A short-term load forecasting method based on aggregated secondary modal decomposition for new energy consumption in power distribution network

Abstract

The application discloses a short-term load forecasting method based on polyphase secondary modal decomposition for new energy consumption of power distribution network, and relates to the technical field of power load forecasting. The method comprises the following steps: collecting multi-source data to construct a net load history sequence, adopting time-varying filtering empirical mode decomposition for first-level decomposition, grouping modal components based on sample entropy and HURST index weighted indicators and K-means clustering, and adopting improved variational modal decomposition for second-level decomposition on non-stationary groups; constructing a double-branch space-time forecasting network based on the topology of the power distribution network, fusing a gated recurrent unit and a graph attention network to extract space-time features, and outputting a predicted value with a confidence interval; and accumulating sub-sequence predicted values and calculating a load fluctuation margin reference value. The application can effectively solve the problem of low prediction accuracy caused by strong non-stationarity of load sequence and node space-time dynamic correlation under high penetration rate of new energy access, and provide risk-controllable decision support for new energy consumption scheduling.

Description

Technical Field

[0001] This application relates to the field of power load forecasting technology, and in particular to a short-term load forecasting method based on aggregated second-mode decomposition for the absorption of new energy in distribution networks. Background Technology

[0002] With the increasing penetration of distributed photovoltaic and wind power in distribution networks, the traditional "source follows load" operation mode has been completely disrupted. The mismatch between source and load timing characteristics has led to increasingly prominent issues such as hindered local consumption of renewable energy and wind / solar curtailment, becoming a core bottleneck restricting the safe and economical operation of new distribution networks. Accurate short-term net load forecasting of distribution networks (i.e., the remaining load after subtracting the output of distributed renewable energy from the total load) is a core prerequisite and key support for achieving accurate source-load matching, optimizing energy storage charging and discharging scheduling, and improving the local consumption capacity of distributed renewable energy. Forecast accuracy directly determines the effectiveness and economy of renewable energy consumption schemes in distribution networks. High-precision net load forecasting can significantly improve the source-load matching degree of distribution networks, reduce power backfeeding and passive curtailment, and simultaneously reduce distribution network line losses and operating costs.

[0003] Current methods for short-term load forecasting in distribution networks can be broadly categorized into three types: statistical learning algorithms, traditional machine learning algorithms, and deep learning algorithms. Statistical learning algorithms, including multiple linear regression, exponential smoothing, and Kalman filtering, offer simple models and low computational cost, but suffer from poor robustness and susceptibility to random factors. In distribution networks with high-penetration renewable energy integration, the net load sequence is significantly affected by both the randomness of renewable energy output and the volatility of user electricity consumption behavior, resulting in a substantial increase in non-stationarity and a significant reduction in the predictive reliability of these methods. Traditional machine learning algorithms, including support vector machines and random forests, can handle nonlinear multidimensional data well, but they are sensitive to outliers and are prone to performance degradation when faced with sudden changes in net load data. Furthermore, they cannot effectively capture the spatial correlation characteristics between distribution network nodes.

[0004] In recent years, deep learning has been widely used in load forecasting. Recurrent neural networks such as Long Short-Term Memory (LSTM) and Gated Recurrent Units (GRUs) have effectively captured the long-term and short-term dependencies in time series, demonstrating excellent performance in time series forecasting. However, most existing studies focus only on the time series characteristics of a single node, neglecting the spatial relationships formed by topological connections between different nodes in the distribution network, resulting in insufficient accuracy in multi-node collaborative forecasting scenarios. Furthermore, most existing methods only output single-point forecast results, failing to quantify forecast uncertainty and thus making it difficult to directly provide risk references for the dispatching of renewable energy consumption in the distribution network.

[0005] Signal decomposition techniques are widely used in load forecasting to address the high non-stationarity and strong volatility of net load sequences. By decomposing the original sequence into multiple sub-components for separate prediction, prediction accuracy can be effectively improved. However, existing decomposition methods still have significant drawbacks: Empirical Mode Decomposition (EMD) is prone to mode aliasing; while Ensemble Empirical Mode Decomposition (EEMD) can alleviate mode aliasing, it suffers from residual noise and pseudo-modes; the high-frequency components generated by primary decomposition are highly non-stationary, leading to significant errors in direct prediction; and secondary decomposition of all components generates a large number of sub-sequences, resulting in a surge in model computation and low training efficiency. Existing methods struggle to achieve an effective balance between prediction accuracy and computational efficiency. Furthermore, most existing decomposition methods use a fixed number of decomposition layers or modes, failing to adaptively adjust according to the sequence's inherent characteristics, resulting in poor adaptability. Summary of the Invention

