Power grid mountain fire prediction method based on causal driving and space-time diagram convolutional network

Abstract

The invention provides a power grid mountain fire prediction method based on causal driving and a space-time diagram convolutional network, and belongs to the technical field of mountain fire prediction. A dynamic feature encoder and a static feature encoder are designed to extract high-dimensional spatial-temporal features of meteorological time sequence information, geographic space environment and power transmission line distribution multi-source heterogeneous data, a causal discovery algorithm is adopted to construct a dynamically evolved causal graph topology, a real causal driven relationship between variables is identified, a causal intensity matrix is decoupled into positive and negative adjacent matrixes, and the dynamic evolved multi-source heterogeneous data is obtained. And designing a causal constrained graph convolution module to aggregate and propagate high-order information, and finally realizing accurate prediction of the power grid forest fire. According to the method, the PCMCI causal discovery algorithm is introduced, so that the real causal driven relationship among multivariate time sequence factors is effectively identified; a causal GCN layer in the CSTGCN model extracts spatial dependence features by using an adjacent matrix constrained by a causal structure, the CSTGCN model embeds causal structure information into a spatio-temporal feature learning framework, and higher prediction precision and generalization performance are achieved.

Description

Technical Field

[0001] This invention provides a power grid wildfire prediction method based on causal driving and spatiotemporal graph convolutional networks, belonging to the field of wildfire prediction technology. Background Technology

[0002] Current research on wildfire prediction by scholars both domestically and internationally mainly focuses on statistical methods and artificial intelligence methods. However, statistical methods lack generalization ability, and artificial intelligence methods have low interpretability and reliability. There is an urgent need to construct a power grid wildfire prediction method that can reflect the high-order interaction process between factors in order to improve prediction accuracy and interpretability.

[0003] One existing technology, which uses statistical methods to construct explicit mathematical models by capturing patterns in historical events, for predicting wildfires in power grids, takes the following form:

[0004] First, statistical methods such as the Spatial Generalized Additive Model (SGAM) and Poisson Regression (PR) were selected. Then, data on meteorological, vegetation, and anthropogenic factors influencing wildfires in the study area were collected and processed. Next, an explicit mathematical model was constructed to fit the quantitative relationship between wildfire probability and multiple influencing factors. The contribution and direction of each factor were analyzed through model parameter estimation and significance testing. Finally, the dataset of wildfire influencing factors in the study area was input into the statistical model to obtain the predicted wildfire rates for the following day in the region.

[0005] This technique often assumes that there is a linear functional relationship between variables and their corresponding coefficients. However, the occurrence of wildfires is affected by the complex interaction of various factors such as meteorological conditions, vegetation type, terrain features and human activities. Models that rely on linear assumptions cannot fully capture and simulate these nonlinear changes, have insufficient generalization ability, and their prediction accuracy needs to be further improved.

[0006] In addition, a second existing technology involves artificial intelligence methods that can automatically mine the nonlinear mapping relationship between wildfires and multiple influencing factors through training and optimization with large amounts of data. The specific form of using artificial intelligence methods for power grid wildfire prediction is as follows:

[0007] First, based on historical wildfire samples and their corresponding multivariate influencing factor data, training and testing datasets were constructed, and features such as meteorology, vegetation, and human activities were standardized and extracted. Then, power grid wildfire prediction models were established using networks such as convolutional neural networks (CNN), convolutional long short-term memory (ConvLSTM), and multi-layer perceptrons (MLP). Next, multi-source datasets were input into the model, and the parameters were optimized using the backpropagation algorithm, iteratively training to minimize prediction errors. Finally, the trained model was applied to new input data to predict future power grid wildfires in the target area.

[0008] The underlying mechanisms of this technology are difficult to reveal, and it often focuses on "post-hoc analysis." For example, it uses methods such as Shapley additive explanations (SHAP), average impurity decrease (AVID), and ermutation importance (PI) to measure the marginal contribution of each influencing factor to the prediction results of the network model. However, it lacks "pre-hoc explanations" of the internal relationships and mechanisms of action of the predictive factors, making it difficult to explain how they interact to cause wildfires, thus limiting the interpretability and reliability of the prediction results. Summary of the Invention

[0009] To address the shortcomings of the first technique, this invention constructs dynamic and static feature encoders to extract corresponding high-dimensional feature representations based on the differences in temporal distribution and spatial structure of multi-source heterogeneous data. It extracts high-dimensional spatiotemporal features of multi-source heterogeneous data such as meteorological time series information, geospatial environment and power transmission line distribution, and constructs a GCN model to aggregate and update the features of adjacent power grid wildfire impact factors through spatial convolutional layers, thereby realizing iterative update of node features and information transmission. Finally, it integrates the spatial correlation and temporal features of multi-source data to predict the risk of power grid wildfires the next day.

