Wide-area multi-bus load forecasting method based on gated spatio-temporal graph neural network

Abstract

The application discloses a wide-area multi-bus load prediction method based on a gated space-time graph neural network, determines weather features strongly related to bus loads through a fast maximum information coefficient, determines time-space coupling correlations among the bus loads through the fast maximum information coefficient, and completes construction of a similarity weight space-time graph through the determined weather features, extracts and mines spatial features of each node of the similarity weight space-time graph in a graph convolution mode, inputs a time sequence formed by a result of a space convolution layer into a gated recurrent unit layer, and realizes time domain feature mining through the gated recurrent unit, so that the problems that the influence of unstructured time-space coupling correlations among the multi-bus loads in a wide-area space on a prediction result is not fully considered and it is difficult to uniformly model multi-bus load prediction are solved, global multi-node feature enhancement is realized, and load prediction precision is effectively improved.

Description

TECHNICAL FIELD

[0001] The application belongs to the technical field of wide-area multi-bus load time-space prediction, and particularly relates to a wide-area multi-bus load prediction method based on a gated time-space graph neural network. BACKGROUND

[0002] Modern power systems are developing in the direction of intelligence, flexibility and networking, showing a complex form of diversity, flexibility and correlation different from the past. Bus load prediction is of great significance to the safety of power grid dispatching and the accuracy of online analysis and decision-making. Compared with system load, bus load has a smaller base. In addition, it is affected by factors such as weather and user behavior in the power supply area, and has greater volatility, stronger randomness and less obvious change trend, making it difficult to accurately predict. At the same time, due to the differences in power consumption behavior of objects supplied by each bus, there are significant differences between bus loads, making it difficult to use a unified prediction model to predict multi-bus loads. Therefore, a unified model needs to be built to accurately predict short-term multi-bus time-space loads in a wide area.

[0003] Short-term load prediction methods are mainly divided into two categories: statistical methods and machine learning methods, such as LSTM, TCN, CNN, ARIMA, etc. Compared with statistical learning, load prediction methods based on machine learning can more effectively capture the complex nonlinear characteristics of load data. In the process of power system operation, if the users in different bus supply areas are of the same type, there may be similar power consumption patterns that are not limited by geographical distance, with potential time-space coupling correlation. Therefore, load prediction considering time-space coupling correlation is more accurate and feasible. In the load prediction based on machine learning, Zhang G et al. use CNN and Seq2Seq methods to analyze and mine multi-load sequences in Euclidean space, effectively improving the accuracy of short-term load prediction. Jiang L proposes a multi-information fusion load prediction method based on CNN-LSTM combination. This method uses LSTM and CNN to learn short-term and long-term power consumption behavior characteristics, effectively improving the accuracy of load prediction. Tae-Young Kim et al. propose an energy consumption prediction method considering time-space correlation information, which extracts spatial features and time domain features through CNN and LSMT, respectively, to improve prediction accuracy. The above researches are focused on households and comprehensive energy, but the methods are limited by the sequential input of data in Euclidean space. This leads to insufficient expression of the time-space correlation between different bus load data, and the prediction ability of the model needs to be improved. Analyzing and mining multi-bus load sequences from non-Euclidean domains not only does not limit the sequential input of data in Euclidean space, but also effectively takes into account the influence of the time-space correlation between the current bus load sequence and multiple specific bus loads on bus load prediction, further improving the prediction accuracy.

[0004] In order to accurately describe the complex time-space coupling correlation between wide-area multi-bus loads in the power system, it is necessary to model the non-grid topology constrained cross-regional time-space coupling of each bus at the time section. The spatial structure data and time series data are fully integrated. The complex coupling relationship between the non-geographical distance constraints and the bus loads that are independent of the grid topology constraints is mapped to the similarity weight time-space graph. Combined with the load characteristics and other characteristics between buses, a similarity weight time-space graph covering a wide spatial range and a long time span is constructed. Due to the different degrees of each node in the similarity weight time-space graph (i.e. the number of neighborhood nodes of each node is different), it is difficult to train the traditional method (such as CNN method) for Euclidean data.