[0006] To address the aforementioned problems, this application provides a short-term load forecasting method based on aggregated second-mode decomposition for renewable energy consumption in distribution networks, comprising the following steps: Step 1: Collect historical load data, distributed renewable energy output data, meteorological data, electricity price data, and distribution network topology information from multiple nodes in the target area of ​​the distribution network. After preprocessing all the data, construct a multi-source historical dataset. Step 2: Subtract the distributed renewable energy output data of the corresponding node from the historical load data of each node at the corresponding time to obtain the distribution network net load value of each node at different times. Define the set of distribution network net load values ​​of all nodes at all time steps as the historical sequence of distribution network net load. Step 3: Use the time-varying filtering empirical mode decomposition algorithm to perform the first-level adaptive decomposition of the historical sequence of net load of the distribution network, and obtain multiple intrinsic mode components and a residual term corresponding to each node; Step 4: For each intrinsic mode component, calculate the non-stationarity comprehensive index composed of the weighted fusion of sample entropy and Hearst exponent. Using the K-means clustering algorithm, divide all intrinsic mode components into non-stationary mode groups and stationary mode groups based on the non-stationarity comprehensive index. Step 5: Perform a second-level decomposition on each intrinsic mode component of the non-stationary mode group using an improved variational mode decomposition algorithm to obtain multiple stationary quadratic decomposition mode components. Step 6: Merge the intrinsic mode components of the stationary mode group with the quadratic decomposition mode components of the non-stationary mode group to form a set of predicted subsequences; Step 7: Construct a node-branch topology map based on the distribution network topology, extract spatial correlation features between nodes using a graph convolutional network, and train and predict all prediction subsequences by combining meteorological data features and electricity price data features. The first branch uses gated cyclic units to extract time-series features of distribution network nodes; The second branch uses a graph attention network to extract the dynamic spatial dependencies between distribution network nodes; The output layer fuses the feature representations of the two branches and outputs the net load prediction value and the corresponding prediction confidence interval for each subsequence. Step 8: Accumulate the net load forecast values ​​of all subsequences to obtain the final short-term net load forecast value of the distribution network, and calculate and generate a load fluctuation margin reference value for the distribution network's new energy consumption and dispatch based on the forecast confidence interval.

[0007] Furthermore, in step 1, the preprocessing includes: removing outliers using the 3σ criterion, filling in missing values ​​using linear interpolation, and then normalizing all data to map the data to the [0,1] interval.

[0008] Furthermore, in step 4, the formula for calculating the comprehensive index of non-stationarity is:

[0009] in, The normalized sample entropy value; The Hearst exponent; is the weighting coefficient, with a value range of [0.4, 0.6].

[0010] Furthermore, in step 5, the improved variational mode decomposition algorithm specifically includes: For the input non-stationary intrinsic mode components, the number of decomposition modes K is adaptively determined using the center frequency observation method; The whale optimization algorithm is introduced, with the minimization of envelope entropy as the objective function, to adaptively optimize the penalty factor and bandwidth parameter of variational mode decomposition. Variational mode decomposition is performed using optimized parameters, outputting multiple quadratic decomposition mode components with narrowband stationary characteristics.

[0011] Furthermore, in step 7, a graph convolutional network is used to extract spatial association features between nodes, specifically including: Based on the node-branch topology of the distribution network, an adjacency matrix A and a degree matrix D are constructed to form a graph structure. , where V is the set of nodes and E is the set of branch edges; Using spectral domain graph convolution to define graph convolution operation, the calculation formula is as follows:

[0012] in, Adjacency matrix for adding self-loops; for The degree matrix; For the first Layer node characteristics; The weight matrix is ​​a learnable weight matrix; For activation functions; By stacking two graph convolutional layers, the second-order neighborhood spatial dependency features of each node are captured, and the spatial feature vector of the fused distribution network topology is output.

[0013] Furthermore, step 7, constructing a dual-branch spatiotemporal prediction network for all predicted subsequences, specifically includes: The meteorological data features and electricity price data features are concatenated with their corresponding subsequences to obtain the time-series input features; In the first branch, the time-series input features are input into the gated recurrent unit, and the long-term and short-term dependency features of the time series are adaptively extracted through the update gate and the reset gate, and the time feature vector is output. In the second branch, the spatial features extracted by the graph convolutional network are used as the initial node representation of the graph attention network. The graph attention network assigns attention weights to different neighbor nodes, dynamically captures the spatial dependency changes between nodes as the net load fluctuates, and outputs a dynamic spatial feature vector. The temporal feature vector and the dynamic spatial feature vector are concatenated and fused, and feature mapping is performed through a fully connected layer; The network is trained using the quantile loss function, and the predicted net load values ​​for each subsequence and the prediction confidence intervals at different confidence levels are output.

[0014] Furthermore, in step 8, the formula for calculating the load fluctuation margin reference value is as follows:

[0015] in, The standard deviation corresponding to the predicted confidence interval; This is the margin adjustment coefficient, with a value range of [1.5, 3]. The load fluctuation margin reference value is used to guide the real-time scheduling of distributed new energy sources and the energy storage charging and discharging strategy of the distribution network.

