KR20210092895A - Method and Server for Compensating a Missing Value in a Power Data - Google Patents

Abstract

The present invention relates to a server for compensating missing values in power data to interpolate the missing values close to actual measurements and a method thereof. According to an embodiment of the present invention, the method comprises the following steps: collecting power measurement data and information on factors affecting the power measurement data; pre-processing the collected power measurement data and the information on the factors in a form for learning; using the pre-processed data to form a learning model which has a structure including sub-networks, receives the information on the factors, and outputs the power measurement data; performing learning for each sub-network forming the learning model by using the measured power measurement data; performing learning on the entire subnetwork included in the learning model; checking whether learning on the learning model is completed; and estimating missing power measurement data by using the learning model completing learning.

Description

{Method and Server for Compensating a Missing Value in a Power Data}

The present invention relates to a missing value complementation server and method for compensating for missing values for power data continuously collected over time. In particular, in consideration of factors that may affect power consumption, It relates to a missing value complementation server and method for compensating for missing values.

The energy management system is equipped with various sensors, including a smart meter, which is a device that measures the amount of used electrical energy consumption, which collects information on management targets such as houses, buildings, and cities, and is used as information for efficient operation of the energy management system.

The smart meter's measurements are used not only as an indicator for current power consumption, but also as an indicator for forecasting future power consumption.

Since the energy management system establishes an energy management plan based on the prediction results, accurate power consumption prediction is an important factor in the operation of the energy management system. In addition, since the prediction accuracy of power consumption is greatly affected by the data quality, well-organized smart meter data collection is essential.

For two-way wireless communication with a structure that sends signals from several sensors of remote facilities or equipment more than several hundred meters to the host, and also sends control signals such as power on/off of remote equipment from the host, the international standard called LoRa is used. Wireless control technology is also applied. In addition, there are various methods such as SigFox, LTE-M, and NB-IoT in addition to LoRa, especially for IoT-dedicated communication methods for transmitting measurement data of smart meters, and major communication service companies compete by adopting different methods. is in a state

When measuring power with a smart meter, power measurement is often missing due to device malfunction or signal transmission error, making it difficult to collect high-quality power data.

When a power measurement is missing, it causes prediction errors due to loss of information, and in particular, the performance of prediction methods based on continuous values such as autoregressive integrated moving average (ARIMA) deteriorates.

The missing part problem is solved by filling in the missing part with appropriate values and supplementing the problem, and various complementary methods are being studied.

Linear interpolation is a method of interpolating using temporally adjacent data, connecting a straight line based on the data before and after the missing value, and substituting the value on this straight line to the missing value. However, there is a disadvantage that the performance is very poor if the length of the missing interval is long.

There is an interpolation method that uses the nearest data using the k-nearest neighbor algorithm. This method finds similar patterns in existing data by using the nearest neighbor algorithm on data points near the missing values. Then, the data of the missing part is interpolated based on the found pattern. However, since the interpolation result depends too much on the past data, the interpolation performance deteriorates when the data pattern is complex.

Republic of Korea Patent Publication No. 10-2018-0112540

An object of the present invention is to provide a server and method for compensating missing values of power data that can interpolate missing values of power data close to actual measured values.

Specifically, an object of the present invention is to provide a missing value complementation server and method for power data capable of performing more accurate missing value correction by identifying a complex power demand pattern.

Specifically, the present invention intends to provide a server and method for compensating for missing values of power data that can configure a variable that can reflect time series patterns well and configure a more optimized interpolation model by ensembles a machine learning algorithm.

In accordance with one aspect of the present invention, there is provided a method for compensating for missing values of power measurement data, the method comprising: collecting information on power measurement data and factors affecting the power measurement data; preprocessing the collected power measurement data and information on the factors into a form for learning; using the preprocessed data to form a learning model having a structure including sub-networks, in which information on the factors is input and the power measurement data is output; performing learning for each sub-network forming the learning model by using the measured power measurement data; performing learning on the entire subnetwork included in the learning model; checking whether learning on the learning model is completed; and estimating the missing power measurement data using the learning model on which learning is completed.

Here, the power measurement data is a measured value for power consumption, and the factors may be weather data and calendar information.

Here, the learning model is an ensemble model, and in the step of pre-processing the measurement data in a form for learning, the input items for each factor are a real value having an absolute value between 0 and 1, a positive real value, or a vector. As a method of converting the values into values, normalization may be performed on the input data of the ensemble learning model by maximal/minimum normalization.

