Power demand multi-algorithm collaborative demand prediction method and system fused with deep learning

Abstract

The invention provides a deep learning-fused power demand multi-algorithm collaborative demand prediction method and system, and the system comprises a short-term power demand prediction platform, a medium and long term power demand prediction platform, a collection management platform, and a display analysis platform. According to the whale optimization algorithm, the generalization ability is improved by introducing chaotic mapping and a nonlinear convergence factor, the weight is optimized in combination with a multi-scale graph convolutional network, the number of nodes, the learning rate and the number of iterations of a model are optimized in combination with a simulated annealing algorithm, the model is constructed, and by setting a medium-and-long-term power demand prediction platform, the power demand prediction efficiency is improved. Based on the Tent chaotic mapping and the dynamic step length factor optimization standard sparrow search algorithm, the prediction result is accurate and reliable by optimizing the key parameters of the composite model combining the convolutional neural network and the bidirectional long and short term memory network with the attention mechanism, and the power demand prediction cost is reduced.

Description

Technical Field

[0001] This invention relates to the field of electricity demand forecasting, specifically to a collaborative forecasting method and system for electricity demand using multiple algorithms that integrate deep learning. Background Technology

[0002] With the rapid development of renewable energy, electricity demand forecasting is becoming increasingly important in ensuring the stable operation of grids with a high proportion of renewable energy. Due to the intermittency and uncertainty of renewable energy generation, accurate forecasting of electricity demand can effectively support power dispatch and energy planning, promote energy structure transformation and address climate change.

[0003] Accurate electricity demand forecasting helps power dispatching departments understand electricity demand in advance, thereby optimizing power dispatching strategies and avoiding power shortages or surpluses caused by demand fluctuations. Simultaneously, by forecasting future electricity demand, power companies can better plan the construction and expansion of power facilities, ensuring the stability and reliability of power supply. It also improves economic and social benefits. Accurate load forecasting allows power companies to rationally formulate generator start-up and shutdown plans within the grid, maintaining the safe and stable operation of the grid and reducing unnecessary spinning reserve capacity. This helps to rationally arrange generator maintenance plans, effectively reducing power generation costs, ensuring normal production and life in society, and improving economic and social benefits. Furthermore, it promotes green and sustainable development. Intelligent electricity demand forecasting helps improve energy efficiency, reduce unnecessary energy waste, and achieve green and sustainable development. Accurate forecasting and optimization of electricity use can reduce energy consumption and emissions, which is of great significance to environmental protection and sustainable development. Finally, it supports the development of the electricity market. Electricity demand forecasting has a significant impact on the electricity market, especially in short-term demand forecasting. It helps marketers, power generation companies, and distribution companies better understand market demand, formulate reasonable market strategies and operating plans, thereby improving market efficiency and competitiveness.

[0004] However, traditional electricity demand forecasting has the following drawbacks: the electricity demand forecasting models in traditional electricity demand forecasting systems may be outdated and inaccurate, leading to frequent errors in forecasting results and increasing the cost of electricity demand forecasting. Summary of the Invention

[0005] This invention addresses the problems existing in the prior art by providing a collaborative prediction method and system for electricity demand using multiple algorithms that integrate deep learning. This solves the problem mentioned in the background art that the electricity demand prediction model in traditional electricity demand prediction systems may be outdated and inaccurate, leading to frequent errors in prediction results and increasing the cost of electricity demand prediction.

[0006] To address the above technical problems, this invention provides the following technical solution: a multi-algorithm collaborative prediction method for electricity demand integrating deep learning, characterized by comprising the following steps:

[0007] S1. Based on the time series data of electricity load, the influencing factors are screened by gradient boosting tree and lasso regression methods, and collinearity is eliminated. The influencing factors include: meteorological data, holiday information, industrial activity index and regional economic indicators.

[0008] S2. Based on the improved Whale Optimization Algorithm (IWOA), Simulated Annealing Algorithm (SA), and Multi-Scale Graph Convolutional Network (MSGNet), a short-term power demand forecasting model IWOA-SA-MSGNet is constructed. The model takes influencing factors as input and outputs short-term power demand forecasting results. The Multi-Scale Graph Convolutional Network (MSGNet) extracts the daily, weekly, and monthly periodic features of the load through frequency domain decomposition.

[0009] An improved Whale Optimization Algorithm (IWOA) is used to optimize the sparsity of the graph adjacency matrix in the multi-scale graph convolutional network MSGNet.

[0010] Simulated annealing (SA) algorithm is used to optimize the number of nodes, learning rate, and number of iterations in the multi-scale graph convolutional network MSGNet.

[0011] S3. Based on the Convolutional Neural Network (CNN), the Bidirectional Long Short-Term Memory (BiLSTM) network, and the Attention mechanism, a medium- to long-term electricity demand forecasting model, ISSA-CNN-BiLSTM-Attention, is constructed. It takes influencing factors as input and outputs medium- to long-term electricity demand forecasting results. The CNN is used to extract local temporal features; the BiLSTM network is used to capture long-distance dependencies; and the Attention mechanism is used to enhance key features. The ISSA-CNN-BiLSTM-Attention model is an improved sparrow search algorithm based on the dynamic step size factor optimization of the Tent chaotic mapping. Specifically, it optimizes the number of convolutional kernels in the CNN, the number of hidden units in the BiLSTM network, and the number of attention heads.

[0012] Furthermore, in step S1 above, the formula for calculating the gradient boosting tree is:

[0013] ,

[0014] in, This represents the initial prediction function of the model at the start of the iteration; It's the learning rate. It is the m-th weak learner, where M represents the total number of weak learners. It is the final predicted value of the gradient boosting tree;

[0015] The formula for calculating using the lasso regression method is:

[0016] ,

[0017] in, It is the predicted value of the target variable Y. It is the first One influencing factor, It is the first Estimated coefficient values ​​of each influencing factor. It is the constant term in the regression model.

[0018] Furthermore, in step S2 above, the whale population location is initialized by improving the whale optimization algorithm IWOA, using the following formula:

[0019] ,

[0020] in, Let be the position of the w-th individual whale, LB and UB be the lower and upper bound vectors of the search space, respectively, and rand be a random vector in the range [0,1].

[0021] The current optimal prey position is defined. Then, the position update formula is:

[0022] , , , ,

[0023] in, and Let represent two random numbers uniformly distributed in the interval [0,1], where Controlling the convergence factor Adjust the search intensity; Control scaling factor Adjusting the direction and magnitude of the gravitational force between the individual and the current optimal solution;

[0024] z represents the number of iterations. It is the position vector of the w-th individual whale at iteration z. It is the position vector of the current optimal solution. It is a distance vector. It is a control parameter vector;

[0025] The formula for calculating the optimized weights of the multi-scale graph convolutional network MSGNet is as follows:

[0026] ,

[0027] in, This represents the output feature vector of node i. Let i represent the set of neighboring nodes. This represents the input feature vector of neighbor node j. This represents a trainable weight matrix. Represents the bias vector;

[0028] The annealing algorithm SA specifically includes:

[0029] Initialization parameters: Initial temperature Lower limit of temperature Initial solution state Each temperature The number of iterations L, the temperature decay coefficient ;

[0030] The current solution state is denoted as ;

[0031] Execute the following loop until the termination condition is met:

[0032] Step a1, for l=1, 2, ..., L, execute:

[0033] Step a2, generating a new solution : ,

[0034] in This represents the amount of perturbation or change in the solution state during the current iteration, i.e., from the current solution... Generate new solutions The introduced fine-tuning perturbation;

[0035] Step a3, calculate the increment of the objective function. f represents the objective function used to evaluate the quality of the current solution state S;

[0036] Step a4, if Then accept the new solution; if Then based on probability Accept new solutions; exp represents the natural exponential function;

[0037] Step a5, update the temperature at time k+1. .

[0038] Furthermore, the aforementioned improved sparrow search algorithm ISSA includes the following steps:

[0039] Step b1, Initialize the population: For the s-th sparrow, s = 1, 2, ..., N, where N represents the total number of sparrows, initialize the position vector of the s-th sparrow. :

[0040] ;

[0041] Step b2, Search for food: Calculate the current position of each sparrow. fitness value;

[0042] Step b3, Discoverer location update:

[0043] ,

[0044] in This indicates that the s-th sparrow is in The position at time t, Q represents a random variable that follows a standard normal distribution Q∼N(0,1);

[0045] Step b4, follower position update: ,

[0046] in Indicates the current number The position vector of the sparrow with the best fitness in the next iteration; d represents the dimension of the current sparrow; A + Let represent a random number that satisfies the normal distribution constraint and satisfies . >0;

[0047] Step b5, Guardian location update:

[0048] ,in This represents the preset disturbance coefficient of the vigilant;

[0049] Step b6: Repeat steps b2 to b5 until the stopping condition is met.

[0050] This invention also provides a multi-algorithm collaborative prediction system for electricity demand integrating deep learning, comprising:

[0051] Short-term electricity demand forecasting platform, medium- and long-term electricity demand forecasting platform, data collection and management platform, and display and analysis platform;

[0052] The short-term power demand forecasting platform forecasts short-term power demand.

[0053] The medium- and long-term electricity demand forecasting platform forecasts medium- and long-term electricity demand.