[0016] This application provides a short-term load forecasting method based on aggregated quadratic mode decomposition for renewable energy consumption in distribution networks. By collecting multi-source historical data and constructing a net load historical sequence, it effectively eliminates source-load coupling interference and focuses on the true supply-demand balance of the power grid. Then, a time-varying filtered empirical mode decomposition algorithm is used for the first-level adaptive decomposition, overcoming the mode aliasing and endpoint effects of traditional decomposition methods. Based on this, a non-stationarity comprehensive index composed of a weighted fusion of sample entropy and Hearst exponent is calculated, and the intrinsic mode components are intelligently divided into non-stationary and stationary mode groups using the K-means clustering algorithm, achieving refined classification of mode components. Subsequently, an improved variational mode decomposition algorithm is applied only to the non-stationary mode group. The algorithm performs a second-level decomposition, significantly reducing computational overhead while maintaining the prediction accuracy of high-frequency components, and merges all stationary components to form a set of predicted subsequences. Furthermore, a node-branch topology graph is constructed based on the distribution network topology, and spatial correlation features are extracted using a graph convolutional network. A dual-branch spatiotemporal prediction network, incorporating meteorological and electricity price features, is built, including a time branch of gated cyclic units and a spatial branch of a graph attention network. This achieves collaborative modeling of node time series features and dynamic spatial dependencies, and outputs a net load prediction value with a prediction confidence interval through a quantile loss function. Finally, the predicted values ​​of each subsequence are summed to obtain the short-term net load prediction value, and a load fluctuation margin reference value is calculated based on the confidence interval. This method effectively solves the problems of low prediction accuracy caused by the strong non-stationarity of load sequences under high-penetration renewable energy access, difficulty in modeling the spatiotemporal dynamic correlation of multiple nodes, and lack of uncertainty quantification support for scheduling decisions. It improves the stability and reliability of short-term load forecasting for distribution networks and provides a quantitative basis with controllable risk boundaries for the local consumption and scheduling of renewable energy. Attached Figure Description

[0017] Figure 1 This is a flowchart illustrating a short-term load forecasting method based on aggregated secondary mode decomposition for renewable energy consumption in distribution networks, as proposed in this application. Detailed Implementation

[0018] The technical solutions of this application will now be clearly and completely described with reference to the accompanying drawings. Obviously, the described embodiments are merely some embodiments of this application, and not all embodiments. The components of this application described and shown in the accompanying drawings can generally be arranged and designed in various different configurations. Therefore, the following detailed description of the embodiments of this application provided in the accompanying drawings is not intended to limit the scope of the claimed application, but merely to illustrate selected embodiments of this application. All other embodiments obtained by those skilled in the art based on the embodiments of this application without inventive effort are within the scope of protection of this application.

[0019] It should be noted that similar reference numerals and letters in the following figures indicate similar items; therefore, once an item is defined in one figure, it does not need to be further defined and explained in subsequent figures. Furthermore, in the description of this application, terms such as "first," "second," etc., are used only to distinguish descriptions and should not be construed as indicating or implying relative importance.

[0020] Example 1 Reference Figure 1 This application proposes a short-term load forecasting method based on aggregated quadratic mode decomposition for renewable energy consumption in distribution networks, comprising the following steps: Step 1: Collect historical load data, distributed renewable energy output data, meteorological data, electricity price data, and distribution network topology information from multiple nodes in the target area of ​​the distribution network. After preprocessing all the data, construct a multi-source historical dataset. Step 2: Subtract the distributed renewable energy output data of the corresponding node from the historical load data of each node at the corresponding time to obtain the distribution network net load value of each node at different times. Define the set of distribution network net load values ​​of all nodes at all time steps as the historical sequence of distribution network net load. Step 3: Use the time-varying filtering empirical mode decomposition algorithm to perform the first-level adaptive decomposition of the historical sequence of net load of the distribution network, and obtain multiple intrinsic mode components and a residual term corresponding to each node; Step 4: For each intrinsic mode component, calculate the non-stationarity comprehensive index composed of the weighted fusion of sample entropy and Hearst exponent. Using the K-means clustering algorithm, divide all intrinsic mode components into non-stationary mode groups and stationary mode groups based on the non-stationarity comprehensive index. Step 5: Perform a second-level decomposition on each intrinsic mode component of the non-stationary mode group using an improved variational mode decomposition algorithm to obtain multiple stationary quadratic decomposition mode components. Step 6: Merge the intrinsic mode components of the stationary mode group with the quadratic decomposition mode components of the non-stationary mode group to form a set of predicted subsequences; Step 7: Construct a node-branch topology map based on the distribution network topology, extract spatial correlation features between nodes using a graph convolutional network, and train and predict all prediction subsequences by combining meteorological data features and electricity price data features. The first branch uses gated cyclic units to extract time-series features of distribution network nodes; The second branch uses a graph attention network to extract the dynamic spatial dependencies between distribution network nodes; The output layer fuses the feature representations of the two branches and outputs the net load prediction value and the corresponding prediction confidence interval for each subsequence. Step 8: Accumulate the net load forecast values ​​of all subsequences to obtain the final short-term net load forecast value of the distribution network, and calculate and generate a load fluctuation margin reference value for the distribution network's new energy consumption and dispatch based on the forecast confidence interval.