Here, in the step of forming the learning model, the number of sub-networks to be included in the learning model may be determined, and after random sampling of the preprocessed data, sub-networks having ensemble weights may be formed.

Here, in the step of performing learning for each sub-network, learning may be performed a predetermined number of times in a direction to minimize a loss function based on a difference between actual power consumption data and expected energy consumption data for each sub-network.

Here, the loss function may be defined according to the following equation.

(where t is the actual electrical energy consumption data, y _i is the expected energy consumption of the subnetwork (SN _{i ))}

Here, in the step of learning the entire sub-network _{, it is checked whether the ensemble weight of the sub-network is smaller than a predefined threshold value (α th} ), and if it is smaller than the threshold value, the corresponding sub-network is trained It can be deleted from the model.

Here, in the step of performing the learning for the entire sub-network, learning may be performed in a direction of minimizing a loss function according to the following equation for the entire sub-network.

(where t is the actual electrical energy consumption data, y _out is the estimated energy consumption of the entire subnetwork)

Here, in the step of checking whether the learning is completed, the result of performing the previously performed learning on each of the sub-networks and the learning of the entire sub-network, and By comparing the results of performing the learning for each sub-network and the learning for the entire sub-network, if the difference between the two results is less than a predetermined reference value, it can be determined that the learning has been completed. .

Power measurement data missing value compensation server according to another aspect of the present invention, the power measurement data and data collection unit for collecting information on factors affecting the power measurement data; a data preprocessing unit that normalizes the collected data and information by maximum and minimum normalization; a learning model forming unit that uses the preprocessed data to form a learning model having a structure including sub-networks, in which information on the factors is input and the power measurement data is output; a learning performing unit for learning the learning model formed using the measured power measurement data; and a data interpolator for estimating missing power measurement data using the learning model on which learning is completed.

Here, the power measurement data is a measured value for power consumption, and the factors may be weather information including temperature and calendar information including month, day, and day of the week.

Here, the data preprocessor converts the weather information into numerical data between 0 and 1 through maximum and minimum normalization, and converts the month and day information into a vector value having an absolute value between 0 and 1, and the day of the week Information can be converted to integer values of 0 and 1.

Here, the learning performing unit performs learning for each sub-network forming the learning model by using the measured power measurement data, and performs learning on the weight of each sub-network on the learning model, It may be determined whether learning of the learning model is completed.

Here, the learning performing unit includes a result of performing the previously performed learning for each of the sub-networks and the learning for the entire sub-network, and a result of the previously performed learning for each of the sub-networks. Comparing the results of performing the learning and the learning of the entire sub-network, if the difference between the two results is less than a predetermined reference value, it may be determined that the learning is completed.

When the server and method for supplementing the missing value of power data according to the present invention according to the above-described configuration are implemented, there is an advantage in that the missing value of the power data can be interpolated close to the actual measured value.

The server and method for supplementing missing values of power data of the present invention have the advantage that an effective input variable can be configured for power demand pattern learning through data preprocessing.

The server and method for supplementing missing values of power data of the present invention train a multi-layer perceptron model by extracting data in an arbitrary period, and use an excellent model among the learned multi-layer perceptron to construct a predictive model optimized for the corresponding power data, It has the advantage of accurately predicting the amount of power in the missing period.

1 is a conceptual diagram of an algorithm showing a process configuration of a system for compensating missing values of power data according to the spirit of the present invention.
2 is a conceptual diagram representing an assembly structure of an ensemble model;
3 is a conceptual diagram illustrating a flow required for calculating EMAk for an intermediate check.
4 is a block diagram illustrating a power measurement data missing value complementation server according to the spirit of the present invention.

In describing the present invention, terms such as first, second, etc. may be used to describe various components, but the components may not be limited by the terms. The terms are only for the purpose of distinguishing one component from another. For example, without departing from the scope of the present invention, a first component may be referred to as a second component, and similarly, a second component may also be referred to as a first component.

When a component is referred to as being connected or connected to another component, it may be directly connected or connected to the other component, but it can be understood that other components may exist in between. .

The terms used herein are used only to describe specific embodiments, and are not intended to limit the present invention. The singular expression may include the plural expression unless the context clearly dictates otherwise.

