Energy storage capacity prediction method based on meteorological parameters
By constructing a graph-structured network model based on meteorological parameters to predict energy storage capacity, this method captures the nonlinear, temporal, and causal relationships between meteorological parameters and energy storage capacity. This solves the problem of neglecting the impact of meteorological changes in existing technologies and achieves high-precision energy storage capacity prediction and system optimization.
Patent Information
- Application Number
- CN202511682036.7
- Authority / Receiving Office
- CN · China
- Patent Type
- Applications(China)
- Current Assignee / Owner
- Filing Date
- 2025-11-17
- Publication Date
- 2026-02-10
AI Technical Summary
Existing energy storage capacity forecasting methods neglect the nonlinear coupling effect of meteorological changes on renewable energy generation, making it difficult to accurately predict short-term energy storage demand at high spatiotemporal resolution. This is especially true in small microgrids or remote areas, leading to redundant configurations or capacity shortages in energy storage systems, which reduces system economy and response performance.
The method for predicting energy storage capacity based on meteorological parameters acquires multi-source meteorological data, constructs a meteorological time series matrix, performs perturbation enhancement processing, establishes a graph structure network model, extracts the dynamic dependency path between meteorological perturbation and changes in energy storage capacity, uses a lightweight convolutional time series prediction model for prediction, and optimizes the window parameters in the perturbation feature tensor generation process by combining prediction errors.
It significantly improves the model's ability to perceive short-term sudden disturbances and its capacity prediction accuracy, enhances the model's generalization ability and convergence speed in different scenarios, and is suitable for real-time scheduling and edge computing, meeting high reliability requirements.
Smart Images

Figure CN121502326A_ABST
Abstract
Description
Technical Field
[0001] This invention relates to the fields of time series modeling and graph neural network technology, specifically to a method for predicting energy storage capacity based on meteorological parameters. Background Technology
[0002] With the widespread integration of new energy sources and the continuous deployment of distributed energy storage systems, the power system's flexible regulation capabilities are facing unprecedented challenges. Especially against the backdrop of a significant increase in the proportion of renewable energy sources such as wind and solar power, energy storage devices are playing an increasingly important role in peak shaving and valley filling, system frequency regulation, and emergency power supply.
[0003] In existing technologies, most energy storage capacity prediction methods rely on historical load data and equipment operating parameters for modeling, neglecting the nonlinear coupling effect of meteorological changes on renewable energy generation. This makes it difficult to accurately predict short-term energy storage demand at high spatiotemporal resolution scales. Especially in small microgrids or remote areas, the limitations of data acquisition accuracy, sparse spatial distribution, and random fluctuations in equipment power lead to frequent redundant configurations or capacity shortages in energy storage systems, reducing the overall economic efficiency and response performance of the system.
[0004] Furthermore, traditional forecasting methods rely heavily on manual parameter selection and lack the ability to deeply explore the dynamic correlations between multi-source heterogeneous data driven by meteorological conditions. Especially in complex terrain areas, local meteorological parameters (such as wind speed shear, solar radiation disturbances, and abnormal surface heat flux) may induce drastic fluctuations in energy storage load, making it difficult for existing methods to extract generalizable capacity forecasting features from them. Summary of the Invention
[0005] The purpose of this invention is to provide a method for predicting energy storage capacity based on meteorological parameters, in order to address the shortcomings of the prior art.
[0006] To achieve the above objectives, the present invention provides the following technical solution: a method for predicting energy storage capacity based on meteorological parameters, comprising: Acquire multi-source meteorological parameter data within the service area of the energy storage system and construct the original meteorological time series matrix; The meteorological time series matrix is subjected to perturbation enhancement processing, and a meteorological perturbation feature tensor is generated using a dynamic weighting strategy based on local fluctuation windows; The meteorological disturbance feature tensor is mapped to a higher-order feature space to construct a multi-scale capacity response sensitive factor set and extract significant influencing factors under the current meteorological state. Based on the set of capacity response sensitive factors, a graph structure network model containing time dependencies is established, and historical energy storage load sequences are embedded in the graph node attributes. The topological adjacency weights of each node in the graph network are adaptively adjusted to construct a causal link graph driven by meteorology, and to extract the dynamic dependency path between meteorological disturbances and changes in energy storage capacity. The dynamic dependency path is used to generate a predicted feature vector, which is then input into a lightweight convolutional time series prediction model to output the energy storage capacity prediction result for the future target time period. Based on the prediction results, and combined with the backpropagation results of the prediction error, the window parameters in the perturbation feature tensor generation process are dynamically optimized.
[0007] Preferably, the multi-source meteorological parameter data includes wind speed, solar radiation, air humidity, surface temperature, and air pressure change rate.
[0008] Preferably, the acquired meteorological parameter data is synchronously calibrated according to timestamps, and a time-series interpolation algorithm is used to fill in any missing data points; Spatial mapping of sensor nodes and regional labeling of local microclimate zones by combining satellite remote sensing data enable geographic regional aggregation of meteorological data. Based on the constructed minute-level time step, the original meteorological time series matrix is generated using a sliding window mechanism.
[0009] Preferably, the meteorological time series matrix is subjected to perturbation enhancement processing, including: The meteorological time series matrix is divided into multiple local fluctuation windows of fixed length according to the time dimension, and each window contains a preset number of time steps; For each local fluctuation window, the first and second derivatives of the meteorological parameter sequence are calculated to measure the fluctuation velocity and fluctuation acceleration within the window, which are used as dynamic disturbance sensitivity indicators. Based on the disturbance sensitivity index, a dynamic weighting coefficient is constructed using an exponential decay function; The weighted perturbation features of each window are concatenated into a three-dimensional tensor structure to form a meteorological perturbation feature tensor, in which the tensor dimensions include the time window sequence, the meteorological parameter dimension, and the perturbation intensity weight dimension.