[0021] As a preferred embodiment of this application: in step 1, the preprocessing includes: removing outliers using the 3σ criterion, filling in missing values ​​using linear interpolation, and then normalizing all data to map the data to the [0,1] interval.

[0022] Specifically, The criteria are to remove outliers that significantly exceed the mean ± 3 standard deviations and mark them as missing; missing values ​​are filled using linear interpolation between the previous and next time points; and the max-min normalization method is used to map all data to the [0,1] interval to eliminate the influence of dimensions.

[0023] As a preferred embodiment of this application: in step 4, the formula for calculating the comprehensive index of nonstationarity is:

[0024] in, The normalized sample entropy value; The Hearst exponent; is the weighting coefficient, with a value range of [0.4, 0.6].

[0025] Specifically, for each intrinsic mode component Where n is the sequence length, and the embedding dimension is set. (Empirical optimal value, balancing computational complexity and sequence feature extraction accuracy), similarity tolerance The calculation process for sample entropy is as follows: Reconstruct the components as indivual Dimensional vector: ; Calculate the Chebyshev distance between any two vectors: ; Statistical satisfaction vector logarithm percentage And calculate its average value. ; Let the embedding dimension be Repeat the above steps to obtain ; Calculate the entropy value of the original sample: ; Minimum-maximum normalization is performed on the original sample entropy values ​​of all components to obtain the normalized sample entropy mapped to the [0,1] interval. .

[0026] The Hurst exponent is calculated using the rescaled range method (R / S method). The calculation process is as follows: The intrinsic mode component sequence of length n is divided into k sequences of length n. continuous subsequences, The range of values ​​is ; Calculate the mean for each subsequence Cumulative deviation series Range and standard deviation ; Calculate the rescaled range And take the average of all subsequences to obtain ; by The x-axis is... When performing linear regression, the slope of the regression line is the Hurst exponent, which ranges from [0,1].

[0027] The K-means clustering algorithm was used to cluster the S-values ​​of all intrinsic mode components, and the number of clusters was set. Cluster centers were initialized using the k-means++ method, and the iteration stopping criterion was that the change in cluster centers was less than 1%. After clustering is completed, the group with larger cluster centers is classified into a non-stationary mode group. The group with smaller cluster centers is classified as the stationary mode group ( ).

[0028] This application utilizes a weighted fusion of two complementary indicators to simultaneously cover both the random fluctuation and trend characteristics of the net load sequence, avoiding the limitations of a single indicator and more accurately characterizing the non-stationarity of different modal components. Clustering based on the comprehensive indicator can group components with similar non-stationary characteristics into one category, avoiding the subjectivity of traditional fixed-threshold grouping.

[0029] As a preferred embodiment of this application: step 5 of the improved variational mode decomposition algorithm specifically includes: For the input non-stationary intrinsic mode components, the number of decomposition modes K is adaptively determined using the center frequency observation method; The whale optimization algorithm is introduced, with the minimization of envelope entropy as the objective function, to adaptively optimize the penalty factor and bandwidth parameter of variational mode decomposition. Variational mode decomposition is performed using optimized parameters, outputting multiple quadratic decomposition mode components with narrowband stationary characteristics.

[0030] Specifically, the improved variational mode decomposition algorithm used in this application adaptively determines the optimal K value through the center frequency observation method: For the non-stationary intrinsic mode components of the input Perform a fast Fourier transform to obtain its power spectral density function. ; A peak detection algorithm is used to identify local maxima in the power spectrum, and significant peaks with amplitudes greater than 5% of the global maximum amplitude are selected. The number of significant peaks is then counted. ; Set the initial search range of K to [ , (Correct peak detection error) Perform VMD decomposition on each candidate K value in sequence; Calculate the center frequency of all decomposed modes for each candidate K value. If the ratio of the center frequencies of adjacent modes is greater than 1.2 (no frequency overlap) and there are no spurious modes (single mode energy percentage is less than 0.1%), then the K value is the optimal number of decomposed modes.

