A power distribution network load prediction and optimization method based on big data

Abstract

The application discloses a power distribution network load prediction and optimization method based on big data, comprising: constructing a power distribution network digital twin in multiple regions and generating virtual load samples; constructing a physical information neural network as a global model, and each region uses local real data and virtual samples for federated learning training; migrating the converged global model to a target region for meta-learning fine-tuning to obtain a target load prediction model; using the digital twin of the target region to simulate and check the physical compliance of the prediction result, and triggering online fine-tuning of the model or generating an optimization strategy for execution according to the checking result; finally, multi-region digital twin collaborative simulation is used to realize cross-regional load scheduling optimization. Through virtual-real data fusion training and twin closed-loop verification, the physical rationality and accuracy of load prediction and the safety and efficiency of cross-regional collaborative scheduling are significantly improved.

Description

Technical Field

[0001] This invention relates to the field of intelligent operation and control technology for power systems, and in particular to a method for load forecasting and optimization of distribution networks based on big data. Background Technology

[0002] With the continuous improvement of the intelligence level of new power systems and distribution networks, the operation mode of distribution networks is gradually showing characteristics of multi-regional collaboration, frequent source-load interaction, and highly dynamic operation status. Especially in the scenario of cross-regional distribution networks, load changes are affected by a variety of factors such as meteorological conditions, distributed power source access, user electricity consumption behavior, and energy interaction between regions, which puts forward higher requirements for load forecasting accuracy and dispatch decision-making capabilities.

[0003] Currently, distribution network load forecasting and operation optimization generally rely on historical operating data and statistical or data-driven models for analysis. However, in practical applications, there are significant differences in data scale, data quality, and operational characteristics between different regions. On the one hand, some regions are limited by infrastructure conditions or load scale, resulting in a relatively limited amount of real operating data available for modeling, making it difficult for model training to fully depict load evolution patterns. On the other hand, cross-regional collaborative analysis is constrained by factors such as data security and privacy protection, making it difficult to achieve direct sharing of raw data, thus creating practical obstacles to multi-regional joint modeling.

[0004] Furthermore, existing load forecasting methods often focus on correlation mining at the data level, neglecting the inherent physical constraints of the distribution network itself, such as power flow balance, equipment capacity limitations, and network topology. This forecasting approach, lacking physical consistency constraints, is prone to producing forecast results that do not match the actual operating conditions, thereby affecting the reliability and security of subsequent scheduling and control strategies.

[0005] On the other hand, with the development of simulation and modeling technologies, digital models have been gradually introduced into power distribution network operation analysis. However, their application is often limited to single-region or offline analysis scenarios, and a deep collaborative mechanism with load forecasting models has not yet been formed. Under cross-regional operation conditions, load changes and operating states between regions are coupled. Without effective collaborative simulation and verification methods, it is difficult to systematically evaluate the physical rationality of the forecast results and their cross-regional impact.

[0006] Meanwhile, in actual operation, the load characteristics of the distribution network exhibit significant time-varying and regional differences. Traditional static models or prediction models trained in a single instance are difficult to adapt to the needs of new regions or operating conditions in a timely manner. When the model is applied to target areas with scarce load data or large differences in operating characteristics, prediction accuracy and stability are often difficult to guarantee, lacking the ability to quickly adapt and correct online.

[0007] Therefore, there is an urgent need for a load forecasting and optimization method that can achieve multi-regional collaborative modeling while ensuring data security and privacy, and taking into account both data-driven capabilities and the physical constraints of the distribution network. At the same time, the method should also have the ability to quickly adapt to new regions and small sample scenarios, and form a closed-loop mechanism through simulation verification and operational optimization to improve the safety, reliability and overall efficiency of cross-regional distribution network operation. Summary of the Invention

[0008] In view of the shortcomings of existing technologies, the purpose of this invention is to provide a big data-based method for load forecasting and optimization of power distribution networks, which can improve the forecasting accuracy of small sample areas, cross-regional coordination efficiency, and the overall safety and reliability of power grid operation.

