CN112348070A - Method and system for forecasting medium and short term loads of smart power grid - Google Patents

Method and system for forecasting medium and short term loads of smart power grid Download PDF

Info

Publication number
CN112348070A
CN112348070A CN202011185160.XA CN202011185160A CN112348070A CN 112348070 A CN112348070 A CN 112348070A CN 202011185160 A CN202011185160 A CN 202011185160A CN 112348070 A CN112348070 A CN 112348070A
Authority
CN
China
Prior art keywords
data
load
load data
short
dcnn
Prior art date
Legal status (The legal status is an assumption and is not a legal conclusion. Google has not performed a legal analysis and makes no representation as to the accuracy of the status listed.)
Pending
Application number
CN202011185160.XA
Other languages
Chinese (zh)
Inventor
卢建刚
张国翊
付佳佳
郑鸿远
张珮明
曾瑛
Current Assignee (The listed assignees may be inaccurate. Google has not performed a legal analysis and makes no representation or warranty as to the accuracy of the list.)
Electric Power Dispatch Control Center of Guangdong Power Grid Co Ltd
Original Assignee
Electric Power Dispatch Control Center of Guangdong Power Grid Co Ltd
Priority date (The priority date is an assumption and is not a legal conclusion. Google has not performed a legal analysis and makes no representation as to the accuracy of the date listed.)
Filing date
Publication date
Application filed by Electric Power Dispatch Control Center of Guangdong Power Grid Co Ltd filed Critical Electric Power Dispatch Control Center of Guangdong Power Grid Co Ltd
Priority to CN202011185160.XA priority Critical patent/CN112348070A/en
Publication of CN112348070A publication Critical patent/CN112348070A/en
Pending legal-status Critical Current

Links

Images

Classifications

    • G—PHYSICS
    • G06—COMPUTING; CALCULATING OR COUNTING
    • G06Q—INFORMATION AND COMMUNICATION TECHNOLOGY [ICT] SPECIALLY ADAPTED FOR ADMINISTRATIVE, COMMERCIAL, FINANCIAL, MANAGERIAL OR SUPERVISORY PURPOSES; SYSTEMS OR METHODS SPECIALLY ADAPTED FOR ADMINISTRATIVE, COMMERCIAL, FINANCIAL, MANAGERIAL OR SUPERVISORY PURPOSES, NOT OTHERWISE PROVIDED FOR
    • G06Q10/00—Administration; Management
    • G06Q10/04—Forecasting or optimisation specially adapted for administrative or management purposes, e.g. linear programming or "cutting stock problem"
    • G—PHYSICS
    • G06—COMPUTING; CALCULATING OR COUNTING
    • G06F—ELECTRIC DIGITAL DATA PROCESSING
    • G06F18/00—Pattern recognition
    • G06F18/20—Analysing
    • G06F18/21—Design or setup of recognition systems or techniques; Extraction of features in feature space; Blind source separation
    • G06F18/214—Generating training patterns; Bootstrap methods, e.g. bagging or boosting
    • G—PHYSICS
    • G06—COMPUTING; CALCULATING OR COUNTING
    • G06N—COMPUTING ARRANGEMENTS BASED ON SPECIFIC COMPUTATIONAL MODELS
    • G06N3/00—Computing arrangements based on biological models
    • G06N3/02—Neural networks
    • G06N3/04—Architecture, e.g. interconnection topology
    • G06N3/045—Combinations of networks
    • G—PHYSICS
    • G06—COMPUTING; CALCULATING OR COUNTING
    • G06N—COMPUTING ARRANGEMENTS BASED ON SPECIFIC COMPUTATIONAL MODELS
    • G06N3/00—Computing arrangements based on biological models
    • G06N3/02—Neural networks
    • G06N3/08—Learning methods
    • G—PHYSICS
    • G06—COMPUTING; CALCULATING OR COUNTING
    • G06Q—INFORMATION AND COMMUNICATION TECHNOLOGY [ICT] SPECIALLY ADAPTED FOR ADMINISTRATIVE, COMMERCIAL, FINANCIAL, MANAGERIAL OR SUPERVISORY PURPOSES; SYSTEMS OR METHODS SPECIALLY ADAPTED FOR ADMINISTRATIVE, COMMERCIAL, FINANCIAL, MANAGERIAL OR SUPERVISORY PURPOSES, NOT OTHERWISE PROVIDED FOR
    • G06Q50/00—Systems or methods specially adapted for specific business sectors, e.g. utilities or tourism
    • G06Q50/06—Electricity, gas or water supply

Abstract

The invention provides a method and a system for forecasting medium and short term loads of a smart grid, wherein the method comprises the following steps: acquiring daily power consumption total load data and hourly user power load data of a user; determining a Deep Convolutional Neural Network (DCNN) prediction model obtained by training based on historical power load data; the historical user load data comprises daily power load data in preset days; and inputting the daily total power consumption load data and the hourly user power load data into the DCNN prediction model for processing, and outputting a short-term load prediction result. The scheme of the invention can simplify the load prediction model and improve the load prediction precision.