[0010] Preferably, the construction of the multi-scale capacity response sensitivity factor set includes: A multi-scale one-dimensional convolutional neural network is used to extract features of meteorological disturbance feature tensors along the time dimension; Batch normalization and ReLU activation function are applied to the feature map output by each convolutional channel to enhance the non-linear expressive power of the features; Feature maps extracted from multiple scales are merged into a unified high-order feature tensor through a stitching operation. By calculating the attention weights between each meteorological parameter channel and historical capacity changes, parameter combinations that have a responsive effect on energy storage capacity are selected. The significant parameter features after weighted fusion are output as a set of sensitive factors for multi-scale capacity response, where each factor corresponds to a high-weight meteorological feature.
[0011] Preferably, establishing a graph-structured network model that includes time dependencies includes: The capacity response sensitivity factor set is divided into several time segments according to time order, and each time segment is used as a node in the graph structure to form a time-ordered node set. In the graph structure, directed edge connections between nodes are established according to a preset time sliding window length, with the direction of the edges pointing from the historical moment to the future moment; The historical energy storage load value corresponding to each node is embedded in the node attribute and jointly encoded with the capacity response sensitivity factor to form a composite node feature representation that includes disturbance information and load status. Graph convolutional networks or graph attention networks are constructed on the graph structure to propagate and fuse perturbation information between time series nodes, and output graph embedding feature vectors.
[0012] Preferably, the adaptive adjustment of the topological adjacency weights of each node in the graph network includes: Construct the adjacency matrix of the initial time series graph, and establish the adjacency relationship based on a fixed time window sliding mechanism; Calculate the change in the influence of each edge in the prediction result, and dynamically adjust the weight of the corresponding edge in the adjacency matrix accordingly; A causal reasoning algorithm is used to score the causal strength of the edge weights between meteorological disturbance nodes and capacity fluctuation nodes, and to screen out disturbance paths with causal influence. Edges with scores above a set threshold are retained and their weights are strengthened to construct a causal link graph driven by weather conditions. The dynamic dependency path between weather disturbances and changes in energy storage capacity is then extracted from the graph structure.
[0013] Preferably, generating the predicted feature vector using the dynamic dependency path includes: The node embedding vectors corresponding to dynamic dependency paths with significant causal strength in the causal link graph are stacked in chronological order to construct a path-aware feature sequence, and the edge weights between nodes are fused as the time weight coefficients of the path features. The path-aware feature sequence is subjected to dimensionality compression and normalization, and a one-dimensional convolutional neural network is used to extract local time-dependent features. By introducing skip connection structures and residual units into the output layer of the convolutional network, perturbation response features at different scales are fused. The final extracted feature vector is input into the fully connected prediction layer, which outputs the predicted energy storage capacity for the target time period.
[0014] Preferably, the window parameters in the process of dynamically optimizing the perturbation feature tensor generation include: During training, the prediction error between the predicted output and the actual capacity data in each round is recorded, and the error is used as a feedback term in the loss function to participate in the backpropagation process. The length of the local fluctuation window, the sliding step size, and the weight parameters of the perturbation index are set as learnable parameters and jointly optimized through the gradient descent mechanism during training. Construct a differentiable mapping relationship between window parameters and the output of the perturbation tensor to achieve automatic adjustment and adaptive evolution of window granularity; When the window parameters change beyond a preset threshold range, update the time step division method of the perturbation tensor.
[0015] Preferred time step division methods for updating the perturbation tensor include: Set the dynamic adjustment threshold range for window length and sliding step size. When the update magnitude of any parameter in consecutive training rounds exceeds its corresponding threshold, the perturbation tensor partitioning and reconstruction mechanism is triggered. In the partitioning and reconstruction mechanism, the length L of the time window and the sliding step S are adjusted according to the current prediction error gradient direction. The updated L and S are applied to the time axis partitioning of the original meteorological time series matrix to regenerate the local fluctuation window sequence, and a new perturbation feature tensor is calculated accordingly.
[0016] The technical effects and advantages provided by the present invention in the above technical solution are as follows: 1. This invention, by introducing a dynamic feature extraction mechanism based on meteorological disturbances and a graph neural network modeling structure, can effectively capture the nonlinear, temporal, and causal relationships between meteorological parameters and energy storage capacity. Compared with traditional methods that rely on static features or linear models, this invention can construct a causal link graph with physical interpretability and predictive capabilities, significantly improving the model's ability to perceive short-term sudden disturbances and the accuracy of capacity prediction, making it particularly suitable for distributed energy storage systems under variable meteorological environments.
[0017] 2. This invention innovatively introduces the perturbation window parameter as a learnable variable and establishes a prediction error-driven dynamic optimization mechanism. This allows the construction method of the perturbation feature tensor to adaptively adjust during model training, thereby achieving deep coupling between data-driven and structural optimization. This mechanism improves the model's generalization ability and convergence speed in different scenarios, possessing advantages such as lightweight structure, flexible deployment, and high prediction accuracy. It is particularly suitable for high-reliability scenarios such as real-time scheduling, edge computing, and dynamic capacity control. Attached Figure Description
[0018] To more clearly illustrate the technical solutions in the embodiments of this application or the prior art, the drawings used in the embodiments will be briefly introduced below. Obviously, the drawings described below are only some embodiments recorded in this invention. For those skilled in the art, other drawings can be obtained based on these drawings.
