A method for power grid load rate prediction using a GAT-LSTM & STGCN-MLP dual spatiotemporal graph model

Abstract

The application provides a kind of method for power grid load rate prediction using GAT-LSTM&STGCN-MLP dual space-time graph model, comprising: obtaining target area load data and weather data from data center and meteorological website;The load data of the target area and the weather data are cleaned, the load data includes date, the load rate of the target area town, and the weather data includes date, the temperature of the target area;The data after data cleaning is processed by feature engineering, to obtain a plurality of dimension data;Based on GAT-LSTM space-time graph algorithm network and STGCN-MLP space-time graph algorithm network, a dual space-time graph multi-objective algorithm model is constructed;The obtained several dimension data are input into the dual space-time graph multi-objective algorithm model for prediction, to obtain the load rate prediction result of the town subordinate to the target area.The application can still maintain stable prediction under various complex and changeable holiday conditions, and enhances the stability of the model.

Description

TECHNICAL FIELD

[0001] The application relates to the field of artificial intelligence algorithms, and particularly relates to a method for predicting power grid load rate by using a GAT-LSTM&STGCN-MLP double spatiotemporal graph model. BACKGROUND

[0002] The power grid load rate is one of important indexes of power grid safety, and if heavy load and overload and the like occur, the power grid burden is increased, the probability of a large-scale power failure accident is increased, and the power use of residents and enterprises is affected. The application takes an S area as an observation object, and actually predicts daily load rates of eight towns under the S area.

[0003] Since the prediction involves multiple targets (i.e. the load rates of multiple towns), the single-target cycle prediction method is too time-consuming and is not considered, and the existing multi-target prediction method is expanded from the initial pure time sequence level to the spatiotemporal graph level (time sequence + space). However, the existing single spatiotemporal graph algorithm mostly only considers the final error, and does not consider whether the error difference between the targets is too large, and the model error is not stable enough when a holiday is encountered, which is not conducive to actual use. SUMMARY

[0004] In order to solve the above problems in the prior art, the application provides a double spatiotemporal graph multi-target prediction model GAT-LSTM&STGCN-MLP for predicting the power grid load rate, and the anti-interference performance is high.

[0005] A method for predicting the power grid load rate by using a GAT-LSTM&STGCN-MLP double spatiotemporal graph model, comprising the following steps:

[0006] Step one, obtaining load data and weather data of a target area from a data center and a weather website;

[0007] Step two, performing data cleaning on the load data and weather data of the target area obtained in step one, wherein the load data comprises date and load rates of towns under the target area, and the weather data comprises date and air temperature of the target area;

[0008] Step three, performing feature engineering processing on the data cleaned in step two, to obtain a plurality of dimension data;

[0009] Step four, constructing a double spatiotemporal graph multi-target algorithm model based on a GAT-LSTM spatiotemporal graph algorithm network and a STGCN-MLP spatiotemporal graph algorithm network;

[0010] Step five, inputting the plurality of dimension data obtained in step three into the double spatiotemporal graph multi-target algorithm model for prediction, to obtain a load rate prediction result of the towns under the target area.

[0011] Further, step three specifically comprises:

[0012] The original air temperature is divided into maximum temperature and minimum temperature, and is transformed, and the formula is as follows:

[0013] Maximum temperature 1 = maximum temperature + 50

[0014] Minimum temperature 1 = minimum temperature + 50

[0015] The load rate, maximum temperature 1 and minimum temperature 1 of the target area subordinate town are standardized to 0-1.

[0016] Further, step four specifically includes:

[0017] Setting GAT-LSTM spatio-temporal graph algorithm network: constructing the graph structure required by the GAT algorithm, including the nodes and edges of the graph, setting each town of the target area as a graph node, and adding maximum temperature 1 and minimum temperature 1 as virtual towns to become two nodes, performing Pearson correlation between the graph nodes, and taking the correlation greater than 0.5 as the edge between the two graph nodes, if the correlation between the two nodes is less than 0.5, then there is no connected edge between the two graph nodes;

