A method for short-term load rate actual prediction by using a multi-target space-time graph convolutional neural network GCN-BILSTM model

Abstract

A kind of method for short-term load rate actual prediction using multi-objective space-time graph convolutional neural network GCN-BILSTM model, comprising: obtaining the load data and weather data of target area from data center and meteorological website;Data cleaning is carried out to the obtained load data and weather data;The data cleaning data is processed by feature engineering, including: the original temperature is split into maximum temperature, minimum temperature, and the maximum temperature and minimum temperature are respectively transformed by increasing a preset value, so that the minimum temperature is greater than or equal to 0;The load rate, maximum temperature and minimum temperature of the town subordinate to the target area are standardized to 0-1, to obtain several dimensional data;GCN-BILSTM model is constructed, and the several dimensional data are input into GCN-BILSTM model to output prediction result.The "linkage" between space nodes is increased in the prediction process, and the mape error is lower compared with Multi-BILSTM algorithm.

Description

TECHNICAL FIELD

[0001] The present application relates to the field of artificial intelligence algorithms, and in particular to a method for short-term load rate actual prediction using a multi-target space-time graph convolutional neural network GCN-BILSTM model. BACKGROUND

[0002] Short-term load rate actual prediction (referring to predicting the true future value) is the basis of economic operation of the power system. If heavy load, overload and other phenomena occur, it will increase the power grid load and increase the probability of power outage accidents, which will have a great impact on the production and life of enterprises and residents. The minimum range of power grid load rate is the transformer area. However, there are hundreds or thousands of transformer areas in each city. If these transformer areas are predicted one by one, it will be time-consuming and laborious. However, predicting the overall load rate of the city is not accurate enough. The present application selects towns as the observation object to predict the daily load rate of 8 towns in S area at the same time.

[0003] Since it involves the prediction of multiple targets (multi-task, i.e. the load rate of multiple towns), there are two traditional prediction methods. The first is single-target cycle prediction, which is time-consuming. The second is time series prediction based on the multi-target LSTM algorithm (MultiTask-LSTM, referred to as Multi-LSTM in the present application, which is different from the commonly used single-target prediction algorithm LSTM; according to the actual function, the multi-target GCN-BILSTM of the present application should be called GCN-MultiTask-BILSTM, but it is still customary to call it GCN-BILSTM, which is the same as the single-target GCN-BILSTM). However, it is only based on the time series level, i.e. the targets exist "in isolation". This model only handles the relationship at the time level, but does not consider the "linkage" between targets, i.e. ignores the spatial relationship. SUMMARY

[0004] In order to solve the defects ignored by traditional methods in multi-target prediction, the present application is based on the multi-target unidirectional LSTM algorithm, and is changed to bidirectional conduction time series Multi-BILSTM, and the "linkage" between spatial nodes is increased in the prediction process. A GCN-BILSTM space-time graph convolutional neural network model is established to process data from the spatial level and the time series level, and each town is taken as a graph node to simultaneously predict the daily load rate of multiple towns in S area.

[0005] A method for short-term load rate actual prediction using a multi-target space-time graph convolutional neural network GCN-BILSTM model, comprising the following steps:

[0006] Step one, obtaining load data and weather data of the target area from the data platform and the weather website, the load data including date, load rate of the town under the target area, the weather data including date and temperature of the target area;

[0007] Step two, data cleaning of the load data and weather data obtained in step one;

[0008] Step three, feature engineering processing of the data after data cleaning in step two, including: splitting the original temperature into maximum temperature and minimum temperature, and respectively transforming the maximum temperature and the minimum temperature by adding a preset value, so that the minimum temperature is greater than or equal to 0; 0-1 standardization of the load rate of the town under the target area, the maximum temperature and the minimum temperature, to obtain several dimension data;

[0009] Step four, constructing a GCN-BILSTM model, inputting the several dimension data obtained in step three into the GCN-BILSTM model to output the prediction result.

[0010] Further, the data cleaning in step two includes outlier deletion, missing value processing, numerical value processing and splicing data, wherein the splicing data is the horizontal splicing of the load data and the weather data according to the date.