[0054] The collection and management platform achieves automated training of targeted datasets through the following methods:

[0055] Multi-source heterogeneous data integration: Integrating electricity load time series data, meteorological data, holiday information, industrial activity index, and regional economic indicators to construct a multi-dimensional dataset; Dynamic feature engineering: Extracting frequency domain features from electricity load time series data through Fast Fourier Transform; Combining sliding window statistical features and external factor interaction features, defining the correlation strength between features and the target variable through Fast Fourier Transform and mutual information method:

[0056] , It refers to the first Variables and the Variables Normalized mutual information; For variables and Mutual information;

[0057] Automated training process: Dynamically optimize model hyperparameters based on the improved Whale Algorithm (IWOA); balance global exploration and local exploitation through simulated annealing (SA); automatically divide the training, validation, and test sets.

[0058] Assistive parameter tuning interface: Provides a visual interface that allows experts to adjust feature weights and inject domain knowledge;

[0059] The visualization and analysis platform provides an intuitive display of prediction results, helping users understand the predictions and make informed decisions.

[0060] Furthermore, the aforementioned short-term electricity demand forecasting platform includes a first influencing factor screening module, a whale optimization algorithm module, and a simulated annealing algorithm module;

[0061] The first influencing factor screening module screens influencing factors using gradient boosting tree and lasso regression methods, and eliminates collinearity;

[0062] The whale optimization algorithm module improves generalization ability by introducing Tent chaotic mapping and nonlinear convergence factor, and optimizes weights by combining multi-scale graph convolutional network.

[0063] The simulated annealing algorithm module is used to optimize the number of nodes, learning rate, and number of iterations of the IWOA-SA-MSGNet model, and to build a high-precision prediction model by co-optimizing hyperparameters;

[0064] The IWOA-SA-MSGNet model is a three-level collaborative optimization architecture that integrates the improved whale optimization algorithm (IWOA), the simulated annealing strategy (SA), and the multi-scale graph convolutional network (MSGNet), including:

[0065] Global parameter optimization layer: Improves the model topology and hyperparameter combination by refining the IWOA (Integrated Whale Optimization) algorithm;

[0066] Local hyperparameter tuning layer: Optimizes dynamic parameters during training using the simulated annealing (SA) strategy;

[0067] Multi-scale feature extraction layer: The frequency domain features and spatiotemporal correlation of power load are captured by the multi-scale graph convolutional network MSGNet.

[0068] The collection and management platform achieves autonomous optimization of electricity demand forecasting through heterogeneous algorithm fusion and dynamic collaboration mechanism. The collection and management platform includes a parameter dynamic coupling module based on streaming clustering, a multimodal data governance module, an interpretable factor interaction engine, a meta-enhanced prediction center, and a causal inference optimization module.

[0069] The parameter dynamic coupling module based on streaming clustering employs an improved K-means algorithm to perform incremental clustering analysis on real-time power load data. It introduces the Dynamic Time Warped (DTW) distance as a similarity measure for load curves and constructs an adaptive weighting function based on historical holiday influence factors. This achieves collaborative optimization of peak period segmentation and user-defined holiday rules. Each time new power load time series data arrives, the K-means algorithm is used to perform clustering analysis, with the following steps:

[0070] SS1, Initialization Cluster centers, Indicates the first Cluster centers, Then calculate each data point Euclidean distance to all cluster centers :

[0071] ,

[0072] in, It is the first Data points To the Cluster centers Euclidean distance, It is the first Data points In the Dimension value, It is the first The cluster centers at the in Dimension value; data points Assigned to the nearest cluster center :

[0073] ,

[0074] in Indicates the first Data points The cluster number assigned to the cluster; This indicates that the objective function Cluster center number when the minimum value is obtained ;

[0075] SS2, Update Cluster Centers : ,

[0076] in, It is the first The set of all data points in a cluster It is the first The number of data points in each cluster;

[0077] SS3 uses Dynamic Time Warping (DTW) distance to measure two power load curves. and Similarities between them:

[0078] , ,

[0079] in Represents a sequence The power load value of the Mth sampling point; Represents a sequence The power load value of the Nth sampling point; Indicates from sequence The starting point To the finish line and from the sequence The starting point To the finish line The cumulative distance of the best matching path formed;

[0080] , It is Euclidean distance. Representing a subsequence and DTW distance between them; where subsequences It is a sequence The subsequence formed by the first point to the i-th point, It is a sequence The subsequence formed by the first point to the j-th point;

[0081] Equivalent to power load time series subsequence of and subsequence of The DTW distance between them is expressed as:

[0082] ;

[0083] The parameter dynamic coupling module (41) based on streaming clustering introduces an adaptive weight function based on historical holiday influencing factors. The weight function is dynamically adjusted in conjunction with holiday rules to achieve collaborative optimization between peak period segmentation and user-defined holiday rules.

[0084] ,

[0085] in This represents the weight at time point t. A rule function indicating whether time point t is a holiday. The adjustment coefficient representing the historical high load factor. This indicates the adjustment coefficient for holiday rules.

[0086] Furthermore, the aforementioned multimodal data governance module establishes a dual-channel anomaly detection architecture based on an attention mechanism. It fuses the path length features of the Isolation Forest algorithm with the improved sliding window 3σ criterion at the feature level, and dynamically selects either LSTM interpolation or robust median imputation strategy through a hybrid interpolation network controlled by a gated recurrent unit (GRU). The specific implementation process includes: establishing a dual-channel anomaly detection architecture based on an attention mechanism. In this architecture, power load time series data is first processed in parallel through two channels. One channel processes standard input data, and the other channel processes features enhanced by multimodal information. In each channel, a weighted method based on an attention mechanism is used to improve the representation of important features, as shown in the formula:

[0087] ,

[0088] in, It is a weighted feature representation. It is the attention coefficient. Indicates the first The importance of each feature The number of features;

[0089] Path length of each tree Data points The degree of isolation for a data point data points Path length The calculation formula is:

[0090] ,

[0091] in, For the number of trees, It is in the Data points in a tree Path length when isolated;

[0092] Using path length Calculate abnormal scores :

[0093] ,

[0094] in, It is related to the set of quantities Size The relevant normalization constants;

[0095] The improved sliding window 3σ criterion is used for real-time detection of outliers in a data stream, with each window... Include 1 data point, 1 window Data points within The process for evaluating whether a data point is an outlier is as follows: Calculate the mean of the data within the window. and standard deviation :

[0096] , ,

[0097] According to the improved sliding window 3σ criterion, if the data points The value satisfies: Then determine the data points This is abnormal data;

[0098] The formula for the Long Short-Term Memory (LSTM) network is:

[0099] ,

[0100] in, For the data at the current time step t, Let be the hidden state vector of the previous time step t-1. Let be the hidden state vector of the previous time step t;

[0101] Robust median imputation strategy: If data points are missing within a window, the median of the non-missing data points within the window is used. As an alternative:

[0102] ,

[0103] in, The interpolated value at time point t; This represents the (t+k)th data point;

[0104] GRU control logic: The gated loop unit (GRU) dynamically selects the interpolation strategy based on the current time step input, using the following formula:

[0105] ,

[0106] in This represents the output of the gated recurrent unit (GRU) at time point t. The input vector representing time point t;

[0107] Furthermore, the aforementioned interpretable factor interaction engine constructs a factor selection model based on a knowledge graph in the power sector. It utilizes a SHAP interpreter optimized through transfer learning to generate a feature importance graph and designs a three-dimensional tensor decomposition structure to capture higher-order interaction effects of meteorological, economic, and social factors. Simultaneously, it embeds a lightweight online learning architecture to achieve real-time differential updates of factor weights. The specific process includes: constructing a factor selection model based on a knowledge graph in the power sector, mapping meteorological factor M, economic factor E, and social factor S to graph nodes, and modeling factor relationships using a graph embedding method. The graph embedding optimization objective is to minimize the distance between the embedded representations of nodes and their neighboring nodes.

[0108] ,

[0109] in, This represents the value of node e that minimizes the objective function;

[0110] and Let n be the embedding representation of node n and let o be the embedding representation of node o. Represents the set of neighboring nodes of node n;

[0111] Using a SHAP interpreter optimized by transfer learning, a contribution map of factors to the prediction results is generated, and the contribution of each factor to the model output is evaluated. The optimization objective of transfer learning is:

[0112] ,

[0113] in, This represents the total loss function in transfer learning; and These are the loss functions for the target task and the source task, respectively. and These are the parameters for the target task and the parameters for the source task, respectively. It is a hyperparameter that balances the source and target tasks;

[0114] 3D Tensor Decomposition: In order to capture the higher-order interaction effects of meteorological factor M, economic factor E, and social factor S, the input core tensor is decomposed. Decomposed into a core tensor G and a factor matrix , , The product of:

[0115] ,

[0116] Where G is the core tensor, , , These are factor matrices representing meteorological, economic, and social factors, with symbols... The tensor represents the tensor in the first place. The product of the modulus and the matrix. The values ​​are 1, 2, and 3; the engine updates the factor weights in real time using a lightweight online learning architecture.

[0117] ,

[0118] in, Indicates the first Factor weight vector for the next iteration It's the learning rate. It is a loss function Regarding the factor weight vector The gradient;

[0119] The meta-enhanced prediction center (44) creates a multi-scale prediction model group containing a spatiotemporal attention mechanism, and simultaneously optimizes the model weights and feature weights using a dual PPO algorithm framework. A reward function R based on the prediction error gradient is designed.

