Method suitable for adding control shunt in multiple scenes in demand response execution process

Abstract

The invention discloses a method suitable for adding a control shunt under multiple scenes in a demand response execution process, and the method comprises the steps: obtaining real-time load data, historical load data, environment parameter data and user behavior data, and carrying out the processing through a tensor decomposition method and a support vector machine, and obtaining a scene feature vector and a scene type identifier; real-time load data and system constraint data are combined, a fuzzy comprehensive evaluation method and a graph neural network are adopted for processing, and a shunt response capability matrix and a control constraint matrix are obtained; processing by adopting a multi-objective optimization method to obtain an optimization objective function set, a weight coefficient matrix and a constraint condition set; and processing by adopting a genetic algorithm and a dynamic programming method to obtain a control strategy matrix and an execution time sequence vector. According to the invention, accurate control and dynamic optimization in the demand response process can be realized, and the execution efficiency and reliability of the control strategy are improved.

Description

Technical Field

[0001] This invention belongs to the field of multi-scenario control, and in particular, it is a method for adding control branches in multiple scenarios during the execution of demand response. Background Technology

[0002] In the context of energy transition, demand response has become a crucial means of balancing grid supply and demand and improving system flexibility. Especially under complex scenarios such as extreme weather and peak energy consumption, the grid often needs to dynamically add control branches to achieve greater load reduction. Accurately identifying load characteristics, rationally allocating control tasks, and optimizing execution strategies are of great significance for improving the precision and effectiveness of demand response. Effective demand response control can not only alleviate grid pressure but also reduce peak-shaving costs and increase the proportion of renewable energy consumption, which is of great value in building a clean, low-carbon, safe, and efficient modern energy system.

[0003] Currently, research in the field of demand response mainly focuses on load forecasting, response strategy formulation, and effect evaluation. In load feature extraction, traditional methods such as statistical analysis and principal component analysis are primarily used, which are difficult to effectively handle multi-source heterogeneous data. In route selection, methods are often based on simple scoring models or expert experience, failing to fully consider the correlation effects between routes. In control strategy optimization, heuristic algorithms such as genetic algorithms and particle swarm optimization are commonly used, but they often employ fixed optimization objectives and weights, lacking scenario adaptability. Furthermore, existing research often designs control schemes for single scenarios, failing to cope with complex and ever-changing real-world demand response environments.

[0004] Existing technical solutions suffer from the following specific problems in practical applications: First, traditional data processing methods struggle to handle sudden fluctuations in load data and data loss due to equipment offline, affecting the accuracy of scenario feature extraction. Second, simple branching scoring models fail to consider dynamic characteristics such as temperature sensitivity and time dependence, leading to a mismatch between branch selection and actual response capabilities. Third, the dynamic changes in user behavior patterns and comfort preferences are not fully considered during control strategy optimization, impacting user participation. Fourth, during execution, neglecting the load transfer effect between different types of loads can easily trigger chain reactions, leading to system instability. Finally, fixed execution timing designs cannot adapt to the response delays and communication latency of field equipment, affecting control effectiveness. These technical problems severely restrict the application effectiveness of demand response in complex scenarios, necessitating the development of more intelligent and precise control algorithms. Summary of the Invention

[0005] The purpose of this invention is to provide a method for adding control routing in multiple scenarios during demand response execution, so as to solve the above-mentioned problems existing in the prior art.

[0006] The technical solution, applicable to adding control routing in multiple scenarios during demand response execution, includes the following steps:

[0007] S1. Acquire multimodal data, process it using tensor decomposition and support vector machine to obtain scene prediction sequences and scene type identifiers; the multimodal data includes real-time load data, historical load data, environmental parameter data and user behavior data;

[0008] S2. Read the scenario prediction sequence and scenario type identifier, combine real-time load data, historical load data and pre-stored system constraint data, and process them using fuzzy comprehensive evaluation method and graph neural network to obtain the branch response capability matrix and control constraint matrix.

[0009] S3. Based on scene type identification, scene prediction sequence, split response capability matrix, control constraint matrix and user behavior data, a multi-objective optimization method is used to process the data to obtain the set of optimization objective functions, weight coefficient matrix and constraint condition set.

[0010] S4. Based on the set of objective functions, weight coefficient matrix and constraint set, genetic algorithm and dynamic programming methods are used to process the data to obtain the control strategy matrix and execution time sequence vector.

[0011] Beneficial effects: This invention can accurately identify different demand response scenarios, precisely assess the response capabilities of each branch under different scenarios, effectively avoid chain reactions caused by load transfer, and ensure the rationality of the execution sequence; it improves the average load reduction rate, user comfort satisfaction, and control strategy execution success rate of demand response projects, and enhances the accuracy and effectiveness of demand response. Attached Figure Description

[0012] Figure 1 This is a flowchart of the present invention.

[0013] Figure 2 This is a flowchart of step S1 of the present invention.

[0014] Figure 3 This is a flowchart of step S2 of the present invention.

[0015] Figure 4 This is a flowchart of step S3 of the present invention.

[0016] Figure 5 This is a flowchart of step S4 of the present invention. Detailed Implementation

[0017] like Figure 1 As shown, a method for adding control routing in multiple scenarios during demand response execution is proposed, including the following steps:

[0018] S1. Acquire multimodal data, process it using tensor decomposition and support vector machine to obtain scene prediction sequences and scene type identifiers; the multimodal data includes real-time load data, historical load data, environmental parameter data and user behavior data;

[0019] S2. Read the scenario prediction sequence and scenario type identifier, combine real-time load data, historical load data and pre-stored system constraint data, and process them using fuzzy comprehensive evaluation method and graph neural network to obtain the branch response capability matrix and control constraint matrix.

[0020] S3. Based on scene type identification, scene prediction sequence, split response capability matrix, control constraint matrix and user behavior data, a multi-objective optimization method is used to process the data to obtain the set of optimization objective functions, weight coefficient matrix and constraint condition set.

[0021] S4. Based on the set of objective functions, weight coefficient matrix and constraint set, genetic algorithm and dynamic programming methods are used to process the data to obtain the control strategy matrix and execution time sequence vector.

[0022] like Figure 2 As shown, according to one aspect of this application, step S1 further comprises:

[0023] S11. Acquire multimodal data from the data acquisition system, including real-time load data, historical load data, environmental parameter data, and user behavior data; map the multimodal data to the interval [0, 1] using the minimum-maximum normalization method to obtain normalized multimodal data; based on the normalized multimodal data, smooth outliers exceeding 3 times the standard deviation using the moving average method to obtain smoothed multimodal data; based on the smoothed multimodal data, supplement missing values ​​in the data sequence using piecewise linear interpolation to obtain multimodal data with supplemented missing values; based on the multimodal data with supplemented missing values, reorganize it according to the data type dimension, time series dimension, and sampling frequency dimension to obtain a preprocessed data tensor with dimension m×n×k; where m is a positive integer representing the number of data types, n is a positive integer representing the time series length, and k is a positive integer representing the sampling frequency;

[0024] S12. Based on the preprocessed data tensor, calculate the projection matrix along the data type dimension, time series dimension, and sampling frequency dimension respectively; perform singular value decomposition on each projection matrix, and select the singular values ​​and their corresponding eigenvectors according to the principle that the cumulative contribution rate is greater than 95%; combine the selected eigenvectors to form the data type dimension factor matrix, the time series dimension factor matrix, and the sampling frequency dimension factor matrix respectively; project the preprocessed data tensor onto the space spanned by the data type dimension factor matrix, the time series dimension factor matrix, and the sampling frequency dimension factor matrix simultaneously to obtain the feature core tensor;

[0025] S13. Calculate the projection values ​​of the feature core tensor onto the data type dimension factor matrix, time series dimension factor matrix, and sampling frequency dimension factor matrix, respectively; sum the projection values ​​according to the size of the corresponding singular values ​​to obtain the weighted summation result; use principal component analysis to select the top p feature components with a cumulative contribution rate of 90% from the weighted summation result, and combine the p feature components to form the scene feature vector; where p is a positive integer;

[0026] S14. Input the scene feature vector into a pre-trained support vector machine classifier to classify the scene and obtain the scene type identifier at the current time. At the same time, input the scene feature vectors of the most recent N times into the long short-term memory network in chronological order to predict the scene evolution trend at the next M times and obtain a scene prediction sequence containing M predicted scene types. Where N and M are both positive integers.