[0005] Graph Convolutional Network (GCN) is based on the fixed point theory, inspired by convolutional neural networks and graph embedding, and provides a new technical route for the diversified expression and feature mining of power system data. It has been successfully applied in electricity price prediction, wind power prediction, photovoltaic output prediction, and household load prediction. Hang S et al. proposed a GCN-based day-ahead market marginal electricity price prediction method. This method fully considers the external factors that affect regional electricity prices under cross-regional power transmission conditions, improving the accuracy of node electricity price prediction. Khodayar et al. combined graph theory, graph convolutional neural networks, and rough sets to achieve accurate wind power prediction. J Simeunovi et al. modeled multi-site photovoltaic time series output data as graph structure data. A GCN-based multi-site PV prediction model is proposed to effectively improve the prediction accuracy of photovoltaic output. Lin W et al. proposed a GCN-based household load prediction method that effectively improves the accuracy of household load prediction, but does not consider the influence of weather, date, and other factors on load fluctuations. In addition, this method has not been applied in bus load prediction. The Spatial Convolution Layer (SCL) of GCN can directly process graph structure data composed of nodes and edges, and deeply mine the time-space coupling correlation contained in load data. Therefore, in bus load prediction, the number of outliers (evaluation indicators) that have a greater impact on scheduling can be effectively reduced. However, using SCL in GCN can only mine the spatial domain features between bus loads, and there is a problem of insufficient mining of time domain features in the similarity weight time-space graph. SUMMARY

[0006] The purpose of the present application is to provide a wide-area multi-bus load prediction method based on a gated time-space graph neural network, which solves the problem of insufficient consideration of the unstructured time-space coupling correlation between multi-bus loads in the wide-area space and the difficulty of unified prediction modeling of multi-bus loads in the prior art, and can realize global multi-node feature enhancement and effectively improve the load prediction accuracy.

[0007] The technical solution adopted in this invention is a wide-area multi-bus load forecasting method based on a gated spatiotemporal graph neural network, which is implemented according to the following steps:

[0008] Step 1: Treat the busbar of a certain area as a node, collect the meteorological characteristics and load characteristics of each node over a period of time, and filter the meteorological characteristics that have the greatest impact on the load characteristic changes based on the fast maximum information coefficient.

[0009] Step 2: Construct a similarity weighted spatiotemporal diagram based on the nodal meteorological characteristics and nodal load variables obtained after screening;

[0010] Step 3: Construct a spatiotemporal graph neural network, enhance the spatial convolutional layer in the spatiotemporal graph neural network, extract and mine the spatial features of each node of the similarity weight spatiotemporal graph through the spatial convolutional layer, and output the high-dimensional feature vector of each time step.

[0011] Step 4: Construct a GRUL with a gating mechanism, input the high-dimensional feature vectors at each time point into the GRUL with a gating mechanism, and output the load prediction results for the node for the next day.

[0012] The invention is further characterized by:

[0013] Step 1, based on the fast maximum information coefficient, involves selecting the meteorological features most influential on the load characteristic variables. The specific process is as follows: Nodal meteorological features include nodal temperature, rainfall, and air pressure. X is defined as nodal temperature, rainfall, and air pressure, and Y is the load characteristic variable. A two-dimensional space composed of X and Y is dynamically programmed and uniformly divided in the X and Y directions to form an x*y grid G. With a fixed number of grid divisions, different mutual information values ​​are obtained by changing the grid division positions. The maximum mutual information value is expressed as:

[0014]

[0015] In the formula: x and y represent the interval division in the direction of characteristic variables X and Y; D represents meteorological characteristic data, D|G represents the probability distribution of data D on G; I(D|G) represents the mutual information of D|G; max (·) represents the maximum value;

[0016] The maximum normalized mutual information of the meteorological feature data D is used to construct the feature matrix M(D). The elements of the feature matrix M(D) are defined as follows:

[0017]

[0018] In the formula, min{·,·} represents the minimum value, and the fast maximum information coefficient is defined as:

[0019]

[0020] Therefore, the correlation Rel(x) between the nodal meteorological characteristic variable X and the nodal load variable Y is:

[0021] Rel(x) = RapidMIC(x,y) (4)

[0022] The degree of correlation between variables is positively correlated with the value of Rel(x);

[0023] Select the meteorological characteristic variable that has the greatest correlation with the nodal load variable.