[0009] To achieve the above objectives, the present invention provides the following technical solution: a method for load forecasting and optimization of distribution networks based on big data, comprising the following steps: Step S1: In at least two regional power distribution networks, construct digital twins of the power distribution network for each region based on real-time operational data collected by sensors in the corresponding regions. Step S2: Based on the digital twin of the distribution network, generate virtual load samples for each region; Step S3: Construct a physical information neural network as the global model for federated learning; each region uses local real-world data and virtual load samples to train the global model locally and obtain local model parameters. Step S4: After privacy processing of the local model parameters of each region, the parameters are uploaded to the coordination server. The coordination server performs federated aggregation of the privacy-processed local model parameters, generates updated global model parameters, and distributes them to each region to iteratively converge the global model. Step S5: The global model that has been converged through iteration in step S4 is transferred to the target area, and the global model is quickly fine-tuned using a small amount of real operating data from the target area through a meta-learning algorithm to obtain the target load prediction model. Step S6: Input the load forecasting results output by the target load forecasting model into the distribution network digital twin of the target area for physical compliance simulation verification; if the error index obtained by the simulation verification exceeds the preset error threshold, the online parameter fine-tuning of the target load forecasting model is triggered; if the error index does not exceed the preset error threshold, the distribution network digital twin generates load allocation and power optimization strategies based on the current power grid operating status and issues them for execution. In step S7, digital twins of distribution networks in multiple regions share boundary operation data through a secure communication channel, perform cross-regional collaborative simulation, and optimize and generate cross-regional load dispatch strategies based on the collaborative simulation results and execute them.

[0010] Furthermore, in step S1, the process of constructing a digital twin of the distribution network also includes: Based on real-time operational data, the structural similarity and physical constraint verification of the constructed digital twin of the distribution network are performed. Among them, the structural similarity verification index is not lower than the preset similarity threshold, and the power balance error of the physical constraint verification does not exceed the preset accuracy threshold. The digital twin of the distribution network constructed in step S1 includes a sensing layer, an analysis layer, and a decision-making layer; The perception layer is used to collect real-time operational data via sensors at millisecond-level frequencies; The analysis layer is used for real-time streaming processing of the collected data; The decision layer is used to run the reinforcement learning decision engine.

[0011] Furthermore, in step S2, the step of generating virtual load samples based on the digital twin of the distribution network specifically includes: Virtual load samples are generated using a physical constraint generative adversarial network, where the generator of the physical constraint generative adversarial network uses the core physical laws of the power system as the mapping function. The quality of virtual load samples is optimized collaboratively by an authenticity discriminator and a physical compliance discriminator until it meets a preset quality threshold.

[0012] Furthermore, in step S3, the loss function of the physical information neural network is composed of a weighted average of the data fitting loss term and the physical constraint regularization loss term; The physical constraint regularization loss term includes at least the nodal power balance equation constraints and the line current carrying capacity constraints. When each region trains the global model locally, the weight of the data fitting loss term in the loss function is the first weight coefficient, and the weight of the physical constraint regularization loss term is the second weight coefficient. The first weight coefficient is greater than the second weight coefficient.

[0013] Furthermore, in step S3, the local training data used by each region is composed of a mixture of the region's real operating data and virtual load samples generated based on the region's digital twin; Among them, virtual load samples are dedicated to expanding the local training set under the federated learning framework to optimize local model parameters; When each region trains the global model locally, the total number of virtual load samples used does not exceed a preset proportion threshold of the total number of local real training samples.

[0014] Furthermore, step S4 specifically includes: After differential privacy processing of the local model parameters for each region, they are uploaded to the coordination server; The coordination server uses a federated averaging algorithm to aggregate the received model parameters from each region, generate updated global model parameters, and distribute them to each region. Repeat step S3 up to this step until the loss function value of the global model drops below the preset convergence threshold, thus completing the iterative convergence.

