Multi-modal data fusion method and system based on energy scheduling and storage medium

Abstract

The invention relates to the technical field of energy scheduling, in particular to a multi-modal data fusion method and system based on energy scheduling and a storage medium. The method comprises the following steps: obtaining multi-modal data, and carrying out abnormal fluctuation feature extraction to obtain a space-time fusion abnormal feature labeling set; deducing a multi-objective optimization path according to the space-time fusion abnormal feature labeling set to obtain a dynamic scheduling decision map; performing edge node game equilibrium calculation according to the dynamic scheduling decision map to obtain a trusted scheduling verification chain; performing digital twinborn constraint optimization on the trusted scheduling verification chain to obtain a closed-loop scheduling digital twinborn body; compiling a dynamic scheduling instruction set based on the closed-loop scheduling digital twin to obtain an anti-disturbance energy scheduling strategy library; and obtaining real-time energy supply and demand data, and performing scheduling deviation tracing on the real-time energy supply and demand data according to the anti-disturbance energy scheduling strategy library to obtain an energy distribution decision. According to the invention, the efficiency and reliability of energy scheduling can be improved.

Description

Technical Field

[0001] The present invention relates to the field of energy scheduling technology, and in particular to a multimodal data fusion method, system and storage medium based on energy scheduling. Background Art

[0002] Energy dispatch is a critical component of modern energy systems, particularly in large-scale power grids and renewable energy grid-connected systems. Real-time energy dispatch and optimization are crucial for improving energy efficiency and ensuring stable system operation. Existing energy dispatch technologies primarily rely on traditional centralized dispatch models, which are typically managed and controlled by a single central node or dispatch center. While this approach simplifies the dispatch process to some extent, with the increasing complexity of energy structures, especially the increasing proportion of renewable energy (such as wind and solar) in power systems, centralized dispatch faces several prominent challenges. First, traditional centralized dispatch relies on a single central node, resulting in high computational complexity and the risk of single points of failure, making it difficult to cope with the dynamic interactive demands of multiple regions. Second, extreme weather (such as typhoons and high temperatures) or fluctuations in renewable energy sources can lead to significant deviations in supply and demand forecasts, making it difficult for existing energy dispatch models to respond and adjust promptly and accurately. Furthermore, traditional centralized dispatch systems often rely on cloud platforms for data processing and storage, which results in significant data transmission delays. Furthermore, centralized processing on cloud platforms is vulnerable to cyberattacks, compromising dispatch reliability. Summary of the Invention

[0003] Based on this, it is necessary for the present invention to provide a multimodal data fusion method, system and storage medium based on energy scheduling to solve at least one of the above technical problems.

[0004] To achieve the above objectives, a multimodal data fusion method based on energy scheduling includes the following steps:

[0005] Step S1: Acquire multimodal data and construct a dynamic spatiotemporal adjacency matrix based on the multimodal data; extract abnormal fluctuation features based on the dynamic spatiotemporal adjacency matrix to obtain a spatiotemporal fusion abnormal feature annotation set;

[0006] Step S2: Perform distributed policy gradient aggregation on the spatiotemporal fusion anomaly feature annotation set to obtain a regional collaborative scheduling vector; deduce a multi-objective optimization path based on the regional collaborative scheduling vector to obtain a dynamic scheduling decision map;

[0007] Step S3: Perform edge node game equilibrium calculation based on the dynamic scheduling decision graph, and perform cross-regional energy mutual assistance verification to obtain a trusted scheduling verification chain;

[0008] Step S4: Perform digital twin constraint optimization on the trusted scheduling verification chain, construct a multi-physics field coupling scheduling unit, and obtain a closed-loop scheduling digital twin; compile a dynamic scheduling instruction set based on the closed-loop scheduling digital twin to obtain an anti-disturbance energy scheduling strategy library;

[0009] Step S5: Acquire real-time energy supply and demand data, and dynamically evolve the anti-disturbance strategy for the real-time energy supply and demand data based on the anti-disturbance energy scheduling strategy library to generate a global scheduling optimization evidence chain; implement scheduling deviation tracing based on the global scheduling evidence chain, and screen highly robust scheduling instructions to obtain energy allocation decisions.

[0010] By integrating multiple technical approaches, this invention effectively addresses the shortcomings of traditional centralized energy scheduling models, improving the efficiency, reliability, and intelligence of energy scheduling. First, by acquiring and processing multimodal data, a comprehensive understanding of the dynamic changes in different energy and environmental factors can be achieved, providing rich real-time information for scheduling decisions. Furthermore, the construction of a dynamic spatiotemporal adjacency matrix helps capture spatiotemporal fluctuation characteristics, thereby identifying abnormal fluctuations and facilitating the timely detection of potential energy supply and demand issues. Distributed policy gradient aggregation technology enables efficient coordinated scheduling between regions, enhancing the system's responsiveness to local demand fluctuations. This allows for the derivation of multi-objective optimization paths and ensures the adaptability of scheduling solutions in different scenarios. Edge node game equilibrium calculations not only accurately optimize inter-regional energy mutual assistance relationships but also verify the credibility of scheduling, preventing the impact of single points of failure and cyberattacks on system stability, and ensuring the security and robustness of scheduling decisions. Based on the implementation of a trusted scheduling verification chain, digital twin constraint optimization provides real-time feedback and simulation for the scheduling process, ensuring that decisions can quickly adapt and adjust to changing operating environments, enhancing the system's adaptability and flexibility. The compilation of the dynamic scheduling instruction set provides a comprehensive set of anti-disturbance strategies, enabling the scheduling system to maintain stable operation in the face of external disturbances. Finally, by constructing a global scheduling optimization evidence chain, the system integrates information from various sources, traces scheduling deviations, ensures the accuracy of scheduling plans, and selects highly robust scheduling instructions, thereby enabling precise energy allocation decisions, optimizing resource utilization efficiency, and ensuring system stability. These steps, in tandem, promote the development of intelligent, distributed energy scheduling systems, providing a new solution for addressing the increasingly complex energy supply and demand environment.

[0011] Optionally, step S1 specifically includes:

[0012] Step S11: Acquire multimodal data, where the multimodal data includes meteorological data, power grid operation status data, and user behavior data; standardize the format of the multimodal data, and perform missing value interpolation and denoising to generate a standardized multimodal energy data set;

[0013] Step S12: constructing a dynamic spatiotemporal adjacency matrix based on the standardized multimodal energy dataset, and performing global attention modeling based on the spatial topological features and time series features in the dynamic spatiotemporal adjacency matrix to generate a spatiotemporal correlation feature map;

[0014] Step S13: performing abnormal pattern recognition on the spatiotemporal correlation feature graph, extracting abnormal fluctuation patterns, screening abnormal regions with high impact factors, and generating a spatiotemporal fusion abnormal feature tensor;

[0015] Step S14: Project the spatiotemporal fusion abnormal feature tensor into a feature space, calculate the abnormal distribution skewness, and obtain the abnormal feature projection matrix;

[0016] Step S15: label the abnormal pattern according to the abnormal feature projection matrix, generate abnormal classification labels, and calculate the abnormal spatiotemporal impact weights, thereby obtaining a spatiotemporal fusion abnormal feature annotation set.

[0017] The present invention comprehensively captures the multiple factors that affect energy scheduling by acquiring meteorological information, power grid operation, and user behavior, providing rich data support for subsequent decision-making. The data is formatted, missing values are interpolated, and denoised to ensure data accuracy and completeness. A dynamic spatiotemporal adjacency matrix is constructed based on the standardized data set, and combined with global attention modeling to capture the spatiotemporal correlations of different regions and time periods, providing support for accurate scheduling decisions. Through abnormal pattern recognition, potential supply and demand fluctuations are discovered in a timely manner, high-impact areas are screened, and scheduling efficiency is improved. Further feature space projection and abnormal distribution skewness analysis enhance the sensitivity of anomaly detection, reveal potential problem trends, and provide forward-looking data support. Finally, anomaly classification labels are generated and spatiotemporal impact weights are calculated to help the system identify the severity of anomalies, conduct reasonable management, and ensure the efficient and stable operation of the energy system.

[0018] Optionally, step S13 is specifically as follows:

[0019] Step S131: performing time series normalization on the spatiotemporal correlation feature map, removing periodic features and overall offsets, and obtaining a normalized spatiotemporal correlation feature map;

[0020] Step S132: performing abnormal pattern recognition on the standardized spatiotemporal correlation feature graph, calculating the local abnormality score of the standardized spatiotemporal correlation feature graph, and performing abnormal pattern classification to generate a preliminary screening abnormal pattern set;

[0021] Step S133: Based on the initial screening abnormal pattern set, calculate the probability distribution of the abnormal pattern in the time and space dimensions, and screen the abnormal regions with high impact factors to generate a matrix of abnormal regions with high impact factors;

[0022] Step S134: performing spatiotemporal feature expansion on the high impact factor abnormal region matrix, constructing a spatiotemporal correlation map of abnormal regions, and calculating the time alignment between abnormal regions in the spatiotemporal correlation map of abnormal regions, thereby generating a spatiotemporal abnormal correlation map;

[0023] Step S135: Perform spatiotemporal feature fusion according to the spatiotemporal anomaly association graph, perform cross-time step feature interaction on the spatiotemporal features, and select the optimal anomaly classification hyperparameters to generate a spatiotemporal fusion anomaly feature tensor.

[0024] The present invention performs time series standardization on the spatiotemporal correlation feature graph, removes periodic features and offsets, ensures data smoothness and true reflection, and improves the accuracy of subsequent analysis. The standardized data is used to identify abnormal patterns, effectively calculate local abnormality scores, accurately identify potential problems, and ensure the comprehensiveness of abnormality detection. By generating a probability distribution of the spatiotemporal dimensions of abnormal patterns, high-impact factor abnormal areas are screened, resource allocation is optimized, and scheduling decisions are concentrated in high-risk areas. The spatiotemporal features of the screened high-impact abnormal areas are expanded and a spatiotemporal correlation graph is constructed to reveal the spatiotemporal characteristics and propagation trends of abnormal fluctuations, providing a basis for decision-making. Finally, through spatiotemporal feature fusion and abnormal classification hyperparameter optimization, the abnormal pattern classification capability is improved, and a high-quality spatiotemporal fusion abnormal feature tensor is generated to provide accurate data support for energy scheduling.

[0025] Optionally, step S2 is specifically:

[0026] Step S21: performing feature space dimensionality reduction on the spatiotemporal fusion abnormal feature tensor, extracting abnormal features, and calculating the Euclidean distance and cosine similarity between abnormal features to generate an abnormal pattern similarity matrix;

[0027] Step S22: constructing an abnormal pattern spatiotemporal network based on the abnormal pattern similarity matrix, and calculating the global impact factor of the abnormal area in the abnormal pattern spatiotemporal network to generate an abnormal pattern impact vector;

[0028] Step S23: Based on the abnormal mode impact vector, the regional collaborative scheduling strategy is optimized using a preset distributed particle swarm optimization algorithm, a particle search space is set, and the agent strategy is adjusted based on the trade-off between local optimality and global optimality, thereby obtaining the regional collaborative scheduling vector;

[0029] Step S24: setting energy scheduling constraints based on the regional collaborative scheduling vector, calculating the Pareto optimal solution set, and generating a multi-objective optimization path set;

[0030] Step S25: Verify the temporal feasibility of the multi-objective optimization path set, evaluate the adaptability of the adversarial training environment path, and optimize the scheduling decision based on the preset scheduling objective function group, thereby generating a dynamic scheduling decision map, where the scheduling objective function group is specifically:

[0031] F(X)=α·F c (X)+β·F s (X)+γ·F ef (X)+δ·F en (X);

[0032]

[0033] Among them, F(X) is the scheduling objective function, F c (X) is the cost objective function, F s (X) is the stability objective function, F ef (X) is the efficiency objective function, F en (X) is the environmental impact objective function, X represents the scheduling decision vector, α is the cost objective weight factor, β is the stability weight factor, γ is the efficiency weight factor, δ is the environmental impact weight factor, T is the total monitoring time, t is the time index, N is the number of all regional nodes, i is the regional node index, C i (t) is the energy cost of the node in the i-th region at time t, P i (t) is the energy allocation ratio of the node in the i-th region at time t, E i (t) is the energy storage state factor of the node in the i-th region at time t, is the preset load of the node in the i-th region at time t, P max (i) is the maximum energy available value of the node in the i-th region, ζ i is the carbon emission factor of the i-th regional node.