[0019] Figure 1 This is a flowchart of the method of the present invention. Detailed Implementation
[0020] To make the objectives, technical solutions, and advantages of the embodiments of the present invention clearer, the technical solutions of the embodiments of the present invention will be clearly and completely described below with reference to the accompanying drawings. Obviously, the described embodiments are only some embodiments of the present invention, not all embodiments. Based on the embodiments of the present invention, all other embodiments obtained by those skilled in the art without creative effort are within the scope of protection of the present invention.
[0021] For examples, please refer to Figure 1 As shown in this embodiment, the energy storage capacity prediction method based on meteorological parameters includes: Acquire multi-source meteorological parameter data within the service area of the energy storage system and construct the original meteorological time series matrix; The meteorological time series matrix is subjected to perturbation enhancement processing, and a meteorological perturbation feature tensor is generated using a dynamic weighting strategy based on local fluctuation windows; The meteorological disturbance feature tensor is mapped to a higher-order feature space to construct a multi-scale capacity response sensitive factor set and extract significant influencing factors under the current meteorological state. Based on the set of capacity response sensitive factors, a graph structure network model containing time dependencies is established, and historical energy storage load sequences are embedded in the graph node attributes. The topological adjacency weights of each node in the graph network are adaptively adjusted to construct a causal link graph driven by meteorology, and to extract the dynamic dependency path between meteorological disturbances and changes in energy storage capacity. The dynamic dependency path is used to generate a predicted feature vector, which is then input into a lightweight convolutional time series prediction model to output the energy storage capacity prediction result for the future target time period. Based on the prediction results, and combined with the backpropagation results of the prediction error, the window parameters in the perturbation feature tensor generation process are dynamically optimized.
[0022] Multiple edge weather sensing nodes are deployed within the coverage area of the energy storage system. These nodes are built based on MEMS micro weather modules and feature low power consumption, high sensitivity, and configurable data reporting frequency. Each sensing node includes at least an anemometer, a solar radiation sensor, a temperature and humidity module, a surface infrared thermal probe, and a barometric pressure sensor, and can simultaneously collect the following parameters: Horizontal wind speed (unit: meters per second); Total solar radiation at ground level (unit: watts per square meter); Relative humidity (%) Earth's surface temperature (unit: degrees Celsius); Rate of change of air pressure (unit: hectopascals / minute).
[0023] Each sensor node collects data with a time resolution of 1 minute and connects to the edge data aggregation unit via LoRa or NB-IoT protocols to ensure low power consumption, wide range, and high density deployment.
[0024] Due to the heterogeneous distribution of multiple nodes, data reporting delays or packet loss may occur. To standardize the data input format, the system introduces a timestamp alignment mechanism on the aggregation unit side and employs the following process for data preprocessing: Based on the node's built-in real-time clock (RTC), a collection timestamp is appended to the data packets; For data with missing time periods, cubic spline interpolation is used to fit and complete the data points. The excessive noise values are smoothed using a median filter to eliminate abrupt changes.
[0025] After the timelines of all nodes are synchronized, a structured meteorological data table with a step size of one minute is formed.
[0026] To construct a meteorological input feature matrix with spatial awareness capabilities, this invention introduces a Geographic Information System (GIS) module. This module spatially maps the geographic location information of the sensor nodes, as follows: The GPS coordinates of each sensor are calibrated to form a two-dimensional geographic coordinate set; Use remote sensing data (such as MODIS or Sentinel-2) to obtain the topography, vegetation and surface thermal characteristics of the area; A K-Means-based geographic partitioning algorithm is used to divide the monitoring area into several microclimate units, with each unit corresponding to a set of representative nodes.
[0027] Within each climate unit, a weighted average strategy is used to fuse nodal data, where the weighting factor is determined by nodal signal strength, historical stability, and spatial density.
[0028] After time alignment and spatial mapping are completed, a sliding window mechanism is used to organize the data into a continuous input sequence. Specifically: Set the window length to 60 (i.e., 60 consecutive minutes) and the step size to 5 minutes; For each time window, extract all meteorological parameters for the current time and the previous 59 minutes; Each window is treated as an input sample, and they are stacked in chronological order to form a three-dimensional matrix with dimensions of "number of samples × number of time steps × types of parameters".
[0029] For example, if five parameters—wind speed, radiation, humidity, temperature, and air pressure—are collected within a one-hour window, each sample matrix will have dimensions of 60 rows × 5 columns. This matrix serves as the input to the subsequent perturbation enhancement module, possessing high temporal accuracy and multi-dimensional environmental perception capabilities.
[0030] For the original meteorological time series matrix, the matrix has a dimension of "M×N", where "M" represents the number of consecutive time steps (e.g., sampling per minute, M can be 1440 to represent one day), and N represents the meteorological parameter dimensions (e.g., wind speed, humidity, solar radiation, temperature, air pressure, etc., usually 5 dimensions).
[0031] To capture short-term fluctuation trends, the matrix is divided into multiple fixed-length local fluctuation windows along the time dimension. Each window is set to a length L (e.g., L=30 means each window is 30 minutes), allowing it to slide across the entire time series to generate... There are multiple windows. Windows can slide in an overlapping manner (e.g., with a step size of 10 minutes) or be divided in a non-overlapping manner, depending on the model's time series feature requirements.
[0032] Each window after partitioning is represented as a submatrix W_i with dimensions “L×N”, where i is the window index, L is the window time length, and N is the number of meteorological parameters.