[0120] ,

[0121] Where MAE represents the mean absolute error. Indicates dynamic interval coverage. and These are the weighting coefficients;

[0122] The meta-enhanced prediction hub implements distributed training through an asynchronous parameter server architecture;

[0123] The causal inference optimization module constructs a dynamic scheme evaluation system based on Bayesian Structured Time Network (BSTM) for quantitatively evaluating the causal effects of power strategy adjustments, specifically including:

[0124] Constructing a Bayesian structured time network: Introducing a structured Bayesian network in the form of a directed acyclic graph (DAG), with random variable nodes. Represents the first time at time t. Due to individual states, network evolution follows the Markov assumption, and the joint probability model is as follows:

[0125] ,

[0126] in Represents a node The set of parent nodes is obtained by the structure learning method; This represents the overall joint probability distribution of all variable nodes from time step 1 to T; This represents the nth variable node at time t. Given its parent node set Conditional probability distribution under given conditions;

[0127] T represents the total number of time steps in the evolution of the dynamic Bayesian network;

[0128] Design a cost-sensitive loss function Introducing class weight vectors Adjust the ordinary cross-entropy loss to a weighted form:

[0129] ,

[0130] Where K represents the category index, ;

[0131] Indicates the first Class weight coefficient; This represents the one-hot encoded value of the true label, if the sample belongs to the first... The value is 1 if the condition is met, and 0 otherwise. The model represents the first The predicted probability of a class;

[0132] Counterfactual reasoning mechanism: Introducing a dual-channel reasoning structure, based on actual observation data... Based on this, through the analysis of decision variables Intervention to construct counterfactual samples The output is obtained through the inference network. The difference between factual and counterfactual outcomes is compared to assess the causal impact of the proposed solutions; among which... This represents the intervention operation in the cause-effect diagram; Indicates the set decision variables;

[0133] VAE-generated adversarial sample robustness verification mechanism: Constructing a variational autoencoder structure:

[0134] ,

[0135] Combined with counter-disturbance Generate adversarial samples ;

[0136] in Represents the variational loss function of a VAE;

[0137] This represents the posterior distribution of the encoder output. The following expectations;

[0138] This represents the reconstruction probability distribution generated by the decoder;

[0139] express Divergence, used to measure the encoder posterior With prior distribution The differences between them.

[0140] Compared with the prior art, the beneficial technical effects of the present invention using the above technical solution are as follows:

[0141] (1) By setting up a short-term power demand forecasting platform, the Whale Optimization Algorithm improves its generalization ability by introducing Tent chaotic mapping and nonlinear convergence factor, and optimizes the weights by combining multi-scale graph convolutional network. Combined with simulated annealing algorithm, the number of nodes, learning rate and number of iterations of the model are optimized, and the IWOA-SA-MSGNet model is constructed. The prediction results are accurate and reliable, and the cost of power demand forecasting is reduced.

[0142] (2) By setting up a medium- and long-term power demand forecasting platform, an improved sparrow search algorithm was proposed based on Tent chaotic mapping and dynamic step size factor optimization of the standard sparrow search algorithm. By optimizing the key parameters of the composite model of convolutional neural network and bidirectional long short-term memory network combined with attention mechanism through ISSA, an ISSA-CNN-BiLSTM-Attention model was constructed. The prediction results are accurate and reliable, and the cost of power demand forecasting is reduced. Attached Figure Description

[0143] Figure 1 This is a schematic diagram of the architecture of the demand forecasting system of the present invention;

[0144] Figure 2 This is a schematic diagram of the architecture of the short-term power demand forecasting platform of the present invention;

[0145] Figure 3 This is a schematic diagram of the architecture of the medium- and long-term power demand forecasting platform of this invention.

[0146] Figure 4 This is a schematic diagram of the architecture of the data collection and management platform of the present invention;

[0147] Figure 5This is a schematic diagram illustrating the architecture of the analysis platform for this invention;

[0148] Figure 6 This is a flowchart of the present invention.

[0149] In the figure, 1-prediction system; 2-short-term power demand forecasting platform; 21-first influencing factor screening module; 22-whale optimization algorithm module; 23-simulated annealing algorithm module; 3-medium- and long-term power demand forecasting platform; 31-second influencing factor screening module; 32-sparrow search algorithm module; 33-ISSA-CNN-BiLSTM-Attention model module; the collection and management platform 4 includes: 41-parameter dynamic coupling module based on streaming clustering; 42-multimodal data governance module; 43-interpretable factor interaction engine; 44-meta-reinforcement prediction center; 45-causal inference optimization module. Detailed Implementation

[0150] To better understand the technical content of the present invention, specific embodiments are described below in conjunction with the accompanying drawings.

[0151] In this invention, various aspects of the invention are described with reference to the accompanying drawings, in which numerous illustrative embodiments are shown. Embodiments of the invention are not limited to those depicted in the drawings. It should be understood that the invention is implemented through any of the various concepts and embodiments described above, as well as the concepts and embodiments described in detail below, because the concepts and embodiments disclosed herein are not limited to any particular implementation. Furthermore, some aspects of the invention disclosed may be used alone or in any suitable combination with other aspects of the invention disclosed.

[0152] refer to Figure 6 This invention provides a multi-algorithm collaborative prediction method for electricity demand integrating deep learning, comprising the following steps:

[0153] S1. Based on the time series data of electricity load, the influencing factors are screened by gradient boosting tree and lasso regression methods, and collinearity is eliminated. The influencing factors include: meteorological data, holiday information, industrial activity index and regional economic indicators.

[0154] S2. Based on the improved Whale Optimization Algorithm (IWOA), Simulated Annealing Algorithm (SA), and Multi-Scale Graph Convolutional Network (MSGNet), construct the short-term electricity demand forecasting model IWOA-SA-MSGNet, input influencing factors, and output short-term electricity demand forecasting results.

[0155] The multi-scale graph convolutional network MSGNet extracts daily, weekly, and monthly periodic features of the load through frequency domain decomposition;

[0156] An improved Whale Optimization Algorithm (IWOA) is used to optimize the sparsity of the graph adjacency matrix in the multi-scale graph convolutional network MSGNet.

[0157] Simulated Annealing (SA) algorithm is used to optimize the number of nodes (number of hidden layer neurons), learning rate (range [0.001, 0.1]), and number of iterations (range [50, 200]) of the multi-scale graph convolutional network MSGNet.

[0158] S3. Based on the multiplicative neural network (CNN), bidirectional long short-term memory network (BiLSTM), and attention mechanism, construct the medium- and long-term electricity demand forecasting model ISSA-CNN-BiLSTM-Attention, input influencing factors, and output medium- and long-term electricity demand forecasting results.

[0159] Convolutional Neural Networks (CNNs) are used to extract local temporal features; Bidirectional Long Short-Term Memory (BiLSTM) networks are used to capture long-range dependencies; and Attention mechanisms are used to enhance key features.

[0160] The long-term electricity demand forecasting model ISSA-CNN-BiLSTM-Attention model is an improved sparrow search algorithm based on the dynamic step size factor optimization of the Tent chaotic mapping. Specifically, it includes optimizing the number of convolutional kernels in the convolutional neural network CNN (range [8,64]), the number of hidden layer units in the bidirectional long short-term memory network BiLSTM (range [32,256]), and the number of attention heads (range [2,8]).

[0161] In a preferred embodiment, the formula for calculating the gradient boosting tree in step S1 is:

[0162] ,

[0163] in, This represents the initial prediction function of the model at the start of the iteration; It is the learning rate (step size), used to control the contribution weight of the weak learner (range [0.1, 0.3]). It is the m-th weak learner, where M represents the total number of weak learners. It is the final predicted value of the gradient boosting tree;

[0164] Lasso regression uses L1 regularization to screen influencing factors (feature variables) and eliminate multicollinearity. The formula for calculating lasso regression is:

[0165] ,

[0166] in, It is the predicted value of the target variable Y. It is the first One influencing factor (characteristic variable) It is the first Estimated coefficient values ​​of each influencing factor. It is the constant term in the regression model, representing the "basic influence" of the feature variable on the target variable Y; The numerical value reflects the direction and strength of the influence of the feature variable on the target variable Y. In Lasso regression, if If the value approaches 0, it indicates that the feature variable is suppressed by the regularization penalty term and is considered a non-important feature.

[0167] In electricity demand forecasting, if all influencing factors (such as temperature and time) are zero, This is based on the basic power load demand; L1 regularization is used to compress some coefficients to zero, thereby achieving feature selection.

[0168] Step S1: Detailed explanation of the influencing factor screening method: In order to improve the accuracy of electricity demand forecasting, key influencing factors are first screened by combining Gradient Boosting Tree (GBT) and Lasso Regression, and the problem of feature collinearity is effectively eliminated.

[0169] (1) Feature importance assessment:

[0170] The initial power load forecasting model is trained using the GBT model, with all candidate features (such as temperature, humidity, wind speed, weekday / holiday identification, number of hours, industrial activity index, etc.) input into the model.

[0171] After the GBT model is trained, it outputs the "importance score" of each feature, which reflects the average gain of a certain feature in all decision tree node splits.

[0172] Set the threshold τ1 = 0.01 (adjustable) and retain only features with importance scores higher than τ1 for the next step of analysis.