[0024] Step 2 is as follows:

[0025] Step 2.1: Map the correlation between the load variable of each node and the meteorological characteristics of the nodes obtained after screening to a node spatiotemporal graph with edge features as weights to obtain multiple node spatiotemporal graphs.

[0026] Step 2.2: Calculate the correlation degree of load features between nodes in the multi-node spatiotemporal graph, and map the correlation degree of load features between nodes to a similarity weight spatiotemporal graph with edge features as weights.

[0027] Step 3 involves enhancing the spatial convolutional layer in the spatiotemporal graph neural network. The specific process is as follows:

[0028] The specific formula for global node feature enhancement based on spatial convolutional layers is shown below:

[0029] H (l+1) =σ(AH (l) W (l) (7)

[0030] In the formula, σ is the activation function ReLU, A is the adjacency matrix, and H is the repetition function. (l) For the output of layer l, W (l) The weights of layer l;

[0031]

[0032] Formula (8) is used to add a self-looping module to the adjacency matrix A;

[0033]

[0034] Formula (9) is the result after adding the self-looping module. The degree matrix;

[0035]

[0036] Using formula (10) to perform symmetric normalization on the nodes of the current feature enhancement, we get:

[0037]

[0038] Step 3 involves extracting and mining the spatial features of each node in the similarity weight spatiotemporal graph using a spatial convolutional layer. The specific process is as follows:

[0039] Constructing two spatial convolutional layers to achieve wide-area multi-node coupled feature extraction, the formula is as follows:

[0040]

[0041] Where W0 is the weight from the input layer to the hidden layer, W1 is the weight from the hidden layer to the output layer, ReLU is the activation function, and X′ is the node feature input to the spatiotemporal graph neural network.

[0042] The beneficial effects of this invention are:

[0043] This invention is a wide-area multi-bus load forecasting method based on gated spatiotemporal graph neural network. It solves the problem that previous Euclidean data could not directly and accurately reflect the pairwise connection relationship of each bus in the non-Euclidean domain. It effectively takes into account the spatiotemporal coupling correlation of multi-bus loads hidden in the non-Euclidean spatial domain and reduces the generation of outliers in the evaluation indicators in load forecasting.

[0044] By extracting multi-source complex high-dimensional features such as the load of neighboring multiple nodes in the weighted spatiotemporal graph through graph convolutional neural networks, the full-domain multi-node feature enhancement is achieved, which effectively improves the accuracy of load prediction.

[0045] The features extracted from the spatial convolutional layers at different times are used to form a time series input into the gated recurrent unit layer to deeply explore the temporal features between the spatiotemporal graph and fully explore the temporal domain features between the spatiotemporal graph of similarity weight. Attached Figure Description

[0046] Figure 1 It is a wide-area multi-bus short-term load forecasting structure diagram invented from a gated spatiotemporal graph neural network;

[0047] Figure 2 This invention relates to the adjacency matrix graph in wide-area multi-bus short-term load forecasting using a gated spatiotemporal graph neural network;

[0048] Figure 3 This is a diagram of the spatial convolutional layer structure in the wide-area multi-bus short-term load prediction of the gated spatiotemporal graph neural network of the present invention.

[0049] Figure 4 This invention relates to the distribution of indicators for each bus in the test set of various methods for wide-area multi-bus short-term load forecasting using gated spatiotemporal graph neural networks.

[0050] Figure 5 This invention relates to the distribution of bus indexes corresponding to the maximum outliers in various methods of wide-area multi-bus short-term load forecasting using gated spatiotemporal graph neural networks.

[0051] Figure 6 This invention relates to the evaluation indices of SMAPE and MAE for each bus in wide-area multi-bus short-term load prediction using a gated spatiotemporal graph neural network.