[0033] Within each local fluctuation window, for each meteorological parameter sequence, its first and second derivatives are calculated as indicators of disturbance velocity and disturbance acceleration to quantify the degree of short-term fluctuation. For any column (i.e., a certain meteorological parameter) in window W_i, let its time series sequence within that window be x_t, where t∈[1,L], then: The first derivative d1_t is calculated using the difference approximation and is expressed as the change between the current time step and the previous time step, i.e. This reflects the instantaneous rate of change of meteorological quantities; The second derivative d²_t represents the rate of change of the rate of change, and is calculated as follows: This reflects the accelerating or decelerating trend of weather changes.
[0034] After calculating the first and second derivative sequences of each parameter, a disturbance index vector P_i can be constructed within each time window, containing the fluctuation sensitivity measure of all parameters within the current window, with a dimension of "L×2N".
[0035] Based on the aforementioned disturbance index vector P_i, to highlight the role of the severely disturbed segment in subsequent feature modeling, a dynamic weighting coefficient matrix W_i_weight is constructed using an exponential decay function based on the disturbance amplitude. Each element of this weighting matrix is calculated according to the following rules: Calculate the absolute values of the first derivative d1_t and the second derivative d2_t, respectively; Using an exponential weighting function , where α is a positive adjustment factor (e.g., α=0.8) used to adjust the intensity of the disturbance response; The resulting weight values are within the range (0,1). The greater the disturbance, the closer the weight is to 1.
[0036] Ultimately, each window W_i corresponds to a disturbance intensity weight matrix W_i_weight with dimensions “L×N”, which corresponds to each time step and each meteorological parameter.
[0037] The original meteorological submatrix W_i in each fluctuation window is multiplied element-wise with the corresponding weight matrix W_i_weight to obtain the weighted window feature matrix W_i_enhanced. All weighted window feature matrices are concatenated in time series order to form a three-dimensional input tensor T_enhance, with dimensions of: number of windows × number of parameters × number of time steps (i.e., K × N × L, where K is the number of windows). This tensor is the meteorological disturbance feature tensor, which, while preserving the original meteorological information, explicitly introduces a dynamic weight adjustment mechanism for the disturbance intensity, possessing the following advantages: The explicit encoding of perturbation structures can be completed at the front end of the model, reducing the difficulty for the model to abstract perturbation features by relying on deep networks; Effectively enhances the model's response to short-term, abrupt meteorological disturbances (such as local storms, sudden temperature drops, and high radiation changes); It supports seamless connection with subsequent time series prediction modules (such as convolutional temporal networks and graph neural networks) to achieve an end-to-end prediction framework design.
[0038] This invention first processes the input meteorological disturbance feature tensor. This tensor is denoted as T_in, and its dimensions are "B×T×C", where: B indicates the sample batch size; T represents the time step (e.g., 60 represents 60 minutes of continuous data). C represents the number of meteorological parameter channels (such as wind speed, humidity, radiation, temperature, air pressure, etc., C=5).
[0039] To capture the potential impact of meteorological disturbances on energy storage capacity at different time scales, multiple one-dimensional convolutional neural networks with different kernel sizes are used for parallel feature extraction along the time dimension of T_in. Let the kernel sizes be 3, 5, and 7, representing the short-term, medium-term, and relatively long-term disturbance mode response windows, respectively.
[0040] Each scale convolutional channel uses the following configuration: Number of input channels: C; Number of output channels: F (e.g., F=32); The step size is 1, and the "same" strategy is used for padding to keep the time dimension unchanged.
[0041] After multi-scale convolution processing, three tensor outputs can be obtained, denoted as T_3, T_5, and T_7, each with a dimension of "B×T×F".
[0042] To improve feature representation capabilities and model training stability, the convolutional output feature maps at each scale are sequentially introduced as follows: The Batch Normalization layer is used to normalize the distribution of activation values within each batch, reducing internal covariate bias. The ReLU (Rectified Linear Unit) activation function is used to introduce nonlinearity, enabling the model to express more complex perturbation response relationships.
[0043] After processing, enhanced scale feature maps T_3', T_5', and T_7' are obtained, which still retain the dimensions "B×T×F".
[0044] By concatenating the feature maps at different scales along the channel dimension (i.e., the third dimension), a unified high-order feature tensor T_fusion with dimensions “B×T×(3F)” is obtained.
[0045] To further extract meteorological disturbance features that have a significant response effect on energy storage capacity, a channel-wise attention module is designed to calculate the importance weight of each meteorological channel in the current forecasting task.
[0046] Attention mechanisms include the following steps: First, the historical capacity change sequence Y_history (with dimensions "B×T") and T_fusion are subjected to global average pooling (GAP) to compress it into vector form; Channel mapping is achieved through a two-layer fully connected network. The first layer outputs a dimension of "(3F) / 4" and uses ReLU as the activation function. The second layer outputs an attention weight vector α of length "3F" and uses Sigmoid as the activation function, which is used to map to the interval (0,1). The attention weight vector is multiplied element-wise with the channel dimension of T_fusion to achieve the weighted fusion operation.
[0047] This mechanism ensures that, at the meteorological channel dimension, disturbance features that are highly correlated with historical capacity changes will receive higher weights and thus be retained in the final set of sensitive factors.
[0048] The higher-order feature tensor after attention-weighted fusion is denoted as T_weighted, and its dimension remains "B×T×(3F)". In order to generate a set of capacity response sensitive factors with a well-defined structure, max pooling or average pooling is performed on T_weighted along the time dimension to reduce its dimension to a two-dimensional vector of "B×(3F)".