[0018] Setting STGCN-MLP spatio-temporal graph algorithm network: first setting STGCN, STGCN is composed of two spatio-temporal graph convolution blocks and an output fully connected layer, the whole process first passes through the time door convolution layer, then passes through the space graph convolution layer, and finally passes through the time door convolution layer again;

[0019] Based on the GAT-LSTM spatio-temporal graph algorithm network and the STGCN-MLP spatio-temporal graph algorithm network, a double spatio-temporal graph multi-objective algorithm model is constructed: according to the historical prediction results, the mape error of all town load rates Y of GAT-LSTM and the mape error of all town load rates Y of STGCN-MLP are calculated; according to the historical error, the weight of all town load rates Y of GAT-LSTM and the weight of all town load rates Y of STGCN-MLP are calculated, and the sum of the two weights of each group is 1; using the GAT-LSTM spatio-temporal graph algorithm network and the STGCN-MLP spatio-temporal graph algorithm network respectively to predict the results of the new date, according to the weight, respectively weighted, output a group of town load rate results and error, and finally calculate the overall error and balance degree.

[0020] The application proposes a method for predicting power grid load rate by using GAT-LSTM & STGCN-MLP double spatio-temporal graph model, which constructs a double spatio-temporal graph multi-objective algorithm model based on GAT-LSTM spatio-temporal graph algorithm network and STGCN-MLP spatio-temporal graph algorithm network, and can still maintain stable prediction under various complex and changeable holiday conditions, and enhances the stability of the model. BRIEF DESCRIPTION OF DRAWINGS

[0021] Figure 1 This is a flowchart illustrating the method for predicting power grid load rate using the GAT-LSTM & STGCN-MLP dual spatiotemporal graph model of the present invention.

[0022] Figure 2 This is a topology network structure diagram of GAT in an embodiment of the present invention. Detailed Implementation

[0023] To make the objectives, technical solutions, and advantages of the embodiments of the present invention clearer, the technical solutions of the embodiments of the present invention will be clearly and completely described below with reference to the accompanying drawings. Obviously, the described embodiments are only some embodiments of the present invention, not all embodiments. Based on the embodiments of the present invention, all other embodiments obtained by those skilled in the art without creative effort are within the scope of protection of the present invention.

[0024] Please see Figure 1 This invention provides a method for predicting power grid load factor using a dual spatiotemporal graph model of GAT-LSTM & STGCN-MLP. Figure 1 The left side shows the multi-objective prediction model GAT-LSTM, and the right side shows STGCN-MLP, which includes the following steps:

[0025] Step 1: Obtain load data and weather data for the target area (e.g., region S) from the data platform and meteorological website. The load data includes the date and the load rate of the 8 townships under region S, and the weather data includes the date and the temperature of region S. The period for the load data and weather data is from 20220101 to 20230731 for region S, with a time granularity of daily.

[0026] Step 2: Perform data cleaning on the load data and weather data obtained in Step 1, specifically including:

[0027] (1) Outlier deletion; Through previewing the data, it was found that the load rate of the data from 20220101 to 20220110 was abnormally high, about 10 times that of the later days, so it was deleted;

[0028] (2) Missing value handling: There are two types of missing values ​​in this embodiment of the invention. One is that the data is directly missing, and the other is that it is marked as " / N". First, all " / N" are replaced with missing values, and then the mean filling method is used to fill all the missing values.

[0029] (3) Numerical processing: The load rate calculation has been generated in the database, but it still needs to be converted from decimal form to percentage form.

[0030] (4) Data splicing: The load data and weather data are spliced ​​horizontally according to the date.

[0031] Step three, the data obtained in step two is subjected to feature engineering processing to obtain several dimensional data. The specific process is as follows:

[0032] Split the air temperature and transform it. The original air temperature is split into the highest temperature and the lowest temperature, and is transformed, and the formula is as follows:

[0033] Highest temperature 1 = highest temperature + 50

[0034] Lowest temperature 1 = lowest temperature + 50

[0035] 0-1 standardization of features, 0-1 standardization of the load rate of town 1-town 8, highest temperature 1, and lowest temperature 1, the field lowest temperature 1 does not exist negative number after the air temperature transformation in the last stage, which does not affect the result of standardization.

