Grid dynamic load forecasting system fused with weather feature enhancement

Abstract

The present application relates to the field of electric digital data processing, specifically to a power grid dynamic load prediction system fused with weather feature enhancement. The system calculates mutual information based on the historical load sequence of each node of the power grid and the multi-dimensional weather element sequence, dynamically updates the weight of the adjacent matrix of the graph structure according to the mutual information, and constructs a weather load dynamic correlation topology. The weather feature is decoupled in time and space dimensions. The time dimension uses the extended causal convolution to extract the long-term trend and short-term fluctuation of the weather. The spatial dimension performs graph convolution based on the above topology to aggregate the heterogeneous features of adjacent nodes affected by the weather. The weather-enhanced features after time-space decoupling and the load historical features are nonlinearly fused in the hidden space, and the dynamic load prediction result is output. The present application establishes a deep nonlinear mapping between weather and load, captures the spatial transmission path of the weather acting on the load, and eliminates the load prediction lag and distortion caused by weather mutations.

Description

Technical Field

[0001] This invention relates to the field of electrical digital data processing, specifically to a power grid dynamic load forecasting system that integrates meteorological features. Background Technology

[0002] In existing power grid dynamic load forecasting technologies, the impact of meteorological factors on load changes is crucial. Conventional power grid dynamic load forecasting schemes typically use acquired multi-dimensional meteorological data as independent feature vectors, which are then simply concatenated or directly linked with historical power grid load data at the input layer. The resulting hybrid features are then input into a deep neural network for time-series forecasting. In such schemes, the correlation between meteorological elements and power grid load nodes is statically preset or directly ignored. The heterogeneous impact of meteorological conditions at different spatial locations on the load cannot be explicitly modeled in the model. Furthermore, the processing of meteorological data is often limited to simple smoothing in the time domain or truncating fixed time windows, without addressing the specific mechanisms by which meteorological changes are transmitted through the power grid topology.

[0003] Existing technologies that simply stitch together meteorological and load data sever the physical connection between meteorological elements and load nodes in the power grid's spatial topology. This results in models that can only learn a shallow linear correspondence between meteorological data and load data, failing to characterize the heterogeneous impacts and spatiotemporal transmission delays of meteorological elements in different spatial regions of the power grid. When meteorological changes occur abruptly, models lacking spatial topology mapping and spatiotemporal decoupling mechanisms cannot track the dynamic transmission path of meteorological effects on the load, leading to the core technical problem of load forecast lag and distortion under meteorological abrupt changes. Summary of the Invention

[0004] The purpose of this invention is to provide a dynamic load forecasting system for power grids that integrates meteorological features, which can effectively solve the problems mentioned in the background art.

[0005] To achieve the above objectives, the technical solution adopted by the present invention is as follows: A dynamic load forecasting system for power grids that integrates meteorological features includes: a component for calculating the mutual information of different meteorological elements on the load of each node under different time windows based on the historical load sequence of each node of the power grid and the multi-dimensional meteorological element sequence, dynamically updating the adjacency matrix weight of the graph structure according to the mutual information, and constructing a dynamic correlation topology of meteorological load; This component is used to decouple the input meteorological features in the time and spatial dimensions. In the time dimension, it uses dilated causal convolution to extract the long-term trend and short-term fluctuation features of meteorological changes. In the spatial dimension, it performs graph convolution based on the dynamic correlation topology of the meteorological load to aggregate the heterogeneous features of adjacent nodes affected by meteorology. A component used to nonlinearly fuse spatiotemporally decoupled meteorological enhancement features with load history features in the latent space, and output dynamic load prediction results through fully connected mapping.

[0006] Preferably, the calculation of the mutual information of different meteorological elements on the load of each node under different time windows, and the dynamic updating of the adjacency matrix weight of the graph structure based on the mutual information include: for any two meteorological elements in the multi-dimensional meteorological elements, calculating the conditional mutual information under a given historical load sequence, and removing meteorological element pairs whose conditional mutual information is less than a preset redundancy threshold. For the retained meteorological elements, a joint entropy expression of multidimensional meteorological joint probability distribution and load node sequence is constructed. The difference between the joint entropy expression and the edge entropy of the load node is solved to obtain the mutual information. The time-varying sequence of mutual information is extracted based on a sliding time window. The time-varying sequence is then input into a hyperbolic tangent activation function for normalization. The normalization result is used as the dynamic weight value of the edge connection to the corresponding node in the adjacency matrix of the graph structure.

[0007] Preferably, the extraction of long-term trend and short-term fluctuation features of meteorological changes by using dilated causal convolution in the time dimension includes: inputting the input meteorological features in parallel to multiple causal convolution branches with different dilation rates, wherein the first branch uses a dilated convolution kernel with an exponentially increasing dilation rate with network depth to capture long-term meteorological evolution patterns, and the second branch uses a one-dimensional convolution kernel with a fixed small dilation rate to extract local transient meteorological fluctuations. The output features of the first branch and the second branch are concatenated along the channel dimension. The concatenated features are then subjected to cross-channel information interaction and dimensionality reduction through depthwise separable convolution. The dimensionality-reduced features are then output as the time dimension decoupled features of the meteorological changes.

[0008] Preferably, in the spatial dimension, performing graph convolution based on the dynamic correlation topology of the meteorological load to aggregate the heterogeneity features of adjacent nodes affected by meteorology includes: determining the central node and its set of neighboring nodes based on the adjacency matrix of the dynamic correlation topology of the meteorological load, and extracting the load feature vector of the central node and the meteorological feature vector of each neighboring node in the set of neighboring nodes; Calculate the cosine similarity between the load feature vector of the central node and the meteorological feature vectors of each neighboring node, and normalize the cosine similarity into a spatial attention coefficient using the softmax function; The spatial attention coefficient is weighted and summed with the meteorological feature vectors of the corresponding neighboring nodes. The weighted sum is then residually concatenated with the original load features of the central node to obtain the heterogeneity features.

