A method and system for predicting net load of a power distribution network with high proportion of new energy access

Abstract

The application discloses a kind of high proportion new energy access power distribution network's net load prediction method and system, method includes: collecting power distribution network historical operation data and regional meteorological data and preprocessing;The causal effect intensity between different characteristics in the data after being calculated is determined the final causal adjacency matrix;Net load prediction model is built, including multilayer trend feature extraction module, multilayer periodic feature coding module and feature fusion coding module;Through net load prediction model, predict load power, photovoltaic and wind power output, net load power;Based on the prediction result, total loss function is calculated and net load prediction model is trained;Through the net load prediction model of well-trained, predict the net load power of future time.This application introduces causal mechanism modeling, trend and cycle decomposition fusion and physical constraint joint optimization mechanism, significantly improves the accuracy, stability and engineering practicability of net load prediction under the scene of high proportion new energy access power distribution network.

Description

Technical Field

[0001] This invention belongs to the field of power load forecasting technology, specifically relating to a net load forecasting method and system for a high proportion of renewable energy access to the distribution network. Background Technology

[0002] With the accelerated pace of the global energy transition, the penetration rate of renewable energy sources such as photovoltaics and wind power in the power system continues to rise, and in some regional power grids, new energy has even become the main power source. This transformation has profoundly changed the load characteristics of the power system, and a new type of load—net load—has gradually become a key research object in power grid operation analysis. Net load refers to the difference between the total load power on the user side and the output of distributed renewable energy at a given moment, and its changing patterns directly determine the power grid personnel's analysis of the power grid system's operation. Because the power output of renewable energy sources such as photovoltaics and wind power is strongly influenced by meteorological conditions (such as solar irradiance, temperature, wind speed, wind direction, and cloud cover), it exhibits significant randomness, volatility, and intermittency. This characteristic causes the user-side net load curve to show new changing patterns rarely seen in traditional power systems, such as the "duck curve" and "anti-peak-shaving characteristics." Especially in scenarios with a high proportion of renewable energy connected to the grid, traditional load forecasting methods struggle to accurately capture the rapid changes in net load, which may lead to system operation risks and even affect the stability and security of the power grid.

[0003] In recent years, deep learning technology has made some progress in the field of load forecasting. Models such as Long Short-Term Memory (LSTM) networks and Gated Recurrent Units (GRUs) can capture long-term dependencies in time series, while the Transformer model has shown advantages in processing long sequences due to its self-attention mechanism.

[0004] Currently, existing load forecasting methods include: Patent CN118195103A provides a Transformer-based integrated load forecasting method and system. This method obtains trend features and seasonal features through multi-scale decomposition operations. After the decomposed features are divided into sequences, different models are used to process them separately. That is, the relatively stable trend features are input into the linear model, and the more volatile seasonal features are input into the Transformer model, thereby improving the accuracy of multi-load forecasting in regional integrated resource systems. Chinese patent CN118572664A discloses a multi-node short-term power load forecasting method based on the fusion of MST_GCN and Transformer. It uses an attribute enhancement unit to enhance the input historical spatiotemporal load matrix and extract the spatiotemporal load matrix and adjacency matrix. The load is then fed into the MST-GCN model to mine the short-term spatiotemporal features of each node's load, and into the Transformer to mine the long-term temporal features of each node's load. Finally, the short-term spatiotemporal features and long-term temporal features are integrated by a gating fusion module as the output of the prediction, which can effectively improve the accuracy and stability of multi-node load prediction. Meanwhile, CN119476605A provides a spatiotemporal load prediction method for electric vehicle battery swapping stations under different couplings based on the GCN-Transformer model. It uses GCN to process the load spatial features of battery swapping stations in different geographical locations and uses Transformer to process the load temporal features of battery swapping stations in different geographical locations, thereby integrating spatial and temporal features. This can capture the complex spatiotemporal dependencies between swapping stations and enhance the accuracy of load prediction.

[0005] However, the loads studied in the above studies are all single-characteristic loads, and the load data itself is not coupled with other types of power information. In high-proportion renewable energy power systems, loads become net loads, and are affected by fluctuations in renewable energy output, exhibiting strong non-stationarity and multi-time-scale coupling characteristics. The difficulty in predicting them lies in two aspects: firstly, there are dynamic and nonlinear coupling relationships between load, renewable energy, and other multi-source characteristics, and existing methods do not adequately model such dynamic characteristic associations; secondly, net load exhibits multi-time-scale temporal fluctuation patterns during renewable energy fluctuations, and existing models struggle to account for both long-term and short-term dependencies, leading to distortion of the prediction curve during periods of rapid change. How to enhance the modeling of dynamic coupling between features and strengthen multi-scale temporal dependency learning to improve the prediction morphology fidelity of the net load sequence during periods of severe renewable energy fluctuations has become an urgent technical problem to be solved. Summary of the Invention

[0006] To address the technical problems in existing technologies, such as the difficulty in effectively modeling the dynamic coupling relationships and multi-scale temporal dependencies among multiple features in the net load sequence, leading to distorted prediction curves and insufficient accuracy during periods of severe power output fluctuations, this invention provides a net load prediction method and system for high-proportion renewable energy access to the distribution network. This method can achieve high-precision prediction of net load by fully considering multi-source information such as photovoltaic and wind power output and meteorological characteristics. The method includes: collecting and preprocessing historical operating data of the distribution network and regional meteorological data; calculating the causal effect strength between different features in the preprocessed data to determine the final causal adjacency matrix; constructing a net load prediction model, including a multi-layer trend feature extraction module, a multi-layer periodic feature encoding module, and a feature fusion encoding module; predicting load power, photovoltaic and wind power output, and net load power using the net load prediction model; calculating the total loss function based on the prediction results and training the net load prediction model; and predicting the net load power at future times using the trained net load prediction model. This invention significantly improves the accuracy, stability, and engineering practicality of net load forecasting in scenarios with a high proportion of renewable energy access to the distribution network by introducing causal mechanism modeling, trend and periodic decomposition fusion, and physical constraint joint optimization mechanism.

[0007] The present invention adopts the following technical solution: A second aspect of the present invention provides a net load forecasting method for high-proportion renewable energy access to the distribution network, comprising: S1. Collect historical operating data of the power distribution network and regional meteorological data of the area where the power distribution network is located, and preprocess them; S2. The preprocessed data is used to generate an observation time series data matrix for each feature through a predefined causal mechanism model. Based on the parameters of the causal mechanism model, the strength of the causal effect between different features is determined, and a causal adjacency matrix is ​​generated. With the goal of minimizing the difference between the results generated by the causal mechanism model and the actual results, as well as the sum of the elements of the causal adjacency matrix, the parameters of the causal mechanism model are solved to determine the final causal adjacency matrix. S3. Construct a net load forecasting model, including a multi-layer trend feature extraction module, a multi-layer periodic feature encoding module, and a feature fusion encoding module; perform sequence decomposition on the preprocessed data in S1 to obtain periodic components and trend components; sequentially perform encoding embedding and position embedding on the periodic components and trend components to generate trend position embedding and periodic position embedding; input the trend position embedding and causal adjacency matrix into the multi-layer trend feature extraction module to generate trend features; input the periodic position embedding into the multi-layer periodic feature encoding module to output periodic encoded features; input the trend features and periodic encoded features into the feature fusion encoding module to predict load power, photovoltaic and wind power output, and net load power; S4. Based on the predicted vectors of load power, photovoltaic output, and wind power output, calculate the estimated vector of net load power; based on the predicted vector and estimated vector of net load power, calculate the total loss function and train the net load prediction model; predict the net load power at future times using the trained net load prediction model.

