factors that may affect the one or more prediction targets, time series data of past observation values obtained for the factor. The update data includes sample data for each of one or more prediction targets for each observation period. For each prediction target, for each observation period, the sample data includes one or more observation values obtained for the prediction target during the observation period and observation values obtained for each of one or more factors during the observation period. The prediction system calculates the degree of difference between the first sample group and the second sample group for each prediction target. For each prediction target, the first sample group is one or more sample data used to identify a prediction model corresponding to the prediction target, and the second sample group is one or more sample data that has not yet been used to identify a prediction model corresponding to the prediction target. For each prediction target, the prediction model is a model that takes as input a factor prediction value, which is data including a predicted value for a future period for each of one or more factors, and outputs a predicted value of the observed value obtained for the prediction target in the future period. The prediction system performs an update necessity determination for each prediction target, which is a determination of whether the degree of difference is equal to or greater than a threshold, and outputs the result of the update necessity determination. A prediction model for which the result of the update necessity determination is true is a prediction model that requires updating. A prediction model for which the result of the update necessity determination is false is a prediction model that does not require updating. The prediction system updates a prediction model that is determined to require updating from one or more prediction models corresponding to one or more prediction targets.

According to the present invention, it is possible to determine the prediction model of a prediction target whose response characteristics have changed with a low load and to update the determined prediction model.

Other issues, configurations, and advantages will become clear from the description of the embodiments below.

FIG. 1 is an overall configuration diagram of a power grid management system. FIG. 2 is a conceptual diagram showing a data flow of a power grid management system. FIG. 2 is a functional configuration diagram of each device constituting the electricity demand forecasting system. FIG. 4 is a functional block diagram of a data extraction unit. FIG. 4 is a functional block diagram of a model update control unit. FIG. 2 is a functional block diagram of a model management unit. FIG. 13 is a configuration diagram of update prediction target data. FIG. 13 is a diagram showing a configuration of updating factor data. FIG. 13 is a diagram showing the configuration of prediction factor data. FIG. 13 is a diagram showing the structure of classification-added updating data. FIG. 2 is a diagram showing the structure of model management data. FIG. 2 is a diagram showing the structure of model update control data. FIG. 2 is a diagram showing the structure of model data. 13 is a flowchart showing a power demand prediction process. FIG. 13 is a diagram illustrating a concept of determining whether or not an update is necessary based on the degree of difference. FIG. 13 is a diagram illustrating a concept of determining whether or not an update is necessary based on the degree of difference. FIG. 13 is a diagram illustrating a model update timing and an update processing load according to a comparative example. FIG. 11 is a diagram illustrating a model update timing and an update processing load according to the embodiment. FIG. 13 is a diagram showing a specific example of a case where a model needs to be updated. FIG. 13 is a diagram showing a specific example of a case in which a model update is not required. FIG. 13 is a diagram showing a specific example of a case in which a model update is not required. FIG. 13 is a functional block diagram of a data extraction unit according to a modified example.

In the following description, an "interface unit" may refer to one or more interface devices. The one or more interface devices may be at least one of the following:
One or more I/O (Input/Output) interface devices. The I/O (Input/Output) interface devices are interface devices to at least one of the I/O devices and a remote display computer. The I/O interface device to the display computer may be a communications interface device. The at least one I/O device may be a user interface device, for example, either an input device such as a keyboard and a pointing device, or an output device such as a display device.
One or more communication interface devices. The one or more communication interface devices may be one or more homogeneous communication interface devices (e.g., one or more NICs (Network Interface Cards)) or two or more heterogeneous communication interface devices (e.g., a NIC and an HBA (Host Bus Adapter)).

In the following description, "memory" refers to one or more memory devices, typically a primary storage device. At least one of the memory devices in the memory may be a volatile memory device or a non-volatile memory device.

In the following description, a "persistent storage device" refers to one or more persistent storage devices. A persistent storage device is typically a non-volatile storage device (e.g., an auxiliary storage device), and more specifically, for example, a hard disk drive (HDD) or a solid state drive (SSD).

In the following description, "storage device" may refer to at least one memory, including memory and persistent storage device.

In the following description, a "processor" refers to one or more processor devices. The at least one processor device is typically a microprocessor device such as a CPU (Central Processing Unit), but may be other types of processor devices such as a GPU (Graphics Processing Unit). The at least one processor device may be a single-core or multi-core. The at least one processor device may be a processor core. The at least one processor device may be a broader processor device such as a hardware circuit that performs part or all of the processing (e.g., an FPGA (Field-Programmable Gate Array) or an ASIC (Application Specific Integrated Circuit)).

In the following description, functions are sometimes described using the expression "yyy part", but the function may be a computer program itself, may be realized by one or more computer programs being executed by a processor, may be realized by one or more hardware circuits (e.g., FPGA or ASIC), or may be realized by a combination of these. When a function is realized by a program being executed by a processor, the function may be at least a part of the processor, since the specified processing is performed using a storage device and/or an interface device, etc. as appropriate. Processing described with a function as the subject may be processing performed by a processor or a device having the processor. A program may be installed from a program source. The program source may be, for example, a program distribution computer or a computer-readable recording medium (e.g., a non-transitory recording medium). The description of each function is an example, and multiple functions may be combined into one function, or one function may be divided into multiple functions.

One embodiment of the present invention will be described in detail below with reference to the drawings.

In this embodiment, for the sake of concrete explanation, the power demand for each substation in a power system is taken as an example to be predicted.
(1) Overall structure

Figure 1 shows the overall configuration of the grid management system 1 in this embodiment.

The grid management system 1 is a system for calculating a predicted value of power demand at each substation at a specified date and time based on factor data including observed data of factors (and predicted data of factors) and observed data of power demand at each substation, and for controlling grid equipment 41 based on the predicted value of power demand at each substation. The grid management system 1 is composed of various devices owned by the power operator 2, the power system 4, and the external attribute information provider 5. Each of the communication paths 60, 61, and 62 is a communication path (typically a communication network) that connects various devices such as a WAN (Wide Area Network) or a LAN (Local Area Network) so that they can communicate with each other.

The factor data is created from the attribute data and includes values of one or more factors. The attribute data is data including information representing elements that can be factors, and includes system attribute data held by the power operator 2 and external attribute data provided by the external attribute information provider 5. A factor is an element that can affect the prediction target. The factor data may include one or more values (one or more factor values) for each of one or more factors. For example, when the factor data is factor data for one day, the factor data may include factor values for each factor for each hour. The one or more factors may be, for example, at least one of the following:
-Weather such as temperature, humidity, solar radiation, wind speed, and air pressure.
- Amount of energy supply from renewable sources such as solar power generation.
- Trading volumes and prices of fuels such as crude oil and natural gas.
- Subjects related to transmission lines, such as transmission capacity.
• Generator operating conditions, including generator operation or maintenance schedules.
- Calendar dates such as date, day of the week, and flag values indicating the type of day that is set arbitrarily.
-Whether or not sudden events such as typhoons or other events occur.
- Economic conditions such as the number of energy consumers, industry trends and business indicators.
- The occupancy rate of express trains, the number of passengers on board, the number of reserved seats, or the movement of people and objects, such as road traffic conditions.
- The number of communication terminals connected to the communication base station.

The power operator 2 is a business operator that controls the system equipment 41. The power operator 2 is provided with a demand forecast data management device 30 that manages data for calculating the predicted value of power demand, such as system data and external attribute data, a power demand forecast device 31 that calculates the predicted value of demand, a system control calculation device 21 that calculates the control contents of the system equipment 41, a system equipment management device 22 that manages and controls the system equipment 41, and an information input/output terminal 20 for exchanging data with these devices. The system data includes system observation data and system attribute data. The system observation data includes measurement data (e.g., measurement values (observation values) such as voltage value, current value, power value, frequency, etc.) of each point of the power system 4 measured by the measurement device 40 of the power system 4. Note that data on the power demand (forecast target) of each substation is included in the system observation data. The system attribute data includes, for example, functional specifications of the system equipment 41 of the power system 4, system configuration information indicating the connection relationship of the system equipment 41, location information indicating the geographical location of the system equipment 41, planning information indicating plans for future setting states and control values of the system equipment 41, and planning information indicating plans for future system configurations. Note that the control of the system equipment 41 is, for example, an operation of changing the tap position of a transformer to change the winding ratio of the transformer and adjust the output voltage.

The power system 4 is a facility that transmits power generated from a power generation facility to a load facility via a power distribution facility. The power system 4 is composed of a system facility 41 that supplies, distributes, and consumes power, and a measuring device 40 that measures the state of each point of the power system 4. The system facility 41 includes at least a power generation facility, a power distribution facility, and a load facility. The power generation facility is a facility that supplies power, and includes, for example, a power generation device that generates power by thermal power generation, hydroelectric power generation, nuclear power generation, solar power generation, wind power generation, etc., a power generation control device that controls the power generation device, and a storage battery that stores energy and supplies power as needed. The power distribution facility includes various facilities related to power transmission and distribution, such as overhead transmission lines, underground transmission lines, transformers, circuit breakers, disconnecting switches, and phase modifying facilities. The load facility is a facility that consumes power, and includes, for example, household electrical appliances, industrial electrical appliances, and storage batteries. The measurement device 40 is a device that measures measurement data at each point in the power system 4, and includes, for example, a voltage measurement device, a current measurement device, a power measurement device, and a frequency measurement device.