[0009] Preferably, the nonlinear fusion of the spatiotemporally decoupled meteorological enhancement features and load history features in the latent space includes: mapping the spatiotemporally decoupled meteorological enhancement features into query vectors and key vectors, and mapping the load history features into value vectors; Calculate the dot product between the query vector and the key vector, divide the dot product by the scaling factor, and then input the result into the softmax function to obtain the latent space attention weight matrix. Multiply the latent space attention weight matrix with the value vector to obtain the meteorological-driven load feature reconstruction vector; The reconstructed load feature vector is multiplied element-wise with the original historical load features, and the multiplication result is input into a multilayer perceptron. The dynamic load prediction result is then output through the nonlinear activation function in the multilayer perceptron.

[0010] Preferably, the construction of a dynamic meteorological load association topology based on the adjacency matrix weights of the graph structure dynamically updated according to the mutual information further includes: applying topological sparsity constraints to the updated adjacency matrix, and removing weak connection edges in the adjacency matrix whose weight values ​​are lower than the active threshold by calculating the L1 norm of the adjacency matrix and introducing Lagrange multipliers. A time smoothing regularization term is introduced, and the Frobenius norm difference between the adjacency matrix at the current time step and the adjacency matrix at the previous time step is calculated. The norm difference is then added as a penalty term to the objective function for updating the adjacency matrix. The adjacency matrix is ​​iteratively optimized based on the updated objective function to ensure the structural continuity and sparsity of the dynamic correlation topology of meteorological load during the time evolution process.

[0011] Preferably, the calculation of conditional mutual information under a given load history sequence includes: estimating the joint probability density using a non-parametric method based on kernel density estimation, taking into account the continuous variable characteristics of the multidimensional meteorological elements and the load history sequence; A Gaussian kernel function is selected, and the bandwidth parameter of the Gaussian kernel function is adaptively determined based on Scott's rule. The bandwidth parameter is then used to smooth the joint probability density. Substitute the joint probability density into the integral expression of the conditional mutual information, and transform the integral expression into a discrete summation operation through Monte Carlo sampling; A locality-sensitive hashing mechanism is introduced to perform approximate nearest neighbor retrieval on sample pairs in the discrete summation operation, eliminating sample pairs whose distance exceeds a set threshold to reduce computational overhead, and outputting the conditional mutual information.

[0012] Preferably, inputting the meteorological features in parallel to multiple causal convolution branches with different dilation rates includes: performing a fast Fourier transform on the input meteorological features to extract the dominant frequency components of meteorological changes in the frequency domain; The corresponding meteorological change cycle length is calculated based on the dominant frequency component, and the meteorological change cycle length is mapped to the dilation rate benchmark value of each of the causal convolution branches. A learnable scaling factor is introduced, and the expansion rate baseline value is multiplied by the learnable scaling factor to obtain the dynamically adjusted target expansion rate; Hollow sampling is performed on the convolution kernel in the causal convolution branch according to the target expansion rate, so that different meteorological elements adaptively match the receptive field of different time scales to capture multi-band meteorological change characteristics.

[0013] Preferably, normalizing the cosine similarity to a spatial attention coefficient using a softmax function includes: after calculating the cosine similarity, introducing a temperature scaling factor, wherein the temperature scaling factor is inversely proportional to the variance of the fluctuation of the current meteorological element; Divide the cosine similarity by the temperature scaling factor and then input it into the softmax function to obtain the smoothed spatial attention coefficient; When the variance of the meteorological element exceeds the stability threshold, the temperature scaling factor is increased to suppress the attention weight spikes caused by extreme weather fluctuations. DropEdge regularization is applied to the spatial attention coefficients, and during the model training phase, some neighboring nodes' edges are randomly dropped with a set probability to prevent the spatial attention coefficients from overfitting to a specific meteorological topology.

[0014] Preferably, outputting the dynamic load prediction result through the nonlinear activation function in the multilayer perceptron includes: performing an affine transformation on the input feature vector in the hidden layer of the multilayer perceptron, and calculating the mean and variance of the feature vector in the channel dimension; The feature vector is standardized based on the mean and variance, and learnable scaling and translation parameters are introduced to restore the expressive power of the feature distribution. The standardized feature vector is input into the parameterized PReLU activation function, where the negative half-axis slope is dynamically initialized based on the signal-to-noise ratio of the meteorological features. The activated feature vector is input into the subsequent hidden layer for nonlinear mapping, and a linear transformation is used in the output layer to map the feature dimension to the prediction duration dimension to obtain the dynamic load prediction result.

[0015] Compared with the prior art, the beneficial effects of the present invention are as follows: 1. This invention calculates the mutual information of different meteorological elements on the load of each node under different time windows, and dynamically updates the adjacency matrix weights of the graph structure based on the mutual information to construct a dynamic meteorological load correlation topology. In the spatial dimension, graph convolution is performed based on this topology to aggregate the heterogeneity features of adjacent nodes affected by meteorology. In the temporal dimension, dilated causal convolution is used to extract the long-term trend and short-term fluctuation of meteorological changes. The meteorological enhancement features after spatiotemporal decoupling and the historical features of the load are nonlinearly fused in the latent space to establish a deep nonlinear mapping relationship between meteorology and load. The spatial transmission path of meteorological effects on the load is captured through dynamic topological mapping, eliminating the load prediction lag and distortion caused by meteorological abrupt changes.