[0010] To address the shortcomings of technique two, this invention employs the PCMCI causal discovery algorithm to mine potential causal relationships between multidimensional dynamic variables, constructs a dynamic evolutionary causal graph structure divided by time period and region, reveals the causal chains of different variables, and uses the causal graph as the topology of a spatiotemporal graph convolutional network to guide the graph convolutional layer to achieve information propagation and feature aggregation under causal constraints. This reduces spurious correlation interference, more accurately characterizes the dynamic coupling effect between influencing factors, and improves the accuracy and interpretability of power grid wildfire prediction.

[0011] The specific technical solution of this invention:

[0012] A method for predicting power grid wildfires based on causal driving and spatiotemporal graph convolutional networks is proposed. Dynamic and static feature encoders are designed to extract high-dimensional spatiotemporal features from multi-source heterogeneous data, including meteorological time-series information, geospatial environment, and transmission line distribution. A causal discovery algorithm is used to construct a dynamically evolving causal graph topology to identify the true causal driving relationships between variables. The causal intensity matrix is ​​decoupled into positive and negative adjacency matrices. A causal constraint graph convolutional module is designed to aggregate and propagate high-order information, ultimately achieving accurate prediction of power grid wildfires.

[0013] Specifically, the following steps are included:

[0014] (1) Problem definition

[0015] Accurate prediction of the study area The probability that each grid pixel k will be affected by a wildfire in the future. [0,1].

[0016] The study area is spatially divided into I sub-regions R. i (i=1,2,…,I), each sub-region has a spatial size of H×W and a total of K pixels; it is divided into J consecutive time periods T. j (j=1,2,…,J), with a time step of S. In each pair of spatiotemporal units (R) i ,T j Within this range, the basic coupling patterns among various wildfire influencing factors are relatively fixed. Based on this division, a multidimensional vector is constructed:

[0017]

[0018] Where: x i,j,k,m For spacetime unit (R) i ,T j The value of the m-th factor at pixel k within a given area; M is the number of factors. The problem of wildfire prediction in power grids can be abstracted as a nonlinear mapping process from a multivariate feature vector to the pixel-level probability of wildfire occurrence:

[0019]

[0020] In the formula: y i,j =[y i,j,1 ,y i,j,2 ,…,y i,j,K ] represents a spacetime unit (R) i ,T j The probability of a wildfire occurring at each pixel within the range; f(·) is the mapping function; and These represent the spatial dependency structure and temporal dynamic characteristics of the variables within the unit, respectively.

[0021] (2) Dataset construction

[0022] A dataset of power grid wildfire impact factors was constructed, encompassing 14 categories of variables across seven aspects: historical fire locations, meteorological characteristics, vegetation conditions, drought conditions, human activities, topographic structure, and power transmission lines.

[0023] Based on temporal resolution, influencing factors are divided into daily-scale dynamic variables and annual-scale static variables. Using continuous T-day dynamic data as input enhances the ability to characterize the temporal dependencies and interaction mechanisms among variables; static data from the h×w grid region near the center point is extracted to capture spatial correlations.

[0024] In the data preprocessing stage, missing value imputation, normalization, and outlier correction are performed on dynamic variables to ensure the continuity and consistency of the time series; spatial pruning, resampling, and standardization are performed on static variables.

[0025] Subsequently, each variable is uniformly encoded and aligned according to its dimension, and saved as a tensor for model use. The preprocessed dynamic and static tensor shapes input to the network are (B, T, C) respectively. d (B,C) s (,h,w). Where B is the batch size; C d and C s These represent the number of dynamic and static variables, respectively.

[0026] (3) Constructing a causal graph generation framework

[0027] A power grid wildfire prediction method based on GSTGCN is adopted. The PCMCI causal discovery algorithm is combined with GCN. The real causal driving relationship is quantitatively identified from multi-source spatiotemporal data, and a causal topology graph is constructed to replace the fixed adjacency matrix based on spatial distance in the standard GCN. Based on the division of spatiotemporal units in the problem definition in (1), for each spatiotemporal unit (R i ,T j By extending the time range forward to L, multi-source data of all pixels within the spatial range H×W and time step L are extracted to construct a spatiotemporal dataset for causal discovery. .