[0008] Preferably, the process of generating the causal adjacency matrix in S2 is as follows: The preprocessed historical operational data and regional meteorological data are concatenated to form an observation time-series data matrix. For any feature in the observation time-series data matrix, based on a predefined causal mechanism model, the observation time-series data matrices of other features are mapped to the observation time-series data matrix corresponding to that feature. A causal adjacency matrix is ​​defined, where each element represents the causal effect strength of the feature corresponding to the column index on the feature corresponding to the row index. This matrix is ​​obtained by calculating the L2 norm of the column corresponding to the row index in the first-level weight matrix of the causal mechanism model corresponding to the column index.

[0009] Preferably, the process of determining the final causal adjacency matrix in S2 is as follows: For each feature, the L2 norm is calculated by subtracting the output of the corresponding causal mechanism model from the observation time series data matrix corresponding to that feature; the L1 norm is calculated for the causal adjacency matrix; the squares of the L2 norm calculation results of all features are summed and weighted with the L1 norm calculation result; the weighted result is divided by the time length of the collected data as the optimization objective; when the graph corresponding to the causal adjacency matrix is ​​a directed acyclic graph, the parameters of the causal mechanism model for different features are solved by minimizing the optimization objective; the final causal adjacency matrix is ​​determined by substituting the solved parameters.

[0010] Preferably, the process of performing sequence decomposition on the preprocessed data from S1 in S3 is as follows: Average pooling is performed on the preprocessed historical operational data and regional meteorological data to obtain the historical operational trend component and the regional meteorological trend component. The corresponding trend component is subtracted from the preprocessed historical operational data and regional meteorological data to obtain the historical operational cycle component and the regional meteorological cycle component. The historical operational trend component and the regional meteorological trend component are concatenated to obtain the trend component. The historical operational cycle component and the regional meteorological cycle component are concatenated to obtain the cycle component.

[0011] Preferably, the specific process of encoding embedding and position embedding in S3 is as follows: Using a sliding window, the trend component and the periodic component are converted into trend time series features and periodic time series features of fixed time length, respectively. Each time series feature includes feature values ​​within the time window corresponding to different time segments. Using an encoding embedding function, the trend time series features and the periodic time series features are mapped to a predetermined dimension space, respectively, to obtain trend encoding embedding and periodic encoding embedding. For any feature in trend encoding embedding and periodic encoding embedding, the encoding embeddings of different time segments of the corresponding feature are input into the position embedding unit to obtain the hidden state of each time segment; the hidden states of all time segments are concatenated and added to the encoding embedding of the corresponding feature to obtain the position embedding of the corresponding feature; for trend component and periodic component, the position embeddings of all features in the corresponding component are concatenated to obtain trend position embedding and periodic position embedding respectively.

[0012] Preferably, the process of generating trend features in S3 is as follows: A parameter matrix is ​​randomly initialized, and scaling and activation operations are performed on the parameter matrix to obtain the node features of the corresponding node; the antisymmetric matrix of the node features is calculated, and the initial adjacency matrix of the node is calculated through the activation function; the initial adjacency matrix of the node is pruned to obtain the corresponding sparse adjacency matrix. The causal adjacency matrix and the sparse adjacency matrix are standardized using a symmetrically normalized Laplace matrix. The transposed trend position is embedded as the input to the multi-layer trend feature extraction module. For any layer in the multi-layer trend feature extraction module, the output feature matrix of the current layer is generated based on the input feature matrix, the standardized causal adjacency matrix, and the sparse adjacency matrix of the current layer. The output feature matrices of all layers are accumulated after passing through a linear layer to obtain the trend features.

[0013] Preferably, the process of generating the output feature matrix of the current layer is as follows: The input feature matrix of the current layer is multiplied by a predefined gating weight matrix and then input into the activation function to obtain the gating matrix of the current layer. The standardized causal adjacency matrix is ​​multiplied by the input feature matrix of the current layer and then multiplied element-wise by the gating matrix of the current layer to obtain the static trend feature. The standardized sparse adjacency matrix is ​​multiplied by the input feature matrix of the current layer and then multiplied element-wise by 1 minus the gating matrix of the current layer to obtain the dynamic trend feature. The sum of the static trend feature and the dynamic trend feature of the current layer is used as the static-dynamic fusion feature of the current layer. Based on the predefined modulation parameters, the static-dynamic fusion feature of the current layer and the input feature matrix are weighted to obtain the output feature matrix of the current layer, which is also the input feature matrix of the next layer.

[0014] Preferably, the process of predicting distribution network operation data under different tasks in S3 is as follows: For any layer in the multi-layer periodic feature encoding module, the periodic encoding features output by the previous layer are input into a predefined multi-head time-sensitive attention mechanism to obtain the periodic attention features of the corresponding layer; the periodic attention features are added to the periodic encoding features output by the previous layer and then passed through the layer normalization function and the feedforward neural network in sequence to output the periodic encoding features output by the current layer; after passing through the multi-layer periodic feature encoding module with a predetermined number of layers, the final periodic encoding features are output. The periodic coding features and trend features are added together and then input into the feature fusion coding module to obtain the fused coding features. The fused coding features are then flattened in dimensions and then output as prediction vectors for load power, photovoltaic power output, wind power output and net load power through multiple prediction heads.

[0015] Preferably, the process of calculating the periodic attention features of different layers is as follows: The periodic encoding features output from the previous layer are input into the linear layer to generate n sets of query vectors, key vectors, and value vectors; a set of periods to be modeled is defined, and each period in the set is assigned to a different attention head. The assigned attention head is used as a non-free head, and the remaining attention heads are used as free heads. For a non-free head, define a time-sensitive mask matrix. For each position in the time-sensitive mask matrix, determine whether the time difference between the query vector and the key vector at the corresponding position in the non-free head is an integer multiple of the period of the corresponding non-free head. If yes, the element value of the corresponding position in the time-sensitive mask matrix is ​​0; otherwise, it is infinity. Multiply the query vector of the corresponding non-free head by the transpose of the key vector, divide by the square root of the key vector dimension, add the corresponding time-sensitive mask matrix, and then pass it through an activation function to obtain the attention score of the corresponding non-free head. Multiply the attention score by the value vector to obtain the attention output of the corresponding non-free head. For each free head, calculate the self-attention of each attention head to obtain the corresponding attention output; concatenate the attention outputs of all attention heads and perform dimensional mapping through a linear layer to output periodic attention features.

[0016] Preferably, in S4, the predicted vectors of photovoltaic power output and wind power output are successively subtracted from the predicted vector of load power to obtain the estimated vector of net load power; the mean square error between the predicted vector and the estimated vector of net load power is calculated as the loss function of the relational term. The mean square error of the net load power prediction vector and the actual net load power is calculated and used as the conventional loss function; the relational loss function and the conventional loss function are weighted to obtain the total loss function.