[0016] 2. By applying a topological sparsification constraint based on the first-order norm and a time smoothing regularization term based on the Frobenius norm difference to the updated adjacency matrix, weak connections are eliminated while maintaining the temporal continuity of the topology, reducing the computational resource overhead of the dynamic graph construction process. By extracting the dominant frequency of meteorological changes through fast Fourier transform and mapping it to the dynamic adjustment target expansion rate of causal convolution branches, different meteorological elements are adaptively matched to the multi-scale temporal receptive field. In the calculation of the spatial attention coefficient, a temperature scaling factor inversely proportional to the variance of meteorological fluctuations is introduced to suppress the weight spikes caused by extreme meteorological fluctuations, enhancing the numerical stability of the prediction system under complex meteorological fluctuations. Attached Figure Description

[0017] Figure 1 This is a flowchart illustrating the overall execution of the power grid dynamic load forecasting system with enhanced meteorological features, as described in this invention. Figure 2 This is a flowchart of the calculation and redundancy elimination of meteorological element condition mutual information in this invention; Figure 3 This is a flowchart of the meteorological feature temporal dimension extended causal convolution feature extraction process of the present invention; Figure 4 This is a flowchart of the spatial dimension graph convolution heterogeneity feature aggregation process of the present invention; Figure 5 This is a flowchart of the nonlinear fusion and prediction output of hidden space meteorological-load characteristics in this invention; Figure 6 This is a flowchart of the dynamic topological adjacency matrix sparsification and time smoothing optimization of the present invention. Detailed Implementation

[0018] 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, not all, of the embodiments of the present invention. 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.

[0019] Please refer to Figure 1 This embodiment provides a dynamic load forecasting system for power grids that integrates meteorological features. First, it acquires the historical load sequence and multi-dimensional meteorological element sequence for each node in the power grid. The nodes include substation outgoing line nodes, distribution transformer nodes, and important user access nodes. The historical load sequence consists of active power samples from each node within continuous time steps, with a sampling interval of 15 minutes and a sequence length of 96 time steps, corresponding to the load data of the past 24 hours. The multi-dimensional meteorological element sequence includes five dimensions: temperature, relative humidity, wind speed, precipitation, and solar radiation intensity. The sampling interval for each meteorological element is consistent with the load data, and a spatial interpolation method is used to map the meteorological station data to the geographical locations of each node in the power grid, forming a node-level meteorological feature matrix that matches the dimensions of the load sequence.

[0020] Before being input into the system, the load history sequences of each node in the power grid require preprocessing. Preprocessing includes outlier detection and repair, missing value imputation, and data standardization. Outlier detection uses the 3σ criterion, marking samples exceeding the mean ± 3 standard deviations as outliers, and then using linear interpolation to imput them. Missing value imputation uses forward imputation, filling missing values ​​with values ​​from the previous valid time step. Data standardization uses Z-score standardization, converting the load data into a standard normal distribution with a mean of 0 and a standard deviation of 1.

[0021] Spatial interpolation of meteorological data employs inverse distance weighted interpolation, mapping observational data from meteorological stations to the geographical locations of nodes in the power grid. The basic principle of inverse distance weighted interpolation is that sample points closer to the interpolation point have a greater influence on the interpolation result, with the weight inversely proportional to the power of the distance. The number of meteorological stations participating in the interpolation is all national-level meteorological observation stations within the region, and the power exponent is set to 2.

[0022] The system first constructs the dynamic correlation topology of meteorological load. For each power grid node, it extracts its corresponding historical load sequence and meteorological element sequences across all dimensions, and calculates the mutual information of different meteorological elements on the node's load under different time windows. The time window lengths are set to four types: 1 hour, 4 hours, 12 hours, and 24 hours, corresponding to different lag time scales of the impact of meteorological changes on the load. Mutual information is used to quantify the statistical correlation between two random variables, and its calculation formula is as follows:

[0023] Where X represents the time series of a certain meteorological element, and Y represents the historical load series of the corresponding node. The marginal entropy of the meteorological element sequence, The marginal entropy of the load sequence, It is the joint entropy of the meteorological element sequence and the load sequence.

[0024] Marginal entropy is calculated using a histogram-based probability density estimation method. The continuous variable is discretized into equal-width intervals, and the frequency of samples occurring within each interval is statistically analyzed and substituted into the entropy calculation formula.

[0025] Where Z is an arbitrary sequence of random variables, and N is the number of discretization intervals. Let be the probability that a sample falls into the i-th interval. The number of discretization intervals is set to 10 to balance the accuracy of probability density estimation with computational complexity.

[0026] Joint Entropy The calculation uses a two-dimensional histogram method, discretizing the meteorological element sequence and the load sequence into M and N intervals respectively, and statistically analyzing the frequency of the two-dimensional joint distribution, which is then substituted into the formula:

[0027] in, Let M and N be the joint probability that a meteorological element falls in the i-th interval and a load falls in the j-th interval. Both M and N are set to 10, consistent with the number of one-dimensional discretization intervals.

[0028] After calculating the mutual information of different meteorological elements under different time windows, the system dynamically updates the adjacency matrix weights of the graph structure based on this mutual information. Each node in the graph structure corresponds to a power grid node, and the edges between nodes represent the correlation between the loads of the two nodes and the influence of the same meteorological element. For any two nodes i and j, the weight of their edge is the weighted average of the mutual information of all meteorological elements on the loads of nodes i and j under all time windows. The weight coefficient is inversely proportional to the length of the time window to highlight the impact of short-term meteorological changes on load correlation. The weight coefficient for a 1-hour window is set to 0.4, for a 4-hour window to 0.3, for a 12-hour window to 0.2, and for a 24-hour window to 0.1.

[0029] To map the mutual information values ​​to the [0,1] interval to meet the numerical requirements of graph convolution, the system inputs the calculated time-varying sequence of mutual information into the hyperbolic tangent activation function for normalization. The normalization formula is as follows:

[0030] in, Let be the weight of the edge connecting node i and node j at time t, and K be the total number of combinations of meteorological elements and time windows. The weight coefficients for the k-th combination are... Let be the mutual information value of the loads of node i and node j under the k-th combination at time t.