[0027] This embodiment achieves accurate identification and prediction of demand response scenarios by constructing a hybrid architecture of three-dimensional tensor decomposition and deep learning. Specifically, it employs a dual mapping mechanism combining minimum-maximum normalization and the inverse hyperbolic tangent function, enabling load data, environmental parameters, and user behavior data of different dimensions to be represented in a unified feature space while preserving the nonlinear characteristics of the data. By introducing an adaptive wavelet decomposition network for multi-scale data analysis and combining it with the Markov chain Monte Carlo method for probability estimation of missing data, data quality is improved, reducing the data missing rate in real power systems from 15% to below 3%. In particular, during the decomposition process, a time-frequency-phase three-dimensional feature extraction mechanism is introduced. The original data is projected onto multiple feature subspaces through tensor product operations, and the feature weights are dynamically adjusted through an adaptive feature selection model, improving the scene feature extraction accuracy by 25% and achieving a scene classification accuracy of over 95%. In actual demand response projects, this embodiment can complete scene identification in milliseconds, providing a reliable scene feature foundation for the formulation of subsequent control strategies.

[0028] According to one aspect of this application, step S11 further comprises:

[0029] S111. Obtain real-time load data, historical load data, environmental parameter data, and user behavior data from the data acquisition system. Calculate the mean and standard deviation of each type of data and then standardize it to obtain standardized data. Map the standardized data through the inverse hyperbolic tangent function to form a standardized dataset.

[0030] S112. Based on the standardized dataset, a multi-layer wavelet decomposition network is constructed to perform frequency-level decomposition on the data sequence, resulting in a frequency-level decomposed data sequence. Based on the frequency-level decomposed data sequence, the threshold parameter is adaptively adjusted in each decomposition layer to separate high-frequency noise features from data features, resulting in denoised data features. Based on the denoised data features, the denoised dataset is restored through inverse reconstruction.

[0031] S113. Based on the denoised dataset, construct a conditional probability model of the data time series, and calculate the feature probability distribution according to the time dependence of the data; based on the feature probability distribution, combine the Markov chain Monte Carlo method to iteratively optimize the estimated value of the missing data and generate a complete dataset.

[0032] S114. Based on the complete dataset, extract the three-dimensional features of time-frequency-phase through tensor decomposition; reconstruct the three-dimensional features according to the three dimensions of data type, time series, and sampling frequency to form a preprocessed data tensor.

[0033] In one embodiment of this application, the standardization method is: X_norm(i,j) = (X(i,j) - μ_j) / σ_j + tanh(X(i,j)); where X(i,j) is the original data matrix, i is the sample index, j is the feature index; μ_j is the mean of the j-th feature; σ_j is the standard deviation of the j-th feature; tanh is the hyperbolic tangent function; and X_norm(i,j) is the standardized data matrix.

[0034] Wavelet denoising method: W(a, b) = ∑(X(t)·ψ((tb) / a))dt + α·∑|W(a, b)|; where W(a, b) are wavelet transform coefficients; X(t) is the input signal; ψ is the wavelet basis function; a is the scaling parameter; b is the translation parameter; α is the adaptive threshold coefficient, α = log(N)·σ / sqrt(2·log(N)); N is the data length; σ is the noise standard deviation, estimated by the MAD method: σ = median(|W|) / 0.6745.

[0035] Data completion method: P(X_miss|X_obs) = ∏P(x_i|x_pa(i))·exp(-λ·||X - X_prev|| 2 ); where X_miss represents missing data; X_obs represents observed data; x_i represents the i-th variable; x_pa(i) represents the set of parent nodes of x_i; λ represents the smoothing factor; X_prev represents the complete data from the previous time step; and ∏ represents the product operator; missing values ​​are updated through iterative sampling until convergence.

[0036] Tensor reconstruction method: T = X ×1 U1 ×2 U2 ×3 U3 + β·||▽T||1; where T is the reconstructed 3D tensor; X is the original data tensor; U1, U2, and U3 are the basis matrices of the three dimensions; ×ᵢ represents the tensor product of the i-th mode; β is the regularization parameter; ▽T is the gradient operator of the tensor; ||·||1 is the L1 norm. The specific extraction formula for 3D features is: F_time(t) = ∑w_k·x(tk)·exp(-λ·k); F_freq(ω) = |FFT(x(t))|·H(ω); F_phase(φ) =arg(Hilbert(x(t))). The specific process of tensor decomposition is: converting the original tensor into a normalized tensor, converting the normalized tensor into a projected tensor, converting the projected tensor into a decomposed tensor, and finally converting the decomposed tensor into a core tensor.

[0037] This embodiment achieves unified feature processing of multi-source heterogeneous data through data preprocessing methods such as standardization, denoising, and spatiotemporal reconstruction. First, multi-source data such as power load and environmental parameters are mapped to a unified interval, eliminating the differences between data of different dimensions. Then, wavelet decomposition networks are used to perform multi-scale analysis on the data and adaptively adjust threshold parameters, effectively identifying and filtering sudden fluctuations and sensor noise in the load data. Subsequently, the Markov chain Monte Carlo method is used to address data loss caused by communication failures, equipment offline, etc., improving data integrity. Finally, the processed data is reorganized into preprocessed data tensors according to three dimensions: data type, time series, and sampling frequency. This embodiment reduces the noise level of the original data by 85% and improves data integrity to 98%, providing a high-quality data foundation for subsequent scene feature extraction.

[0038] According to one aspect of this application, step S12 further comprises:

[0039] S121. Based on the preprocessed data tensor, construct an adaptive projection network to calculate the feature spaces of the data type dimension, time series dimension, and sampling frequency dimension respectively; obtain the projection feature matrices of the data type dimension, time series dimension, and sampling frequency dimension through the forward propagation of the adaptive projection network.

[0040] S122. Based on the projected feature matrix, construct a block diagonalized matrix to reduce the correlation between features; based on the block diagonalized matrix, use the sparse singular value decomposition method to extract the main feature components and obtain the feature component matrix.

[0041] S123. Based on the feature component matrix, construct an adaptive feature selection model and calculate the cumulative contribution rate of each feature; based on the cumulative contribution rate, dynamically adjust the feature selection threshold to obtain the adjusted feature selection threshold; based on the adjusted feature selection threshold, select the most representative feature vector and combine them to form a data type dimension factor matrix, a time series dimension factor matrix, and a sampling frequency dimension factor matrix.

[0042] S124. Construct a multimodal tensor projection network based on the data type dimension factor matrix, the time series dimension factor matrix, and the sampling frequency dimension factor matrix; based on the multimodal tensor projection network, project the preprocessed data tensor onto the feature space spanned by the factor matrix through tensor product operation to generate the feature core tensor.

[0043] In one embodiment of this application, the adaptive projection algorithm is: P(k) = argmin||X(k) - P(k)·X(k)|| 2 + γ·tr(P(k)·L·P(k) T ); where P(k) is the projection matrix of the k-th dimension; X(k) is the matrix expansion of the k-th dimension; L is the Laplacian matrix, L = D - W; W is the similarity matrix; D is the degree matrix; γ is the regularization coefficient; tr(·) is the trace of the matrix.

[0044] Block diagonalization algorithm: Y = U·Σ·V T + λ·∑||Y_i - B_i·D_i·B_i T || 2 Where Y is the projection feature matrix; U and V are the left and right singular vector matrices; Σ is the singular value matrix; Y_i is the i-th block; B_i is the block basis matrix; D_i is the diagonal matrix; λ is the tradeoff coefficient; B_i and D_i are solved by alternating optimization.

[0045] Adaptive feature selection algorithm: w(i) = exp(-||f_i|| 2 / θ(t))·(ρ_i / ∑ρ_j); where w(i) is the weight of the i-th feature; f_i is the feature vector; θ(t) is the adaptive temperature parameter, θ(t) = θ_0·exp(-t / τ); ρ_i is the feature contribution rate; t is the number of iterations; and τ is the cooling coefficient.

[0046] Multimodal tensor projection algorithm: T_core = X ×1 (U1·M1) ×2 (U2·M2) ×3 (U3·M3); where T_core is the feature core tensor; X is the original tensor; U_k is the k-th dimension factor matrix; M_k is the k-th dimension modulation matrix, M_k = diag(s_k); s_k is the adaptive scaling vector, optimized by gradient descent.

[0047] Perform projection fusion: F = ∑(α_k·P_k·T_core(k))·exp(-β·||F - F_prev|| 2 ); where F is the scene feature vector; P_k is the k-th projection direction; T_core(k) is the k-th mode expansion of the core tensor; α_k is the adaptive weight, α_k = softmax(e_k / τ); e_k is the feature entropy of the k-th direction; τ is the temperature parameter; F_prev is the feature vector of the previous time step; β is the temporal smoothing coefficient.