The external attribute information provider 5 is a business entity that provides past observation data and future prediction data of external attribute data. The external attribute information provider 5 is provided with an external attribute data distribution device 50 for distributing external attribute data. The external attribute data includes first information indicating an element having a certain periodicity that is the same as or similar to the periodicity of the power demand, and second information indicating an element having a small periodicity but that may have a correlation with the power demand. The first information may include, for example, information on the weather, such as temperature, humidity, cloud cover, and solar radiation, and information on the calendar, such as holidays. The second information may include event information, such as sports competitions, mobile information, such as GPS of cars and mobile terminals, consumer information, such as the operating status of factories, and information on social activities, such as industrial trends. The factor data created from the attribute data includes data created from the first information (for example, factor data of a factor having a certain periodicity) and data created from the second information (for example, factor data of a factor having a periodicity smaller than the certain periodicity).

Next, the grid management system 1 will be described with reference to Figure 2.

The grid management system 1 is composed of a measurement device 40, a grid equipment management device 22, an external attribute data distribution device 50, a power demand forecasting system 3, a grid control calculation device 21, and grid equipment 41.

The measuring device 40 measures measurement data at each point in the power system 4 and transmits the measurement data at each point to the system equipment management device 22.

The system equipment management device 22 transmits system data 3006A to the power demand forecasting system 3. The system data 3006A includes system observation data including measurement data of each point of the power system 4 received from the measurement device 40, and system attribute data held by the system equipment management device 22.

The external attribute data distribution device 50 transmits external attribute data to the power demand forecasting system 3. The external attribute data includes external attribute observation data 3007A, which is observation data of the external attribute data held by the external attribute data distribution device 50, and external attribute forecast data 3008A, which is forecast data of the external attribute data.

The power demand forecasting system 3 calculates a forecast value of the power demand of each substation based on the system data 3006A received from the system equipment management device 22 and the external attribute data received from the external attribute data distribution device 50, and transmits the forecast value of the power demand of each substation to the system control calculation device 21.

The system control calculation device 21 generates control data for the system equipment 41 based on the predicted power demand of each substation received from the power demand forecasting device 31, and transmits the control data to the system equipment management device 22.

The system equipment management device 22 controls the system equipment 41 based on the control data of the system equipment 41 received from the system control calculation device 21.

In this embodiment, the power demand forecasting system 3 is a physical computer system (one or more physical computers), but it may also be a logical computer system based on a physical computer system (e.g., a cloud computing system). The power demand forecasting system 3 is composed of a demand forecast data management device 30 and a power demand forecasting device 31.

The demand forecast data management device 30 stores the system data 3006A received from the system equipment management device 22, and the external attribute observation data 3007A and external attribute forecast data 3008A contained in the external attribute data received from the external attribute data distribution device 50, and transmits these data 3006A, 3007A, and 3008A to the power demand forecast device 31.

The power demand forecasting device 31 calculates the predicted power demand of each substation based on the data 3006A, 3007A, and 3008A received from the demand forecast data management device 30, and transmits the predicted power demand of each substation to the system control calculation device 21.

Furthermore, in the power demand prediction device 31, the data extraction unit 3106 receives the data 3006A, 3007A, and 3008A, extracts the observed value of the power demand of each prediction target included in the system data 3006A, and outputs the observed value of the power demand of each prediction target used for updating the prediction model as update prediction target data 3110A. The data extraction unit 3106 also extracts factor data of a predetermined factor from the external attribute observation data 3007A and the external attribute prediction data 3008A, and outputs the factor data used for updating the prediction model from the extracted factor data as update factor data 3111A and the factor data used for calculating the predicted value as prediction factor data 3112A. The update factor data 3111A includes the observed value of one or more factors correlated with the update prediction target data 3110A. The prediction factor data 3112A includes the predicted value of one or more factors at a predetermined date and time. The model update control unit 3107 receives the update prediction target data 3110A and the update factor data 3111A from the data extraction unit 3106, and receives the model management data 3115A from the model management unit 3108. The model update control unit 3107 judges whether or not each prediction model (specifically, the prediction model for each substation) needs to be updated, and outputs the model update control data 3114A describing the judgment result for each prediction model. The model management unit 3108 receives the model update control data 3114A, the update prediction target data 3110A, and the update factor data 3111A. The model management unit 3108 updates the model data 3116A, which is the data of each prediction model used for calculating the predicted value, and the model management data 3115A describing the update history information of each prediction model, based on the received data 3114A, 3110A, and 3111A, and outputs the updated model management data 3115A and the model data 3116A. The prediction value calculation unit 3109 receives the model data 3116A output by the model management unit 3108 and the prediction factor data 3112A output by the data extraction unit 3106, calculates a prediction value of the power demand for each prediction target at a specified date and time based on the received data 3115A and 3116A, and transmits the prediction value of the power demand to the system control calculation device 21.
(2) Internal structure

Figure 3 shows the functional configuration of each device that makes up the electricity demand forecasting system 3.

The demand forecast data management device 30 is composed of an information processing device such as a personal computer or a server computer, and includes a CPU 3001 that controls the overall operation of the demand forecast data management device 30, an input device 3002, an output device 3003, a communication device 3004, and a storage device 3005. The CPU 3001 is connected to the input device 3002, the output device 3003, the communication device 3004, and the storage device 3005.

The input device 3002 is composed of a keyboard or a mouse, and the output device 3003 is composed of a display or a printer. The communication device 3004 is composed of a NIC for connecting to a wireless LAN or a wired LAN. The storage device 3005 may include memory such as a RAM (Random Access Memory) and a ROM (Read Only Memory).

The storage device 3005 includes a system data storage area 3006, an external attribute observation data storage area 3007, and an external attribute prediction data storage area 3008.

The system data storage area 3006 stores system data 3006A. The system data 3006A is data (e.g., a database) including system observation data including data such as observed values of power demand at each substation, and system attribute data including attribute data such as system configuration information indicating the connection relationships of the system equipment 41.

The external attribute observation data storage area 3007 stores external attribute observation data 3007A. The external attribute observation data 3007A is data (e.g., a database) that includes observation data such as meteorological observation data among external attribute data such as meteorological data.

The external attribute prediction data storage area 3008 stores external attribute prediction data 3008A. The external attribute prediction data 3008A is data (e.g., a database) that includes prediction data such as weather forecast data from external attribute data such as weather data.

The power demand forecasting device 31 is composed of an information processing device such as a personal computer or a server computer, and includes a CPU 3101 that controls the overall operation of the power demand forecasting device 31, an input device 3102, an output device 3103, a communication device 3104, and a storage device 3105. The CPU 3101 is connected to the input device 3102, the output device 3103, the communication device 3104, and the storage device 3105. The CPU 3101 (and 3001) is an example of a processor. The communication device 3104 (and 3004) is an example of an interface device.

The input device 3102, the output device 3103, the communication device 3104, and the storage device 3105 may have the same configuration as the input device 3002, the output device 3003, the communication device 3004, and the storage device 3005.

The storage device 3105 stores computer programs such as a data extraction unit 3106, a model update control unit 3107, a model management unit 3108, and a prediction value calculation unit 3109. These computer programs are executed by the CPU 3101.

The data extraction unit 3106 extracts the observed value of the power demand for a specified past period of a specified substation and a specified factor of a specified substation from the system data 3006A received from the system equipment management device 22 and the external attribute data received from the external attribute data distribution device 50. The data extraction unit 3106 generates update prediction target data 3110A, update factor data 3111A, and prediction factor data 3112A for calculating a predicted value, which are used to update the prediction model.

The model update control unit 3107 generates classified update data 3113A from the update prediction target data 3110A and update factor data 3111A generated by the data extraction unit 3106, and model management data 3115A, which is update history information for each prediction model. The classified update data 3113A is data in which each data (sample data) in the update data (data combining the update prediction target data 3110A and the update factor data 3111A) is assigned a classification as either a first sample group or a second sample group. The model update control unit 3107 determines whether or not each prediction model needs to be updated based on the classified update data 3113A, and generates model update control data 3114A that records the determination result.

The model management unit 3108 updates the model data 3116A and the model management data 3115A from the model update control data 3114A generated by the model update control unit 3107 and the update prediction target data 3110A and update factor data 3111A generated by the data extraction unit 3106 (generating updated model data 3116A and model management data 3115A).