[0028] For each factor m, the spatial average of all pixels is taken over time steps to obtain a multidimensional time series reflecting the overall evolutionary characteristics of the region. Using this sequence as input to the PCMCI algorithm, the algorithm quantitatively identifies the true causal driving relationship between the factors corresponding to the spatiotemporal unit, and outputs a causal strength matrix and a causal delay matrix representing the strength of the causal relationship between the factors.

[0029] Then, using time step S as a sliding window, the causal relationship of the next spatiotemporal unit is calculated step by step along the time dimension, realizing the dynamic update of the causal graph and providing interpretable causal input for subsequent power grid wildfire prediction models.

[0030] (4) Construct a power grid wildfire prediction model

[0031] After obtaining a dynamically updatable causal topology graph, a GSTGCN power grid wildfire prediction model is designed, consisting of three modules: a dynamic feature encoder, a static feature encoder, and a causal GCN layer. First, a dynamic feature encoder is designed to extract high-level semantic features of time-series dynamic variables. One-dimensional grouped convolutions are used to encode the dynamic variables, treating the data channels of each node as independent convolutional groups to ensure the independence of feature extraction. After the convolution operation, the features are mapped to a high-dimensional latent space, forming a node state sequence. Where F represents the high-level semantic feature dimension of the node at each time step.

[0032] To quantitatively characterize the lagged causal effects between nodes, based on the causal graph generation framework (3), time-delay information is extracted from the node state sequence through the causal time-delay matrix, and weighted aggregation is performed in combination with the causal strength matrix to obtain the node representation that integrates causal dependencies. The operation is shown in Equation (3):

[0033]

[0034] In the formula: Z a (t) represents the representation of node a after fusing its causal neighborhood historical information at time step t; S ab The causal influence strength of node b on node a is given by the causal strength matrix; This represents the corresponding causal lag order.

[0035] In the causal GCN layer, the causal strength matrix is ​​used as the adjacency matrix to perform graph convolution operation on the node representation, so that the local causal information of the node propagates along the causal topology, thereby capturing the high-order interaction relationship between the nodes.

[0036] Decompose the causality strength matrix into a positive matrix A. + and negative matrix A - As an adjacency matrix, it is used for weighted propagation in graph convolution. Positive edges enhance the facilitation effect between nodes, while negative edges independently model the suppression effect, avoiding the loss of symbolic information in nonlinear activation functions such as ReLU. The operation is shown in equation (4):

[0037]

[0038] In the formula: f ReLU(·) represents the ReLU activation function; B(·) represents the batch normalization layer; , These are the positive and negative normalized causal adjacency matrices, respectively; f Drop (·) represents a random deactivation layer; , The weight matrices are for positive and negative subgraphs; , This is a bias term.

[0039] The static feature encoder is designed to progressively increase the number of channels and reduce the spatial size using multiple convolutional blocks, extracting spatial representations from local to global perspectives. Batch normalization and ReLU activation functions are added after each convolutional layer to stabilize the training process and enhance nonlinear expressive power. After multiple convolutional processing, static variables are mapped to a high-dimensional latent space with the same dimension as dynamic variables.

[0040] Finally, the dynamic and static features are concatenated to obtain a fused feature vector. After processing by a fully connected layer, the output result is controlled to the [0,1] interval by the Sigmoid activation function.

[0041] The technical effects of the present invention are as follows:

[0042] (1) The introduction of the PCMCI causal discovery algorithm can effectively identify the real causal driving relationship between multiple time series factors, enabling the model to focus on the key path during information dissemination, reduce pseudo-correlation interference, and thus more accurately characterize the dynamic coupling effect between influencing factors.

[0043] (2) The causal GCN layer in the CSTGCN model uses the adjacency matrix constrained by causal structure to extract spatial dependency features, enabling nodes to adaptively aggregate neighborhood information with causal relationships during feature propagation and update, which is more interpretable than traditional deep learning models.

[0044] (3) The CSTGCN model embeds causal structure information into a spatiotemporal feature learning framework, which has higher prediction accuracy and generalization performance. It can effectively identify high-risk areas and provide early warnings, providing reliable support for power grid wildfire prevention and control, operation risk assessment and emergency dispatch decision-making. Attached Figure Description

[0045] Figure 1 This is the causal graph generation framework of the present invention;

[0046] Figure 2 This is the GSTGCN network structure of the present invention;

[0047] Figure 3 This is a causal structure diagram of the present invention;

[0048] Figure 4This is a power grid wildfire risk distribution map for the present invention. Detailed Implementation

[0049] The specific technical solutions of the present invention will be described with reference to the embodiments.