[0034] The present invention constructs a scheduling objective function group by consulting relevant materials and based on expert experience to achieve supply and demand matching optimization. The function group includes the cost objective function F c (X), stability objective function F s (X), efficiency objective function F ef (X) and the environmental impact objective function F en (X), by introducing different weight factors (α, β, γ) to adjust the influence of each objective function, the scheduling system can find a balance between energy cost, stability, efficiency and environmental impact. The weights of different objectives can be adjusted according to actual needs to adapt to specific scheduling requirements and optimization goals, thereby achieving optimal scheduling in a dynamically changing energy supply and demand environment.

[0035] Cost objective function Fc (X) calculates the energy cost of regional nodes in the entire scheduling process, including the energy consumption and energy storage status of each regional node. First, by To calculate the energy allocation cost of each regional node at each time point t, where C i (t) is the energy cost of the node in the i-th region at time t, P i (t) is the energy distribution ratio of the node in the i-th region at time t. A regularization term is added Energy storage state factor E i (t) is penalized so that areas with excessive energy storage status bear more costs. This cost objective function reduces unnecessary expenditures while ensuring energy supply by accurately controlling costs and optimizes the overall cost by adjusting the energy storage status.

[0036] Stability objective function F s (X) measures the energy allocation ratio P of each regional node i (t) and preset load The difference between the two, using the absolute difference The difference is measured and summed at each time point t. By reducing these differences, scheduling can ensure optimized energy scheduling stability and avoid excessive fluctuations, which is crucial for ensuring user load stability. This helps ensure that energy supply at each regional node is as close as possible to expected demand, improving system stability and reducing supply and demand imbalances caused by unexpected situations.

[0037] Efficiency objective function F ef (X) Pass Measure the energy allocation efficiency of each regional node. By normalizing the energy allocation ratio P of each regional node i (t) and its maximum available energy P max (i) The system can evaluate the efficiency of each node in maximizing its resource utilization. This helps to improve the overall efficiency of energy distribution by optimizing the efficient use of energy and avoiding energy waste.

[0038] Environmental impact objective function F en (X) Pass The impact of energy distribution on the environment is measured by calculating the carbon emissions of each regional node. The carbon emission factor ζ for each region i and energy distribution ratio P i The product of (t) reflects the environmental burden of energy use in a region. Optimizing this goal can reduce carbon emissions and promote sustainable development. This helps guide energy scheduling strategies toward green, low-carbon development by comprehensively considering environmental factors, thereby reducing negative environmental impacts.

[0039] The present invention performs feature space dimensionality reduction on the spatiotemporal fusion anomaly feature tensor, extracts core anomaly features, and calculates Euclidean distance and cosine similarity to identify the similarity of anomaly patterns and avoid redundant information interfering with scheduling decisions. A spatiotemporal network of anomaly patterns is constructed and global impact factors are calculated to reveal the actual impact of anomaly patterns on the system, providing data support for optimizing scheduling strategies. A distributed particle swarm optimization algorithm is used to dynamically adjust regional collaborative scheduling strategies, optimize energy scheduling efficiency, balance local and global optimality, and improve system scheduling performance. Multi-objective optimization is performed by calculating the Pareto optimal solution set to ensure the best energy scheduling path set, and time series feasibility verification and path adaptability evaluation are performed to enhance system flexibility. The final generated dynamic scheduling decision map accurately reflects the scheduling path effect, ensuring that the system can adjust and optimize the energy scheduling plan in a timely and accurate manner.

[0040] Optionally, step S3 specifically includes:

[0041] Step S31: constructing an edge node game model based on the dynamic scheduling decision graph, setting the game subject and strategy space, and calculating the strategy payoff matrix of different game subjects to generate an initial game equilibrium solution set;

[0042] Step S32: Acquire historical regional energy dispatch data, and perform energy dispatch mode identification on the historical regional energy dispatch data to obtain historical energy dispatch mode data;

[0043] Step S33: Dynamically adjust the initial game equilibrium solution set, introduce a preset non-cooperative game mechanism, evaluate the optimal response strategies of different game players under energy scheduling constraints, and perform reinforcement learning optimization based on historical energy scheduling pattern data to generate a reinforcement game equilibrium solution set;

[0044] Step S34: Calculate the cross-regional energy mutual assistance optimization parameters based on the reinforcement game equilibrium solution set, set the mutual assistance triggering conditions, calculate the cross-regional energy mutual assistance benefits, and generate the cross-regional energy mutual assistance optimization plan;

[0045] Step S35: Conduct credible verification on the cross-regional energy mutual assistance optimization plan, calculate the mutual assistance contribution of each region, and perform adaptive credit evaluation based on the non-cooperative game mechanism to optimize the energy mutual assistance plan of each region and generate a credible scheduling verification chain.

[0046] The present invention constructs an edge node game model, simulates the behavior of different regions in energy scheduling, calculates the strategy benefit matrix, and generates an initial game equilibrium solution set, which provides a theoretical basis for multi-region coordination and energy scheduling, ensuring reasonable modeling and prediction of regional scheduling needs. By identifying historical energy scheduling patterns and extracting potential laws and patterns, the decision-making quality of the scheduling system is improved, and it helps to accurately predict future energy needs. Dynamically adjusting the game equilibrium solution set, introducing non-cooperative game mechanisms and reinforcement learning optimization, improves the adaptability, accuracy and flexibility of scheduling decisions. The setting of cross-regional energy mutual assistance optimization parameters and trigger conditions enables efficient resource sharing between regions and optimizes overall energy utilization efficiency. On the basis of trusted verification, adaptive credit evaluation is performed in combination with non-cooperative game mechanisms to ensure the credibility and reliability of the energy mutual assistance process.

[0047] Optionally, the step S4 of constructing a multi-physics field coupling scheduling unit is as follows:

[0048] Extract meteorological features, power grid operation features, and user behavior features from the trusted dispatch verification chain, and structure and layer them to generate a multi-physics field raw data set.

[0049] The standardized multimodal energy dataset is time-aligned, the time step is set to 5 minutes and interpolation is performed to obtain a time-aligned multi-physics field dataset;

[0050] A multi-layer non-Euclidean graph model was constructed based on the time-aligned multi-physics dataset. A topological mapping was established based on the time-aligned multi-physics dataset. The twin constraint parameter was set to 0.7. A preset digital twin verification mechanism was embedded in the topological mapping process. Graph neural network feature extraction was performed to generate an initial multi-physics coupling graph structure model.

[0051] Spectral clustering analysis was performed on the initial multi-physics coupling diagram structure model. The first 10 eigenvectors were selected for Laplace eigendecomposition. The feature truncation threshold was set to 0.85, and high-dimensional feature dimensionality reduction was performed to generate a spectrally smoothed multi-physics coupling diagram.

[0052] Perform a standardized transformation on the spectral smoothed multi-physics field coupling diagram, set the normalization interval [0,1] and perform time series trend correction to generate a standardized multi-physics field scheduling unit;

[0053] The standardized multi-physics field scheduling unit is corrected by Bayesian optimal difference data, and the confidence interval is set to 0.95 for confidence interpolation completion. The historical energy scheduling mode data is combined to perform distributed time series prediction to generate a multi-physics field coupling scheduling unit.

[0054] The present invention extracts and structures the characteristics of meteorology, power grid operation and user behavior, successfully integrates multi-physics field data to generate original data sets, provides all-round support for energy scheduling, and enhances the system's predictive ability for energy resource allocation. Time series alignment and missing value completion ensure data integrity and timeliness, avoid errors caused by missing data, and improve the accuracy of subsequent analysis. A multi-layer non-Euclidean graph model is constructed based on the time-aligned data set, which enhances the description of the relationship between meteorology, power grid and user behavior, and improves the adaptability of the system through the digital twin verification mechanism. Graph neural network feature extraction deeply explores complex nonlinear relationships, spectral clustering and Laplace eigendecomposition optimize computational efficiency, reduce dimensionality and reduce redundant information, and improve the efficiency and stability of multi-physics field coupling graphs. The coupling graph after spectral smoothing undergoes standardized transformation and time series correction to ensure the scientificity and consistency of scheduling decisions. The combination of Bayesian optimal difference correction and time series prediction improves the prediction accuracy of the scheduling unit and ensures that the system responds quickly and accurately in complex and dynamic environments.

[0055] Optionally, the compilation of the dynamic scheduling instruction set in step S4 is specifically:

[0056] Based on the energy scheduling constraints, the scheduling parameter boundaries are set for the multi-physics field coupling scheduling unit to generate an initial scheduling parameter set;

[0057] Dynamic scheduling optimization modeling is performed on the initial scheduling parameter set, the scheduling optimization objective function is set, and the distributed particle swarm optimization algorithm is used for strategy search. The particle swarm size is set to 50 and the iteration step length is set to 20s to generate the global optimal scheduling parameter set.

[0058] Compile the scheduling instruction set based on the global optimal scheduling parameter set, optimize the scheduling execution strategy, set the reward feedback factor to 0.8, and generate an enhanced scheduling instruction set;

[0059] Parallel simulations were conducted on the enhanced dispatch instruction set. The dispatch execution efficiency was evaluated under different combinations of meteorological characteristics, grid operating status characteristics, and user behavior characteristics. The execution efficiency threshold was set at 85%, and the task throughput was calculated. The instruction set execution path was optimized to generate an optimized dispatch instruction set.

[0060] Instruction set compatibility mapping is performed based on the optimized scheduling instruction set, and a cross-platform compatibility threshold is set to 90% to generate a cross-platform compatible dynamic scheduling instruction set.

[0061] The present invention ensures that the scheduling scheme meets actual needs and constraints, thereby improving system efficiency and operability by setting energy scheduling constraints and generating an initial scheduling parameter set. By using a distributed particle swarm optimization algorithm for dynamic scheduling optimization, it is possible to find the optimal solution among multiple schemes, optimize the objective function to ensure that the scheduling strategy accurately responds to changing energy demands, enhance search efficiency, and reduce system costs and risks. The generation of a global optimal scheduling parameter set optimizes the scheduling instruction set, and the introduction of reward feedback factors promotes continuous optimization of the system, improving long-term scheduling execution efficiency and stability. The scheduling instruction set is verified by parallel simulation to ensure efficient system operation in complex environments and extreme weather conditions, and timely adjust the task execution path to optimize the execution effect. Compatibility mapping and cross-platform compatibility settings ensure that the dynamic scheduling instruction set is smoothly executed between different platforms, which improves the flexibility of the scheduling system, promotes cross-regional and cross-platform joint scheduling, and ensures the efficient and stable operation of the energy system.

[0062] Optionally, the scheduling deviation tracing in step S5 is specifically as follows:

[0063] Analyze the global scheduling evidence chain, extract the planned scheduling value, actual execution value and adjustment instructions within 5 minutes of each scheduling cycle, calculate the scheduling deviation vector, and generate a scheduling execution deviation information database;

[0064] Based on the multi-physics field coupling scheduling unit and the scheduling execution deviation information library, a dynamic time warping causal analysis of scheduling deviation is performed. The significance level is set to 0.05 to calculate the influence weight of each factor in the multi-physics field coupling scheduling unit on the scheduling deviation, thereby scheduling the impact factor matrix;

[0065] The scheduling state deviation transition probability is calculated based on the scheduling impact factor matrix, and the time window is set to 30 minutes. The deviation propagation path is constructed using a time-series sliding method. The influence weights between each state are calculated based on the deviation propagation path, the deviation causal relationship is inferred, and a deviation spatiotemporal traceability model is generated.

[0066] Based on the deviation spatiotemporal traceability model, the optimal response strategy of each regional dispatching entity is set, and dynamic optimization is performed to screen dispatching instructions with a robustness score higher than 0.85 to obtain a highly robust dispatching instruction set.

[0067] Based on the highly robust scheduling instruction set, the energy scheduling optimization objective function is defined, and the energy scheduling optimization parameters are calculated to obtain the optimized energy scheduling parameter matrix;

[0068] The energy supply and demand matching degree of each region is calculated based on the optimized energy scheduling parameter matrix, and the energy distribution equilibrium optimization is performed based on the reinforcement game equilibrium solution set. The energy allocation decision is generated and uploaded to the regional energy scheduling platform to execute the energy allocation task.