The predicted value calculation unit 3109 calculates a predicted value of the power demand for each substation at a specified date and time from the model data 3116A generated by the model management unit 3108 and the prediction factor data 3112A generated by the data extraction unit 3106, and transmits the predicted value of the power demand for each substation to the system control calculation device 21.

The storage device 3105 also stores an update prediction target data memory area 3110, an update factor data memory area 3111, a prediction factor data memory area 3112, a classified update data memory area 3113, a model update control data memory area 3114, a model management data memory area 3115, and a model data memory area 3116.

The update prediction target data storage area 3110 stores update prediction target data 3110A. The update prediction target data 3110A includes prediction target sample data for each substation (for each prediction model) to be used to update the prediction model. For one substation (prediction model), the prediction target sample data is data obtained by dividing the observation data of the power demand of the substation in a predetermined arbitrary time range into a predetermined time unit, and is data (e.g., a database) including an identifier of the prediction model corresponding to the substation and a value representing the observation date and time. For example, the update prediction target data 3110A is data in which the power demand data for the past year of each substation is divided by day, and an identifier that identifies the substation and the observation date and time are assigned to each data.

The update factor data memory area 3111 stores update factor data 3111A. The update factor data 3111A is data (e.g., a database) including factor values of one or more factors corresponding to the update prediction target data 3110A (each prediction target sample data to be used for updating the prediction model).

The prediction factor data storage area 3112 stores prediction factor data 3112A. The prediction factor data 3112A is data (e.g., a database) that includes one or more factors for a predetermined arbitrary date and time, for use in calculating the predicted value of the power demand of each substation.

The classified update data memory area 3113 stores classified update data 3113A. The classified update data 3113A is data (e.g., a database) in which each sample data of the update data (data obtained by combining the corresponding data of update prediction target data 3110A and update factor data 3111A) is assigned a classification (the classification of the sample, which is either a first sample group or a second sample group) based on the observation date and time of the sample represented by the sample data.

The model update control data storage area 3114 stores model update control data 3114A. The model update control data 3114A is data (e.g., a database) that includes an update control value that indicates whether each prediction model needs to be updated.

The model management data storage area 3115 stores model management data 3115A. The model management data 3115A is data (e.g., a database) that includes information about the samples used for updating, such as the last update date and time of each prediction model and the observation date and time range of the samples.

The model data storage area 3116 stores model data 3116A. The model data 3116A is data (e.g., a database) including data (e.g., mathematical expressions and parameters) of each prediction model obtained by identification of the prediction model.
(3) Details of each processing unit (program)

Hereinafter, the details of each processing unit of the power demand forecasting device 31 will be described with reference to the drawings.
(3-1) Data Extraction Unit 3106

Figure 4 shows the configuration of the data extraction unit 3106.

The data extraction unit 3106 is composed of an update data extraction unit 31061 and a prediction data extraction unit 31062.

The update data extraction unit 31061 extracts the observed values of the power demand for each substation for a predetermined time range (e.g., one day) from the system data 3006A, and divides the extracted observed values into a predetermined time granularity (e.g., one hour). The update data extraction unit 31061 generates prediction target sample data for each substation (prediction model), which is data obtained by adding an identifier of the prediction model and an observation date and time to the divided data (e.g., data including the observed values of the power demand for each hour in one day). The update data extraction unit 31061 generates update prediction target data 3110A, which is data including one or more prediction target sample data. The update data extraction unit 31061 may use values obtained by performing processing such as standardization and normalization on the observed values of the power demand extracted from the system data 3006A to generate the update prediction target data 3110A.

Figure 7 shows the configuration of the update prediction target data 3110A. Specifically, for example, the update prediction target data 3110A is a table, and has a row (entry) for each pair of a prediction model and an observation date and time. The row has data such as a model ID 3110A1 representing an identifier (model ID) of the prediction model, a sample observation date and time 3110A2 representing the observation date and time of the prediction target sample, and a power demand observation value 3110A3 representing the observation value of the power demand for each time point in a specified time range. Therefore, in the case of Figure 7, each row contains one prediction target sample data in which the observation value of the power demand is divided by the time granularity of one day. The data contained in the first row is the prediction target sample data used to update the prediction model of model ID "M0001", and the observation value of the power demand of the substation corresponding to model ID "M0001" on December 10, 2020 is "0.9" at 0:00 and "0.8" at 0:05. In the example of Figure 7, the unit of the observation date and time is the date for simplicity, but it is not limited to the date, and the unit of the observation date and time may include more detailed information such as hours, minutes, and seconds.

Next, the update data extraction unit 31061 extracts observed values of one or more specified factors within a specified time range from the system data 3006A and the external attribute observation data 3007A, and divides them at a specified time granularity so as to correspond to each prediction target sample of the update prediction target data 3110A. The update data extraction unit 31061 generates factor sample data, which is data in which the same prediction model identifier and observation date and time as each prediction target sample data are assigned to each divided data, and generates update factor data 3111A, which is data including one or more factor sample data. Note that the update data extraction unit 31061 may use values obtained by performing processing such as standardization and normalization on the observed values of the factors extracted from the system data 3006A and the external attribute observation data 3007A to generate the update factor data 3111A.

Figure 8 shows the configuration of the update factor data 3111A. Specifically, for example, the update factor data 3111A is a table, and has a row (entry) for each pair of a prediction model and an observation date and time. The row has data such as a model ID 3111A1 representing an identifier of the prediction model, a sample observation date and time 3111A2 representing the observation date and time of the factor sample, and a factor observation value 3111A3 including the observation value of each factor. Therefore, in the case of Figure 8, each row contains one factor sample data in which the observation value of each factor is divided by a time granularity of one day. The data contained in the first row is a factor sample with a model ID of "M0001" and a sample observation date and time of "2020-12-10", and therefore is the factor sample data corresponding to the prediction target sample data shown in the first row of Figure 7. The observation values of each factor are "0.4" for factor A and "0.1" for factor B. In the example of Figure 8, for simplicity, the unit of the observation date and time of the factor sample is the date, but it is not limited to the date, and for example, the unit of the observation date and time of the factor sample may include more detailed units such as hours, minutes, and seconds.

Next, the prediction data extraction unit 31062 extracts predicted values of one or more specified factors for a specified prediction date for each prediction model from the system data 3006A and the external attribute prediction data 3008A. The prediction data extraction unit 31062 generates prediction factor sample data in which the prediction values of the extracted factors are assigned an identifier of the prediction model and a prediction target date and time, and generates prediction factor data 3112A, which is data including one or more prediction factor samples. Note that the prediction data extraction unit 31062 may use values obtained by performing processing such as standardization and normalization on the predicted values of the factors extracted from the system data 3006A and the external attribute prediction data 3008A to generate prediction factor data 3112A.

FIG. 9 shows a conceptual diagram of the prediction factor data 3112A. Specifically, for example, the prediction factor data 3112A is a table, and has a row (entry) for each pair of a prediction model and an observation date and time. The row has data such as a model ID 3112A1 representing an identifier (model ID) of the prediction model, a prediction target date and time 3112A2 representing the prediction target date and time, and a factor predicted value 3112A3 including the predicted value of each factor. Therefore, in the case of FIG. 9, each row is prediction factor sample data for a predetermined prediction target date and time of each prediction model, and the first row is prediction factor sample data used to calculate a predicted value by a prediction model with model ID "M0001", and the predicted values of each factor on the prediction target date, December 11, 2020, are "0.3" for factor A and "0.2" for factor B. Note that in the example of FIG. 9, for simplicity, the unit of the prediction target date and time of the prediction factor sample is the date, but it is not limited to the date, and may include more details such as hours, minutes, and seconds, for example.
(3-2) Model update control unit 3107

Figure 5 shows the configuration of the model update control unit 3107.

The model update control unit 3107 is composed of an update data classification unit 31071, a dissimilarity calculation unit 31072, and a model update determination unit 31073.

First, the update data classification unit 31071 generates update data by combining data of each factor sample in the update factor data 3111A for each prediction target sample in the update prediction target data 3110A. Next, the update data classification unit 31071 generates classified update data 3113A in which each sample data in the update data is classified as a first sample group or a second sample group based on information on the samples used in the previous identification of each prediction model in the model management data 3115A and information on the observation date and time of each sample in the update data. Note that the "first sample group" refers to samples in the sample group used in the previous identification of the prediction model, and the "second sample group" refers to samples in the sample group newly acquired after the date and time of the previous identification of the prediction model.

Figure 10 shows the configuration of the classified update data 3113A. Specifically, for example, the classified update data 3113A is a table, and has a row (entry) for each pair of a prediction model and an observation date and time. The row has data such as a model ID 3113A1 representing an identifier of the prediction model, a classification 3113A2 representing a classification of sample data, a sample observation date and time 3113A3 representing the observation date and time of the sample, a factor observation value 3113A4 including the observation value of each factor, and a power demand observation value 3113A5 including the observation value of the power demand at each time. Note that the value "1" of the classification 3113A2 means the first sample group, and the value "2" of the classification 3113A2 means the second sample group. The factor observation value 3113A4 may be the same as the factor observation value 3111A3 described above, and the power demand observation value 3113A5 may be the same as the power demand observation value 3110A3 described above.