[0049] In this vector, each element represents a high-weighted perturbation feature that responds to energy storage capacity fluctuations under different scales and parameter combinations, denoted as the capacity response sensitivity factor set V_response, where: Each factor corresponds to a specific combination of convolutional kernel size and channel number; The top few significant factors can be selected based on a threshold λ (e.g., λ=0.8) (e.g., Top-k strategy). The selection criteria are a comprehensive index of attention weight and perturbation intensity, which has physical interpretability.
[0050] The capacity response sensitivity factor set output from the previous module is divided chronologically to form a set of equally spaced time segments. Let the entire prediction window be T_total minutes, and the division interval be Δt (e.g., Δt = 5 minutes). Then the factor sequence can be divided into N = T_total / Δt time segments.
[0051] Each time segment corresponds to a time node, denoted as v_i, where i∈{1,2,...,N}. This node represents the capacity response characteristic state within the current time window. The node set V={v_1,v_2,...,v_N} constitutes the basic structure of the graph.
[0052] For each node v_i, its initial feature vector is represented by the mean of the capacity response sensitivity factor set within the corresponding time period, denoted as f_i, with dimension D, where D represents the dimension of the higher-order perturbation factor (e.g., 96 dimensions, i.e., the feature vector obtained after fusing multiple convolutional channels).
[0053] To characterize the evolution path and influence relationship of the perturbation factor in the time dimension, a directed edge connection relationship is established between nodes based on the time sliding window mechanism.
[0054] The specific method is as follows: Set the sliding window span to W (for example, W = 3), indicating that each node is connected to at most the previous W historical nodes; For any time node v_i (i ≥ 2), establish directed edges between it and the previous W nodes v_(i - 1), v_(i - 2),..., v_(i - W); The direction of the edge always points from the historical node to the current node, that is, the edge set E = {(v_j → v_i)|i - W ≤ j < i}, satisfying j < i; Each edge can be attached with an edge weight, which can be initially set to 1 uniformly, or attenuated according to the time interval. For example, set the edge weight to , where γ is the edge weight attenuation factor (such as γ = 0.9).
[0055] The graph G = (V, E) constructed in this way is a directed graph with a clear time - dependent direction, which can characterize the dynamic evolution structure of the perturbation factor over time.
[0056] To jointly model the perturbation information and the operating state of the energy storage system, the historical energy storage load value is introduced into the feature representation of each graph node to achieve feature - level fusion.
[0057] The specific operation is as follows: For each time node v_i, obtain the actual load value p_i (unit: kWh) of the energy storage system during its corresponding time period, or use the average value of the load values during this period; Concatenate p_i with the capacity response sensitivity factor vector f_i to form an extended feature vector x_i = [f_i; p_i], with a dimension of D + 1; Perform this operation on all nodes to obtain the composite node feature matrix X = {x_1, x_2,..., x_N} that combines the perturbation features and the load state.
[0058] This joint encoding method can enhance the physical meaning interpretability of the graph model during the node aggregation process and improve the prediction accuracy of the model.
[0059] After completing the construction of the graph structure and the initialization of node features, a graph neural network is constructed on the graph structure for node information propagation and aggregation. According to specific task requirements, one of the following two mainstream structures can be selected: Based on the adjacency matrix and Laplace transform, the graph convolutional network realizes the smooth propagation of graph signals. Its core propagation formula is: ; where: Let A_ij represent the representation of node j in layer l; A_ij is the connection relationship in the adjacency matrix (which may include edge weights); W is the trainable weight matrix; and σ is the activation function, usually ReLU. This approach is suitable for scenarios with relatively balanced structures and no strong attention bias.
[0060] Graph attention networks introduce learnable attention coefficients to achieve differentiated modeling of the degree of influence on neighboring nodes. Its core process includes: For each node pair (i,j), calculate the attention score. Update the node representation using a weighted summation method: This method can more accurately identify the historical disturbance nodes that have the greatest impact on the current node, and is suitable for scenarios with highly nonlinear and multi-scale disturbance propagation.
[0061] After multiple layers of propagation through the graph neural network, the final embedding vector z_i at each time point can be obtained, with dimension D', i.e., z_i=GNN(x_i,G), where GNN represents the aforementioned GCN or GAT structure.
[0062] By concatenating or pooling the embedded features z_i of all nodes (such as time-dimensional average pooling), we can obtain the representation vector Z_global of the overall perturbation state of the current graph structure, which can be used as the input for subsequent capacity prediction modules (such as time series prediction and regression networks).
[0063] In a graph neural network structure, nodes represent time-divided segments of capacity response sensitive factors, and the edge connections of the graph structure determine the direction and intensity of information propagation over time.
[0064] Therefore, the initial adjacency matrix A0 is constructed first, and the steps are as follows: The time series is divided into N equal-length segments, each segment representing a node; Define a fixed-length time sliding window W (e.g., W=3), representing the influence each node can receive from its previous W time segments; In the adjacency matrix, for each node i and its W preceding nodes j ( )set up , indicating that there exists a directed edge from j to i; Alternatively, the initial edge weights can be set to exponential decay: A0[i][j] = , where γ∈(0,1) is the time decay coefficient, for example γ=0.9.
[0065] The adjacency matrix A0 described above forms the basic time-series graph topology, which is used for subsequent graph neural network propagation.
[0066] To ensure that the edge weights in the graph structure reflect the true contribution of perturbation features to the capacity prediction results, an edge weight impact evaluation mechanism based on backpropagation of prediction errors is introduced to dynamically adjust the adjacency matrix.