[0050] A power grid wildfire prediction method based on causal driving and spatiotemporal graph convolutional networks includes the following steps:

[0051] (1) Problem definition

[0052] Wildfire prediction in power grids is essentially a sophisticated spatiotemporal binary classification problem. Its core objective is to fuse multi-source data to accurately predict wildfires in the study area. The probability that each grid pixel k will be affected by a wildfire in the future. [0,1]. Due to the complex nonlinear coupling relationships between meteorological, environmental, and other factors influencing wildfires, and the variations in this coupling across different spatiotemporal scales, the local features of a single pixel are insufficient to reveal the complex mechanisms of wildfire occurrence. Directly constructing a unified prediction model for all pixels globally makes accurate prediction difficult. Therefore, it is necessary to introduce spatial and temporal partitioning before prediction to systematically characterize the interactions between factors.

[0053] The study area is spatially divided into I sub-regions R. i (i=1,2,…,I), each sub-region has a spatial size of H×W and a total of K pixels; it is divided into J consecutive time periods T. j (j=1,2,…,J), with a time step of S. In each pair of spatiotemporal units (R) i ,T j Within this range, the basic coupling patterns among various wildfire influencing factors are relatively fixed. Based on this division, a multidimensional vector is constructed:

[0054]

[0055] Where: x i,j,k,m For spacetime unit (R) i ,T j The value of the m-th factor at pixel k within a given area; M is the number of factors. The problem of wildfire prediction in power grids can be abstracted as a nonlinear mapping process from a multivariate feature vector to the pixel-level probability of wildfire occurrence:

[0056]

[0057] In the formula: y i,j =[y i,j,1 ,y i,j,2 ,…,y i,j,K ] represents a spacetime unit (R)i ,T j The probability of a wildfire occurring at each pixel within the range; f(·) is the mapping function; and These represent the spatial dependency structure and temporal dynamic characteristics of the variables within the unit, respectively. Therefore, how to construct the mapping function f(·) for joint learning... Understanding the spatial distribution characteristics and temporal evolution patterns of these factors to reflect the higher-order interaction processes between them is key to accurately predicting wildfires in power grids.

[0058] (2) Dataset construction

[0059] To achieve effective learning of the mapping function f(·), a high-quality dataset of power grid wildfire impact factors needs to be constructed. This invention constructs a dataset of power grid wildfire impact factors covering 14 categories of variables across seven aspects: historical fire points, meteorological characteristics, vegetation conditions, drought conditions, human activities, terrain structure, and transmission lines. Based on temporal resolution, the impact factors are divided into daily-scale dynamic variables and annual-scale static variables. T consecutive days of dynamic data are used as input to enhance the representation of temporal dependencies and interaction mechanisms between variables; static data from the h×w grid region near the center point is extracted to capture spatial correlations. In the data preprocessing stage, missing value imputation, normalization, and outlier correction are performed on the dynamic variables to ensure the continuity and consistency of the time series; spatial pruning, resampling, and standardization are performed on the static variables. Subsequently, each variable is uniformly encoded and aligned according to its dimension and saved as a tensor for model use. After preprocessing, the shapes of the dynamic and static tensors input to the network are (B, T, C). d (B,C) s (,h,w). Where B is the batch size; C d and C s These represent the number of dynamic and static variables, respectively. The attributes of the input dataset are shown in Table 1.

[0060] Table 1 Input Dataset Attributes

[0061]

[0062] Among them, the normalized difference vegetation index (NDVI) reflects the vegetation growth status and cover density. The higher the NDVI value, the higher the leaf cover, which provides sufficient combustible material for wildfires. The standardized precipitation evapotranspiration index (SPEI) is a comprehensive indicator of dryness and wetness. It shows a low value under continuous drought conditions, reflecting reduced air humidity and soil moisture deficit, which increases the risk of wildfires.