[0036] Step four, build GAT-LSTM & STGCN-MLP model, divided into three parts, ① set GAT-LSTM network, then use the several dimensional data obtained in step three for training and prediction, ② set STGCN-MLP network, then use the several dimensional data obtained in step three for training and prediction, ③ fuse the two; wherein ① and ② are performed simultaneously.

[0037] The specific process of step four is as follows:

[0038] (1) Set GAT-LSTM spatio-temporal graph algorithm network

[0039] The spatio-temporal graph network GAT-LSTM is composed of GAT and LSTM. First, the graph structure required by the GAT algorithm is constructed, including the nodes and edges of the graph. In the process of constructing the graph nodes, there are 8 towns in the embodiment of the application, each of which is set as a graph node, a total of 8 graph nodes, plus the highest temperature and the lowest temperature as virtual towns to become 2 nodes, a total of 10 nodes. Then, the 8 graph nodes are pairwise correlated, and the correlation greater than 0.5 is taken as the edge between the two graph nodes. If the correlation between two nodes is less than 0.5, there is no connecting edge between the two graph nodes. Since the 8 towns under S area are closely connected in economic activities, the daily load rate of each town is also closely related, and the Spearman correlation is greater than 0.5, that is, each graph node has 7 edges. In addition, each node has an attention score for measuring the degree of association between neighboring nodes. These attention scores are obtained through learning and can be dynamically adjusted according to the features between nodes. Figure 2 is the topological network structure of GAT, a total of 10 nodes, of which nodes 0-7 represent the 8 towns of S area, and nodes 8-9 represent the highest temperature and the lowest temperature.

[0040] Subsequently, the data is transmitted into a bidirectional long short-term memory network LSTM. The graph attention network is input into the long short-term memory neural network in a time sequence manner for learning time features, using ReLU as an activation function, batch_size is 30, 14 consecutive days are taken as a history period, and 8 fully connected layers are taken as synchronous output of 8 township daily load rates per day, and daily single-step cycle rolling prediction is performed.

[0041] (2) Set the STGCN-MLP spatiotemporal graph algorithm network

[0042] The spatiotemporal graph network STGCN-MLP is composed of STGCN and MLP. First, the STGCN is set, which is composed of two spatiotemporal graph convolution blocks (ST-ConvBlock) and an output fully connected layer (OutputLayer). The whole process first passes through a time gate convolution layer, then a spatial graph convolution layer, and finally a time gate convolution layer. The two TemporalGated-Conv use 1-D convolution, which is similar to the processing of a one-dimensional time sequence signal by CNN, that is, convolution in the time domain dimension; SpatialGraph-Conv performs convolution in the spatial domain, and the model is GCN; finally, the result is output through the MLP multilayer perceptron network in the fully connected layer.

[0043] The composition of the STGCN-MLP algorithm is mainly realized by the PyGTemporal library of pytorch. The core of the STGCN algorithm is constructed by calling 2 ST-ConvBlock, and then the MLP network is connected in the fully connected layer to form the STGCN-MLP algorithm. The input sequence input_size is set to 8, the sequence length seq_len is 14, the learning rate lr is 0.008, and the output sequence output_size is 1, for daily single-step cycle rolling prediction.

[0044] (3) Construct a double spatiotemporal graph multi-objective algorithm model based on the GAT-LSTM spatiotemporal graph algorithm network and the STGCN-MLP spatiotemporal graph algorithm network.

[0045] The double spatiotemporal graph multi-objective prediction algorithm model is the key point of the present application, and its core is to fuse the two single spatiotemporal graph multi-objective prediction algorithms. Unlike the fusion of two single-target time domain algorithms, this method is relatively mature; but for the fusion of two single spatiotemporal graph multi-objective algorithms, it is still blank. The present application embodiment first proposes to fuse the two single spatiotemporal graph multi-objective algorithms based on the multi-objective error reciprocal method (Multiobjective error reciprocal method), which can be divided into three steps:

[0046] 1. Calculate historical errors respectively. Calculate the mape errors of the 8 township load rates Y of GAT-LSTM and the mape errors of the 8 township load rates Y of STGCN-MLP according to the historical prediction results;

[0047] 2. Calculate weights. Calculate the weights of the 8 Y of GAT-LSTM and the weights of the 8 Y of STGCN-MLP according to the historical errors, and the sum of the two weights of each group is 1, for example: the weights of Y1 (township 1 load rate) of the two algorithms add up to 1, the weights of Y2 (township 2 load rate) of the two algorithms also add up to 1, and so on, and the weights of Y8 (township 8 load rate) of the two algorithms also add up to 1;

[0048] 3. Calculate the final error and balance degree of the fusion model. The results of the new date are predicted by the two algorithms respectively, and the results are output according to the weights, and the final error and balance degree are calculated.