Description

Method and system for forecasting medium and short term loads of smart power grid
Technical Field
The invention relates to the technical field of information processing in a smart power grid, in particular to a method and a system for predicting medium and short term loads of the smart power grid.
Background
The power load prediction is a technology for establishing a mathematical relationship model between historical power loads and future power loads based on multi-dimensional historical data in the aspects of weather, date, society, economic development and the like. The power load prediction is very important to the construction, maintenance and management of the smart grid. Data processing is an important part of model building.
As an important component of the power load prediction, the short-term load prediction is a recent power load prediction. Typically, this near future is 1 hour to 7 days in the future. From a cost perspective, accurate short-term load predictions may yield considerable revenue, saving hundreds of thousands or even millions of dollars per one percentile reduction in average error of short-term load predictions.
With the development of the internet of things (IoT), many sensor devices are present in smart grids. The application of smart meter infrastructure plays an important role in the conversion from the traditional power grid to the smart grid. The smart electric meter can acquire power consumption information of a user at a certain frequency through the technology of the internet of things and transmit the power consumption information to an electric power company. In this case, a large amount of data is always generated in the smart grid. From these large data, the accuracy of load prediction can be improved by mining useful information.
Fifth generation (5G) communication technologies enable larger amounts of data in smart grids compared to smart grids that are not equipped with 5G. Accordingly, data processing is more difficult. 5G has become one of the core technologies of a new round of global scientific and technical revolution and industrial transformation. It is a strategic communication infrastructure for realizing national digitization, intelligence and everything interconnection. To meet the 5G application, three schemes of enhanced mobile broadband (eMBB), ultra-reliable and low-latency communication (urrllc) and large-scale machine type communication (mtc) are proposed.
Due to the heterogeneous, flexible wireless part in smart grids, 5G presents a complex and enormous challenge to the grid. For example, the backbone network of a smart grid consists of optical networks. Distributed networks that connect thousands of users to the backbone network are not monitored in real time. Real-time monitoring of distributed networks is impractical due to the cost of optical network deployment and the large number of users. Real-time monitoring of distributed networks places high demands on wireless communication. The average delay is within 15ms, the time service is less than 1 mus, and the reliability is 99.999%. 5G is considered a solution for distributed network real-time monitoring.
However, weather, data, user behavior, etc. make load prediction still a difficult problem. Most of the existing load prediction methods have some limitations. The traditional statistical method is difficult to process nonlinear data, and the intelligent calculation method has some problems in the aspects of artificial feature extraction, limited learning capability, difficult interpretation of training results and the like.
Disclosure of Invention
The invention aims to provide a method and a system for forecasting medium and short term loads of an intelligent power grid. The load prediction model can be simplified, and the load prediction precision can be improved.
In order to solve the technical problems, the technical scheme of the invention is as follows:
a method for forecasting medium-short term load of a smart grid, the method comprising the following steps:
acquiring daily power consumption total load data and hourly user power load data of a user;
determining a Deep Convolutional Neural Network (DCNN) prediction model obtained by training based on historical power load data; the historical user load data comprises daily power load data in preset days;
and inputting the daily total power consumption load data and the hourly user power load data into the DCNN prediction model for processing, and outputting a short-term load prediction result.
Optionally, determining a deep convolutional neural network DCNN prediction model obtained by training based on historical electrical load data includes:
acquiring an input data set, wherein the input data set comprises daily power load data in preset days;
and determining the DCNN prediction model according to the data set.
Optionally, the data set includes: at least one of maximum daily temperature, minimum daily temperature, humidity, precipitation weather parameters.
Optionally, determining the DCNN prediction model according to the data set includes:
preprocessing the data set to obtain a processing result;
performing characteristic selection based on mutual information to obtain mutual information between daily total load and the characteristics of the influence factors;
and determining the DCNN prediction model according to the processing result and the mutual information.
Optionally, preprocessing the data set to obtain a processing result, including:
according to the following formula:
Figure BDA0002751203780000031
preprocessing the data set to obtain a processing result;
where x (i) is the ith data record in the initial data set, xminIs the smallest data record, xmaxIs the maximum record of the data that can be recorded,
Figure BDA0002751203780000032
is the normalized result of the ith data record.
Optionally, the mutual information-based feature selection is performed to obtain mutual information between the daily total load and the influence factor features, where the mutual information includes:
mutual information between the input characteristic X and the daily total load Y is defined as:
Figure BDA0002751203780000033
wherein n and m are the sample numbers of random variables X and Y respectively; the probability of each possible value in X is p (X), the probability of each possible value in Y is p (Y), and p (xi, yj) is a joint probability density function of X and Y.
Optionally, determining the DCNN prediction model according to the processing result and the mutual information includes:
selecting different node numbers in training, comparing prediction results, and determining the number of hidden layer nodes and the number of hidden layers;