Figure 11 shows the configuration of model management data 3115A. Specifically, for example, model management data 3115A is a table, and has a row (entry) for each prediction model. The row has data such as model ID 3115A1 indicating the identifier of the prediction model, identification date and time 3115A2 indicating the identification date and time of the prediction model, and last used sample date and time 3115A3 indicating the observation date and time of the last sample used for identification.

Therefore, in the example of Figure 10, each row is a sample combining each prediction target sample in the update prediction target data 3110A and each factor sample in the update factor data 3111A. The first row of the update data in Figure 10 contains sample data used to update the prediction model with model ID "M0001", and the sample observation date and time is "2020-12-10". The update data classification unit 31071 references the model ID 3115A1 "M0001" and the latest used sample date and time 3115A3 "2020-12-9" in the first row of the model management data 3115A in Figure 11. The sample observation date and time 3113A3 in the first row of the update data in Figure 10 is after the date and time represented by the latest used sample date and time 3115A3. Therefore, the update data classification unit 31071 classifies the sample data in the first row of the update data 3113A in FIG. 10 into the second sample group, and sets the value of the classification 3113A2 of the row to "2". Also, in the second row of the update data in FIG. 10, the sample observation date and time 3113A3 of the model ID "M0001" is "2020-12-09", which is not after the latest used sample date and time 3115A3 of the model ID "M0001", "2020-12-09". Therefore, the update data classification unit 31071 classifies the sample data in the second row of the update data 3113A in FIG. 10 into the first sample group, and sets the value of the classification 3113A2 of the row to "1".

The difference degree calculation unit 31072 calculates the difference degree, which is an index value indicating the difference between the samples classified into the first sample group and the samples classified into the second sample group of each prediction model of the classified update data 3113A. The difference degree is an index value indicating the difference between the two sample groups. Note that the calculation of the difference degree may use a representative value of the index value indicating the difference between the individual samples of the first sample group and the second sample group. The representative value of the index value indicating the difference between the individual samples includes the average value, maximum value, and minimum value of the index value indicating the difference between the individual samples, the average value, maximum value, and minimum value of the index value indicating the difference between the representative sample and the individual samples, and the index value indicating the difference between the representative sample and the representative sample. The representative sample is, for example, an average sample obtained by averaging the values of the observed values of all the samples constituting each sample group, or the latest sample with the latest sample observation date and time. Note that a known method such as a distance or similarity calculation method may be applied to the calculation of the index value indicating the difference between the individual samples. Examples of well-known methods include methods for calculating distances in a sample space, such as Euclidean distance and Manhattan distance, and methods for calculating similarity between samples, such as the inner product.

Using the classified update data 3113A shown in FIG. 10, an example of calculating the degree of dissimilarity when using the Euclidean distance between the average sample of the first sample group and the latest sample of the second sample group will be described. For simplicity, only the degree of dissimilarity of model ID "M0001" is calculated, and the second sample group is composed of only the samples in the first row, and the first sample group is composed of only the samples in the second and third rows. In addition, the factor observation value 3113A4 is composed of only the observation values of "Factor A", "Factor B", and "Factor C", and the power demand observation value 3113A5 is composed of only the observation values of "00:00", "00:05", and "00:10".

The average sample of the first sample group is the average of the values of each column of the factor observation values 3113A4 and the power demand observation values 3113A5 of the samples that make up the first sample group. For example, the average value of the column of factor A is "(0.4 + 0.3) / 2 = 0.35", so the value of factor A of the average sample of the first sample group is "0.35". By calculating the average value of each column in the same way, the values of each column of the average sample are obtained as "0.35, 0.15, 0.75, 0.95, 0.9, 0.9" (the composition is "Factor A", "Factor B", "Factor C", "00:00", "00:05", "00:10"). The Euclidean distance between the average sample and the samples of the second sample group is the squared average of the squared differences of the values of each column of the two samples. The differences in the values of each column between the average sample "0.35, 0.15, 0.75, 0.95, 0.9, 0.9" and the samples "0.4, 0.1, 0.9, 0.9, 0.8, 0.7" of the second sample group are "-0.05, 0.05, -0.15, 0.05, 0.1, 0.2", and the square of the average of the squared differences in the values of each column is "0.12". As described above, the degree of difference between the first sample group consisting of the samples in the second and third rows of Figure 10 and the second sample group consisting of the sample in the first row of Figure 10 is "0.12".

The model update determination unit 31073 determines whether an update is necessary based on the result of comparing the degree of difference of each prediction model with a predetermined threshold, and generates model update control data 3114A indicating the determination result. The determination of whether an update is necessary may be made by determining whether the degree of difference is equal to or greater than the threshold. A prediction model whose degree of difference is equal to or greater than the threshold is determined to require updating, and a prediction model whose degree of difference is less than the threshold is determined not to require updating. The threshold value is set in advance for each prediction model by input from the information input/output terminal 20, for example.

FIG. 12 shows the configuration of the model update control data 3114A. Specifically, for example, the model update control data 3114A is a table, and has a row (entry) for each prediction model. The row has data such as a model ID 3114A1 representing the identifier of the prediction model, a difference degree 3114A2 representing the difference degree, a threshold value 3114A3 representing the threshold value of the difference degree, and a judgment 3114A4 representing the judgment result of whether or not the prediction model needs to be updated. Therefore, in the case of FIG. 12, according to the first row, for the prediction model with model ID "M0001", the difference degree is "0.12" which is equal to or greater than the threshold value "0.1". Therefore, it is determined that the prediction model needs to be updated, and the value of the judgment 3114A4 is "necessary". Also, according to the second row, for the prediction model with model ID "M0002", the difference degree "0.3" is less than the threshold value "0.6". Therefore, it is determined that the prediction model does not need to be updated, and the value of the judgment 3114A4 is "not required".
(3-3) Model management unit 3108

Figure 6 shows the configuration of the model management unit 3108.

The model management unit 3108 is composed of a model identification unit 31081 and a model storage unit 31082.

First, the model identification unit 31081 identifies a prediction model for which a model update is determined to be necessary based on the model update control data 3114A, using samples of each prediction model included in the update prediction target data 3110A and the update factor data 3111A.

Therefore, according to the model update control data 3114A illustrated in FIG. 12, the prediction model with model ID "M0001" and the prediction model with model ID "M0003" are identified because judgment 3114A4 is "necessary." The prediction model with model ID "M0002" is not identified because judgment 3114A4 is "unnecessary." In addition, for example, the prediction model with model ID "M0001" is identified using the prediction target sample data with model ID "M0001" in the update prediction target data 3110A in FIG. 8 and the factor sample data with model ID "M0001" in the update factor data 3111A in FIG. 9.

Then, the model storage unit 31082 receives data on each prediction model identified by the model identification unit 31081, and replaces (updates) model data 3116A such as parameters of the identified prediction model with the existing model data 3116A. Furthermore, the model storage unit 31082 replaces (updates) model management data 3115A such as the observation date and time of the specimen used to identify the identified prediction model and the identification date and time of the prediction model with the existing model management data 3115A based on the updated prediction target data 3110A and the updated factor data 3111A.

FIG. 13 shows the configuration of the model data 3116A. Specifically, for example, the model data 3116A is a table, and has a row (entry) for each prediction model. The row has data such as a model ID 3116A1 representing an identifier of the prediction model, a formula 3116A2 representing a formula (an example of a model configuration) of the prediction model, and a parameter 3116A3 including a parameter value of each parameter item of the formula. Therefore, in the example of FIG. 10, the model storage unit 31082 receives data of the prediction model with the model ID "M0001" and the model ID "M0003", and rewrites the values of the formula 3116A2 and the parameter 3116A3 of the model ID "M0001" and the model ID "M0003" of the model data 3116A of FIG. 13. Furthermore, the model storage unit 31082 rewrites the values of the identification date and time 3115A2 and the latest specimen use date and time 3115A3 of the model ID "M0001" and the model ID "M0003" of the model management data 3115A in FIG.
(3-4) Prediction value calculation unit 3109

The prediction value calculation unit 3109 receives the model data 3116A and the prediction factor data 3112A updated by the model management unit 3108, inputs the factor prediction value 3112A3 in the prediction factor data 3112A to the prediction model represented by the model data 3116A, calculates the predicted value of the power demand for a specified date and time at the substation corresponding to the prediction model, and transmits the predicted value of the power demand of each substation to the system control calculation device 21.

This completes the power demand prediction process of the power demand prediction device 31.