[0069] This method analyzes the global scheduling evidence chain to extract scheduling deviation information, gaining a deep understanding of the differences between planned scheduling values, actual execution values, and adjustment instructions. It generates a scheduling deviation vector, providing a detailed record for subsequent optimization and problem analysis. Through dynamic time-warping causal analysis of scheduling deviations, significance levels are set and factor influence weights are calculated to accurately identify key factors influencing scheduling deviations. A scheduling influence factor matrix is generated, providing a scientific basis for optimizing scheduling strategies. Deviation propagation paths are constructed and influence weights are calculated to reveal the deviation transmission mechanism between scheduling states. A deviation spatiotemporal traceability model is generated, and the causal relationship between deviations is inferred, providing an efficient response strategy for dealing with uncertainty and emergencies. By screening scheduling instruction sets with high robustness scores, the stability and reliability of scheduling instructions are ensured, improving system resilience. By defining the optimization objective function and calculating energy scheduling parameters, the system energy configuration is precisely adjusted to achieve optimal energy supply and demand matching. Energy is rationally allocated through enhanced game equilibrium optimization, improving scheduling efficiency. The resulting energy allocation decision is uploaded to the regional scheduling platform, ensuring that the system efficiently and accurately executes energy allocation tasks and maintains stable operation.

[0070] Optionally, this specification further provides a multimodal data fusion system based on energy scheduling, which is used to execute the multimodal data fusion method based on energy scheduling as described above. The multimodal data fusion system based on energy scheduling includes:

[0071] The abnormal fluctuation feature annotation module is used to obtain multimodal data and construct a dynamic spatiotemporal adjacency matrix based on the multimodal data; the abnormal fluctuation features are extracted according to the dynamic spatiotemporal adjacency matrix to obtain a spatiotemporal fusion abnormal feature annotation set;

[0072] The scheduling decision deduction module is used to perform distributed policy gradient aggregation on the spatiotemporal fusion anomaly feature annotation set to obtain the regional collaborative scheduling vector; based on the regional collaborative scheduling vector, it deduces the multi-objective optimization path to obtain the dynamic scheduling decision map;

[0073] The game equilibrium calculation module is used to perform edge node game equilibrium calculations based on the dynamic scheduling decision graph, and perform cross-regional energy mutual assistance verification to obtain a trusted scheduling verification chain;

[0074] The dynamic scheduling instruction compilation module is used to perform digital twin constraint optimization on the trusted scheduling verification chain, build a multi-physics field coupling scheduling unit, and obtain a closed-loop scheduling digital twin. Based on the closed-loop scheduling digital twin, the dynamic scheduling instruction set is compiled to obtain a disturbance-resistant energy scheduling strategy library.

[0075] The scheduling deviation tracing module is used to obtain real-time energy supply and demand data, and dynamically evolve the anti-disturbance strategy for the real-time energy supply and demand data based on the anti-disturbance energy scheduling strategy library to generate a global scheduling optimization evidence chain; based on the global scheduling evidence chain, scheduling deviation tracing is implemented, and highly robust scheduling instructions are screened to obtain energy allocation decisions.

[0076] The multimodal data fusion system based on energy scheduling of the present invention can implement any one of the multimodal data fusion methods based on energy scheduling of the present invention, and is used to combine the operation and signal transmission medium between various modules to complete the multimodal data fusion method based on energy scheduling. The internal modules of the system cooperate with each other, thereby helping to improve the efficiency and reliability of energy scheduling.

[0077] Optionally, the present specification also provides a computer-readable storage medium storing a computer program, which, when executed, implements the multimodal data fusion method based on energy scheduling as described above. BRIEF DESCRIPTION OF THE DRAWINGS

[0078] Other features, objects and advantages of the present invention will become more apparent upon reading the detailed description of non-limiting embodiments thereof made with reference to the following drawings:

[0079] Figure 1 Schematic diagram of the steps of the multimodal data fusion method based on energy scheduling of the present invention;

[0080] Figure 2 Detailed step flow diagram of step S1 in the present invention;

[0081] Figure 3 Detailed step flow diagram of step S2 in the present invention;

[0082] The purpose, features and advantages of the present invention will be further described with reference to the accompanying drawings and in conjunction with the embodiments. DETAILED DESCRIPTION

[0083] The following is a clear and complete description of the technical method of the present invention in conjunction with the accompanying drawings. It is obvious that the embodiments described are part of the embodiments of the present invention, but not all of the embodiments. Based on the embodiments of the present invention, all other embodiments obtained by those skilled in the art without making any creative efforts are within the scope of protection of the present invention.

[0084] In addition, the accompanying drawings are merely schematic illustrations of the present invention and are not necessarily drawn to scale. Identical reference numerals in the figures denote identical or similar parts, and thus repetitive descriptions thereof will be omitted. Some of the block diagrams shown in the accompanying drawings are functional entities that do not necessarily correspond to physically or logically separate entities. These functional entities may be implemented in software, in one or more hardware modules or integrated circuits, or in different network and / or processor and / or microcontroller approaches.

[0085] It should be understood that although the terms "first," "second," and the like may be used herein to describe various elements, these elements should not be limited by these terms. These terms are used solely to distinguish one element from another. For example, a first element may be referred to as a second element, and similarly, a second element may be referred to as a first element, without departing from the scope of the exemplary embodiments. The term "and / or" as used herein includes any and all combinations of one or more of the listed associated items.

[0086] To achieve this, please refer to Figures 1 to 3 The present invention provides a multimodal data fusion method based on energy scheduling, the method comprising the following steps:

[0087] Step S1: Acquire multimodal data and construct a dynamic spatiotemporal adjacency matrix based on the multimodal data; extract abnormal fluctuation features based on the dynamic spatiotemporal adjacency matrix to obtain a spatiotemporal fusion abnormal feature annotation set;

[0088] In this embodiment, multimodal data is collected from the regional energy dispatching platform, including meteorological data, power grid operation status data, and user behavior data. By fusing these data, a dynamic spatiotemporal adjacency matrix is constructed. In this process, a spatiotemporal clustering algorithm is applied to calibrate the correlation between different time points and spatial regions. A sliding window technique is used to analyze the dynamic characteristics of the adjacency matrix over time, and the window size is set to 5 minutes. This method can dynamically capture the mutual dependence between nodes in time and space and form a matrix containing spatiotemporal features. Then, through the matrix-based abnormal fluctuation feature extraction algorithm, abnormal fluctuations caused by external disturbances or system failures are identified, and a spatiotemporal fusion abnormal feature annotation set is generated. In the abnormal fluctuation feature extraction process, the annotation accuracy requirement is set to 95% to ensure that the extracted abnormal features meet the predetermined standards and the accuracy of data annotation is guaranteed.

[0089] Step S2: Perform distributed policy gradient aggregation on the spatiotemporal fusion anomaly feature annotation set to obtain a regional collaborative scheduling vector; deduce a multi-objective optimization path based on the regional collaborative scheduling vector to obtain a dynamic scheduling decision map;

[0090] In this embodiment, a distributed policy gradient aggregation method is adopted to pass the spatiotemporal fusion anomaly feature annotation set into multiple regional nodes to generate regional collaborative scheduling vectors. In this process, the policy update between regional nodes is based on the gradient descent method to ensure that in the multi-region collaborative scheduling process, the policy optimization between nodes can be carried out towards the global optimum. Each node calculates its policy gradient and collaboratively optimizes the scheduling decision through information sharing between regions. A reinforcement learning algorithm is used to deduce the multi-objective optimization path, and a dynamic scheduling decision map is generated based on the evaluation results of the path. In this process, the learning rate is set to 0.01 to ensure that the model can converge quickly in a large-scale node system, and the number of training rounds is 1000 times each time, so that the policy optimization gradually tends to the optimal.

[0091] Step S3: Perform edge node game equilibrium calculation based on the dynamic scheduling decision graph, and perform cross-regional energy mutual assistance verification to obtain a trusted scheduling verification chain;

[0092] In this embodiment, edge node game equilibrium calculations are performed based on a dynamic scheduling decision graph to ensure that the scheduling strategies between multiple regional nodes are optimized and to achieve cross-regional energy mutual assistance verification. In the game model, the resource allocation strategy of each node is considered, and the equilibrium solution of the game is calculated using non-cooperative game theory. In this process, the game strategy space includes five different scheduling strategies for each regional node, and the game equilibrium accuracy requirement does not exceed 1%, ensuring that the strategy adjustments during the game process meet the optimization requirements of the system. At the same time, the cross-regional energy mutual assistance verification process will evaluate the energy supply and demand matching between different regions in real time, verify whether the energy scheduling plans of each region can be effectively coordinated, and ensure the optimization of energy distribution between regions. Ultimately, a trusted scheduling verification chain is generated to verify the accuracy and feasibility of scheduling decisions.

[0093] Step S4: Perform digital twin constraint optimization on the trusted scheduling verification chain, construct a multi-physics field coupling scheduling unit, and obtain a closed-loop scheduling digital twin; compile a dynamic scheduling instruction set based on the closed-loop scheduling digital twin to obtain an anti-disturbance energy scheduling strategy library;

[0094] In this embodiment, digital twin technology is applied to the trusted scheduling verification chain to perform closed-loop optimization of the scheduling process. The digital twin can reflect the changes in physical quantities in the multi-physics field coupling scheduling unit, such as current, voltage, load, etc. in real time. Through two-way interaction with the physical system, the digital twin can continuously adjust the scheduling strategy based on real-time data feedback. The optimization process uses the digital twin constraint optimization algorithm to generate a closed-loop scheduling digital twin, and compiles a dynamic scheduling instruction set based on the body. The instruction set contains multiple scheduling strategies and generates an anti-disturbance strategy library based on the disturbance conditions of the energy system. The anti-disturbance strategy library covers emergency scheduling solutions under various disturbance scenarios to ensure that the scheduling instructions can be stably executed in the event of an emergency or system failure. During the optimization process, the accuracy of the digital twin is required to reach 99%, and data synchronization is performed every 30 seconds to ensure the real-time and accuracy of the data.

[0095] Step S5: Acquire real-time energy supply and demand data, and dynamically evolve the anti-disturbance strategy for the real-time energy supply and demand data based on the anti-disturbance energy scheduling strategy library to generate a global scheduling optimization evidence chain; implement scheduling deviation tracing based on the global scheduling evidence chain, and screen highly robust scheduling instructions to obtain energy allocation decisions.

[0096] In this example, a regional energy dispatch platform acquires real-time energy supply and demand data, including regional power load, user electricity usage behavior, and meteorological data. User profile data is also collected, capturing historical consumption habits, device types, and preferences. Subsequently, a regional energy demand forecasting model based on a long short-term memory (LSTM) network is used to predict regional energy demand for the next 24 hours, using a 5-minute time step. The target forecast accuracy is over 95%. Based on the forecast results, a segmented compensation mechanism is designed. For example, a high subsidy policy is implemented during off-peak hours (e.g., from 10:00 PM to 6:00 AM the following day). Dynamic electricity price adjustments, with the subsidy percentage updated every minute, incentivize users to shift some of their electricity load to off-peak hours, thereby balancing grid load. Furthermore, user profile data is leveraged to generate personalized energy-saving solutions for different users, such as intelligently recommending optimized appliance operating modes, adjusting air conditioning temperatures, or providing energy-saving reminders, thereby encouraging users to actively participate in energy-saving regulation. All real-time supply and demand data, forecast results, compensation strategies, and user responses are integrated. Subsequently, relying on the anti-disturbance energy scheduling strategy library, the integrated real-time data is dynamically evolved through reinforcement learning and distributed optimization algorithms, and the anti-disturbance strategy is automatically adjusted to cope with load fluctuations and abnormal disturbances, ensuring that the scheduling plan can remain efficient and stable in a changing environment. All real-time data, predicted outputs, compensation strategies, and user responses are converted into different blocks, recorded in real time, and integrated into the global scheduling optimization evidence chain. This evidence chain uses blockchain technology to achieve data immutability and full-link traceability, providing a transparent and reliable basis for subsequent scheduling decisions. Next, based on the global evidence chain, the scheduling deviation is traced between real-time data and historical scheduling data. By comparing the actual execution value with the predicted value, the deviation vector is calculated, and the deviation propagation path modeling technology (for example, based on the Markov chain state transition model) is combined with the Bayesian network to infer the deviation causal relationship. The deviation transmission mechanism in the scheduling process is identified, thereby screening high-robustness scheduling instructions with a robustness score above 85%. Finally, based on the screened high-robustness scheduling instruction set, the system defines and calculates the energy scheduling optimization objective function, comprehensively considers multiple indicators such as cost, stability, efficiency and environmental impact, obtains the optimized energy scheduling parameter matrix, and calculates the energy supply and demand matching degree of each region based on this, and finally generates a global energy allocation decision.