[0031] Through the above process, the system constructs a dynamic topology of meteorological load that evolves over time. This topology can reflect in real time the degree of influence of different meteorological elements on the load of each node of the power grid at different time scales, as well as the correlation between nodes affected by meteorological conditions.

[0032] Next, the system performs a spatiotemporal decoupling extraction process for meteorological features. First, the input node-level meteorological feature matrix is ​​decoupled in the time and spatial dimensions to extract the temporal and spatial features of meteorological changes, respectively.

[0033] refer to Figure 3 In the time dimension, the system employs dilated causal convolution to extract long-term trends and short-term fluctuations in meteorological changes. Dilated causal convolution expands the receptive field without increasing the number of parameters by introducing holes into the convolution kernel. Simultaneously, causal constraints ensure that the convolution operation utilizes only historical data from the current moment and prior, avoiding the leakage of future information. The system sets up two parallel causal convolution branches: the first branch uses a dilated convolution kernel with an exponentially increasing dilation rate with network depth to capture long-term meteorological evolution patterns; the second branch uses a one-dimensional convolution kernel with a fixed small dilation rate to extract local transient meteorological fluctuations.

[0034] The expansion rate of the first branch is set to ,in The depth of the network layers increases from 0. The first branch contains four causal convolutional layers with dilation rates of 1, 2, 4, and 8, respectively. For the input meteorological feature sequence... ,in The time step length, As a meteorological feature dimension, the first branch is... The output features of the layer are calculated through a one-dimensional causal convolution operation with a kernel size of 3 and 32 output channels.

[0035] The second branch adopts a fixed expansion rate. The one-dimensional causal convolution kernel has a kernel size of 3, contains 2 layers of causal convolution, and has 16 output channels.

[0036] The system concatenates the output features of the first and second branches along the channel dimension to obtain the concatenated features. Subsequently, the system uses depthwise separable convolution to perform cross-channel information interaction and dimensionality reduction on the concatenated features. Depthwise separable convolution consists of two steps: depthwise convolution and pointwise convolution. Depthwise convolution performs convolution operations independently on each channel, while pointwise convolution uses a 1×1 convolution kernel to achieve cross-channel information fusion. The output of the depthwise separable convolution has 32 output channels, and the output is the decoupled feature representing the time dimension of meteorological changes. .

[0037] refer to Figure 4 In the spatial dimension, the system performs graph convolution based on the constructed dynamic topology of meteorological loads to aggregate the heterogeneous features of neighboring nodes affected by meteorological conditions. The graph convolution operation takes the meteorological feature vectors and dynamic adjacency matrix of a node as input, updating the feature representation of the central node by aggregating the features of neighboring nodes. The system employs a graph convolution method approximated by Chebyshev polynomials to reduce the computational complexity of graph convolution; the order of the Chebyshev polynomial is set to 3 to balance computational complexity and feature extraction capability.

[0038] Specifically, the system first uses the adjacency matrix of the dynamic correlation topology of meteorological loads. Determine the central node and its neighbor node set ,in Extract the central node. Load characteristic vector Each neighbor node in the neighbor node set meteorological feature vector .

[0039] The system calculates the cosine similarity between the load feature vector of the central node and the meteorological feature vectors of each neighboring node. The cosine similarity measures the degree of similarity between two vectors in direction, and its calculation formula is as follows:

[0040] in, As the central node with neighboring nodes Cosine similarity between them This represents the dot product operation of vectors. This represents the L2 norm of a vector.

[0041] The system normalizes the calculated cosine similarity into spatial attention coefficients using the softmax function. The normalization formula is as follows:

[0042] in, As the central node For neighboring nodes Spatial attention coefficient, This represents an exponential function.

[0043] The system performs a weighted summation of the spatial attention coefficient and the meteorological feature vectors of the corresponding neighboring nodes to obtain the aggregated neighboring feature vector. Subsequently, the system performs a residual concatenation between the weighted summation result and the original load features of the central node to obtain the heterogeneity features. The residual connection maps the dimension of the neighboring meteorological features to the same dimension as the load features of the central node through a linear transformation matrix.

[0044] Through the above decoupling and extraction process of the time and space dimensions, the system obtains the meteorological enhancement features after spatiotemporal decoupling. .

[0045] Next, refer to Figure 5 The system nonlinearly fuses the spatiotemporally decoupled meteorological enhancement features with historical load features in the latent space, outputting dynamic load prediction results. The system first maps the historical load sequence to historical load features through a one-dimensional convolutional layer. The kernel size is 3, and the number of output channels is 32.

[0046] The system enhances meteorological features after spatiotemporal decoupling. Mapped to a query vector through linear transformation With key vector , load historical characteristics Mapped to a value vector through a linear transformation The linear transformation is achieved through a learnable weight matrix, with dimensions of [dimensions to be filled in]. and .

[0047] The system calculates the dot product of the query vector and the key vector, and then divides the dot product by the scaling factor. The latent space attention weight matrix is ​​then obtained by inputting the softmax function. The scaling factor prevents the dot product from becoming too large, which could cause the softmax function to enter the gradient saturation region. The formula for calculating the latent space attention weight matrix is:

[0048] in, Let K be the transpose of the key vector K. This specifies the dimensions of the query vector and key vector.

[0049] The system multiplies the latent space attention weight matrix with the value vector to obtain the meteorological-driven load feature reconstruction vector. Subsequently, the system performs element-wise multiplication of the reconstructed load feature vector with the original historical load features to obtain the fused feature vector. The element-wise multiplication operation can achieve dynamic modulation of meteorological characteristics on historical load characteristics.