[0017] A second aspect of the present invention provides a net load forecasting system for a high proportion of renewable energy access to a distribution network, and a net load forecasting method for a high proportion of renewable energy access to a distribution network, comprising: The data acquisition module collects historical operating data of the power distribution network and regional meteorological data of the area where the power distribution network is located, and preprocesses the data. The causal relationship determination module generates an observation time series data matrix for each feature by passing the preprocessed data through a predefined causal mechanism model; based on the parameters of the causal mechanism model, it determines the causal effect strength between different features and generates a causal adjacency matrix; with the goal of minimizing the difference between the results generated by the causal mechanism model and the actual results, as well as the sum of the elements of the causal adjacency matrix, it solves for the parameters of the causal mechanism model and determines the final causal adjacency matrix. The net load forecasting model construction module constructs a net load forecasting model, including a multi-layer trend feature extraction module, a multi-layer periodic feature encoding module, and a feature fusion encoding module. It performs sequence decomposition on the preprocessed data (S1) to obtain periodic and trend components. It then sequentially performs encoding embedding and position embedding on the periodic and trend components to generate trend position embeddings and periodic position embeddings. The trend position embeddings and the causal adjacency matrix are input into the multi-layer trend feature extraction module to generate trend features. The periodic position embeddings are input into the multi-layer periodic feature encoding module to output periodic encoded features. Finally, the trend features and periodic encoded features are input into the feature fusion encoding module to predict load power, photovoltaic and wind power output, and net load power. The net load forecasting module calculates the estimated vector of net load power based on the forecast vectors of load power, photovoltaic power output, and wind power output; it calculates the total loss function and trains the net load forecasting model based on the forecast vector and estimated vector of net load power; and it predicts the net load power at future times using the trained net load forecasting model.

[0018] A third aspect of the present invention provides a terminal, including a processor and a storage medium; The storage medium is used to store instructions; The processor is configured to operate according to the instructions to execute steps of a net load forecasting method based on a high proportion of new energy sources connected to the distribution network.

[0019] A fourth aspect of the invention provides a computer-readable storage medium having a computer program stored thereon, which, when executed by a processor, implements the steps of a method for net load forecasting with a high proportion of new energy sources connected to a distribution network.

[0020] Compared with the prior art, the beneficial effects of the present invention include at least the following: 1. This invention constructs a causal adjacency matrix based on a causal mechanism model, explicitly introducing the causal effect strength between different features into the net load forecasting model, effectively overcoming the limitation of traditional time series forecasting methods that rely solely on correlation modeling. By minimizing the joint optimization objective of the model output error and the sparse constraint term of the causal adjacency matrix, and solving the parameters under the constraint of a directed acyclic graph, the resulting causal structure possesses physical interpretability and structural stability. This improves the accuracy and robustness of load forecasting in scenarios with a high proportion of renewable energy access, maintaining high forecast accuracy even under conditions of severe weather fluctuations or large fluctuations in renewable energy output.

[0021] 2. This invention achieves the separate modeling of long-term trend changes and multi-scale periodic fluctuations by performing sequence decomposition of historical operational data and regional meteorological data into trend and periodic components, and constructing multi-layer trend feature extraction and multi-layer periodic feature encoding modules respectively. The trend feature extraction module combines causal adjacency matrices and sparse adjacency matrices, and achieves adaptive fusion of static causal relationships and dynamic structural relationships through a gating mechanism; the periodic feature encoding module uses a time-sensitive multi-head attention mechanism to model different periodic patterns in a targeted manner, thereby significantly enhancing the model's ability to express complex temporal features and improving the prediction stability and generalization ability in multi-period superposition scenarios.

[0022] 3. This invention introduces physical constraints on net load power during the training phase. The predicted results of load power, photovoltaic output, and wind power output are constructed into a net load estimation vector, which, together with the net load prediction vector, forms a relational term loss function. This function is then weighted with a conventional mean square error loss function to form the total loss function. This design embeds the physical constraints of energy balance in the power system into the model optimization process, effectively reducing inconsistencies between multi-task predictions and improving the coordination and physical rationality of multi-objective prediction results. This enhances the overall prediction model's application value and engineering feasibility in actual distribution network operation and scheduling. Attached Figure Description

[0023] Figure 1 This is a structural diagram of the net load prediction model provided by the present invention; Figure 2 This is a schematic diagram of the sequence decomposition process in an embodiment of the present invention; Figure 3 This is a schematic diagram of the timing constraint position encoding process in an embodiment of the present invention; Figure 4 This is a schematic diagram of GCN feature extraction in an embodiment of the present invention. Detailed Implementation

[0024] To make the objectives, technical solutions, and advantages of this invention clearer, the technical solutions of this invention will be clearly and completely described below with reference to the accompanying drawings of the embodiments of this invention. The embodiments described in this application are merely some embodiments of this invention, and not all embodiments. Based on the spirit of this invention, all other embodiments obtained by those skilled in the art without creative effort are within the protection scope of this invention.

[0025] Example 1 Embodiment 1 of the present invention provides a net load forecasting method for high-proportion renewable energy access to the distribution network, comprising the following steps: S1. Collect historical operating data of the power distribution network and regional meteorological data of the area where the power distribution network is located, and preprocess them.

[0026] Specifically, historical operating data of each node in the distribution network is collected, including load power, photovoltaic and wind power output, and net load power. Correlation analysis is used to screen key meteorological features related to net load power from regional meteorological data, such as humidity, irradiance, temperature, and air density. Time scale expansion is used to align historical operating data and the selected key meteorological features across different sampling frequencies. Interpolation methods are used to complete missing data in the time-aligned data. Outlier detection and correction are performed on the completed data based on statistical detection to improve data continuity and consistency, providing high-quality input for subsequent models.

[0027] S2. The preprocessed data is used to generate an observation time series data matrix for each feature through a predefined causal mechanism model. Based on the parameters of the causal mechanism model, the causal effect strength between different features is determined, and a causal adjacency matrix is ​​generated. With the goal of minimizing the difference between the results generated by the causal mechanism model and the actual results, as well as the sum of the elements of the causal adjacency matrix, the parameters of the causal mechanism model are solved to determine the final causal adjacency matrix.

[0028] The trend components corresponding to each distinct feature are used as corresponding feature nodes, and the latent feature representation is learned by a graph convolutional network. Traditional methods typically use symmetric indices (such as Euclidean metrics or dot products) to quantify node similarity, which inherently assume bidirectional relationships. However, this symmetry becomes problematic in net load forecasting because asymmetric causal relationships dominate load power, wind power, photovoltaic power, and weather factors. Wind power output exhibits a strong unidirectional dependence on weather conditions, while showing a negligible correlation with net load power. Conversely, wind power output has a significant impact on net load dynamics. Regional meteorological data remains strictly exogenous, entirely controlled by atmospheric processes, and remains unaffected by changes in net load. To explicitly encode these causal relationships, this embodiment employs the nonlinear NOTEARS algorithm to model causal discovery as a constrained continuous optimization problem. The nonlinear NOTEARS algorithm learns the causal structure from observational data through a continuous optimization framework, utilizes a neural network to parameterize the structural equation model (SEM), and transforms the directed acyclic graph (DAG) constraint into a differentiable algebraic penalty.

[0029] Specifically, the preprocessed historical operational data and regional meteorological data are stitched together to form an observation time series data matrix. ,in The sample size (in this embodiment, the corresponding time length L for data collection) is the sample size. For feature dimensions; For any feature in the observation time series data matrix, based on a predefined causal mechanism model, the observation time series data matrices of other features are mapped to the observation time series data matrix corresponding to that feature; a causal adjacency matrix is ​​defined, where each element represents the causal effect strength of the feature corresponding to the column index on the feature corresponding to the row index, which is obtained by performing L2 calculation on the column corresponding to the row index in the first layer weight matrix of the causal mechanism model corresponding to the column index; For each feature, the L2 norm is calculated by subtracting the output of the corresponding causal mechanism model from the observation time series data matrix corresponding to that feature; the L1 norm is calculated for the causal adjacency matrix; the squares of the L2 norm calculation results of all features are summed and weighted with the L1 norm calculation result; the weighted result is divided by the time length of the collected data as the optimization objective; when the graph corresponding to the causal adjacency matrix is ​​a directed acyclic graph, the parameters of the causal mechanism model for different features are solved by minimizing the optimization objective; the final causal adjacency matrix is ​​determined by substituting the solved parameters.