performing training from t 1 to t τ;
for a time step t, neuron parameters in layer i are updated according to the following formula:
Figure BDA0002751203780000034
Figure BDA0002751203780000035
Figure BDA0002751203780000036
Figure BDA0002751203780000037
Figure BDA0002751203780000038
wherein x is(t)Is tthInput data of step, y(t)Is the result of the corresponding prediction(s),
Figure BDA0002751203780000039
is the output actual value;
Figure BDA00027512037800000310
is at tthSharing state of layers; σ is laserAn excitation function;
Figure BDA00027512037800000311
is the input characteristic of layer l in the t-th cycle, W1And WNRespectively, the weather subset, U1And UNThe weighting coefficients of the first step and the Nth step are respectively, L is an error function, b1 and bNThe first step and the Nth step are respectively weight constants, and the loss _ function is used for iteratively adjusting the weight coefficients.
Optionally, inputting the daily total power consumption load data and the hourly user power load data into the DCNN prediction model for processing, and outputting a short-term load prediction result, including:
and when the system error delta is larger than the prediction error R and the training round R is smaller than the preset maximum training round R, inputting the daily total power consumption load data and the hourly user power load data into the DCNN prediction model for processing, and outputting a short-term load prediction result.
Optionally, the method for predicting the medium-short term load of the smart grid further includes:
acquiring load data root mean square error RMSE and average absolute percentage error MAPE of n predicted points;
evaluating a short-term load prediction result based on the RMSE and the MAPE.
The embodiment of the invention also provides a system for forecasting the medium-short term load of the smart grid, which comprises the following steps:
the data transmission layer is used for acquiring daily power consumption total load data and hourly user power load data of the users;
the intelligent electric surface layer SM is used for classifying the historical electric load data;
the intelligent substation IS IS used for processing the historical power load data;
the master station layer is used for training based on the historical power load data to obtain a Deep Convolutional Neural Network (DCNN) prediction model; the historical user load data comprises daily power load data in preset days;
and the data processing layer is used for inputting the daily total power consumption load data and the hourly user power load data into the DCNN prediction model for processing and outputting a short-term load prediction result.
The scheme of the invention at least comprises the following beneficial effects:
the scheme of the invention provides a daily time gradual load prediction method based on deep learning, which can accurately predict the short-term load of a user; a Deep Circulation Neural Network (DCNN) model is built for the total daily and hourly load of the user. The total daily load of the user is first predicted using the high-dimensional features, and then an hourly load prediction is made based on the total daily load and the selected low-dimensional features. Simulation results show that the method can simplify the load prediction model and improve the load prediction precision.
Drawings
FIG. 1 is a schematic flow chart of a short-term load forecasting method in a smart grid according to an embodiment of the invention;
FIG. 2 is a detailed flow diagram of short term load forecasting in a smart grid according to an embodiment of the present invention;
FIG. 3 is a schematic diagram of a DCNN network model according to an embodiment of the present invention;
FIG. 4 is a graphical illustration of the load prediction efficiency of different algorithms in an embodiment of the invention;
fig. 5 is a schematic architecture diagram of a short-term load prediction system in a smart grid according to an embodiment of the present invention.
Detailed Description
Exemplary embodiments of the present disclosure will be described in more detail below with reference to the accompanying drawings. While exemplary embodiments of the present disclosure are shown in the drawings, it should be understood that the present disclosure may be embodied in various forms and should not be limited to the embodiments set forth herein. Rather, these embodiments are provided so that this disclosure will be thorough and complete, and will fully convey the scope of the disclosure to those skilled in the art.
As shown in fig. 1, an embodiment of the present invention provides a method for predicting short-term load in a smart grid, where the method includes:
step 11, acquiring daily power consumption total load data of a user and hourly power consumption load data of the user;
step 12, determining a Deep Convolutional Neural Network (DCNN) prediction model obtained by training based on historical power load data; the historical user load data comprises daily power load data in preset days;
and step 13, inputting the daily total power consumption load data and the hourly user power load data into the DCNN prediction model for processing, and outputting a short-term load prediction result.
According to the embodiment of the invention, the daily electricity consumption total load data and the hourly user electricity load data of the user are obtained; determining a Deep Convolutional Neural Network (DCNN) prediction model obtained by training based on historical power load data; the historical user load data comprises daily power load data in preset days; and inputting the daily total power consumption load data and the hourly user power load data into the DCNN prediction model for processing, and outputting a short-term load prediction result, so that the short-term load of the user can be accurately predicted.
In an alternative embodiment of the present invention, step 12 may include:
step 121, acquiring an input data set, wherein the input data set comprises daily power load data in preset days; optionally, the data set includes: at least one of maximum daily temperature, minimum daily temperature, humidity, precipitation weather parameters, although other weather parameters may be included.
And step 122, determining the DCNN prediction model according to the data set.
In an optional embodiment of the present invention, step 122 may specifically include:
1221, preprocessing the data set to obtain a processing result;
specifically, the following formula can be used:
Figure BDA0002751203780000061
preprocessing the data set to obtain a processing result; where x (i) is the ith in the initial data setData record, xminIs the smallest data record, xmaxIs the maximum record of the data that can be recorded,