[0015] Furthermore, in step S5, the global model is rapidly fine-tuned using a meta-learning algorithm, specifically as follows: The model-independent meta-learning algorithm is adopted. Based on only a small amount of real running data in the target region, the parameters of the global model are updated within a preset threshold of gradient update times to adapt to the load characteristics of the target region. The number of samples of the small amount of real running data is lower than the preset threshold of a small amount of samples.

[0016] Furthermore, in step S6, the physical compliance simulation verification specifically includes: The load prediction results output by the target load prediction model are fed back to the digital twin of the distribution network in the same target area for real-time simulation. By calculating the voltage deviation and system power imbalance of key nodes in the digital twin, error indicators are generated to characterize the physical feasibility of the prediction results.

[0017] Furthermore, in step S6, the physical compliance simulation verification further includes: Based on error metrics, a dynamic feedback closed loop is formed with the digital twin as its core. When the error index exceeds the preset error threshold, the feedback path for online parameter fine-tuning of the target load prediction model is triggered. When the error index does not exceed the preset error threshold, the feedback path is triggered based on the real-time operation status of the power grid simulated by the digital twin, and load allocation and power optimization strategies are generated and distributed through the optimization decision algorithm.

[0018] Further, in step S7: The secure communication channel is an encrypted channel that supports secure transmission over long distances; The process of sharing boundary operation data includes: Based on the preset security verification principles, the boundary operation data to be shared in each area is verified to ensure that the physical quantity corresponding to the data does not exceed the limit capacity of the interconnection channel; The process of cross-regional collaborative simulation includes: Based on the shared boundary operation data, the overall network operation status after load transfer is simulated in the digital twins of each region to accurately identify the load coupling relationship between regions.

[0019] The beneficial effects of this invention are: Compared with the prior art, the present invention has at least the following beneficial effects: 1. Improve the accuracy and stability of distribution network load forecasting: This invention constructs a digital model in a multi-regional distribution network and combines it with a multi-regional collaborative training mechanism, enabling the load forecasting model to simultaneously learn the operating characteristics and load change patterns of different regions. This effectively mitigates the impact of insufficient data from a single region on forecast accuracy. Especially in scenarios with small load data scales or complex operating conditions, it can significantly improve the accuracy and stability of forecast results and enhance the robustness of the model under complex operating conditions and extreme load changes.

[0020] 2. Enhance the physical consistency and operational safety of prediction results: By introducing physical constraints of the power grid during model training and prediction, and using digital models to simulate and verify the prediction results, this invention can effectively avoid prediction outputs that do not conform to the actual operating rules of the distribution network, ensure that the load prediction results meet the requirements of network structure and operation constraints, thereby reducing safety risks such as overload and exceeding limits during the operation of the distribution network, and improving the reliability of scheduling and operation decisions.

[0021] 3. Achieving cross-regional collaborative modeling while ensuring data security: This invention adopts a multi-regional local training and parameter-level collaboration approach to achieve collaborative optimization of models between different regions without sharing original operating data. This fully utilizes multi-regional operating information to improve overall model performance while ensuring data security and privacy in each region, thus solving the problem of limited data sharing in cross-regional collaborative prediction.

[0022] 4. Enhance the model's ability to adapt quickly to new regions and scenarios: By migrating the converged global model to the target region and making rapid adjustments using a small amount of local operating data, this invention can obtain a load prediction model suitable for the target region in a short time, reduce the cost of model retraining, and improve the system's adaptability to new region access and changes in new operating scenarios.

[0023] 5. Achieving a closed-loop operation mechanism for prediction, verification, and scheduling optimization: This invention incorporates load forecasting results into a digital model for physical compliance simulation, and dynamically adjusts model parameters or generates operational load allocation and power optimization strategies based on the simulation results, forming a closed-loop operation process of prediction-verification-optimization. Simultaneously, through cross-regional collaborative simulation and scheduling optimization, it improves the coordination of load allocation and resource utilization between regions, enhancing the overall operational efficiency of cross-regional distribution networks.