[0030] In this embodiment, it is assumed that the data generation process follows an additive noise model: ; In the formula, Indicates the first Features Second observation; Indicates the difference between features The observation time series data matrix corresponding to other features; As an independent noise term; For models of unknown causal mechanisms, This indicates that the function will Dimensional mapping to 1 dimension, express The algorithm employs a multilayer perceptron (MLP) to parameterize each dimension in order to model nonlinear relationships. , ,in , indicating the first All weights and bias parameters of a causal mechanism function; To quantify the strength of causal effects among features and construct an optimizable graph structure, a causal adjacency matrix is ​​defined. ,element By calculating the first The first layer weight matrix of the MLP ( (where the hidden layer dimension is the first) The 2-norm is obtained as follows: ; In the formula, Represents the first of the matrix Columns, corresponding to input variables Connection weights to hidden layers; This is the Euclidean norm, also known as the second norm; For the first The element value in the u-th row and k-th column of the first layer weight matrix of an MLP; Indicates the existence of features arrive The magnitude of the edge reflects the strength of the causal effect; To ensure that the learned graph is a directed acyclic graph (DAG), an acyclicity constraint based on matrix exponents is introduced; the constraint function is defined as follows: ; In the formula, This represents the Hadamard product (element-by-element multiplication). It is a matrix exponential function. The trace of the matrix (the sum of its diagonal elements); The theoretical basis for this constraint lies in the Taylor expansion of the matrix exponent. ,in Proportional to the length in the figure The number of closed paths, therefore If and only if For acyclic structures, solve the following constrained optimization problem: ; In the formula, For the parameter set of all MLPs, For sparse regularization coefficients, for Norms induce sparsity; The above problem is solved by using the augmented Lagrange multiplier method to obtain the parameters of each causal mechanism model, and the causal adjacency matrix between features is determined based on the parameters.

[0031] S3, Construction as follows Figure 1 The net load forecasting model shown includes a multi-layer trend feature extraction module, a multi-layer periodic feature encoding module, and a feature fusion encoding module. The preprocessed data (S1) is decomposed into a sequence to obtain periodic and trend components. The periodic and trend components are then encoded and embedded sequentially to generate trend-position embeddings and periodic-position embeddings. The trend-position embeddings and the causal adjacency matrix are input into the multi-layer trend feature extraction module to generate trend features. The periodic-position embeddings are input into the multi-layer periodic feature encoding module to output periodic encoded features. Finally, the trend features and periodic encoded features are input into the feature fusion encoding module to predict distribution network operation data under different tasks.

[0032] As a preferred implementation method, such as Figure 2 As shown, the temporal characteristics of the sequence fundamentally encapsulate the strong correlation between adjacent time points, manifesting as time patterns such as trend segments or periodic loads. Average pooling is performed on the preprocessed historical operational data and regional meteorological data to obtain the historical operational trend component and the regional meteorological trend component, representing the low-frequency or trend components in the sequence. Subtracting the corresponding trend components from the preprocessed historical operational data and regional meteorological data yields the historical operational cycle component and the regional meteorological cycle component, which can be considered as the irregular fluctuation components remaining after removing the trend. This process helps subsequent models to handle short-term fluctuations and long-term trends in the time series respectively. A time series with a certain characteristic... For example, by using kernels of various sizes, it can be decomposed into more predictable periodic and trend components; specifically represented as: ; In the formula, For the first part of the preprocessed input data i The original time series of each characteristic, such as pre-processed load power, photovoltaic output, wind power output, humidity, irradiance, temperature, etc. For the first i The first feature in the original time series L Data collected at each time point; LIndicates the length of the time series; Indicates the fill function; This is the average pooling function; For the first j The kernel size is set by the sub-average pooling method; J This represents the total number of pooling operations. For the first i The trend component of each feature; For the first i The periodic components of each feature; By concatenating the historical trend components and the regional meteorological trend components, the trend component is obtained. The historical operational cycle components and regional meteorological cycle components are spliced ​​together to obtain the cycle components. ; In this embodiment, It is the original time series of any feature in the preprocessed historical operating data and regional meteorological data, including load power, photovoltaic output, wind power output, net load power, and key meteorological features related to net load power.

[0033] As a preferred implementation, the specific processes of encoding embedding and positional embedding are as follows: Using a sliding window, the trend component and the periodic component are converted into trend time series features of a fixed time length, respectively. and periodic time series characteristics Each time-series feature contains feature values ​​within the time window corresponding to N time segments, which facilitates the subsequent model to extract local time features; Given a length of L sequence Divide it into N There are 3 non-overlapping time segments (Pach); the length of each Patch is 1. P Then the number of patches N for: ; Where S is the sliding step size; when the sliding step size is... S = P At that time, there is no overlap between patches; Using encoding embedding functions, trend time-series features and periodic time-series features are mapped to a predetermined dimension space, respectively, to obtain trend encoding embedding and periodic encoding embedding; in this embodiment, the length is... P Time segment feature sequence mapping to D In 3D space; let the number of input data features be... C Embedding dimension is D The specific formula is: ; ; In the formula, For trend encoding embedding after encoding embedding; This is a periodic encoding embedding after encoding embedding; Encode the embedding function; For parameter matrices; This is for embedding bias terms.

[0034] Because Transformers lack a recurrent structure, positional information must be explicitly injected to enable the model to perceive the temporal sequence order. Traditional sine and cosine positional encoding uses sine and cosine functions for odd and even dimensions respectively, leveraging the periodicity of trigonometric functions to allow the model to learn relative positional relationships. However, time-series data exhibits strong temporal continuity and periodicity. To enable the model to utilize the strict temporal order of time series, we consider using a multi-layer recurrent convolutional neural network (RNN) to inject enhanced positional contextual information into the sequence symbols. Instead of directly using the output of the multi-layer RNN as the final embedding, the result is added to the original sequence symbols, thus preserving the semantics while explicitly carrying strict temporal order information.

[0035] like Figure 3 As shown, for any feature in trend encoding embedding and periodic encoding embedding, the encoding embeddings of different time segments of the corresponding feature are input into the position embedding unit (in this embodiment, a multi-layer recurrent convolutional neural network unit) to obtain the hidden state of each time segment; after concatenating the hidden states of all time segments, they are added to the encoding embedding of the corresponding feature to obtain the position embedding of the corresponding feature; for trend components and periodic components, the position embeddings of all features in the corresponding components are concatenated to obtain the trend position embedding and periodic position embedding, respectively; the specific process of generating position embeddings for any feature in trend encoding embedding and periodic encoding embedding is as follows: ; in, For RNN functions, it can be any type of RNN model, such as RNN, LSTM, and GRU; R Indicates the number of layers in an RNN unit; The encoding embedding corresponding to the i-th feature is from the i-th feature. r The first in the layer j The hidden states obtained by the first RNN unit, and the first layer of the first RNN unit j Input of each RNN unit ; It is the first after encoding and embedding i The first feature j Trend encoding embedding or periodic encoding embedding for a time segment; Is it the first corresponding to trend encoding embedding or periodic encoding embedding? iThe final hidden state of each feature; It is the first i The positional embedding of each feature; for the trend component, the trend positional embedding is defined as... For periodic components, the periodic position embedding is defined as follows: ; This embodiment uses embeddings generated at previous positions to recursively generate embeddings at subsequent positions. The resulting position embeddings ensure positional differences at each time step and introduce strict temporal order information.