Figure BDA0002751203780000062
is the normalized result of the ith data record.
Step 1222, selecting features based on mutual information to obtain mutual information between daily total load and influencing factor features;
in a specific implementation, the mutual information between the input characteristic X and the daily total load Y is defined as:
Figure BDA0002751203780000063
wherein n and m are the sample numbers of random variables X and Y respectively; the probability of each possible value in X is p (X), the probability of each possible value in Y is p (Y), and p (xi, yj) is a joint probability density function of X and Y.
And 1223, determining the DCNN prediction model according to the processing result and the mutual information.
Specifically, the step may include: selecting different node numbers in training, comparing prediction results, and determining the number of hidden layer nodes and the number of hidden layers;
performing training from t 1 to t τ;
for a time step t, neuron parameters in layer i are updated according to the following formula:
Figure BDA0002751203780000064
Figure BDA0002751203780000065
Figure BDA0002751203780000066
Figure BDA0002751203780000067
Figure BDA0002751203780000068
wherein x is(t)Is tthInput data of step, y(t)Is the result of the corresponding prediction(s),
Figure BDA0002751203780000069
is the output actual value;
Figure BDA00027512037800000610
is at tthSharing state of layers; σ is the excitation function;
Figure BDA00027512037800000611
is the input characteristic of layer l in the t-th cycle, W1And WNRespectively, the weather subset, U1And UNThe weighting coefficients of the first step and the Nth step are respectively, L is an error function, b1 and bNThe first step and the Nth step are respectively weight constants, and the loss _ function is used for iteratively adjusting the weight coefficients.
The following describes a specific implementation process of the above embodiment of the present invention with reference to the steps of the load prediction method shown in fig. 2 and the training network shown in fig. 3:
first, data input is performed. The input data is a one-dimensional dataset having a plurality of elements. The data set of the daily load data set is relatively simple.
Secondly, model training is completed. Several specific parameters of the DCNN model are determined by iterative calculations based on the existing input data set.
Thirdly, feature selection based on mutual information is performed. The daily total load data already contains information of some input elements, so it is necessary to filter the elements. The basic idea is to calculate mutual information between the daily total load and the characteristics of the influencing factors.
Finally, a load prediction model having a plurality of parameters is determined.
And if the precision is smaller than a preset threshold value, obtaining a load prediction model. Otherwise, the calculation results are used as input for model training together with the initial input data set.
The data set needs to be preprocessed before the data is input.
The data set W includes daily maximum temperature, minimum temperature, humidity, precipitation and other weather parameters. Whether the date is divided into holidays (including the solar calendar festival and the female calendar festival), workdays and holidays. The supplement of missing data, the correction of erroneous data and the normalization should be performed in the data preprocessing. To standardize
Figure BDA0002751203780000071
Where x (i) is the ith data record in the initial data set, xminIs the smallest data record, xmaxIs the maximum record of the data that can be recorded,
Figure BDA0002751203780000072
is the normalized result of the ith data record.
For model training, the number of nodes in the input layer in the training network is related to the selected load-related data. The final output result is the total load value for the predicted day with one output node.
The number of hidden layer nodes and hidden layers is typically selected by trial and error.
By selecting different node numbers in training, the prediction results are compared, and then the relatively proper hidden layer node number and hidden layer number are determined.
Training is performed from t 1 to t τ. For a time step t, neuron parameters in layer i are updated according to the following formula:
Figure BDA0002751203780000081
Figure BDA0002751203780000082
Figure BDA0002751203780000083
Figure BDA0002751203780000084
Figure BDA0002751203780000085
x(t)is tthInput data of step, y(t)Is the result of the corresponding prediction(s),
Figure BDA0002751203780000086
is to output the actual value.
Figure BDA0002751203780000087
Is at tthAnd sharing state of the layer. σ is the excitation function, i.e., tanh function.
Figure BDA0002751203780000088
Is the input characteristic of layer i in the t-th cycle.
Mutual information between the input characteristic X and the daily total load Y is defined as:
Figure BDA0002751203780000089
where n, m are the number of samples of the random variables X and Y, respectively. The probability of each possible value in X is p (X), the probability of each possible value in Y is p (Y), and p (xi, yj) is a joint probability density function of X and Y.
The load prediction algorithm is as follows:
inputs: < W, H, D > (input data set W, H, D)
Begin (Begin)
3.While(δ>R&&r<R)
4.The historical data set<W,H,D>was pre-processed
Data set is partitioned into tracking set T and verification set V; (data set S is divided into training set T and validation set V)
Base on DCNN, the total load C of forecast day is output; (predicting the Total load C as output based on the DCNN model)
Basic on organizational information, a subset of weather data set W ', historical load data set H, date type D is selected to form a new input data set < W', H, D >; (based on mutual information, weather subset W', historical load set H, select date type D to form a new input data set)
The data set is partitioned into a training set T2 and a verification set V2 (data set is divided into a training set T2 and a verification set V2);
data set T2 is input into DCNN network for pre-training (T2 input DCNN network pre-training);
the verification set V2 is used for cross validation and super parameter optimization of The pre-trained model (cross validation and hyper-parameter optimization of pre-trained model);
the < W2', H2', D ' > of The required day input into The trained model for The required day, and The time load series L of The required day input into The trained model for The prediction day to obtain The time series L of The prediction day;