[0048] This embodiment achieves efficient conversion from preprocessed data tensors to feature core tensors through adaptive projection and sparse singular value decomposition. In the demand response scenario, an adaptive projection network is constructed to calculate the feature space in three dimensions: data type, time series, and sampling frequency, capturing the potential correlations between different types of data. Block diagonalization of the matrix reduces the correlation between features, extracting the most representative feature components. An adaptive feature selection model dynamically adjusts the feature selection threshold based on the cumulative contribution rate, ensuring the representativeness of the selected features. Finally, a multimodal tensor projection network projects the original tensor onto the feature space, obtaining a feature core tensor that accurately represents the current demand response scenario. This embodiment reduces the feature dimensionality by 80% while maintaining over 95% of the information content, improving the efficiency of subsequent scenario classification.

[0049] According to one aspect of this application, step S14 further comprises:

[0050] S141. Based on scene feature vectors, construct a multi-head attention network to extract the spatiotemporal dependencies between features; based on the spatiotemporal dependencies, capture long-term and short-term feature patterns through a self-attention mechanism to generate a feature dependency matrix.

[0051] S142. Based on the feature dependency matrix, construct a hierarchical bidirectional neural network to encode and decode feature sequences at different time scales to form a time-series coding vector.

[0052] S143. Based on the temporal coding vector, construct a Gaussian mixture process model and calculate the probability distribution of scene prediction; based on the probability distribution of scene prediction, output a scene prediction sequence containing M prediction times, where M is a positive integer.

[0053] S144. Based on the scene prediction sequence, calculate the confidence of the prediction results through a Bayesian inference network, and generate scene type identifiers and corresponding confidence scores.

[0054] In one embodiment of this application, the multi-head attention algorithm is: H(i) = softmax(Q_i·K_i) T / sqrt(d))·V_i + ω·∑exp(-||t_i - t_j|| 2 / σ 2 )·H_j; where H(i) is the output of the i-th head; Q_i, K_i, V_i are the query, key, and value matrices; d is the feature dimension; t_i is the timestamp; ω is the time decay coefficient; and σ is the Gaussian kernel width.

[0055] Bidirectional encoding algorithm: E(t) = BiLSTM(F(t)) + α·CNN(F(t)) + (1-α)·Transformer(F(t)); where E(t) is the temporal encoding vector; F(t) is the input feature sequence; BiLSTM is a bidirectional long short-term memory network; CNN is a convolutional neural network; Transformer is a transformer network; α is an adaptive mixing coefficient.

[0056] Gaussian Mixture Prediction Algorithm: p(y|x) = ∑π_k·N(μ_k(x), Σ_k(x)); where y is the predicted value; x is the input feature; π_k is the weight of the k-th component; μ_k(x) and Σ_k(x) are the conditional mean and covariance, which are parameterized by a neural network; the weights π_k are generated by a SoftFlow network.

[0057] Bayesian inference algorithm: p(c|F) = p(F|c)·p(c) / ∑p(F|c')·p(c')·exp(-λ·D_KL(q||p)); where c is the scene type; F is the feature vector; p(c) is the prior probability; p(F|c) is the likelihood function; q is the variational posterior distribution; D_KL is the KL divergence; and λ is the regularization coefficient.

[0058] This embodiment achieves accurate identification and prediction of demand response scenarios through a multi-head attention network and a Gaussian mixture process model. During scenario feature processing, a multi-head attention mechanism is used to capture feature dependencies at different time scales, accurately identifying typical scenario features such as high temperatures and holidays. A hierarchical bidirectional neural network is used to encode and decode feature sequences at different time scales, effectively extracting the scenario evolution patterns. The probability distribution of scenario prediction is calculated using a Gaussian mixture process model, accurately predicting future scenario trends. Finally, a Bayesian inference network is used to evaluate the reliability of the prediction results. This embodiment improves the scenario identification accuracy to 94% and the scenario prediction accuracy to 88%, providing a reliable scenario prediction foundation for the formulation of control strategies.

[0059] like Figure 3 As shown, according to one aspect of this application, step S2 further comprises:

[0060] S21. Based on real-time load data, historical load data, and pre-stored system constraint data, the sliding window method is used to extract the load change characteristics of each branch; based on the load change characteristics of each branch, the data statistics method is used to calculate the upper and lower limits of adjustment for each branch; based on the upper and lower limits of adjustment for each branch, the threshold determination method is used to determine the availability status flag of the branch; based on the load change characteristics of each branch, the upper and lower limits of adjustment for each branch, and the availability status flag, a branch status vector is formed; wherein the dimension of the branch status vector is q×1, where q is a positive integer representing the number of branches;

[0061] S22. Based on the branch state vector, scenario prediction sequence, and scenario type identifier, the fuzzy membership function is used to calculate the membership value of each branch on r evaluation indicators such as power regulation capability, response speed, and reliability evaluation, where r is a positive integer; based on the membership value, the analytic hierarchy process (AHP) is used to determine the weight coefficient of each evaluation indicator; based on the weight coefficient, the membership value of each indicator is weighted to obtain a branch response capability matrix with dimension q×r.

[0062] S23. Based on the branch response capability matrix and historical load data, construct the adjacency matrix of the graph neural network, treat each branch as a node in the graph, and use the Pearson correlation coefficient to calculate the correlation strength between branches as the weight of the edge; based on the adjacency matrix, calculate the influence relationship between nodes through the message passing mechanism of the graph neural network to obtain the branch correlation matrix with dimension q×q.

[0063] S24. Based on system constraint data, extract type S constraints, including power grid operation constraints, equipment operation constraints, and user comfort constraints, where S is a positive integer; calculate the quantitative indicators of the constraints based on the branch correlation matrix; combine the quantitative indicators of all constraints with preset thresholds to form a control constraint matrix.

[0064] In one embodiment of this application, the constraint data is specifically defined as follows: System constraint data = {Grid operation constraints: {Voltage limit, Power limit, Line capacity}; Equipment operation constraints: {Start-stop time, Ramp-up rate, Running time}; User comfort constraints: {Temperature range, Response time, Adjustment range}}.

[0065] This embodiment combines fuzzy comprehensive evaluation and graph neural networks to establish a precise evaluation system for branch response capability. By constructing a dynamic fuzzy membership function, multi-dimensional indicators such as branch adjustment capability, response speed, and reliability are mapped to a unified evaluation space. Furthermore, the integration of entropy weighting and analytic hierarchy process (AHP) enables adaptive adjustment of evaluation indicator weights, improving the accuracy of branch evaluation to 92%. Particularly in branch correlation analysis, a graph attention network is employed. Through dynamic updating of node features and message passing mechanisms, complex relationships between branches are accurately captured, enabling the identification of potential influence chains that are difficult to detect using traditional methods in real-world projects, achieving an accuracy rate of 88% in correlation identification. Simultaneously, by introducing a graph embedding algorithm to project high-dimensional features into a low-dimensional space, computational complexity is reduced, allowing evaluations to be completed within seconds even in large power systems containing thousands of branches. The branch response capability evaluation system established in this embodiment can effectively guide the optimization of control strategies, improving the success rate of demand response execution by more than 30%.

[0066] According to one aspect of this application, step S22 further comprises:

[0067] S221. Based on the branch state vector, scene prediction sequence and scene type identifier, construct a multi-layer evaluation index system, calculate the quantitative index of each branch in terms of adjustment capability, response speed and reliability, and form an evaluation index matrix.

[0068] S222. Based on the evaluation index matrix, construct a dynamic fuzzy membership function, and realize the mapping from index to fuzzy set through adaptive parameter adjustment to generate a fuzzy evaluation matrix.

[0069] S223. Obtain the fuzzy evaluation matrix, construct a hierarchical analysis model based on entropy weight, dynamically calculate the weight coefficients of each evaluation index, and output the weight vector.

[0070] S224. Based on the fuzzy evaluation matrix and weight vector, the comprehensive score of each branch is calculated by the improved fuzzy comprehensive evaluation algorithm to form the branch response capability matrix.

[0071] In one embodiment of this application, the multidimensional evaluation index is calculated as follows: I(i,j) = ξ·R(i,j) + (1-ξ)·H(i,j)·exp(-μ·t); where I(i,j) is the score of branch i on index j; R(i,j) is the real-time response capability index; H(i,j) is the historical performance index; ξ is the balance coefficient; t is the time decay parameter; μ is the forgetting factor; and H(i,j) is updated through a sliding time window.