[0097] Optionally, step S1 specifically includes:

[0098] Step S11: Acquire multimodal data, where the multimodal data includes meteorological data, power grid operation status data, and user behavior data; standardize the format of the multimodal data, and perform missing value interpolation and denoising to generate a standardized multimodal energy data set;

[0099] In this embodiment, a regional energy dispatching platform is used to collect various types of data in real time, including regional meteorological data (such as temperature, humidity, wind speed, etc.), power grid operation status data (such as voltage, current, load, etc.), and user behavior data (such as electricity usage habits, consumption patterns, response behaviors, etc.). The collection frequency of these data is set to once per minute to ensure sufficient real-time performance. During the data processing process, a unified format is first used for standardization to convert data from different sources into the same timestamp and unit (for example, all voltage data is converted to volts, and all temperature data is converted to degrees Celsius). For data with missing values, linear interpolation and time series interpolation methods are used to fill in the missing values; for noisy data, the Kalman filter is used to denoise the power grid status and user behavior data. After these processes, a standardized multimodal energy data set is generated, which includes various types of data every minute, covering all records in the past 24 hours, and providing basic data for subsequent steps.

[0100] Step S12: constructing a dynamic spatiotemporal adjacency matrix based on the standardized multimodal energy dataset, and performing global attention modeling based on the spatial topological features and time series features in the dynamic spatiotemporal adjacency matrix to generate a spatiotemporal correlation feature map;

[0101] In this embodiment, the generated standardized multimodal energy data set is used to first construct a dynamic spatiotemporal adjacency matrix. This matrix not only represents the spatial topological structure between different regions, but also contains spatiotemporal characteristics. Specifically, the spatial topological characteristics are calculated based on the grid connection relationship (such as the distribution of substations and the connectivity of grid nodes), while the time series characteristics are modeled based on the electricity consumption of each region, changes in meteorological data, etc. For the calculation of the spatiotemporal adjacency matrix, the time window is set to 15 minutes to ensure that the data within each time window is fully considered. On this basis, a global attention mechanism is used to model the dynamic spatiotemporal adjacency matrix. Through the self-attention mechanism, a dynamic weight is assigned to each regional node, which represents the relative importance of the node in the overall system. For each node, the influence of its adjacent regions and time steps is considered to form a spatiotemporal correlation feature map. This graph structure effectively captures the interactive relationship between different regions and their dynamic characteristics over time.

[0102] Step S13: performing abnormal pattern recognition on the spatiotemporal correlation feature graph, extracting abnormal fluctuation patterns, screening abnormal regions with high impact factors, and generating a spatiotemporal fusion abnormal feature tensor;

[0103] In this embodiment, based on the generated spatiotemporal correlation feature map, the abnormal fluctuation pattern is extracted by a multidimensional abnormal pattern recognition algorithm (for example, an abnormality detection algorithm based on local weighted regression). By calculating the standard deviation of the load fluctuation and power consumption pattern of each area, the areas with significant fluctuations or irregular behaviors in a specific time period are identified. These fluctuations are caused by power grid failures, meteorological anomalies or changes in user behavior. In order to further screen out abnormal areas that have a greater impact on the overall system, an impact factor calculation rule is set - for example, the abnormal fluctuation amplitude of the area is greater than a set threshold, or the abnormal duration exceeds a certain time window (such as 30 minutes). After screening out these high-impact factor areas, a spatiotemporal fusion abnormality feature tensor is generated, which combines the spatiotemporal features with the abnormal fluctuation pattern to provide comprehensive information for subsequent anomaly labeling and impact assessment.

[0104] Step S14: Project the spatiotemporal fusion abnormal feature tensor into a feature space, calculate the abnormal distribution skewness, and obtain the abnormal feature projection matrix;

[0105] In this embodiment, for the generated spatiotemporal fusion anomaly feature tensor, feature space projection is first performed. High-dimensional data is mapped to low-dimensional space through dimensionality reduction techniques such as principal component analysis (PCA), retaining the main features and reducing redundant information. After projection, statistical methods such as skewness and kurtosis are used to calculate the skewness of the abnormal data. These indicators are used to measure the distribution form of the data. For example, skewness reflects the symmetry of the data distribution, and kurtosis indicates the steepness of the data distribution. Through this process, it is possible to effectively identify which abnormal patterns have long-term and strong skewed distributions, thereby inferring their potential impact on the entire power grid scheduling and energy management.

[0106] Step S15: label the abnormal pattern according to the abnormal feature projection matrix, generate abnormal classification labels, and calculate the abnormal spatiotemporal impact weights, thereby obtaining a spatiotemporal fusion abnormal feature annotation set.

[0107] In this embodiment, abnormal pattern labeling is performed based on the calculated abnormal feature projection matrix. By setting a threshold (such as abnormal fluctuations with a skewness coefficient greater than 1 are marked as high-impact abnormalities), each abnormal pattern is divided into different levels according to these labeling rules: mild abnormalities (little impact on the power grid system), moderate abnormalities (impact on some areas), and severe abnormalities (may cause power grid instability or user service interruption). At the same time, the spatiotemporal impact weight of each abnormal area is calculated, which represents the degree of impact of the abnormality on other areas and the entire power grid. For example, if the abnormal pattern of a certain area simultaneously causes instability in the power grid status of multiple adjacent areas, the spatiotemporal impact weight of the area will be increased. Through these calculations, a spatiotemporal fusion abnormal feature labeling set is finally obtained, which provides important reference data for subsequent scheduling decisions, enabling the system to optimize the power grid scheduling strategy based on abnormal characteristics.

[0108] Optionally, step S13 is specifically as follows:

[0109] Step S131: performing time series normalization on the spatiotemporal correlation feature map, removing periodic features and overall offsets, and obtaining a normalized spatiotemporal correlation feature map;

[0110] In this embodiment, the time series of the spatiotemporal correlation feature map is normalized to remove periodic features and overall offsets. Specifically, for the feature data of each regional node, the sliding window mean normalization (the window size is set to 24 hours) is used to eliminate the influence of the daily cycle, and the mean of each feature is normalized to 0 and the variance is normalized to 1 by the Z-score normalization method to ensure that data of different dimensions can be analyzed on the same scale. In addition, in order to eliminate the overall trend offset, a differential transformation (for example, a first-order difference) is used to remove the global offset effect that changes with time, so that the data mainly reflects local dynamic characteristics. After standardization, a standardized spatiotemporal correlation feature map is obtained, which can more accurately depict the relative abnormality of each regional node, rather than the original data affected by long-term trends or offsets.

[0111] Step S132: performing abnormal pattern recognition on the standardized spatiotemporal correlation feature graph, calculating the local abnormality score of the standardized spatiotemporal correlation feature graph, and performing abnormal pattern classification to generate a preliminary screening abnormal pattern set;

[0112] In this embodiment, a local anomaly factor (LOF) algorithm based on a dynamic threshold is used to calculate the local anomaly score of each regional node. The LOF algorithm evaluates the degree of deviation by comparing the feature similarity of a node with other nodes in its neighborhood, and sets the range of local anomaly scores to 0 to 1, where nodes exceeding 0.75 are considered to be highly abnormal nodes. Subsequently, DBSCAN (density-based spatial clustering method) is used to classify abnormal patterns. This method can effectively distinguish isolated abnormal points from continuous abnormal patterns, thereby screening out abnormal patterns with consistency. Finally, a preliminary screening abnormal pattern set is generated, which contains all identified abnormal areas and their categories (such as sudden anomalies, periodic anomalies, random anomalies, etc.).

[0113] Step S133: Based on the initial screening abnormal pattern set, calculate the probability distribution of the abnormal pattern in the time and space dimensions, and screen the abnormal regions with high impact factors to generate a matrix of abnormal regions with high impact factors;

[0114] In this embodiment, the time intervals of abnormal occurrences in each region are counted, and its time series probability distribution is estimated based on the Gaussian mixture model (GMM), and the spatial impact factor (such as the probability of the abnormal pattern propagating in the adjacent area) is calculated at the same time. For the spatial impact factor, the impact radius is set (such as 5 kilometers), the possibility of the abnormal pattern propagating within the impact radius is calculated, and the degree of its spatiotemporal impact is adjusted by the sliding average of the time window (the window size is set to 30 minutes). Subsequently, the high impact factor abnormal area is screened according to the set impact threshold (such as the spatiotemporal impact probability is greater than 0.8), and a high impact factor abnormal area matrix is constructed. The matrix is used to characterize the spatiotemporal distribution characteristics of high abnormal areas in the entire energy network.

[0115] Step S134: performing spatiotemporal feature expansion on the high impact factor abnormal region matrix, constructing a spatiotemporal correlation map of abnormal regions, and calculating the time alignment between abnormal regions in the spatiotemporal correlation map of abnormal regions, thereby generating a spatiotemporal abnormal correlation map;

[0116] In this embodiment, for each high-abnormal area, the degree of overlap between it and other abnormal areas in the time dimension is calculated, and the time alignment between different abnormal areas is calculated based on the dynamic time warping (DTW) method. A time window is set (such as the past 24 hours), and time series data is extracted for each high-impact factor abnormal area, and the DTW distance between it and the time series of other abnormal areas is calculated. Time alignment is used to measure whether the anomalies of two areas have similar time evolution patterns. For example, if the anomaly time alignment of two areas exceeds 0.9, they are considered to belong to the same type of spatiotemporal anomaly pattern. On this basis, a weighted spatiotemporal graph convolutional network (ST-GCN) is used to construct a spatiotemporal correlation graph of abnormal areas. The high-impact factor abnormal areas are used as nodes in the graph, and the time alignment calculated by DTW is used as the weight of the edge to construct a spatiotemporal correlation graph of abnormal areas. Spatial adjacency information (such as the power grid topology or geographical distance of adjacent areas) is further introduced to ensure that the correlation of abnormal areas not only depends on time synchronization, but also combines spatial distribution characteristics. Anomaly patterns are hierarchically clustered based on the temporal alignment between different anomaly regions. Using an agglomerative clustering approach, similar anomaly regions are gradually merged from the bottom up, ensuring that highly correlated anomaly patterns are identified in the spatiotemporal dimensions to identify spatiotemporal anomaly patterns with global impact. Clustering termination conditions are set (e.g., clustering is stopped when the temporal alignment is greater than 0.85 and the spatial proximity is less than a set threshold), thereby forming multiple consistent anomaly clusters. Finally, a spatiotemporal anomaly correlation map is generated, which is used to characterize the spatiotemporal correlation between anomaly regions to assist in subsequent anomaly feature fusion and classification optimization.

[0117] Step S135: Perform spatiotemporal feature fusion according to the spatiotemporal anomaly association graph, perform cross-time step feature interaction on the spatiotemporal features, and select the optimal anomaly classification hyperparameters to generate a spatiotemporal fusion anomaly feature tensor.

[0118] In this embodiment, the spatiotemporal features are interacted across time steps, and a bidirectional long short-term memory network (Bi-LSTM) is used to model the temporal evolution relationship of the abnormal pattern, so that the abnormal pattern of the current time step can be predicted in combination with the information of the historical time step. Secondly, the feature weights of different time steps are dynamically adjusted based on the attention mechanism to improve the temporal consistency of the abnormal pattern. At the same time, a hyperparameter search algorithm (such as Bayesian optimization) is used to select the hyperparameters of the optimal abnormal classification model, including the learning rate (search range: 0.001-0.01), hidden layer dimension (search range: 64-256), etc., to optimize the classification accuracy. Finally, a spatiotemporal fusion abnormal feature tensor is generated, which contains the spatiotemporal features of all abnormal patterns and can be used for subsequent energy scheduling optimization and abnormal response strategy generation.