[0050] Finally, the system inputs the fused feature vector into a multilayer perceptron (MLP), and outputs the dynamic load prediction result through the nonlinear activation function in the MLP. The MLP consists of two hidden layers and one output layer. The hidden layers use the ReLU activation function, and the output layer uses a linear transformation to map the feature dimension to the prediction duration dimension. The prediction duration is set to 24 hours, corresponding to 96 time steps. The number of neurons in the two hidden layers are 128 and 64, respectively. The mutual information values ​​of typical meteorological elements and nodal loads under different time windows are shown in Table 1.

[0051] Table 1. Mutual information values ​​of typical meteorological elements and nodal loads under different time windows.

[0052] Table 1 lists the calculated mutual information of three nodes in a typical regional power grid under four time windows for different meteorological factors and loads on a typical summer day.

[0053] This embodiment constructs a dynamic topology for meteorological load association, realizing dynamic modeling of the spatial relationship between meteorological elements and power grid node loads. By using a spatiotemporal decoupling feature extraction method, it captures the long-term trend and short-term fluctuation characteristics of meteorological changes, as well as spatial heterogeneity characteristics. Through the latent spatial attention mechanism, it realizes the deep nonlinear fusion of meteorological features and load features, and establishes a deep nonlinear mapping relationship between meteorology and load.

[0054] In a preferred embodiment, reference Figure 2 When calculating the mutual information of different meteorological elements on the load of each node under different time windows, the system first calculates the conditional mutual information of any two meteorological elements in the multi-dimensional meteorological elements under the given historical load sequence. Meteorological element pairs with conditional mutual information less than the preset redundancy threshold are removed to eliminate redundant information between meteorological elements.

[0055] Conditional mutual information is used to quantify the statistical correlation between two variables given a third variable. Its calculation formula is as follows:

[0056] Where Z is a given load history sequence, Let X be the conditional entropy of X under condition Z. Let Y be the conditional entropy under condition Z. Let X be the joint conditional entropy of X and Y under condition Z.

[0057] To address the continuous variable characteristics of multi-dimensional meteorological elements and historical load sequences, the system employs a non-parametric method based on kernel density estimation to estimate the joint probability density, thus avoiding the assumptions about data distribution imposed by parametric methods. The formula for kernel density estimation is as follows:

[0058] in, Let be the probability density estimate of random variable Z at point z, where n is the sample size and h is the bandwidth parameter of the kernel function. This is the kernel function.

[0059] The system selects the Gaussian kernel function as the kernel function for kernel density estimation. The expression for the Gaussian kernel function is:

[0060] Where u is the input variable of the kernel function.

[0061] The system adaptively determines the bandwidth parameters of the Gaussian kernel function based on Scott's rule. The bandwidth calculation formula for Scott's rule is as follows:

[0062] in, This is an estimate of the standard deviation of the sample.

[0063] After smoothing the joint probability density using the aforementioned bandwidth parameters, the system substitutes the joint probability density into the integral expression of the conditional mutual information, and transforms the integral expression into a discrete summation operation through Monte Carlo sampling. The number of samples for Monte Carlo sampling is set to 10,000 to ensure the accuracy of the integral approximation. The system randomly selects 10,000 sample pairs from the joint distribution of meteorological element sequences and load sequences, and substitutes them into the discrete summation formula of the conditional mutual information for calculation.

[0064] To reduce computational overhead, the system introduces a Locality Sensitive Hashing (LSH) mechanism to perform approximate nearest neighbor retrieval on sample pairs in discrete summation operations. Sample pairs with a distance exceeding a set threshold are discarded, retaining only those with closer distances for the summation operation. The LSH mechanism employs a random projection-based LSH method, mapping high-dimensional sample vectors to a low-dimensional hash space. This ensures that sample vectors that are close in the original space have a higher probability of being mapped to the same hash bucket in the hash space. The system generates 10 sets of random projection vectors, performs a dot product operation on each sample vector and the random projection vector, and generates a 10-bit hash code based on the sign of the dot product result. For two sample vectors, if their hash codes are the same, they are considered approximate neighbors and participate in the conditional mutual information calculation; if their hash codes are different, they are considered far apart and are discarded.

[0065] After calculating the conditional mutual information of all meteorological element pairs, the system ensures that the conditional mutual information is less than a preset redundancy threshold. Meteorological elements are eliminated, and a redundancy threshold is preset. Set to 0.05. For the eliminated meteorological element pairs, only the meteorological element with the larger mutual information with the load is retained, thereby eliminating redundant information between meteorological elements and reducing the complexity of subsequent calculations.

[0066] For the retained meteorological elements, the system constructs a joint entropy expression for the multidimensional meteorological joint probability distribution and the load node sequence, and solves for the difference between the joint entropy expression and the marginal entropy of the load nodes to obtain the mutual information. The estimation of the multidimensional meteorological joint probability distribution adopts the multidimensional kernel density estimation method, which extends the above one-dimensional kernel density estimation to a multidimensional space. The bandwidth parameter adopts the multidimensional extension form of Scott's rule.

[0067] After calculating the mutual information, the system extracts a time-varying sequence of the mutual information based on a sliding time window. The length of the sliding time window is set to 24 hours, and the sliding step size is 15 minutes, consistent with the data sampling interval. The system inputs the time-varying sequence of the mutual information into a hyperbolic tangent activation function for normalization, and uses the normalization result as the dynamic weight value of the edge connection between corresponding nodes in the adjacency matrix of the graph structure.

[0068] Further, refer to Figure 6 The system applies topological sparsity constraints to the updated adjacency matrix. This is achieved by calculating the L1 norm of the adjacency matrix and introducing Lagrange multipliers to remove weak connections with weights below the activity threshold. The objective function for the topological sparsity constraints is:

[0069] Where A is the optimized adjacency matrix. The initial adjacency matrix, Describe the Frobenius norm. Describing the L1 norm, The Lagrange multiplier for the L1 regularization term is set to 0.01.

[0070] The formula for calculating the L1 norm is:

[0071] in, This represents the number of nodes in the graph structure.