[0063] (3) Constructing a causal graph generation framework

[0064] This invention proposes a power grid wildfire prediction method based on GSTGCN. It combines the PCMCI causal discovery algorithm with GCN. Since the node information propagation of standard GCN is based on a fixed graph topology, the equal-weight aggregation mechanism of aggregating the features of all adjacent nodes with equal weights through a standardized adjacency matrix is ​​insufficient to characterize the dynamic non-uniform coupling relationship between wildfire influencing factors. Therefore, this invention introduces the PCMCI causal discovery algorithm to quantitatively identify the true causal driving relationship from multi-source spatiotemporal data and construct a causal topology graph to replace the fixed adjacency matrix based on spatial distance in standard GCN. Based on the division of spatiotemporal units in the problem definition in (1), for each spatiotemporal unit (R... i ,T j Since the time step is relatively short, it is difficult to accurately estimate causal relationships based solely on data within a single cell. Therefore, the time range is extended forward to L, and multi-source data from all pixels within the spatial range H×W and time step L are extracted to construct a spatiotemporal dataset for causal discovery. To reduce computational complexity and capture the overall dynamic characteristics of the sub-region, for each factor m, spatial averaging is performed over all pixels at each time step to obtain a multidimensional time series reflecting the overall evolutionary characteristics of the region. The sequence is used as input to the PCMCI algorithm to quantitatively identify the true causal driving relationships between the factors corresponding to the spatiotemporal unit, outputting a causal strength matrix characterizing the strength of the causal relationship between the factors and a causal delay matrix representing the lag order. Then, using time step S as a sliding window, the causal relationship of the next spatiotemporal unit is calculated step by step along the time dimension, realizing the dynamic updating of the causal graph and providing interpretable causal input for subsequent power grid wildfire prediction models. The causal graph generation framework is as follows: Figure 1 As shown.

[0065] (4) Construct a power grid wildfire prediction model

[0066] After obtaining a dynamically updatable causal topology graph, this invention designs a GSTGCN power grid wildfire prediction model consisting of three modules: a dynamic feature encoder, a static feature encoder, and a causal GCN layer. First, a dynamic feature encoder is designed to extract high-level semantic features of time-series dynamic variables. One-dimensional grouped convolutions are used to encode the dynamic variables, treating the data channels of each node as independent convolutional groups to ensure the independence of feature extraction. After the convolution operation, the features are mapped to a high-dimensional latent space, forming a node state sequence. Where F represents the high-level semantic feature dimension of the node at each time step.

[0067] To quantitatively characterize the lagged causal effects between nodes, based on the causal graph generation framework (3), time-delay information is extracted from the node state sequence through the causal time-delay matrix, and weighted aggregation is performed in combination with the causal strength matrix to obtain the node representation that integrates causal dependencies. The operation is shown in Equation (3):

[0068]

[0069] In the formula: Z a (t) represents the representation of node a after fusing its causal neighborhood historical information at time step t; S ab The causal influence strength of node b on node a is given by the causal strength matrix; The corresponding causal lag order is represented by the causal intensity matrix. In the causal GCN layer, the causal intensity matrix is ​​used as the adjacency matrix to perform graph convolution on the node representation, allowing the local causal information of the nodes to propagate along the causal topology, thereby capturing higher-order interactions between nodes. Considering the directionality of causal effects, the positive and negative values ​​in the causal intensity matrix represent promoting and inhibiting effects, respectively. To fully express the structural characteristics and propagation patterns of causal relationships in different directions, this invention decomposes the causal intensity matrix into a positive matrix A. + and negative matrix A - As an adjacency matrix, it is used for weighted propagation in graph convolution. Positive edges enhance the facilitation effect between nodes, while negative edges independently model the suppression effect, avoiding the loss of symbolic information in nonlinear activation functions such as ReLU. The operation is shown in equation (4):

[0070]

[0071] In the formula: f ReLU (·) represents the ReLU activation function; B(·) represents the batch normalization layer; , These are the positive and negative normalized causal adjacency matrices, respectively; f Drop (·) represents a random deactivation layer; , The weight matrices are for positive and negative subgraphs; , This is a bias term.

[0072] Static variables mainly include spatial attributes that do not change daily, such as topography, land type, and the spatial distribution of power transmission lines. They exhibit significant local correlation and regional differences in spatial dimension. Therefore, a static feature encoder is designed to progressively increase the number of channels and reduce the spatial size using multi-layer convolutional blocks to extract spatial representations from local to global perspectives. Batch normalization and ReLU activation functions are added after each convolutional layer to stabilize the training process and enhance nonlinear expressive power. After multi-layer convolution processing, the static variables are mapped to a high-dimensional latent space consistent with the dimension of the dynamic variables.