[0036] As a preferred implementation method, the process of generating trend features is as follows: During training, the causal relationships between different feature factors may change under different time scales and operating conditions. The statically predefined causal adjacency matrix obtained by S2 cannot self-correct the causal direction and connection strength between feature nodes according to the dynamic changes in data. To solve this problem, this embodiment introduces an adaptive graph. The structure of this graph (usually the adjacency matrix) is not pre-defined, but is automatically learned from the data during training. In this way, the model can simultaneously utilize prior knowledge (predefined causal adjacency matrix) and data-driven discovery patterns (adaptive graph), thereby capturing spatial correlations more comprehensively and flexibly.

[0037] First, a learnable parameter matrix is ​​randomly initialized. The parameter matrix is ​​scaled and activated to obtain the node features of the corresponding nodes; the antisymmetric matrix of the node features is calculated, and the initial adjacency matrix of the nodes is calculated using the activation function; the initial adjacency matrix of the nodes is pruned to obtain the corresponding sparse adjacency matrix; the specific formula is as follows: ; ; ; ; In the formula, and It is the parameter matrix of the graphical model. C is the dimension of the graph node embedding, and C is the number of features; Convert to intermediate parameter matrix and ; For a control from the original embedded E To the middle indicates M Factors affecting the intensity of information flow; The activation function is usually Tanh; Represents the linear rectification activation function; The superscript indicates the transpose of the matrix; The first row of the j-th row in the initial adjacency matrix k A set of indices of the maximum values; Returns the index of the maximum value in the vector. To reduce the computational cost of subsequent graph convolutions, the index of the first value is selected. k The nearest node is selected as the neighbor; while retaining the weights of connected nodes, the weights of non-connected nodes are set to zero, resulting in a sparse adjacency matrix. ; During the graph learning phase, a symmetric-normalized Laplacian matrix is ​​used to model the causal adjacency matrix. and sparse adjacency matrix Standardization processing; For example, the specific formula is: ; In the formula, It is the identity matrix. It is a causal adjacency matrix The degree matrix, It is the standardized causal adjacency matrix; A gating mechanism is employed to dynamically and selectively fuse information from predefined and adaptive graphs, flexibly adjusting the dependence on different graph structures based on specific input samples. When the information conveyed by the two graphs conflicts, gating can act as a reconciliation and selection mechanism. The final gating value can serve as an analytical metric to explain when the model relies more on prior knowledge and when it trusts patterns learned from the data more.

[0038] The transposed trend position is embedded as the input to the multi-layer trend feature extraction module. For any layer in the multi-layer trend feature extraction module, the input feature matrix of the current layer is multiplied by a predefined gate weight matrix and then input into the activation function to obtain the gate matrix of the current layer. The standardized causal adjacency matrix is ​​multiplied by the input feature matrix of the current layer and then multiplied element-wise by the gate matrix of the current layer to obtain the static trend feature. The standardized sparse adjacency matrix is ​​multiplied by the input feature matrix of the current layer and then multiplied element-wise by 1 minus the gate matrix of the current layer to obtain the dynamic trend feature. The sum of the static trend feature and the dynamic trend feature of the current layer is used as the static-dynamic fusion feature of the current layer. Based on the predefined modulation parameters, the static-dynamic fusion feature of the current layer and the input feature matrix are weighted to obtain the output feature matrix of the current layer, which is also the input feature matrix of the next layer. The specific formula is as follows: ; In the formula, Let v be the gating matrix of the v-th layer; It is the sigmoid activation function; Let be the learnable gating weight matrix of the v-th layer; For Hadamah accumulation; This is the standardized sparse adjacency matrix; It is the output feature matrix of the v-th layer; It is the output feature matrix of the v-th layer, which is also the input feature matrix of the (v-1)-th layer, and the input feature matrix of the first layer is the transpose of the trend position embedding. ; The static and dynamic fusion features of the vth layer; These are modulation parameters used to adjust the static and dynamic fusion features and the input feature matrix of the current layer; like Figure 4 As shown, the output feature matrices of all layers are accumulated after passing through a linear layer, that is, the information of multiple convolutional layers is aggregated to obtain the trend features: ; In the formula, This is the output of the multi-layer trend feature extraction module, i.e., the extracted trend features; This represents the total number of layers in the multi-layer trend feature extraction module. Let be the linear layer weight matrix of the v-th layer.

[0039] The Transformer is generally divided into two parts: encoder and decoder. However, this structure has high computational complexity, especially when processing long sequences or large-scale data, the computational burden increases significantly, resulting in a decrease in model training speed. To simplify the model architecture, this embodiment uses a Transformer structure that only retains the encoder module.

[0040] As a preferred implementation method, the process of predicting distribution network operation data under different tasks is as follows: For any layer in the multi-layer periodic feature encoding module, the periodic encoding features output by the previous layer are input into a predefined multi-head time-sensitive attention mechanism to obtain the periodic attention features of the corresponding layer. The periodic attention features are then added to the periodic encoding features output by the previous layer, and subsequently passed through a layer normalization function and a feedforward neural network to output the periodic encoding features of the current layer. Specifically, this is expressed as follows:

[0041] In the formula, For encoding layer functions; For layer normalization function; It is a feedforward neural network; The first layer's input is the periodic encoded feature output from the previous layer, and the input to the first layer is the periodic position embedding. After passing through a predetermined number of multi-layer periodic feature encoding modules, the final periodic encoded features are output. After the Transformer performs multi-layer stacked learning of time dependencies, and the GCN performs multi-layer stacked learning of interaction dependencies, the outputs are summed and then input into the Transformer network again to further enhance feature representation, thereby comprehensively extracting intra-sequence and inter-sequence dependencies in the payload data. Specifically, the periodic encoding features and trend features are added and then input into the feature fusion encoding module (using the Transformer network in this embodiment) to obtain the fused encoding features; specifically represented as follows: ; In the formula, The encoding layer function of layer E in the periodic feature encoding module is nested to output periodic encoded features; The feature matrix is ​​the sum of periodic coding features and trend features; To fuse coding features; The fused encoded features are flattened using the Flatten operation, and then multiple prediction heads are used to output prediction vectors for load power, photovoltaic output, wind power output, and net load power, respectively. The specific process can be represented as follows: ; In the formula, These are the fusion-encoded features after dimensionality reduction; The dimension flattening function; It is the first x Prediction vectors for each task (e.g., net load prediction); It maps flattened features to the first... x Learnable parameters output by each task; It is the first x Learnable biases for each task.

[0042] Because the scaling dot product attention mechanism processes time series information in parallel, each time segment uses the same calculation score. Considering the significant time-varying and periodic characteristics of power grid load data, a time-sensitive attention weight adjustment mechanism is introduced into the self-attention calculation process. Based on the predicted target time point, it automatically identifies historical segments in the historical sequence that are in the same time period or the same operating interval, adds attention to them, and masks attention for non-periodic time periods.