[0119] Optionally, step S2 is specifically:

[0120] Step S21: performing feature space dimensionality reduction on the spatiotemporal fusion abnormal feature tensor, extracting abnormal features, and calculating the Euclidean distance and cosine similarity between abnormal features to generate an abnormal pattern similarity matrix;

[0121] In this embodiment, the feature space dimensionality reduction is performed on the spatiotemporal fusion anomaly feature tensor. The t-SNE (t-Distributed Stochastic Neighbor Embedding) or PCA (Principal Component Analysis) method is used to map high-dimensional anomaly features to low-dimensional space, ensuring that the data retains key pattern information during the dimensionality reduction process. Specifically, the dimensionality reduction target dimension is set to 20, and the anomaly features in different regions are standardized to eliminate dimensional differences. In the feature space after dimensionality reduction, high-impact anomaly pattern features and medium-impact anomaly pattern features are extracted, and the Euclidean distance between the anomaly pattern features is further calculated to measure the spatial similarity between the anomaly features. In addition, the cosine similarity of the anomaly features is calculated, and the similarity threshold is set to 0.85. When the similarity of two anomaly features is higher than this threshold, they are considered to belong to the same pattern. Finally, an anomaly pattern similarity matrix is constructed based on the above calculation results, and the elements of the matrix represent the similarity strength between different anomaly features.

[0122] Step S22: constructing an abnormal pattern spatiotemporal network based on the abnormal pattern similarity matrix, and calculating the global impact factor of the abnormal area in the abnormal pattern spatiotemporal network to generate an abnormal pattern impact vector;

[0123] In this embodiment, an abnormal pattern spatiotemporal network is constructed based on the abnormal pattern similarity matrix. Among them, the abnormal patterns are used as nodes, and the similarities between patterns are used as edge weights to form a weighted undirected graph. After constructing the network, the global impact factor of each abnormal area is calculated, and the PageRank algorithm or the HITS (Hyperlink-Induced Topic Search) algorithm is used to measure the importance of each abnormal pattern in the entire network. During the specific calculation, the iteration step size is set to 10 times and the convergence threshold is set to 0.001 to ensure that the calculation results are stable. Finally, the abnormal pattern with the highest impact factor is extracted, and an abnormal pattern influence vector is formed, where the vector dimension corresponds to all abnormal areas in the system, and each element represents the degree of influence of the abnormal pattern in the area on the global system.

[0124] Step S23: Based on the abnormal mode impact vector, the regional collaborative scheduling strategy is optimized using a preset distributed particle swarm optimization algorithm, a particle search space is set, and the agent strategy is adjusted based on the trade-off between local optimality and global optimality, thereby obtaining the regional collaborative scheduling vector;

[0125] In this embodiment, the regional collaborative scheduling strategy is optimized based on the abnormal pattern impact vector. An improved distributed particle swarm optimization (DPSO) algorithm is used to enable the intelligent agents in each region to autonomously optimize local resources. In specific implementation, the algorithm is improved based on selected parameters (such as energy supply and demand imbalance, scheduling cost, and carbon emission targets). This includes dynamic inertia weight adjustment, adaptive expansion of the particle search space, dynamic correction of local optimality, and adaptive optimization of the global optimal weight to ensure search efficiency and convergence stability. The particle search space range is set to [0, 1], the particle swarm size is 50, and the maximum number of iterations is 100. A dynamic inertia weight w = 0.9-0.4 is introduced (decreasing with the number of iterations), and the local and global optimal update weights are set to c1 = 1.5 and c2 = 1.7, respectively, to achieve a balance between exploration and exploitation. During algorithm execution, the intelligent agents dynamically adjust the local search direction based on the regional energy consumption pattern and the abnormal pattern impact vector to ensure that the optimization results converge to an efficient scheduling strategy. Ultimately, a regional collaborative scheduling vector is obtained, in which each element corresponds to the optimal energy scheduling strategy parameters for a region.

[0126] Step S24: setting energy scheduling constraints based on the regional collaborative scheduling vector, calculating the Pareto optimal solution set, and generating a multi-objective optimization path set;

[0127] In this embodiment, based on the regional collaborative scheduling vector, energy scheduling constraints are set, mainly including: (1) energy supply and demand balance constraints, ensuring that the energy supply of each region meets the demand; (2) equipment operation safety constraints, ensuring that the load rate of each device does not exceed the set threshold (such as 90%); (3) carbon emission constraints, controlling the total carbon emissions to not exceed the target upper limit (such as 500 tons of CO2 / day). Under the premise of meeting these constraints, a multi-objective genetic algorithm (NSGA-II) is used to calculate the Pareto optimal solution set, and the optimization objectives include minimizing energy loss, minimizing scheduling costs, and maximizing supply and demand matching. During the calculation process, the population size is set to 100, the mutation probability is 0.1, the crossover probability is 0.9, and 50 rounds of iteration are performed to ensure the stability of the optimal solution. Finally, a set of multi-objective optimization paths is generated, and each path corresponds to a feasible energy scheduling strategy.

[0128] Step S25: Verify the temporal feasibility of the multi-objective optimization path set, evaluate the adaptability of the adversarial training environment path, and optimize the scheduling decision based on the preset scheduling objective function group, thereby generating a dynamic scheduling decision map, where the scheduling objective function group is specifically:

[0129] F(X)=α·F c (X)+β·F s (X)+γ·F ef (X)+δ·Fen (X);

[0130]

[0131]

[0132] Among them, F(X) is the scheduling objective function, F c (X) is the cost objective function, F s (X) is the stability objective function, F ef (X) is the efficiency objective function, F en (X) is the environmental impact objective function, X represents the scheduling decision vector, α is the cost objective weight factor, β is the stability weight factor, γ is the efficiency weight factor, δ is the environmental impact weight factor, T is the total monitoring time, t is the time index, N is the number of all regional nodes, i is the regional node index, C i (t) is the energy cost of the node in the i-th region at time t, P i (t) is the energy allocation ratio of the node in the i-th region at time t, E i (t) is the energy storage state factor of the node in the i-th region at time t, is the preset load of the node in the i-th region at time t, P max (i) is the maximum energy available value of the node in the i-th region, ζ i is the carbon emission factor of the i-th regional node.

[0133] In this embodiment, the temporal feasibility of the multi-objective optimization path set is verified, and the path adaptability is evaluated using an adversarial training environment. Specifically, adversarial samples are constructed based on historical supply and demand data, extreme energy demand fluctuations are simulated through GAN (generative adversarial network), and the adaptability of the optimized path is tested. After the adaptability evaluation, the optimized path is further adjusted based on a preset set of scheduling objective functions (such as scheduling delay minimization objective and load balancing objective) to improve its execution effect in the actual environment. Finally, based on the optimized multi-objective optimization path set, four types of scheduling objective functions, namely cost, stability, efficiency, and environmental impact, are integrated to construct a dynamic scheduling decision map. A spatiotemporal graph neural network (ST-GNN) is used to extract energy flow characteristics, and combined with a Markov decision process (MDP) to optimize regional scheduling strategies, a global scheduling solution is finally generated that includes information such as regional energy scheduling strategies, cross-regional energy flows, and energy load change trends. A spatiotemporal correlation model is constructed based on the ST-GNN, taking the multi-objective optimization path as input to extract energy scheduling characteristics of each region, such as local load distribution, cross-regional energy interaction, and the degree of supply and demand imbalance. A graph attention mechanism (GAT) is used to calculate the energy dispatch weight for each region and dynamically model energy flows. Secondly, a Markov decision process (MDP) is used to optimize the dispatch strategy. The state space is defined as the energy supply and demand status and historical dispatch plans for each region, and the action space is defined as the optional energy allocation strategies (such as local load balancing, cross-regional allocation, and peak shaving and valley filling). The reward function is composed of four objective functions: cost, stability, efficiency, and environmental impact. The MDP strategy is optimized using a reinforcement learning (RL) algorithm, enabling the agent to dynamically adjust the dispatch plan under varying supply and demand conditions, achieving comprehensive optimization of the four objectives. Furthermore, a time series prediction (LSTM) model is used to model energy load trends. A sliding window approach is used to analyze historical load data and simulate future load changes based on the dispatch strategy. Finally, the ST-GNN spatiotemporal correlation, MDP optimization strategy, and LSTM trend prediction are integrated to construct a dynamic dispatch decision graph, achieving global optimization of regional energy dispatch and providing decision support for real-time scheduling.

[0134] Optionally, step S3 specifically includes:

[0135] Step S31: constructing an edge node game model based on the dynamic scheduling decision graph, setting the game subject and strategy space, and calculating the strategy payoff matrix of different game subjects to generate an initial game equilibrium solution set;

[0136] In this embodiment, an edge node game model is constructed based on a dynamic scheduling decision graph. First, the game subjects are set, including regional energy management centers, distributed energy suppliers, regional load regulation centers, etc., and the strategy space is set, that is, the optional strategies of each subject in energy scheduling, such as local load balancing, cross-regional energy allocation, energy storage optimization, etc. The profit matrix calculation method is used to calculate the benefits of different game subjects under different scheduling strategies, such as energy costs, load stability, scheduling delay, etc., and then the Nash equilibrium method in game theory is used to solve the strategy equilibrium of the game subjects and calculate the initial game solution set. Each game subject adjusts its strategy in multiple rounds of games in order to maximize its own long-term benefits, and dynamically adjusts the game strategy according to environmental changes and demand changes. Finally, the preliminary game equilibrium solution set of each subject is obtained, which provides a basis for subsequent strategy optimization.

[0137] Step S32: Acquire historical regional energy dispatch data, and perform energy dispatch mode identification on the historical regional energy dispatch data to obtain historical energy dispatch mode data;

[0138] In this embodiment, historical regional energy dispatch data, including regional electricity load data, renewable energy output data, energy storage charging and discharging data, cross-regional energy flow data, etc., are obtained through the regional energy management center, and analyzed using a time series pattern mining algorithm. First, the historical data is denoised using wavelet transform, and the autoregressive integrated moving average (ARIMA) model is used for time series modeling to identify long-term trends and seasonal variation characteristics. Then, a clustering algorithm is used to perform pattern recognition on the dispatch data to identify common energy dispatch patterns, such as peak load periods and valley dispatch periods. Based on these historical dispatch patterns, a pattern library is constructed, and the patterns are further classified so that each pattern can correspond to a certain dispatch strategy. The historical pattern data will provide important input for the game model, helping to optimize the strategy selection of the game subject.

[0139] Step S33: Dynamically adjust the initial game equilibrium solution set, introduce a preset non-cooperative game mechanism, evaluate the optimal response strategies of different game players under energy scheduling constraints, and perform reinforcement learning optimization based on historical energy scheduling pattern data to generate a reinforcement game equilibrium solution set;

[0140] In this embodiment, based on the initial game equilibrium solution set, a non-cooperative game mechanism is introduced. It is assumed that each regional entity has a certain degree of competitiveness in energy allocation, and its optimal response strategy benefits under different scheduling constraints are calculated, including energy costs, energy storage utilization, regional energy supply stability, etc. Reinforcement learning algorithms are used, especially the deep Q network (DQN) model based on Q-learning, to optimize the decision-making strategies of the game entities through reinforcement learning. For each game entity, the performance of the current strategy is evaluated based on historical regional energy scheduling pattern data and real-time scheduling data, and updated according to the learned strategy. In reinforcement learning, by setting reward mechanisms (such as reducing energy scheduling costs and improving stability) and punishment mechanisms (such as load imbalance and over-reliance on external energy), the game entities are encouraged to adjust their strategies and gradually obtain a reinforced game equilibrium solution set.

[0141] Step S34: Calculate the cross-regional energy mutual assistance optimization parameters based on the reinforcement game equilibrium solution set, set the mutual assistance triggering conditions, calculate the cross-regional energy mutual assistance benefits, and generate the cross-regional energy mutual assistance optimization plan;

[0142] In this embodiment, the cross-regional energy mutual assistance optimization parameters are calculated based on the reinforcement game equilibrium solution set. First, the mutual assistance trigger conditions are set. For example, when the capacity of the energy storage equipment in a certain area is lower than the preset 5% threshold or the load fluctuation in the area exceeds the set 10% range, the cross-regional energy mutual assistance mechanism is enabled. Then, considering the heterogeneity of energy supply and demand in different regions, a multi-objective optimization method (such as the Pareto frontier method) is used to optimize the cross-regional energy scheduling scheme. By setting the objective function, the optimal solution for cross-regional energy mutual assistance is determined by comprehensively considering multiple dimensions such as the stability of energy supply, the balance of load demand, the cost of energy scheduling, and the impact on environmental protection. Genetic algorithm (GA) or particle swarm optimization algorithm (PSO) is used to search for the optimal solution and optimize the energy flow between regions to maximize the benefits of the overall system.