[0072] By minimizing the above objective function, the system can maintain the similarity between the adjacency matrix and the initial adjacency matrix while making most elements in the adjacency matrix approach 0, thus achieving topology sparsity. The system assigns weights in the adjacency matrix below an activity threshold. Setting the element to 0 removes weak connections and sets the activity threshold. Set it to 0.1.

[0073] Simultaneously, the system introduces a time-smoothing regularization term, calculating the Frobenius norm difference between the adjacency matrix at the current time and the adjacency matrix at the previous time. This norm difference is added as a penalty term to the objective function for updating the adjacency matrix, ensuring the structural continuity of the dynamic topology of meteorological loads during temporal evolution. The expression for the time-smoothing regularization term is:

[0074] in, Let be the adjacency matrix at time t. for The adjacency matrix at time 10:00. The weighting coefficient for the time smoothing regularization term is set to 0.1.

[0075] Combining topological sparsity constraints and a time-smoothing regularization term, the system constructs the final adjacency matrix update objective function. The system iteratively optimizes the adjacency matrix based on this objective function, with 10 iterations, using gradient descent to find the minimum value of the objective function. The learning rate is set to 0.01. In each iteration, the system calculates the gradient of the objective function with respect to the elements of the adjacency matrix and then updates the values ​​of the adjacency matrix elements based on the gradient. During iteration, the system restricts the values ​​of the adjacency matrix elements to the interval [0,1] to ensure the effectiveness of the weight values. After the iteration ends, the system removes elements from the adjacency matrix whose weight values ​​are below an activity threshold. The elements are set to 0 to complete the topology sparsity processing. Through iterative optimization, the system obtains a dynamic adjacency matrix that simultaneously satisfies sparsity and temporal continuity, and constructs a dynamic correlation topology for meteorological loads. The comparison of the adjacency matrix structure parameters before and after topology sparsity constraints is shown in Table 2.

[0076] Table 2 Comparison of Adjacency Matrix Structure Parameters Before and After Topological Sparsity Constraints

[0077] Table 2 shows a comparison of the structural parameters of the adjacency matrix before and after topology sparsification constraints for a typical regional power grid under typical summer days.

[0078] This embodiment introduces conditional mutual information to eliminate redundant information between meteorological elements, uses a non-parametric method based on kernel density estimation to accurately estimate the probability density of continuous variables, and optimizes the structure of the adjacency matrix through topological sparsity constraints and time smoothing regularization terms. While reducing computational complexity, it ensures the structural continuity and sparsity of the dynamic association topology of meteorological loads during the time evolution process.

[0079] In another preferred embodiment, when the system extracts the long-term trend and short-term fluctuation features of meteorological changes using dilated causal convolution in the time dimension, it first performs a Fast Fourier Transform (FFT) on the input meteorological features to extract the dominant frequency components of meteorological changes in the frequency domain. The FFT converts the time-domain signal into a frequency-domain signal, clearly showing the intensity of different frequency components in the signal. The system uses a radix-2 FFT algorithm, padding the sequence length with zeros to the nearest power of 2 to improve computational efficiency. For a meteorological feature sequence of length 96, the system pads it with zeros to 128 time steps before performing the FFT.

[0080] When extracting the dominant frequency components, the system first calculates the amplitude spectrum of the frequency domain sequence, and selects the top three frequency components with the largest amplitudes in the amplitude spectrum as the dominant frequency components of meteorological changes. The corresponding frequency values ​​are as follows: ,and The length of the corresponding meteorological variation cycle is calculated based on the dominant frequency component. The formula for calculating the cycle length is:

[0081] in, Let be the period length corresponding to the m-th dominant frequency component. Let m be the frequency value of the dominant frequency component. The sequence length is given.

[0082] The system maps the length of the meteorological change cycle to the baseline value of the dilation rate of each causal convolution branch. The mapping relationship is that the baseline value of the dilation rate equals the cycle length divided by the convolution kernel size. For the first branch, the system selects the longest cycle length. As the benchmark value for the expansion rate; for the second branch, the system selects the shortest period length. As a benchmark value for expansion rate.

[0083] The system introduces a learnable scaling factor, multiplying the baseline dilation rate by the learnable scaling factor to obtain a dynamically adjusted target dilation rate. The initial value of the learnable scaling factor is set to 1 and optimized during model training using the backpropagation algorithm. The system performs dilated sampling on the convolution kernels in the causal convolution branch according to the target dilation rate, enabling different meteorological elements to adaptively match the receptive field at different time scales, thereby capturing multi-band meteorological change characteristics.

[0084] In the spatial dimension, the system performs graph convolution based on the dynamic correlation topology of meteorological load. When aggregating the heterogeneous features of neighboring nodes affected by meteorological conditions, after calculating the cosine similarity between the load feature vector of the central node and the meteorological feature vectors of each neighboring node, a temperature scaling factor is introduced. This temperature scaling factor is inversely proportional to the variance of the current meteorological element fluctuation. The formula for calculating the temperature scaling factor is:

[0085] in, This is the temperature scaling factor. The reference temperature coefficient is set to 1.0. This represents the variance of current meteorological elements. It is a very small positive number, used to prevent the denominator from being 0, and is set to... .

[0086] The system divides the cosine similarity by a temperature scaling factor and then inputs it into a softmax function to obtain a smoothed spatial attention coefficient. When the variance of meteorological elements exceeds a stationarity threshold... At that time, temperature scaling factor Decreasing this threshold increases the range of input values ​​for the softmax function, resulting in a smoother distribution of the output attention coefficients. (Stability threshold) Set to twice the historical variance of meteorological elements.