[0067] The specific steps are as follows: During the training phase of the graph neural network, the increase or decrease trend of the model prediction error L (such as mean squared error Loss) under the change of edge weights before and after each round is recorded. For each edge (j→i), observe the influence of its existence in the current graph structure on the prediction error L, ΔL_ij. If L increases significantly after temporarily removing the edge, it indicates that the edge is important for prediction and its weight should be increased. Update the edge weights to: A[i][j]←A[i][j]+η×ΔL_ij, where η is the learning rate hyperparameter (e.g., η=0.05); After all edge weights are updated, normalization is performed to ensure that the sum of the weights of the incoming edges of each node is 1.
[0068] The above method enables the edge weights in the graph structure to adaptively reflect the relative strength of the influence of each node in the actual prediction.
[0069] To further extract perturbation paths with real physical meaning from the adjusted graph structure, a causal reasoning mechanism is introduced to determine the causality of edge weights. This invention employs the following two methods: For each pair of nodes (j→i), construct the following two models: Model 1: Predict the current value of node i using its historical capacity data; Model 2: Predict the current value using historical data of node i and historical perturbation features of node j; If Model 2 is statistically significant better than Model 1, and the error decreases by more than the set threshold ε (e.g., ε=10%), then it is considered that there is a Granger causal relationship between j and i. Assign a higher causal score C_ij∈[0,1] to all edges that satisfy the causal condition.
[0070] In the graph attention network structure, the attention coefficient α_ij for each edge is calculated; Normalize and threshold the value of α_ij. If α_ij > τ (threshold, e.g., τ = 0.6), then retain the edge. Edges with high attention scores are considered key transmission channels in the perturbation information flow path and are included in the causal link graph.
[0071] Based on the causal scoring results, the final causal link graph G_causal is constructed: For each edge (j→i), if its causal score C_ij is higher than the set threshold θ (e.g., θ=0.7), then the edge is retained and the edge weight A[i][j]←A[i][j]×(1+C_ij); Edges below a certain threshold are masked or weakened to reduce their interference with the propagation of the graph structure. The final graph structure G_causal contains only perturbation paths with significant causal relationships, and the edge weights explicitly reflect the perturbation strength and capacity response.
[0072] This invention first filters out paths that are significantly causally related to changes in energy storage capacity from the causal link graph, denoted as P={p1,p2,...,pk}, where each path pi represents a time-ordered sequence of multiple disturbance nodes, representing the propagation process of disturbance from occurrence to affecting capacity changes.
[0073] For each path pi, the node representation vector it contains (obtained by graph neural network propagation) is denoted as: Zpi={z1,z2,...,zn}, where zt represents the embedding representation of the node at time step t in the path, with dimension d; Meanwhile, the edge weight vector in this path is Wpi={w1→2,w2→3,...,w{n-1}→n}, which is used to measure the transmission strength of the disturbance between different nodes.
[0074] Stack the node embedding vectors in chronological order to form the path feature sequence Tp_i∈ To enhance the model's ability to perceive the intensity of path perturbations, a path weighting mechanism is introduced. This involves multiplying the embedding z_t of each node by its incoming edge weight w_{t-1}→t to obtain a weighted embedding representation. The final path-aware prediction feature sequence is obtained as follows: This constitutes the prediction vector of the input model.
[0075] Features from multiple paths can be concatenated or averaged to form a single input feature tensor T_in with dimensions B×n×d, where B is the batch size.
[0076] To adapt to the aforementioned structured time series inputs, this invention designs a Lightweight Convolutional Temporal Network (LCTN) to achieve high-efficiency, low-complexity short-term capacity prediction. The network structure has the following characteristics: One-dimensional convolution (Conv1D) is used to model the time dimension and extract the temporal dependency features of the perturbation response; The kernel size is set to a small window between 3 and 5 (e.g., 3, 5, 7) to cover the range of local perturbation changes; The network depth is limited to 3 layers or less to avoid over-parameterization and improve deployment efficiency; Each convolutional layer is followed by batch normalization (BatchNorm) and ReLU activation function to improve feature stability and non-linear expressive power.
[0077] The convolutional network layer structure is as follows: Layer 1: Conv1D(kernel_size=3, filters=64) → BatchNorm → ReLU; Layer 2: Conv1D(kernel_size=5, filters=32)→BatchNorm→ReLU; Layer 3: Conv1D(kernel_size=3,filters=16)→BatchNorm→ReLU.
[0078] The network described above maps the input sequence T_in to an intermediate feature tensor of dimension B×n×f, where f is the number of output channels.
[0079] To further enhance the model's ability to respond to multi-scale perturbation trends, residual connections and skip connections are introduced into the convolutional network structure, enabling the network to retain the original perturbation trend while extracting local features.
[0080] Specifically as follows: A residual path is introduced between every two convolutional layers, calculated as follows: ; where F_l(T) represents the output after the l-th convolution, and T is the input feature of this layer.
[0081] Channel concatenation is performed on the feature maps output by different convolution kernel sizes, and 1×1 convolution is used for compression and fusion to obtain a comprehensive feature representation at a uniform scale.
[0082] This mechanism allows the network to incorporate perturbation patterns at different time granularities while maintaining a lightweight structure, thereby improving the model's predictive robustness in variable weather environments.
[0083] In the final stage of the model, the feature tensor after convolution and fusion is fed into a fully connected prediction layer to output the capacity prediction value.
[0084] Let the final extracted feature vector be V∈ Input to the fully connected (Dense Layer): Where W is the weight matrix and b is the bias term. It predicts capacity values (in kWh); it can be expanded to a multi-step prediction form with an output length of T_pred (e.g., predicting 6-12 capacity points in the next 30 minutes to 1 hour); it supports a sliding window prediction mechanism to achieve continuous prediction capability. The target prediction time period ΔT can be configured to 5 minutes, 15 minutes, 30 minutes, or 60 minutes, etc., to meet the granularity requirements of different energy storage systems.