[0024] 6. Possesses good engineering feasibility and promotional value: The method of this invention can be deployed in conjunction with existing power distribution network monitoring, control and management systems, requires minimal modification to existing hardware and system architecture, is applicable to power distribution networks of different sizes and regional types, and has strong engineering feasibility and prospects for promotion and application. Attached Figure Description

[0025] Figure 1 This is a flowchart of the steps in the big data-based distribution network load prediction and optimization method of the present invention. Detailed Implementation

[0026] The present invention will be further described in detail below with reference to the accompanying drawings and embodiments. Identical components are denoted by the same reference numerals. It should be noted that the terms "front," "rear," "left," "right," "upper," and "lower" used in the following description refer to directions in the accompanying drawings, and the terms "bottom surface," "top surface," "inner," and "outer" refer to directions toward or away from the geometric center of a specific component, respectively.

[0027] Example 1, refer to Figure 1 This is the first embodiment of the present invention, which provides a method for load prediction and optimization of distribution networks based on big data.

[0028] I. System architecture and deployment environment of Implementation Example 1; In this embodiment, to achieve load forecasting and operation optimization of multi-regional distribution networks, the following system architecture is deployed in the distribution networks of at least two regions.

[0029] (a) Hardware deployment environment; Regional hardware equipment: Deployed in each distribution network area: Edge computing terminals: used for local model training and inference, supporting the operation of lightweight neural network models; Quantum encryption authentication terminal: used for secure communication between regions and between regions and the coordination server, supporting a hybrid wireless and fiber optic transmission method with a transmission distance of not less than 125 kilometers; Sensor array: Deployed on the power generation side, transmission side and load side, covering operating parameters such as node voltage, branch current, load power and new energy output.

[0030] Central hardware: A federated coordination server is set up to receive model parameters uploaded from each region, perform federated aggregation calculations, and distribute updated global model parameters to each region.

[0031] The aforementioned hardware deployment enables each region to operate independently and perform collaborative computing, providing a foundation for subsequent distributed modeling and cross-regional collaboration.

[0032] (ii) Software architecture environment; The software layer adopts a multi-module collaborative architecture, including: A multi-agent communication system based on the VOLTTRON™ platform; The federated learning module uses a federated averaging algorithm to aggregate parameters. The load forecasting model uses a physical information neural network (PINN architecture). The virtual sample generation module uses the PC-Generator generation model; The model is rapidly adapted using the MAML meta-learning algorithm; The runtime optimization module uses the DDPG reinforcement learning algorithm; Data processing uses the Apache Kafka + Flink real-time stream processing framework.

[0033] The aforementioned software modules, through decoupled deployment, enable the coordinated operation of data acquisition, model training, simulation verification, and scheduling optimization.

[0034] II. Specific implementation steps and working principle of Example 1; Step S1: Data acquisition and construction of a digital twin of the distribution network; Each region acquires real-time operational data such as node voltage, branch current, load power, and new energy output through sensor arrays at a sampling frequency of 50ms, and classifies and labels them according to industrial load, residential load, and commercial load.

[0035] Based on the collected data, a corresponding digital twin of the distribution network is constructed for each region. This digital twin is used to map the topology and operating status of the real distribution network, and its construction process includes: The structural similarity between the digital twin and the real power distribution network is verified using the multi-scale flatness norm method, with a required similarity of no less than 95%. Physical consistency verification is performed by embedding power balance constraints to ensure that the power balance error of the twin simulation does not exceed 0.5%.

[0036] Technical benefits: This step provides a high-fidelity physical foundation for subsequent virtual sample generation, prediction result verification, and scheduling simulation, preventing model training and optimization from deviating from the actual operation of the power grid.

[0037] Step S2: Generate virtual load samples based on digital twins; Within each region, virtual load samples are generated using the existing digital twin of the distribution network and a PC-Generator model to expand the scale of training data.