In this specification, the terms include or include are intended to designate that a feature, number, step, operation, component, part, or combination thereof described in the specification exists, and includes one or more other features or numbers, It may be understood that the existence or addition of steps, operations, components, parts or combinations thereof is not precluded in advance.

In addition, shapes and sizes of elements in the drawings may be exaggerated for clearer description.

1 is a conceptual diagram illustrating a process configuration of a system for compensating missing values of power data according to the spirit of the present invention.

The missing value correction method according to the illustrated process may consist of five steps, including data set preparation, model preparation, enhancement stage, adjustment stage, and interim inspection. Detailed examples for each step will be described later.

Specifically, the illustrated power measurement data missing value supplementation method includes the steps of: collecting power measurement data and information on factors affecting the power measurement data (S10); pre-processing the collected power measurement data and information on the factors into a form for learning (S15) (here, it is possible to configure a data set for learning the ensemble model); using the pre-processed data to form a learning model having a structure including sub-networks, in which information on the factors is input and the power measurement data is output (S20); performing learning for each sub-network forming the learning model by using the measured power measurement data (S30); performing learning on the entire subnetwork included in the learning model (S40); checking whether learning on the learning model is completed (S50); and estimating the missing power measurement data using the learning model on which learning is completed (S60).

In the step of collecting the data/information (S10), as shown in FIG. 1, power-related measurement values are recorded from a plurality of distributed power measuring devices, and weather information is collected and recorded as factors affecting power consumption. In a system having a central collection device, it may be performed in a manner of reading the measured values recorded in the central collection device. In another implementation, in the step S10, the measured value may be directly transmitted from each measuring device or a weather measuring device.

Next, a process of preparing a data set for learning performed in the step S15 of pre-processing in the form for learning will be described.

In performing the steps S10 and S15, the power measurement data is a measured value for power consumption, and the factors are concretely described in the case of weather data and calendar information.

The method proposed according to the spirit of the present invention may be based on an ensemble model with several MLPs. Therefore, the first step is the process of collecting data and constructing a data set for training the ensemble model. For example, when collecting electrical energy consumption data, the collected data becomes an output variable of the ensemble model. We collect descriptive data such as weather and calendar to construct input variables of the ensemble model. These data must be measured or processed for each measurement period of the SM (sensor module). For example, if the SM records energy consumption every hour, the descriptive data should contain information every hour . Weather data may include temperature, humidity, wind speed, temperature humidity index (THI) and wind temperature index (WCI). Since these factors are actually used to predict energy consumption, they can be selected as input variables. Calendar data may include information about times of measurement, such as timestamps, managed work schedules, and holiday dates.

Calendar data needs some preprocessing. First, find values for season, month, day, hour, minute, and day of the week through timestamps. Depending on the implementation, minutes may not be considered if the measurement interval is 1 hour or more. Next, time units such as month, day, and hour can be preprocessed to reflect periodicity. The equations used here are Equation 1 and Equation 2 below.

Here, time is a time unit to be converted, and period _time is a period. For example, if you want to convert months, period _time becomes 12. If you want to convert days, the period _time may vary from 28 to 31 depending on the month.

Next, the season and the day of the week are expressed as a one-hot vector with only 1 element.

Next, add an indication as to whether the measured time is included in the managed subject's working time.

Finally, we add a variable indicating whether the measurement time is a holiday. After preprocessing the calendar data, 26 input variables are prepared, and a list of these variables may be configured as shown in Table 1 below.

The weather-related variable may be greater than 1. You can make these variables range from 0 to 1 by performing min-max normalization to fit within similar ranges as other variables. The collected electrical energy consumption data can also be normalized for the same reason. As a result, one output variable and 26 input variables are generated for each measurement time and the entire data set is prepared. Of all data points in the data set, this data point is included in the training data set if the values of the output variables are known, and then the ensemble model learns from these data points.

In the specific example described above, the learning model is an ensemble model, and in the step (S15) of pre-processing the measurement data into a form for learning, the input items for each factor are real values having an absolute value between 0 and 1. , it can be seen that normalization is performed by maximal/minimum normalization of the input data of the ensemble learning model in a manner of converting it into a positive real value or a vector value.

Next, a learning model preparation process performed in the step S20 of forming the learning model will be described.