[0049] The specific derivation formula is as follows:

[0050]

[0051] Where n represents which multi-objective prediction algorithm, here n = 1 represents GAT-LSTM, n = 2 represents STGCN-MLP, i represents the i-th prediction target, since the S area in the invention is subordinate to 8 townships, here i ranges from 1 to 8, t represents the true value, p represents the predicted value, e represents the error, and y represents the result, for example: e ni represents the error of the i-th prediction target of algorithm n, y nit represents the predicted value of the i-th y of algorithm n, y nip represents the true value of the i-th y of algorithm n.

[0052] When n = 1, the weights of each target of GAT-LSTM are:

[0053]

[0054] When n = 2, the weights of each target of STGCN-MLP are:

[0055]

[0056] Substitute (1) into equations (2) and (3) to obtain equations (4) and (5):

[0057]

[0058] W 1i represents the weight of GAT-LSTM at target i, for example, W 11is the weight of GAT-LSTM at Y1, W 2i represents the weight of STGCN-MLP at target i.

[0059] Combined with the true prediction results of each k-day, since the prediction is 7 days, here k ranges from [1, 7], and the final prediction value is calculated according to the dynamic weight value:

[0060]

[0061] where W 1,i is formula (4), W 1,i,k represents the weight of the prediction target i of GAT-LSTM at the kth day, and the target i ranges from [1, 8], for example, W 1，1，1 is the weight of the prediction target Y1 of GAT-LSTM at the 1st day; f1 i,k refers to the new prediction value of the prediction target i of GAT-LSTM algorithm at the kth day. The above process of STGCN-MLP algorithm is the same.

[0062] Establish the objective function, and F i Take the average to get the overall error F:

[0063]

[0064] Substitute formula (6) into formula (7) to get the final overall error F:

[0065]

[0066] Here the final overall error where Y true,i,k is the true value of target i on the kth day. represents the 7-day average error of the eight prediction targets of the dual spatiotemporal graph model.

[0067] In addition, it is not comprehensive to evaluate the entire model with only one or the same type of indicator. Here, the concept of balance degree is first proposed, and the reciprocal of the standard deviation is used as the calculation method, denoted by B:

[0068]

[0069] where N represents the number of samples, here there are 8 Ys, i.e. 8 mape error samples, so N is 8, x i is the mape error value of each sample, i takes 8, and u is the mean of the mape error values of these samples.

[0070] It should be noted that the overall error and error balance are independent between the two, and each evaluates the model from a different perspective. Overall error: used to evaluate the overall situation of the model prediction, the more stable the overall error each time (cannot be high and low), the better, and then the lower the better; balance: used to evaluate the good and bad situation of each target prediction error of the model, the higher the balance each time the better. The model is mainly used in actual production environment, so it has higher reliability and stability.

[0071] Error balance can be expressed by the derivation of some statistical indicators, such as the reciprocal of standard deviation, the reciprocal of variance, etc., and if the standard deviation or variance is directly used as the balance, the value is basically between the percentage and the thousandth, which is very unfavorable for comparison and observation. Which way to use depends on the actual distribution of the data, and in the data situation of the embodiment of the present application, the reciprocal of the standard deviation is used as the calculation method, which is more appropriate, but it does not mean that the error balance is always equal to the reciprocal of the standard deviation.

