An interconnected transformer area distributed energy output and load prediction method, device and storage medium

Abstract

The application discloses an interconnected transformer area distributed energy output and load prediction method and device and a storage medium, and belongs to the technical field of power systems. The method first realizes distributed energy field group total power prediction, and then extends to transformer area load collaborative prediction to form a source-load integrated prediction system. The distributed energy field group total power prediction defines three types of nodes and constructs input samples, calculates dynamic space-time correlation tensors by using an improved normalized mutual information, and builds a hierarchical directed graph to explicitly model physical causal relationships. Based on the spatial domain GCN, a power prediction model is constructed, and cross-domain collaborative training is realized by combining the FedProx algorithm and differential privacy. The load prediction is realized by constructing a dynamic space-time information graph, and multi-step collaborative prediction is realized based on the TGCN. The application breaks through the limitations of traditional methods, accurately depicts dynamic space-time correlation, considers data privacy and prediction accuracy, improves the robustness and generalization ability of the model, and provides technical support for safe and efficient operation of interconnected transformer areas.

Description

Technical Field

[0001] This invention belongs to the field of distributed energy technology, and in particular relates to a method, device and storage medium for predicting the output and load of distributed energy in interconnected power distribution areas. Background Technology

[0002] With the continuous advancement of energy structure transformation and the construction of new power systems, distributed clean energy sources, represented by photovoltaics and wind power, are widely used and connected to the grid in distribution network areas due to their environmental friendliness and flexible deployment. To improve the reliability and economy of regional energy supply and overcome limitations in resource regulation, interconnecting multiple geographically adjacent or electrically connected distribution substations through technical or management means to form a networked and collaborative interconnected substation system has become an important trend in distribution network development. Interconnected systems enable cross-regional energy complementarity and flexible dispatch, which is of great significance for improving overall energy utilization efficiency.

[0003] However, while interconnecting distribution substations brings synergistic benefits, it also increases the complexity of system operation. On the one hand, the output of distributed energy resources is highly dependent on natural weather conditions, exhibiting intermittency, volatility, and randomness. On the other hand, the power load between interconnected substations displays strong spatiotemporal correlation and is highly nonlinear, influenced by multiple factors (such as weather, time of day, user behavior, and electricity prices). The dual uncertainties of distributed power output and load demand superimposed on each other pose challenges to voltage stability control, energy optimization allocation, and operational safety of interconnected substation systems. Therefore, achieving accurate and reliable prediction of future distributed energy output and load demand has become a key technological prerequisite for ensuring the safe and stable operation of interconnected substation systems and optimizing energy dispatch.

[0004] For energy output and load forecasting, traditional statistical methods such as time series analysis and regression analysis, which were used in the early stages, are ineffective in characterizing the complex nonlinear dynamic characteristics of output and load data and have been gradually phased out. Machine learning methods, represented by BP neural networks and support vector machines, have certain nonlinear fitting capabilities, but when processing time series data, they often ignore the inherent temporal correlation of the data and have inherent drawbacks such as slow convergence speed and susceptibility to getting trapped in local optima. Subsequently, deep learning models, represented by gated recurrent units and long short-term memory networks, are able to better uncover the deep temporal dependencies in sequence data. To further integrate spatiotemporal information, strategies such as hybrid convolutional neural networks and recurrent neural networks have emerged, making initial attempts at the joint extraction of spatiotemporal features.

[0005] Despite this, existing forecasting methods still have several key limitations when applied to the complex scenario of interconnected power distribution areas: First, most methods only model and forecast single, isolated power distribution areas, neglecting the dynamic spatiotemporal correlations and energy interactions formed between interconnected areas under the influence of factors such as electrical connections, geographical proximity, and similar meteorological conditions, resulting in incomplete forecast information. Second, while some methods using graph neural networks introduce spatial structure, they are usually based on static graphs constructed from fixed prior topologies, failing to adaptively characterize the dynamic spatiotemporal correlations between power distribution areas that evolve with operating status, meteorological propagation, and other factors. Third, existing methods mostly employ linear or shallow feature selection and fusion strategies, making it difficult to fully explore the complex nonlinear coupling relationships and interaction mechanisms between distributed energy output, diverse loads, and external influencing factors. Fourth, models typically use output and load as parallel inputs, failing to explicitly model the physical causal relationship that "resources determine output, and output and load jointly affect system state," affecting the interpretability of the model and lacking robustness in the face of data noise or anomalies.

[0006] Furthermore, interconnected regional systems often involve multiple different ownership or management entities in actual operation, resulting in a natural "data silo" problem. The traditional prediction paradigm of centrally collecting all data for training is prone to risks of user privacy and trade secret leakage, while most existing methods do not take into account data privacy protection needs, which greatly restricts cross-regional and cross-entity data collaboration and in-depth feature mining. Summary of the Invention

[0007] The purpose of this invention is to provide a method, device, and storage medium for predicting the output and load of distributed energy in interconnected power distribution areas. First, it achieves accurate prediction of the total power of the distributed energy field group in interconnected power distribution areas, and then extends the power prediction logic to the coordinated prediction of the load in the power distribution area, forming an integrated "source-load" prediction system, which meets the needs of safe, stable and efficient operation of interconnected power distribution areas.

[0008] To achieve the above objectives, the present invention is implemented using the following technical solution:

[0009] In a first aspect, the present invention provides a method for forecasting the output and load of distributed energy resources in interconnected power distribution areas, comprising:

[0010] Acquire the energy data and load data to be measured for each interconnected power distribution area; wherein, the energy data to be measured includes the power generation resource data and power generation data of the distributed energy nodes in each interconnected power distribution area; the load data to be measured includes the historical load sequence data and topology information of each interconnected power distribution area;

[0011] The energy data to be measured is input into a pre-trained energy output prediction model to obtain the total power prediction value of distributed energy nodes in each interconnected area. The energy output prediction model constructs a hierarchical directed graph based on the energy data to be measured and extracts dynamic spatiotemporal correlation features between distributed energy nodes. The hierarchical directed graph and the dynamic spatiotemporal correlation features are then fused to predict the total power. Each node in the hierarchical directed graph represents power generation resource data, power generation data, and the total power prediction value, respectively. The hierarchical directed graph contains the causal relationships between the nodes.

[0012] A load spatiotemporal information map is constructed based on the aforementioned topology information and historical load sequence data. This load spatiotemporal information map is then input into a pre-trained load prediction model to obtain load prediction values ​​for each interconnected transformer area at multiple future time points. The load prediction model extracts dynamic spatial features between nodes in the load spatiotemporal information map and extracts time-dependent features based on historical load sequence data. Multi-step load prediction is then performed by fusing the dynamic spatial features and time-dependent features. Each node in the load spatiotemporal information map represents the overall load of each interconnected transformer area.

[0013] The energy output prediction model and the load prediction model are both trained collaboratively using a hierarchical federated learning framework.

[0014] Optionally, the power generation resource data of each distributed energy node is defined as an energy node, the power generation data is defined as an energy power node, and the total power prediction of the interconnected power grid area is defined as a target node;

[0015] The layered directed graph includes a first directed edge, a second directed edge, an energy input layer, an energy power input layer, and a target node;

[0016] The energy input layer contains the same number of energy nodes as the distributed energy nodes, the energy power input layer contains the same number of energy power nodes as the distributed energy nodes, and the number of target nodes is one.

[0017] Each energy node in the energy input layer is connected to the corresponding energy power node in the energy power input layer through a first directed edge, representing a one-way causal relationship from power generation resource data to power generation data; wherein, the weight of the first directed edge is calculated based on the causal strength between power generation resource data and power generation data;

[0018] Each energy power node in the energy power input layer is connected to the target node through a second directed edge, representing the convergence relationship between power generation data and total power prediction value; wherein, the weight of the second directed edge is calculated based on the contribution and correlation of the power generation node to the total power prediction value.

[0019] Optionally, the method for obtaining the dynamic spatiotemporal correlation features includes:

[0020] Construct the initial spatial correlation matrices for energy nodes and energy power nodes respectively;

[0021] An improved normalized mutual information model incorporating kernel density estimation and skewness correction is used to nonlinearly enhance and correct the distance decay of the initial spatial correlation matrix.

[0022] The enhanced and corrected correlation matrices of consecutive moments are stacked along the time dimension, and time decay weights and derived features are incorporated to form dynamic spatiotemporal correlation features.