[0143] Step S35: Conduct credible verification on the cross-regional energy mutual assistance optimization plan, calculate the mutual assistance contribution of each region, and perform adaptive credit evaluation based on the non-cooperative game mechanism to optimize the energy mutual assistance plan of each region and generate a credible scheduling verification chain.

[0144] In this embodiment, a credible verification is performed on the cross-regional energy mutual aid optimization plan, and the mutual aid contribution of each region is calculated. That is, in the historical mutual aid records, the number of times each region has provided energy support, the amount of support provided, and the degree of positive impact on the overall scheduling. Based on the non-cooperative game mechanism, an adaptive credit evaluation system is set up for each region, giving higher credit scores to regions that provide long-term mutual aid support, while restricting regions that frequently consume mutual aid resources but make insufficient contributions. At the same time, a blockchain smart contract is used to build a trusted scheduling verification chain, and the mutual aid behavior, contribution, credit score and other data of each region are uploaded to the chain to ensure the transparency and traceability of the scheduling plan, optimize the energy mutual aid plan of each region, and finally generate a trusted scheduling verification chain.

[0145] Optionally, the step S4 of constructing a multi-physics field coupling scheduling unit is as follows:

[0146] Extract meteorological features, power grid operation features, and user behavior features from the trusted dispatch verification chain, and structure and layer them to generate a multi-physics field raw data set.

[0147] In this embodiment, to extract meteorological characteristics, grid operation characteristics, and user behavior characteristics from the trusted scheduling verification chain, the verification chain must first be decrypted and parsed. For meteorological characteristics, the chain's timestamp data and encrypted meteorological data are analyzed to extract meteorological elements such as temperature, humidity, and wind speed. For grid operation characteristics, relevant characteristics are extracted by parsing grid operational data, including operational status information such as load, frequency, and voltage. Finally, user behavior characteristics can be extracted by analyzing energy consumption patterns to extract user electricity usage behavior data, such as peak and valley periods and load demand. This data extraction process can be accomplished using the fingerprint data stored in the chain, combined with encryption and decryption techniques, ensuring data security while obtaining the required characteristic information. The different parameters contained in the meteorological data (such as temperature, humidity, and wind speed) are considered independent physical layers and are therefore listed as a separate layer. Each meteorological characteristic data is grouped by timestamp, and meteorological data at different time points are used as the time series input for that layer. Because meteorological characteristic data spans a large time span and may experience significant seasonal fluctuations, these temporal factors must be taken into account when stratifying. Dynamic data such as voltage, current, and power factor, including grid operation characteristics, are also stratified by timestamp. To capture the changing trends of the grid under different loads and operating conditions, each grid characteristic dataset is processed as independent time series data and normalized to ensure that the data values are within the same range. Furthermore, because the real-time status of grid load and grid operation may be affected by the external environment, the structural stratification of grid characteristics should focus on their interaction with meteorological characteristics and user behavior characteristics. User behavior characteristics include temporal patterns of electricity demand, frequency of electricity use, and load fluctuations, which reflect user electricity usage habits. User behavior data is stratified by different periods, such as daily, weekly, and seasonal, and normalized to ensure data consistency. During stratification, special consideration is given to the relationship between electricity demand during different time periods and meteorological changes, as meteorological changes can significantly affect user electricity demand. After these data are structurally stratified, a multi-physics raw dataset is ultimately formed. This dataset integrates various layers of meteorological, grid, and user behavior data and uniformly arranges them by timestamp, ensuring that the time series of each characteristic data type is synchronized with the other data types.

[0148] The standardized multimodal energy dataset is time-aligned, the time step is set to 5 minutes and interpolation is performed to obtain a time-aligned multi-physics field dataset;

[0149] In this embodiment, after obtaining the standardized multimodal energy data set, a time alignment process is performed. To ensure synchronization between different data sources, the time step is set to 5 minutes, and the missing timestamp data is supplemented by a linear interpolation method. In the time alignment process, the data of each physical field are first timestamp aligned, and it is ensured that all data are consistent at the same time step. For example, if meteorological data is collected once an hour and grid operation data is collected every 10 minutes, then through the interpolation method, all data will be aligned at a time step of every 5 minutes. The time-aligned multi-physics field data set processed in this way ensures the synchronization and integrity of the multidimensional data, and provides standardized input data for subsequent modeling and prediction.

[0150] A multi-layer non-Euclidean graph model was constructed based on the time-aligned multi-physics dataset. A topological mapping was established based on the time-aligned multi-physics dataset. The twin constraint parameter was set to 0.7. A preset digital twin verification mechanism was embedded in the topological mapping process. Graph neural network feature extraction was performed to generate an initial multi-physics coupling graph structure model.

[0151] In this embodiment, a multi-layer non-Euclidean graph model is constructed based on a time-aligned multi-physics field data set. A topological mapping graph is constructed for each type of feature (meteorological, power grid operation status, user behavior), and the twin constraint parameter is set to 0.7 to ensure that the similarity between the graphs satisfies the twin constraint. In the process of establishing the graph, a graph neural network (GNN) is first used to extract features from different feature data, and the spatial and temporal characteristics of each feature are mapped to the graph structure. The embedded digital twin verification mechanism ensures the association and consistency between each data level. For example, various features are weightedly fused through the graph convolution layer, and the consistency of the data in the topological structure is verified. This process can automatically identify the correlation between different physical fields, construct an initial multi-physics field coupling graph structure model, and generate a graph structure that can be further optimized.

[0152] Spectral clustering analysis was performed on the initial multi-physics coupling diagram structure model. The first 10 eigenvectors were selected for Laplace eigendecomposition. The feature truncation threshold was set to 0.85, and high-dimensional feature dimensionality reduction was performed to generate a spectrally smoothed multi-physics coupling diagram.

[0153] In this embodiment, spectral clustering analysis is performed on the initial multi-physics field coupling graph structure model to extract the most representative eigenvectors. The first 10 eigenvectors are selected for Laplace eigendecomposition, and the feature cutoff threshold is set to 0.85 to perform dimensionality reduction processing on high-dimensional data. This process mainly reduces the data dimension by eigendecomposition of the graph Laplace matrix while retaining the most representative feature information. In this way, high-dimensional data can be mapped to a low-dimensional space, thereby reducing computational complexity and improving the interpretability of the model. The graph structure after dimensionality reduction can more clearly show the interactions and associations between different physical fields, providing a more concise representation for subsequent data analysis and modeling.

[0154] Perform a standardized transformation on the spectral smoothed multi-physics field coupling diagram, set the normalization interval to [0,1], and perform time series trend correction to generate a standardized multi-physics field scheduling unit;

[0155] In this embodiment, a normalization transformation is performed on the spectrally smoothed multi-physics field coupling diagram. First, the numerical values of the diagram are normalized to the interval [0,1] to ensure that the numerical values between different physical fields are compared on the same scale. Then, a time series trend correction is performed to eliminate the impact of long-term trends on the data and ensure that the fluctuations in the time series data reflect actual changes. The normalization transformation helps to eliminate the dimensional differences between different data sources and ensure that the data is processed on the same scale. In this process, a sliding window method is used to correct the trend and avoid data offsets caused by long-term changing trends, making the subsequent model's prediction of the time series more accurate.

[0156] The standardized multi-physics field scheduling unit is corrected by Bayesian optimal difference data, and the confidence interval is set to 0.95 for confidence interpolation completion. The historical energy scheduling mode data is combined to perform distributed time series prediction to generate a multi-physics field coupling scheduling unit.

[0157] In this embodiment, the standardized multi-physics field scheduling unit is corrected by Bayesian optimal difference data, and the confidence interval is set to 0.95 for confidence interpolation completion. The Bayesian interpolation method predicts missing data by using the probability distribution of existing data, and determines the reliability of the interpolation by setting a confidence interval. In practical applications, the historical energy scheduling mode data is first analyzed to find out the laws and characteristics therein, such as load peak period and grid stability factors. Then, data correction is performed based on the Bayesian method, and future data is completed in combination with the distributed time series prediction method. Based on the combination of Bayesian interpolation completion and distributed time series prediction, the scheduling unit generated by the joint optimization model not only contains the data of a single physical field, but also integrates the mutual influence and synergy between multiple physical fields. These multi-physics field data are coupled into a comprehensive scheduling unit to form a multi-physics field coupling scheduling unit. It covers the interaction of multiple factors such as meteorology, energy load, user behavior, and grid operation, and can provide reliable data support for further regional energy scheduling.

[0158] Optionally, the compilation of the dynamic scheduling instruction set in step S4 is specifically:

[0159] Based on the energy scheduling constraints, the scheduling parameter boundaries are set for the multi-physics field coupling scheduling unit to generate an initial scheduling parameter set;

[0160] In this embodiment, the boundaries of key parameters in the multi-physics coupling scheduling unit are set based on energy scheduling constraints. Scheduling constraints for the multi-physics coupling scheduling unit are set. Specifically, the regional collaborative scheduling vector is composed of energy production and consumption data for each region, load demand forecasts, and inter-regional energy interoperability relationships. Based on this data, scheduling parameter boundaries are set, such as each region's maximum power generation capacity, maximum load demand, load balancing constraints, and cross-regional energy transmission restrictions. These constraints ensure that the generated scheduling solution is feasible and stable during subsequent scheduling optimization and meets the requirements of inter-regional energy interoperability and scheduling coordination. Based on these constraints, an initial set of scheduling parameters is generated, providing a reasonable solution space for the optimization algorithm. For example, maximum and minimum limits for grid load, constraints on user behavior models (such as rate limits on user demand changes), and the impact range of meteorological characteristics (such as wind speed and light intensity) on energy output are set. Based on these constraints, a set of scheduling parameters is generated that includes the characteristics of each physical field, where each parameter meets the established upper and lower limits. This set serves as the initial parameters in the optimization process, ensuring that the scheduling solution is feasible in the actual environment and meets the various constraints. Operational parameters may include the fluctuation range of meteorological data, load growth rate, etc.

[0161] Dynamic scheduling optimization modeling is performed on the initial scheduling parameter set, the scheduling optimization objective function is set, and the distributed particle swarm optimization algorithm is used for strategy search. The particle swarm size is set to 50 and the iteration step length is set to 20s to generate the global optimal scheduling parameter set.

[0162] In this embodiment, an improved distributed particle swarm optimization (PSO) algorithm is used to optimize the initial scheduling parameter set. First, the scheduling optimization objective function is defined, including but not limited to cost minimization, energy efficiency maximization, and environmental impact minimization. The PSO algorithm is used to perform particle search in the defined objective function space, and each particle represents a potential scheduling strategy. The particle swarm size is set to 50, and the speed and position of the particles are continuously updated according to the fitness function. The step size in the iterative process is set to 20s to ensure that the optimization process can quickly obtain the optimal solution when time is tight. In each round of iteration, the search process is guided by evaluating the fitness value of the particle (that is, the value of the objective function) until it converges to the global optimal scheduling parameter set.

[0163] Compile the scheduling instruction set based on the global optimal scheduling parameter set, optimize the scheduling execution strategy, set the reward feedback factor to 0.8, and generate an enhanced scheduling instruction set;

[0164] In this embodiment, the scheduling instruction set is compiled based on the global optimal scheduling parameter set obtained from the particle swarm optimization algorithm. Through the optimized scheduling parameters, specific execution strategies are formulated, such as scheduling timing, resource allocation, priority sorting, etc. At the same time, in order to improve the adaptability and response speed of the scheduling system, the reward feedback factor is set to 0.8 to guide the real-time adjustment of the scheduling strategy. The role of the reward factor is to give the scheduling system corresponding rewards when the system achieves the preset goals (such as improved execution efficiency, maximization of energy efficiency, etc.), thereby promoting it to give priority to effective strategies in the next task. The generated enhanced scheduling instruction set will serve as the basis for subsequent scheduling execution.