[0011] Further, step four specifically includes:

[0012] (1) constructing a GCN graph structure network: the graph structure includes nodes and edges of the graph, each town is set as a graph node, all nodes are firstly correlated with each other by Spearman, the several dimension data obtained in step three are input into the GCN network, and the correlation greater than 0.5 is taken as the edge between two graph nodes, so as to establish the nodes and edges of the GCN network;

[0013] (2) constructing a BILSTM network: two LSTM layers are constructed to process forward and reverse sequence data respectively, the outputs of the two LSTM layers are combined together to provide more comprehensive sequence modeling capability, then a fully connected layer is constructed, the fully connected layer has 8 nodes for outputting results;

[0014] (3) constructing a GCN-BILSTM model: firstly, convolution processing of the graph structure, through convolution operation, each node updates its own feature vector, or considers the feature vector of its neighbor node; then, the nodes in the graph structure are aggregated through pooling operation; finally, the pooled result is transmitted into the BILSTM network, thereby forming the GCN-BILSTM model for time series prediction;

[0015] (4) After the GCN-BILSTM model is constructed, the GCN-BILSTM model is trained and corrected, and real prediction is performed by using the GCN-BILSTM model, ReLU is used as an activation function, batch_size is 30, and 14 consecutive days are used as a history period, the daily target area subordinate township daily load rate is synchronously output, and daily single-step cycle rolling training is performed; wherein the historical data set is divided into 6:2:2, the first 6 and the middle 2 are used for training the model, and the last 2 are the correction set; after the model is trained and corrected, the model is used for real prediction: the last 1 history period of the correction set is used for prediction, the output of the 1 row of results is added to the last 1 row of the historical training set, and the cycle is added for 7 times, at this time, the data set becomes the historical data set + 7 rows of new data, and the 7 rows of new data are the final real prediction results.

[0016] The application firstly uses the Spearman correlation to construct the edges of the GCN algorithm graph node, firstly proposes a multi-target GCN-BILSTM based on the GCN-LSTM, and is used for multi-target actual prediction of the power grid load rate; in addition, the application uses a single GCN-BILSTM to simultaneously predict multiple targets (simultaneously predict multiple Ys), compared with the prior art GCN-BILSTM which can only predict a single target (predict one Y each time), under the same hardware conditions and the same number of prediction targets, the running speed of the application is more advantageous; the "linkage" between the space nodes in the prediction process of the application is reflected in the process of constructing the edges of the graph nodes in the GCN through the Spearman correlation, the historical correlation coefficients between the subordinate township prediction results can be calculated through the training set, and the new correlation coefficients between the prediction results should also be close to the historical correlation coefficients when the prediction is performed, and the Multi-BILSTM does not have this process, and only predicts the daily load rates from the time sequence level. BRIEF DESCRIPTION OF DRAWINGS

[0017] Figure 1 is a flowchart of a method for short-term load rate actual prediction by using a multi-target space-time graph convolutional neural network GCN-BILSTM model according to an embodiment of the application;

[0018] Figure 2 is a graph structure of 8 townships in S area according to an embodiment of the application;

[0019] Figure 3 is a mape error map of two algorithms for predicting 20230601-20230607 according to an embodiment of the application;

[0020] Figure 4 is a mape error map of two algorithms for predicting 202306024-20230630 according to an embodiment of the application. DETAILED DESCRIPTION

[0021] In order to make the purposes, technical solutions and advantages of the embodiments of the present application clearer, the technical solutions in the embodiments of the present application will be described clearly and completely below with reference to the drawings in the embodiments of the present application. Obviously, the described embodiments are only a part of the embodiments of the present application, rather than all the embodiments. Based on the embodiments in the present application, all other embodiments obtained by a person of ordinary skill in the art without creative work fall within the protection scope of the present application.

[0022] Please refer to Figure 1 The embodiments of the present application provide a method for short-term load rate actual prediction by using a multi-target space-time graph convolutional neural network GCN-BILSTM model, comprising the following steps:

[0023] Step one, obtain the load data and weather data of the target area, the data set is derived from the data center and the meteorological website, the period is S area 20220101-20230531, and the time granularity is per day. The load data includes date, load rate of the target area (8 towns under S area, such as Figure 2 The weather data includes date and temperature of the target area S area.

[0024] Step two, data cleaning is performed on the load data and weather data obtained in step one, specifically including:

[0025] (1) Abnormal value deletion; through previewing data, it is found that the data of 20220101-20220110 is about 10 times higher than the load rate, which is deleted;

[0026] (2) Missing value processing, the missing date data is filled by using the convenient mean filling method;

[0027] (3) Numerical value processing, the load rate calculation has been generated in the database, but it still needs to be converted from decimal form to percentage form;

[0028] (4) Splicing data, the load data and weather data are horizontally spliced according to date.