[0052] Figure 7 This invention relates to the decrease (not the difference) of the SMAPE index percentage after conversion between the time-space graph neural network* and the time-space graph neural network method in the wide-area multi-bus short-term load forecasting of the gated time-space graph neural network. Detailed Implementation

[0053] The present invention will now be described in detail with reference to the accompanying drawings and specific embodiments.

[0054] This invention relates to a wide-area multi-bus load forecasting method based on a gated spatiotemporal graph neural network, such as... Figure 1 As shown, this includes node feature selection based on fast maximum information coefficient and construction of a similarity-weighted spatiotemporal graph with edge features as weights. Based on this, global node feature enhancement based on spatial convolutional layers is implemented. Then, temporal feature mining between spatiotemporal graphs is performed through gated recurrent unit layers. Finally, wide-area multi-node bus load spatiotemporal prediction is completed; specifically, it is implemented according to the following steps:

[0055] Step 1: Treat the busbar of a certain area as a node, collect the meteorological characteristics and load characteristics of each node over a period of time, and filter the meteorological characteristics that have the greatest impact on the load characteristic changes based on the fast maximum information coefficient.

[0056] The specific process for selecting the meteorological features with the greatest influence on load characteristic variables based on the fast maximum information coefficient is as follows: Nodal meteorological features include nodal temperature, rainfall, and air pressure. X is defined as nodal temperature, rainfall, and air pressure, and Y is the load characteristic variable. A two-dimensional space composed of X and Y is dynamically programmed and uniformly divided in the X and Y directions to form an x*y grid G. With a fixed number of grid divisions, different mutual information values ​​are obtained by changing the grid division positions. The maximum mutual information value is expressed as:

[0057]

[0058] In the formula: x and y represent the interval division in the direction of characteristic variables X and Y; D represents meteorological characteristic data, D|G represents the probability distribution of data D on G; I(D|G) represents the mutual information of D|G; max (·) represents the maximum value;

[0059] The maximum normalized mutual information of the meteorological feature data D is used to construct the feature matrix M(D). The elements of the feature matrix M(D) are defined as follows:

[0060]

[0061] In the formula, min{·,·} represents the minimum value, and the fast maximum information coefficient is defined as:

[0062]

[0063] Therefore, the correlation Rel(x) between the nodal meteorological characteristic variable X and the nodal load variable Y is:

[0064] Rel(x) = RapidMIC(x,y) (4)

[0065] The degree of correlation between variables is positively correlated with the value of Rel(x);

[0066] Select the meteorological characteristic variable that has the greatest correlation with the nodal load variable.

[0067] The fast maximum information coefficient was used to analyze the coupling relationship between complex meteorological characteristics such as temperature, rainfall, and air pressure and wide-area multi-bus loads, and to determine the characteristic variables strongly coupled with the power load. The fast maximum information coefficients of each meteorological variable and the load were calculated using formulas 1 to 4, and the results are shown in Table 1.

[0068] Table 1

[0069]

[0070] Step 2: Map the spatiotemporal coupling relationships between the bus loads existing in non-Euclidean space to a similarity-weighted spatiotemporal graph with edge features as weights. Each bus is abstracted as a node within the similarity-weighted spatiotemporal graph, and the meteorological and other features corresponding to each bus are the features between nodes within the graph. The correlation between bus loads is characterized using the edge features within the graph. Construct an adjacency matrix, as shown in formula (5). Because there are significant differences between bus loads, the connection relationships between nodes within the similarity-weighted spatiotemporal graph determined by the adjacency matrix are low-dimensional and sparse.

[0071]

[0072] In the formula, ε is the value corresponding to formula (4) between the load sequence of bus i and the load sequence of bus j. Since the degree of correlation between each bus varies, ε, which reflects the load between each bus, is used as the edge feature and as the similarity weight of the spatiotemporal graph edge.

[0073] Formula (5) represents the similarity weight time-space graph.

[0074] G t =(V t (6)

[0075] in Let V be the set of power bus loads in the power system at time t. t ∈RN×f . It consists of the meteorological feature set, historical load feature set, and date feature set of the Nth power bus at time t. E = {e1, e2, ..., e...} M} represents the similarity weight spatiotemporal graph data edge feature set, e M The edge characteristic is the ε value between bus loads.