[0031] Furthermore, the decomposition effect of VMD is highly dependent on the penalty factor. (Control modal bandwidth; the larger the value, the narrower the modal bandwidth) and bandwidth parameters (Controlling noise tolerance, the smaller the value, the stronger the noise resistance), the Whale Optimization Algorithm (WOA) is introduced to optimize parameters with the goal of minimizing envelope entropy. The steps are as follows: Set the whale population size N=30, the maximum number of iterations T=50, and the parameters to be optimized. The search range is [1000, 10000]. The search range is [0, 1]; The objective function is to minimize the average envelope entropy of all modal components after decomposition.

[0032] in, For the k-th decomposition mode component, for The envelope entropy is calculated using the following formula:

[0033] in, for The probability distribution of the envelope signal after normalization obtained by Hilbert transform; the smaller the envelope entropy, the better the narrowband stationarity of the modal components and the lower the information redundancy. Simulates the shrinking encirclement mechanism, spiral update mechanism, and random search mechanism of whales to iteratively update the position of each whale in the population; After the iteration, the position of the whale with the smallest objective function value corresponds to... and This is the optimal VMD parameter.

[0034] The optimal number of decomposition modes Optimal penalty factor and optimal bandwidth parameters Input the VMD algorithm, perform variational mode decomposition, and output... A quadratic decomposition mode component with narrowband stationary characteristics .

[0035] As a preferred embodiment of this application: In step 7, a graph convolutional network is used to extract spatial association features between nodes, specifically including: Based on the node-branch topology of the distribution network, an adjacency matrix A and a degree matrix D are constructed to form a graph structure. , where V is the set of nodes and E is the set of branch edges; Using spectral domain graph convolution to define graph convolution operation, the calculation formula is as follows:

[0036] in, Adjacency matrix for adding self-loops; for The degree matrix; For the first Layer node characteristics; The weight matrix is ​​a learnable weight matrix; For activation functions; By stacking two graph convolutional layers, the second-order neighborhood spatial dependency features of each node are captured, and the spatial feature vector of the fused distribution network topology is output.

[0037] Specifically, an undirected graph is constructed based on the node-branch topology of the distribution network. ,in: Node set N is the total number of distribution network nodes; Edge set If node With nodes If there is a direct physical branch connection between them, then ,otherwise ; Constructing an adjacency matrix based on a graph structure ,in To avoid the loss of node features during graph convolution, self-loops are added to obtain an enhanced adjacency matrix. ,in Given the identity matrix; calculate the degree matrix of the enhanced adjacency matrix. ,in , .

[0038] The spectral domain graph convolution, based on the Chebyshev first-order approximation, is used as the fundamental graph convolution operation, and its calculation formula is as follows:

[0039] Stacking two layers of graph convolutional layers to extract spatial features: The first graph convolutional layer takes the initial features of the nodes (including basic attributes such as node net load magnitude, photovoltaic installed capacity, and load type) as input and outputs 64 dimensions, capturing the first-order direct spatial dependencies of the nodes' neighborhoods. The second graph convolutional layer takes 64-dimensional features from the first layer as input and outputs 64 dimensions, capturing the second-order neighborhood indirect spatial dependencies of nodes. The final output is a spatial feature vector of the fused topology. This serves as the initial node representation for subsequent graph attention networks.

[0040] The spatial feature vectors output by the two-layer graph convolution are batch normalized to map the feature values ​​to a standard normal distribution with a mean of 0 and a variance of 1, thereby accelerating network training convergence and preventing overfitting.

[0041] As a preferred embodiment of this application: step 7, constructing a dual-branch spatiotemporal prediction network for all predicted subsequences, specifically includes: The meteorological data features and electricity price data features are concatenated with their corresponding subsequences to obtain the time-series input features; In the first branch, the time-series input features are input into the gated recurrent unit, and the long-term and short-term dependency features of the time series are adaptively extracted through the update gate and the reset gate, and the time feature vector is output. In the second branch, the spatial features extracted by the graph convolutional network are used as the initial node representation of the graph attention network. The graph attention network assigns attention weights to different neighbor nodes, dynamically captures the spatial dependency changes between nodes as the net load fluctuates, and outputs a dynamic spatial feature vector. The temporal feature vector and the dynamic spatial feature vector are concatenated and fused, and feature mapping is performed through a fully connected layer; The network is trained using the quantile loss function, and the predicted net load values ​​for each subsequence and the prediction confidence intervals at different confidence levels are output.

[0042] Specifically, for each prediction subsequence, the time-aligned meteorological data features and electricity price data features are horizontally concatenated to construct a time-series input feature that integrates multiple external factors: A two-layer stacked gated recurrent unit (GRU) is used to extract long-short-term dependency features of time series. GRU solves the gradient vanishing problem of traditional RNN by updating and resetting gates, and is simpler and more computationally efficient than LSTM.

[0043] The 64-dimensional static spatial features output by the graph convolutional network are used as the initial node representation of the graph attention network (GAT). By assigning adaptive attention weights to different neighboring nodes, GAT dynamically captures the spatial dependency changes between nodes as the net load fluctuates, thus solving the defect that GCN can only learn fixed weights.