In the case of ensemble model configuration, the number of subnetworks constituting the ensemble model represented by N is determined. The more subnetworks the model has, the more likely it is to get a better model at the end of training. However, the learning time becomes longer in proportion to N. Experimentally, it can be seen that the proposed model can achieve sufficient results when N is set to 20 or more.

After setting N, random sampling is generally performed to construct a new data set for each subnetwork. Sampling generates N different sets of new data, each subnetwork owns its own data set. Subnetworks are trained by the data set, and after learning, different prediction results are obtained from different subnetworks to enhance the effectiveness of the ensemble. This method is commonly referred to as bootstrap aggregation or bagging . A famous example of bagging is random forest (RF). Random Forest (RF) is a tree-based ensemble model that uses bagging to grow various trees. Excellent generalization performance and accurate prediction performance have been reported in several studies, which gives an advantage to adopt bagging for the model.

In the step of forming the learning model ( S20 ), the number of sub-networks to be included in the learning model is determined, and after random sampling of the pre-processed data, sub-networks having ensemble weights are formed.

2 shows the assembly structure of the ensemble model.

N _SN subnetworks are created after random sampling. Except for the weight and bias of the MLP, each subnetwork (SN _i , i = 1, 2, 3,…N) additionally has an ensemble weight a _i . a _i is used to compute the results of the ensemble model, which is the weighted average of the outputs of _{each subnetwork (y i ).} However, to guarantee a weighted average, the ensemble weights must satisfy the following two conditions: 1) All ensemble weights must be included in the interval [0, 1] (a _i ∈[0, 1]). 2) The sum of the ensemble weights must be 1 (Σa _i = 1).

When designing a deep learning model, it is difficult to find a way to train the model while meeting these two conditions, but the function of softmax can always satisfy these conditions used for classification tasks. Therefore, by applying the softmax function, _{a i} can be determined by adopting a _{new variable w i .} The softmax function may be defined by Equation 3 below.

Due to the properties of the softmax function described above, the ensemble weight always satisfies two conditions. Therefore, in the training of the model, a _i is not directly adjusted, but _{w i} is adjusted instead of _{a i .} Changes in the w _i will be able to find the appropriate value of the influence on a _i a _i. In the rest of the description, for convenience, a _i will be expressed roughly as if it were directly adjusted.

In this model, two parts, 1) sub-network and 2) ensemble weights, need to be optimized. Initially, all parameters, including the parameters of the subnetwork and ensemble weights, are randomly assigned, but the two parts are optimized as the next two-step training proceeds.

Next, a process of improving the formed learning model, that is, performing learning on each of the sub-networks (S30) will be described.

In the step (S30) of performing learning for each sub-network, learning is performed a predetermined number of times in the direction of minimizing a loss function based on a difference between actual power consumption data and expected energy consumption data for each sub-network.

Specifically, in the learning (first learning) of step S30, each sub-network is independently learned. The loss function (loss _ES,i ) of SN _i may be defined by Equation 4 below.

Here, t is the actual electrical energy consumption data, and y _i is the expected energy consumption of _{SN i .} Without interfering with each other, all subnetworks learn their own data sets to minimize lossy functions. Therefore, in the above step, only the parameters of the MLP are changed while the ensemble weight is fixed. When all the subnetworks have _{learned the number of epoch ESs} , this stage is completed and the second learning, the adjustment stage, begins.

Next, an adjustment process for the entire sub-network, that is, a step (S40) of learning the entire sub-network included in the learning model will be described.

In this process, ensemble weights are adjusted, and unnecessary subnetworks are removed according to these weights. In the ensemble weight adjustment, unlike the enhancement phase, only the ensemble weights can be changed during learning, and the parameters of the subnetwork are fixed. This means that y _i cannot be changed in this process. The loss function of the adjustment step is as shown in Equation 5 below.

The data set used for training is the original data set. That is, this data set contains all the data points from the training data set.

Ensemble weight learning can be performed with _{Epoch AS number.} Then, it is checked whether the ensemble weight of each subnetwork _{is less than a predefined threshold (α th ).} If it is less than the threshold, this subnetwork is removed from the ensemble model. The reason for performing this step is that a sub-network with a low ensemble weight hardly contributes to an accurate prediction compared to a sub-network with a large ensemble weight, so it is removed to reduce unnecessary learning time.