[0023] Optionally, the energy output prediction model uses a spatial domain graph convolutional network to fuse hierarchical directed graphs and dynamic spatiotemporal correlation features to predict total power.

[0024] The spatial domain graph convolutional network includes multi-layer graph convolutional residual blocks and a temporal attention convolution and gated loop module serially connected between each adjacent graph convolutional residual block;

[0025] The graph convolutional residual block is used for feature propagation based on the adjacency matrix of the hierarchical directed graph; wherein the adjacency matrix of the hierarchical directed graph is obtained by normalizing the edge weights of the hierarchical directed graph.

[0026] The graph convolutional residual block includes an edge attention mechanism and a node attention mechanism. The edge attention mechanism is used to multiply the edge attention weights of the dynamic spatiotemporal correlation features with the adjacency matrix element by element-wise multiplication to fuse the hierarchical directed graph and the dynamic spatiotemporal correlation features. The node attention mechanism is used to adaptively adjust the importance of each node feature in the hierarchical directed graph.

[0027] The temporal attention convolution is used to extract local temporal dependency features of each node in the hierarchical directed graph through dynamic convolution kernels and dilation rates. The gated loop module uses a gating mechanism to control the information flow of each node feature in the hierarchical directed graph.

[0028] Optionally, the load prediction model includes a sequentially connected time-graph convolutional network and an output layer;

[0029] The temporal graph convolutional network is used to fuse dynamic spatial features and temporal dependent features to obtain spatiotemporal fusion features, including parallel graph convolutional networks and gated recurrent units;

[0030] The graph convolutional network is used to extract spatial features between load nodes based on the load spatiotemporal information graph.

[0031] The gated loop unit is used to extract time-dependent features based on the historical load sequence data of each load node;

[0032] The output layer is used to obtain multi-step load prediction values ​​based on spatiotemporal fusion feature mapping.

[0033] Optionally, the graph convolutional network includes a serially connected two-layer graph convolutional structure and an adaptive weight adjustment unit;

[0034] The dual-layer graph convolutional structure is used to extract spatial features between load nodes based on the adjacency matrix of the load spatiotemporal information graph. The adjacency matrix of the load spatiotemporal information graph is obtained by normalizing electrical distance, geographical distance, load correlation, and time-dynamic weights. The electrical distance is calculated based on the inter-station impedance, the geographical distance is the straight-line distance between the inter-stations, the load correlation is determined by fusing improved normalized mutual information with the Pearson correlation coefficient, and the time-dynamic weights are calculated based on the load peak ratio and the correlation coefficient with ambient temperature.

[0035] The adaptive weight adjustment unit is used to introduce spatial attention weights to adjust the spatial feature extraction process.

[0036] Optionally, the load node includes historical load sequences and higher-order derived features;

[0037] The higher-order derived features are obtained by standardizing the peak load, ambient temperature, and temperature correlation coefficient.

[0038] Optionally, the hierarchical federated learning framework includes:

[0039] Each distributed energy node optimizes its local model parameters based on local data and uploads them to the coordinator on the corresponding interconnected area side for regional aggregation, generating a regional-level global model for each interconnected area.

[0040] Each interconnected station area coordinator uploads the corresponding regional-level global model parameters to the global coordinator for global aggregation, generates a global model, and then distributes the global model parameters to the corresponding regional-level global model of each interconnected station area.

[0041] The global model parameters are distributed to the local models of each distributed energy node through the regional coordinator.

[0042] Secondly, the present invention provides an interconnected distributed energy output and load forecasting device for power distribution areas, comprising:

[0043] Data acquisition module: used to acquire the energy data and load data to be measured for each interconnected power distribution area; wherein, the energy data to be measured includes the power generation resource data and power generation data of the distributed energy nodes in each interconnected power distribution area; the load data to be measured includes the historical load sequence data and topology information of each interconnected power distribution area;

[0044] Energy output prediction module: used to input the energy data to be measured into a pre-trained energy output prediction model to obtain the total power prediction value of distributed energy nodes in each interconnected area; wherein, the energy output prediction model constructs a hierarchical directed graph based on the energy data to be measured and extracts the dynamic spatiotemporal correlation features between distributed energy nodes, and integrates the hierarchical directed graph and the dynamic spatiotemporal correlation features to predict the total power; each node of the hierarchical directed graph represents the power generation resource data, power generation data and the total power prediction value respectively, and the hierarchical directed graph contains the causal relationship between the nodes;

[0045] Load forecasting module: This module constructs a load spatiotemporal information map based on the topology information and historical load sequence data. The load spatiotemporal information map is then input into a pre-trained load forecasting model to obtain load forecast values ​​for each interconnected transformer area at multiple future time points. The load forecasting model extracts dynamic spatial features between nodes in the load spatiotemporal information map and extracts time-dependent features based on historical load sequence data. After fusing the dynamic spatial features and time-dependent features, multi-step load forecasting is performed. Each node in the load spatiotemporal information map represents the overall load of each interconnected transformer area.

[0046] The energy output prediction model and the load prediction model are both trained collaboratively using a hierarchical federated learning framework.

[0047] Thirdly, the present invention provides a computer storage medium having a computer program stored thereon, which, when executed by a processor, implements the distributed energy output and load forecasting method for interconnected power distribution areas as described in any of the first aspects.

[0048] Compared with the prior art, the beneficial effects achieved by the present invention are as follows:

[0049] 1. By constructing a hierarchical directed graph that integrates dynamic spatiotemporal correlation and physical causal relationship, and using a spatial domain graph convolutional network for feature extraction, it is possible to accurately characterize the dynamic correlation between interconnected stations as time and operating status evolve. This overcomes the shortcomings of traditional static graph models and prediction models that ignore causal relationships in modeling nonlinear and time-varying characteristics, thereby achieving higher accuracy in predicting the total output of distributed energy.

[0050] 2. In distributed energy power prediction, by introducing dynamic correlation measurement based on mutual information and hierarchical directed graph modeling, the model can adaptively learn the correlation strength between intermediate intervals in complex environments, without relying on fixed prior topology, which enhances the model's adaptability to different interconnection scenarios and operating conditions. At the same time, explicit causal modeling reduces the interference of irrelevant noise and improves the robustness of the model when data fluctuates or outliers exist.

[0051] 3. Based on the cross-domain federated collaborative learning framework, each interconnected station area is allowed to use private data for model training locally, only exchanging encrypted model parameters or intermediate gradients, without sharing original sensitive data (such as specific user loads and generator operation details). This effectively solves the data collaboration problem caused by data silos and privacy concerns in the interconnected system, and protects the data security and privacy rights of all participants while fully exploring cross-domain spatiotemporal characteristics.

[0052] 4. In distributed energy power forecasting, by introducing a single target node to directly output the total power of the power grid, multiple independent forecasting error sources in traditional methods are aggregated into one, which theoretically reduces error accumulation and optimizes the overall forecasting performance. In load forecasting, the time-map convolutional network realizes synchronous and coordinated multi-step forecasting of the load of all interconnected transformer areas, providing a more consistent and reliable decision basis for system optimization scheduling. Attached Figure Description

[0053] Figure 1 The diagram shown is a framework diagram of a wind farm cluster power prediction method in one embodiment of this application.

[0054] Figure 2 The diagram shown is a schematic diagram of a hierarchical directed graph structure in one embodiment of this application;

[0055] Figure 3 The figure shown is an example of the simulation results of wind energy forecast based on a wind farm 4 hours later, according to one embodiment of this application.

[0056] Figure 4 The diagram shown is a comparison of RMSE and MAE of different methods in the test set in one embodiment of this application.

[0057] Figure 5 The diagram shown is a schematic representation of the spatiotemporal information of node load in an interconnected area according to one embodiment of this application.

[0058] Figure 6 The diagram shown is a schematic diagram of the TGCN structure in one embodiment of this application;

[0059] Figure 7 The diagram shown is a comparison of RMSE of different models in multi-step prediction in one embodiment of this application.

[0060] Figure 8 The figure shown is an example of the prediction simulation results for steps one, two, and three in one embodiment of this application. Detailed Implementation

[0061] The present invention will be further described below with reference to the accompanying drawings. The following embodiments are only used to more clearly illustrate the technical solution of the present invention, and should not be used to limit the scope of protection of the present invention.