[0173] (2) Collinearity elimination and redundant feature removal:

[0174] For the features preserved above, Lasso regression is used for sparse modeling, and L1 regularization is introduced to compress some coefficients to 0.

[0175] If the coefficient of a certain feature approaches zero, it is considered redundant or highly collinear with other features and is removed; multiple cross-validations are used to determine the Lasso regularization parameter λ to prevent overfitting.

[0176] (3) Final screening results:

[0177] Taking the historical load data of Jiangsu Province's actual power grid in 2023 as an example, the following 10 key influencing factors were finally selected as inputs for subsequent modeling:

[0178] 1. Average temperature (°C)

[0179] 2. Relative humidity (%)

[0180] 3. Wind speed (m / s)

[0181] 4. Holiday signage (0 / 1)

[0182] 5. Weekday indicator (0 / 1)

[0183] 6. Hours (0–23)

[0184] 7. Calendar days (1–365)

[0185] 8. Industrial Activity Index (Unit Index Value)

[0186] 9. GDP quarter-on-quarter growth rate (%)

[0187] 10. 24-hour moving average load (MW)

[0188] This screening method balances model interpretability and generalization ability, and provides a unified set of input features for short-term and medium-to-long-term load forecasting, ensuring that the input variables are both representative and numerically stable in subsequent deep learning modeling.

[0189] In a preferred embodiment of the present invention, step S2 involves initializing and iteratively updating the whale population location using the improved whale optimization algorithm IWOA, specifically including:

[0190] (1) The formula for initializing the position of the Tent chaotic map is:

[0191] , ,

[0192] Let be an initial uniformly distributed random variable;

[0193] in, Let be the position of the w-th individual whale, LB and UB be the lower and upper bound vectors of the search space, respectively, and rand be a random vector in the range [0,1].

[0194] (2) Location update

[0195] The current optimal prey position is defined. Then, the position update formula is:

[0196] , , , ,

[0197] in, and Let represent two random numbers uniformly distributed in the interval [0,1], where Controlling the convergence factor Adjust the search intensity; Control scaling factor Adjusting the direction and magnitude of the gravitational force between the individual and the current optimal solution;

[0198] z represents the number of iterations. It is the position vector of the w-th individual whale at iteration z. It is the position vector of the current optimal solution. It is a distance vector. It is a control parameter vector;

[0199] (3) Improve the design of convergence factors

[0200] In traditional WOA, the convergence factor decreases linearly. This invention replaces it with a nonlinear cosine function to enhance the local search capability in the later stages.

[0201]

[0202] Where: z is the current iteration number; T is the maximum iteration number.

[0203] The IWOA algorithm integrates Tent chaotic initialization and nonlinear convergence factor scheduling mechanism, which effectively improves the convergence speed, global search capability and generalization stability of the algorithm in the short-term electricity demand forecasting problem.

[0204] In a preferred embodiment of the present invention, the formula for calculating the optimized weights of the multi-scale graph convolutional network MSGNet is as follows:

[0205] ,

[0206] in, This represents the output feature vector of node i. Let i represent the set of neighboring nodes. This represents the input feature vector of neighbor node j. This represents a trainable weight matrix. Represents the bias vector;

[0207] The annealing algorithm SA specifically includes:

[0208] Initialization parameters: Initial temperature Lower limit of temperature Initial solution state Each temperature The number of iterations L, the temperature decay coefficient ;

[0209] The current solution state is denoted as ;

[0210] Execute the following loop until the termination condition is met:

[0211] Step a1, for l=1, 2, ..., L, execute:

[0212] Step a2, generating a new solution : ,

[0213] in This represents the amount of perturbation or change in the solution state during the current iteration, i.e., from the current solution... Generate new solutions The introduced fine-tuning perturbations are usually generated by perturbation functions or random perturbation strategies to escape local optima in the solution space;

[0214] Step a3, calculate the increment of the objective function. f represents the objective function used to evaluate the quality of the current solution state S; it is typically related to prediction error, loss value, or fitting accuracy. The simulated annealing algorithm uses this function f as the basis for optimization, guiding the search direction in the solution space.

[0215] Step a4, if Then accept the new solution; if Then based on probability Accept new solutions; exp represents the natural exponential function;

[0216] Step a5, update the temperature at time k+1. .

[0217] As a preferred embodiment of the present invention, the improved sparrow search algorithm ISSA includes the following steps:

[0218] Step b1, Initialize the population: For the s-th sparrow, s = 1, 2, ..., N, where N represents the total number of sparrows, initialize the position vector of the s-th sparrow. :

[0219] ;

[0220] Step b2, Searching for food (fitness assessment): Calculate the current location of each sparrow. fitness value;

[0221] Step b3, Discoverer location update (global exploration):

[0222] ,

[0223] in This indicates that the s-th sparrow is in The position at time Q represents a random variable (mean 0, standard deviation 1) that conforms to the standard normal distribution Q∼N(0,1). It is used to introduce the random perturbation component of the discoverer individual during the position update process, thereby increasing the diversity and jumpiness of the global search.

[0224] Step b4, follower position update (partial development):

[0225] ,

[0226] in Indicates the current number The position vector of the sparrow with the best fitness in the next iteration, i.e., the position of the current global optimal solution; d represents the dimension of the current sparrow (i.e., the dimension of the solution vector), used to scale the update magnitude in local development to avoid excessive high-dimensional perturbations; A + Let represent a random number that satisfies the normal distribution constraint and satisfies . >0 is used to control the direction and intensity of local disturbances;

[0227] Step b5, Guardian location update:

[0228] ,

[0229] in This represents the preset warning jump step size coefficient, used to enable sparrows to escape quickly when facing danger, increasing the search radius and improving the algorithm's ability to escape local optima. Its value is usually a constant or randomly selected within a certain range [0.5, 1.5].

[0230] Step b6: Repeat steps b2 to b5 until the stopping conditions are met. The stopping conditions include: reaching the maximum number of iterations Tmax, and no significant improvement in the globally optimal fitness value for several consecutive generations (convergence threshold). ( ), or the average positional change of the searched individuals is lower than a set threshold. The convergence threshold is... This represents the termination criterion of the optimization algorithm; the algorithm is considered convergent when the change in population fitness or objective function is less than ε. This invention will... Settings Within the interval (i.e., from) to The specific value is obtained by adjusting the algorithm type and data fluctuation sensitivity.

[0231] like Figure 1-5 As shown, the present invention also provides a multi-algorithm collaborative prediction system for electricity demand integrating deep learning 1, characterized in that it includes:

[0232] 2. Short-term electricity demand forecasting platform; 3. Medium- and long-term electricity demand forecasting platform; 4. Data collection and management platform; and 5. Display and analysis platform.

[0233] The short-term power demand forecasting platform 2 forecasts short-term power demand;

[0234] The medium- and long-term electricity demand forecasting platform 3 forecasts medium- and long-term electricity demand.

[0235] The collection and management platform 4 achieves automated training of targeted datasets through the following methods:

[0236] Multi-source heterogeneous data integration: Integrate power load time series data, meteorological data (temperature / humidity / wind speed), holiday information, industrial activity index and regional economic indicators to construct a multi-dimensional dataset;

[0237] Frequency domain features (such as daily / weekly / monthly periodic components) of electricity load time series data are extracted using Fast Fourier Transform (FFT). Combining sliding window statistical features (such as the past 3-hour mean and standard deviation) and external factor interaction features (such as the "temperature × holiday" combination feature), the correlation strength between the features and the target variable (electricity load) is defined using FFT and mutual information methods. , It refers to the first Variables and the Variables Normalized Mutual Information (NMI) is used to measure and The relative correlation strength, with a value range of [0, 1], indicates a stronger correlation as the value is larger. For variables and Mutual information is used to measure... and The degree of absolute interdependence, when the two are completely independent. ;

[0238] because The numerical value of mutual information is affected by the information entropy of the variable itself, making it impossible to directly compare the correlation between different features; therefore, this invention uses normalized mutual information. As the final feature selection metric.

[0239] in, It is a variable Entropy is used to measure The greater the value, the greater the uncertainty. The higher the randomness; It is a variable entropy, It is the geometric mean of entropy, used to normalize mutual information and eliminate the influence of the uncertainty of the variable itself;

[0240] Automated training process: Dynamically optimize model hyperparameters (such as the number of convolutional kernels and the number of attention heads) based on the improved Whale Algorithm (IWOA); balance global exploration and local exploitation through simulated annealing (SA) to avoid getting trapped in local optima; automatically divide the training set / validation set / test set (with rolling time window splitting, supporting incremental training);

[0241] Parameter tuning interface: Provides a visual interface that allows experts to adjust feature weights (such as removing features affected by extreme weather events) and inject domain knowledge;

[0242] The display and analysis platform 5 provides an intuitive display of prediction results, helping users understand the prediction results and make decisions.

[0243] The short-term power demand forecasting platform 2 includes a first influencing factor screening module 21, a whale optimization algorithm module 22, and a simulated annealing algorithm module 23;

[0244] The first influencing factor screening module 21 screens influencing factors using gradient boosting tree and lasso regression methods, and eliminates collinearity;

[0245] The whale optimization algorithm module 22 improves generalization ability by introducing Tent chaotic mapping and nonlinear convergence factor, and optimizes weights by combining multi-scale graph convolutional network.