[0044] Time feature vector With dynamic space feature vectors The features are concatenated along the feature dimension to obtain the fused feature vector. .

[0045] Two fully connected layers are stacked to perform non-linear mapping on the fused features: the first fully connected layer has a dimension of 64 and the activation function is ReLU; the second fully connected layer has a dimension of 3, corresponding to 3 quantile outputs, and has no activation function.

[0046] The network is trained using a quantile loss function to predict net load values ​​at different quantile levels. The formula for calculating the quantile loss function is as follows:

[0047] in, To predict quantiles, the range is (0,1), indicating that there are... The true probability is less than or equal to the predicted value of that quantile; This represents the actual net load value. for The net load forecast value corresponding to the quantile.

[0048] This application selects three key quantiles for joint training: , , The total loss function is the arithmetic mean of the three quantile losses, and the calculation formula is:

[0049] in: The corresponding median forecast is used as the final point forecast result. Compared with the mean forecast optimized by the traditional mean squared error (MSE) loss, the median forecast is more robust to outliers in the net load data (such as sudden power outages and load fluctuations caused by extreme weather).

[0050] and These correspond to the lower and upper limits of the prediction interval, respectively, and together they form a 95% confidence interval, indicating that there is a 95% probability that the actual net load value will fall within this interval, thus quantifying the uncertainty of the prediction result.

[0051] As a preferred embodiment of this application: the formula for calculating the load fluctuation margin reference value in step 8 is as follows:

[0052] in, The standard deviation corresponding to the predicted confidence interval; This is the margin adjustment coefficient, with a value range of [1.5, 3]. The load fluctuation margin reference value is used to guide the real-time scheduling of distributed new energy sources and the energy storage charging and discharging strategy of the distribution network.

[0053] Specifically, the net load fluctuation range is constructed based on the load fluctuation margin. ,in The net load point prediction value at time t. Here is the reference value for load fluctuation margin at time t. This range represents the probable fluctuation range of net load under the current forecast uncertainty and is used for guidance. Distributed new energy real-time scheduling: when When the line's current carrying capacity or transformer capacity is exceeded, the output of some non-priority renewable energy sources should be restricted in advance to avoid equipment overload; when When the load is below the minimum demand of the power grid, increase the space for renewable energy consumption and reduce curtailment of solar power. Energy storage charging and discharging strategy optimization: Adjust the energy storage charging and discharging plan according to the fluctuation range. Increase energy storage charging during lower periods, Increase energy storage discharge during higher periods, while reserving space for future use. The corresponding energy storage capacity serves as a backup to cope with sudden load fluctuations.

[0054] This application transforms the abstract prediction confidence interval into an intuitive load fluctuation margin index, providing dispatchers with a clear risk boundary. Compared to traditional experience-based fixed margin methods, it can dynamically adjust the margin size according to the prediction accuracy, avoiding insufficient renewable energy absorption due to excessive conservatism or grid security risks due to excessive aggressiveness.

[0055] Example 2 Historical data from the past six months was collected from 10 key nodes (including 4 load nodes, 3 distributed photovoltaic grid-connected nodes, and 3 energy storage access nodes) in a local power distribution network, with a data sampling interval of 15 minutes. The collected data includes: Historical load data for each node ; Historical power output data of each distributed photovoltaic node ; Meteorological data: including temperature, humidity, light intensity, and cloud cover recorded by meteorological stations in the area every 15 minutes; Electricity price data: Time-of-use electricity price signals ; Distribution network topology information: node connection matrix.

[0056] The collected raw data is preprocessed for each node. And every moment Calculate the net load value using the following formula: (If the node has no new energy source, then) ) Define the set of net load values ​​for all nodes and all time steps as the historical sequence of net load for the distribution network. , where n is the total number of time steps.

[0057] Net load sequence for each node Perform TVF-EMD decomposition separately.

[0058] Taking a high photovoltaic penetration node as an example, its net load sequence is decomposed into 7 intrinsic mode components (IMF1 to IMF7) and one residual component Res. The frequencies of each component are arranged from high to low, which effectively avoids the mode aliasing phenomenon of traditional EMD. A total of 80 intrinsic mode components are obtained for all 10 nodes.

[0059] For all intrinsic mode components, the comprehensive nonstationarity index S is calculated for each component. After obtaining the S value for all components, the K-means clustering algorithm is used to group them according to the S value. After clustering, the group with larger cluster centers (S>0.5) is classified as the nonstationary mode group, which contains 32 components (mainly IMF1 to IMF3 of each node, corresponding to high-frequency random fluctuation components, which are significantly affected by meteorological changes, random electricity consumption behavior of users, and fluctuations in photovoltaic output). The group with smaller cluster centers (S<0.5) is classified as the stationary mode group, which contains 48 components (mainly IMF4 to IMF7 of each node and residual components, corresponding to daily periodic, weekly periodic, and long-term trend components).