[0085] During the training process of the energy storage capacity prediction model, the system continuously records the model's prediction output. The error between the actual capacity data y_true and the predicted loss is calculated using a common loss function (such as mean squared error, MSE). , where i represents the predicted sample number and N is the total number of samples.
[0086] This loss function serves as the core objective of training optimization and participates in the backpropagation process of the entire model. In traditional models, the gradient only updates the weight parameters within the neural network. However, in this invention, the construction process of the perturbation feature tensor is also considered part of the model learning, and the window parameter participates in gradient propagation as a learnable parameter.
[0087] The construction of the perturbation tensor depends on the partitioning method of the original meteorological time series matrix, specifically including: Window length L: Represents the number of time steps contained in each local disturbance window (unit: minutes or step length). Sliding step size S: represents the span of the window sliding on the time axis, affecting the coverage density of the samples; Perturbation calculation weight α: affects the proportion of the first and second derivatives in the calculation of the perturbation sensitivity index.
[0088] This invention sets the three parameters (L, S, α) as learnable variables and optimizes them jointly during neural network training using a gradient descent mechanism. This requires constructing a differentiable path that starts from these parameters and ultimately affects the model's predicted output, ensuring that error gradients can be effectively propagated.
[0089] For example, the derivative of each time window in the perturbation tensor can be expressed as: , , Where α is the learnable coefficient of the perturbation index.
[0090] The window length L and step size S control the way the tensor is partitioned, and in practice, it is modeled as a differentiable variable using discretization approximation functions (such as Gumbel-softmax or Sigmoid-based approximation).
[0091] To ensure that the window parameters can be effectively updated during training, the system needs to construct the following mapping path: Input parameters: L, S, α; Generation process: sliding window partitioning → derivative calculation → perturbation index weighting → tensor generation; Output target: perturbation feature tensor T_disturb∈ Where D is the dimension of the meteorological parameter; Loss feedback: The error in the model prediction output is backpropagated to T_disturb, and L, S, and α are updated through the differentiable mapping chain.
[0092] This process requires that the operations in the tensor partitioning and perturbation processes be differentiable or approximately differentiable, so that automatic differentiation engines (such as TensorFlow or PyTorch) can perform backpropagation.
[0093] During consecutive training rounds, the system monitors the changes in the parameters of the monitoring window, setting the following change threshold range: ΔL_max: The maximum number of steps allowed for the window length to change (e.g., ΔL_max=5); ΔS_max: The maximum number of steps allowed for the sliding step size (e.g., ΔS_max=3). Δα_max: Threshold for perturbation weight change (e.g., Δα_max=0.1).
[0094] When any of the following conditions are met: The change in L compared to the previous update exceeds ΔL_max; The current value of S has changed more than ΔS_max compared to the previous update; The current change in α exceeds Δα_max; This triggers the perturbation tensor partitioning and reconstruction mechanism, which performs the following operations: Parameter update: Based on the current error descent direction and gradient information, fine-tune the integer values of L and S (e.g., adjust L from 30 steps to 35 steps, and S from 10 steps to 5 steps) to enhance the perturbation expression density. Window re-division: The original meteorological time series matrix is re-sliding windowed with new parameters to generate a new set of local fluctuation windows; Feature tensor reconstruction: Based on the new partition, recalculate the derivative and perturbation sensitivity to construct a new perturbation feature tensor; Performance monitoring: Continue training for several rounds. If the prediction error decreases significantly and the convergence trend is good, then solidify the new parameters.
[0095] The above description is merely a specific embodiment of this application, but the scope of protection of this application is not limited thereto. Any changes or substitutions that can be easily conceived by those skilled in the art within the scope of the technology disclosed in this application should be included within the scope of protection of this application.
Claims
1. A method for predicting energy storage capacity based on meteorological parameters, characterized in that: include: Acquire multi-source meteorological parameter data within the service area of the energy storage system and construct the original meteorological time series matrix; The meteorological time series matrix is subjected to perturbation enhancement processing, and a meteorological perturbation feature tensor is generated using a dynamic weighting strategy based on local fluctuation windows; The meteorological disturbance feature tensor is mapped to a higher-order feature space to construct a multi-scale capacity response sensitive factor set and extract significant influencing factors under the current meteorological state. Based on the set of capacity response sensitive factors, a graph structure network model containing time dependencies is established, and historical energy storage load sequences are embedded in the graph node attributes. The topological adjacency weights of each node in the graph network are adaptively adjusted to construct a causal link graph driven by meteorology, and to extract the dynamic dependency path between meteorological disturbances and changes in energy storage capacity. The dynamic dependency path is used to generate a predicted feature vector, which is then input into a lightweight convolutional time series prediction model to output the energy storage capacity prediction result for the future target time period. Based on the prediction results, and combined with the backpropagation results of the prediction error, the window parameters in the perturbation feature tensor generation process are dynamically optimized.
2. The energy storage capacity prediction method based on meteorological parameters according to claim 1, characterized in that: The multi-source meteorological parameter data includes wind speed, solar radiation, air humidity, surface temperature, and air pressure change rate.
3. The energy storage capacity prediction method based on meteorological parameters according to claim 2, characterized in that: The acquired meteorological parameter data are synchronously calibrated according to timestamps, and time-series interpolation algorithms are used to fill in any missing data points; Spatial mapping of sensor nodes and regional labeling of local microclimate zones by combining satellite remote sensing data enable geographic regional aggregation of meteorological data. Based on the constructed minute-level time step, the original meteorological time series matrix is generated using a sliding window mechanism.