[0062] Example 1

[0063] This embodiment proposes a method for forecasting the output and load of distributed energy resources in interconnected power distribution areas. Taking the ultra-short-term total power forecasting problem of wind farm clusters as an example, it explains how to apply the proposed hierarchical directed graph and dynamic spatiotemporal correlation matrix to construct and train a spatial domain graph convolutional network (GCN), and build an energy output forecasting model based on the spatial domain GCN model. At the same time, based on the power forecasting logic, it extends to the collaborative forecasting of power distribution area loads, and performs spatiotemporal forecasting of the short-term power load of the interconnected power distribution area system. It demonstrates how to combine the spatial domain GCN and the gated recurrent unit (GRU) to construct a temporal graph convolutional network (TGCN), and then build a load forecasting model based on the TGCN to achieve collaborative multi-step forecasting of multi-node loads.

[0064] In the specific implementation process, the construction of the energy output prediction model and load prediction model proposed in this application includes the following common steps: Data acquisition and preprocessing: collecting historical operating data of each node (such as output unit and load point) within the interconnected area and performing normalization and other standardization processes to provide features of a unified scale for model input; Graph structure modeling: defining node types according to the prediction task (output or load) and constructing an adjacency matrix or spatiotemporal correlation tensor reflecting the spatiotemporal relationship between nodes based on dynamic measurement methods such as mutual information or fixed electrical topology; Model construction and training: designing a corresponding graph neural network model based on the constructed graph structure and iteratively optimizing the model parameters using a pre-defined training set; Real-time prediction and performance evaluation: using the trained model, inputting the latest real-time data to generate prediction results, and using standard indicators such as root mean square error and mean absolute error to quantitatively evaluate the prediction accuracy. The specific steps are as follows:

[0065] S1, Node Definition and Input Sample Construction: Based on different distributed renewable energy sources such as wind farms and photovoltaic power plants, three types of nodes are defined: energy nodes (representing wind speed and irradiance), energy power nodes (representing wind power and photovoltaic output), and target nodes (referring to the total power prediction node). The number of energy nodes and energy power nodes is the same as the number of electric fields N in the interconnected power field cluster, with each node corresponding to one renewable energy power plant. There is only one target node, used to output the total power prediction value of the renewable energy power field cluster within the interconnected power field area. Simultaneously, historical data for each node is collected, preprocessed, and feature-encoded, and then used with an adaptive step-size sliding window technique to generate training samples.

[0066] S2, Dynamic Spatiotemporal Correlation Matrix Calculation: Centered on mutual information measurement and tensor construction, this method accurately characterizes the dynamic correlations between nodes. An improved normalized mutual information (KDE-Normalized Mutual Information) model, incorporating kernel density estimation and skewness correction, is employed to precisely quantify the strength of nonlinear dynamic correlations between nodes. Initial spatial correlation matrices for energy and power nodes are constructed separately, and their weights are optimized through nonlinear enhancement and distance decay correction. The enhanced matrices from consecutive time steps are stacked along the time dimension, incorporating time decay weights and derived features to form a spatiotemporal correlation tensor, comprehensively capturing the spatiotemporal correlation characteristics between nodes.

[0067] S3, Hierarchical Directed Graph Construction: Construct a hierarchical directed graph consisting of an energy input layer, an energy power input layer, a target node, a first directed edge, and a second directed edge. This structure clarifies the mapping relationship from input variables to output variables, and by introducing a target node, reduces the number of error sources in the prediction process from N to 1. At the same time, the causal relationship is clarified through directed edges.

[0068] S4, Distributed Energy Output Model Training and Prediction: A distributed energy power prediction model integrating spatiotemporal features and causal relationships is constructed based on the spatial domain GCN; node attention, edge attention, and residual connections are introduced to optimize feature propagation, and temporal attention convolution and gated recurrent modules are inserted between spatial domain GCN layers to extract time-dependent features; FedProx algorithm combined with differential privacy protection is used for federated training, each distributed energy node optimizes the proximal loss function based on local data, and the coordinator updates global parameters weighted by data volume and model performance; at new prediction times, S1-S3 are repeated to generate real-time input, which is then input into the trained model to obtain the total power prediction value.

[0069] S5. Constructing a load forecasting model: Based on the power forecasting logic, the model extends to the coordinated load forecasting of distribution areas. First, when constructing the spatiotemporal information map of load in interconnected distribution areas, the load of each interconnected distribution area is regarded as a graph node. A dynamic graph structure integrating topology, load characteristics, and external influencing factors is constructed to lay the foundation for spatiotemporal feature extraction. Subsequently, GCN is used to extract the spatial features between load nodes, and GRU is used to extract the temporal features of historical load data. On this basis, TGCN is constructed, and historical data is used to train the spatiotemporal forecasting model to optimize parameters. After training, the latest distribution area load sequence and a fixed topology adjacency matrix are input, and the model can synchronously output the load forecast values ​​of all interconnected distribution areas at multiple future times, realizing system-level coordinated multi-step forecasting.

[0070] S6, Prediction and Evaluation: The prediction results are evaluated using root mean square error, mean absolute error, and mean relative percentage error to verify the model's load prediction performance for interconnected station areas.

[0071] In this embodiment, S1 specifically includes:

[0072] S11, Data Collection and Preprocessing: Data collection and preprocessing are the core prerequisites for data analysis and modeling, directly determining the reliability of subsequent results. In this embodiment, historical data from each distributed energy node in the interconnected area is collected and then normalized. An improved adaptive weighted Z-Score normalization process is used to eliminate dimensional differences and suppress the influence of outliers.

[0073] ,

[0074] ,

[0075] ,

[0076] in, The data is after normalization; The deviation after removing local means; It represents a local fluctuation scale, similar to a local standard deviation; The local standard deviation reflects the overall dispersion of the data. The standard deviation of the data reflects the discreteness of the current data; , which is the offset coefficient, used to adjust the distribution of the normalized data to a more suitable range for model training; Raw data (such as wind speed, power generation). For the first in the window Data points at each time point; The current moment; The step size is 72. As a time decay weight; For the first The original data values ​​at each moment; This is the index variable for the inner summation, used to iterate through the data at each time point within the sliding window and calculate the weighted average. For the first The original data values ​​at each moment; To avoid smooth terms with a denominator of zero.

[0077] S12, Node Definition and Feature Encoding: Define three types of nodes: energy nodes (representing wind speed, light intensity, etc.), energy generation power nodes, and target nodes; also define the input samples for each node. The input samples for energy nodes and energy generation power nodes are historical data. Similarly, the input samples for the target node contain historical total power. Step. Define three types of nodes in time. The input samples can be:

[0078] ,

[0079] ,

[0080] ,

[0081] In summary, the problem of power forecasting for power plant clusters can be described as follows:

[0082] ,

[0083] in, Indicates the first One energy node input sample; Indicates the first In the input samples of each energy node at time step Energy data; Indicates the first Input samples for each energy power node; Indicates the first In the input samples of each energy power node at time step Energy power data; This represents the input sample for the target node; Indicates at time step Total energy output data; This refers to the task of power forecasting for power plant clusters.

[0084] S13, Sliding Window Sample Generation: An adaptive step-size sliding window technique is used to extract training samples from the time series data. This fully utilizes the temporal correlation of the data and adapts to data fluctuation characteristics, improving sample quality and model training efficiency. The formula for calculating the total number of samples is:

[0085] ,

[0086] ,

[0087] in, This represents the total number of training samples extracted in the end. This represents the total length of the original time-series data; The length of the historical sequence of the input sample; To predict the number of steps; For the first One sample window; To adapt the sliding step size, , ; The indicator function (1 if the condition is met, 0 otherwise) is used to filter effective samples with small fluctuations. This is the data fluctuation threshold (default 0.15 times the global standard deviation), used to control the strictness of sample selection; represents the standard deviation of the interval data.