[0073] Finally, the dynamic and static features are concatenated to obtain a fused feature vector. After processing by a fully connected layer, the output is controlled to the [0,1] interval by a Sigmoid activation function. The CSTGCN network structure diagram is shown below. Figure 2 As shown.

[0074] The following example will verify this:

[0075] 1. Training Process

[0076] California, USA, was selected as the study area. Wildfire samples from 2012-2016 were used as the training set, and wildfire samples from 2017 were used as the test set. The above input dataset was combined with the GSTGCN network to achieve power grid wildfire prediction. The overall training steps of this invention are as follows:

[0077] 1) Collected 14 types of power grid wildfire impact factor datasets in California from 2012 to 2017, and spatially mapped them to a 500×500 grid, with each grid pixel being approximately 2.53km×2.53km.

[0078] 2) Based on the causal graph generation framework, the study area is divided into several spatiotemporal units with a spatial range of 50×50 grids and a time step of 1 day. For each spatiotemporal unit, the time range is extended forward to the past 90 days, and the spatial average value of each factor within the unit is extracted as the input of the PCMCI algorithm to calculate the corresponding causal time delay matrix and causal intensity matrix.

[0079] 3) Centering each fire point on a pixel, extract the dynamic data from the previous 30 days for that point and the static data from a 9×9 grid surrounding the center point. This yields a dynamic tensor of 4096×30×8 and a static tensor of 4096×7×9×9, which are used as positive samples. Considering the class imbalance problem in actual power grid wildfire prediction and to ensure the model can learn the distribution differences between positive and negative samples, non-fire points are randomly selected at a 1:2 ratio, and their dynamic and static features are extracted using the same method to form negative samples. Subsequently, the positive and negative samples are divided into training and test sets.

[0080] 4) Import the training set data and the corresponding causal delay matrix and causal strength matrix into the CSTGCN model for training. Select the binary cross-entropy loss function (BCELoss) as the loss function, use the Adam optimizer to iteratively update the network weights and parameters, and calculate evaluation metrics to monitor the network training effect.

[0081] 5) After reaching the maximum number of iterations, save the network weights and parameters, input the test set into the trained model, and generate the power grid wildfire risk prediction results.

[0082] Therefore, the pseudocode of the power grid wildfire prediction algorithm based on the CSTGCN model is shown in Table 2:

[0083] Table 2. Pseudocode of the power grid wildfire prediction algorithm based on the CSTGCN model.

[0084]

[0085] 2. Causal relationship analysis

[0086] Using typical spatiotemporal units during periods of high wildfire incidence, the causal findings are visualized. Figure 3 The causal structure generated based on the PCMCI algorithm is shown.

[0087] in, Figure 3 (a) is a causal directed acyclic graph (DAG). The directed arrows on the edges indicate the direction of causal action; the color intensity corresponds to the strength of the causal action, with red indicating a promoting effect and blue indicating an inhibiting effect; the numbers on the edges represent the lag time step of the causal relationship; and the node color indicates the strength of the self-causal influence of the factor's historical state on its current state. The graph shows that the self-causal strength of the fire point density is high, indicating that the historical state of the fire points has a strong and continuous impact on the current fire situation. Without effective human intervention, the fire is unlikely to extinguish itself, leading to a continuous increase in fire point density. Simultaneously, fire point density exhibits significant causal relationships with multiple factors. Increased temperature, decreased dew point, and reduced precipitation synergistically reduce air humidity, making combustibles dry and flammable, thus triggering wildfires and increasing fire point density. Furthermore, frequent wildfires reduce air humidity through the combustion of combustibles, leading to a decrease in the SPEI (Special Purpose Energy Index) and exacerbating regional drought. Drought further promotes temperature increases, forming a "wildfire-drought-warming" cycle, continuously increasing the risk of wildfires. An increase in NDVI indicates lush vegetation and high cover, providing ample fuel for wildfires and increasing the probability of wildfires. Further selection of the maximum hysteresis length... A time-delay causal graph is drawn to illustrate the transmission characteristics of causal effects over time, complementing the causal directed acyclic graph, such as... Figure 3 As shown in (b), the variables causally related to fire point density have a time lag of 3-7 steps, indicating that changes in fire point density are not an immediate response, but rather the result of the synergistic effects of meteorological, drought, and vegetation factors over time. These factors form a dynamic causal network through complex mutual coupling, collectively leading to lagged changes in fire point density.