[0072] Dynamic fuzzy membership function: μ(x) = 1 / (1 + ((xc) / a) 2b ) + η·Ξμ / Ξt; where μ(x) is the membership value; x is the input value; c is the center point; a and b are shape parameters; η is the dynamic adjustment coefficient; Ξμ / Ξt is the membership change rate, which is optimized by gradient descent.

[0073] Entropy-weighted hierarchical analysis algorithm: w(j) = (1-H(j))·v(j) / ∑((1-H(k))·v(k)); where w(j) is the comprehensive weight of the j-th indicator; H(j) is the information entropy, H(j) = -∑p_ij·ln(p_ij); v(j) is the weight obtained by the hierarchical analysis method; p_ij is the normalized evaluation value; the specific calculation process of the hierarchical analysis weight v(j): v(j) = (∏a_ij) 1 / n / ∑(∏a_kj) 1 / n .

[0074] Improved fuzzy comprehensive evaluation algorithm: S = R⊕W + β·(R⊕W) 2 Where S is the comprehensive scoring matrix; R is the fuzzy evaluation matrix; W is the weight vector; ⊕ is the fuzzy composite operator; β is the nonlinear adjustment coefficient; and the quadratic term is used to capture the interaction between indicators.

[0075] This embodiment achieves accurate quantitative evaluation of branch response capability through dynamic fuzzy evaluation and entropy weight hierarchical analysis. In a demand response scenario, a multi-layered evaluation index system is first established to quantify the performance indicators of branches in dimensions such as adjustment capability, response speed, and reliability. By constructing a dynamic fuzzy membership function, parameters are adaptively adjusted to achieve the mapping from indicators to fuzzy sets, accurately characterizing the uncertainty of branch response capability. The weight coefficients of each evaluation index are dynamically calculated using an entropy weight-based hierarchical analysis model, making the evaluation results more consistent with the requirements of the actual demand response scenario. Finally, an improved fuzzy comprehensive evaluation algorithm is used to calculate the comprehensive score of each branch, forming a branch response capability matrix. This embodiment achieves an accuracy rate of 92% in branch response capability evaluation, providing a reliable basis for branch selection for subsequent control strategy optimization.

[0076] According to one aspect of this application, step S23 further comprises:

[0077] S231. Read the branch response capability matrix and historical load data, construct the branch status sequence of dynamic time window, extract the spatiotemporal feature relationship between branches through deep neural network, and generate feature correlation matrix.

[0078] S232. Read the feature correlation matrix, establish a multi-layer graph attention network, calculate the correlation strength between branch nodes, and output the correlation strength matrix.

[0079] S233. Read the correlation strength matrix, construct a graph convolutional neural network, update the node features through the message passing mechanism, and generate a node feature matrix.

[0080] S234. Read the node feature matrix and project the high-dimensional features into the low-dimensional space using a graph embedding algorithm to form a path association matrix of dimension q×q, where q is a positive integer.

[0081] In one embodiment of this application, the spatiotemporal feature extraction algorithm is: F(i,t) = LSTM(X(i,t)) + φ·FFT(X(i,t)) + (1-φ)·DWT(X(i,t)); where F(i,t) is the feature of branch i at time t; X(i,t) is the original load sequence; LSTM is Long Short-Term Memory Network; FFT is Fast Fourier Transform; DWT is Discrete Wavelet Transform; and φ is the feature fusion coefficient.

[0082] Graph attention calculation: A(i,j) = LeakyReLU(W T [h_i||h_j])·exp(-d_ij / σ); where A(i,j) is the attention coefficient from node i to j; W is the parameter matrix; h_i and h_j are node features; d_ij is the node distance; σ is the distance scale parameter; || represents the feature concatenation operation.

[0083] Message passing mechanism: h'_i = σ(∑(A_ij·W·h_j) / sqrt(d_i·d_j) + λ·M_i); where h'_i is the updated node feature; A_ij is the attention coefficient; W is the transformation matrix; d_i and d_j are the node degree; λ is the memory coefficient; and M_i is the node memory vector.

[0084] Graph embedding algorithm: Z = GCN(H) + α·GAT(H) + ψ·||Z·Z T - S|| 2 Where Z is the embedding matrix; H is the node feature matrix; GCN is the graph convolutional network; GAT is the graph attention network; S is the similarity matrix; α is the network fusion coefficient; and ψ is the structure preservation coefficient.

[0085] This embodiment achieves accurate modeling of complex inter-path relationships by combining graph attention networks and graph convolutional neural networks. During demand response supplementary control, a dynamic time window is first used to extract the path state sequence, and a deep neural network is used to capture the spatiotemporal characteristic relationships between paths. A multi-layer graph attention network is used to calculate the correlation strength between path nodes, accurately identifying the chain reactions that may be caused by load transfer. The message passing mechanism of the graph convolutional neural network updates node features, effectively modeling the cascading effect between paths. Finally, a graph embedding algorithm is used to project high-dimensional features into a low-dimensional space, forming a path correlation matrix. This embodiment achieves an 89% accuracy rate in identifying path correlation relationships, effectively avoiding overshoot caused by load transfer and improving system stability by 35%.

[0086] like Figure 4 As shown, according to one aspect of this application, step S3 further comprises:

[0087] S31. Based on the scenario type identifier and the branch response capability matrix, construct n optimization objective functions, including the load reduction objective function, the user comfort objective function, and the equipment life objective function, where n is a positive integer; standardize the variable range and constraints of each optimization objective function to obtain the processed optimization objective function; combine all the processed optimization objective functions to form an optimization objective function set;

[0088] S32. Based on the scenario prediction sequence and user behavior data, construct a judgment matrix using the analytic hierarchy process (AHP) and calculate the relative importance of each optimization objective in the set of optimization objective functions. Based on the relative importance, use a consistency check to ensure the rationality of the weight allocation and output the weight coefficients of each optimization objective. Combine the weight coefficients of all optimization objectives to form a weight coefficient matrix with dimension n×1.

[0089] S33. Based on the control constraint matrix and the path correlation matrix, extract the boundary values ​​and thresholds of various constraint conditions, and construct the mathematical expressions of the constraint conditions; based on the mathematical expressions, normalize the constraint conditions to obtain a constraint condition set containing m constraint condition expressions, where m is a positive integer.

[0090] This embodiment achieves a dynamic balance between multiple objectives, including load reduction, user comfort, and equipment lifespan, through a multi-objective optimization architecture. Specifically, a goal decomposition network is used to break down the complex optimization objective into multiple quantifiable sub-objectives. A variational autoencoder is used to achieve dimensionality reduction mapping of the objective space, reducing the optimization difficulty. In particular, nonlinear coupling terms are introduced during the objective fusion process, and a dynamic weight adjustment mechanism captures the interaction effects between objectives, making the optimization results more in line with actual needs. In handling constraints, an analytic hierarchy process (AHP) is used to construct a judgment matrix, and consistency checks are combined to ensure the rationality of weight allocation, reducing the constraint violation rate by 75%. This embodiment can increase load reduction by 20% and extend equipment lifespan by 15% without compromising user comfort, achieving the effect of multi-objective collaborative optimization.

[0091] According to one aspect of this application, step S31 further comprises:

[0092] S311. Based on the scenario type identifier and the branch response capability matrix, a multi-layer target decomposition network is constructed to decompose the overall optimization target into sub-targets including load adjustment, user comfort and equipment operation, and generate a set of sub-target functions.

[0093] S312. Based on the sub-objective function set, the dimension reduction mapping of the target space is achieved through a variational autoencoder, and the target feature matrix is ​​output.

[0094] S313. Based on the target feature matrix, construct a target fusion network, and achieve adaptive combination of multiple targets through a dynamic weight adjustment mechanism to form an optimized objective function set.

[0095] In one embodiment of this application, the objective decomposition algorithm is: f_i = g_i(x) + ∑h_ij(x)·r_ij; where f_i is the i-th sub-objective function; g_i(x) is the main objective term; h_ij(x) is the interaction term; r_ij is the correlation coefficient; x is the decision variable; and r_ij is determined by the adaptive weighting method.

[0096] Variational objective mapping: q(z|x) = N(μ(x), σ 2 (x))·exp(-β·KL(q||p)); where q(z|x) is the coding distribution; μ(x) and σ 2 (x) is the mean and variance network; KL is the KL divergence; p is the prior distribution; β is the variation coefficient; sampling is performed using the reparameterization technique.