[0043] As a preferred implementation, the process of calculating the periodic attention features of different layers is as follows: The periodic encoded features output from the previous layer are input into the linear layer to generate... n Group query vectors, key vectors, and value vectors ( Q , K , V Specifically, it is expressed as: ; In the formula, These are the projection matrices of the y-th attention head, representing the query vector, key vector, and value vector, respectively. , These are the query vector, key vector, and value vector of the y-th attention head, respectively; Define a set of periodicity sets to be modeled ,Will o Each cycle is assigned to o Each attention head is assigned as a non-free attention head; if The remaining Each attention head is treated as a free head without explicit periodic constraints, allowing it to freely learn non-periodic or residual patterns in the data. For non-free heads, define a time-sensitive mask matrix. This mask matrix allows for time steps t and ( z Attention is focused between positive integers, while other connections are masked; for each position of the time-sensitive mask matrix, it is determined whether the time difference between the query vector and the key vector at the corresponding position in the non-free head is an integer multiple of the period of the corresponding non-free head; if so, the element value of the corresponding position in the time-sensitive mask matrix is ​​0; otherwise, it is infinity; the specific formula is expressed as:

[0044] In the formula, Let be the value of the element in the b-th row and c-th column of the time-sensitive mask matrix of the non-free head corresponding to the y-th period; , These represent the collection times for the indexes at the corresponding positions of the query vector and key vector, respectively. This is the y-th period; The attention score for the corresponding non-free head is obtained by multiplying the query vector of the corresponding non-free head by the transpose of the key vector, dividing by the square root of the key vector dimension, adding the corresponding time-sensitive mask matrix, and then applying the result through an activation function. This attention score is then multiplied by the value vector to obtain the attention output for the corresponding non-free head. Specifically: ; In the formula, An activation function transforms a vector into a probability distribution. For the first y The attention output of a non-free head; The dimension of the key vector; For a free head, calculate the self-attention of each attention head (i.e., without adding a time-sensitive mask matrix) to obtain the corresponding attention output; The attention outputs of all attention heads are concatenated, and then dimension-mapped back through a linear layer. D Dimension), output periodic attention features: ; In the formula, The parameter matrix of the linear layer; It is a matrix concatenation operation; The multi-head time-sensitive attention output assigned to the period, i.e., the periodic attention feature.

[0045] S4. Based on the predicted vectors of load power, photovoltaic output, and wind power output, calculate the estimated vector of net load power; based on the predicted vector and estimated vector of net load power, calculate the total loss function and train the net load prediction model; predict the net load power at future times using the trained net load prediction model.

[0046] Net load typically refers to the proportion of electrical load remaining after subtracting the output power of renewable energy from the total load (i.e., the total electricity consumed by all electrical equipment); this relationship can be mathematically expressed as:

[0047] In the formula, It is the function for calculating net load. , , and These represent net load, load, and power output from photovoltaic and wind power generation, respectively.

[0048] The relation loss module is built upon a neural network model. By introducing prior knowledge from observational data, empirical data, and physical laws into the network structure, relational constraints are imposed on the neural network's learning process. Through this design, the neural network not only relies on empirical learning from historical samples but also effectively utilizes the system's physical characteristics and energy conservation constraints to achieve highly reliable predictions of net load in complex renewable energy integration scenarios.

[0049] Specifically, the predicted vectors of photovoltaic power output and wind power output are successively subtracted from the predicted vector of load power to obtain the estimated vector of net load power; the mean square error between the predicted vector and the estimated vector of net load power is calculated and used as the loss function of the relational term. The specific design is as follows:

[0050] In the formula, For the relational term loss function; Forecast net load power; , and For predicted load power, photovoltaic and wind power output; This represents the total number of prediction points. The mean square error of the predicted net load power vector and the actual net load power is calculated and used as the conventional loss function. The relational loss function and the conventional loss function are weighted to obtain the total loss function. The following design is possible: ; In the formula, This represents a conventional loss function that measures the distance between the predicted output and the target output in a supervised learning manner. For the overall net load forecasting model; for t Time before L Input data of length; Indicates advance Actual net load power during the time period; This is the loss factor, used to control the weights of the relation terms; The net load prediction model is trained based on the total loss function to obtain the trained net load prediction model. Historical operational data and regional meteorological data of a predetermined length prior to the current moment are collected, preprocessed, and then input into a trained net load prediction model to predict net load power at future moments.

[0051] In this embodiment, the preprocessed dataset (including historical net load, meteorological characteristics, etc.) is first divided into a training set, a validation set, and a test set; a time-series partitioning method is used to ensure the temporal continuity of the training data and avoid future information leakage. Then, multidimensional data (net load, load, power generation, key meteorological variables) from the previous time window (such as the past 24 hours) are used as input to predict the net load value for a specific future period (such as the next 15 minutes to 24 hours) as output. During training, backpropagation and adaptive moment estimation (Adam) optimizers are used to iteratively update network weight parameters by minimizing the difference between predicted and true values. To prevent overfitting, a validation set is used for monitoring during training, and early stopping and regularization strategies (such as Dropout) are integrated. When the performance on the validation set no longer improves over several consecutive training epochs, the model is considered sufficiently trained and converged, the training process is terminated, and the optimal model parameters are saved.

[0052] This embodiment conducts multivariate time series forecasting analysis based on historical operation data of a regional power grid. Photovoltaic output and electricity load data collected in 2020 at 15-minute intervals were selected. During the feature selection phase, correlation analysis identified several predictive variables that significantly influence net load changes, including humidity, top-of-atmosphere irradiance, surface horizontal irradiance, air temperature, and air density. The normalized dataset was divided into training, validation, and test sets in a 7:1:2 ratio. The hyperparameters and optimization schemes in the experimental setup are detailed in Table 1. To evaluate the forecasting performance, three indicators—mean squared error (MSE), mean absolute percentage error (MAPE), and mean squared percentage error (MSPE)—were used, and the results were compared with existing benchmark methods.

[0053] Table 1. Hyperparameters for training the prediction model

[0054] The proposed net load prediction model is compared with benchmark methods such as Inverted Transformer (iTransformer), Transformer, LSTM-Attention, Non-stationary Linear (NLinear), and Multilayer Perceptron (MLP), with results shown in Table 2. Experiments show that the proposed model achieves the best performance in all prediction accuracy metrics, reducing the MAPE error by approximately 7.5% compared to the Transformer model. Among the various comparison models, Transformer-based architectures (such as iTransformer) outperform linear models (NLinear) and LSTM models with attention mechanisms (LSTM-Attention) overall. This is mainly due to its multi-head attention mechanism, which can adaptively focus on key temporal features and effectively capture dependencies between long time steps.

[0055] Table 2 Comparison of Predictive Performance of Different Models

[0056] In this embodiment, a rolling time window is used for online net load forecasting. The specific steps are as follows: After the model completes training and passes testing, it can be deployed for future rolling multi-step predictions. During deployment, the latest measured data from the previous day is first organized into an input vector according to the format required during training, and the model directly outputs the predicted net load point for the next day. To dynamically update the predictions, a rolling mechanism is used: after each new measured data point is obtained, the input window slides accordingly to incorporate that data, and an updated prediction is generated. This process allows the model to adapt to the latest state of the system, thereby effectively improving prediction accuracy and practical application value.