[0087] Simultaneously, the system applies DropEdge regularization to the spatial attention coefficients. During model training, it randomly discards some edges of neighboring nodes with a set probability p to prevent the spatial attention coefficients from overfitting to a specific meteorological topology. The dropout probability p for DropEdge regularization is set to 0.2. In each training batch, for each central node's set of neighboring nodes, some neighboring nodes are randomly discarded with probability p, retaining only the remaining neighboring nodes for spatial attention coefficient calculation and feature aggregation. During model inference, DropEdge regularization is turned off, and all neighboring nodes participate in the calculation.

[0088] When nonlinearly fusing the spatiotemporally decoupled meteorological enhancement features with historical load features in the latent space, the system inputs the fused feature vector into a multilayer perceptron (MLP). The LPP outputs dynamic load prediction results through a nonlinear activation function. In the hidden layers of the LPP, the system performs an affine transformation on the input feature vector. The system calculates the mean and variance of the feature vector along the channel dimension and standardizes the feature vector based on the mean and variance. During the training phase, the system calculates the mean and variance based on the samples in the current batch; during the inference phase, the system uses the moving average mean and moving average variance accumulated during training for standardization. The momentum coefficient is set to 0.9.

[0089] The system introduces learnable scaling parameters. With translation parameters The system restores the expressive power of the feature distribution. The standardized feature vector is input into a parameterized PReLU activation function, where the slope of the negative half-axis is dynamically initialized based on the signal-to-noise ratio of the meteorological features. The expression for the PReLU activation function is:

[0090] in, The negative half-axis slope parameter can be learned during model training. The initial value of the negative half-axis slope is set to the reciprocal of the signal-to-noise ratio (SNR) of the meteorological feature, which is the ratio of the signal power to the noise power of the meteorological feature.

[0091] The system inputs the activated feature vectors into subsequent hidden layers for nonlinear mapping, and then uses a linear transformation in the output layer to map the feature dimension to the prediction duration dimension, thus obtaining the dynamic load prediction result. The multilayer perceptron contains three hidden layers, with 256, 128, and 64 neurons in each layer, respectively. The number of neurons in the output layer is equal to the number of time steps corresponding to the prediction duration.

[0092] During model training, the system uses mean squared error as the loss function and the Adam optimizer for model training. The learning rate is set to 0.001, the batch size to 32, and the number of training epochs to 100. An early stopping mechanism is employed during training; training is terminated early if the validation set loss does not decrease for 10 consecutive epochs. The receptive field coverage of meteorological time features extracted under different expansion rate configurations is shown in Table 3.

[0093] Table 3. Sensitive field coverage range for meteorological time features extracted under different expansion rate configurations.

[0094] Table 3 lists the receptive field size of the causal convolution branch and the corresponding meteorological change cycle coverage under different combinations of expansion rates.

[0095] This embodiment extracts the dominant frequency of meteorological changes through Fast Fourier Transform and maps it to the dynamic adjustment target expansion rate of the causal convolution branch, enabling different meteorological elements to adaptively match the multi-scale temporal receptive field. In the calculation of spatial attention coefficient, a temperature scaling factor inversely proportional to the variance of meteorological fluctuations is introduced to suppress the weight spikes caused by extreme meteorological fluctuations. Batch standardization and parameterized PReLU activation function are introduced into the multilayer perceptron to achieve stable model training.

Claims

1. A power grid dynamic load forecasting system integrating meteorological features, characterized in that, include: This component is used to calculate the mutual information of different meteorological elements on the load of each node under different time windows based on the historical load sequence of each node of the power grid and the multi-dimensional meteorological element sequence, and to dynamically update the adjacency matrix weight of the graph structure according to the mutual information, thereby constructing a dynamic correlation topology of meteorological load. This component is used to decouple the input meteorological features in the time and spatial dimensions. In the time dimension, it uses dilated causal convolution to extract the long-term trend and short-term fluctuation features of meteorological changes. In the spatial dimension, it performs graph convolution based on the dynamic correlation topology of the meteorological load to aggregate the heterogeneous features of adjacent nodes affected by meteorology. A component used to nonlinearly fuse spatiotemporally decoupled meteorological enhancement features with load history features in the latent space, and output dynamic load prediction results through fully connected mapping.

2. The power grid dynamic load forecasting system integrating meteorological feature enhancement as described in claim 1, characterized in that, The calculation of the mutual information of different meteorological elements on the load of each node under different time windows, and the dynamic updating of the adjacency matrix weight of the graph structure based on the mutual information include: for any two meteorological elements in the multi-dimensional meteorological elements, calculating the conditional mutual information under a given historical load sequence, and removing meteorological element pairs whose conditional mutual information is less than a preset redundancy threshold. For the retained meteorological elements, a joint entropy expression of multidimensional meteorological joint probability distribution and load node sequence is constructed. The difference between the joint entropy expression and the edge entropy of the load node is solved to obtain the mutual information. The time-varying sequence of mutual information is extracted based on a sliding time window. The time-varying sequence is then input into a hyperbolic tangent activation function for normalization. The normalization result is used as the dynamic weight value of the edge connection to the corresponding node in the adjacency matrix of the graph structure.

3. The power grid dynamic load forecasting system integrating meteorological feature enhancement as described in claim 2, characterized in that, The extraction of long-term trend and short-term fluctuation features of meteorological changes using dilated causal convolution in the time dimension includes: inputting the meteorological features in parallel to multiple causal convolution branches with different dilation rates, wherein the first branch uses a dilated convolution kernel with an exponentially increasing dilation rate with network depth to capture long-term meteorological evolution patterns, and the second branch uses a one-dimensional convolution kernel with a fixed small dilation rate to extract local transient meteorological fluctuations. The output features of the first branch and the second branch are concatenated along the channel dimension. The concatenated features are then subjected to cross-channel information interaction and dimensionality reduction through depthwise separable convolution. The dimensionality-reduced features are then output as the time dimension decoupled features of the meteorological changes.