[0246] The simulated annealing algorithm module 23 is used to optimize the number of nodes (i.e. the number of hidden layer neurons in the neural network), learning rate and number of iterations of the IWOA-SA-MSGNet model, and to build a high-precision prediction model by co-optimizing hyperparameters;

[0247] The IWOA-SA-MSGNet model is a three-level collaborative optimization architecture that integrates the Improved Whale Optimization Algorithm (IWOA), Simulated Annealing (SA), and Multi-Scale Graph Convolutional Network (MSGNet), including:

[0248] Global parameter optimization layer: Optimizes model topology and hyperparameter combination by improving the Innovative Whale Optimization (IWOA) algorithm;

[0249] Local hyperparameter tuning layer: Optimizes dynamic parameters (such as learning rate and momentum factor) during training through simulated annealing (SA) strategy.

[0250] Multi-scale feature extraction layer: Captures the frequency domain features and spatiotemporal correlation of power load through a multi-scale graph convolutional network (MSGNet).

[0251] The medium- and long-term electricity demand forecasting platform 3 includes a second influencing factor screening module 31, a sparrow search algorithm module 32, and an ISSA-CNN-BiLSTM-Attention model module 33;

[0252] The ISSA-CNN-BiLSTM-Attention model 33 is a composite model based on an improved sparrow search algorithm (ISSA) optimized convolutional neural network (CNN), bidirectional long short-term memory network (BiLSTM), and attention mechanism (Attention); the second influencing factor screening module 31 screens influencing factors through gradient boosting tree and lasso regression methods and eliminates collinearity;

[0253] The sparrow search algorithm module 32 is based on the Tent chaotic mapping and dynamic step size factor to optimize the standard sparrow search algorithm.

[0254] The collection and management platform 4 achieves autonomous optimization of power demand forecasting through heterogeneous algorithm fusion and dynamic collaboration mechanism. The collection and management platform 4 includes a parameter dynamic coupling module 41 based on streaming clustering, a multimodal data governance module 42, an interpretable factor interaction engine 43, a meta-enhanced prediction center 44, and a causal inference optimization module 45.

[0255] The parameter dynamic coupling module 41 based on streaming clustering employs an improved K-means algorithm to perform incremental clustering analysis on real-time power load data. It introduces Dynamic Time Warping (DTW) distance as a similarity measure of load curves and constructs an adaptive weighting function based on historical holiday influence factors. This achieves collaborative optimization of peak period segmentation and user-defined holiday rules. Each time new power load time series data arrives, the K-means algorithm is used to perform clustering analysis. First, initialization... Cluster centers, Indicates the first Cluster centers, Then calculate each data point Euclidean distance to all cluster centers :

[0256] ,

[0257] in It is the first Data points To the Cluster centers Euclidean distance, It is the first Data points In the Dimension value, It is the first The cluster centers at the in Dimension value; data points Assigned to the nearest cluster center :

[0258] ,

[0259] in Indicates the first Data points The assigned cluster category (or cluster center) number;

[0260] This indicates that the objective function Cluster center number when the minimum value is obtained ;

[0261] Update cluster center : ,

[0262] in, It is the first The set of all data points in a cluster It is the first The number of data points in each cluster;

[0263] To account for the time-series characteristics of the power load curves, the Dynamic Time Warping (DTW) distance is used to measure the distance between the two power load curves. and Similarities between them:

[0264] ,

[0265] ,

[0266] in Represents a sequence The power load value of the Mth sampling point; Represents a sequence The power load value of the Nth sampling point; Indicates from sequence The starting point To the finish line and from the sequence The starting point To the finish line The cumulative distance of the best matching path (i.e., DTW distance) is used to measure the overall similarity between two power load curves;

[0267] , Euclidean distance is usually used. Representing a subsequence and DTW distance between them; where subsequences It is a sequence The subsequence formed by the first point to the i-th point, It is a sequence The subsequence formed by the first point to the j-th point;

[0268] Equivalent to power load time series subsequence of and subsequence of The DTW distance between them is expressed as:

[0269] ,

[0270] To achieve coordinated optimization of peak-hour segmentation and user-defined holiday rules, the parameter dynamic coupling module 41 based on streaming clustering introduces an adaptive weighting function based on historical holiday influencing factors. The weight function is dynamically adjusted according to holiday rules:

[0271] ,

[0272] in This represents the weight at time point t. A rule function indicating whether time point t is a holiday (1 if yes, 0 otherwise). and Here is the adjustment coefficient, where The adjustment coefficient representing the historical high load factor. This represents the adjustment coefficient for holiday rules. and The value range is set in [0,2], and is adjusted through cross-validation according to the prediction accuracy requirements. The parameter dynamic coupling module 41 based on streaming clustering can effectively optimize and analyze real-time power load data through the above incremental clustering analysis, DTW similarity measurement and dynamic weight adjustment, further improve the load prediction accuracy, and ensure the prediction capability during holidays and peak periods.

[0273] The multimodal data governance module 42 establishes a dual-channel anomaly detection architecture based on an attention mechanism. It fuses the path length features of the isolated forest algorithm with the improved sliding window 3σ criterion (adaptive window length adjustment) at the feature level. A hybrid interpolation network controlled by a gated recurrent unit (GRU) dynamically selects either Long Short-Term Memory (LSTM) interpolation or robust median imputation strategies. The specific implementation includes: establishing a dual-channel anomaly detection architecture based on an attention mechanism. In this architecture, power load time series data is first processed in parallel through two channels. One channel processes standard input data, and the other channel processes features enhanced by multimodal information. In each channel, a weighted approach based on an attention mechanism is used to enhance the representation of important features, as shown in the formula:

[0274] ,

[0275] in, It is a weighted feature representation. It is the attention coefficient. Indicates the first The importance of each feature The number of features; the attention coefficient It is learned through the training process. By calculating the weight of each feature, the model can focus on those more important modalities, thereby improving the accuracy of anomaly detection;

[0276] The Isolation Forest algorithm is used for anomaly detection. Its key idea is to "isolate" data points by constructing a tree structure. The path length of each tree... Data points The degree of "isolation" a data point exhibits depends on the path length; the shorter the path, the more likely it is to be an outlier. For a given data point... data points Path length The calculation formula is:

[0277] ,

[0278] in, For the number of trees, It is in the Data points in a tree Path length when isolated;

[0279] Using path length Calculate abnormal scores : ,

[0280] in, It is related to the set of quantities Size The relevant normalization constants;

[0281] The improved sliding window 3σ criterion is used for real-time detection of outliers in a data stream, with each window... Include 1 data point, 1 window Data points within The process for evaluating whether a data point is an outlier is as follows: Calculate the mean of the data within the window. and standard deviation :

[0282] , ,

[0283] According to the improved sliding window 3σ criterion, if the data points The value satisfies: Then determine the data points This is abnormal data; the window length is adjusted to adapt to different scenarios. It can adapt and adjust itself.

[0284] Hybrid Imputation Network (GRU-Controlled LSTM and Robust Median Imputation Strategy): Gated Recurrent Units (GRUs) are used to control the selection of the imputation strategy. The GRU network can learn the temporal patterns of missing data and decide when to use a Long Short-Term Memory (LSTM) network for imputation or a robust median imputation strategy. LSTM Imputation Strategy: LSTM predicts missing values ​​by learning the long-term dependencies of time-series data. The formula for LSTM is:

[0285] ,

[0286] in, For the data at the current time step t, Let be the hidden state vector of the previous time step t-1. Let be the hidden state vector of the previous time step t;

[0287] Robust median imputation strategy: This strategy imputes missing values ​​based on the median of neighboring data points. If a data point is missing within a window, the median of the non-missing data points within that window is used. As an alternative:

[0288] ,

[0289] in, The interpolated value at time point t; This represents the (t+k)th data point;

[0290] GRU control logic: The gated recurrent unit (GRU) dynamically selects the interpolation strategy based on the current time step input (including historical data and missing information), using the following formula:

[0291] ,

[0292] in This represents the output of the gated recurrent unit (GRU) at time point t. The input vector at time point t contains the observation data and related auxiliary features at the current moment, which is used to guide GRU to perform dynamic interpolation or prediction.

[0293] Then through the output The decision is made on whether to use LSTM interpolation or robust median imputation strategy; the multimodal data governance module 42 realizes anomaly detection and data repair through the above process.

[0294] The interpretable factor interaction engine 43 constructs a factor selection model based on a knowledge graph in the power industry. It utilizes a SHAP interpreter optimized through transfer learning to generate a feature importance graph and designs a three-dimensional tensor decomposition structure (Tucker decomposition mode) to capture higher-order interaction effects of meteorological, economic, and social factors. Simultaneously, it embeds a lightweight online learning architecture to achieve real-time differential updates of factor weights. The specific process includes: constructing a factor selection model based on a knowledge graph in the power industry, mapping meteorological factor M, economic factor E, and social factor S to graph nodes, and modeling factor relationships using a graph embedding method (Node2Vec). The graph embedding optimization objective is to minimize the distance between a node and its neighboring node embeddings.