calculating the accuracy of daily load prediction A
13.IF(A<F)；
Re-select a subset of weather data (reselecting a subset of weather data);
15.Repeat 7-12；
16.End If；
17.End while；
18.End。
the system error is larger than the prediction error R, and the training round R is smaller than the preset maximum training round R. A load prediction model is trained within the circulation.
In an optional embodiment of the present invention, the inputting the daily total power consumption load data and the hourly user power load data into the DCNN prediction model for processing, and outputting a short-term load prediction result includes:
and when the system error delta is larger than the prediction error R and the training round R is smaller than the preset maximum training round R, inputting the daily total power consumption load data and the hourly user power load data into the DCNN prediction model for processing, and outputting a short-term load prediction result.
In an optional embodiment of the present invention, the method for predicting short-term load in a smart grid may further include:
acquiring load data root mean square error RMSE and average absolute percentage error MAPE of n predicted points;
evaluating a short-term load prediction result based on the RMSE and the MAPE.
The above-described embodiments of the present invention were tested using the public data set of the EUNITE network. The EUNITE network data set is the active power load data with a sampling interval of 15 minutes. That is, 96 sample points per day for a total of 28800 load points. The sampling time is, for example, 2016, 7, 19 days to 2016, 8, 9 days. The probability of neuron inactivation in the exit layer of the deep learning neural network was set to 0.5. The load data is preprocessed to ensure data quality. The average value is used for making up the missing value and judging an abnormal value. ARIMA and Support Vector Machines (SVM) are used as control groups.
Table 1 lists the parameters in this experiment.
Parameters Value
Network model DCNN
Hidden layer number 5
Neuron number in each layer 512
The Root Mean Square Error (RMSE) and the Mean Absolute Percentage Error (MAPE) were selected as evaluation indicators. RMSE calculates the square of the difference between the actual and predicted values. The larger the RMSE, the larger the error between the predicted value and the actual value. MAPE was chosen for use because it provides a percentage-based error for direct visualization. In addition, a trained DCNN model was developed for each algorithm using the data collected for two weeks, new test data was selected from random days in the week, and then predictions were made for that particular day. The calculation formula for MAPE and RMSE is:
Figure BDA0002751203780000101
Figure BDA0002751203780000102
where n is the number of prediction points, yiIs the actual load value at the ith prediction point,
Figure BDA0002751203780000103
is the predicted load value at the ith prediction point.
On the same training data set, ARIMA algorithm, SVM algorithm and our proposed DCNN were performed, respectively. Then, for example, the load point curve is continuously predicted every 15 minutes from 8-month 1-day to 8-month 7-day. The RMSE and MAPE indices for the three algorithms were calculated.
Figure BDA0002751203780000104
Figure BDA0002751203780000111
It can be seen that the DCNN algorithm proposed by the embodiments of the present invention is superior to the other two algorithms in terms of RMSE and MAPE. In addition, the DCNN algorithm has stable performance. Therefore, the algorithm has certain generalization capability. Within these 7 days, 8 months and 3 days are the rest days. The ARIMA algorithm has a significant performance degradation. This results in the ARIMA algorithm performing poorly on special days.
The load prediction efficiency of the different algorithms at different iteration numbers is shown in fig. 4. When ARIMA iterates to 20 rounds and the SVM algorithm iterates to 15 rounds, the prediction error starts to converge. The proposed DCNN algorithm starts to converge at round 70. Through analysis, the ARIMA and SVM algorithms are relatively simple, and a load prediction result can be quickly obtained. The prediction errors of the ARIMA and SVM algorithms converge to 15% and 12.3%, respectively. The DCNN algorithm converges to 6.5%. Therefore, the DCNN algorithm has high accuracy.
The above embodiment of the present invention proposes a time-of-day gradual load prediction method based on deep learning to cope with data explosion caused by 5G. A Deep Circulation Neural Network (DCNN) model is built for the total daily and hourly load of the user.
As shown in fig. 5, an embodiment of the present invention further provides a system for predicting short-term load in a smart grid, including:
the data transmission layer is used for acquiring daily power consumption total load data and hourly user power load data of the users;
the intelligent electric surface layer SM is used for classifying the historical electric load data;
the intelligent substation IS IS used for processing the historical power load data;
the master station layer is used for training based on the historical power load data to obtain a Deep Convolutional Neural Network (DCNN) prediction model; the historical user load data comprises daily power load data in preset days;
and the data processing layer is used for inputting the daily total power consumption load data and the hourly user power load data into the DCNN prediction model for processing and outputting a short-term load prediction result.
Determining a Deep Convolutional Neural Network (DCNN) prediction model obtained by training based on historical electrical load data, wherein the DCNN prediction model comprises the following steps:
acquiring an input data set, wherein the input data set comprises daily power load data in preset days;
and determining the DCNN prediction model according to the data set.
Optionally, the data set includes: at least one of maximum daily temperature, minimum daily temperature, humidity, precipitation weather parameters.
Optionally, determining the DCNN prediction model according to the data set includes:
preprocessing the data set to obtain a processing result;
performing characteristic selection based on mutual information to obtain mutual information between daily total load and the characteristics of the influence factors;
and determining the DCNN prediction model according to the processing result and the mutual information.
Optionally, preprocessing the data set to obtain a processing result, including:
according to the following formula:
Figure BDA0002751203780000121