[0072] The smaller the change of overall error F each month, the smaller the error fluctuation of each month, and the more stable the model. Balance B is used to measure whether the error difference between each target Y is too large, and the smaller B is, the larger the error difference between each Y. Under the premise that each township (i.e. Y) is equally important, the error of these towns will be covered up by the average value, which is not conducive. As a more extreme example, the error of Y1 in January of a multi-target model is 5%, the error of Y2 is 5%, the error of Y1 in February is 1%, and the error of Y2 is 9%. If only the overall error F is considered, the model is very stable at 5% for two months, but in fact the error of the load rate of Y2 in February is covered up, and the purpose of prediction is to facilitate early warning, so as to make the Y2 area make wrong decisions in the future.

[0073] Step five, data execution, can be divided into four parts.

[0074] (1) Take July as an example and use GAT-LSTM for prediction, the data set is 20220101-20230630, take the first 6, the middle 2 as the training set, and the last 2 as the correction set. The next is the actual prediction: the last time segment (number 14) of the data set is predicted by rolling cycle, the newly generated row data (8 columns in total, representing Y1-Y8) is added to the data set, and the cycle is predicted 7 times. The final data is the original data set + 7-day new data, which is the prediction result of 20230701-20230707 (structure 7*8).

[0075] (2) Take July as an example and use STGCN-MLP for prediction, the data set and process are the same as above.

[0076] (3) Calculate the weight. Take the prediction results of the correction set stage to calculate the errors of the 8 Ys of GAT-LSTM and the errors of the 8 Ys of STGCN-MLP respectively, and calculate the weight (2 groups of 7*8), which is required for the fusion model.

[0077] (4) Prediction. With the prediction results of the two algorithms from 20230701 to 20230707 (2 groups of 7*8) and the weight (2 groups of 7*8), the prediction results (1 group of 7*8) are obtained according to the weight, and the final overall error and balance degree are calculated by comparing the true results.

[0078] Result analysis:

[0079] Take 20220101-20230531 as training test, and perform actual prediction on 20230601-20230607; algorithm 1 is GAT-LSTM, algorithm 2 is STGCN-MLP, township 1-8 is the error and average error of each target prediction value and actual value, and the balance degree is the inverse of the standard deviation of the error of township 1 to the error of township 8 (see formula 9 in step 6.5), and the fusion model GAT-LSTM&STGCN-MLP does not participate in this prediction, and the results are as shown in Table 1:

[0080] Table 1 Prediction results in June

[0081]

[0082] Take 20220101-20230630 as training test, and perform actual prediction on 20230701-20230707; algorithm 1 is GAT-LSTM, algorithm 2 is STGCN-MLP; the fusion model GAT-LSTM&STGCN-MLP generates weight according to the errors of the two algorithms in June in the above table (see formulas 1-5 in step 6.5), and the prediction results generated by the two algorithms from 0701 to 0707 are weighted (see formula 7 in step 6.5), to obtain the prediction error of the fusion model from 0701 to 0707 (see formula 8 in step 6.5), and the balance degree is calculated as above, and the results are as shown in Table 2:

[0083] Table 2 Prediction results in July

[0084]

[0085] Take 20220101-20230731 as training test, and perform actual prediction on 20230801-20230807, and the fusion model error and average error, balance degree calculation method is the same as above, and is not described again, and the results are as shown in Table 3:

[0086] Table 3 Prediction results in August

[0087]

[0088] From the above results, if single spatio-temporal graph algorithm is used for multi-target prediction, the results fluctuate greatly in different months. For example, the errors of GAT-LSTM in June, July and August are 8.8%, 12% and 13.7% respectively, and the errors of STGCN-MLP in June, July and August are 11.4%, 6.3% and 9.7% respectively. The average error fluctuation is huge, which is very unfavorable for use in production environment.

[0089] However, when using double spatio-temporal graph for multi-target prediction, the model results are obviously stable, and the errors of GAT-LSTM&GCN-MLP in July and August are 9% and 10.1% respectively. At the same time, combined with the balance degree, the balance degrees of the fusion model in July and August are 22.2 and 24.1 respectively, and the balance degree is improved. Therefore, compared with single spatio-temporal graph in different months, the double spatio-temporal graph model not only has more stable overall error in different months, but also has stable target error (there is no target Y1 error that is too small to lower the overall average error, and it can still remain stable in different months).