[0076] The specific process is as follows:

[0077] Step 2.1: Map the correlation between the load variable of each node and the meteorological characteristics of the nodes obtained after screening to a node spatiotemporal graph with edge features as weights to obtain multiple node spatiotemporal graphs.

[0078] Step 2.2: Calculate the correlation degree of load features between nodes in the multi-node spatiotemporal graph, and map the correlation degree of load features between nodes to a similarity weight spatiotemporal graph with edge features as weights.

[0079] Step 3: Construct a spatiotemporal graph neural network, enhance the spatial convolutional layer in the spatiotemporal graph neural network, extract and mine the spatial features of each node of the similarity weight spatiotemporal graph through the spatial convolutional layer, and output the high-dimensional feature vector of each time step.

[0080] To address the challenge of handling non-Euclidean graph data using methods like CNNs, a spatial convolutional layer is constructed to extract high-dimensional features, such as multi-node loads within the K-order neighborhood of each node in the similarity-weighted spatiotemporal graph. The spatial convolutional layer utilizes the structural features of node-edge connections and the attribute features associated with the graph to extract implicit graph features. Under the constraints of the node connectivity relationships in the similarity-weighted spatiotemporal graph and the receptive field of the spatial convolutional layer, the complex spatiotemporal coupling features of each node are abstracted into feature transfer among nodes within the similarity-weighted spatiotemporal graph. This process exhibits similar computational relationships to power flow in power systems.

[0081] Spatial convolutional layers focus on the node features within a K-order neighborhood centered on a given node. Figure 3 The receptive field of a spatial convolutional layer is illustrated using node ① as an example. A single spatial convolutional layer can only capture multivariate, complex, and high-dimensional features such as the load of first-order neighboring nodes. To extract extensive node features from the similarity-weighted spatiotemporal graph, multiple spatial convolutional layers are stacked. Figure 3 As shown, taking node ① as an example, a single spatial convolutional layer can only extract features such as the differential loads of the neighboring nodes ②, ③, ④, and ⑤. After two spatial convolutional layers, node ① obtains multi-dimensional, complex, and high-dimensional features such as the differential loads of its neighboring nodes ②, ③, ④, ⑤, ⑥, and ⑦.

[0082] The specific process of enhancing the spatial convolutional layer in a spatiotemporal graph neural network is as follows:

[0083] The specific formula for global node feature enhancement based on spatial convolutional layers is shown below:

[0084] H (l+1) =σ(AH (l) W (l) (7)

[0085] In the formula, σ is the activation function ReLU, A is the adjacency matrix, and H is the repetition function. (l) For the output of layer l, W (l) The weights of layer l;

[0086]

[0087] Formula (8) is used to add a self-looping module to the adjacency matrix A;

[0088]

[0089] Formula (9) is the result after adding the self-looping module. The degree matrix;

[0090]

[0091] Using formula (10) to perform symmetric normalization on the nodes of the current feature enhancement, we get:

[0092]

[0093] The specific process of extracting and mining the spatial features of each node in the similarity weight spatiotemporal graph through spatial convolutional layers is as follows:

[0094] To prevent oversmoothing, two spatial convolutional layers are constructed to achieve wide-area multi-node coupled feature extraction, as shown in the formula:

[0095]

[0096] Where W0 is the weight from the input layer to the hidden layer, W1 is the weight from the hidden layer to the output layer, ReLU is the activation function, and X′ is the node feature input to the spatiotemporal graph neural network.

[0097] Step 4: Construct a GRUL with a gating mechanism, input the high-dimensional feature vectors at each time point into the GRUL with a gating mechanism, and output the load prediction results for the node for the next day.

[0098] To address the challenge of insufficient temporal feature mining between similarity weight time-space graphs by spatial convolutional layers, a GRUL with a gating mechanism is constructed. The high-dimensional feature vectors output by the spatial convolutional layers at each time step are used as the input of GRUL to achieve load prediction for wide-area multi-bus systems.