For example, in the situation where a1 = 0.01, a2 = 0.05, a3 = 0.15, and a4 = 0.79, if αth = 0.02, SN1 is deleted from the ensemble model. When α _th = 0.1, SN1 and SN2 are deleted together in the above situation. Therefore, α _th affects the number of remaining subnetworks after learning. If you need a lightweight ensemble model, you can set α high, but if α _th is too large, the accuracy can be greatly reduced.

Next, the step of checking whether the learning is completed (S50) will be described.

After the sub-networks with low influence are deleted in the process of step S40, a process for performing an intermediate check may be started. This intermediate check process may be repeated NMAX times, and as the iteration continues, the number of subnetworks decreases. Also, the time required to perform each step is reduced due to the deletion of the subnetwork. If the ensemble model is considered sufficiently trained during iteration, all learning processes can be terminated immediately.

Whenever each adjustment process by learning of each order for the entire subnetwork is completed until the number of iterations reaches NMAX, it is checked whether learning is completed. That is, as a kind of intermediate check, the goal is to determine that the ensemble model does not require any more training iterations and to determine the end of training.

In this process, if the ensemble model satisfies the condition, the iteration loop is terminated. Otherwise, the step of performing learning for each of the sub-networks is performed again.

Since the iterative loop of the learning phase takes more time compared to other parts of the machine learning algorithm, in the case of the learning algorithm, an intermediate check process as described above is necessary to shorten the total time required to train the model. do.

3 is a conceptual diagram illustrating a flow required _{for calculating EMA k} for an intermediate check.

With reference to FIG. 3 , the conditions for checking in the step S50 of checking whether the learning is completed will be described.

An ensemble model no longer needs to be trained once the model converges. When the ensemble weights do not change any more, it can be said that the ensemble model converges. Accordingly, the condition to be checked in the step of checking whether the learning is completed may be whether the model converges or not, and the required calculation flow is shown in FIG. 3 .

We can denote all ensemble weights as ew _{k in the kth iteration.} As described in the step of forming the learning model above, the ensemble weights satisfy two conditions due to the characteristics of the softmax function. Therefore, _{the sum of all elements of ew k} is 1, and each element is included in the interval [0, 1]. This is also a property of a probability distribution, so we _{can treat ew k} like a probability distribution.

When the ensemble model converges, ew _k does not differ from ew _{k-1 ,} and the ensemble weight at the k-1th iteration is determined. Here, in order to measure the difference, Kullback-Leibler divergence (KLD) may be applied as a measure of how one probability distribution differs from another probability distribution. If the given probability distributions are similar, the KLD is low, and vice versa, the KLD is large. Also, if two given items are exactly the same, KLD is 0. Therefore, _{when the KLD between ew k-1} and ew _k expressed as _{KL k} is 0, it is possible to check the weight in the state in which the learning is completed by finding k.

Equation 6 below shows the equation _{of KL k .} Here, D _KL (x||y) means the KLD between two probability distributions x and y. a _i ,k represents the ensemble weight _{of SN i} in iteration k.

However, _{the situation in which KL k} becomes 0 occurs only in the ideal case, and in reality, KL _k converges to a specific value that is not 0 and close to it. Therefore, we can consider this value as a threshold value instead of 0, but it is difficult to control or predict because this value is affected by many factors such as the number of subnetworks, random initialization, and training data set. To alleviate this problem, a subtraction between _{KL k} and KL _{k-1 is applied.} If the subtraction result is 0, there is no difference between _{KL k} and KL _{k-1 , and it can be determined that KL k} converges to a specific value. DIFF _k is the result of subtraction _{of KL k} and KL _k-1.

However, empirically, DIFF _k becomes 0 after several iterations, so it is difficult to use _{DIFF k as a reference.} DIFF _k may cause insufficient training time for each subnetwork, resulting in inaccurate iterations (loops). Therefore, we adopt the average of exponential shifts of _{DIFF k} to ensure sufficient learning time for the subnetwork. _{The average of the exponential movement of EMA k} in the k-th iteration is as shown in Equation 7 below.

In the above equation, α∈ [0,1] is the update coefficient, and in summary, when EMA _k becomes 0, the learning iteration (loop) is completed. If α is large, the learning iteration is completed early. Conversely, the required iteration time may be close to NMAX, but with better accuracy than in the previous case. However, if N is small, EMA _k may become 0 after several iterations. This is a result of a larger change in the ensemble weights in the initial iteration. In this case, α should be larger or set the minimum number of iterations.