Then, based on the received forecasted power demand of each substation, the system control calculation device 21 generates an optimal voltage setting value at each point in the power system at a specified future time, generates control instructions for each system equipment 41 to realize the optimal voltage setting value, and transmits control data including the control instructions for each system equipment 41 to the system equipment management device 22. Note that the optimal voltage setting value is, for example, a voltage value that maximizes the amount of reduction possible in transmission loss throughout the system, or a voltage value that suppresses failures of each device in the system equipment 41. Note that the control instructions for each system equipment 41 are, for example, tap switching of a transformer or increasing or suppressing reactive power output by a phase modifying device.

Next, the system equipment management device 22 operates the system equipment 41, for example, by changing the transformer tap from No. 1 to No. 2, based on the received control data of the system equipment 41. The system equipment 41 changes its operation due to the operation, and a physical change occurs, for example, the output voltage of the transformer becomes larger than before the change.
(4) Flowchart

Figure 14 shows the procedure for the power demand forecasting process. This process begins when the power demand forecasting device 31 receives an input operation from the power operator 2, and the power demand forecasting device 31 executes steps S101 to S106.

In practice, the processing is performed based on various computer programs stored in the CPU 3101 and storage device 3105 of the power demand forecasting device 31. For ease of explanation, the processing is described as being performed by the various computer programs possessed by the demand type management device.

First, the data extraction unit 3106 extracts the observed value of the power demand of a specified substation from the system data 3006A, and generates update prediction target data 3110A. Furthermore, the data extraction unit 3106 extracts the observed value and predicted value of a specified factor from the system data 3006A, the external attribute observation data 3007A, and the external attribute prediction data 3008A, and generates update factor data 3111A and prediction factor data 3112A (S101).

Next, the model update control unit 3107 determines whether or not each prediction model that predicts the power demand of each substation needs to be updated based on the update prediction target data 3110A, the update factor data 3111A, and the model management data 3115A, and generates model update control data 3114A indicating the determination result (S102).

Next, the model management unit 3108 determines whether or not there is a prediction model to be updated based on the judgment 3114A4 of each prediction model in the model update control data 3114A (S103). If the judgment result of S103 is true (S103: Yes), the model management unit 3108 identifies each prediction model to be updated using the update prediction target data 3110A and the update factor data 3111A, and updates the model management data 3115A and the model data 3116A (S104).

Furthermore, the model management unit 3108 transmits the model data 3116A to the prediction value calculation unit 3109 (S105).

Next, the prediction value calculation unit 3109 inputs the prediction values of the factors in the prediction factor data 3112A into the data of each prediction model in the model data 3116A, and calculates the prediction value of the power demand of each substation (S106).

With the above processing, the power demand forecasting processing in this embodiment is completed.
(5) Effects of this embodiment

Figures 15A and 15B show the concept of determining whether or not an update is necessary based on the degree of difference. Figure 16A shows the model update timing and update processing load for a comparative example. Figure 16B shows the model update timing and update processing load for this embodiment.

In Figures 15A and 15B, the plots on the graphs with factors on the horizontal axis and power demand on the vertical axis show sample data constituting the first sample group of a certain prediction model, or sample data constituting the second sample group. In this paragraph, "sample data" refers to data that includes factor observation value 3113A4 and power demand observation value 3113A5 in one entry. The result of determining whether or not a model needs to be updated differs depending on the relative positions of the sample data in the first sample group and the second sample group. Note that here, there is only one sample data constituting the second sample group.

According to FIG. 15A, the first sample group and the second sample group are close in position on the graph, i.e., the degree of difference is small. This means that there is little change in the response characteristics between the first sample group, which are the samples used in the previous update of the prediction model, and the second sample group, which are the samples obtained after the previous update of the prediction model. For this reason, it is determined that updating the prediction model is not necessary.

On the other hand, according to FIG. 15B, the first sample group and the second sample group are far apart in position on the graph, i.e., the degree of difference is large. This means that there is a large change in the response characteristics between the first sample group, which are the samples used in the previous update of the prediction model, and the second sample group, which are the samples obtained after the previous update of the prediction model. For this reason, a determination result is obtained that the prediction model needs to be updated to follow the change in the response characteristics.

In Figures 16A and 16B, the horizontal axis of the graph indicates the update timing of the prediction model, the vertical axis of the graph indicates the processing load of the update process of the prediction model, and the dotted line indicates the upper limit of the processing load that can be processed within a specified time.

According to FIG. 16A, all prediction models are updated at each update timing. Because all prediction models are constantly updated, the processing load exceeds the processing load that can be processed within the time regardless of the update timing, and the update of all prediction models, including prediction models whose response characteristics have changed, is not completed within the time.

On the other hand, according to FIG. 16B, forecast models are updated only for those that detect changes in the response characteristics of factors and demand (for which the degree of difference is equal to or greater than a threshold). As a result, it is possible to complete the update of forecast models whose response characteristics have changed within the allotted time.

Furthermore, we will explain why the model updating method according to this embodiment can suppress deterioration in prediction accuracy.

A prediction model that needs to be updated is not a sample that shows an abnormal value, but a prediction model in which sample data that shows a different trend from previous ones within the normal range appears, that is, a prediction model that indicates the possibility of a change in response characteristics. As shown in FIG. 17A, when sample data that shows a change in the trend of the response characteristics within the normal range is obtained from the second sample group, the response characteristics can be tracked by updating prediction model 1700a to new prediction model 1700b, thereby improving prediction accuracy.

On the other hand, a prediction model that does not require updating is a model in which sample data showing a similar trend to the first sample group, which is the sample data observed before the update, or sample data showing an abnormal value (outlier) is obtained. When the former sample data (sample data showing a similar trend to the first sample group, which is the sample observed before the update) is obtained, as shown in FIG. 17B, even if the prediction model 1700a is updated to a new prediction model 1701b, the prediction model does not change substantially, and therefore the prediction accuracy does not improve. When the latter sample data (sample data showing an abnormal value (outlier)) is obtained, as shown in FIG. 17C, when the prediction model 1700a is updated to a new prediction model 1702b, the prediction model will follow the abnormal value, and therefore the prediction accuracy will deteriorate.

As described above, a prediction model that is likely to improve in prediction accuracy is determined to require updating (i.e., the prediction model is updated), and a prediction model that is likely to have no change in prediction accuracy or to deteriorate is determined not to require updating (i.e., the prediction model is not updated). As a result, deterioration in prediction accuracy is suppressed.

In addition, by completing the update of prediction models whose response characteristics have changed within a certain time period and suppressing deterioration in the prediction accuracy of many prediction models, highly accurate prediction values for the power demand of many substations can be calculated, and the system equipment 41 of the power system 4 can be controlled over a wide area based on the predicted values of the power demand of many substations, enabling efficient operation of the system equipment 41, such as reducing transmission losses in the power system 4.

The above explanation can be summarized as follows, for example:

The prediction system (e.g., power demand prediction system 3) includes an interface device (e.g., communication device 3004) that accepts input of prediction target observation data and factor observation data, a storage device (e.g., storage device 3005) in which the prediction target observation data and factor observation data are stored, and a processor (e.g., CPU 3101) that is connected to the interface device and the storage device and updates one or more prediction models corresponding to one or more prediction targets that are determined to require updating.

For each prediction target, the prediction model (e.g., the model represented by formula 3116A2) is a model that takes as input factor prediction values (e.g., factor prediction values 3112A3), which are data including prediction values for a future period for each of one or more factors that may affect one or more prediction targets, and outputs predicted values of observed values obtained for the prediction target in the future period. The factor prediction values 3112A3 may be data extracted from external attribute prediction data 3008A.

The prediction target observation data includes, for each of one or more prediction targets, time series data of past observation values obtained for the prediction target. The prediction target observation data may be, for example, update prediction target data 3110A, or may be data extracted from external attribute observation data 3007A.

The factor observation data includes, for each of one or more factors, time series data of past observation values obtained for that factor. The factor observation data may be, for example, update factor data 3111A, or data extracted from external attribute observation data 3007A.

The processor generates update data from the prediction target observation data and the factor observation data. The update data includes sample data for each observation period for one or more prediction targets. For each prediction target, for each observation period, the sample data includes one or more observation values (e.g., power demand observation value 3113A5) obtained during the observation period for the prediction target, and observation values (e.g., factor observation value 3113A4) obtained during the observation period for each of one or more factors.

The processor calculates the degree of difference between the first sample group and the second sample group for each prediction target. For each prediction target, the first sample group is one or more sample data used to identify a prediction model corresponding to the prediction target (e.g., sample data in which the time represented by the sample observation date and time 3113A3 is the same as or earlier than the time represented by the last used sample date and time 3115A3). The second sample group is one or more sample data that has not yet been used to identify a prediction model corresponding to the prediction target (e.g., sample data in which the time represented by the sample observation date and time 3113A3 is later than the time represented by the last used sample date and time 3115A3).

The processor performs an update necessity determination for each prediction target, which is a determination of whether the degree of difference is equal to or greater than a threshold, and outputs the result of the update necessity determination. A prediction model for which the result of the update necessity determination is true is a prediction model that needs to be updated. A prediction model for which the result of the update necessity determination is false is a prediction model that does not need to be updated.