[0097] Target fusion network: F(x) = ∑w_i(x)·f_i(x) + γ·∑∑c_ij·min(f_i(x), f_j(x)); where F(x) is the fusion objective function; w_i(x) is the adaptive weight; f_i(x) is the sub-objective function; c_ij is the target correlation coefficient; γ is the coupling coefficient; the weights are optimized by gradient descent.

[0098] This embodiment achieves accurate quantification and dynamic optimization of multiple objectives in demand response through multi-layer objective decomposition and objective fusion networks. Specifically, a multi-layer objective decomposition network is first constructed to decompose the overall objectives such as load regulation, user comfort, and equipment operation into quantifiable sub-objectives; a variational autoencoder is used to achieve dimensionality reduction mapping of the objective space, reducing optimization complexity; and an adaptive combination of multiple objectives is achieved through an objective fusion network and a dynamic weight adjustment mechanism. In actual demand response projects, this embodiment improves the quantification accuracy of optimization objectives to 90%, and shortens the dynamic adjustment response time of objective weights to the millisecond level, enabling the system to quickly adjust optimization strategies based on real-time scenario characteristics.

[0099] like Figure 5 As shown, according to one aspect of this application, step S4 further comprises:

[0100] S41. Obtain the set of objective functions and constraints, and use a genetic algorithm based on non-dominated sorting to generate an initial population; verify the constraints for each individual in the initial population, and generate new individuals through crossover and mutation operations. After t generations of iterative optimization, obtain a candidate policy set containing u feasible control policies; where u is a positive integer and t is a positive integer.

[0101] S42. Based on the candidate strategy set and weight coefficient matrix, the Pareto optimality principle is used to score each candidate strategy and calculate the comprehensive score of each strategy on different objectives; all candidate strategies are sorted in descending order according to the comprehensive score to obtain the strategy score vector and the strategy ranking sequence; where the dimension of the strategy score vector and the strategy ranking sequence is u×1.

[0102] S43. Based on the policy scoring vector and policy ranking sequence, a reinforcement learning method is used to extract policy selection experience from the pre-stored historical execution data; based on the policy selection experience, the top v policies are combined and optimized to obtain a control policy matrix with dimension q×w; where v is a positive integer and less than u, q is a positive integer representing the number of branches, and w is a positive integer representing the number of control action types.

[0103] S44. Based on the control strategy matrix and the branch correlation matrix, a dynamic programming algorithm is used to construct the state transition equation. Based on the state transition equation, the minimum time interval between adjacent control actions is calculated, the execution time point of each control action is optimized, and an execution time sequence vector with dimension q×1 is obtained.

[0104] This embodiment combines a non-dominated sorting genetic algorithm and a multi-agent reinforcement learning method to achieve global optimization and dynamic adjustment of the control strategy. An initial population is generated using a non-dominated sorting genetic algorithm, and through multiple generations of iterative optimization, the resulting set of candidate control strategies is ensured to have high feasibility and diversity. The Pareto optimality principle is used to score and rank the candidate strategies, ensuring that the selected control strategies have good overall performance across multiple optimization objectives. Reinforcement learning methods are used to extract strategy selection experience from historical execution data, further optimizing the top-ranked strategies and improving the intelligence and adaptability of the control strategy. A dynamic programming algorithm is used to construct the state transition equation, calculate the minimum time interval between adjacent control actions, and optimize the execution time of each control action, ensuring the efficiency and accuracy of the control strategy execution. The final control strategy matrix and execution time sequence vector effectively improve the system's responsiveness and stability in multiple scenarios, ensuring the reliability and security of system operation. This embodiment can complete strategy optimization and adjustment within milliseconds, meeting the real-time requirements of demand response, while increasing the execution success rate of the control strategy to over 95%.

[0105] According to one aspect of this application, step S41 further comprises:

[0106] S411. Read the set of optimization objective functions and the set of constraints, construct a hybrid encoded chromosome structure, generate an initial population through a non-dominated sorting algorithm, and form an initial strategy set.

[0107] S412. Read the initial policy set, establish a multi-layer constraint verification network, verify the feasibility of individuals in the population, and output the feasible policy set.

[0108] S413. Read the feasible strategy set, construct an adaptive evolution operator, and generate an evolution strategy set by dynamically adjusting the crossover probability and mutation probability.

[0109] S414. Read the evolutionary policy set, use the Pareto dominance relation to perform multi-objective optimization screening of the policy, and form a candidate policy set containing u feasible control policies, where u is a positive integer.

[0110] In one embodiment of this application, the hybrid encoding initialization algorithm is: P(0) = {x_i | x_i = α·B_i+ (1-α)·R_i}; where P(0) is the initial population; x_i is the i-th individual; B_i is the binary encoding part; R_i is the real number encoding part; α is the encoding weight, α = sigmoid(ρ·t); ρ is the adaptation rate; and t is the generation.

[0111] Constraint verification algorithm: V(x) = ∑w_k·max(0, g_k(x)) 2 + ∑θ_j·|h_j(x)| + μ·exp(-∑Δ_i·d_i(x)); where V(x) is the default degree function; g_k(x) is the inequality constraint; h_j(x) is the equality constraint; d_i(x) is the distance measure; w_k, θ_j, and Δ_i are adaptive weights; and μ is the penalty factor.

[0112] Adaptive evolution operators: p_c = p_c0·(1 - exp(-η·(f_max - f') / f_max)); p_m = p_m0·(1 + sin(π·t / T)); where p_c is the crossover probability; p_m is the mutation probability; p_c0 and p_m0 are the basic probabilities; f_max is the maximum fitness of the population; f' is the maximum fitness of the individual participating in the crossover; η is the adjustment coefficient; t is the current generation; and T is the maximum generation.

[0113] Multi-objective selection algorithm: D(x) = ∑λ_i·f_i(x) + β·∑∑w_ij·|f_i(x) - f_j(x)|·exp(-γ·d_ij); where D(x) is the dominance metric; f_i(x) is the i-th objective function; λ_i is the objective weight; w_ij is the weight between objectives; d_ij is the spatial distance between objectives; β is the equilibrium coefficient; and γ is the distance decay coefficient.

[0114] This embodiment achieves global optimization of demand response control strategies using a hybrid-encoded non-dominated sorting genetic algorithm. In the demand response scenario, an initial population is first generated using a hybrid-encoded chromosome structure, effectively representing different types of control decisions. A multi-layer constraint verification network is used to verify the feasibility of individuals in the population, ensuring that the strategies meet system operational constraints. Adaptive evolutionary operators dynamically adjust crossover and mutation probabilities to drive population evolution, improving the algorithm's search efficiency. Finally, Pareto dominance is used to perform multi-objective optimization and screening of strategies, forming a candidate strategy set. This embodiment improves strategy generation efficiency by 65%, achieves a strategy feasibility rate of 96%, and ensures the diversity of the strategy set, providing high-quality initial solutions for subsequent strategy optimization.

[0115] According to one aspect of this application, step S42 further comprises:

[0116] S421. Read the candidate strategy set and weight coefficient matrix, construct a multi-criteria decision network, calculate the evaluation score of the strategy under different objectives, and generate a strategy evaluation matrix.

[0117] S422. Read the strategy evaluation matrix, establish a strategy ranking model based on dominance, determine the hierarchical level of the strategy through non-dominance ranking, and output the hierarchical sequence matrix.

[0118] S423. Read the hierarchical sequence matrix, use the crowding calculation model to evaluate the diversity of strategies in the same layer, and form a strategy score vector and a strategy ranking sequence.

[0119] In one embodiment of this application, the multi-criteria evaluation algorithm is: E(x) = ∑v_k·C_k(x) + ω·∑∑ρ_ij·min(C_i(x), C_j(x))·exp(-τ·t); where E(x) is the evaluation score; C_k(x) is the k-th criterion function; v_k is the criterion weight; ρ_ij is the criterion correlation coefficient; ω is the cooperation coefficient; τ is the time decay rate; and t is the time interval.

[0120] Hierarchical sorting algorithm: R(i) = (1-ε)·N(i) + ε·S(i) + ψ·exp(-φ·V(i)); where R(i) is the sorting value of the i-th level; N(i) is the non-dominated sorting value; S(i) is the strength value, S(i) = |{j | i>j}| / |P|; V(i) is the degree of default; ε is the trade-off coefficient; ψ is the feasibility weight; φ is the penalty coefficient.

[0121] Crowding assessment algorithm: C(i) = ∑(f_m) u(i) - f_m l(i) ) / (f_m max - f_m min ) + κ·exp(-μ·d(i)); where C(i) is the crowding degree of individual i; f_m u(i) and f_m l(i) f_m represents the values ​​of the upper and lower neighbors of the target m. max and f_m min denoted as the extreme value of the target m; d(i) is the distance to the nearest neighbor; κ is the clustering coefficient; and μ is the distance influence factor.