[0099] Compared to LSTM, GRU, as a variant of LSTM, solves the problems of complex structure and long training time of LSTM. GRUL consists of an update gate u t Reset door r t It consists of two multiplication gates. GRUL combines the features extracted by the spatial convolutional layer with the hidden layer output h at time t-1. t-1 As input, thus outputting h t The specific formula is as follows:

[0100] u t =σ(W u [gc(X′ t ,A),h t-1 ]+b u (13)

[0101] r t =σ(W r [gc(X′ t ,A),h t-1 ]+b r (14)

[0102] c t =tanh(W c [gc(X′ t ,A),(r t *h t-1 )]+b c (15)

[0103] h t =u t *h t-1 +(1-u t )*c t (16)

[0104] In the formula: σ represents the sigmoid activation function, tanh is the activation function, and W u Represents the update gate u t The weight matrix, gc(·) represents graph convolution operation, X′ t The node features represent the input of the spatial convolutional layer, A represents the adjacency matrix containing edge features, and h represents the node features. t-1 b represents the output at time t-1. u b r and b c W represents the bias vector. r Represents the reset gate r t The weight matrix, W c Represents memory c t The weight matrix, h t This represents the load sequence output at time t.

[0105] The model loss function is shown below:

[0106]

[0107] In the formula: Y t Represents the actual value of the power load. λ represents the predicted power load value, λ is the hyperparameter, and L2 represents regularization using weight decay.

[0108] Example

[0109] This study uses load data (15-minute timeframe) from 15 busbars in a region of Northeast China throughout 2018, along with local meteorological data. The load on each busbar in this region fluctuates dramatically throughout the year, exhibiting significant variability and strong randomness. The training and test sets are divided in a 9:1 ratio. Experimental parameters are as follows: 1000 training iterations, learning rate of 0.01, batch size of 128, and hidden units of 128.

[0110] To compare the prediction performance of the spatiotemporal graph neural network method with other methods, comparative experiments were conducted using CNN-LSTM, GRU, SVR, and the spatiotemporal graph neural network* (without meteorological or holiday features). The input features for the comparative experiments were consistent with the spatiotemporal graph node features of the spatiotemporal graph neural network.

[0111] Box plot ( Figure 4 This study characterizes the index distribution of prediction results from various methods from a holistic perspective of multiple busbars. Each box contains 15 busbars and 32 days of predicted indices (480 sets of data in total). Compared to CNN-LSTM, GRU, and SVR methods, the spatiotemporal graph neural network method shows significantly lower values ​​for upper outlier nodes, lower lower bounds, and a box structure defined by the mean line, upper quartile, and lower quartile in the SMAPE index compared to other methods. Furthermore, compared to spatiotemporal graph neural networks that do not incorporate meteorological and date features, the spatiotemporal graph neural network that considers multiple complex features can effectively improve load forecasting accuracy.

[0112] In actual power grid dispatching, the worst-case indicator (i.e., the largest outlier) of the predicted load is significantly affected. Because the spatiotemporal graph neural network method performs convolution in the non-Euclidean domain, it fully extracts and mines the features of each node within the spatiotemporal graph. Therefore, this method has fewer and lower-valued outliers. The buses corresponding to the largest outliers for the spatiotemporal graph neural network*, CNN-LSTM, GRU, and SVR are bus 13, bus 11, and bus 2, respectively. An analysis of the buses corresponding to the largest outliers for each method is as follows... Figure 5 As shown. (Through) Figure 5It can be seen that the maximum outlier values ​​for the Spatiotemporal Graph Neural Network (SBR) and SBR* in bus 13 are 8.78% and 14.72%, respectively; the maximum outlier values ​​for the SBR and CNN-LSTM in bus 11 are 8.70% and 10.20%, respectively; and the maximum outlier values ​​for the SBR, GRU, and SVR in bus 2 are 10.07%, 11.89%, and 12.08%, respectively. From the above analysis, it can be seen that the SMAPE indexes for the bus methods corresponding to the maximum outliers of the SBR*, CNN-LSTM, GRU, and SVR are all higher than those of the SBR method.