In summary, in the step (S50) of checking whether the learning is completed, the results of performing the previously performed learning on each of the sub-networks and the learning of the entire sub-network are performed; Comparing the results of performing the learning for each sub-network and the learning for the entire sub-network performed this time, if the difference between the two results is less than a predetermined reference value, learning is completed can judge

In the step (S60) of estimating the missing power measurement data, the learning of the model composed of the power data as output is completed by inputting the factors (weather information, calendar information) influencing the performance results up to the step S50 as input. It is possible to estimate/acquire power data at the missing time by reflecting the information on the factors affecting the missing time in the model as input.

For example, by merging the thus-obtained deficit data on the power consumption with the total measured power consumption data, it is possible to obtain a mid/long-term expected power consumption pattern.

4 shows a power measurement data missing value complementation server according to the spirit of the present invention.

The illustrated power measurement data missing value supplementation server 100 includes: a data collection unit 110 for collecting power measurement data and information on factors affecting the power measurement data; a data pre-processing unit 115 for normalizing the collected data and information by maximum and minimum normalization; a learning model forming unit 120 using the preprocessed data to form a learning model having a structure including sub-networks, in which information on the factors is input and the power measurement data is output; a learning performing unit 130 for learning the learning model formed using the measured power measurement data; and a data interpolator 160 for estimating the missing power measurement data using the learning model on which the learning has been completed.

The data collection unit 110 performs the step (S10) of collecting the data/information described above in FIG. 1, and the data preprocessor 115 performs the step (S15) of pre-processing in the form for learning described above. And, the learning model forming unit 120 performs the above-described learning model forming step (S20).

The learning performing unit 130 performs learning for each sub-network forming the above-described learning model (S30); A step (S40) of learning the entire sub-network included in the learning model is performed, and a step (S50) of confirming whether or not the learning of the learning model is completed is performed.

The data interpolator 160 performs the step of estimating the missing power measurement data described above in FIG. 1 ( S60 ).

The data collection unit 110 collects weather information and past electrical load data for a predetermined period. For example, in the case of a system having a central collection device 200 that records power-related measured values from a plurality of distributed power meters installed in the system, and collects and records weather information as factors affecting power consumption, the data As illustrated, the collection unit 110 may acquire data/information recorded in the central collection device 200 as a separate DB server.

The preprocessor 115 changes the variables to values between 0 and 1 through Min-Max normalization for learning speed and accuracy, and month, day, hour and minute data to reflect periodicity. By applying the sin and cos functions, preprocessing such as Equations 1 and 2 may be performed. More specifically, period means the period of the variable, and when monthly data is substituted for time, period becomes 12, which is the period of the substituted data, and when hour data is substituted, the period value is set to 24, Pre-processing can be performed. As a result, variables after pretreatment as shown in Table 1 can be obtained.

The learning model forming unit 120 may perform data sampling on the preprocessed data/information to form a learning model including a plurality of sub-networks.

The training model forming unit 120 selects an arbitrary period of power consumption data, and the power consumption data of the selected period is used to learn the multilayer perceptron model in the single model learning unit. In the above process, the weight adjustment unit increases the weight of the excellent multi-layer perceptron model (ie, the subnetwork) after a certain iteration. The execution process is the same as in FIG. 2 described above.

When viewed functionally, the learning performing unit 130 includes a single model learning unit that performs learning for each sub-network forming the above-described learning model; It can be seen that the model includes a weight adjustment unit that adjusts weights for each sub-network according to learning of the entire sub-network included in the learning model, and a model verification unit that checks whether learning of the learning model is completed.

The ensemble model constructed by the learning model forming unit 120 is verified by the model verifying unit. The verification process is the same as in FIG. 3 described above. In the verification, a probability distribution can be calculated through Coolback Leibler divergence in each learning step, and the learning completion degree of the model can be verified using this.

Those skilled in the art to which the present invention pertains should understand that the present invention may be embodied in other specific forms without changing the technical spirit or essential characteristics thereof, so the embodiments described above are illustrative in all respects and not restrictive. only do The scope of the present invention is indicated by the following claims rather than the detailed description, and all changes or modifications derived from the meaning and scope of the claims and their equivalent concepts should be construed as being included in the scope of the present invention. .