[0038] Among them: the virtual sample generation process is controlled by the physical constraints of the digital twin; the proportion of virtual samples in the local training data does not exceed 50%.

[0039] Technical effect: By introducing physically constrained virtual load samples, the problem of insufficient samples in areas with scarce load data is effectively alleviated, while avoiding data distortion caused by unconstrained generation.

[0040] Step S3: Construction and local training of the global federated learning model; A federated learning global model is constructed with a physical information neural network as its core. The loss function of this model consists of a data error term and a physical constraint term.

[0041] The physical constraints include nodal power balance equations and line current carrying capacity limits, which are used to constrain the model output to meet the operating rules of the distribution network.

[0042] Each region uses local real-world operational data and virtual load samples to train the global model locally, obtaining local model parameters. After local training is complete, the model parameters undergo differential privacy processing before being uploaded to the federated coordination server.

[0043] Technical effect: By introducing physical constraints during the model training phase, the prediction model can improve accuracy while maintaining physical consistency, thus reducing the risk of subsequent operation.

[0044] Step S4: Federated aggregation and global iteration of model parameters; The federal coordination server receives the privacy-processed local model parameters uploaded by each region and aggregates them using a federal averaging algorithm to generate updated global model parameters.

[0045] The aggregated global model parameters are then distributed to each region, and the above local training and aggregation process is repeated 100–200 times until the following conditions are met: The model is considered convergent if the global loss function value is no higher than 0.01.

[0046] Technical effect: It enables collaborative learning of operational characteristics across multiple regions without sharing the original data, balancing data privacy and model performance.

[0047] Step S5: Target region model transfer and rapid fine-tuning; The converged global model is then transferred to the target region (such as a remote area where load data is scarce). Using 5–10 sets of real load data from that region, the model is rapidly fine-tuned using the MAML meta-learning algorithm. The fine-tuning process involves 5–10 gradient updates.

[0048] Technical benefits: Enables the model to quickly adapt to the load characteristics of the target region, avoiding the time cost and insufficient accuracy problems caused by training from scratch.

[0049] Step S6: Validation of prediction results and execution of closed-loop optimization; The fine-tuned target load prediction model outputs load prediction results, which are then input into the digital twin of the distribution network in the target area for physical compliance simulation verification, and error indicators such as voltage deviation and power imbalance are calculated.

[0050] Specifically: if the voltage deviation exceeds ±5%, the error index is determined to exceed the preset error threshold, triggering online fine-tuning of the model parameters; If the error index does not exceed the threshold, the digital twin generates load allocation and power optimization strategies based on the current power grid operating status using the DDPG reinforcement learning algorithm, and then sends them to the dispatch system for execution.

[0051] Technical benefits: It forms a dynamic closed loop of "prediction-verification-optimization", improving the executability and operational safety of load forecasting results.

[0052] Step S7: Cross-regional collaborative simulation and scheduling optimization; Digital twins of power distribution networks in multiple regions share boundary operation data through quantum-encrypted communication channels, conduct cross-regional collaborative simulations, and identify power coupling relationships in tie lines.

[0053] Based on the results of collaborative simulation, a cross-regional load dispatching strategy is generated to realize the energy transfer from the renewable energy surplus area to the load center area without exceeding the line carrying capacity limit, and cross-regional dispatching instructions are generated for execution.

[0054] Technical benefits: Enhances the coordinated operation capability of cross-regional power distribution networks and the level of renewable energy consumption, while reducing overall operational risks.

[0055] III. Verification results of Example 1; The verification scenario uses three interconnected distribution network areas in the Yangtze River Delta region, including a remote area with scarce load data. The interconnected distribution network areas specifically include Area A (industrial load dominant, data abundant), Area B (residential + commercial load, data moderate), and Area C (mountainous area, distributed photovoltaic + small hydropower, data scarce).

[0056] During the model training phase: Region C generates 3000 sets of virtual samples through twins to participate in federated training; after the global model is transferred to Region C, it is fine-tuned with 8 sets of real samples to complete the adaptation.