[0088] S2 specifically includes:

[0089] S21, Dynamic Correlation Measurement Based on Mutual Information: NMI is an information metric in information theory. It can be viewed as the amount of information contained in one random variable about another. In the prediction scenario of distributed energy output and transformer load, data such as wind speed, irradiance, and load often exhibit nonlinear fluctuations, and the correlation between transformers changes dynamically with weather and time periods. Improving the measurement accuracy in nonlinear and dynamic correlation scenarios allows the model to more accurately quantify these complex and time-varying correlations (such as the nonlinear causal relationship between wind speed and power, and the dynamic coupling of cross-transformer loads), thereby improving prediction accuracy. Kernel density estimation and skewness correction are introduced:

[0090] ,

[0091] The formula for calculating KDE-NMI is:

[0092] ,

[0093] in, All are multi-dimensional time series, with each series corresponding to specific monitoring data of a node, and the time dimension is consistent with the data sampling frequency; Variables calculated based on kernel density estimation Information entropy, measured in bits, is used to more accurately describe the uncertainty of variables; , Variables The , Each sample value; The bandwidth parameter for kernel density estimation is used to ensure the reasonableness of the kernel density estimation; Variables calculated based on kernel density estimation Information entropy; Variables calculated based on kernel density estimation and The joint information entropy; For improved normalized mutual information, the value range is [0,1]. The closer the value is to 1, the better the variable... and The stronger the correlation, the more directly it can be used later. Substitute , for and Skewness is used to describe the asymmetry of the distribution of a variable; This is a skewness correction factor used to improve the accuracy of correlation measurements between asymmetric distributed variables.

[0094] S22, Construct the spatial correlation matrix: For each time step Calculate the KDE-NMI between all pairs of input samples from all energy nodes to form the initial energy spatial correlation matrix. ,in, Similarly, the normalized mutual information of input samples between all energy power nodes is calculated to form the initial power space correlation matrix. By quantifying the correlation strength between two time series variables using normalized mutual information, and to strengthen the weights of strongly correlated nodes and suppress interference from weak correlations, the initial matrix undergoes nonlinear enhancement and normalization processing, as shown in the following formula:

[0095] ,

[0096] ,

[0097] in, To enhance post-energy nodes and At any moment Spatial correlation weights; This is the nonlinear enhancement coefficient, with a default value of 15. This is the correlation threshold parameter, with a default value of 0.4; For energy nodes and The initial associated value; , The initial energy space correlation matrix at time 10 is respectively The maximum and minimum values; To represent the row index in the spatial incidence matrix, represents the row index of the first row. One node; To represent the column index in the spatial incidence matrix, represents the first... One node; For energy nodes and The corresponding geographical distance between them; For energy power nodes and The corresponding electrical distance between them; This is the distance attenuation factor, with a default value of 0.3; To enhance post-energy power nodes and At any moment Spatial correlation weights; For power nodes and The initial associated value; In the initial correlation matrix of the power space, the first... The node and the first Initial association values ​​between nodes; , They are respectively.

[0098] S23, Constructing the Spatiotemporal Correlation Matrix: The enhanced spatial correlation matrices from multiple consecutive time points are stacked along the time dimension, and time decay weights are incorporated to form a three-dimensional spatiotemporal correlation tensor. Simultaneously, derived features from the time dimension are introduced to comprehensively characterize the spatiotemporal correlation properties between nodes.

[0099] Spatiotemporal correlation tensor:

[0100] ,

[0101] ,

[0102] in, For the corresponding node in the energy spatiotemporal correlation tensor and At time step The element value; This is the time step index, corresponding to the th time step in a series of H consecutive time steps. At any moment ( The value range is from the first time step to the Hth time step, used to traverse all data within the historical window, such as historical sequences of wind speed and load, to provide complete time series input for the model. For energy nodes and At any moment Enhanced spatial correlation weights; This is the time decay coefficient, with a default value of H / 3, used to give higher weight to the spatial correlation of recent times; The Pearson correlation coefficient is used to measure the consistency of spatial correlation across different time periods. For the corresponding node in the power spatiotemporal correlation tensor and At time step The element value; For energy power nodes and At any moment Enhanced spatial correlation weights.

[0103] The energy input layer mentioned in S3 consists of all energy nodes and their corresponding energy spatiotemporal correlation matrices; the energy power input layer consists of all energy power nodes and their corresponding energy power spatiotemporal correlation matrices; the target node is used to directly output the predicted total power of the electric field group; the first directed edge points from each energy node in the energy input layer to the corresponding energy power node in the energy power input layer, representing the unidirectional causal relationship of "energy output determines power", ensuring that information flows only unidirectionally from the energy output layer to the power layer; the second directed edge points from each energy power node in the energy power input layer to the target node, used to aggregate the power information of each power plant. The layered directed graph models the data flow of "energy → power → predicted total power" through directed edges, specifically defined as follows:

[0104] Hierarchical directed graph ,in, Represents a set of nodes. Denotes the set of edges. Represents the set of weights;

[0105] Node set: ,in, For energy node set, For energy power node set, For the target node;

[0106] Edge set: ,in, This is the first directed edge (energy → energy power). This is the second directed edge (energy power → target).

[0107] Weight set In the above, the first directed edge weight is calculated based on the energy-power causality strength, and the second directed edge weight is calculated based on the power node contribution and correlation, as follows:

[0108] ,

[0109] ,

[0110] in, For the first The weight of the first directed edge reflects the energy node. For energy power nodes The strength of the causal influence; For power nodes The variance of the input samples reflects the degree of fluctuation in the power data; For energy nodes The variance of the input sample reflects the degree of fluctuation in energy resource data; For the first The weight of the second directed edge reflects the power node. Contribution to the target node; No. Rated power of each node.

[0111] S4 specifically includes:

[0112] S41, Construction of the Overall Adjacency Matrix: The hierarchical directed graph structure constructed above clearly describes the mapping relationship between the input data and the future total power. After obtaining the directed graph, the corresponding adjacency matrix is ​​defined for training the spatial domain GCN. The specific definition is as follows:

[0113] ,

[0114] For adjacency matrix Symmetric normalization and dynamic scaling are performed to improve the stability and effectiveness of graph convolution:

[0115] ,

[0116] in, for The degree matrix, ; Degree matrix The trace, that is, the sum of the diagonal elements; Degree matrix The average of the diagonal elements is used to dynamically scale the adjacency matrix. The values ​​of the adjacency matrix reflect the relationship between two different variables, therefore, in this formula... and They are not equal.

[0117] S42, Multi-level Graph Convolutional Design: The proposed model utilizes a Geographic Convolutional Network (GCN) to achieve power forecasting for power plant clusters. GCNs connect power plants in different distribution areas and can be categorized into spectral domain GCNs and spatial domain GCNs. However, spectral domain-based GCNs have significant limitations; for example, they can only handle small-scale and static graph structures and cannot handle directed graphs. These limitations mean that the input to spectral domain GCNs typically only contains time series data of various influencing factors. Furthermore, information on the various spatiotemporal correlations and causal relationships between these input variables cannot be used as model input. These shortcomings negatively impact the accuracy of power forecasting. In view of these shortcomings, this application employs an L-layer spatial domain GCN, introducing node attention, edge attention, and residual connections to optimize feature propagation. layer( The propagation formula is:

[0118] ,

[0119] ,

[0120] ,

[0121] Where GELU is the activation function; For the first Layer feature matrix, For the first Layer feature dimension; This is an element-wise multiplication method that performs element-wise multiplication between the adjacency matrix and the attention score matrix. This is the edge attention weight matrix, used to calculate the attention score of each edge. This is the activation function used to normalize the attention score; This is the residual connection weight matrix, which helps alleviate the gradient vanishing problem in deep networks. The residual coefficient controls the contribution strength of residual connections. Layer normalization coefficients improve model stability; This is a layer normalization operation that standardizes the feature distribution. This is the node attention weight vector, used to adaptively adjust the importance of each node's features; This is the node attention weight matrix; For node attention bias; For the first Feature matrix after layer attention enhancement.

[0122] S43, Temporal Dimension Feature Extraction: Temporal attention convolution and gated recurrent modules are inserted between spatial domain GCN layers to extract temporally dependent features, achieving deep fusion of spatiotemporal features.

[0123] ,

[0124] ,

[0125] ,

[0126] ,

[0127] in, This is the temporary temporal feature matrix after convolution; This is a one-dimensional convolution operation used to extract local temporal dependent features; The kernel size is dynamically adjusted according to the number of network layers, enhancing the adaptability of temporal feature extraction; The expansion rate increases the receptive field of the convolution as the number of network layers increases; This is the temporal attention weight matrix; For temporal attention and gating weight matrix; For temporal attention and gating bias terms; It is the weight matrix of the previous layer or the previous time state; Assign higher weights to the recent time steps as the temporal attention decay coefficient; For gating matrix; The output range is [0,1], which controls the gate switch; For the first The temporal feature matrix after layer fusion.