[0060] For example, the IMF1 (S=0.82) of the high photovoltaic penetration node belongs to the non-stationary mode group; the IMF5 (S=0.23) belongs to the stationary mode group; and the residual component (S=0.11) belongs to the stationary mode group.

[0061] For the 32 intrinsic mode components of the non-stationary mode group, the improved variational mode decomposition (VMD) algorithm is used for the second-level decomposition. On average, each component is decomposed into 2.8 quadratic mode components, resulting in a total of 89 quadratic decomposed mode components with narrowband stationary characteristics.

[0062] The 48 intrinsic mode components of the stationary mode group and the 89 quadratic decomposition mode components of the non-stationary mode group were merged to form a set containing 137 prediction subsequences.

[0063] Construct an undirected node-branch graph based on the topology of the distribution network. , where the set of nodes The edge set E is determined based on the physical branch connections between nodes. A two-layer stacked graph convolutional network (GCN) is used to extract spatial correlation features, resulting in a spatial feature vector that integrates the topological structure.

[0064] A two-branch spatiotemporal prediction network was constructed for all 137 predicted subsequences and trained for prediction. For each prediction subsequence, the meteorological data features (temperature, humidity, light intensity, cloud cover) and electricity price data features aligned with the time dimension are horizontally concatenated to obtain a time-series input feature with a dimension of 6.

[0065] The temporal input features are fed into a two-layer stacked gated recurrent unit (GRU), with the hidden layer dimension set to 64. The system adaptively extracts long-term and short-term dependency features of the time series through update and reset gates, outputting a temporal feature vector. .

[0066] The 64-dimensional static spatial features output by the GCN are used as the initial node representation of the Graph Attention Network (GAT). Two layers of GAT are configured, each with 8 attention heads and an output dimension of 64. The GAT assigns adaptive attention weights to different neighboring nodes, dynamically capturing the spatial dependencies between nodes as net load fluctuates. For example, during periods of sudden changes in photovoltaic output due to cloudy weather, the attention weight between photovoltaic nodes and adjacent load nodes increases from 0.12 to 0.38. The output is a dynamic spatial feature vector. .

[0067] Time feature vector With dynamic space feature vectors By concatenating the features along the feature dimension, a fused feature vector is obtained. Two fully connected layers are stacked for nonlinear mapping.

[0068] The quantile loss function is used for training. The dataset is divided as follows: training set 70%, validation set 15%, and test set 15%. The batch size is set to 32, the optimizer is Adam, and the initial learning rate is... An early stopping mechanism was adopted (training was stopped if the validation set loss did not decrease for 10 consecutive rounds). After training, the training set loss converged to 0.012 and the validation set loss converged to 0.015, with no obvious overfitting.

[0069] The net load forecasts of all subsequences are summed to obtain the final short-term net load forecast of the distribution network (in terms of...). The corresponding median prediction is used as the point prediction result, based on and Based on the predicted values, construct a 95% prediction confidence interval. Use the margin adjustment coefficient. Calculate the reference value of load fluctuation margin at each time point, and construct the net load fluctuation range based on the point forecast value and fluctuation margin.

[0070] The prediction results and load fluctuation margin reference values ​​of this method were applied to the real-time dispatching of distributed new energy sources and the optimization of energy storage charging and discharging strategies in this distribution network. The actual operation results show that no equipment overload or voltage over-limit events occurred after adopting this method, which effectively improved the new energy absorption capacity and operational safety of the distribution network.

[0071] While the specific embodiments of this application have been described above in conjunction with the accompanying drawings, this is not intended to limit the scope of protection of this application. Those skilled in the art should understand that various modifications or variations that can be made by those skilled in the art without creative effort based on the technical solutions of this application are still within the scope of protection of this application.