[0113] Table 2

[0114]

[0115] Table 2 shows the mean values ​​of the evaluation indicators for each method. Figure 5 The evaluation metrics for each method are presented. The benchmarks are CNN-LSTM, GRU, SVR, and the spatiotemporal graph neural network* (excluding meteorological and holiday features). The absolute values ​​of SMAPE and MAE errors of the spatiotemporal graph neural network method decreased by 0.25%–0.64% (corresponding to a percentage decrease of 5.36%–12.67%) and 0.5MW–1.05MW (corresponding to a percentage decrease of 6.48%–12.70%), respectively.

[0116] pass Figure 6 In the study, the spatiotemporal neural network and the SMAPE index show that the impact of date and meteorological factors on load forecasting varies among different busbars. Figure 7 It can be seen that each bus is affected by factors such as date and weather. Among them, bus 4 is least affected by date and weather, with a SMAPE difference of 0.04% between the spatiotemporal graph neural network* method and the spatiotemporal graph neural network method (equivalent to a 0.97% reduction in the SMAPE percentage of the spatiotemporal graph neural network method compared to the spatiotemporal graph neural network* method). Bus 15 is most affected, with a SMAPE difference of 0.85% between the spatiotemporal graph neural network* method and the spatiotemporal graph neural network method (equivalent to a 17.4% reduction in the SMAPE percentage of the spatiotemporal graph neural network method compared to the spatiotemporal graph neural network* method).

[0117] Methods such as GRU and SVR can only explain the temporal fluctuations of power load, failing to account for the impact of complex spatiotemporal coupling relationships between buses within the power grid on the load fluctuations of each node. While the CNN-LSMT method analyzes the spatiotemporal correlations of each bus, it does not adequately explore the paired connections between buses existing in non-Euclidean space. In contrast, the spatiotemporal graph neural network method effectively integrates spatial structure data and temporal data by constructing an unstructured similarity-weighted spatiotemporal graph, and utilizes spatial convolutional layers to enhance the features of nodes across the entire domain. Based on this, GRUL is used to mine temporal features between similarity-weighted spatiotemporal graphs, ultimately achieving wide-area multi-bus short-term load forecasting. Furthermore, the spatiotemporal graph neural network with negative bus features exhibits better prediction performance compared to a spatiotemporal graph neural network that only considers load features. The spatiotemporal graph neural network method effectively considers the multidimensional spatiotemporal correlations of multi-bus loads in a wide-area space, as well as the impact of meteorological and date features on load forecasting results. This method effectively improves load forecasting accuracy and overcomes the shortcomings of previous methods that only consider the time domain. It can more reasonably predict and interpret load fluctuations, and accurately reflect the load fluctuations of each bus.

[0118] Power load data originating from SCADA systems is susceptible to channel errors, measurement errors, and sudden accidents, leading to abnormal load data. In practical engineering, it is often difficult to achieve refined identification and repair of abnormal loads. Furthermore, the results of abnormal load identification and repair vary across different scenarios, all of which affect the accuracy of short-term load forecasting. Therefore, it is necessary to evaluate the robustness of new methods in scenarios containing abnormal load data.

[0119] The experiment simulated real-world scenarios with abnormal load data by constructing zero-point, continuous constant-value load, and abnormal step load simulations, without performing any abnormal load identification or processing. The experiment was conducted in six groups, with the proportion of abnormal load data in the test set increasing by 5% sequentially to 30%. The proportion of each type of abnormal load data was randomized.

[0120] Table 3

[0121]

[0122] Table 3 shows the SMAPE metrics of 15 bus lines for each method under different abnormal load percentage scenarios. Compared with the metrics of CNN-LSTM, GRU, and SVR, the absolute value of the SMAPE error of the spatiotemporal graph neural network method decreases by 0.1%–0.75% (i.e., a corresponding percentage decrease in error of 1.92%–14.07%). This demonstrates that the spatiotemporal graph neural network method has strong robustness compared to other methods under different abnormal load percentage scenarios. By constructing spatial convolutional layers to extract high-dimensional features such as the load of neighboring multiple nodes in the spatiotemporal graph, full-domain multi-node feature enhancement is achieved. Therefore, the spatiotemporal graph neural network method effectively addresses the challenge of reduced prediction accuracy caused by abnormal load data.