A prediction target whose degree of difference is equal to or greater than the threshold is a prediction target whose response characteristics have changed. Such a prediction target does not require learning of a prediction model, and can be identified from the relationship between the degree of difference between the first sample group and the second sample group and the threshold. In other words, the prediction model for a prediction target whose response characteristics have changed can be determined with low load.

In addition, when there are multiple prediction targets, the prediction models to be updated are narrowed down to those determined to need updating (prediction models corresponding to prediction targets whose response characteristics have changed), increasing the likelihood that the update process can be completed within the specified time.

In addition, even if the second sample group contains sample data that includes outliers, the degree of difference between the first sample group and the second sample group does not necessarily exceed the threshold (for example, the second sample group contains more sample data that does not contain outliers than sample data that includes outliers). This reduces the possibility of unnecessary updates to the prediction model.

For each prediction target, the determination of whether or not an update is necessary may be made by determining whether or not the degree of difference is equal to or greater than a first threshold and less than a second threshold (second threshold > first threshold).

Furthermore, for at least one prediction target among the one or more prediction targets, the processor may calculate a representative value of an index value indicating the difference between each sample data of the first sample group and each sample data of the second sample group, and calculate the degree of difference based on the representative value. This is expected to make the degree of difference more appropriate.

In addition, for at least one prediction target among the one or more prediction targets, the processor calculates the degree of dissimilarity based on the distance or similarity between the first sample group and the second sample group. This is expected to result in a more appropriate degree of dissimilarity.

In the above embodiment, the one or more factors include at least one of weather and energy supply from renewable energy, and each of the one or more prediction targets is the power demand of a substation. The power demand of a substation has a relationship specific to the power industry with weather such as temperature and solar radiation (and/or renewable energy such as solar power generation), and the sample data includes observed values having such a relationship (observed values of power demand and observed values of factors), so that it is possible to identify a prediction target whose response characteristics have changed and update a prediction model corresponding to the prediction target, that is, to realize a model update that follows the change in response characteristics.
(6) Modifications (other embodiments)
(6-1) Modification of the prediction target

Instead of electricity demand (e.g., electricity consumption), the prediction target may be any of the following: demand (e.g., consumption) for energy other than electricity (e.g., gas or water), output (e.g., power generation) of energy such as solar power generation or wind power generation, trading volume or trading price of energy traded at a wholesale exchange, communication volume measured at a communication base station, and location information history of a mobile object such as an automobile. The time series data may include energy demand such as gas or heat, commodity demand, industrial data such as factory production volume, and meteorological data such as temperature and solar radiation. Time series data of types other than electricity demand values may include many targets such as factories and regions, and therefore falls within the scope of application of the present technology, making it possible to perform predictions while suppressing deterioration of the prediction accuracy of many prediction models.
(6-2) Modification of the prediction target

The processor may calculate the similarity between prediction targets based on the observation data of the prediction targets, classify the multiple prediction targets into a predetermined number of groups so that prediction targets with high similarity based on the similarity are in the same group, and for each group, integrate the sample data of all prediction targets belonging to that group. In this case, in the update data, for each of one or more prediction targets, the sample data for each observation period may be integrated sample data obtained for the group to which the prediction target belongs. Specifically, for example, it may be as follows.

The prediction target may be a power demand obtained by integrating the power demands of multiple substations. The integration means converting multiple data series into one data series by statistical processing such as summation or averaging. A method for selecting multiple substations to be integrated is to select substations that tend to be similar. As a method for selecting substations that tend to be similar, for example, a method for grouping highly similar substations by clustering substations using features such as the average daily demand transition curve and the maximum daily power value of each substation, and selecting substations classified into the same group as targets for integration is available. A known method for clustering may be used, for example, hierarchical clustering such as the Ward method, or non-hierarchical clustering such as k-means. This reduces the number of prediction targets, thereby reducing the processing load, and reduces the deterioration of prediction accuracy due to the law of large numbers.
(6-3) Modification of the Data Extraction Unit 3106

The processor may generate a factor prediction value for each prediction target based on factor prediction data (e.g., data extracted from external attribute prediction data 3008A) including a prediction value for each of one or more factors, and the factor observation data. Specifically, for example, the following may be used.

The prediction data extraction unit 31062 of the data extraction unit 3106 may generate prediction factor data 3112A based on the observation value of a predetermined factor extracted from the system data 3006A and the external attribute observation data 3007A. The predetermined factor here is the setting value of the system equipment 41 several hours ago, the power demand value several hours ago, the temperature several hours ago, etc. That is, the prediction data extraction unit 31062 extracts a prediction value of one or more predetermined factors for a predetermined prediction target date for each prediction model from the system data 3006A, the external attribute observation data 3007A, and the external attribute prediction data 3008A, and shapes the prediction value of the extracted factor into a prediction factor sample by adding an identifier of the prediction model and a prediction target date and time, thereby generating prediction factor data 3112A that is data including one or more prediction factor samples. The factors may include factors that affect the prediction target with a time delay, such as the temperature several hours ago affecting the power demand. When making predictions using factors that affect the prediction target with a time lag, the observed values of the factors are available at the time of calculating the predicted value of the prediction target, and the predicted value of the prediction target can be calculated based on the observed values of the factors. This makes it possible to reduce the influence of prediction errors in the predicted values of the factors, thereby improving the prediction accuracy.
(6-4) Modifications of the Data Extraction Unit 3106

For at least one prediction target among one or more prediction targets, the processor may detect unsuitable sample data, which is sample data having an abnormal value, from the update data, and may implement a process to deal with the unsuitable sample data (e.g., a process to remove the unsuitable sample data from the update data). Specifically, for example, the process may be as follows.

The data extraction unit 3106 may further include an unsuitable sample processing unit 31063 as shown in FIG. 18. The unsuitable sample processing unit 31063 receives the update prediction target data 3110A and the update factor data 3111A output by the update data extraction unit 31061, and detects unsuitable samples in the update prediction target data 3110A and the update factor data 3111A. The unsuitable sample processing unit 31063 performs a countermeasure process such as deleting the detected unsuitable samples in the update prediction target data 3110A and the update factor data 3111A, and outputs the update prediction target data 3110A and the update factor data 3111A that have been subjected to the countermeasure process. Note that an unsuitable sample is a prediction target sample and a factor sample that have an abnormal value compared to other samples, and is, for example, a prediction target sample that has a power demand value that is N times (N is an integer of 2 or more) larger than the average value of the power demand values of other prediction target samples having the same prediction model identifier. One method for detecting unsuitable samples is to compare each sample with a predetermined threshold and detect samples with values equal to or greater than the threshold as unsuitable samples. This makes it possible to avoid unnecessary model updates due to temporary abnormalities in observed data and to improve prediction accuracy.
(6-5) Modification of the model update determination unit 31073

For at least one prediction target among one or more prediction targets, the processor may identify a threshold value at which the performance index value, which is a value based on at least one of the processing load of updating the prediction model corresponding to the prediction target and the prediction accuracy of the prediction model, is optimal, and set the identified threshold value as the difference degree threshold. Specifically, for example, it may be as follows.

The model update determination unit 31073 may calculate a threshold value of the degree of difference in advance by a parameter calculation method based on a predetermined performance index value, and may determine whether or not each prediction model needs to be updated based on the calculated threshold value. As a parameter calculation method, a grid search method in which all the results of performance index values for several threshold values are calculated in advance and a threshold value when the performance index value is the best among the calculated values may be calculated, or a known method such as Bayesian optimization in which a threshold value that is likely to be the best performance index value is sequentially and probabilistically searched and determined may be used. The performance index value is an index value indicating the processing load of updating the prediction model and the prediction accuracy of the prediction model, and is, for example, the sum of the processing time required for the update process of each prediction model or the average value of the prediction error, which is the difference between the predicted value and the observed value of the power demand calculated by each prediction model. As a result, it is possible to accurately determine whether or not the prediction model needs to be updated, thereby suppressing deterioration of the prediction accuracy.
(6-6) Modification of the model update determination unit 31073

For at least one prediction target among the one or more prediction targets, the processor may set a value obtained by multiplying the standard deviation of the distance between sample data in the first sample group by a predetermined multiple as the threshold value of the degree of dissimilarity. Specifically, for example, it may be as follows.