[0088] 3. Analysis of Prediction Results

[0089] Three consecutive days of samples were selected from the test set, and the predicted probabilities output by the model were divided into five levels—low risk, lower risk, medium risk, higher risk, and high risk—based on equal intervals. A spatial distribution map of wildfire risk was then drawn accordingly. Figure 4As shown in the enlarged view, the risk distribution in the concentrated fire point area is illustrated. The CSTGCN model outputs a high or high risk level at the actual fire point locations, accurately identifying potential wildfire areas. This demonstrates that the model effectively identifies the causal driving relationships between variables, achieving a reliable nonlinear mapping from massive amounts of data to wildfire probability. The number of actual fire points in the enlarged view gradually increased over three consecutive days, and the CSTGCN model accurately responded to this change. The high-risk area gradually expanded over time, consistent with the fire point evolution trend.

[0090] In summary, causal relationship modeling plays a crucial role in improving the performance of wildfire risk prediction. The PCMCI causal discovery algorithm effectively identifies the true lagged causal relationships between variables in multi-source time-series data, thus providing a physically interpretable causal structure for the GCN layer of the power grid wildfire prediction model, making the information it transmits in the spatial dimension more consistent with the actual laws of fire risk propagation. Based on this structure, the GCN can fully explore the spatial correlation characteristics between influencing factors, achieving efficient aggregation and hierarchical propagation of information in the causal direction. Therefore, the CSTGCN model has high accuracy and good generalization ability in complex nonlinear backgrounds, effectively characterizing spatiotemporal evolution laws and accurately identifying wildfire risks, demonstrating the comprehensive advantages of the proposed method in causal-driven spatiotemporal modeling. The model prediction results are in good agreement with the actual fire situation, indicating that the model performs well in capturing spatial differences in risk areas and identifying potential high-risk points. The wildfire risk distribution map generated by the model can intuitively reflect the regional risk distribution and provide a reference for key section inspections and power grid disaster prevention scheduling.

Claims

1. A power grid wildfire prediction method based on causal driving and spatiotemporal graph convolutional networks, characterized in that, We designed dynamic and static feature encoders to extract high-dimensional spatiotemporal features from multi-source heterogeneous data, used a causal discovery algorithm to construct a dynamically evolving causal graph topology, identified the real causal driving relationship between variables, decoupled the causal intensity matrix into positive and negative adjacency matrices, designed a graph convolution module with causal constraints to aggregate and propagate high-order information, and finally achieved accurate prediction of power grid wildfires.

2. The power grid wildfire prediction method based on causal driving and spatiotemporal graph convolutional networks according to claim 1, characterized in that, Specifically, the following steps are included: (1) Problem definition Accurate prediction of the study area The probability that each grid pixel k will be affected by a wildfire in the future. [0,1]; The study area is spatially divided into I sub-regions R. i (i=1,2,…,I), each sub-region has a spatial size of H×W and a total of K pixels; it is divided into J consecutive time periods T. j (j=1,2,…,J), with a step size of S for each time interval; In each pair of spacetime units (R) i ,T j Within this range, the basic coupling patterns among various wildfire influencing factors are relatively fixed; based on this division, a multidimensional vector is constructed: ； Where: x i,j,k,m For spacetime unit (R) i ,T j The value of the m-th factor at pixel k; M is the number of factors; the problem of wildfire prediction in power grids can be abstracted as a nonlinear mapping process from multivariate feature vectors to pixel-level wildfire occurrence probabilities: ； In the formula: y i,j =[y i,j,1 ,y i,j,2 ,…,y i,j,K ] represents a spacetime unit (R) i ,T j The probability of a wildfire occurring at each pixel within the range; f(·) is the mapping function; and These represent the spatial dependency structure and temporal dynamic characteristics of the variables within the unit, respectively. (2) Dataset construction A dataset of power grid wildfire impact factors was constructed, covering 14 categories of variables in seven aspects: historical fire points, meteorological characteristics, vegetation conditions, drought conditions, human activities, topographic structure, and transmission lines. Based on time resolution, influencing factors are divided into daily-scale dynamic variables and annual-scale static variables; dynamic data of continuous T days are used as input to enhance the ability to characterize the temporal dependence and interaction mechanism between variables; static data of the h×w grid near the center point are extracted to capture spatial correlation. (3) Constructing a causal graph generation framework A power grid wildfire prediction method based on GSTGCN is adopted; the PCMCI causal discovery algorithm is combined with GCN; the real causal driving relationship is quantitatively identified from multi-source spatiotemporal data, and a causal topology graph is constructed to replace the fixed adjacency matrix based on spatial distance in the standard GCN; based on the division of spatiotemporal units in the problem definition in (1), for each spatiotemporal unit (R i ,T j By extending the time range forward to L, multi-source data of all pixels within the spatial range H×W and time step L are extracted to construct a spatiotemporal dataset for causal discovery. ; For each factor m, the spatial average of all pixels is taken over time steps to obtain a multidimensional time series reflecting the overall evolutionary characteristics of the region. The sequence is used as input to the PCMCI algorithm to quantitatively identify the real causal driving relationship between the factors corresponding to the spatiotemporal unit, and outputs the causal strength matrix and the causal delay matrix representing the strength of the causal relationship between the factors. Then, using time step S as a sliding window, the causal relationship of the next spatiotemporal unit is calculated step by step along the time dimension, realizing the dynamic update of the causal graph and providing interpretable causal input for the subsequent power grid wildfire prediction model; (4) Construct a power grid wildfire prediction model After obtaining a dynamically updatable causal topology graph, a GSTGCN power grid wildfire prediction model is designed, consisting of three modules: a dynamic feature encoder, a static feature encoder, and a causal GCN layer.