[0029] Step three, feature engineering processing is performed on the data after data cleaning in step two, to obtain a plurality of dimension data, specifically as follows:

[0030] (1) Splitting temperature and transformation, the original temperature is split into maximum temperature and minimum temperature, and the transformation is performed, the formula is as follows:

[0031] Maximum temperature 1 = maximum temperature + 50

[0032] Minimum temperature 1 = minimum temperature + 50

[0033] (2) 0-1 standardization of features, the load rate of town 1-town 8, the maximum temperature 1, the minimum temperature 1, the 10 variables are 0-1 standardized, after the temperature transformation in the previous stage, the field minimum temperature 1 does not have negative numbers at this time, which does not affect the results of standardization.

[0034] Up to now, the fields to be input into the model have a total of 10 columns, which are: town 1 load rate, town 2 load rate, town 3 load rate, town 4 load rate, town 5 load rate, town 6 load rate, town 7 load rate, town 8 load rate, maximum temperature 1, and minimum temperature 1. Among them, town 1 load rate-town 8 load rate are both dependent variables Y and independent variables X (because 8 town load rates are predicted at the same time each time, but the load rates of the 8 towns are mutually influenced and are subject to historical correlation, so they can also be regarded as X), maximum temperature 1 and minimum temperature 1 are independent variables X.

[0035] Step four, build GCN-BILSTM model, input the dimension data obtained in step three into GCN-BILSTM model to output the prediction result.

[0036] First, the graph structure required by the GCN algorithm is constructed, including the nodes and edges of the graph. Input the 10-dimensional data (8 town load rates, maximum temperature, minimum temperature) into the GCN-BILSTM model, select the Spearman correlation method to correlate the 10 dimensions with each other, set the correlation greater than 0.5 as the threshold, only the correlation of the 8 town load rates meets the requirements, if the correlation between 2 nodes is less than 0.5, there is no connecting edge between the 2 graph nodes. Since the 8 towns under the S area are closely connected in economic activities, the daily load rates of the towns are closely related, and the Spearman correlation is greater than 0.5, that is, each graph node has 7 edges (see Figure 2 , node 0-node 7 represents town 1-town 8). But the Spearman correlation between the 2 variables of maximum temperature and minimum temperature and the other 8 variables is not strong, so they do not enter the subsequent modeling process, that is, these 2 variables do not have an effective effect on the final result of the model (but data cleaning is still necessary at the beginning, because it is not clear at the beginning that the role of these 2 variables, and it cannot be said that they do not have an effect at the beginning).

[0037] Then perform convolution processing, in the graph structure, convolution operation can be understood as information transmission between nodes. Through convolution operation, each node can update its own feature vector, and also consider the feature vector of its neighbor node.

[0038] Then perform pooling processing, the purpose of using the pooling operation is to reduce the dimension of the graph structure. Through the pooling operation, the nodes in the graph structure can be aggregated, thereby reducing the calculation amount and improving the generalization ability of the model.

[0039] Finally, the results of the pooling are passed into the BILSTM network. With ReLU as the activation function, batch_size is 30, and 14 consecutive days are taken as a history period. Eight fully connected layer nodes are used to output the daily load rate of eight townships synchronously. The single-step rolling prediction of each day is performed.

[0040] Specifically, the input_size is 10 in the initial args parameter setting, but since the highest temperature and the lowest temperature do not meet the threshold of Spearman correlation, the input_size is actually 8 in the subsequent modeling, that is, x=(batch_size=30, input_size=8, hidden_size=64). After convolution, 30(8, 64) is obtained, and then the maximum pooling is performed, and the result is (batch_size=30, seq_len=14, input_size=8). The BILSTM network is input.

[0041] Step five, result analysis

[0042] The data set date is S area 20220101-20230531, and the ratio of training set: test set: correction set is 6:2:2. The actual prediction date is 20230601-20230607 (the June data is not added to the data set in advance, and it is added afterwards in order to calculate the mape). The prediction results are shown in Table 1 and Figure 3 .

[0043] The calculation formula of the mean absolute percentage error is:

[0044]

[0045] Table 1 Mape error of two algorithms for predicting 20230601-20230607

[0046] Algorithm results Town 1 Town 2 Town 3 Town 4 Town 5 Town 6 Town 7 Town 8 Average Multi-BILSTM 14.6％ 24.8％ 7.0％ 13.9％ 2.8％ 10.5％ 15.4％ 4.7％ 11.7％ GCN-BILSTM 5.7％ 24.6％ 7.3％ 11.9％ 3.9％ 4.8％ 15.5％ 2.1％ 9.5％

[0047] In order to ensure the accuracy of the model results, the date is changed for re-prediction. The data set date is S area 20220101-20230623, and the ratio of training set: test set: correction set is 6:2:2. The actual prediction date is 20230624-20230630 (the June data is not added to the data set in advance, and it is added afterwards in order to calculate the mape). The prediction results are shown in Table 2 and Figure 4 .