The model update determination unit 31073 may determine whether or not each prediction model needs to be updated by calculating the standard deviation of the intra-group sample distances of the first sample group of each prediction model, and using a value obtained by multiplying the standard deviation by a predetermined multiple as a threshold value. The standard deviation of the intra-group sample distances of the first sample group is obtained by focusing on each sample of the first sample, excluding the sample from the first sample group as the other sample group, calculating the distance between the sample and the other sample group for each sample, and obtaining the standard deviation of the data series of the distances calculated for the number of samples. The predetermined multiple may be a value determined in advance by input from the information input/output terminal 20 or a value determined in advance by a parameter search method. As a parameter calculation method, a grid search method in which all the results of performance index values for several threshold values are calculated in advance, and a threshold value when the performance index value is the best among the calculated values is calculated, or a known method such as Bayesian optimization in which a threshold value that is likely to be the best performance index value is sequentially and probabilistically searched for and determined may be used. The performance index value may be the same as the performance index value described in (6-5). This makes it possible to suppress deterioration of prediction accuracy by accurately determining whether or not the prediction model needs to be updated.
(6-7) Modification of the model update determination unit 31073

The processor may determine whether a re-evaluation start condition is satisfied, in which the number of prediction models determined to require updating exceeds the upper model limit, which is the number of prediction models that can be updated within a specified time. If the re-evaluation start condition is satisfied, the processor may output data representing prediction models that are determined to require updating and that are equal to or less than the upper model limit and have a relatively high degree of difference, as the result of the update necessity determination. Specifically, for example, the following may be used.

The model update determination unit 31073 generates a first determination result in which the necessity of updating is determined based on the result of comparing the degree of difference between the model and a threshold value, and if the first determination result satisfies a predetermined re-determination start condition, generates a second determination result in which the necessity of updating each model is determined again based on the first determination result and other update determination criteria, and generates model update control data 3114A indicating the new second determination result. The re-determination start condition is, for example, a condition in which the number of prediction models determined to require updating in the first determination result exceeds the number of prediction models that can be updated within the time. The number of prediction models that can be updated within the time is determined in advance based on the processing load that can be processed within the time. The other update determination criteria are, for example, a determination criterion in which, among the prediction models determined to require updating in the first determination result, the number of prediction models that can be updated within the time is determined to require updating in order of the degree of difference, and the remaining prediction models and prediction models determined not to require updating in the first determination result are determined not to require updating. As a result, even if the response characteristics of a large number of prediction targets change, it is possible to avoid a significant deterioration in prediction accuracy by preferentially updating prediction models that particularly require updating.
(6-8) Modification of the model update determination unit 31073

The processor may determine the presence or absence of an inappropriate prediction target, which is a prediction target for which a predicted value output using a prediction model is predicted not to satisfy a condition, based on at least one of the factor prediction value and the factor prediction data including the predicted values of one or more factors, and may determine whether a re-evaluation start condition that an inappropriate prediction target exists is satisfied. If the re-evaluation start condition is satisfied, data representing the prediction models excluding the prediction models corresponding to the inappropriate prediction targets from among the prediction models determined to require updating may be output as a result of the update necessity determination. Specifically, for example, it may be as follows.

The model update determination unit 31073 may generate a first determination result in which a determination is made as to whether or not an update is necessary based on a comparison result between the degree of difference between the model and a threshold value, and calculate a prediction result for determination to be used for determining a predetermined re-determination start condition described later. The model update determination unit 31073 may determine whether or not a predetermined re-determination start condition is satisfied based on the prediction result for determination, and if the re-determination start condition is satisfied, may generate a second determination result in which a determination is made as to whether or not an update is necessary for each model again based on the first determination result and other update determination criteria, and generate model update control data 3114A indicating the new second determination result. The prediction result for determination is, for example, a prediction result of whether or not the power demand of each substation will deviate from a predetermined value range due to the influence of the amount of power generated by photovoltaic power generation at a predetermined prediction target date and time, which is predicted by the model update determination unit 31073 based on a predicted value of a factor such as the amount of solar radiation included in the prediction factor data 3112A. Specifically, for example, if the value of the solar radiation amount included in the prediction factor data 3112A is less than a predetermined value, it is predicted that the power demand at the predetermined prediction target date and time will not deviate from the predetermined value range, and if the value is equal to or greater than the predetermined value, it is predicted that the power demand at the predetermined prediction target date and time will deviate from the predetermined value range. The re-determination start condition is, for example, a condition that a judgment prediction result is obtained indicating that there is a substation whose power demand does not deviate from the predetermined value range due to the influence of the amount of power generated by photovoltaic power generation at the predetermined prediction target date and time. The other update determination criterion is, for example, a determination criterion that, among the prediction models determined to require update in the first determination result, refers to the judgment prediction result and determines that update is unnecessary for substations whose power demand is predicted to not deviate from the predetermined value range due to the influence of the amount of power generated by photovoltaic power generation at the predetermined prediction target date and time, and determines that update is unnecessary for the remaining prediction models and prediction models determined to require no update in the first determination result. This makes it possible to avoid unnecessary updates of prediction models that do not deteriorate the control accuracy of the grid equipment 41 even if no updates are performed, and reduces the processing load.
(6-9) Modification of the model identification unit 31081

For at least one prediction target among one or more prediction targets, the processor may set a sample weighting value, which is the relative importance of the sample data to other sample data, for each sample data for the prediction target based on the degree of difference calculated for the prediction target, and identify a prediction model corresponding to the prediction target based on the set sample weighting value. Specifically, for example, it may be as follows.

The model identification unit 31081 may set a sample weighting value for each sample when identifying a prediction model, and may perform identification based on the sample weighting value. The sample weighting value is a value indicating the relative importance with respect to other samples. Identification based on the sample weighting value is to identify a prediction model so as to be particularly suitable for a sample with a large sample weighting value. The method of setting the sample weighting value includes a time weighting method in which the sample weighting value of a sample with a newer observation date and time is set to a large value based on the observation date and time of the sample, and a dissimilarity weighting method in which the sample weighting value of a sample with a newer observation date and time is set to a large value when the index value indicating the degree of dissimilarity of the model update control data 3114A is equal to or greater than a predetermined value. The method of setting the sample weighting value may be a combination of both the time weighting method and the dissimilarity weighting method. Identification based on the sample weighting value set by the time weighting method makes it possible to update the prediction model to one that more closely follows the latest response characteristics, thereby making it possible to suppress deterioration of prediction accuracy. By using the identification method based on the sample weights set by the dissimilarity weighting method, when there is a large change in the response characteristics, it is possible to update the prediction model to one that more closely tracks the latest response characteristics, thereby making it possible to suppress deterioration in prediction accuracy.
(6-10) Modification of the predicted value calculation unit 3109

For at least one prediction target among one or more prediction targets, the processor may calculate a predicted value of the prediction target using the identified prediction model corresponding to the prediction target. If a predetermined condition regarding the degree of difference is satisfied, the processor may perform a correction process on the calculated predicted value to calculate a new predicted value. Specifically, for example, the following may be performed.

The forecast value calculation unit 3109 calculates the forecast value of the power demand at a specified date and time of each prediction target based on each prediction model and factor data of the prediction factor data 3112A, and may perform a correction process on the calculated forecast value of the power demand if a condition is satisfied. The condition is a condition based on an index value indicating the degree of difference, for example, an index value indicating the degree of difference of each prediction model of the model update control data 3114A is equal to or greater than a specified value. That is, the forecast value calculation unit 3109 receives the model management data 3115A, the model data 3116A, and the prediction factor data 3112A, inputs specified factor data corresponding to each prediction model of the prediction factor data 3112A to each prediction model, and calculates the forecast value of the power demand at a specified date and time of each prediction target. Next, the forecast value calculation unit 3109 performs a correction process on the calculated forecast value of the power demand of the prediction model that satisfies the condition, calculates a new forecast value, and transmits the new forecast value to the system control calculation device 21. The correction process is a process of correcting the predicted value to follow the observed value of the current power demand, for example, by calculating the difference between the average predicted value of the power demand in a specified range and the observed value of the current power demand, and then adding the calculated difference to the predicted value of the power demand to calculate a new predicted value. This makes it possible to suppress deterioration in prediction accuracy by adapting the predicted value of the power demand to the current observed value, even in situations where the prediction accuracy of the prediction model may decrease due to a sudden change in the response characteristics.

As a correction process, a process may be used in which an observed value of power demand showing the same trend as the most recent observed value of power demand is extracted from the data of observed values of power demand so that the time range of the predicted value of power demand is included, and a new predicted value is calculated by taking a weighted average of the predicted value and the extracted observed value of power demand showing the same trend. This makes it possible to suppress deterioration of the prediction accuracy by adapting the predicted value of power demand to the observed value showing a similar trend in the past, even in a situation where the prediction accuracy of the prediction model may decrease due to a sudden change in the response characteristics.
(6-11) Other Modifications

In addition, in the electricity demand forecasting system 3 of this embodiment, a display unit is omitted for simplicity of explanation, but the calculation results of each processing unit and intermediate results of each processing unit may be calculated as appropriate through an output device such as a display or printer.

Although one embodiment of the present invention and several modified examples have been described above, these are merely examples for the purpose of explaining the present invention, and are not intended to limit the scope of the present invention to these embodiments and modified examples. The present invention can also be implemented in various other forms.

For example, any two or more of the above-mentioned modifications can be combined.