preprocessing the data set to obtain a processing result;
where x (i) is the ith data record in the initial data set, xminIs the smallest data record, xmaxIs the maximum record of the data that can be recorded,
Figure BDA0002751203780000122
is of the ith data recordAnd normalizing the result.
Optionally, the mutual information-based feature selection is performed to obtain mutual information between the daily total load and the influence factor features, where the mutual information includes:
mutual information between the input characteristic X and the daily total load Y is defined as:
Figure BDA0002751203780000123
wherein n and m are the sample numbers of random variables X and Y respectively; the probability of each possible value in X is p (X), the probability of each possible value in Y is p (Y), and p (xi, yj) is a joint probability density function of X and Y.
Optionally, determining the DCNN prediction model according to the processing result and the mutual information includes:
selecting different node numbers in training, comparing prediction results, and determining the number of hidden layer nodes and the number of hidden layers;
performing training from t 1 to t τ;
for a time step t, neuron parameters in layer i are updated according to the following formula:
Figure BDA0002751203780000131
Figure BDA0002751203780000132
Figure BDA0002751203780000133
Figure BDA0002751203780000134
Figure BDA0002751203780000135
wherein x is(t)Is tthInput data of step, y(t)Is the result of the corresponding prediction(s),
Figure BDA0002751203780000136
is the output actual value;
Figure BDA0002751203780000137
is at tthSharing state of layers; σ is the excitation function;
Figure BDA0002751203780000138
is the input characteristic of layer l in the t-th cycle, W1And WNRespectively, the weather subset, U1And UNThe weighting coefficients of the first step and the Nth step are respectively, L is an error function, b1 and bNThe first step and the Nth step are respectively weight constants, and the loss _ function is used for iteratively adjusting the weight coefficients.
Optionally, inputting the daily total power consumption load data and the hourly user power load data into the DCNN prediction model for processing, and outputting a short-term load prediction result, including:
and when the system error delta is larger than the prediction error R and the training round R is smaller than the preset maximum training round R, inputting the daily total power consumption load data and the hourly user power load data into the DCNN prediction model for processing, and outputting a short-term load prediction result.
Optionally, the data processing layer further includes:
acquiring load data root mean square error RMSE and average absolute percentage error MAPE of n predicted points;
evaluating a short-term load prediction result based on the RMSE and the MAPE.
The system can accurately predict the load and realize safe and economic operation and scientific management of the power system.
1) A data transmission layer: the data transmission layer mainly focuses on data acquisition and screening of load prediction. The sensor devices transmit temperature, humidity, weather conditions and other data to the Smart Meter (SM) layer. The SM mainly records load data each time. At this level, the load data is typically based on the home. As a first level of edge devices, SMs not only have the function of summarizing and perceiving data, but also can perceive and classify the content of the data. Finally, the SM packages and transmits the preliminary data to the Intelligent Substation (IS) layer.
The IS collects data transmitted by the SM layer. The data acquired by the IS layer IS usually one area or several areas. There is also a large difference between is and SM. A large number of electric power system instruments and facilities are deployed in the intelligent substation, and electric parameters such as voltage, current and phase can be obtained. Meanwhile, the IS IS used as a second stage of the edge device and bears partial data processing tasks. And the IS layer transmits the processed data and the data to be processed in the cloud to the master station layer.
As a cloud data center, the main distribution station layer mainly takes the tasks of large data storage and data operation. And the IS transmits the calculation result to the main distribution station for storage, and transmits data required by the task needing to be put into the cloud calculation to the main station. In addition, the primary distribution site layer sends the historical data to the IS, while sending deep learning control information to the IS.
2) A data processing layer: the data processing layer mainly studies data processing of load prediction in the smart grid under the edge computing environment. Since all users are directly load predicted at a time, the number of neurons in the hidden layer will be excessive. This will result in a reduction in computational efficiency. In an edge computing environment, data processing tasks with high energy consumption and low time delay requirements are deployed on a main distribution substation layer, and data processing tasks with low energy consumption and high time delay requirements are deployed in edge equipment such as an IS. Therefore, the data transmission can be reduced by transferring the calculation task to the edge device, and the edge device is closer to renewable energy sources, so that a more green effect can be achieved.
3) And (3) a task control layer: the task control layer mainly focuses on the scenes of the edge devices (IS, SM and the like) and the cloud operation load prediction tasks. By scheduling load prediction tasks, system latency and power consumption may be reduced to improve performance of the prediction system.
The embodiment of the invention provides a time-per-day gradual load prediction method based on deep learning, aiming at the problem that the dimension of the short-term load prediction model parameter of a user is too high. A Deep Circulation Neural Network (DCNN) model is built for the total daily and hourly load of the user. The total daily load of the user is first predicted using the high-dimensional features, and then an hourly load prediction is made based on the total daily load and the selected low-dimensional features. Simulation results show that the method can simplify the load prediction model and improve the load prediction precision.
While the foregoing is directed to the preferred embodiment of the present invention, it will be understood by those skilled in the art that various changes and modifications may be made without departing from the spirit and scope of the invention as defined in the appended claims.