Claims

1. A wide-area multi-bus load forecasting method based on a gated space-time graph neural network, characterized in that, The following steps are implemented in detail: Step 1: Take the busbar of a certain area as a node, collect the meteorological characteristics of each node and the load characteristics of each node within a period of time, and select the meteorological characteristics that have the greatest impact on the load characteristic change amount based on the fast maximum information coefficient; Step 2: Construct a similar weight space-time graph based on the selected node meteorological characteristics and node load variables; Step 3: Construct a space-time graph neural network, enhance the spatial convolution layer in the space-time graph neural network, extract and mine the spatial characteristics of each node in the similar weight space-time graph through the spatial convolution layer, and output the high-dimensional feature vector at each time; Step 4: Construct a GRUL with a gating mechanism, input the high-dimensional feature vector at each time into the GRUL with a gating mechanism, and output the load prediction result of the node for the next day; The specific process for selecting the meteorological features with the greatest influence on load characteristic variables based on the fast maximum information coefficient in step 1 is as follows: Nodal meteorological features include nodal temperature, rainfall, and air pressure, defined as follows: X =Nodal temperature, rainfall, air pressure Y As load characteristic variables, for those derived from X and Y The two-dimensional space formed X and Y The directions are formed by dynamic programming and uniform partitioning respectively. grid G With a fixed number of grid divisions, different mutual information values ​​are obtained by changing the grid division positions. The maximum mutual information value is represented as: (1) wherein: x and y denotes the characteristic variable X and Y interval division of the direction; D denotes the weather characteristic data, denotes the data D distribution probability on G ; denotes the mutual information; max (·) denotes the maximum value; The weather feature data D A maximum norm mutual information of the weather feature data is constructed as a feature matrix M(D), and an element in the feature matrix M(D) is defined as: (2) wherein The fast maximum information coefficient is defined as: (3) Correlation between the weather characteristic variable of the sink node and the load variable of the sink node X Correlation between the weather characteristic variable of the sink node and the load variable of the sink node Y Correlation between the weather characteristic variable of the sink node and the load variable of the sink node Correlation between (4) where the degree of correlation between the variables and values is positively correlated; Select the meteorological characteristic variable that has the greatest correlation with the node load variable.

2. The wide-area multi-bus load forecasting method based on the gated spatio-temporal graph neural network according to claim 1, characterized in that, The specific process of step 2 is as follows: Step 2.1: Map the correlation between the load variable of each node and the selected node meteorological characteristics to the node space-time graph with edge characteristics as the weight, and obtain multiple node space-time graphs; Step 2.2: Calculate the correlation degree of the load characteristics between nodes in the multiple node space-time graphs, and map the correlation degree of the load characteristics between nodes to the similar weight space-time graph with edge characteristics as the weight.

3. The method of claim 1, wherein the method further comprises: The specific process of enhancing the spatial convolution layer in the space-time graph neural network in step 3 is as follows: The global node feature enhancement based on the spatial convolution layer is specifically shown in the following formula: (7) wherein, is the activation function ReLU, A is the adjacency matrix, is the output of the l-th layer, is l the weight of the l-th layer; (8) Taking formula (8) to the adjacency matrix Increase the self-loop module; (9) Equation (9) is the degree matrix after adding the self-loop module of the degree matrix; (10) Symmetric normalization is performed on the current feature-enhanced node using formula (10) to obtain: (11)。 4. The wide-area multi-bus load forecasting method based on the gated spatio-temporal graph neural network according to claim 3, characterized in that, The specific process of extracting and mining the spatial characteristics of each node in the similar weight space-time graph through the spatial convolution layer in step 3 is as follows: Two layers of spatial convolution layers are constructed to realize wide-area multi-node coupling feature extraction, and the formula is as follows: (12) wherein, is a weight from the input layer to the hidden layer, is a weight from the hidden layer to the output layer, is an activation function, is a node feature input to the spatio-temporal graph neural network.

Images