100: missing value complement server
110: data collection unit
115: data preprocessor
120: learning model forming unit
130: learning execution unit
160: data interpolation unit
200: central collection device

Claims (14)

collecting power measurement data and information on factors affecting the power measurement data;
preprocessing the collected power measurement data and information on the factors into a form for learning;
using the preprocessed data to form a learning model having a structure including sub-networks, in which information on the factors is input and the power measurement data is output;
performing learning for each sub-network forming the learning model by using the measured power measurement data;
performing learning on the entire subnetwork included in the learning model;
checking whether learning on the learning model is completed; and
Estimating missing power measurement data using the learning model on which learning is completed
A method of complementing missing values of power measurement data, including
According to claim 1,
The power measurement data is a measured value for power consumption,
The above factors are meteorological data and calendar information, a method of supplementing missing power measurement data.
According to claim 1,
The learning model is an ensemble model,
In the step of pre-processing the measurement data in a form for learning,
Normalizing the input data of the ensemble learning model by maximum and minimum normalization in a way that converts the input items for each factor into a real value having an absolute value between 0 and 1, a positive real value, or a vector value How to compensate for missing power measurement data.
4. The method of claim 3,
In the step of forming the learning model,
Determining the number of subnetworks to be included in the learning model,
After random sampling of the preprocessed data,
A method of compensating for missing values of power measurement data to form subnetworks having ensemble weights.
3. The method of claim 2,
In the step of performing learning for each sub-network,
A method of compensating for missing values of power measurement data in which learning is performed a predetermined number of times in the direction of minimizing a loss function based on the difference between actual power consumption data and expected energy consumption data for each subnetwork.
6. The method of claim 5,
The loss function is a power measurement data missing value compensation method, characterized in that according to the following equation.

(where t is the actual electrical energy consumption data, y _i is the expected energy consumption of the subnetwork (SN _{i ))}
According to claim 1,
In the step of performing learning for the entire subnetwork,
A method for compensating for missing values in power measurement data by checking whether the ensemble weight of the sub-network _{is smaller than a predefined threshold value (α th} ), and deleting the corresponding sub-network from the learning model if it is smaller than the threshold value.
According to claim 1,
In the step of performing learning for the entire subnetwork,
A method for compensating for missing values of power measurement data for performing learning in a direction of minimizing a loss function according to the following equation for the entire subnetwork.

(where t is the actual electrical energy consumption data, y _out is the estimated energy consumption of the entire subnetwork)
According to claim 1,
In the step of checking whether the learning is completed,
A result of performing the previously performed learning for each sub-network and performing the learning for the entire sub-network;
By comparing the results of the step of performing the learning on each of the sub-networks performed this time and the learning of the entire sub-network,
If the difference between the two results is less than a predetermined reference value, it is determined that the learning is completed.
a data collection unit for collecting power measurement data and information on factors affecting the power measurement data;
a data preprocessing unit that normalizes the collected data and information by maximum and minimum normalization;
a learning model forming unit that uses the preprocessed data to form a learning model having a structure including sub-networks, in which information on the factors is input and the power measurement data is output;
a learning performing unit for learning the learning model formed using the measured power measurement data; and
A data interpolator for estimating the missing power measurement data using the learning model on which learning has been completed
Compensation server for missing power measurement data comprising a.
11. The method of claim 10,
The power measurement data is a measured value for power consumption,
The above factors are meteorological information including temperature and calendar information including month, day, and day of the week, power measurement data missing value supplement server.
12. The method of claim 11,
The data preprocessor,
The weather information is converted into numerical data between 0 and 1 through maximum and minimum normalization,
The month and day information is converted into a vector value having an absolute value between 0 and 1,
The day information is a power measurement data missing value supplement server that converts integer values of 0 and 1.
11. The method of claim 10,
The learning execution unit,
By using the measured power measurement data, learning for each sub-network forming the learning model is performed,
performing learning on the weight of each sub-network on the learning model,
A power measurement data missing value supplementation server for determining whether learning for the learning model is completed.
14. The method of claim 13,
The learning execution unit,
A result of performing the previously performed learning for each sub-network and performing the learning for the entire sub-network;
By comparing the results of the step of performing the learning for each of the sub-networks performed this time and the step of performing the learning on the entire sub-network,
If the difference between the two results is less than a predetermined reference value, the power measurement data missing value supplementation server for determining that the learning is completed.

Images