[0165] Parallel simulations were conducted on the enhanced dispatch instruction set. The dispatch execution efficiency was evaluated under different combinations of meteorological characteristics, grid operating status characteristics, and user behavior characteristics. The execution efficiency threshold was set at 85%, and the task throughput was calculated. The instruction set execution path was optimized to generate an optimized dispatch instruction set.

[0166] In this embodiment, the enhanced scheduling instruction set will be simulated and verified in parallel in multiple virtual environments to evaluate its execution efficiency. During the simulation process, various combination scenarios such as different meteorological characteristics (such as wind speed, light intensity), power grid operation status (such as load fluctuations, power grid stability) and user behavior characteristics (such as changes in electricity demand) will be simulated. The execution efficiency threshold is set to 85%, and the execution effect of the instruction set is determined by evaluating indicators such as task throughput, response time and stability of scheduling execution. If the execution efficiency is lower than the preset threshold, the instruction set needs to be optimized and adjusted. The results of the simulation verification help adjust the scheduling execution path so that it can adapt to different operating environments and demand changes. The optimized instruction set can ensure that the system operates efficiently and stably in various environments.

[0167] Instruction set compatibility mapping is performed based on the optimized scheduling instruction set, and a cross-platform compatibility threshold is set to 90% to generate a cross-platform compatible dynamic scheduling instruction set.

[0168] In this embodiment, the optimized scheduling instruction set is mapped for cross-platform compatibility to ensure that it can run on different hardware platforms or computing environments. The process of instruction set compatibility mapping includes checking the compatibility between different hardware platforms (such as cloud platforms, edge computing nodes, etc.) and solving possible problems such as interface and protocol differences. The cross-platform compatibility threshold is set at 90%, and mapping technology is used to ensure the seamless connection and execution of instruction sets between different platforms. This operation usually includes adapting the instruction set so that it can be executed in a unified manner in different devices and systems, and finally generating a cross-platform compatible dynamic scheduling instruction set. This instruction set can be used for collaborative work between different systems to ensure the real-time and accuracy of energy scheduling.

[0169] Optionally, the scheduling deviation tracing in step S5 is specifically as follows:

[0170] Analyze the global scheduling evidence chain, extract the planned scheduling value, actual execution value and adjustment instructions within 5 minutes of each scheduling cycle, calculate the scheduling deviation vector, and generate a scheduling execution deviation information database;

[0171] In this embodiment, by parsing the global scheduling evidence chain, key data within the scheduling cycle in the chain is extracted, specifically the planned scheduling value, actual execution value, and adjustment instructions for each 5-minute time period. The planned scheduling value refers to the predetermined scheduling plan generated according to the pre-set scheduling strategy, the actual execution value is the actual system response data, and the adjustment instruction represents the scheduling adjustment due to environmental or system changes. By comparing the planned scheduling value with the actual execution value, the scheduling deviation vector is calculated to reflect the accuracy of the scheduling. The deviation vector is stored in the scheduling execution deviation information library, which provides an important historical scheduling execution data foundation for subsequent scheduling optimization, ensuring that future scheduling can be more accurately predicted and adjusted based on this data.

[0172] Based on the multi-physics field coupling scheduling unit and the scheduling execution deviation information library, a dynamic time warping causal analysis of scheduling deviation is performed. The significance level is set to 0.05 to calculate the influence weight of each factor in the multi-physics field coupling scheduling unit on the scheduling deviation, thereby scheduling the impact factor matrix;

[0173] In this embodiment, the dynamic time warping (DTW) algorithm is used to analyze the relationship between the multi-physics field coupling scheduling unit and the scheduling execution deviation information library to perform a causal analysis of the scheduling deviation. By comparing the time series differences between the changing trends of each factor and the scheduling deviation, it is analyzed which factors (such as meteorological characteristics, power grid operation status, user behavior, etc.) have a significant impact on the scheduling deviation. The DTW algorithm can identify the similarities between the deviation and feature data at different time points, thereby revealing the causal relationship. The significance level is set to 0.05, and a statistical significance test is performed to ensure the reliability of the analysis results. By calculating the influence weight of each factor on the scheduling deviation in the multi-physics field coupling scheduling unit, a statistical test (such as a chi-square test) is used to determine whether the influence of each factor on the scheduling deviation is significant. The analysis results will generate a "scheduling impact factor matrix", which contains the influence weight of each factor on the scheduling deviation, as the basis for subsequent optimization of the scheduling plan. The calculation of the weight can be based on a weighted regression model or principal component analysis to ensure that important factors are fully considered during the optimization process.

[0174] The scheduling state deviation transition probability is calculated based on the scheduling impact factor matrix, and the time window is set to 30 minutes. The deviation propagation path is constructed using a time-series sliding method. The influence weights between each state are calculated based on the deviation propagation path, the deviation causal relationship is inferred, and a deviation spatiotemporal traceability model is generated.

[0175] In this embodiment, the transition probability of scheduling state deviations is further analyzed based on the scheduling impact factor matrix. By setting the time window to 30 minutes, a time-series sliding method is used to construct the deviation propagation path between each state. Assuming that each scheduling state can be regarded as a node, the change in deviation (such as the difference between the scheduled execution value and the planned value) will transfer from one state to another, forming a deviation propagation path. The deviation changes within each window are analyzed, and the transition probability between each state is calculated. This process uses a Markov chain model to describe state transitions and estimates the transition probability based on historical data. By calculating the deviation propagation path, it is possible to understand how the deviation is transmitted between different scheduling states, thereby providing a basis for subsequent deviation optimization and correction. This process helps the system understand and predict scheduling state changes under different environmental changes, ensuring that these potential deviation propagation paths are considered when making scheduling decisions. Using the previously calculated deviation propagation path, the influence weights between the scheduling states are further analyzed. Graph theory algorithms (such as the PageRank algorithm or the maximum flow algorithm) are used to evaluate the degree of mutual influence between states and calculate the influence during the state transition process. This process reveals the dominant role of certain states in the overall deviation of the dispatching system and infers the causal chain of deviations, namely, which state changes directly lead to increases or decreases in deviations. Based on these causal relationships, a spatiotemporal deviation traceability model is constructed to describe the spatiotemporal causal connections between various dispatching states. This model can predict the deviations that may occur in the system under a specific state and trace them back through time and space, helping dispatchers quickly identify the root cause of the problem.

[0176] Based on the deviation spatiotemporal traceability model, the optimal response strategy of each regional dispatching entity is set, and dynamic optimization is performed to screen dispatching instructions with a robustness score higher than 0.85 to obtain a highly robust dispatching instruction set.

[0177] In this embodiment, the optimal response strategy of each regional dispatching entity is set according to the results of the deviation spatiotemporal tracing model. These strategies are based on the results inferred from the deviation propagation path and causal relationship, and are dynamically optimized for factors such as energy demand, load fluctuations, and meteorological changes in different regions. This process is achieved by establishing a dynamic optimization model, in which the robustness score (such as obtained by regression analysis) is used to evaluate the reliability of the dispatching instructions. For dispatching instructions with a robustness score lower than 0.85, adjustments or elimination are made to ensure that the final selected dispatching instruction set has a higher ability to cope with deviations. The algorithms involved in the optimization process include particle swarm optimization, genetic algorithms, etc., which can ensure that the optimization results are globally optimal and can operate stably in the face of changes.

[0178] Based on the highly robust scheduling instruction set, the energy scheduling optimization objective function is defined, and the energy scheduling optimization parameters are calculated to obtain the optimized energy scheduling parameter matrix;

[0179] In this embodiment, an energy scheduling optimization objective function is defined based on an optimized, highly robust scheduling instruction set. The objective function covers multiple scheduling objectives, such as maximizing energy efficiency, minimizing energy costs, and ensuring system stability. A set of optimized energy scheduling parameters is calculated using scheduling optimization algorithms (such as linear programming and nonlinear programming) to minimize the scheduling objective function. During the calculation process, an optimized energy scheduling parameter matrix is obtained based on constraints such as the energy supply and demand situation in each region and grid stability, ensuring that the energy scheduling plan for each region can maximize efficiency and meet the system's stability requirements during execution.

[0180] The energy supply and demand matching degree of each region is calculated based on the optimized energy scheduling parameter matrix, and the energy distribution equilibrium optimization is performed based on the reinforcement game equilibrium solution set. The energy allocation decision is generated and uploaded to the regional energy scheduling platform to execute the energy allocation task.

[0181] In this embodiment, the energy supply and demand matching degree of each region is calculated based on the optimized energy scheduling parameter matrix. This step determines the energy matching degree of each region by evaluating the gap between the actual energy demand and the available resources. Then, energy distribution balance optimization is performed in combination with the reinforcement game equilibrium solution set. The theoretical model of reinforcement game is used to simulate the energy distribution game between regions, and the optimal energy distribution decision is calculated through the game equilibrium solution set. This decision ensures the efficiency of energy transmission and scheduling between regions and optimizes the balance of the overall energy network. Ultimately, the generated energy distribution decision will be uploaded to the regional energy scheduling platform to execute the allocation task and achieve the scheduling goals.

[0182] Optionally, this specification further provides a multimodal data fusion system based on energy scheduling, which is used to execute the multimodal data fusion method based on energy scheduling as described above. The multimodal data fusion system based on energy scheduling includes:

[0183] The abnormal fluctuation feature annotation module is used to obtain multimodal data and construct a dynamic spatiotemporal adjacency matrix based on the multimodal data; the abnormal fluctuation features are extracted according to the dynamic spatiotemporal adjacency matrix to obtain a spatiotemporal fusion abnormal feature annotation set;

[0184] The scheduling decision deduction module is used to perform distributed policy gradient aggregation on the spatiotemporal fusion anomaly feature annotation set to obtain the regional collaborative scheduling vector; based on the regional collaborative scheduling vector, it deduces the multi-objective optimization path to obtain the dynamic scheduling decision map;

[0185] The game equilibrium calculation module is used to perform edge node game equilibrium calculations based on the dynamic scheduling decision graph, and perform cross-regional energy mutual assistance verification to obtain a trusted scheduling verification chain;

[0186] The dynamic scheduling instruction compilation module is used to perform digital twin constraint optimization on the trusted scheduling verification chain, build a multi-physics field coupling scheduling unit, and obtain a closed-loop scheduling digital twin. Based on the closed-loop scheduling digital twin, the dynamic scheduling instruction set is compiled to obtain a disturbance-resistant energy scheduling strategy library.

[0187] The scheduling deviation tracing module is used to obtain real-time energy supply and demand data, and dynamically evolve the anti-disturbance strategy for the real-time energy supply and demand data based on the anti-disturbance energy scheduling strategy library to generate a global scheduling optimization evidence chain; based on the global scheduling evidence chain, scheduling deviation tracing is implemented, and highly robust scheduling instructions are screened to obtain energy allocation decisions.

[0188] Optionally, the present specification also provides a computer-readable storage medium storing a computer program, which, when executed, implements the multimodal data fusion method based on energy scheduling as described above.

Claims

1. A multimodal data fusion method based on energy scheduling, characterized in that: The following steps are involved: Step S1: Acquire multimodal data and construct a dynamic spatiotemporal adjacency matrix based on the multimodal data; extract abnormal fluctuation features based on the dynamic spatiotemporal adjacency matrix to obtain a spatiotemporal fusion abnormal feature annotation set; Step S2: Perform distributed policy gradient aggregation on the spatiotemporal fusion anomaly feature annotation set to obtain a regional collaborative scheduling vector; deduce a multi-objective optimization path based on the regional collaborative scheduling vector to obtain a dynamic scheduling decision map; Step S3: Perform edge node game equilibrium calculation based on the dynamic scheduling decision graph, and perform cross-regional energy mutual assistance verification to obtain a trusted scheduling verification chain; Step S4: Perform digital twin constraint optimization on the trusted scheduling verification chain, construct a multi-physics field coupling scheduling unit, and obtain a closed-loop scheduling digital twin; compile a dynamic scheduling instruction set based on the closed-loop scheduling digital twin to obtain an anti-disturbance energy scheduling strategy library; Step S5: Acquire real-time energy supply and demand data, and dynamically evolve the anti-disturbance strategy for the real-time energy supply and demand data based on the anti-disturbance energy scheduling strategy library to generate a global scheduling optimization evidence chain; implement scheduling deviation tracing based on the global scheduling evidence chain, and screen highly robust scheduling instructions to obtain energy allocation decisions.