Claims (10)

1. A method for forecasting medium-short term load of a smart grid is characterized by comprising the following steps:
acquiring daily power consumption total load data and hourly user power load data of a user;
determining a Deep Convolutional Neural Network (DCNN) prediction model obtained by training based on historical power load data; the historical user load data comprises daily power load data in preset days;
and inputting the daily total power consumption load data and the hourly user power load data into the DCNN prediction model for processing, and outputting a short-term load prediction result.
2. The method for predicting the short-term load in the smart grid according to claim 1, wherein determining a Deep Convolutional Neural Network (DCNN) prediction model obtained by training based on historical electric load data comprises:
acquiring an input data set, wherein the input data set comprises daily power load data in preset days;
and determining the DCNN prediction model according to the data set.
3. The method for forecasting short-term load in a smart grid according to claim 2, wherein the data set comprises: at least one of maximum daily temperature, minimum daily temperature, humidity, precipitation weather parameters.
4. The method for predicting short-term load in a smart grid according to claim 1, wherein determining the DCNN prediction model according to the data set comprises:
preprocessing the data set to obtain a processing result;
performing characteristic selection based on mutual information to obtain mutual information between daily total load and the characteristics of the influence factors;
and determining the DCNN prediction model according to the processing result and the mutual information.
5. The method for predicting the short-term load in the smart grid according to claim 4, wherein the preprocessing is performed on the data set to obtain a processing result, and the method comprises the following steps:
according to the following formula:
Figure FDA0002751203770000011
preprocessing the data set to obtain a processing result;
where x (i) is the ith data record in the initial data set, xminIs the smallest data record, xmaxIs the maximum record of the data that can be recorded,
Figure FDA0002751203770000012
is the normalized result of the ith data record.
6. The method for forecasting the short-term load in the smart grid according to claim 4, wherein the mutual information-based feature selection is performed to obtain mutual information between the daily total load and the influence factor features, and the method comprises the following steps:
mutual information between the input characteristic X and the daily total load Y is defined as:
Figure FDA0002751203770000021
wherein n and m are the sample numbers of random variables X and Y respectively; the probability of each possible value in X is p (X), the probability of each possible value in Y is p (Y), and p (xi, yj) is a joint probability density function of X and Y.
7. The method for predicting the medium-short term load in the smart grid according to claim 4 or 6, wherein the determining the DCNN prediction model according to the processing result and the mutual information comprises:
selecting different node numbers in training, comparing prediction results, and determining the number of hidden layer nodes and the number of hidden layers;
performing training from t 1 to t τ;
for a time step t, neuron parameters in layer i are updated according to the following formula:
Figure FDA0002751203770000022
Figure FDA0002751203770000023
Figure FDA0002751203770000024
Figure FDA0002751203770000025
Figure FDA0002751203770000026
wherein x is(t)Is tthInput data of step, y(t)Is the result of the corresponding prediction(s),
Figure FDA0002751203770000027
is the output actual value;
Figure FDA0002751203770000028
is at tthSharing state of layers; σ is the excitation function;
Figure FDA0002751203770000029
is the input characteristic of layer l in the t-th cycle, W1And WNRespectively, the weather subset, U1And UNThe weighting coefficients of the first step and the Nth step are respectively, L is an error function, b1 and bNThe first step and the Nth step are respectively weight constants, and the loss _ function is used for iteratively adjusting the weight coefficients.
8.The method for predicting the short-term load in the intelligent power grid according to claim 1, wherein the step of inputting the daily total load data of power consumption and the hourly user load data of the power consumption into the DCNN prediction model for processing and outputting a short-term load prediction result comprises the steps of:
and when the system error delta is larger than the prediction error R and the training round R is smaller than the preset maximum training round R, inputting the daily total power consumption load data and the hourly user power load data into the DCNN prediction model for processing, and outputting a short-term load prediction result.
9. The method for forecasting the short-term load in the smart grid according to claim 1, further comprising:
acquiring load data root mean square error RMSE and average absolute percentage error MAPE of n predicted points;
evaluating a short-term load prediction result based on the RMSE and the MAPE.
10. A system for forecasting medium and short term load of a smart grid is characterized by comprising the following components:
the data transmission layer is used for acquiring daily power consumption total load data and hourly user power load data of the users;
the intelligent electric surface layer SM is used for classifying the historical electric load data;
the intelligent substation IS IS used for processing the historical power load data;
the master station layer is used for training based on the historical power load data to obtain a Deep Convolutional Neural Network (DCNN) prediction model; the historical user load data comprises daily power load data in preset days;
and the data processing layer is used for inputting the daily total power consumption load data and the hourly user power load data into the DCNN prediction model for processing and outputting a short-term load prediction result.
CN202011185160.XA 2020-10-30 2020-10-30 Method and system for forecasting medium and short term loads of smart power grid Pending CN112348070A (en)