3. The power grid wildfire prediction method based on causal driving and spatiotemporal graph convolutional networks according to claim 2, characterized in that, In step (2), during the data preprocessing stage, missing value imputation, normalization and outlier correction are performed on dynamic variables to ensure the continuity and consistency of the time series. Spatial pruning, resampling, and standardization are performed on static variables; Subsequently, each variable is uniformly encoded and aligned according to its dimension, and saved as a tensor for model use; the preprocessed dynamic and static tensor shapes input to the network are (B, T, C) respectively. d (B,C) s (,h,w); where B is the batch size; C d and C s These represent the number of dynamic and static variables, respectively.

4. The power grid wildfire prediction method based on causal driving and spatiotemporal graph convolutional networks according to claim 2, characterized in that, The specific method for step (4) is as follows: First, a dynamic feature encoder is designed to extract high-level semantic features of temporal dynamic variables. Then, one-dimensional grouped convolutions are used to encode the dynamic variables, treating each node's data channel as an independent convolutional group to ensure the independence of feature extraction. After the convolution operation, the features are mapped to a high-dimensional latent space, forming a node state sequence. Where F represents the dimension of the high-level semantic features of the node at each time step; To quantitatively characterize the lagged causal effects between nodes, based on the (3) causal graph generation framework, time-delay information is extracted from the node state sequence through the causal time-delay matrix, and weighted aggregation is performed in combination with the causal strength matrix to obtain a node representation that integrates causal dependencies. The operation is shown in the equation: ； In the formula: Z a (t) represents the representation of node a after fusing its causal neighborhood historical information at time step t; S ab The causal influence strength of node b on node a is given by the causal strength matrix; This represents the corresponding causal lag order; In the causal GCN layer, the causal strength matrix is ​​used as the adjacency matrix to perform graph convolution operation on the node representation, so that the local causal information of the node propagates along the causal topology, thereby capturing the high-order interaction relationship between nodes. Decompose the causality strength matrix into a positive matrix A. + and negative matrix A - As an adjacency matrix, it is used for weighted propagation in graph convolution. Positive edges enhance the facilitation effect between nodes, while negative edges independently model the suppression effect, avoiding the loss of symbolic information in nonlinear activation functions such as ReLU; the operation is shown in the equation: ； In the formula: f ReLU (·) represents the ReLU activation function; B(·) represents the batch normalization layer; , These are the positive and negative normalized causal adjacency matrices, respectively; f Drop (·) represents a random deactivation layer; , The weight matrices are for positive and negative subgraphs; , For bias terms; The static feature encoder is designed to gradually increase the number of channels and reduce the spatial size by using multiple convolutional blocks to extract spatial representations from local to global. Batch normalization and ReLU activation functions are added after each convolutional layer to stabilize the training process and enhance nonlinear expressive power. After multiple convolutional processing, static variables are mapped to a high-dimensional latent space with the same dimension as dynamic variables. Finally, the dynamic and static features are concatenated to obtain a fused feature vector. After processing by a fully connected layer, the output result is controlled to the [0,1] interval by the Sigmoid activation function.

Images