4. The energy storage capacity prediction method based on meteorological parameters according to claim 1, characterized in that: The meteorological time series matrix is subjected to perturbation enhancement processing, including: The meteorological time series matrix is divided into multiple local fluctuation windows of fixed length according to the time dimension, and each window contains a preset number of time steps; For each local fluctuation window, the first and second derivatives of the meteorological parameter sequence are calculated to measure the fluctuation velocity and fluctuation acceleration within the window, which are used as dynamic disturbance sensitivity indicators. Based on the disturbance sensitivity index, a dynamic weighting coefficient is constructed using an exponential decay function; The weighted perturbation features of each window are concatenated into a three-dimensional tensor structure to form a meteorological perturbation feature tensor, in which the tensor dimensions include the time window sequence, the meteorological parameter dimension, and the perturbation intensity weight dimension.
5. The energy storage capacity prediction method based on meteorological parameters according to claim 1, characterized in that: The construction of the multi-scale capacity response sensitivity factor set includes: A multi-scale one-dimensional convolutional neural network is used to extract features of meteorological disturbance feature tensors along the time dimension; Batch normalization and ReLU activation function are applied to the feature map output by each convolutional channel to enhance the non-linear expressive power of the features; Feature maps extracted from multiple scales are merged into a unified high-order feature tensor through a stitching operation. By calculating the attention weights between each meteorological parameter channel and historical capacity changes, parameter combinations that have a responsive effect on energy storage capacity are selected. The significant parameter features after weighted fusion are output as a set of sensitive factors for multi-scale capacity response, where each factor corresponds to a high-weight meteorological feature.
6. The energy storage capacity prediction method based on meteorological parameters according to claim 1, characterized in that: The establishment of a graph-structured network model containing time dependencies includes: The capacity response sensitivity factor set is divided into several time segments according to time order, and each time segment is used as a node in the graph structure to form a time-ordered node set. In the graph structure, directed edge connections between nodes are established according to a preset time sliding window length, with the direction of the edges pointing from the historical moment to the future moment; The historical energy storage load value corresponding to each node is embedded in the node attribute and jointly encoded with the capacity response sensitivity factor to form a composite node feature representation that includes disturbance information and load status. Graph convolutional networks or graph attention networks are constructed on the graph structure to propagate and fuse perturbation information between time series nodes, and output graph embedding feature vectors.
7. The energy storage capacity prediction method based on meteorological parameters according to claim 1, characterized in that: The adaptive adjustment of the topological adjacency weights of each node in the graph network includes: Construct the adjacency matrix of the initial time series graph, and establish the adjacency relationship based on a fixed time window sliding mechanism; Calculate the change in the influence of each edge in the prediction result, and dynamically adjust the weight of the corresponding edge in the adjacency matrix accordingly; A causal reasoning algorithm is used to score the causal strength of the edge weights between meteorological disturbance nodes and capacity fluctuation nodes, and to screen out disturbance paths with causal influence. Edges with scores above a set threshold are retained and their weights are strengthened to construct a causal link graph driven by weather conditions. The dynamic dependency path between weather disturbances and changes in energy storage capacity is then extracted from the graph structure.
8. The energy storage capacity prediction method based on meteorological parameters according to claim 1, characterized in that: The step of generating the predicted feature vector using the dynamic dependency path includes: The node embedding vectors corresponding to dynamic dependency paths with significant causal strength in the causal link graph are stacked in chronological order to construct a path-aware feature sequence, and the edge weights between nodes are fused as the time weight coefficients of the path features. The path-aware feature sequence is subjected to dimensionality compression and normalization, and a one-dimensional convolutional neural network is used to extract local time-dependent features. By introducing skip connection structures and residual units into the output layer of the convolutional network, perturbation response features at different scales are fused. The final extracted feature vector is input into the fully connected prediction layer, which outputs the predicted energy storage capacity for the target time period.
9. The energy storage capacity prediction method based on meteorological parameters according to claim 1, characterized in that: The window parameters in the dynamic optimization perturbation feature tensor generation process include: During training, the prediction error between the predicted output and the actual capacity data in each round is recorded, and the error is used as a feedback term in the loss function to participate in the backpropagation process. The length of the local fluctuation window, the sliding step size, and the weight parameters of the perturbation index are set as learnable parameters and jointly optimized through the gradient descent mechanism during training. Construct a differentiable mapping relationship between window parameters and the output of the perturbation tensor to achieve automatic adjustment and adaptive evolution of window granularity; When the window parameters change beyond a preset threshold range, update the time step division method of the perturbation tensor.
10. The energy storage capacity prediction method based on meteorological parameters according to claim 9, characterized in that: The time step division method for updating the perturbation tensor includes: Set the dynamic adjustment threshold range for window length and sliding step size. When the update magnitude of any parameter in consecutive training rounds exceeds its corresponding threshold, the perturbation tensor partitioning and reconstruction mechanism is triggered. In the partitioning and reconstruction mechanism, the length L of the time window and the sliding step S are adjusted according to the current prediction error gradient direction. The updated L and S are applied to the time axis partitioning of the original meteorological time series matrix to regenerate the local fluctuation window sequence, and a new perturbation feature tensor is calculated accordingly.
Citation Information
Cited By
Mine hoist main shaft device fault detection method and system based on data driving
CN122065166A
Model iteration method and device, electronic equipment and storage medium
CN122347235A