Priority Applications (1)

Application Number Priority Date Filing Date Title
CN202011185160.XA CN112348070A (en) 2020-10-30 2020-10-30 Method and system for forecasting medium and short term loads of smart power grid

Applications Claiming Priority (1)

Application Number Priority Date Filing Date Title
CN202011185160.XA CN112348070A (en) 2020-10-30 2020-10-30 Method and system for forecasting medium and short term loads of smart power grid

Publications (1)

Publication Number Publication Date
CN112348070A true CN112348070A (en) 2021-02-09

Family

ID=74355721

Family Applications (1)

Application Number Title Priority Date Filing Date
CN202011185160.XA Pending CN112348070A (en) 2020-10-30 2020-10-30 Method and system for forecasting medium and short term loads of smart power grid

Country Status (1)

Country Link
CN (1) CN112348070A (en)

Citations (7)

* Cited by examiner, † Cited by third party
Publication number Priority date Publication date Assignee Title
CN108334988A (en) * 2018-02-08 2018-07-27 吕欣 A kind of short-term Load Forecasting based on SVM
CN109359786A (en) * 2018-12-05 2019-02-19 国网江苏省电力有限公司扬州供电分公司 A kind of power station area short-term load forecasting method
CN109754113A (en) * 2018-11-29 2019-05-14 南京邮电大学 Load forecasting method based on dynamic time warping Yu length time memory
CN110674993A (en) * 2019-09-26 2020-01-10 广东电网有限责任公司 User load short-term prediction method and device
CN111126663A (en) * 2019-11-25 2020-05-08 黑龙江电力调度实业有限公司 Short-term power load prediction method
CN111582548A (en) * 2020-04-14 2020-08-25 广东卓维网络有限公司 Power load prediction method based on multivariate user behavior portrait
CN111783953A (en) * 2020-06-30 2020-10-16 重庆大学 24-point power load value 7-day prediction method based on optimized LSTM network

Patent Citations (7)

* Cited by examiner, † Cited by third party
Publication number Priority date Publication date Assignee Title
CN108334988A (en) * 2018-02-08 2018-07-27 吕欣 A kind of short-term Load Forecasting based on SVM
CN109754113A (en) * 2018-11-29 2019-05-14 南京邮电大学 Load forecasting method based on dynamic time warping Yu length time memory
CN109359786A (en) * 2018-12-05 2019-02-19 国网江苏省电力有限公司扬州供电分公司 A kind of power station area short-term load forecasting method
CN110674993A (en) * 2019-09-26 2020-01-10 广东电网有限责任公司 User load short-term prediction method and device
CN111126663A (en) * 2019-11-25 2020-05-08 黑龙江电力调度实业有限公司 Short-term power load prediction method
CN111582548A (en) * 2020-04-14 2020-08-25 广东卓维网络有限公司 Power load prediction method based on multivariate user behavior portrait
CN111783953A (en) * 2020-06-30 2020-10-16 重庆大学 24-point power load value 7-day prediction method based on optimized LSTM network

Non-Patent Citations (3)

* Cited by examiner, † Cited by third party
Title
RODRIGUES FILIPE 等: "The daily and hourly energy consumption and load forecasting using artificial neural network method: a case study using a set of 93 households in Portugal", 《6TH INTERNATIONAL CONFERENCE ON SUSTAINABILITY IN ENERGY AND BUILDINGS (SEB)》 *
余国晓: "机器学习在短期电力负荷预测中的应用研究", 《中国优秀博硕士学位论文全文数据库(硕士)》 *
王彬 等: "基于短期预测信息和长期值函数近似的大规模电动汽车实时随机优化调度算法", 《电力系统保护与控制》 *

Similar Documents

Publication Publication Date Title
Li et al. Predicting short-term electricity demand by combining the advantages of arma and xgboost in fog computing environment
Ghadimi et al. PSO based fuzzy stochastic long-term model for deployment of distributed energy resources in distribution systems with several objectives
Dou et al. Hybrid model for renewable energy and loads prediction based on data mining and variational mode decomposition
CN110766200A (en) Method for predicting generating power of wind turbine generator based on K-means mean clustering
CN108717579B (en) Short-term wind power interval prediction method
CN110380444B (en) Capacity planning method for distributed wind power orderly access to power grid under multiple scenes based on variable structure Copula
CN115293326A (en) Training method and device of power load prediction model and power load prediction method
Kaplan et al. A novel method based on Weibull distribution for short-term wind speed prediction
Hu et al. Short-term photovoltaic power prediction based on similar days and improved SOA-DBN model
CN117013527A (en) Distributed photovoltaic power generation power prediction method
CN115758151A (en) Combined diagnosis model establishing method and photovoltaic module fault diagnosis method
Fan et al. Spatial-temporal genetic-based attention networks for short-term photovoltaic power forecasting
CN114021848A (en) Generating capacity demand prediction method based on LSTM deep learning
CN111697560B (en) Method and system for predicting load of power system based on LSTM
CN112348236A (en) Abnormal daily load demand prediction system and method for intelligent power consumption monitoring terminal
Shang et al. Loss of life estimation of distribution transformers considering corrupted AMI data recovery and field verification
CN116167465A (en) Solar irradiance prediction method based on multivariate time series ensemble learning
Yuan et al. Short-term wind speed forecasting using STLSSVM hybrid model
CN116128211A (en) Wind-light-water combined short-term optimization scheduling method based on wind-light uncertainty prediction scene
CN113762591B (en) Short-term electric quantity prediction method and system based on GRU and multi-core SVM countermeasure learning
CN115544871A (en) Distributed robust energy storage planning method considering renewable power supply space correlation
CN112348070A (en) Method and system for forecasting medium and short term loads of smart power grid
Prashanthi et al. A comparative study of the performance of machine learning based load forecasting methods
CN113379116A (en) Cluster and convolutional neural network-based line loss prediction method for transformer area
Dong et al. Image-based with peak load ensemble prediction system for demand response in smart grid

Legal Events

Date Code Title Description
PB01 Publication
PB01 Publication
SE01 Entry into force of request for substantive examination
SE01 Entry into force of request for substantive examination
RJ01 Rejection of invention patent application after publication

Application publication date: 20210209

RJ01 Rejection of invention patent application after publication