[0128] S44, Model Training: In the federated learning scenario for distributed energy prediction: Each distributed energy node optimizes its local model parameters based on local data and uploads them to the corresponding substation-side coordinator for regional aggregation, generating a regional-level global model for each interconnected substation; the substation-side coordinator uploads the corresponding regional-level global model parameters to the global coordinator for global aggregation, generating a global model, and then distributes the global model parameters to the corresponding regional-level global models for each interconnected substation; the substation-side coordinator distributes the global model parameters to the local models of each distributed energy node. The FedProx algorithm addresses the instability of model training caused by heterogeneous data across substations by introducing a near-end term; differential privacy adds noise during model parameter interaction, protecting the privacy of the original energy / load data of each substation while ensuring the global model balances accuracy and data security, adapting to the data collaboration needs of interconnected substations. The following is the training of the model based on the federated learning framework and algorithm:

[0129] First, initialize the regional-level global model parameters. Set differential privacy noise parameters This is used to control the balance between the strength of privacy protection and data availability;

[0130] Secondly, each distributed energy node in each interconnected power grid uses adaptive learning rate SGD to optimize the proximal loss function based on local data:

[0131] ,

[0132] ,

[0133] in, For the first The number of participating distributed energy nodes in the 1st The local model parameters of the wheel; For the first The region-level global model parameters of the wheel; For the first The number of participating distributed energy nodes in the 1st The adaptive learning rate of the wheel; To calculate the gradient of the model parameters; A hybrid loss function that comprehensively considers prediction error, regularization, graphical smoothing, and total variation constraints; , where is the proximal regularization coefficient, which alleviates the training bias caused by Non-IID data; Gaussian noise is used to achieve differential privacy protection; The standard deviation of differential privacy noise; For sequence The total variation; For the sequence at the 1st The value at each moment.

[0134] Hybrid loss function Defined as:

[0135] ,

[0136] in, For the first The amount of local data of each participating distributed energy node; This is the mean squared error loss term, which measures the deviation between the predicted value and the actual value. Distributed energy nodes The characteristic output; Distributed energy nodes The characteristic output; for Regularization coefficients are used to prevent the model from overfitting. To smooth the loss coefficients in the graph and enhance the consistency of features across the graph structure; This is the total variation loss coefficient, which improves the smoothness of the predicted sequence.

[0137] Adaptive learning rate The calculation formula is:

[0138] ,

[0139] in, The initial learning rate; For the first During rounds, local dataset upper loss function with respect to parameters The gradient; For the first The variance of local data of each participating distributed energy node.

[0140] Differential privacy noise standard deviation The calculation formula is:

[0141] ,

[0142] in, This is the gradient Lipschitz constant, with a default value of 1.0; Number of participants; This refers to the total number of training rounds.

[0143] Then, upload the optimized local model parameters. The corresponding interconnected area-side coordinator updates the regional-level global model parameters using a weighted average based on data volume and model performance.

[0144] ,

[0145] in, For the first The region-level global model parameters of the wheel; For the first The local model of the participating distributed energy nodes in the 1st year The root mean square error on the validation set; As a performance weighting coefficient, it assigns higher weight to the local model with better prediction accuracy.

[0146] Finally, each interconnected zone coordinator uploads the corresponding regional-level global model parameters to the global coordinator for global aggregation. The global coordinator updates the global model parameters by weighted average based on data volume and model performance, and distributes them to each regional-level global model. Then, each interconnected zone coordinator distributes them to the local model corresponding to each distributed energy node.

[0147] S45, Real-time Prediction: For a new prediction time, repeat steps S1-S3 to generate real-time input samples. Spatiotemporal correlation tensor By inputting the trained model, the initial prediction value can be obtained. The specific process is as follows:

[0148] First, the raw data is normalized by S11 and encoded by S12 nodes to form an initial feature matrix. (Including the initial characteristics of energy nodes, power nodes, and target nodes); secondly, the dynamic spatiotemporal correlation tensor calculated by S2 and the hierarchical directed graph constructed by S3 are transformed into a normalized adjacency matrix. This provides spatiotemporal correlation and causal constraints for feature propagation; then, through L-layer graph convolution and temporal feature extraction, the target node features are iteratively enhanced, ultimately forming... Finally, the future is output through multi-layer perceptron (MLP) mapping. Initial predicted value of total power of the field group at time 1 This completes the entire process from feature input to model computation and then to power output. Prediction output formula:

[0149] ,

[0150] in, For the multilayer perceptron (output layer of the energy output prediction model), it realizes the mapping from the high-dimensional features of the target node to specific power values; The final high-dimensional feature vector of the target node after passing through the L-layer spatial domain GCN and spatiotemporal feature fusion; The weight matrix of the MLP output layer maps the hidden layer output to the power prediction dimension. is the weight matrix of the MLP hidden layer, which performs a nonlinear transformation on the final features of the target node. This is the bias vector of the MLP hidden layer, used to adjust the offset of the hidden layer output. This is the bias vector for the MLP output layer, used to adjust the offset of the final power prediction value.

[0151] In this embodiment, the adjacency matrix construction mentioned in S5 for building the load forecasting model is based on the fusion of four factors: electrical distance, geographical distance, load correlation, and time dynamic weights, to construct a time-varying adjacency matrix. ( (where the number of nodes is the number of nodes), the formula is as follows:

[0152] ,

[0153] Normalize the adjacency matrix:

[0154] ,

[0155] in, For a moment Taiwan and The original adjacency weights between them; The four-factor weighting coefficients satisfy the following conditions: The value is determined through optimization using the validation set. Taiwan District and Electrical distance; The maximum electrical distance between all stations; This is the electrical distance attenuation coefficient; Taiwan District and At any moment The load power; Taiwan District and The geographical straight-line distance; Standard deviation of geographical distance; To measure covariance, the area of ​​the station is... and The degree of linear correlation of the load sequences; To improve normalized mutual information, the nonlinear correlation strength of load sequences is measured; Taiwan District Peak values ​​of historical load sequences; This represents the highest historical load for all transformer substations. Taiwan District and At any moment Ambient temperature; Pearson correlation coefficient, which measures the degree of correlation between temperature series; For the normalization of the back-end area and The adjacency weights between rows are used to ensure that the sum of the weights for each row is 1.

[0156] The aforementioned feature matrix construction: The feature vector of each load node contains multi-dimensional load-related information and higher-order derived features, and the feature matrix... , ( For feature dimension, The feature matrix is ​​standardized using the historical time step as follows:

[0157] ,

[0158] in, These are the original feature values ​​at the corresponding positions; , The first The global mean and standard deviation of the dimensional features.

[0159] GCN employs a two-layer graph convolutional network and introduces an adaptive weight adjustment process for spatial feature extraction.

[0160] ,

[0161] ,

[0162] ,

[0163] ,

[0164] in, , These are the base spatial feature matrices output by the first and second layer graph convolutions, respectively. For activation functions; This is the normalized adjacency matrix; The convolution kernel weight matrix is... For bias terms; The weight matrix is ​​used for skip connections to alleviate gradient vanishing; Spatial attention weights, For bias terms, The second layer of graph convolution outputs the feature dimensions; This is the final spatial enhancement feature matrix; Transform the attention weight vector into a diagonal matrix.

[0165] GRU consists of reset gates and update gates. The formulas for calculating the reset gate, update gate, candidate output, and GRU output of a single-layer GRU are as follows:

[0166] ,

[0167] ,

[0168] ,

[0169] ,

[0170] in, To update the gate, control the ratio of old information retention to new information addition; To reset the door, control whether historical information is forgotten; The candidate hidden state is the new state calculated based on the reset historical information; For the first The final hidden state of the step; is a trainable weight matrix, which is applied to the current input and the historical hidden state, respectively; For the first The node input features of the step; For the first The hidden state of the node in the step; For bias terms; The Sigmoid activation function has an output range of [0,1]. It is a hyperbolic tangent activation function with an output range of [−1, 1].

[0171] The model output layer uses an MLP to achieve multi-step prediction:

[0172] ,

[0173] in, For the future A sequence of load forecast values ​​for all interconnected transformer areas at any given time; To predict the number of steps; For length is The time feature sequence within the historical window (extracted by the GRU module).

[0174] The load prediction model is trained using a federated learning framework, similar to step S44, to achieve cross-domain collaborative training.