[0090] As can be seen from Table 2 and Table 3, the errors of the double spatio-temporal graph fusion model in July and August are 9% and 10.1% respectively, and the error is relatively stable, but the single spatio-temporal graph error fluctuation is very large, for example, the error of GAT-LSTM in July is 12%, and the error in August reaches 13.7%; the error of STGCN-MLP in July is 6.3%, and the error in August reaches 9.7%. Such a continuously stable multi-target fusion model has strong anti-interference ability, and can still maintain stable prediction (not only the overall target error is stable, but also each target error is stable) in response to various complex and variable situations such as holidays and emergencies, which is conducive to practical application in production environment.

[0091] The embodiment of the present application proposes the concept of error balance degree, which is used to measure whether the error difference between each Y of the multi-target prediction algorithm is too large (to avoid the "rich-poor gap" between the targets being too large); the embodiment of the present application proposes a double spatio-temporal graph multi-target prediction model GAT-LSTM&STGCN-MLP, which has strong anti-interference ability and can still maintain stable prediction in various complex and variable holiday situations, thereby enhancing the stability of the model.

[0092] The above is only a specific embodiment of the present application, but the protection scope of the present application is not limited thereto, any changes or replacements within the technical scope disclosed by the present application can be easily thought of by any person skilled in the art, which should be covered within the protection scope of the present application. Therefore, the protection scope of the present application should be subject to the protection scope of the claims.

Claims

1. A method for power grid load rate prediction using a GAT-LSTM&STGCN-MLP dual spatio-temporal graph model, characterized in that, Comprise the following steps: Step one, obtain target area load data and weather data from data platform and weather website; Step two, data cleaning is carried out on the load data and weather data of the target area obtained in step one, the load data includes date, load rate of the target area under township, and the weather data includes date, temperature of the target area; Step three, the data cleaned in step two is processed by feature engineering to obtain several dimension data; Step four, a double spatiotemporal graph multi-objective algorithm model is constructed based on GAT-LSTM spatiotemporal graph algorithm network and STGCN-MLP spatiotemporal graph algorithm network; Step five, input the several dimension data obtained in step three into the double spatiotemporal graph multi-objective algorithm model for prediction to obtain the load rate prediction result of the subordinate townships of the target area; Step four specifically comprises: Setting GAT-LSTM spatiotemporal graph algorithm network: constructing the graph structure required by GAT algorithm, including graph nodes and edges, setting each township of the target area as a graph node, plus the highest temperature 1 and the lowest temperature 1 as virtual townships to become two nodes, taking the correlation greater than 0.5 as the edge between two graph nodes, if the correlation between two nodes is less than 0.5, there is no connecting edge between the two graph nodes; Setting STGCN-MLP spatiotemporal graph algorithm network: first setting STGCN, STGCN consists of two spatiotemporal graph convolution blocks and an output fully connected layer, the whole process first passes through the time gate convolution layer, then passes through the space graph convolution layer, and finally passes through the time gate convolution layer again; Based on GAT-LSTM spatiotemporal graph algorithm network and STGCN-MLP spatiotemporal graph algorithm network, a double spatiotemporal graph multi-objective algorithm model is constructed: according to the historical prediction result, the mape error of all township load rate Y of GAT-LSTM and the mape error of all township load rate Y of STGCN-MLP are calculated; according to the historical error, the weight of all township load rate Y of GAT-LSTM and the weight of all township load rate Y of STGCN-MLP are calculated, the sum of the two weights of each group is 1; using GAT-LSTM spatiotemporal graph algorithm network and STGCN-MLP spatiotemporal graph algorithm network to predict the result of new date respectively, and according to the weight, a group of township load rate results and errors are output, and finally the overall error and balance degree are calculated.

2. The method for power grid load rate prediction using a GAT-LSTM&STGCN-MLP dual spatio-temporal graph model according to claim 1, wherein: Step three specifically comprises: The original temperature is split into highest temperature and lowest temperature, and transformed, the formula is as follows: Highest temperature 1= highest temperature + 50; Lowest temperature 1= lowest temperature + 50; The load rate of the subordinate townships of the target area, highest temperature 1 and lowest temperature 1 are standardized to 0-1.

Images