[0295] ,

[0296] in, This represents the value of node e that minimizes the objective function; and Let n be the embedding representation of node n and let o be the embedding representation of node o. Represents the set of neighboring nodes of node n;

[0297] SHAP Explainer: Utilizing a SHAP (Shapley Additive exPlanations) interpreter optimized through transfer learning, it generates a contribution map of factors to the prediction results, evaluating the contribution of each factor to the model output. The optimization objective of transfer learning is:

[0298] ,

[0299] in, This represents the total loss function in transfer learning, used to comprehensively measure the joint optimization objective of the target task and the source task during the training process;

[0300] and These are the loss functions for the target task and the source task, respectively. and These are the parameters for the target task and the parameters for the source task, respectively. It is a hyperparameter that balances the source and target tasks;

[0301] 3D Tensor Decomposition (Tucker Decomposition): To capture the higher-order interaction effects of meteorological factor M, economic factor E, and social factor S, the input core tensor is decomposed... Decomposed into a core tensor G and a factor matrix (meteorological), (economy), (Social) product:

[0302] ,

[0303] Where G is the core tensor, , , These are factor matrices representing meteorological, economic, and social factors, with symbols... The tensor represents the tensor in the first place. The product of the modulus and the matrix. The values ​​are 1, 2, and 3; the engine updates the factor weights in real time using a lightweight online learning architecture.

[0304] ,

[0305] in, Indicates the first Factor weight vector for the next iteration It's the learning rate. It is a loss function Regarding the factor weight vector The gradient;

[0306] The meta-enhanced prediction hub 44 creates a multi-scale prediction model group (LSTM-CNN-Transformer heterogeneous ensemble) that includes a spatiotemporal attention mechanism. It employs a dual PPO (Proximal Policy Optimization) algorithm framework to simultaneously optimize model weights and feature weights, and designs a reward function R based on the prediction error gradient.

[0307] ,

[0308] Where MAE represents the mean absolute error. Indicates dynamic interval coverage. and The weight coefficients are used; the meta-enhanced prediction center (44) implements distributed training through an asynchronous parameter server architecture;

[0309] The causal inference optimization module 45 constructs a dynamic scheme evaluation system based on Bayesian Structural Temporal Model (BSTM) to quantitatively evaluate the causal effects after power strategy adjustments, specifically including:

[0310] Constructing a Bayesian structured time network: Introducing a structured Bayesian network in the form of a directed acyclic graph (DAG), with random variable nodes. Represents the first time at time t. Due to individual states, network evolution follows the Markov assumption, and the joint probability model is as follows:

[0311] ,

[0312] in Represents a node The set of parent nodes is obtained by a structure learning method (a heuristic algorithm based on BIC scoring);

[0313] This represents the overall joint probability distribution of all variable nodes (factor state nodes) from time step 1 to T; This represents the nth variable node at time t. Given its parent node set The conditional probability distribution under the given conditions; T represents the total number of time steps in the evolution of the dynamic Bayesian network, i.e., the total time window spanned by the model evaluation.

[0314] Design a cost-sensitive loss function Introducing class weight vectors Adjust the ordinary cross-entropy loss to a weighted form:

[0315] Where K represents the category index, ;

[0316] Indicates the first The class weight coefficient represents the importance of that class in the total loss (classes with fewer samples are usually given greater weight). This represents the one-hot encoded value of the true label, if the sample belongs to the first... The value is 1 if the condition is met, and 0 otherwise. The model represents the first The predicted probability of a class;

[0317] The purpose of this method is to use weighted cross-entropy loss to increase attention to the critical minority classes and make the model more robust in class imbalance scenarios.

[0318] Cost-sensitive loss function This is used to emphasize the impact of high-cost error prediction categories on the overall evaluation metrics, and significantly improve the responsiveness to key minority scenarios;

[0319] Counterfactual reasoning mechanism: Introducing a dual-channel reasoning structure, based on actual observation data... Based on this, through the analysis of decision variables Intervention to construct counterfactual samples The output is obtained through the inference network. The difference between factual and counterfactual outcomes is compared to assess the causal impact of the proposed solutions; among which... The term "intervention" in a causal diagram originates from Pearl's theory of causal reasoning, which involves artificially setting variable values ​​to assess causal effects. This represents the set values ​​of decision variables (used to construct counterfactual scenarios);

[0320] The above is used to construct a scenario different from the current observations to assess the model's sensitivity to changes in the intervention variable and its causal response.

[0321] VAE-generated adversarial sample robustness verification mechanism: Constructing a variational autoencoder structure:

[0322] ,

[0323] Combined with counter-disturbance Generate adversarial samples This is used to test the performance changes of the model under extreme inputs and improve the robustness of the solution.

[0324] in Represents the variational loss function of a VAE; This represents the posterior distribution of the encoder output. The following expectations; This represents the reconstruction probability distribution (or likelihood distribution) generated by the decoder. express Divergence, used to measure the encoder posterior With prior distribution The differences between them; This represents the prior distribution of the latent variable;

[0325] The display and analysis platform 5 includes an intuitive display module 51 and an error analysis module 52; the intuitive display module 51 provides a display of prediction results; the error analysis module 52 is used to help users better understand the prediction results and make decisions accordingly.

[0326] In this embodiment, the first influencing factor screening module 21 screens influencing factors using gradient boosting trees and lasso regression, and eliminates collinearity, providing a solid foundation for subsequent predictions. The whale optimization algorithm module 22 improves its generalization ability by introducing a chaotic initialization mechanism based on Tent mapping and a nonlinear convergence factor, and optimizes the weights by combining a multi-scale graph convolutional network. The simulated annealing algorithm module 23 optimizes the number of nodes, learning rate, and number of iterations of the model, constructing the IWOA-SA-MSGNet model. The second influencing factor screening module 31 screens influencing factors using gradient boosting trees and lasso regression, and eliminates collinearity, providing a solid foundation for subsequent predictions. The sparrow search algorithm module 32 optimizes the standard sparrow search algorithm based on Tent chaotic mapping and dynamic step size factor. The improved sparrow search algorithm-convolutional neural network-bidirectional long short-term memory network-attention mechanism integrated model module (ISSA-CNN-BiLSTM-Attention) module 33 optimizes the key parameters of the integrated attention mechanism of the convolutional neural network and bidirectional long short-term memory network using ISSA, constructing the ISSA-CNN-BiLSTM-Attention model. The parameter dynamic coupling module 41 based on streaming clustering provides basic parameter settings for load periods, public holidays, and personalized holiday management according to actual conditions, improving prediction accuracy; the prediction data management module 42 assists customers in easily managing the data to be predicted, providing functions such as data anomaly analysis and processing, automatic interpolation, outlier extraction, and trend analysis; the factor management module 43 assists in managing relevant factor data for prediction, providing functions such as trend analysis, anomaly analysis, and special factor definition; the data prediction module 44 provides functions such as the rationality of future data predictions, management of prediction schemes and plans, setting prediction objects and required factor data, and management of the weights of prediction factors; and provides batch prediction and automated prediction; the result analysis module 45 performs comparative analysis of prediction result data; provides training set analysis of the data to be predicted, prediction result set analysis, comparative analysis of multiple prediction schemes, multi-dimensional and comprehensive functions; and assists customers in analyzing and adjusting prediction results, automatically selecting the optimal prediction scheme.

[0327] In this embodiment, to verify the timeliness and stability of the model under actual power grid conditions, rolling forecasts and trend forecasts of power load are performed at short-term and medium-to-long-term scales, respectively:

[0328] Short-term forecasting: The IWOA-SA-MSGNet model was used to perform daily rolling 24-hour forecasts of the daily load curves of a provincial power grid during the summer peak electricity consumption period (July 10-15, 2024). Specifically, historical characteristic data was input daily, and the load forecast value for the next 24 hours (the following day) was output, continuously rolling for 6 days to obtain 6 days of comparative forecast data (see Tables 1 and 2).

[0329] For the medium- and long-term forecasting section: The ISSA-CNN-BiLSTM-Attention model is used to predict the electricity consumption trend for the next 30 days based on the input feature data on the first day of each month, accumulating six consecutive months of forecast data from January to June 2024 (see Table 3). The model predicts the monthly trend average and peak-valley load indicators, which are used for medium- and long-term energy planning analysis.

[0330] Table 1

[0331]

[0332] Table 2

[0333]

[0334] As shown in Table 2, the special scenario test (extreme high temperature on July 14th) reveals, under extreme high temperature conditions, the IWOA-SA-MSGNet model exhibits a significant improvement in prediction accuracy compared to the baseline model MSGNet. The average error across the three time periods decreased from -8.5% for MSGNet to -2.4% for IWOA-SA-MSGNet, a reduction of 71% (from -7.2% to -2.2%, -10.2% to -2.8%, and -8.1% to -2.3%). The improved model's predictions are closer to the actual load, particularly during the extreme load fluctuation period at 15:00 (error improved from -10.2% to -2.8%).

[0335] IWOA-SA-MSGNet enhances the model's global search capability and parameter optimization efficiency by introducing an improved Whale Optimization (IWOA) algorithm and simulated annealing (SA). The nonlinear characteristics of loads under high-temperature scenarios (such as sudden increases in air conditioning load) may be underestimated by the baseline model, while the collaborative optimization mechanism of IWOA-SA improves the model's ability to capture complex dynamics and reduces systematic bias. Both models exhibit negative errors (predicted values ​​are lower than actual values), indicating that predicting load surges caused by extreme high temperatures is challenging. The improved model has higher error stability (standard deviation reduced from 3.0% to 0.3%), indicating stronger adaptability across different time periods. This characteristic is significant for real-time grid dispatch, reducing the risk of wasted reserve capacity or insufficient power supply due to prediction errors. Accurate load forecasting is crucial for ensuring the safe and economical operation of the power grid under extreme weather conditions. IWOA-SA-MSGNet's error is controlled within 3%, significantly outperforming traditional models and providing more reliable decision support for power systems. Future research can further validate its robustness over longer periods and across multiple scenarios, and explore model interpretability to guide engineering practices. IWOA-SA-MSGNet effectively improves the load forecasting accuracy in extreme high-temperature scenarios through algorithm fusion, providing a new approach for power system optimization in complex environments.