2. The multimodal data fusion method based on energy scheduling according to claim 1 is characterized in that: Step S1 is specifically as follows: Step S11: Acquire multimodal data, where the multimodal data includes meteorological data, power grid operation status data, and user behavior data; standardize the format of the multimodal data, and perform missing value interpolation and denoising to generate a standardized multimodal energy data set; Step S12: constructing a dynamic spatiotemporal adjacency matrix based on the standardized multimodal energy dataset, and performing global attention modeling based on the spatial topological features and time series features in the dynamic spatiotemporal adjacency matrix to generate a spatiotemporal correlation feature map; Step S13: performing abnormal pattern recognition on the spatiotemporal correlation feature graph, extracting abnormal fluctuation patterns, screening abnormal regions with high impact factors, and generating a spatiotemporal fusion abnormal feature tensor; Step S14: Project the spatiotemporal fusion abnormal feature tensor into a feature space, calculate the abnormal distribution skewness, and obtain the abnormal feature projection matrix; Step S15: label the abnormal pattern according to the abnormal feature projection matrix, generate abnormal classification labels, and calculate the abnormal spatiotemporal impact weights, thereby obtaining a spatiotemporal fusion abnormal feature annotation set.

3. The multimodal data fusion method based on energy scheduling according to claim 2 is characterized in that: Step S13 is specifically as follows: Step S131: performing time series normalization on the spatiotemporal correlation feature map, removing periodic features and overall offsets, and obtaining a normalized spatiotemporal correlation feature map; Step S132: performing abnormal pattern recognition on the standardized spatiotemporal correlation feature graph, calculating the local abnormality score of the standardized spatiotemporal correlation feature graph, and performing abnormal pattern classification to generate a preliminary screening abnormal pattern set; Step S133: Based on the initial screening abnormal pattern set, calculate the probability distribution of the abnormal pattern in the time and space dimensions, and screen the abnormal regions with high impact factors to generate a matrix of abnormal regions with high impact factors; Step S134: performing spatiotemporal feature expansion on the high impact factor abnormal region matrix, constructing a spatiotemporal correlation map of abnormal regions, and calculating the time alignment between abnormal regions in the spatiotemporal correlation map of abnormal regions, thereby generating a spatiotemporal abnormal correlation map; Step S135: Perform spatiotemporal feature fusion according to the spatiotemporal anomaly association graph, perform cross-time step feature interaction on the spatiotemporal features, and select the optimal anomaly classification hyperparameters to generate a spatiotemporal fusion anomaly feature tensor.

4. The multimodal data fusion method based on energy scheduling according to claim 1 is characterized in that: Step S2 is specifically as follows: Step S21: performing feature space dimensionality reduction on the spatiotemporal fusion abnormal feature tensor, extracting abnormal features, and calculating the Euclidean distance and cosine similarity between abnormal features to generate an abnormal pattern similarity matrix; Step S22: constructing an abnormal pattern spatiotemporal network based on the abnormal pattern similarity matrix, and calculating the global impact factor of the abnormal area in the abnormal pattern spatiotemporal network to generate an abnormal pattern impact vector; Step S23: Based on the abnormal mode impact vector, the regional collaborative scheduling strategy is optimized using a preset distributed particle swarm optimization algorithm, a particle search space is set, and the agent strategy is adjusted based on the trade-off between local optimality and global optimality, thereby obtaining the regional collaborative scheduling vector; Step S24: setting energy scheduling constraints based on the regional collaborative scheduling vector, calculating the Pareto optimal solution set, and generating a multi-objective optimization path set; Step S25: Verify the temporal feasibility of the multi-objective optimization path set, evaluate the adaptability of the adversarial training environment path, and optimize the scheduling decision based on the preset scheduling objective function group, thereby generating a dynamic scheduling decision map, where the scheduling objective function group is specifically: F(X)=α·F c (X)+β·F s (X)+γ·F ef (X)+δ·F en (X); Among them, F(X) is the scheduling objective function, F c (X) is the cost objective function, F s (X) is the stability objective function, F ef (X) is the efficiency objective function, F en (X) is the environmental impact objective function, X represents the scheduling decision vector, α is the cost objective weight factor, β is the stability weight factor, γ is the efficiency weight factor, δ is the environmental impact weight factor, T is the total monitoring time, t is the time index, N is the number of all regional nodes, i is the regional node index, C i (t) is the energy cost of the node in the i-th region at time t, P i (t) is the energy allocation ratio of the node in the i-th region at time t, E i (t) is the energy storage state factor of the node in the i-th region at time t, is the preset load of the node in the i-th region at time t, P max (i) is the maximum energy available value of the node in the i-th region, ζ i is the carbon emission factor of the i-th regional node.

5. The multimodal data fusion method based on energy scheduling according to claim 1 is characterized in that: Step S3 is specifically as follows: Step S31: constructing an edge node game model based on the dynamic scheduling decision graph, setting the game subject and strategy space, and calculating the strategy payoff matrix of different game subjects to generate an initial game equilibrium solution set; Step S32: Acquire historical regional energy dispatch data, and perform energy dispatch mode identification on the historical regional energy dispatch data to obtain historical energy dispatch mode data; Step S33: Dynamically adjust the initial game equilibrium solution set, introduce a preset non-cooperative game mechanism, evaluate the optimal response strategies of different game players under energy scheduling constraints, and perform reinforcement learning optimization based on historical energy scheduling pattern data to generate a reinforcement game equilibrium solution set; Step S34: Calculate the cross-regional energy mutual assistance optimization parameters based on the reinforcement game equilibrium solution set, set the mutual assistance triggering conditions, calculate the cross-regional energy mutual assistance benefits, and generate the cross-regional energy mutual assistance optimization plan; Step S35: Conduct credible verification on the cross-regional energy mutual assistance optimization plan, calculate the mutual assistance contribution of each region, and perform adaptive credit evaluation based on the non-cooperative game mechanism to optimize the energy mutual assistance plan of each region and generate a credible scheduling verification chain.

6. The multimodal data fusion method based on energy scheduling according to claim 1 is characterized in that: The construction of the multi-physics field coupling scheduling unit described in step S4 is specifically as follows: Extract meteorological features, power grid operation features, and user behavior features from the trusted dispatch verification chain, and structure and layer them to generate a multi-physics field raw data set. The standardized multimodal energy dataset is time-aligned, the time step is set to 5 minutes and interpolation is performed to obtain a time-aligned multi-physics field dataset; A multi-layer non-Euclidean graph model was constructed based on the time-aligned multi-physics dataset. A topological mapping was established based on the time-aligned multi-physics dataset. The twin constraint parameter was set to 0.

7. A preset digital twin verification mechanism was embedded in the topological mapping process. Graph neural network feature extraction was performed to generate an initial multi-physics coupling graph structure model. Spectral clustering analysis was performed on the initial multi-physics coupling diagram structure model. The first 10 eigenvectors were selected for Laplace eigendecomposition. The feature truncation threshold was set to 0.85, and high-dimensional feature dimensionality reduction was performed to generate a spectrally smoothed multi-physics coupling diagram. Perform a standardized transformation on the spectral smoothed multi-physics field coupling diagram, set the normalization interval to [0,1], and perform time series trend correction to generate a standardized multi-physics field scheduling unit; The standardized multi-physics field scheduling unit is corrected by Bayesian optimal difference data, and the confidence interval is set to 0.95 for confidence interpolation completion. The historical energy scheduling mode data is combined to perform distributed time series prediction to generate a multi-physics field coupling scheduling unit.

7. The multimodal data fusion method based on energy scheduling according to claim 1 is characterized in that: The compilation dynamic scheduling instruction set described in step S4 is specifically: Based on the energy scheduling constraints, the scheduling parameter boundaries are set for the multi-physics field coupling scheduling unit to generate an initial scheduling parameter set; Dynamic scheduling optimization modeling is performed on the initial scheduling parameter set, the scheduling optimization objective function is set, and the distributed particle swarm optimization algorithm is used for strategy search. The particle swarm size is set to 50 and the iteration step length is set to 20s to generate the global optimal scheduling parameter set. Compile the scheduling instruction set based on the global optimal scheduling parameter set, optimize the scheduling execution strategy, set the reward feedback factor to 0.8, and generate an enhanced scheduling instruction set; Parallel simulations were conducted on the enhanced dispatch instruction set. The dispatch execution efficiency was evaluated under different combinations of meteorological characteristics, grid operating status characteristics, and user behavior characteristics. The execution efficiency threshold was set at 85%, and the task throughput was calculated. The instruction set execution path was optimized to generate an optimized dispatch instruction set. Instruction set compatibility mapping is performed based on the optimized scheduling instruction set, and a cross-platform compatibility threshold is set to 90% to generate a cross-platform compatible dynamic scheduling instruction set.

8. The multimodal data fusion method based on energy scheduling according to claim 1 is characterized in that: The scheduling deviation tracing in step S5 is specifically as follows: Analyze the global scheduling evidence chain, extract the planned scheduling value, actual execution value and adjustment instructions within 5 minutes of each scheduling cycle, calculate the scheduling deviation vector, and generate a scheduling execution deviation information database; Based on the multi-physics field coupling scheduling unit and the scheduling execution deviation information library, a dynamic time warping causal analysis of scheduling deviation is performed. The significance level is set to 0.05 to calculate the influence weight of each factor in the multi-physics field coupling scheduling unit on the scheduling deviation, thereby scheduling the impact factor matrix; The scheduling state deviation transition probability is calculated based on the scheduling impact factor matrix, and the time window is set to 30 minutes. The deviation propagation path is constructed using a time-series sliding method. The influence weights between each state are calculated based on the deviation propagation path, the deviation causal relationship is inferred, and a deviation spatiotemporal traceability model is generated. Based on the deviation spatiotemporal traceability model, the optimal response strategy of each regional dispatching entity is set, and dynamic optimization is performed to screen dispatching instructions with a robustness score higher than 0.85 to obtain a highly robust dispatching instruction set. Based on the highly robust scheduling instruction set, the energy scheduling optimization objective function is defined, and the energy scheduling optimization parameters are calculated to obtain the optimized energy scheduling parameter matrix; The energy supply and demand matching degree of each region is calculated based on the optimized energy scheduling parameter matrix, and the energy distribution equilibrium optimization is performed based on the reinforcement game equilibrium solution set. The energy allocation decision is generated and uploaded to the regional energy scheduling platform to execute the energy allocation task.

9. A multimodal data fusion system based on energy scheduling, characterized in that: For executing the multimodal data fusion method based on energy scheduling according to claim 1, the multimodal data fusion system based on energy scheduling comprises: The abnormal fluctuation feature annotation module is used to obtain multimodal data and construct a dynamic spatiotemporal adjacency matrix based on the multimodal data; the abnormal fluctuation features are extracted according to the dynamic spatiotemporal adjacency matrix to obtain a spatiotemporal fusion abnormal feature annotation set; The scheduling decision deduction module is used to perform distributed policy gradient aggregation on the spatiotemporal fusion anomaly feature annotation set to obtain the regional collaborative scheduling vector; based on the regional collaborative scheduling vector, it deduces the multi-objective optimization path to obtain the dynamic scheduling decision map; The game equilibrium calculation module is used to perform edge node game equilibrium calculations based on the dynamic scheduling decision graph, and perform cross-regional energy mutual assistance verification to obtain a trusted scheduling verification chain; The dynamic scheduling instruction compilation module is used to perform digital twin constraint optimization on the trusted scheduling verification chain, build a multi-physics field coupling scheduling unit, and obtain a closed-loop scheduling digital twin. Based on the closed-loop scheduling digital twin, the dynamic scheduling instruction set is compiled to obtain a disturbance-resistant energy scheduling strategy library. The scheduling deviation tracing module is used to obtain real-time energy supply and demand data, and dynamically evolve the anti-disturbance strategy for the real-time energy supply and demand data based on the anti-disturbance energy scheduling strategy library to generate a global scheduling optimization evidence chain; based on the global scheduling evidence chain, scheduling deviation tracing is implemented, and highly robust scheduling instructions are screened to obtain energy allocation decisions.

10. A computer-readable storage medium, characterized in that A computer program is stored therein, and when the computer program is executed, the multimodal data fusion method based on energy scheduling as described in any one of claims 1 to 8 is implemented.

Images