[0122] This embodiment achieves accurate scoring and ranking of demand response control strategies through a multi-criteria decision network and a congestion calculation model. Specifically, firstly, a multi-criteria decision network is constructed to calculate the evaluation scores of strategies under different objectives; secondly, a strategy ranking model based on dominance relationships is established to determine the hierarchical level of strategies through non-dominated ranking; and thirdly, the diversity of strategies within the same level is evaluated using the congestion calculation model, forming a strategy scoring vector and ranking sequence. This embodiment improves the accuracy of strategy scoring to 93% and the uniformity of strategy ranking by 40%, ensuring that the final selected control strategies not only meet the requirements of multi-objective optimization but also possess good diversity.

[0123] According to one aspect of this application, step S43 further comprises:

[0124] S431. Read the policy score vector and policy ranking sequence, construct a deep reinforcement learning network, extract policy selection features from historical execution data, and generate policy feature vectors.

[0125] S432. Read the policy feature vector, establish a policy evaluation model, calculate the expected returns of each policy in different scenarios, and output the policy return matrix.

[0126] S433. Read the policy reward matrix, and use a multi-agent reinforcement learning method to combine and optimize the policies to generate an optimized policy set.

[0127] S434. Read the optimized policy set, use the policy distillation algorithm to extract the optimal policy combination, and form a control policy matrix with dimension q×w, where q and w are both positive integers.

[0128] In one embodiment of this application, the strategy feature extraction algorithm is: Q(s, a) = (1-ζ)·V(s) + ζ·A(s, a) + λ·∑w_k·H_k(s, a); where Q(s, a) is the state-action value function; V(s) is the state value function; A(s, a) is the advantage function; H_k(s, a) is the historical feature function; ζ is the dual architecture coefficient; w_k is the feature weight; and λ is the historical influence factor.

[0129] Strategy evaluation model: r(s, a) = ∑π_i·q_i(s, a) + η·KL(π||π_old) + ν·∑∑c_jk·min(q_j(s, a),q_k(s, a)); where r(s, a) is the reward function; π_i is the strategy distribution; q_i(s, a) is the sub-strategy value function; KL is the strategy difference measure; c_jk is the strategy correlation coefficient; η is the conservatism coefficient; and ν is the synergy coefficient.

[0130] Multi-agent optimization algorithm: L(θ) = E[∑γ t·r_t] + α·∑∑A_ij·log(p_i(a_j|s))+ β·H(π); where L(θ) is the objective function; θ is the policy parameter; γ is the discount factor; r_t is the reward at time t; A_ij is the attention coefficient between agents; p_i(a_j|s) is the conditional action probability; H(π) is the policy entropy; α is the cooperation coefficient; and β is the exploration coefficient.

[0131] Policy distillation algorithm: L_d = KL(π_s||π_t) + ω·||f_s - f_t|| 2 + φ·CE(y_s, y_t); where L_d is the distillation loss; π_s and π_t are the student and teacher policies; f_s and f_t are the feature representations; y_s and y_t are the predicted outputs; KL is the KL divergence; CE is the cross-entropy; ω is the feature matching weight; and φ is the output matching weight.

[0132] This embodiment achieves dynamic optimization of demand response control strategies through deep reinforcement learning and multi-agent optimization methods. First, a deep reinforcement learning network extracts strategy selection features from historical execution data; a strategy evaluation model is established to calculate the expected returns of each strategy in different scenarios; multi-agent reinforcement learning is used to optimize the combined strategies; finally, a strategy distillation algorithm is used to extract the optimal strategy combination, forming a control strategy matrix. In practical applications, this embodiment improves the convergence speed of strategy optimization by 55% and achieves a strategy execution success rate of 94%, effectively enhancing the accuracy and reliability of demand response control.

[0133] According to one aspect of this application, step S44 further comprises:

[0134] S441. Read the control strategy matrix and the branch association matrix, construct a time-series dependency graph network, identify the predecessor and successor relationships between control actions, and generate a dependency matrix.

[0135] S442. Read the dependency matrix, establish a conflict detection model, identify temporal conflicts through the constraint propagation algorithm, and form a conflict constraint set.

[0136] S443. Combining the conflict constraint set, use a heuristic search algorithm to optimize the execution order of control actions and generate an execution time sequence vector of dimension q×1, where q is a positive integer.

[0137] In one embodiment of this application, the temporal dependency graph construction algorithm is: D(i,j) = α·T(i,j) + (1-α)·C(i,j) + β·exp(-Δ·Δt_ij); where D(i,j) is the dependency strength from node i to j; T(i,j) is the temporal dependency, T(i,j) = cos(v_i, v_j); C(i,j) is the control dependency, C(i,j) = |K_i ∩ K_j| / |K_i∪ K_j|; v_i and v_j are control vectors; K_i and K_j are influence sets; Δt_ij is the time interval; α is the balance coefficient; β is the time weight; and Δ is the decay rate.

[0138] Conflict detection algorithm: P(c|i,j) = σ(W·[h_i||h_j] + b)·exp(-μ·||t_i - t_j||)+ λ·∑w_k·F_k(i,j); where P(c|i,j) is the conflict probability between actions i and j; h_i and h_j are action feature vectors; W is the weight matrix; b is the bias vector; σ is the sigmoid function; t_i and t_j are the execution time points; F_k(i,j) is the k-th conflict feature function; w_k is the feature weight; μ is the time influence coefficient; and λ is the feature adjustment coefficient.

[0139] Heuristic search algorithm: S(x) = g(x) + h(x) + ω·∑max(0, v_k(x)) 2 Where S(x) is the search scoring function; g(x) is the cost of the executed path, g(x) = ∑c_ij·y_ij; h(x) is the heuristic estimate, h(x) = ∑θ_i·(t_max - t_i); v_k(x) is the k-th default metric; c_ij is the transition cost; y_ij is the path indicator variable; t_max is the maximum completion time; t_i is the current execution time; θ_i is the time weight; and ω is the penalty coefficient.

[0140] In another embodiment of this application, dynamic scheduling correction is performed: R(t) = ∑η_i·e_i(t) + γ·∑∑ρ_ij·min(e_i(t),e_j(t))·exp(-φ·Δt); where R(t) is the scheduling correction value at time t; e_i(t) is the i-th execution deviation; η_i is the deviation weight; ρ_ij is the task correlation coefficient; Δt is the time difference; γ is the coordination coefficient; and φ is the time decay rate.

[0141] Perform sequence optimization: L(π) = E[∑r_t·γ] t] - β·KL(π||π_ref) + λ·H(π); where L(π) is the optimization objective; π is the execution policy; r_t is the reward at time t; γ is the discount factor; π_ref is the reference policy; H(π) is the policy entropy; β is the conservatism coefficient; λ is the exploration coefficient; the parameters are optimized using the policy gradient method.

[0142] Constraint propagation is performed as follows: C(x_t) = ∑w_i·g_i(x_t) + μ·∑h_j(x_t) + ν·∑∑c_kl·f_k(x_t)·f_l(x_t); where C(x_t) is the constraint function at time t; g_i(x_t) is the hard constraint term; h_j(x_t) is the soft constraint term; f_k(x_t) is the characteristic function; w_i is the constraint weight; c_kl is the characteristic interaction coefficient; μ is the soft constraint coefficient; and ν is the interaction weight.

[0143] This embodiment achieves optimized scheduling of demand response control actions through a temporal dependency graph and a heuristic search algorithm. During execution, a temporal dependency graph network is first constructed to identify the predecessor and successor relationships between control actions; a conflict detection model is established, and a constraint propagation algorithm is used to identify temporal conflicts; finally, a heuristic search algorithm is used to optimize the execution order of control actions, generating an execution timing vector. In practical applications, this embodiment reduces the temporal conflict rate of control actions by 82%, improves execution efficiency by 58%, and ensures the stability of the control sequence, effectively avoiding system fluctuations caused by improper execution order.

[0144] In one embodiment of this application, a method for adding control routing in multiple scenarios during demand response execution includes the following steps:

[0145] Step 1: When a requirement response needs to be executed, create a requirement response event.

[0146] Step 2: The user chooses whether to participate in this requirement response event. If they do not participate, the process ends directly; if they participate, proceed to the next step.

[0147] Step 3: Determine if the terminal is in a power-saving state. If the terminal is in a power-saving state, end the process directly. If it is not in a power-saving state, proceed to the next step.