[0336] Table 3

[0337]

[0338] As shown in Table 3, within the 6-month prediction period, the mean absolute percentage error (MAPE) decreased from 16.5% of the baseline model to 10.8%, representing a relative improvement of 34.9%. The root mean square error (RMSE) decreased from 319 to 206, a reduction of 35.4%, indicating that the prediction model performed very well at these time points.

[0339] While the present invention has been described above with reference to preferred embodiments, it is not intended to limit the invention. Those skilled in the art can make various modifications and refinements without departing from the spirit and scope of the invention. Therefore, the scope of protection of the present invention shall be determined by the claims.

Claims

1. A multi-algorithm collaborative forecasting method for electricity demand integrating deep learning, characterized in that, Includes the following steps: S1. Based on the time series data of electricity load, the influencing factors are screened by gradient boosting tree and lasso regression methods, and collinearity is eliminated. The influencing factors include: meteorological data, holiday information, industrial activity index and regional economic indicators. S2. Based on the improved Whale Optimization Algorithm (IWOA), Simulated Annealing Algorithm (SA), and Multi-Scale Graph Convolutional Network (MSGNet), construct the short-term electricity demand forecasting model IWOA-SA-MSGNet, input influencing factors, and output short-term electricity demand forecasting results. The multi-scale graph convolutional network MSGNet extracts daily, weekly, and monthly periodic features of the load through frequency domain decomposition; An improved Whale Optimization Algorithm (IWOA) is used to optimize the sparsity of the graph adjacency matrix in the multi-scale graph convolutional network MSGNet. Simulated annealing (SA) algorithm is used to optimize the number of nodes, learning rate, and number of iterations in the multi-scale graph convolutional network MSGNet. S3. Based on the multiplicative neural network (CNN), bidirectional long short-term memory network (BiLSTM), and attention mechanism, construct the medium- and long-term electricity demand forecasting model ISSA-CNN-BiLSTM-Attention, input influencing factors, and output medium- and long-term electricity demand forecasting results. Convolutional Neural Networks (CNNs) are used to extract local temporal features; Bidirectional Long Short-Term Memory (BiLSTM) networks are used to capture long-range dependencies; Attention mechanisms are used to enhance key features; The long-term electricity demand forecasting model ISSA-CNN-BiLSTM-Attention model is an improved sparrow search algorithm based on the dynamic step size factor optimization of the Tent chaotic map. Specifically, it includes optimizing the number of convolutional kernels in the convolutional neural network (CNN), the number of hidden layer units in the bidirectional long short-term memory network (BiLSTM), and the number of attention heads.

2. The method for collaborative prediction of electricity demand using multiple algorithms incorporating deep learning, as described in claim 1, is characterized in that... In step S1, the formula for calculating the gradient boosting tree is: ， in, This represents the initial prediction function of the model at the start of the iteration; It's the learning rate. It is the m-th weak learner, where M represents the total number of weak learners. It is the final predicted value of the gradient boosting tree; The formula for calculating using the lasso regression method is: ， in, It is the predicted value of the target variable Y. It is the first One influencing factor, It is the first Estimated coefficient values ​​of each influencing factor. It is the constant term in the regression model.

3. The method for collaborative prediction of electricity demand using multiple algorithms incorporating deep learning, as described in claim 1, is characterized in that... In step S2, the whale population location is initialized using the improved whale optimization algorithm IWOA, with the following formula: ， in, Let be the position of the w-th individual whale, LB and UB be the lower and upper bound vectors of the search space, respectively, and rand be a random vector in the range [0,1]. The current optimal prey position is defined. Then, the position update formula is: ， ， ， ， in, and Let represent two random numbers uniformly distributed in the interval [0,1], where Controlling the convergence factor Adjust the search intensity; Control scaling factor Adjusting the direction and magnitude of the gravitational force between the individual and the current optimal solution; z represents the number of iterations. It is the position vector of the w-th individual whale at iteration z. It is the position vector of the current optimal solution. It is a distance vector. It is a control parameter vector; The formula for calculating the optimized weights of the multi-scale graph convolutional network MSGNet is as follows: ， in, This represents the output feature vector of node i. Let i represent the set of neighboring nodes. This represents the input feature vector of neighbor node j. This represents a trainable weight matrix. Represents the bias vector; The annealing algorithm SA specifically includes: Initialization parameters: Initial temperature Lower limit of temperature Initial solution state Each temperature The number of iterations L, the temperature decay coefficient ; The current solution state is denoted as ; Execute the following loop until the termination condition is met: Step a1, for l=1, 2, ..., L, execute: Step a2, generating a new solution : ， in This represents the amount of perturbation or change in the solution state during the current iteration, i.e., from the current solution... Generate new solutions The introduced fine-tuning perturbation; Step a3, calculate the increment of the objective function. f represents the objective function used to evaluate the quality of the current solution state S; Step a4, if Then accept the new solution; if Then based on probability Accept new solutions; exp represents the natural exponential function; Step a5, update the temperature at time k+1. .

4. The method for collaborative prediction of electricity demand using multiple algorithms incorporating deep learning, as described in claim 1, is characterized in that... The improved Sparrow Search Algorithm ISSA includes the following steps: Step b1, Initialize the population: For the s-th sparrow, s = 1, 2, ..., N, where N represents the total number of sparrows, initialize the position vector of the s-th sparrow. : ； Step b2, Search for food: Calculate the current position of each sparrow. fitness value; Step b3, Discoverer location update: ， in This indicates that the s-th sparrow is in The position at time t, Q represents a random variable that follows a standard normal distribution Q∼N(0,1); Step b4, follower position update: ， in Indicates the current number The position vector of the sparrow with the best fitness in the next iteration; d represents the dimension of the current sparrow. Let represent a random number that satisfies the normal distribution constraint and satisfies . >0; Step b5, Guardian location update: ， in This represents the preset disturbance coefficient of the vigilant; Step b6: Repeat steps b2 to b5 until the stopping condition is met.

5. A multi-algorithm collaborative prediction system for electricity demand integrating deep learning (1), characterized in that, include: Short-term electricity demand forecasting platform (2), medium- and long-term electricity demand forecasting platform (3), data collection and management platform (4), and display and analysis platform (5); The short-term power demand forecasting platform (2) forecasts short-term power demand. The medium- and long-term electricity demand forecasting platform (3) forecasts medium- and long-term electricity demand. The collection and management platform (4) achieves targeted automated training of datasets in the following ways: Multi-source heterogeneous data integration: Integrating power load time series data, meteorological data, holiday information, industrial activity index and regional economic indicators to construct a multi-dimensional dataset; Dynamic feature engineering: Extracting frequency domain features from power load time series data using fast Fourier transform; Combining sliding window statistical features and external factor interaction features, the correlation strength between features and the target variable is defined using Fast Fourier Transform and mutual information method: , It refers to the first Variables and the Variables Normalized mutual information; For variables and Mutual information; Automated training process: Dynamically optimize model hyperparameters based on the improved Whale Algorithm (IWOA); balance global exploration and local exploitation through simulated annealing (SA); automatically divide the training, validation, and test sets. Assistive parameter tuning interface: Provides a visual interface that allows experts to adjust feature weights and inject domain knowledge; The display and analysis platform (5) provides an intuitive display of prediction results, helping users understand the prediction results and make decisions.

6. The power demand collaborative prediction system (1) integrating deep learning as described in claim 5, characterized in that, The short-term power demand forecasting platform (2) includes a first influencing factor screening module (21), a whale optimization algorithm module (22), and a simulated annealing algorithm module (23). The first influencing factor screening module (21) screens influencing factors using gradient boosting tree and lasso regression methods and eliminates collinearity; The whale optimization algorithm module (22) improves generalization ability by introducing Tent chaotic mapping and nonlinear convergence factor, and optimizes weights by combining multi-scale graph convolutional network; The simulated annealing algorithm module (23) is used to optimize the number of nodes, learning rate and number of iterations of the IWOA-SA-MSGNet model, and to build a high-precision prediction model by co-optimizing hyperparameters; The IWOA-SA-MSGNet model is a three-level collaborative optimization architecture that integrates the improved whale optimization algorithm (IWOA), the simulated annealing strategy (SA), and the multi-scale graph convolutional network (MSGNet), including: Global parameter optimization layer: Improves the model topology and hyperparameter combination by refining the IWOA (Integrated Whale Optimization) algorithm; Local hyperparameter tuning layer: Optimizes dynamic parameters during training using the simulated annealing (SA) strategy; Multi-scale feature extraction layer: The frequency domain features and spatiotemporal correlation of power load are captured by the multi-scale graph convolutional network MSGNet.