[0048] Table 2 Mape error of two algorithms for predicting 202306024-20230630

[0049]

[0050] Figure 3 The mape error map of two algorithms for predicting 20230601-20230607 is as follows, Figure 4 The mape error map of two algorithms for predicting 202306024-20230630 is as follows.

[0051] From the above results, although the prediction results of the two periods are different, after increasing the node “linkage” in the space level, the mape error of the GCN-BILSTM model is always lower than that of the Multi-BILSTM model.

[0052] It should also be noted that the difficulty of multi-target prediction is generally higher than that of single-target prediction, and there is inevitably a large distribution difference between multiple prediction results Y, for example, town 2 and town 7 in table 1 have a large difference in data distribution from other towns, resulting in a high mape.

[0053] The application proposes a GCN-BILSTM model for multi-target short-term load rate actual prediction, in order to solve the defects ignored by traditional methods in multi-target prediction, according to the Spearman correlation between two nodes, the “linkage” between spatial nodes is increased in the prediction process, and compared with the Multi-BILSTM algorithm, the mape error is lower.

[0054] The above is only a specific embodiment of the application, but the protection scope of the application is not limited thereto, any changes or replacements within the technical scope disclosed by the application can be easily thought of by any person skilled in the art, and should be covered within the protection scope of the application. Therefore, the protection scope of the application should be subject to the protection scope of the claims.

Claims

1. A method for short-term load rate actual prediction using a multi-objective spatio-temporal graph convolutional neural network (GCN-BILSTM) model, characterized in that, Comprise the following steps: Step one, obtain the load data and weather data of the target area from the data platform and the weather website, the load data includes date, load rate of the target area subordinate town, the weather data includes date and temperature of the target area; Step two, data cleaning is carried out on the load data and weather data obtained in step one; Step three, feature engineering processing is carried out on the data after data cleaning in step two, including: the original temperature is split into maximum temperature and minimum temperature, and the maximum temperature and minimum temperature are respectively transformed by adding a preset value, so that the minimum temperature is greater than or equal to 0; The load rate of the target area subordinate town, the maximum temperature and the minimum temperature are standardized to 0-1, and a plurality of dimension data are obtained; Step four, build GCN-BILSTM model, input the plurality of dimension data processed in step three into GCN-BILSTM model to output prediction result; Step four specifically comprises: (1) build GCN graph structure network: the graph structure includes nodes and edges of the graph, each town is set as a graph node, the plurality of dimension data processed in step three are input into the GCN network, and the correlation greater than 0.5 is taken as the edge between the two graph nodes, so as to establish the nodes and edges of the GCN network; (2) build BILSTM network: two LSTM layers are built to process forward and reverse sequence data respectively, the outputs of the two LSTM layers are combined together to provide more comprehensive sequence modeling capability, and then a fully connected layer is built, which has 8 nodes for output result; (3) build GCN-BILSTM model: first, the convolution processing of the graph structure, through the convolution operation, each node updates its own feature vector, or considers the feature vector of its neighbor node; Then, the nodes in the graph structure are aggregated through the pooling operation; Finally, the pooled result is transmitted into the BILSTM network, thereby forming the GCN-BILSTM model for time series prediction; (4) after building the GCN-BILSTM model, start training, correcting the GCN-BILSTM model and using the GCN-BILSTM model for real prediction, take ReLU as the activation function, take batch_size as 30, take the consecutive 14 days as a history period, and output the daily load rate of the target area subordinate town every day, and carry out daily single step cycle rolling training; In which, the historical data set is divided into 6:2:2, the first 6 and the middle 2 are used for training model, and the last 2 are the correction set; After training and correcting the model, the model is used for real prediction: the last 1 history period of the correction set is taken for prediction, the output of the 1 row result is added to the last 1 row of the historical training set, and the cycle is added for 7 times, at this time the data set becomes historical data set+7 new data, which is the final real prediction result.

2. The method for short-term load rate actual prediction using a multi-target space-time graph convolutional neural network GCN-BILSTM model according to claim 1, characterized in that: The data cleaning in step two includes outlier deletion, missing value processing, numerical value processing and splicing data, wherein the splicing data is horizontally spliced according to date.

Images