4. The power grid dynamic load forecasting system integrating meteorological feature enhancement as described in claim 1, characterized in that, In the spatial dimension, graph convolution is performed based on the dynamic correlation topology of the meteorological load to aggregate the heterogeneity features of adjacent nodes affected by meteorology, including: determining the central node and its set of neighboring nodes based on the adjacency matrix of the dynamic correlation topology of the meteorological load, and extracting the load feature vector of the central node and the meteorological feature vector of each neighboring node in the set of neighboring nodes. Calculate the cosine similarity between the load feature vector of the central node and the meteorological feature vectors of each neighboring node, and normalize the cosine similarity into a spatial attention coefficient using the softmax function; The spatial attention coefficient is weighted and summed with the meteorological feature vectors of the corresponding neighboring nodes. The weighted sum is then residually concatenated with the original load features of the central node to obtain the heterogeneity features.

5. The power grid dynamic load forecasting system integrating meteorological feature enhancement according to claim 1, characterized in that, The nonlinear fusion of the spatiotemporally decoupled meteorological enhancement features and load history features in the latent space includes: mapping the spatiotemporally decoupled meteorological enhancement features into query vectors and key vectors, and mapping the load history features into value vectors; Calculate the dot product between the query vector and the key vector, divide the dot product by the scaling factor, and then input the result into the softmax function to obtain the latent space attention weight matrix. Multiply the latent space attention weight matrix with the value vector to obtain the meteorological-driven load feature reconstruction vector; The reconstructed load feature vector is multiplied element-wise with the original historical load features, and the multiplication result is input into a multilayer perceptron. The dynamic load prediction result is then output through the nonlinear activation function in the multilayer perceptron.

6. The power grid dynamic load forecasting system integrating meteorological feature enhancement according to claim 1, characterized in that, Based on the mutual information, the adjacency matrix weights of the graph structure are dynamically updated. The construction of the dynamic association topology of meteorological load also includes: applying topological sparsity constraints to the updated adjacency matrix, calculating the L1 norm of the adjacency matrix and introducing Lagrange multipliers, and removing weak connection edges in the adjacency matrix whose weight values ​​are lower than the active threshold. A time smoothing regularization term is introduced, and the Frobenius norm difference between the adjacency matrix at the current time step and the adjacency matrix at the previous time step is calculated. The norm difference is then added as a penalty term to the objective function for updating the adjacency matrix. The adjacency matrix is ​​iteratively optimized based on the updated objective function to ensure the structural continuity and sparsity of the dynamic correlation topology of meteorological load during the time evolution process.

7. The power grid dynamic load forecasting system integrating meteorological feature enhancement according to claim 2, characterized in that, The calculation of conditional mutual information under a given load history sequence includes: estimating the joint probability density using a non-parametric method based on kernel density estimation, taking into account the continuous variable characteristics of the multidimensional meteorological elements and the load history sequence; A Gaussian kernel function is selected, and the bandwidth parameter of the Gaussian kernel function is adaptively determined based on Scott's rule. The bandwidth parameter is then used to smooth the joint probability density. Substitute the joint probability density into the integral expression of the conditional mutual information, and transform the integral expression into a discrete summation operation through Monte Carlo sampling; A locality-sensitive hashing mechanism is introduced to perform approximate nearest neighbor retrieval on sample pairs in the discrete summation operation, eliminating sample pairs whose distance exceeds a set threshold to reduce computational overhead, and outputting the conditional mutual information.

8. The power grid dynamic load forecasting system integrating meteorological feature enhancement according to claim 3, characterized in that, The input meteorological features are fed in parallel to multiple causal convolutional branches with different dilation rates, including: performing a fast Fourier transform on the input meteorological features to extract the dominant frequency components of meteorological changes in the frequency domain; The corresponding meteorological change cycle length is calculated based on the dominant frequency component, and the meteorological change cycle length is mapped to the dilation rate benchmark value of each of the causal convolution branches. A learnable scaling factor is introduced, and the expansion rate baseline value is multiplied by the learnable scaling factor to obtain the dynamically adjusted target expansion rate; Hollow sampling is performed on the convolution kernel in the causal convolution branch according to the target expansion rate, so that different meteorological elements adaptively match the receptive field of different time scales to capture multi-band meteorological change characteristics.

9. The power grid dynamic load forecasting system integrating meteorological feature enhancement according to claim 4, characterized in that, Normalizing the cosine similarity to a spatial attention coefficient using the softmax function includes: after calculating the cosine similarity, introducing a temperature scaling factor, which is inversely proportional to the variance of the fluctuation of the current meteorological elements; Divide the cosine similarity by the temperature scaling factor and then input it into the softmax function to obtain the smoothed spatial attention coefficient; When the variance of the meteorological element exceeds the stability threshold, the temperature scaling factor is increased to suppress the attention weight spikes caused by extreme weather fluctuations. DropEdge regularization is applied to the spatial attention coefficients, and during the model training phase, some neighboring nodes' edges are randomly dropped with a set probability to prevent the spatial attention coefficients from overfitting to a specific meteorological topology.

10. The power grid dynamic load forecasting system integrating meteorological feature enhancement according to claim 5, characterized in that, Outputting the dynamic load prediction result through the nonlinear activation function in the multilayer perceptron includes: performing an affine transformation on the input feature vector in the hidden layer of the multilayer perceptron, and calculating the mean and variance of the feature vector in the channel dimension; The feature vector is standardized based on the mean and variance, and learnable scaling and translation parameters are introduced to restore the expressive power of the feature distribution. The standardized feature vector is input into the parameterized PReLU activation function, where the negative half-axis slope is dynamically initialized based on the signal-to-noise ratio of the meteorological features. The activated feature vector is input into the subsequent hidden layer for nonlinear mapping, and a linear transformation is used in the output layer to map the feature dimension to the prediction duration dimension to obtain the dynamic load prediction result.

Images