7. The power demand collaborative prediction system (1) integrating deep learning as described in claim 6, characterized in that, The medium- and long-term power demand forecasting platform (3) includes a second influencing factor screening module (31), a sparrow search algorithm module (32), and an ISSA-CNN-BiLSTM-Attention model (33). The ISSA-CNN-BiLSTM-Attention model (33) is a composite model based on the improved sparrow search algorithm ISSA optimized convolutional neural network CNN, bidirectional long short-term memory network BiLSTM and attention mechanism Attention; The second influencing factor screening module (31) screens influencing factors using gradient boosting tree and lasso regression methods and eliminates collinearity; The sparrow search algorithm module (32) is based on the Tent chaotic mapping and dynamic step size factor optimization standard sparrow search algorithm.

8. The power demand collaborative prediction system (1) integrating deep learning as described in claim 7, characterized in that, The collection management platform (4) achieves autonomous optimization of power demand forecasting through heterogeneous algorithm fusion and dynamic collaboration mechanism. The collection management platform (4) includes a parameter dynamic coupling module (41) based on streaming clustering, a multimodal data governance module (42), an interpretable factor interaction engine (43), a meta-enhanced prediction center (44), and a causal inference optimization module (45). The parameter dynamic coupling module (41) based on streaming clustering uses an improved K-means algorithm to perform incremental clustering analysis on real-time power load data. It introduces dynamic time warping (DTW) distance as a similarity measure of load curves and constructs an adaptive weight function based on historical holiday influence factors to achieve collaborative optimization of peak period division and user-defined holiday rules. When new power load time series data arrives, the K-means algorithm is used to perform clustering analysis on the data. The steps are as follows: SS1, Initialization Cluster centers, Indicates the first Cluster centers, Then calculate each data point Euclidean distance to all cluster centers : ， in, It is the first Data points To the Cluster centers Euclidean distance, It is the first Data points In the Dimension value, It is the first The cluster centers at the in Dimension value; data points Assigned to the nearest cluster center : ， in Indicates the first Data points The cluster number assigned to the cluster; This indicates that the objective function Cluster center number when the minimum value is obtained ; SS2, Update Cluster Centers : ， in, It is the first The set of all data points in a cluster It is the first The number of data points in each cluster; SS3 uses Dynamic Time Warping (DTW) distance to measure two power load curves. and Similarities between them: ， ， in Represents a sequence The power load value of the Mth sampling point; Represents a sequence The power load value of the Nth sampling point; Indicates from sequence The starting point To the finish line and from the sequence The starting point To the finish line The cumulative distance of the best matching path formed; , It is Euclidean distance. Representing a subsequence and DTW distance between them; where subsequences It is a sequence The subsequence formed by the first point to the i-th point, It is a sequence The subsequence formed by the first point to the j-th point; Equivalent to electricity load time series subsequence of and subsequence of The DTW distance between them is expressed as: ； The parameter dynamic coupling module (41) based on streaming clustering introduces an adaptive weight function based on historical holiday influencing factors. The weight function is dynamically adjusted in conjunction with holiday rules to achieve collaborative optimization between peak period segmentation and user-defined holiday rules. ， in This represents the weight at time point t. A rule function indicating whether time point t is a holiday. The adjustment coefficient representing the historical high load factor. This indicates the adjustment coefficient for holiday rules.

9. A multi-algorithm collaborative prediction system for electricity demand integrating deep learning as described in claim 8 (1), characterized in that, The multimodal data governance module (42) establishes a dual-channel anomaly detection architecture based on the attention mechanism, integrates the path length feature of the isolated forest algorithm with the improved sliding window 3σ criterion at the feature level, and dynamically selects the long short-term memory network LSTM interpolation or robust median filling strategy through the hybrid interpolation network controlled by the gated recurrent unit GRU. The specific implementation process includes: establishing a dual-channel anomaly detection architecture based on an attention mechanism. In this architecture, power load time series data is first processed in parallel through two channels. One channel processes standard input data, and the other channel processes features enhanced by multimodal information. In each channel, an attention-based weighting method is used to improve the representation of important features, as shown in the formula: ， in, It is a weighted feature representation. It is the attention coefficient. Indicates the first The importance of each feature The number of features; Path length of each tree Data points The degree of isolation for a data point data points Path length The calculation formula is: ， in, For the number of trees, It is in the Data points in a tree Path length when isolated; Using path length Calculate abnormal scores : ， in, It is related to the set of quantities Size The relevant normalization constants; The improved sliding window 3σ criterion is used for real-time detection of outliers in a data stream, with each window... Include 1 data point, 1 window Data points within The process for evaluating whether a data point is an outlier is as follows: Calculate the mean of the data within the window. and standard deviation : ， ， According to the improved sliding window 3σ criterion, if the data points The value satisfies: Then determine the data points This is abnormal data; The formula for the Long Short-Term Memory (LSTM) network is: ， in, For the data at the current time step t, Let be the hidden state vector of the previous time step t-1. Let be the hidden state vector of the previous time step t; Robust median imputation strategy: If data points are missing within a window, the median of the non-missing data points within the window is used. As an alternative: ， in, The interpolated value at time point t; This represents the (t+k)th data point; GRU control logic: The gated loop unit (GRU) dynamically selects the interpolation strategy based on the current time step input, using the following formula: ， in This represents the output of the gated recurrent unit (GRU) at time point t. The input vector represents time point t.

10. A multi-algorithm collaborative prediction system for electricity demand integrating deep learning as described in claim 8 (1), characterized in that, The interpretable factor interaction engine (43) constructs a factor screening model based on the knowledge graph of the power field, uses the SHAP interpreter optimized by transfer learning to generate a feature importance graph, and designs a three-dimensional tensor decomposition structure to capture the high-order interaction effects of meteorological, economic and social factors. At the same time, it embeds a lightweight online learning architecture to realize the real-time differential update of factor weights. The specific process includes: constructing a factor selection model based on a knowledge graph in the power sector, mapping meteorological factor M, economic factor E, and social factor S to graph nodes, and modeling factor relationships using graph embedding methods. The optimization objective of graph embedding is to minimize the distance between the embedded representations of a node and its neighboring nodes. ， in, This represents the value of node e that minimizes the objective function; and Let n be the embedding representation of node n and let o be the embedding representation of node o. Represents the set of neighboring nodes of node n; Using a SHAP interpreter optimized by transfer learning, a contribution map of factors to the prediction results is generated, and the contribution of each factor to the model output is evaluated. The optimization objective of transfer learning is: ， in, This represents the total loss function in transfer learning; and These are the loss functions for the target task and the source task, respectively. and These are the parameters for the target task and the parameters for the source task, respectively. It is a hyperparameter that balances the source and target tasks; 3D Tensor Decomposition: In order to capture the higher-order interaction effects of meteorological factor M, economic factor E, and social factor S, the input core tensor is decomposed. Decomposed into a core tensor G and a factor matrix , , The product of: ， Where G is the core tensor, , , These are factor matrices representing meteorological, economic, and social factors, with symbols... The tensor represents the tensor in the first place. The product of the modulus and the matrix. The values ​​are 1, 2, and 3; the engine updates the factor weights in real time using a lightweight online learning architecture. ， in, Indicates the first Factor weight vector for the next iteration It's the learning rate. It is a loss function Regarding the factor weight vector The gradient; The meta-enhanced prediction center (44) creates a multi-scale prediction model group containing a spatiotemporal attention mechanism, and simultaneously optimizes the model weights and feature weights using a dual PPO algorithm framework, designing a reward function R based on the prediction error gradient: ， Where MAE represents the mean absolute error. Indicates dynamic interval coverage. and These are the weighting coefficients; The meta-enhanced prediction center (44) implements distributed training through an asynchronous parameter server architecture; The causal inference optimization module (45) constructs a dynamic scheme evaluation system based on Bayesian structured time network (BSTM) to quantitatively evaluate the causal effects after power strategy adjustment, specifically including: Constructing a Bayesian structured time network: Introducing a structured Bayesian network in the form of a directed acyclic graph (DAG), with random variable nodes. Represents the first time at time t. Due to individual states, network evolution follows the Markov assumption, and the joint probability model is as follows: ， in Represents a node The set of parent nodes is obtained by the structure learning method; This represents the overall joint probability distribution of all variable nodes from time step 1 to T; This represents the nth variable node at time t. Given its parent node set Conditional probability distribution under given conditions; T represents the total number of time steps in the evolution of the dynamic Bayesian network; Design a cost-sensitive loss function Introducing class weight vectors Adjust the ordinary cross-entropy loss to a weighted form: ， Where K represents the category index, ; Indicates the first Class weight coefficient; This represents the one-hot encoded value of the true label, if the sample belongs to the first... The value is 1 if the condition is met, otherwise it is 0. The model represents the first The predicted probability of a class; Counterfactual reasoning mechanism: Introducing a dual-channel reasoning structure, based on actual observation data... Based on this, by analyzing the decision variables Intervention to construct counterfactual samples The output is obtained through the inference network. The difference between factual and counterfactual outcomes is compared to assess the causal impact of the proposed solutions; among which... This represents the intervention operation in the cause-effect diagram; Indicates the set decision variables; VAE-generated adversarial sample robustness verification mechanism: Constructing a variational autoencoder structure: ， Combined with counter-disturbance Generate adversarial examples ; in Represents the variational loss function of a VAE; This represents the posterior distribution of the encoder output. The following expectations; This represents the reconstruction probability distribution generated by the decoder; express Divergence, used to measure the encoder posterior With prior distribution The differences between them.

Images