The prediction model may be constructed by any of the following methods.
- Linear models such as ridge regression, lasso regression, and elastic net.
- Tree models such as regression trees, random forests, and boosted trees.
- Kernel methods such as support vector regression, kernel ridge regression, and Gaussian process regression.
-Nonlinear models such as recurrent networks and neural networks such as Long Short-Term Memory.

1. Power grid management system 3. Power demand forecasting system

Claims (15)

an interface device that receives input of prediction target observation data and factor observation data;
a storage device in which the prediction target observation data and the factor observation data are stored;
a processor connected to the interface device and the storage device, the processor updating a prediction model determined to require updating among one or more prediction models corresponding to one or more prediction objects,
For each prediction target, the prediction model is a model that receives as input a factor prediction value, which is data including a prediction value for a future period for each of one or more factors that may affect the one or more prediction targets, and outputs a prediction value of an observation value obtained for the prediction target in the future period;
the prediction target observation data includes, for each of the one or more prediction targets, time series data of past observation values obtained for the prediction target;
the factor observation data includes, for each of the one or more factors, time series data of past observation values obtained for the factor;
The processor generates update data from the prediction target observation data and the factor observation data;
The update data includes sample data for each observation period for each of the one or more prediction targets,
For each prediction object, for each observation period, the sample data includes one or more observations obtained in that observation period for the prediction object and an observation obtained in that observation period for each of the one or more factors;
The processor calculates a degree of difference between the first sample group and the second sample group for each prediction target;
For each prediction target,
The first sample group is one or more sample data used to identify a prediction model corresponding to the prediction target,
The second sample group is one or more sample data that have not yet been used to identify a prediction model corresponding to the prediction target,
The processor performs an update necessity determination, which is a determination of whether or not a degree of difference is equal to or greater than a threshold, for each prediction target, and outputs a result of the update necessity determination;
A prediction model for which the result of the update necessity determination is true is a prediction model that needs to be updated,
A prediction model for which the result of the update necessity determination is false is a prediction model that does not need to be updated.
Prediction system. For at least one prediction object of the one or more prediction objects, the processor:
Calculating a representative value of an index value indicating a difference between each sample data of the first sample group and each sample data of the second sample group;
Calculating the degree of difference based on the representative value.
The prediction system of claim 1 . For at least one prediction target among the one or more prediction targets, the processor identifies a threshold value when a performance index value, which is a value based on at least one of a processing load of updating a prediction model corresponding to the prediction target and a prediction accuracy of the prediction model, is optimal, and sets the identified threshold value as a threshold value of a degree of discrepancy.
The prediction system of claim 1 . For at least one prediction target among the one or more prediction targets, the processor calculates a dissimilarity based on a distance or similarity between a first sample group and a second sample group.
The prediction system of claim 1 . For at least one prediction target among the one or more prediction targets, the processor detects unsuitable sample data, which is sample data having an abnormal value, from the update data, and performs a process to deal with the unsuitable sample data.
The prediction system of claim 1 . The processor generates the factor predicted value for each prediction target based on factor prediction data including a predicted value of each of the one or more factors and the factor observation data.
The prediction system of claim 1 . For at least one prediction object of the one or more prediction objects, the processor:
Based on the calculated degree of difference for the prediction target, a sample weight is set for each sample data for the prediction target, which is the relative importance of the sample data to other sample data;
Identifying a prediction model corresponding to the prediction target based on the set sample weighting value;
The prediction system of claim 1 . For at least one prediction object of the one or more prediction objects, the processor:
Calculating a predicted value of the prediction target using the identified prediction model corresponding to the prediction target;
If a predetermined condition regarding the degree of difference is satisfied, a correction process is performed on the calculated predicted value to calculate a new predicted value.
The prediction system of claim 1 . The processor calculates a similarity between the prediction objects based on the prediction object observation data, classifies the plurality of prediction objects into a predetermined number of groups such that prediction objects having high similarity based on the similarity are grouped together, and for each group, integrates sample data of all prediction objects belonging to the group;
In the update data, for each of the one or more prediction targets, sample data for each observation period is integrated sample data obtained for a group to which the prediction target belongs.
The prediction system of claim 1 . For at least one prediction target among the one or more prediction targets, the processor sets a value obtained by multiplying the standard deviation of the distance between sample data in the first sample group by a predetermined multiple as a threshold value of the degree of dissimilarity.
The prediction system of claim 1 . The processor,
A determination is made as to whether or not a re-determination start condition is satisfied, that is, the number of prediction models determined to require updating exceeds an upper limit number of models, which is the number of prediction models that can be updated within a predetermined time period;
When the re-determination start condition is satisfied, data representing a prediction model that is equal to or less than the upper limit number of models and has a relatively high degree of difference among the prediction models determined to require updating is output as a result of the update necessity determination.
The prediction system of claim 1 . The processor,
determining whether or not there is an inappropriate prediction target, which is a prediction target for which a predicted value outputted using a prediction model is predicted not to satisfy a condition, based on at least one of the factor predicted value and factor prediction data including a predicted value of each of the one or more factors;
A determination is made as to whether or not a re-evaluation start condition, that is, that there is an inappropriate prediction target, is satisfied;
When the re-evaluation start condition is satisfied, data representing prediction models excluding prediction models corresponding to inappropriate prediction targets from among the prediction models determined to require updating is output as a result of the update necessity determination.
The prediction system of claim 1 . The one or more factors include at least one of weather and an amount of energy supply by renewable energy;
Each of the one or more prediction targets is a power demand of a substation.
The prediction system of claim 1 . A prediction method for making a prediction using a prediction model,
A computer generates update data from the prediction target observation data and the factor observation data,
The prediction target observation data includes, for each of one or more prediction targets, time series data of past observation values obtained for the prediction target,
The factor observation data includes, for each of one or more factors that may affect the one or more prediction targets, time series data of past observation values obtained for the factor,
The update data includes sample data for each observation period for each of the one or more prediction targets,
For each prediction object, for each observation period, the sample data includes one or more observations obtained in that observation period for the prediction object and an observation obtained in that observation period for each of the one or more factors;
The computer calculates the degree of difference between the first sample group and the second sample group for each prediction target;
For each prediction target,
The first sample group is one or more sample data used to identify a prediction model corresponding to the prediction target,
The second sample group is one or more sample data that have not yet been used to identify a prediction model corresponding to the prediction target,
For each prediction target, the prediction model is a model that receives as input a factor prediction value, which is data including a prediction value for a future period for each of the one or more factors, and outputs a prediction value of an observation value obtained for the prediction target in the future period;
The computer performs an update necessity determination for each prediction target, which is a determination of whether or not the degree of difference is equal to or greater than a threshold, and outputs the result of the update necessity determination;
A prediction model for which the result of the update necessity determination is true is a prediction model that needs to be updated,
The prediction model for which the result of the update necessity determination is false is a prediction model that does not need to be updated,
The computer updates a prediction model that is determined to require updating among one or more prediction models corresponding to the one or more prediction targets.
Forecasting methods. System equipment in a power system;
a prediction system that outputs a prediction value obtained for each of the one or more prediction objects;
a system control arithmetic device that generates control data including a control instruction based on the prediction value output for each of the one or more prediction objects, and outputs the control data;
a system equipment management device that controls the system equipment based on the control data,
The prediction system comprises:
an interface device that receives input of prediction target observation data and factor observation data;
a storage device in which the prediction target observation data and the factor observation data are stored;
a processor connected to the interface device and the storage device, the processor updating a prediction model determined to require updating among one or more prediction models corresponding to one or more prediction objects,
For each prediction target, the prediction model is a model that receives as input a factor prediction value, which is data including a prediction value for a future period for each of one or more factors that may affect the one or more prediction targets, and outputs a prediction value of an observation value obtained for the prediction target in the future period;
the prediction target observation data includes, for each of the one or more prediction targets, time series data of past observation values obtained for the prediction target;
the factor observation data includes, for each of the one or more factors, time series data of past observation values obtained for the factor;
The processor generates update data from the prediction target observation data and the factor observation data;
The update data includes sample data for each observation period for each of the one or more prediction targets,
For each prediction object, for each observation period, the sample data includes one or more observations obtained in that observation period for the prediction object and an observation obtained in that observation period for each of the one or more factors;
The processor calculates a degree of difference between the first sample group and the second sample group for each prediction target;
For each prediction target,
The first sample group is one or more sample data used to identify a prediction model corresponding to the prediction target,
The second sample group is one or more sample data that have not yet been used to identify a prediction model corresponding to the prediction target,
The processor performs an update necessity determination, which is a determination of whether or not a degree of difference is equal to or greater than a threshold, for each prediction target, and outputs a result of the update necessity determination;
A prediction model for which the result of the update necessity determination is true is a prediction model that needs to be updated,
A prediction model for which the result of the update necessity determination is false is a prediction model that does not need to be updated.
Grid management system.

Images