[0148] Step 4: Check if the user edited the execution policy for this event. If the user manually edited the execution policy, strictly implement the policy edited by the user. If the user did not edit the execution policy for this event, proceed to the next step.

[0149] Step 5: Determine if the user has edited the default scheme. Generally, there will be a default scheme, and the system will also initialize three default schemes for users to use: temperature adjustment scheme, load ratio adjustment scheme, and soft shutdown scheme. Users can configure the default scheme themselves, including creating a new default scheme and removing a scheme from the default schemes. If a default scheme exists, the system will synchronize the response capabilities of the third-party system before selecting the scheme; otherwise, it will select the circuit based on the absence of a default scheme.

[0150] Step 6: If a default solution exists, the response capability of each default solution will be calculated and sorted from smallest to largest. The solution's response capability will be compared with the load gap at the current time. The first default solution with a response capability larger than the load gap will be selected. If the response capability of all solutions is smaller than the current load gap, the default solution with the largest response capability will be selected. Additionally, because the temperature control solution has a longer ramp-up time, it will only be used before the event begins; it will not be included when selecting the default solution during execution.

[0151] Step 7: Select branch lines from the chosen schemes in sequence and send control commands. For a rotating shutdown scheme, only one multi-unit branch line can be selected per round of control; for a non-rotating shutdown scheme, there is no such restriction. Branch line selection follows the order within the scheme. Furthermore, branches already controlled in this event will not be selected again unless the available branch line response capacity is less than the current load gap, in which case a branch line will be selected from the previously controlled branches for execution.

[0152] Step 8: If the event ends, control ends; otherwise, wait for the next control time to expire before resuming control.

[0153] Step 9: If no default solution was found in Step 5, this step will proceed. Here, all branches in the system will be sorted from highest to lowest response capacity. First, the branch with the highest response capacity will be selected, and it will be checked whether the current load gap is met. If so, no new branch will be added; otherwise, the branch with the lowest response capacity will be added, and the process will repeat. The highest response capacity branch will be added from the remaining branches, and this process will continue until the current load gap is met or all branches have been added. After determining the added branches, a control command will be sent to initiate control. Then, it will be checked whether the event has ended. If it has ended, control will stop; otherwise, the process will wait for the next control time to occur before resuming control in this step.

[0154] This embodiment utilizes an automatic control algorithm that flexibly adapts to user-defined strategies, ensuring full respect for and efficient implementation of personalized user needs. It boasts high compatibility, seamlessly integrating with user-defined strategy templates and prioritizing user-edited strategies. When users actively participate in strategy customization, the system accurately executes user intentions, ensuring that demand response activities better align with actual user needs. When users do not edit strategies, a preset intelligent execution mode is automatically activated, improving response efficiency, reducing the user's decision-making burden during demand response, and enhancing the execution effect. This not only enhances user participation and control but also improves the system's intelligence level. The system integrates three efficient strategies by default: temperature control, load ratio adjustment, and soft shutdown. These strategies can seamlessly interface with third-party systems (such as air conditioning control systems and intelligent lighting systems), enabling these devices to actively participate in demand response and optimize energy consumption. Users can also customize and edit strategy schemes according to their own power consumption characteristics and needs. Through simple operations, users can set their edited schemes as the default execution schemes. When a demand response event is triggered, the system will automatically employ these customized solutions to ensure that the demand response activity is both user-friendly and efficiently executed. This direct, automatic execution function, requiring no manual intervention, provides users with an exceptionally convenient experience. Even without user-edited execution strategies or default settings, the system can still operate independently and automatically execute demand response tasks. This feature is designed with user time value and ease of operation in mind, allowing users to complete the entire demand response process without any manual intervention. This fully automated execution mode not only reduces user workload and improves user satisfaction but also ensures the efficient and stable execution of demand response activities, providing strong support for the safe operation of the power system.

[0155] This invention constructs a complete multi-scenario additional control routing decision system by organically combining tensor decomposition, graph neural networks, and reinforcement learning during the demand response execution process. First, tensor decomposition extracts scenario features from multi-source heterogeneous data, accurately identifying different demand response scenarios such as high temperature, holidays, and peak-shifting electricity usage. Second, graph neural networks analyze the correlations between routing routes, accurately evaluating the response capabilities of each route under different scenarios, effectively avoiding chain reactions caused by load shifting. Third, a multi-objective optimization method is used to balance multiple objectives such as load reduction and user comfort, dynamically adjusting optimization weights based on real-time scenario characteristics. Finally, a genetic algorithm generates a control strategy sequence, and a temporal dependency graph ensures the rationality of the execution order. Practical applications show that this invention increases the average load reduction rate of demand response projects to 18%, user comfort satisfaction to 92%, and the control strategy execution success rate to 95%, improving the accuracy and effectiveness of demand response.

[0156] The preferred embodiments of the present invention have been described in detail above. However, the present invention is not limited to the specific details in the above embodiments. Within the scope of the technical concept of the present invention, various equivalent transformations can be made to the technical solutions of the present invention, and these equivalent transformations all fall within the protection scope of the present invention.

Claims

1. A method for adding control branches in multiple scenarios during the execution of demand response, characterized in that: The steps include: S1. Obtain multimodal data, process it using tensor decomposition method and support vector machine, and obtain scene prediction sequence and scene type identification; wherein the multimodal data includes real-time load data, historical load data, environmental parameter data and user behavior data; S2, read the scenario prediction sequence and scenario type identification, combine the real-time load data, historical load data and pre-stored system constraint data, use fuzzy comprehensive evaluation method and graph neural network for processing, and obtain the branch response capability matrix and control constraint matrix; S3, based on the scene type identification, scene prediction sequence, branch response capability matrix, control constraint matrix and user behavior data, a multi-objective optimization method is used to obtain an optimization objective function set, a weight coefficient matrix and a constraint condition set; S4. Based on the optimization objective function set, weight coefficient matrix and constraint condition set, genetic algorithm and dynamic programming method are used to obtain the control strategy matrix and execution timing vector.

2. The method for adding control branches in multiple scenarios during the demand response execution process according to claim 1 is characterized in that: Step S1 is further as follows: S11. Acquire multimodal data from a data acquisition system, including real-time load data, historical load data, environmental parameter data, and user behavior data; use a minimum-maximum normalization method to map the multimodal data to the interval [0, 1] to obtain normalized multimodal data; based on the normalized multimodal data, use a moving average method to smooth outliers exceeding three times the standard deviation to obtain smoothed multimodal data; based on the smoothed multimodal data, use a piecewise linear interpolation method to supplement missing values ​​in the data sequence to obtain multimodal data with supplemented missing values; based on the multimodal data with supplemented missing values, reorganize according to the data type dimension, time series dimension, and sampling frequency dimension to obtain a preprocessed data tensor; S12, based on the preprocessed data tensor, calculating the projection matrix along the data type dimension, time series dimension and sampling frequency dimension respectively; Perform singular value decomposition on each projection matrix, and select singular values ​​and their corresponding eigenvectors based on the principle that the cumulative contribution rate is greater than 95%; combine the selected eigenvectors to form a data type dimension factor matrix, a time series dimension factor matrix, and a sampling frequency dimension factor matrix; project the preprocessed data tensor simultaneously into the space formed by the data type dimension factor matrix, the time series dimension factor matrix, and the sampling frequency dimension factor matrix to obtain the characteristic core tensor; S13, respectively calculating the projection values ​​of the feature core tensor in the direction of the data type dimension factor matrix, the time series dimension factor matrix and the sampling frequency dimension factor matrix; performing weighted summation on the projection values ​​according to the corresponding singular value sizes to obtain a weighted summation result; using the principal component analysis method to select the first p feature components with a cumulative contribution rate of 90% from the weighted summation result, and combining the p feature components to form a scene feature vector; Where p is a positive integer; S14. Input the scene feature vector into a pre-trained support vector machine classifier for scene classification to obtain the scene type identification at the current moment; at the same time, input the scene feature vectors of the most recent N moments into the long short-term memory network in chronological order to predict the scene evolution trend at the next M moments, and obtain a scene prediction sequence containing M predicted scene types; wherein N and M are both positive integers.