[0057] Example 2 Embodiment 2 of the present invention provides a net load forecasting system for a high proportion of renewable energy access to a distribution network, and applies a net load forecasting method for a high proportion of renewable energy access to a distribution network, including: The data acquisition module collects historical operating data of the power distribution network and regional meteorological data of the area where the power distribution network is located, and preprocesses the data. The causal relationship determination module generates an observation time series data matrix for each feature by passing the preprocessed data through a predefined causal mechanism model; based on the parameters of the causal mechanism model, it determines the causal effect strength between different features and generates a causal adjacency matrix; with the goal of minimizing the difference between the results generated by the causal mechanism model and the actual results, as well as the sum of the elements of the causal adjacency matrix, it solves for the parameters of the causal mechanism model and determines the final causal adjacency matrix. The net load forecasting model construction module constructs a net load forecasting model, including a multi-layer trend feature extraction module, a multi-layer periodic feature encoding module, and a feature fusion encoding module. It performs sequence decomposition on the preprocessed data (S1) to obtain periodic and trend components. It then sequentially performs encoding embedding and position embedding on the periodic and trend components to generate trend position embeddings and periodic position embeddings. The trend position embeddings and the causal adjacency matrix are input into the multi-layer trend feature extraction module to generate trend features. The periodic position embeddings are input into the multi-layer periodic feature encoding module to output periodic encoded features. Finally, the trend features and periodic encoded features are input into the feature fusion encoding module to predict load power, photovoltaic and wind power output, and net load power. The net load forecasting module calculates the estimated vector of net load power based on the forecast vectors of load power, photovoltaic power output, and wind power output; it calculates the total loss function and trains the net load forecasting model based on the forecast vector and estimated vector of net load power; and it predicts the net load power at future times using the trained net load forecasting model.

[0058] This disclosure can be a system, method, and / or computer program product. A computer program product may include a computer-readable storage medium having computer-readable program instructions loaded thereon for causing a processor to implement various aspects of this disclosure.

[0059] Computer-readable storage media can be tangible devices capable of holding and storing instructions for use by an instruction execution device. Computer-readable storage media can be, for example—but not limited to—electrical storage devices, magnetic storage devices, optical storage devices, electromagnetic storage devices, semiconductor storage devices, or any suitable combination of the foregoing. More specific examples (a non-exhaustive list) of computer-readable storage media include: portable computer disks, hard disks, random access memory (RAM), read-only memory (ROM), erasable programmable read-only memory (EPROM or flash memory), static random access memory (SRAM), portable compact disc read-only memory (CD-ROM), digital multifunction disc (DVD), memory sticks, floppy disks, mechanical encoding devices, such as punch cards or recessed protrusions storing instructions thereon, and any suitable combination of the foregoing. The computer-readable storage media used herein are not to be construed as transient signals themselves, such as radio waves or other freely propagating electromagnetic waves, electromagnetic waves propagating through waveguides or other transmission media (e.g., light pulses through fiber optic cables), or electrical signals transmitted through wires.

[0060] The computer-readable program instructions described herein can be downloaded from computer-readable storage media to various computing / processing devices, or downloaded via a network, such as the Internet, local area network, wide area network, and / or wireless network, to an external computer or external storage device. The network may include copper transmission cables, fiber optic transmission, wireless transmission, routers, firewalls, switches, gateway computers, and / or edge servers. A network adapter card or network interface in each computing / processing device receives the computer-readable program instructions from the network and forwards them to the computer-readable storage media in the respective computing / processing device.

[0061] Computer program instructions used to perform the operations of this disclosure may be assembly instructions, instruction set architecture (ISA) instructions, machine instructions, machine-dependent instructions, microcode, firmware instructions, status setting data, or source code or object code written in any combination of one or more programming languages, including object-oriented programming languages ​​such as Smalltalk, C++, etc., and conventional procedural programming languages ​​such as the "C" language or similar programming languages. The computer-readable program instructions may execute entirely on the user's computer, partially on the user's computer, as a standalone software package, partially on the user's computer and partially on a remote computer, or entirely on a remote computer or server. In cases involving a remote computer, the remote computer may be connected to the user's computer via any type of network—including a local area network (LAN) or a wide area network (WAN)—or may be connected to an external computer (e.g., via the Internet using an Internet service provider). In some embodiments, electronic circuitry, such as programmable logic circuitry, field-programmable gate arrays (FPGAs), or programmable logic arrays (PLAs), is personalized by utilizing the status information of the computer-readable program instructions to implement various aspects of this disclosure.

[0062] Finally, it should be noted that the above embodiments are only used to illustrate the technical solutions of the present invention and not to limit it. Although the present invention has been described in detail with reference to the above embodiments, those skilled in the art should understand that modifications or equivalent substitutions can still be made to the specific implementation of the present invention. Any modifications or equivalent substitutions that do not depart from the spirit and scope of the present invention should be covered within the protection scope of the claims of the present invention.

Claims

1. A method for predicting net load with a high proportion of new energy sources integrated into the distribution network, characterized in that, include: S1. Collect historical operating data of the power distribution network and regional meteorological data of the area where the power distribution network is located, and preprocess them; S2. The preprocessed data is used to generate an observation time series data matrix for each feature through a predefined causal mechanism model. Based on the parameters of the causal mechanism model, the strength of the causal effect between different features is determined, and a causal adjacency matrix is ​​generated. With the goal of minimizing the difference between the results generated by the causal mechanism model and the actual results, as well as the sum of the elements of the causal adjacency matrix, the parameters of the causal mechanism model are solved to determine the final causal adjacency matrix. S3. Construct a net load forecasting model, including a multi-layer trend feature extraction module, a multi-layer periodic feature encoding module, and a feature fusion encoding module; The data preprocessed by S1 is decomposed into periodic and trend components. The periodic component and the trend component are encoded and embedded sequentially to generate trend position embedding and periodic position embedding. The trend location embedding and causal adjacency matrix are input into the multi-layer trend feature extraction module to generate trend features; The periodic position is embedded into the input multi-layer periodic feature encoding module, and the periodic encoded feature is output. Input trend features and periodic coding features into the feature fusion coding module to predict load power, photovoltaic and wind power output, and net load power; S4. Based on the predicted vectors of load power, photovoltaic output, and wind power output, calculate the estimated vector of net load power; based on the predicted vector and estimated vector of net load power, calculate the total loss function and train the net load prediction model; predict the net load power at future times using the trained net load prediction model.

2. The net load forecasting method for high-proportion renewable energy access to the distribution network according to claim 1, characterized in that: The process of generating the causal adjacency matrix in S2 is as follows: The preprocessed historical operational data and regional meteorological data are concatenated to form an observation time-series data matrix. For any feature in the observation time-series data matrix, based on a predefined causal mechanism model, the observation time-series data matrices of other features are mapped to the observation time-series data matrix corresponding to that feature. A causal adjacency matrix is ​​defined, where each element represents the causal effect strength of the feature corresponding to the column index on the feature corresponding to the row index. This matrix is ​​obtained by calculating the L2 norm of the column corresponding to the row index in the first-level weight matrix of the causal mechanism model corresponding to the column index.

3. The net load forecasting method for high-proportion renewable energy access to the distribution network according to claim 1, characterized in that: The process of determining the final causal adjacency matrix in S2 is as follows: For each feature, the L2 norm is calculated by subtracting the output of the corresponding causal mechanism model from the observation time series data matrix corresponding to that feature; the L1 norm is calculated for the causal adjacency matrix; the squares of the L2 norm calculation results of all features are summed and weighted with the L1 norm calculation result; the weighted result is divided by the time length of the collected data as the optimization objective; when the graph corresponding to the causal adjacency matrix is ​​a directed acyclic graph, the parameters of the causal mechanism model for different features are solved by minimizing the optimization objective; the final causal adjacency matrix is ​​determined by substituting the solved parameters.

4. The net load forecasting method for high-proportion renewable energy access to the distribution network according to claim 1, characterized in that: The process of sequence decomposition of the data preprocessed in S1 in S3 is as follows: The preprocessed historical operational data and regional meteorological data were averaged and pooled to obtain the historical operational trend component and the regional meteorological trend component. Subtract the corresponding trend components from the preprocessed historical operational data and regional meteorological data to obtain the historical operational cycle components and the regional meteorological cycle components. The historical trend component and the regional meteorological trend component are spliced ​​together to obtain the trend component. The historical operating cycle component and the regional meteorological cycle component are spliced ​​together to obtain the cycle component.