[0057] The forecast results show that the predicted load MAE for regions A, B, and C are 2.1%, 2.3%, and 3.5%, respectively, which are 16%, 18%, and 32% lower than those of the traditional model.

[0058] The scheduling optimization results show that after the cross-regional collaborative scheduling in step S7, the renewable energy consumption rate in region C increased from 75% to 92%, the overall line overload risk in the Yangtze River Delta region decreased by 90%, and the power supply capacity during peak hours increased by 3%.

[0059] Example 2 is the second embodiment of the present invention. This embodiment is a further refinement and optimization based on Example 1, focusing on the construction and verification mechanism of the digital twin of the distribution network, the physical compliance generation mechanism of virtual load samples, and the federated meta-learning training framework with enhanced physical constraints.

[0060] I. Overall working principle of Example 2; This embodiment constructs digital twins of the power distribution network in each region, enabling them to sense, analyze, and make decisions. Based on this, virtual load samples that meet physical constraints are generated. Then, real operating data and virtual load samples are introduced into a federated learning framework. Without sharing the original data, a meta-learning model with embedded power system physical constraints is used for collaborative training. This improves the accuracy of load forecasting while ensuring the physical consistency and engineering feasibility of the forecast results.

[0061] II. Construction and Verification Process of Digital Twin of Distribution Network; (a) Construction of the hierarchical structure of digital twins; In step S1, a corresponding digital twin of the distribution network is constructed for each regional distribution network. The digital twin includes a sensing layer, an analysis layer, and a decision-making layer.

[0062] Among them, the sensing layer is used to collect real-time operating data such as node voltage, current, power, and new energy output at millisecond frequency through sensors deployed on the generation side, transmission side, and load side. The number of sensor nodes can reach more than 100,000. The analysis layer is used to perform real-time streaming processing on the data collected by the perception layer. Apache Kafka and Flink are used to implement data cleaning, time alignment and feature extraction. The decision layer is used to run the reinforcement learning decision engine, providing decision support for subsequent load forecasting verification and scheduling optimization.

[0063] (ii) Structural similarity verification and physical constraint verification; After the digital twin is constructed, its structural similarity is verified and its physical constraints are checked based on real-time running data.

[0064] Among them, structural similarity is calculated using multi-scale flatness norm, and its verification index is no less than the preset similarity threshold of 95%; Physical constraint verification is performed through power balance constraints, requiring that the simulation power balance error of the digital twin does not exceed a preset accuracy threshold of 0.5%.

[0065] When the structural similarity or power balance error does not meet the above threshold requirements, the parameters of the digital twin are adjusted until the requirements are met.

[0066] Technical effects: This verification mechanism ensures that the digital twin is highly consistent with the real power distribution network in terms of structure and physical behavior, providing a reliable simulation basis for subsequent virtual sample generation and prediction result verification.

[0067] III. Generation mechanism of physically compliant virtual load samples; (a) The principle of generating virtual load samples; In step S2, virtual load samples are generated based on the digital twin of the distribution network, specifically using a physical constraint generative adversarial network.

[0068] Among them, the generator of the generative adversarial network uses the core physical laws of the power system as the mapping function to map random disturbances into virtual load samples that satisfy the node power balance relationship and the line current carrying capacity limit.

[0069] (ii) Virtual sample quality control mechanism; The generated virtual load samples are quality-assessed using the following dual discrimination mechanism: The authenticity discriminant is used to determine the consistency between the statistical distribution of the virtual load sample and the real load data; The physical compliance discriminator is used to determine whether virtual load samples meet the physical operating constraints of the distribution network.

[0070] Through collaborative optimization of the generator and discriminator, the virtual payload samples can only be used as valid training samples after they meet the preset quality threshold.

[0071] Technical effect: This mechanism avoids the problem of generating samples without physical meaning in traditional data augmentation methods, so that the virtual load samples have both statistical authenticity and physical feasibility.