Claims

1. A short-term load forecasting method based on aggregated quadratic mode decomposition for renewable energy absorption in distribution networks, characterized in that, Includes the following steps: Step 1: Collect historical load data, distributed renewable energy output data, meteorological data, electricity price data, and distribution network topology information from multiple nodes in the target area of ​​the distribution network. After preprocessing all the data, construct a multi-source historical dataset. Step 2: Subtract the distributed renewable energy output data of the corresponding node from the historical load data of each node at the corresponding time to obtain the distribution network net load value of each node at different times. Define the set of distribution network net load values ​​of all nodes at all time steps as the historical sequence of distribution network net load. Step 3: Use the time-varying filtering empirical mode decomposition algorithm to perform the first-level adaptive decomposition of the historical sequence of net load of the distribution network, and obtain multiple intrinsic mode components and a residual term corresponding to each node; Step 4: For each intrinsic mode component, calculate the non-stationarity comprehensive index composed of the weighted fusion of sample entropy and Hearst exponent. Using the K-means clustering algorithm, divide all intrinsic mode components into non-stationary mode groups and stationary mode groups based on the non-stationarity comprehensive index. Step 5: Perform a second-level decomposition on each intrinsic mode component of the non-stationary mode group using an improved variational mode decomposition algorithm to obtain multiple stationary quadratic decomposition mode components. Step 6: Merge the intrinsic mode components of the stationary mode group with the quadratic decomposition mode components of the non-stationary mode group to form a set of predicted subsequences; Step 7: Construct a node-branch topology map based on the distribution network topology, extract spatial correlation features between nodes using a graph convolutional network, and train and predict all prediction subsequences by combining meteorological data features and electricity price data features. The first branch uses a gated cyclic unit to extract the time-series features of the distribution network nodes; The second branch uses a graph attention network to extract the dynamic spatial dependencies between distribution network nodes; The output layer fuses the feature representations of the two branches and outputs the net load prediction value and the corresponding prediction confidence interval for each subsequence. Step 8: Accumulate the net load forecast values ​​of all subsequences to obtain the final short-term net load forecast value of the distribution network, and calculate and generate a load fluctuation margin reference value for the distribution network's new energy consumption and dispatch based on the forecast confidence interval.

2. The short-term load forecasting method based on aggregated quadratic mode decomposition for renewable energy consumption in distribution networks according to claim 1, characterized in that: In step 1, the preprocessing includes: removing outliers using the 3σ criterion, filling in missing values ​​using linear interpolation, and then normalizing all data to map the data to the [0,1] interval.

3. The short-term load forecasting method based on aggregated quadratic mode decomposition for renewable energy consumption in distribution networks according to claim 1, characterized in that: In step 4, the formula for calculating the comprehensive index of nonstationarity is: in, The normalized sample entropy value; The Hearst exponent; is the weighting coefficient, with a value range of [0.4, 0.6].

4. The short-term load forecasting method based on aggregated quadratic mode decomposition for renewable energy consumption in distribution networks according to claim 1, characterized in that: Step 5 of the improved variational mode decomposition algorithm specifically includes: For the input non-stationary intrinsic mode components, the number of decomposition modes K is adaptively determined using the center frequency observation method; The whale optimization algorithm is introduced, with the minimization of envelope entropy as the objective function, to adaptively optimize the penalty factor and bandwidth parameter of variational mode decomposition. Variational mode decomposition is performed using optimized parameters, outputting multiple quadratic decomposition mode components with narrowband stationary characteristics.

5. The short-term load forecasting method based on aggregated quadratic mode decomposition for renewable energy consumption in distribution networks according to claim 1, characterized in that: In step 7, a graph convolutional network is used to extract spatial association features between nodes, specifically including: Based on the node-branch topology of the distribution network, an adjacency matrix A and a degree matrix D are constructed to form a graph structure. , where V is the set of nodes and E is the set of branch edges; Using spectral domain graph convolution to define graph convolution operation, the calculation formula is as follows: in, Adjacency matrix for adding self-loops; for The degree matrix; For the first Layer node characteristics; The weight matrix is ​​a learnable weight matrix; For activation functions; By stacking two graph convolutional layers, the second-order neighborhood spatial dependency features of each node are captured, and the spatial feature vector of the fused distribution network topology is output.

6. The short-term load forecasting method based on aggregated quadratic mode decomposition for renewable energy consumption in distribution networks according to claim 1, characterized in that: Step 7, specifically, involves constructing a dual-branch spatiotemporal prediction network for all predicted subsequences, including: The meteorological data features and electricity price data features are concatenated with their corresponding subsequences to obtain the time-series input features; In the first branch, the time-series input features are input into the gated recurrent unit, and the long-term and short-term dependency features of the time series are adaptively extracted through the update gate and the reset gate, and the time feature vector is output. In the second branch, the spatial features extracted by the graph convolutional network are used as the initial node representation of the graph attention network. The graph attention network assigns attention weights to different neighbor nodes, dynamically captures the spatial dependency changes between nodes as the net load fluctuates, and outputs a dynamic spatial feature vector. The temporal feature vector and the dynamic spatial feature vector are concatenated and fused, and feature mapping is performed through a fully connected layer; The network is trained using the quantile loss function, and the predicted net load values ​​for each subsequence and the prediction confidence intervals at different confidence levels are output.

7. The detection method for a short-term load forecasting method based on aggregated quadratic mode decomposition for renewable energy consumption in distribution networks according to claim 1, characterized in that: In step 8, the formula for calculating the load fluctuation margin reference value is as follows: in, The standard deviation corresponding to the predicted confidence interval; This is the margin adjustment coefficient, with a value range of [1.5, 3]. The load fluctuation margin reference value is used to guide the real-time scheduling of distributed new energy sources and the energy storage charging and discharging strategy of the distribution network.

Images