5. The net load forecasting method for high-proportion renewable energy access to the distribution network according to claim 1, characterized in that: The specific processes for encoding embedding and positional embedding in S3 are as follows: Using a sliding window, the trend component and the periodic component are converted into trend time series features and periodic time series features of fixed time length, respectively. Each time series feature includes feature values ​​within the time window corresponding to different time segments. By using encoding embedding functions, trend time series features and periodic time series features are mapped to a predetermined dimension space, respectively, to obtain trend encoding embedding and periodic encoding embedding; For any feature in trend encoding embedding and periodic encoding embedding, the encoding embeddings of different time segments of the corresponding feature are input into the position embedding unit to obtain the hidden state of each time segment; the hidden states of all time segments are concatenated and added to the encoding embedding of the corresponding feature to obtain the position embedding of the corresponding feature; for trend component and periodic component, the position embeddings of all features in the corresponding component are concatenated to obtain trend position embedding and periodic position embedding respectively.

6. The net load forecasting method for high-proportion renewable energy access to the distribution network according to claim 1, characterized in that: The process of generating trend features in S3 is as follows: Randomly initialize a parameter matrix, perform scaling and activation operations on the parameter matrix, and obtain the node features of the corresponding node; Calculate the antisymmetric matrix of node features and calculate the initial adjacency matrix of the node using an activation function; perform pruning operation on the initial adjacency matrix of the node to obtain the corresponding sparse adjacency matrix. The causal adjacency matrix and the sparse adjacency matrix are standardized using a symmetrically normalized Laplace matrix. The transposed trend position is embedded as input to the multi-layer trend feature extraction module; For any layer in the multi-layer trend feature extraction module, the output feature matrix of the current layer is generated based on the input feature matrix of the current layer, the standardized causal adjacency matrix, and the sparse adjacency matrix. The trend features are obtained by summing the output feature matrices of all layers after passing them through a linear layer.

7. The net load forecasting method for high-proportion renewable energy access to the distribution network according to claim 6, characterized in that: The process of generating the output feature matrix of the current layer is as follows: The input feature matrix of the current layer is multiplied by a predefined gate weight matrix and then input into the activation function to obtain the gate matrix of the current layer; the standardized causal adjacency matrix is ​​multiplied by the input feature matrix of the current layer and then multiplied element-wise with the gate matrix of the current layer to obtain the static trend feature. The standardized sparse adjacency matrix is ​​multiplied by the input feature matrix of the current layer, and then multiplied element-wise by 1 minus the gating matrix of the current layer to obtain the dynamic trend feature. The sum of the static trend feature and the dynamic trend feature of the current layer is used as the static-dynamic fusion feature of the current layer. Based on the predefined modulation parameters, the static-dynamic fusion feature of the current layer and the input feature matrix are weighted to obtain the output feature matrix of the current layer, which is the input feature matrix of the next layer.

8. The net load forecasting method for high-proportion renewable energy access to the distribution network according to claim 1, characterized in that: The process of predicting distribution network operation data under different tasks in S3 is as follows: For any layer in the multi-layer periodic feature encoding module, the periodic encoding features output by the previous layer are input into a predefined multi-head time-sensitive attention mechanism to obtain the periodic attention features of the corresponding layer; the periodic attention features are added to the periodic encoding features output by the previous layer and then passed through the layer normalization function and the feedforward neural network in sequence to output the periodic encoding features output by the current layer; after passing through the multi-layer periodic feature encoding module with a predetermined number of layers, the final periodic encoding features are output. The periodic coding features and trend features are added together and then input into the feature fusion coding module to obtain the fused coding features; The fused coding features are flattened in dimension, and then multiple prediction heads are used to output prediction vectors for load power, photovoltaic output, wind power output, and net load power, respectively.

9. The net load forecasting method for high-proportion renewable energy access to the distribution network according to claim 8, characterized in that: The process of calculating the periodic attention features of different layers is as follows: The periodic encoding features output from the previous layer are input into the linear layer to generate n sets of query vectors, key vectors, and value vectors; a set of periods to be modeled is defined, and each period in the set is assigned to a different attention head. The assigned attention head is used as a non-free head, and the remaining attention heads are used as free heads. For non-free headers, define a time-sensitive mask matrix; for each position of the time-sensitive mask matrix, determine whether the time difference between the query vector and the key vector indexed at the corresponding position in the non-free header is an integer multiple of the period of the corresponding non-free header; If yes, the element at the corresponding position in the time-sensitive mask matrix is ​​0; otherwise, it is infinity. Multiply the query vector of the corresponding non-free head by the transpose of the key vector, divide by the square root of the key vector dimension, add the corresponding time-sensitive mask matrix, and then pass the activation function to obtain the attention score of the corresponding non-free head. Multiply the attention score by the value vector to get the attention output for the corresponding non-free head; For each free head, calculate the self-attention of each attention head to obtain the corresponding attention output; concatenate the attention outputs of all attention heads and perform dimensional mapping through a linear layer to output periodic attention features.

10. The net load forecasting method for high-proportion renewable energy access to the distribution network according to claim 1, characterized in that: In S4, the predicted vectors of photovoltaic power output and wind power output are successively subtracted from the predicted vector of load power to obtain the estimated vector of net load power; the mean square error between the predicted vector and the estimated vector of net load power is calculated as the loss function of the relational term. Calculate the mean square error of the net load power prediction vector and the actual net load power as the conventional loss function; weight the relational term loss function and the conventional loss function to obtain the total loss function.

11. A net load forecasting system for high-proportion renewable energy access to a distribution network, using the method described in any one of claims 1-10, characterized in that, include: The data acquisition module collects historical operating data of the power distribution network and regional meteorological data of the area where the power distribution network is located, and preprocesses the data. The causal relationship determination module generates an observation time series data matrix for each feature by passing the preprocessed data through a predefined causal mechanism model; based on the parameters of the causal mechanism model, it determines the causal effect strength between different features and generates a causal adjacency matrix; with the goal of minimizing the difference between the results generated by the causal mechanism model and the actual results, as well as the sum of the elements of the causal adjacency matrix, it solves for the parameters of the causal mechanism model and determines the final causal adjacency matrix. The net load forecasting model construction module constructs a net load forecasting model, including a multi-layer trend feature extraction module, a multi-layer periodic feature encoding module, and a feature fusion encoding module. The data preprocessed by S1 is decomposed into periodic and trend components. The periodic component and the trend component are encoded and embedded sequentially to generate trend position embedding and periodic position embedding. The trend location embedding and causal adjacency matrix are input into the multi-layer trend feature extraction module to generate trend features; The periodic position is embedded into the input multi-layer periodic feature encoding module, and the periodic encoded feature is output. Input trend features and periodic coding features into the feature fusion coding module to predict load power, photovoltaic and wind power output, and net load power; The net load forecasting module calculates the estimated vector of net load power based on the forecast vectors of load power, photovoltaic power output, and wind power output; it calculates the total loss function and trains the net load forecasting model based on the forecast vector and estimated vector of net load power; and it predicts the net load power at future times using the trained net load forecasting model.

12. A terminal, comprising a processor and a storage medium; characterized in that: The storage medium is used to store instructions; The processor is configured to operate according to the instructions to perform the steps of the method according to any one of claims 1-10.

13. A computer-readable storage medium having a computer program stored thereon, characterized in that, When executed by a processor, the program implements the steps of the method according to any one of claims 1-10.

Images