[0175] To fully demonstrate the technical solution and its effects proposed in this application, specific implementation methods will be described in detail through examples:

[0176] This example aims to predict the integrated output of distributed energy resources and the power load in an interconnected distribution area. Specifically, it focuses on predicting the output of distributed energy resources, primarily wind farms. The data used in this case study comes from 32 different wind farms, including historical wind speeds and wind power generation for each farm from June 1st to December 30th, 2020. The time resolution is 15 minutes. Details of the training, validation, and test sets are shown in Table 1.

[0177] Table 1

[0178]

[0179] like Figure 1 The flowchart shown illustrates the power forecasting process for ultra-short-term wind farm clusters. Load forecasting involves selecting 10 typical distribution areas interconnected by tie lines in the distribution network and collecting hourly load data from the past year. The data collection uses the first 330 days as the training set and the last 35 days as the test set to predict the load for the next 1, 2, and 3 hours. This includes the following steps:

[0180] Step S1, Node Definition and Input Sample Construction: Based on wind farms, three types of nodes are defined: wind speed nodes, wind power nodes, and target nodes. The number of wind speed nodes and wind power nodes corresponds to the number of wind farms in the wind farm cluster. Similarly, each node corresponds to one wind farm; there is only one target node, used to output the predicted total power of the wind farm group. This includes S11~S13:

[0181] Step S11, Data Collection and Preprocessing: Collect data for the wind farm cluster from the start point of time. arrive Historical operational data, including wind speed data for 32 wind farms, active power data for each wind farm, and total active power data for the entire wind farm cluster, all with a time resolution of [missing information]. The raw data from 32 wind farms were normalized to bring all features to the same scale.

[0182] Step S12, Node definition and feature encoding:

[0183] This example defines three types of nodes and constructs a fixed-length historical sequence as a feature for each node:

[0184] Wind speed node : Corresponding to the A wind farm, with node characteristics as follows: Step (24-hour) historical wind speed sequence .

[0185] Wind power node : Corresponding to the A wind farm, with node characteristics as follows: Step historical power sequence .

[0186] target node : Characterized by the total power of the field group Step historical sequence .

[0187] The initial feature matrix of all nodes can be represented as:

[0188] ,

[0189] in, This represents the initial hidden layer state or initial feature matrix of the model, with the superscript (0) indicating step 0 (initial time). For the first Moment The features consist of 32 dimensions (referring to wind speed data); For the first Moment Class features, totaling 32 dimensions (referring to power process parameters); For the first Additional characteristics of time; Transpose the entire vector to convert a row vector into a column vector; This indicates that the dimension of the feature matrix is ​​65×H.

[0190] Step S13, Sliding Window Sample Generation: Training samples are extracted from the time series using the sliding window technique. For any given time... Define the input time window and output time window By sliding these windows along the time axis, a large number of training sample pairs can be obtained.

[0191] Step S2, dynamic spatiotemporal correlation matrix calculation: First, the spatial correlation of the input samples at each time step needs to be obtained. Then, the spatiotemporal correlation matrix is ​​constructed by arranging the obtained spatially relevant information in chronological order. Specifically, this includes:

[0192] Step S21, Dynamic correlation measurement based on mutual information: Traditional methods often use geographical distance or correlation coefficient to measure inter-field association, but cannot capture non-linear and dynamic dependencies.

[0193] Step S22, construct the spatial correlation matrix: for each prediction time... Calculate the dynamic correlation matrix between all wind speed nodes within the current time window. ; and thus obtain These correlation matrices together form the spatial correlation matrix.

[0194] Step S23, construct the spatiotemporal correlation matrix: connect the continuous The spatial correlation matrices at each moment are stacked along the time dimension to form the spatiotemporal correlation tensor.

[0195] Step S3, constructing the hierarchical directed graph structure: as follows Figure 2 In this example, a hierarchical directed graph structure is constructed, which includes a wind speed input layer, a wind power input layer, a target node, a first directed edge, and a second directed edge. This structure clarifies the mapping relationship from input variables to output variables, and by introducing a target node, the error sources in the prediction process are reduced from multiple to one. At the same time, the causal relationship is clarified through the directed edges.

[0196] In this example, the directed edge I clarifies the causal relationships between input variables of different physical attributes. The causal relationship referred to here is a one-way influence relationship. For example, wind speed and wind force are two variables with different physical attributes; wind speed can determine wind force, but wind force cannot affect wind speed. However, previous studies have almost completely ignored this causal relationship. Their common approach is to integrate historical data of different physical attributes into time series data and input them into a model for prediction. The core of this method is the assumption that historical wind speed and historical wind power are completely independent and jointly influence future wind force, but this is unrealistic. To address this issue, a directed edge I is used to connect the wind speed input layer and the wind power input layer, ensuring that historical wind speed information flows only unidirectionally to the wind power input layer.

[0197] In this example, the function of the directional edge II and the target node is to reduce error sources in training the model and power prediction. Specifically, if the target node is not set, the total predicted power at each time step will be equal to all... The sum of the predicted power of each wind farm, which means there exists The presence of multiple error sources is extremely detrimental to reducing prediction errors in spatial domain GCNs. However, after setting the target node and directed edges II, the total prediction power at each time step can be obtained directly from the target node, meaning that the number of error sources will decrease from the initial... Reducing the error to 1 helps improve the accuracy of the prediction model. Regarding error source analysis, the error in predicting total power using traditional methods can be decomposed into:

[0198] ,

[0199] in, This represents the total energy loss of the system. To The summation of individual energy losses represents the energy loss of each independent component or process; This is a correction term for energy loss, used to compensate for measurement errors, model simplifications, or additional losses not covered by the sub-terms.

[0200] This example directly predicts the total power by introducing a target node, and the error expression is simplified to:

[0201] ,

[0202] in, This represents the power prediction error.

[0203] Reduced For each independent error source, the theoretical reduction in error variance is:

[0204] ,

[0205] in, The variance represents the energy of the model to be validated, reflecting the degree of fluctuation of that energy; The variance representing the baseline energy serves as a reference benchmark; when As the value increases, the proportion of the variance of the method to be verified relative to the benchmark decreases, reflecting an improvement in the stability of the method.

[0206] Step S4, Graph Convolutional Network Model Training and Prediction: Based on the hierarchical directed graph structure, a spatial domain graph convolutional network prediction model is constructed, specifically including:

[0207] Step S41, Construction of the overall adjacency matrix: Combine all nodes and edges to construct the complete adjacency matrix. Adjacency matrix.

[0208] Step S42, Multi-level graph convolutional design: Use an L-layer graph convolutional network, with each layer containing the following operations:

[0209] Graph convolution operations: ,

[0210] in, Ensure that the contributions of nodes of different degrees are more balanced when aggregating neighbor features; Each node aggregates its own features and those of its neighbors; then multiplies these features by a weight matrix. The nonlinear representation of the learning features is used to obtain the output of this layer. .

[0211] Step S43, Temporal Dimension Feature Extraction: Insert a one-dimensional convolutional layer between spatial domain GCN layers to extract temporal features.

[0212] Step S44, Model Training: Design the loss function by combining mean squared error and smoothness constraints:

[0213] ,

[0214] in, The loss function for the entire model measures the deviation between the predicted and the true values, and the goal is to minimize this value. The total number of samples in the training set; For sample index; The time step included for each sample; For time step index; For the first The sample at the th The actual observations at each time step; For the first The sample at the th Model predictions at each time step; This is the regularization coefficient, used to balance the weights of prediction error and smoothness penalty.

[0215] Step S45, Real-time prediction: For a new prediction time, repeat steps S1-S3 to generate real-time input samples and spatiotemporal correlation tensors; and input them into the trained energy output prediction model to obtain the initial prediction value.

[0216] Step S5: Constructing a load forecasting model: Based on the power forecasting logic, this extends to the coordinated load forecasting of transformer substations. First, when constructing the spatiotemporal information graph of load in interconnected transformer substations, the load of each interconnected substation is considered as a graph node. A dynamic graph structure integrating topology, load characteristics, and external influencing factors is constructed to lay the foundation for spatiotemporal feature extraction. Subsequently, GCN is used to extract the spatial features between load nodes, while GRU is used to extract the temporal features of historical load data. Based on this, a load forecasting model based on TGCN is constructed. A federated learning framework is used to train the load forecasting model with historical data to optimize parameters. After training, the latest transformer substation load sequence and a fixed topological adjacency matrix are input, and the load forecasting model can synchronously output the load forecast values ​​of all interconnected transformer substations at multiple future times, achieving system-level coordinated multi-step forecasting. Details are as follows:

[0217] Constructing a spatiotemporal load information map of interconnected distribution transformer areas: This involves mapping multiple interconnected distribution transformer areas to be predicted, such as... Figure 5As shown, the left side is a load spatiotemporal information diagram containing the first T time points; the middle is the TGCN prediction stage; and the right side is a load spatiotemporal information diagram containing the next P time points. Based on the electrical connection relationships of distribution substations provided by the distribution automation system or the power grid geographic information system, an adjacency matrix reflecting the spatial coupling relationship between substations is constructed. Based on the historical load data of each substation node, a feature matrix is ​​constructed to form a graph structure.