3. The method for adding control branches in multiple scenarios during demand response execution according to claim 2 is characterized in that: Step S2 is further as follows: S21, based on the real-time load data, historical load data and pre-stored system constraint data, a sliding window method is used to extract the load change characteristics of each branch; Based on the load variation characteristics of each branch, the upper and lower limits of each branch are calculated by using the data statistics method; based on the upper and lower limits of each branch, the available status flag of the branch is determined by using the threshold determination method; Based on the load variation characteristics of each branch, the upper and lower limit ranges of each branch and the available state flag, a branch state vector is formed; S22, based on the branch state vector, the scenario prediction sequence and the scenario type identifier, a fuzzy membership function is used to calculate the membership value of each branch on the power regulation capability, response speed and reliability evaluation indicators; based on the membership value, a hierarchical analysis method is used to determine the weight coefficient of each evaluation indicator; Based on the weight coefficient, the membership value of each indicator is weighted and calculated to obtain the branch response capability matrix; S23. Based on the branch response capacity matrix and historical load data, an adjacency matrix of the graph neural network is constructed, each branch is regarded as a node in the graph, and the Pearson correlation coefficient is used to calculate the correlation strength between branches as the weight of the edge; Based on the adjacency matrix, the influence relationship between nodes is calculated through the message passing mechanism of the graph neural network to obtain the branch association matrix; S24, extracting constraint conditions based on system constraint data, including power grid operation constraints, equipment operation constraints and user comfort constraints; Based on the branch association matrix, the quantitative indicators of the constraint conditions are calculated; the quantitative indicators of all constraint conditions and the preset thresholds are combined to form a control constraint matrix.

4. The method for adding control branches in multiple scenarios during demand response execution according to claim 3 is characterized in that: Step S3 is further as follows: S31. Based on the scenario type identification and the branch response capability matrix, construct an optimization objective function, including a load reduction objective function, a user comfort objective function, and an equipment life objective function; standardize the variable range and constraint conditions of each optimization objective function to obtain a processed optimization objective function; combine all processed optimization objective functions to form an optimization objective function set; S32. Based on the scenario prediction sequence and user behavior data, a judgment matrix is ​​constructed using the hierarchical analysis method to calculate the relative importance of each optimization objective in the optimization objective function set; Based on the relative importance, consistency check is used to ensure the rationality of weight allocation, and the weight coefficient of each optimization objective is output; Combine the weight coefficients of all optimization objectives to form a weight coefficient matrix; S33, based on the control constraint matrix and the branch association matrix, extracting the boundary values ​​and thresholds of various constraint conditions, and constructing mathematical expressions of the constraint conditions; Based on mathematical expressions, the constraints are normalized to obtain a constraint set.

5. The method for adding control branches in multiple scenarios during the demand response execution process according to claim 4 is characterized in that: Step S4 is further as follows: S41, obtaining an optimization objective function set and a constraint condition set, and using a genetic algorithm based on non-dominated sorting to generate an initial population; verifying the constraint conditions for each individual in the initial population, and generating new individuals through crossover and mutation operations, and obtaining a candidate strategy set containing feasible control strategies after t generations of iterative optimization; wherein t is a positive integer; S42. Based on the candidate strategy set and the weight coefficient matrix, each candidate strategy is scored using the Pareto optimal principle, and the comprehensive score of each strategy on different objectives is calculated; all candidate strategies are sorted in descending order according to the comprehensive score to obtain a strategy scoring vector and a strategy sorting sequence; S43, based on the strategy scoring vector and the strategy sorting sequence, using a reinforcement learning method to extract strategy selection experience from pre-stored historical execution data; Based on the experience of strategy selection, the top-ranked predetermined strategies are combined and optimized to obtain the control strategy matrix; S44, based on the control strategy matrix and the branch association matrix, a dynamic programming algorithm is used to construct a state transfer equation; Based on the state transfer equation, the minimum time interval between adjacent control actions is calculated, the execution time point of each control action is optimized, and the execution timing vector is obtained.

6. The method for adding control branches in multiple scenarios during demand response execution according to claim 5, characterized in that: Step S11 is further as follows: S111, acquiring real-time load data, historical load data, environmental parameter data, and user behavior data from a data acquisition system, calculating the mean and standard deviation of each type of data and then standardizing them to obtain standardized data; transforming the standardized data through an inverse hyperbolic tangent function mapping to form a standardized data set; S112. Based on the standardized data set, a multi-layer wavelet decomposition network is constructed to perform frequency stratification on the data sequence to obtain a frequency stratified data sequence; based on the frequency stratified data sequence, a threshold parameter is adaptively adjusted in each layer of decomposition to separate high-frequency noise features from data features to obtain denoised data features; based on the denoised data features, the denoised data set is restored by inverse reconstruction; S113. Based on the denoised data set, a conditional probability model of the data time series is constructed, and feature probability distribution is calculated according to the time dependency of the data; Based on the feature probability distribution, the Markov chain Monte Carlo method is combined to iteratively optimize the estimated value of the missing data to generate a complete data set; S114. Based on the complete data set, extract the three-dimensional features of time-frequency-phase through tensor decomposition; reconstruct the three-dimensional features according to the three dimensions of data type, time series, and sampling frequency to form a preprocessed data tensor.

7. The method for adding control branches in multiple scenarios during the demand response execution process according to claim 5 is characterized in that: Step S12 is further as follows: S121. Based on the preprocessed data tensor, an adaptive projection network is constructed to calculate the feature space of the data type dimension, the time series dimension, and the sampling frequency dimension respectively; through the forward propagation of the adaptive projection network, the projection feature matrix of the data type dimension, the time series dimension, and the sampling frequency dimension is obtained; S122. Constructing a block diagonalization matrix based on the projection feature matrix; Based on the block diagonalized matrix, the sparse singular value decomposition method is used to extract the main characteristic components and obtain the characteristic component matrix; S123. Based on the feature component matrix, an adaptive feature selection model is constructed to calculate the cumulative contribution rate of each feature; based on the cumulative contribution rate, the feature selection threshold is dynamically adjusted to obtain an adjusted feature selection threshold; based on the adjusted feature selection threshold, the most representative feature vector is selected to form a data type dimension factor matrix, a time series dimension factor matrix, and a sampling frequency dimension factor matrix in combination; S124. Based on the data type dimension factor matrix, the time series dimension factor matrix and the sampling frequency dimension factor matrix, a multimodal tensor projection network is constructed; based on the multimodal tensor projection network, the preprocessed data tensor is projected into the feature space formed by the factor matrix through tensor product operation to generate a feature core tensor.

8. The method for adding control branches in multiple scenarios during demand response execution according to claim 5, characterized in that: Step S14 is further as follows: S141. Based on the scene feature vector, a multi-head attention network is constructed to extract the spatiotemporal dependency between features. Based on the spatiotemporal dependency, long-term and short-term feature patterns are captured through the self-attention mechanism to generate a feature dependency matrix. S142. Based on the feature dependency matrix, a hierarchical bidirectional neural network is constructed to encode and decode feature sequences of different time scales to form a time series encoding vector; S143, constructing a mixed Gaussian process model based on the time series coding vector, and calculating the probability distribution of the scene prediction; based on the probability distribution of the scene prediction, outputting a scene prediction sequence including M prediction moments, where M is a positive integer; S144. Based on the scene prediction sequence, the credibility of the prediction result is calculated through the Bayesian inference network to generate a scene type identification and a corresponding confidence score.

9. The method for adding control branches in multiple scenarios during the demand response execution process according to claim 5, characterized in that: Step S22 is further as follows: S221, based on the branch state vector, the scene prediction sequence and the scene type identifier, construct a multi-layer evaluation index system, calculate the quantitative index of each branch in the adjustment capability, response speed and reliability dimensions, and form an evaluation index matrix; S222, based on the evaluation index matrix, construct a dynamic fuzzy membership function, realize the mapping of index to fuzzy set through adaptive parameter adjustment, and generate a fuzzy evaluation matrix; S223, obtaining a fuzzy evaluation matrix, constructing a hierarchical analysis model based on entropy weight, dynamically calculating the weight coefficient of each evaluation index, and outputting a weight vector; S224. Based on the fuzzy evaluation matrix and the weight vector, the comprehensive score of each branch is calculated by an improved fuzzy comprehensive evaluation algorithm to form a branch response capability matrix.

10. The method for adding control branches in multiple scenarios during the demand response execution process according to claim 5, characterized in that: Step S31 is further as follows: S311. Based on the scenario type identification and the branch response capability matrix, a multi-layer target decomposition network is constructed to decompose the overall optimization target into sub-targets including load regulation, user comfort and equipment operation, and a sub-target function set is generated; S312, based on the sub-objective function set, realize the dimension reduction mapping of the target space through the variational autoencoder, and output the target feature matrix; S313. Based on the target feature matrix, a target fusion network is constructed, and the adaptive combination of multiple targets is realized through a dynamic weight adjustment mechanism to form an optimized target function set.

Images