[0218] Spatiotemporal feature extraction: GCN is used to extract spatial features between nodes, and GRU is used to extract temporal features of historical load data.

[0219] Model building and training: such as Figure 6 A TGCN is constructed, and the spatiotemporal prediction model is trained using historical data to optimize the model parameters. After training, the latest transformer area load sequence and a fixed topological adjacency matrix are input, and the model can simultaneously output the load prediction values ​​of all interconnected transformer areas at multiple future time points, realizing system-level collaborative multi-step prediction.

[0220] Step S6, Prediction and Evaluation: The prediction results are evaluated using the root mean square error (RMSE), mean absolute error (MAE), and mean absolute percentage error (MAPE) to verify the model's load prediction performance for interconnected station intervals.

[0221] Simulation results:

[0222] like Figure 3 As shown, the RMSE and MAE of the proposed wind farm total power prediction are compared with those of nine other benchmark methods. The prediction curves of the True Value, the Proposed Method, and the other nine benchmark methods (Benchmark 1-9) are displayed. Figure 3 As can be seen, the prediction results of the method proposed in this application are closer to the actual effect in terms of both the trend and the amplitude. Figure 4The paper presents the RMSE and MAE of all methods across different forecast horizons in the test set. The proposed method exhibits the lowest error across each forecast horizon. Specifically, (a) shows a comparison of the RMSE of the proposed method with other benchmarks 1-9 for a 1-hour power forecast horizon; (b) shows a comparison of the MAE of the proposed method with other benchmarks 1-9 for a 1-hour power forecast horizon; (c) shows a comparison of the RMSE of the proposed method with other benchmarks 1-9 for a 2-hour power forecast horizon; (d) shows a comparison of the MAE of the proposed method with other benchmarks 1-9 for a 2-hour power forecast horizon; (e) shows a comparison of the RMSE of the proposed method with other benchmarks 1-9 for a 3-hour power forecast horizon; (f) shows a comparison of the MAE of the proposed method with other benchmarks 1-9 for a 3-hour power forecast horizon; (g) shows a comparison of the RMSE of the proposed method with other benchmarks 1-9 for a 4-hour power forecast horizon; and (h) shows a comparison of the RMSE of the proposed method with other benchmarks 1-9 for a 5-hour power forecast horizon. A diagram showing the MAE comparison between Method and other Benchmarks 1-9.

[0223] like Figure 8 As shown, (a) is a schematic diagram comparing the actual load, the load forecast values ​​obtained by the proposed model method, and other model methods in the one-step forecast; (b) is a schematic diagram comparing the actual load, the load forecast values ​​obtained by the proposed model method, and other model methods in the two-step forecast; (c) is a schematic diagram comparing the actual load, the load forecast values ​​obtained by the proposed model method, and other model methods in the three-step forecast; it can be seen that the load forecasting model proposed in this application has a significant improvement in forecasting accuracy, reaching over 90% accuracy in the three-step forecast. Combined with Table 2 and... Figure 7 The longitudinal comparison results of different models show that this load forecasting model has demonstrated excellent performance in both two-step and three-step forecasting tasks.

[0224] Table 2

[0225]

[0226] The above results demonstrate that the distributed energy output prediction model proposed in this application performs excellently in predicting the power of wind farm clusters based on interconnected distribution areas. The GCN and GRU modules integrated in the proposed load prediction model can accurately mine the spatiotemporal features contained in the power load data, maintaining high prediction accuracy even in multi-step prediction scenarios. Moreover, the accuracy improvement of the model is more significant as the number of prediction steps increases.

[0227] Example 2

[0228] This embodiment provides a distributed energy output and load forecasting device for interconnected power distribution areas, including:

[0229] Data acquisition module: used to acquire the energy data and load data to be measured for each interconnected power distribution area; wherein, the energy data to be measured includes the power generation resource data and power generation data of the distributed energy nodes in each interconnected power distribution area; the load data to be measured includes the historical load sequence data and topology information of each interconnected power distribution area;

[0230] Energy output prediction module: used to input the energy data to be measured into a pre-trained energy output prediction model to obtain the total power prediction value of distributed energy nodes in each interconnected area; wherein, the energy output prediction model constructs a hierarchical directed graph based on the energy data to be measured and extracts the dynamic spatiotemporal correlation features between distributed energy nodes, and integrates the hierarchical directed graph and the dynamic spatiotemporal correlation features to predict the total power; each node of the hierarchical directed graph represents the power generation resource data, power generation data and the total power prediction value respectively, and the hierarchical directed graph contains the causal relationship between the nodes;

[0231] Load forecasting module: This module constructs a load spatiotemporal information map based on the topology information and historical load sequence data. The load spatiotemporal information map is then input into a pre-trained load forecasting model to obtain load forecast values ​​for each interconnected transformer area at multiple future time points. The load forecasting model extracts dynamic spatial features between nodes in the load spatiotemporal information map and extracts time-dependent features based on historical load sequence data. After fusing the dynamic spatial features and time-dependent features, multi-step load forecasting is performed. Each node in the load spatiotemporal information map represents the overall load of each interconnected transformer area.

[0232] The energy output prediction model and the load prediction model are both trained collaboratively using a hierarchical federated learning framework.

[0233] The device provided in this embodiment can execute the distributed energy output and load prediction method for interconnected substations provided in any step of Embodiment 1, and has the corresponding functional modules and beneficial effects of the execution method.

[0234] Example 3

[0235] This embodiment provides a computer storage medium storing a computer program. When the computer program is executed by a processor, it implements the distributed energy output and load forecasting method for interconnected substations as provided in any step of Embodiment 1.

[0236] Those skilled in the art will understand that embodiments of this application can be provided as methods, systems, or computer program products. Therefore, this application can take the form of a completely hardware embodiment, a completely software embodiment, or an embodiment combining software and hardware aspects. Furthermore, this application can take the form of a computer program product embodied on one or more computer-usable storage media (including but not limited to disk storage, CD-ROM, optical storage, etc.) containing computer-usable program code.

[0237] This application is described with reference to flowchart illustrations and / or block diagrams of methods, apparatus (systems), and computer program products according to embodiments of this application. It will be understood that each block of the flowchart illustrations and / or block diagrams, and combinations of blocks in the flowchart illustrations and / or block diagrams, can be implemented by computer program instructions. These computer program instructions can be provided to a processor of a general-purpose computer, special-purpose computer, embedded processor, or other programmable data processing apparatus to produce a machine, such that the instructions, which execute via the processor of the computer or other programmable data processing apparatus, generate instructions for implementing the flowchart... Figure 1 One or more processes and / or boxes Figure 1 A device that provides the functions specified in one or more boxes.

[0238] These computer program instructions may also be stored in a computer-readable storage medium that can direct a computer or other programmable data processing device to function in a particular manner, such that the instructions stored in the computer-readable storage medium produce an article of manufacture including instruction means, which are implemented in a process Figure 1 One or more processes and / or boxes Figure 1 The function specified in one or more boxes.

[0239] These computer program instructions may also be loaded onto a computer or other programmable data processing equipment to cause a series of operational steps to be performed on the computer or other programmable equipment to produce a computer-implemented process, thereby providing instructions that execute on the computer or other programmable equipment for implementing the process. Figure 1 One or more processes and / or boxes Figure 1 The steps of the function specified in one or more boxes.

[0240] The embodiments of the present invention have been described above with reference to the accompanying drawings. However, the present invention is not limited to the specific embodiments described above. The specific embodiments described above are merely illustrative and not restrictive. Those skilled in the art can make many other forms under the guidance of the present invention without departing from the spirit and scope of the claims. All of these forms are within the protection scope of the present invention.

Claims

1. A method for forecasting the output and load of distributed energy resources in interconnected power distribution areas, characterized in that, include: Acquire the energy data and load data to be measured for each interconnected power distribution area; wherein, the energy data to be measured includes the power generation resource data and power generation data of the distributed energy nodes in each interconnected power distribution area; the load data to be measured includes the historical load sequence data and topology information of each interconnected power distribution area; The energy data to be measured is input into a pre-trained energy output prediction model to obtain the total power prediction value of distributed energy nodes in each interconnected area. The energy output prediction model constructs a hierarchical directed graph based on the energy data to be measured and extracts dynamic spatiotemporal correlation features between distributed energy nodes. The hierarchical directed graph and the dynamic spatiotemporal correlation features are then fused to predict the total power. Each node in the hierarchical directed graph represents power generation resource data, power generation data, and the total power prediction value, respectively. The hierarchical directed graph contains the causal relationships between the nodes. A load spatiotemporal information map is constructed based on the aforementioned topology information and historical load sequence data. This load spatiotemporal information map is then input into a pre-trained load prediction model to obtain load prediction values ​​for each interconnected transformer area at multiple future time points. The load prediction model extracts dynamic spatial features between nodes in the load spatiotemporal information map and extracts time-dependent features based on historical load sequence data. Multi-step load prediction is then performed by fusing the dynamic spatial features and time-dependent features. Each node in the load spatiotemporal information map represents the overall load of each interconnected transformer area. The energy output prediction model and the load prediction model are both trained collaboratively using a hierarchical federated learning framework.

2. The distributed energy load forecasting method for interconnected distribution areas according to claim 1, characterized in that, The power generation resource data of each distributed energy node is defined as the energy node, the power generation data as the energy power node, and the total power prediction of the interconnected power grid area is defined as the target node. The layered directed graph includes a first directed edge, a second directed edge, an energy input layer, an energy power input layer, and a target node; The energy input layer contains the same number of energy nodes as the distributed energy nodes, the energy power input layer contains the same number of energy power nodes as the distributed energy nodes, and the number of target nodes is one. Each energy node in the energy input layer is connected to the corresponding energy power node in the energy power input layer through a first directed edge, representing a one-way causal relationship from power generation resource data to power generation data; wherein, the weight of the first directed edge is calculated based on the causal strength between power generation resource data and power generation data; Each energy power node in the energy power input layer is connected to the target node through a second directed edge, representing the convergence relationship between power generation data and total power prediction value; wherein, the weight of the second directed edge is calculated based on the contribution and correlation of the power generation node to the total power prediction value.

3. The method for predicting distributed energy load in interconnected distribution areas according to claim 2, characterized in that, The method for obtaining the dynamic spatiotemporal correlation features includes: Construct the initial spatial correlation matrices for energy nodes and energy power nodes respectively; An improved normalized mutual information model incorporating kernel density estimation and skewness correction is used to nonlinearly enhance and correct the distance decay of the initial spatial correlation matrix. The enhanced and corrected correlation matrices of consecutive moments are stacked along the time dimension, and time decay weights and derived features are incorporated to form dynamic spatiotemporal correlation features.

4. The method for predicting distributed energy load in interconnected distribution areas according to claim 1, characterized in that, The energy output prediction model uses a spatial domain graph convolutional network to fuse hierarchical directed graphs and dynamic spatiotemporal correlation features to predict total power. The spatial domain graph convolutional network includes multi-layer graph convolutional residual blocks and a temporal attention convolution and gated loop module serially connected between each adjacent graph convolutional residual block; The graph convolutional residual block is used for feature propagation based on the adjacency matrix of the hierarchical directed graph; wherein the adjacency matrix of the hierarchical directed graph is obtained by normalizing the edge weights of the hierarchical directed graph. The graph convolutional residual block includes an edge attention mechanism and a node attention mechanism. The edge attention mechanism is used to multiply the edge attention weights of the dynamic spatiotemporal correlation features with the adjacency matrix element by element-wise multiplication to fuse the hierarchical directed graph and the dynamic spatiotemporal correlation features. The node attention mechanism is used to adaptively adjust the importance of each node feature in the hierarchical directed graph. The temporal attention convolution is used to extract local temporal dependency features of each node in the hierarchical directed graph through dynamic convolution kernels and dilation rates. The gated loop module uses a gating mechanism to control the information flow of each node feature in the hierarchical directed graph.

5. The method for predicting distributed energy load in interconnected distribution areas according to claim 1, characterized in that, The load prediction model includes a sequentially connected time-graph convolutional network and an output layer; The temporal graph convolutional network is used to fuse dynamic spatial features and temporal dependent features to obtain spatiotemporal fusion features, including parallel graph convolutional networks and gated recurrent units; The graph convolutional network is used to extract spatial features between load nodes based on the load spatiotemporal information graph. The gated loop unit is used to extract time-dependent features based on the historical load sequence data of each load node; The output layer is used to obtain multi-step load prediction values ​​based on spatiotemporal fusion feature mapping.

6. The method for predicting distributed energy load in interconnected distribution areas according to claim 5, characterized in that, The graph convolutional network includes a serially connected two-layer graph convolutional structure and an adaptive weight adjustment unit; The dual-layer graph convolutional structure is used to extract spatial features between load nodes based on the adjacency matrix of the load spatiotemporal information graph. The adjacency matrix of the load spatiotemporal information graph is obtained by normalizing electrical distance, geographical distance, load correlation, and time-dynamic weights. The electrical distance is calculated based on the inter-station impedance, the geographical distance is the straight-line distance between the inter-stations, the load correlation is determined by fusing improved normalized mutual information with the Pearson correlation coefficient, and the time-dynamic weights are calculated based on the load peak ratio and the correlation coefficient with ambient temperature. The adaptive weight adjustment unit is used to introduce spatial attention weights to adjust the spatial feature extraction process.

7. The method for predicting distributed energy load in interconnected distribution areas according to claim 1, characterized in that, The load nodes include historical load sequences and higher-order derived features; The higher-order derived features are obtained by standardizing the peak load, ambient temperature, and temperature correlation coefficient.

8. The method for predicting distributed energy load in interconnected distribution areas according to claim 1, characterized in that, The hierarchical federated learning framework includes: Each distributed energy node optimizes its local model parameters based on local data and uploads them to the coordinator on the corresponding interconnected area side for regional aggregation, generating a regional-level global model for each interconnected area. Each interconnected station area coordinator uploads the corresponding regional-level global model parameters to the global coordinator for global aggregation, generates a global model, and then distributes the global model parameters to the corresponding regional-level global model of each interconnected station area. The global model parameters are distributed to the local models of each distributed energy node through the regional coordinator.

9. A distributed energy output and load forecasting device for interconnected transformer substations, characterized in that, include: Data acquisition module: used to acquire the energy data and load data to be measured for each interconnected power distribution area; wherein, the energy data to be measured includes the power generation resource data and power generation data of the distributed energy nodes in each interconnected power distribution area; the load data to be measured includes the historical load sequence data and topology information of each interconnected power distribution area; Energy output prediction module: used to input the energy data to be measured into a pre-trained energy output prediction model to obtain the total power prediction value of distributed energy nodes in each interconnected area; wherein, the energy output prediction model constructs a hierarchical directed graph based on the energy data to be measured and extracts the dynamic spatiotemporal correlation features between distributed energy nodes, and integrates the hierarchical directed graph and the dynamic spatiotemporal correlation features to predict the total power; each node of the hierarchical directed graph represents the power generation resource data, power generation data and the total power prediction value respectively, and the hierarchical directed graph contains the causal relationship between the nodes; Load forecasting module: This module constructs a load spatiotemporal information map based on the topology information and historical load sequence data. The load spatiotemporal information map is then input into a pre-trained load forecasting model to obtain load forecast values ​​for each interconnected transformer area at multiple future time points. The load forecasting model extracts dynamic spatial features between nodes in the load spatiotemporal information map and extracts time-dependent features based on historical load sequence data. After fusing the dynamic spatial features and time-dependent features, multi-step load forecasting is performed. Each node in the load spatiotemporal information map represents the overall load of each interconnected transformer area. The energy output prediction model and the load prediction model are both trained collaboratively using a hierarchical federated learning framework.

10. A computer storage medium having a computer program stored thereon, characterized in that, When the computer program is executed by the processor, it implements the distributed energy output and load forecasting method for interconnected transformer areas as described in any one of claims 1-8.

Images