[0072] IV. A Federated Meta-Learning Training Framework with Enhanced Physical Constraints; (a) Three-level model architecture and training process; In step S3, a three-level training architecture of "global model - local model - physical constraint layer" is constructed.

[0073] The global model aggregates local model parameters from multiple regions using a federated averaging algorithm; The local model is trained based on the MAML meta-learning algorithm to enable the model to quickly adapt to the load characteristics of different regions. The physical constraint layer is used to embed the core physical equations of the distribution network into the model training process.

[0074] Local training data for each region consists of a mixture of real operational data from that region and virtual load samples generated based on the digital twin of that region, wherein: Virtual load samples are only used to augment the local training set within the federated learning framework. The total number of virtual load samples shall not exceed a preset percentage threshold of 50% of the total number of local real training samples.

[0075] (II) Loss Function Design and Formula Interpretation; In this embodiment, the loss function of the physical information neural network is designed as follows: ; in, It is the overall training objective function of the model, and is a comprehensive loss value used to measure the model's predictive performance and physical consistency. This is a coefficient used to adjust the weight of the data fitting term in the overall loss, and its value ranges from 0.6 to 0.8. Used to control the degree of attention the model pays to the accuracy of load forecasting. It is the mean squared error loss between the predicted load value and the actual load value, used to characterize the model's ability to fit historical and real-time load data. This is a coefficient used to adjust the weight of the physical constraint regularization term in the overall loss. Its value ranges from 0.2 to 0.4, and it is used to balance the influence of physical constraints on model training. It is a penalty term for the degree to which the model output violates the physical operating constraints of the distribution network. It is used to quantify the degree to which the node power balance equation and line current carrying capacity constraints are violated, thereby guiding the model output to meet the physical laws of the power system.

[0076] Technical effect: Through the above loss function design, the model can simultaneously take into account prediction accuracy and physical compliance during the federated meta-learning training process, fundamentally reducing the generation of physically unreasonable prediction results.

[0077] V. Privacy Protection and Virtual Participant Mechanisms; During federated learning, each region only uploads model parameters processed with differential privacy and does not share raw payload data; cross-regional boundary data is transmitted using quantum encryption, supporting long-distance secure communication of over 125 kilometers.

[0078] Meanwhile, the digital twins of the distribution networks in each region are connected to the federated learning alliance as virtual participants. The virtual samples only participate in training locally and are not transmitted to the outside world, which increases the amount of effective training data in small sample areas to more than 300% of the original scale.

[0079] Technical effect: Significantly improves model training quality and prediction stability in data-scarce regions without increasing the risk of data leakage.

[0080] VI. Explanation of the overall effects of Example 2; The method in Example 2 enables cross-regional collaborative training of distribution network load forecasting models while ensuring data privacy and communication security. It also significantly improves the physical consistency and engineering feasibility of forecasting results through digital twins and physical constraint mechanisms, making it particularly suitable for regional distribution network scenarios with complex load structures or scarce data.

[0081] In this embodiment, the global model in the federated learning framework uses the following unified prediction-physical consistency joint mapping formula to model and infer the load status of multi-regional distribution networks: ; The above formula is used to construct the core mapping relationship of the global model of federated meta-learning in Example 2. It achieves a unified balance between statistical fitting accuracy, physical consistency and cross-regional generalization ability of load prediction results by jointly modeling real operation data collected from multiple regions, virtual load samples generated by digital twins and embedded power system physical constraints.

[0082] Among them, the integral structure is used to characterize the continuous evolution of load over time, the normalization term is used to eliminate the influence of load scale differences in different regions, the Gamma function and Riemann Zeta function are used to model the fluctuation characteristics of heavy-tailed loads and cross-scale energy distribution, the Bessel function and elliptic function are used to describe the oscillatory behavior caused by the distribution network topology and tie-line power coupling, the error function is used to suppress the interference of abnormal sampling points on the model output, and the exponential function is used to strengthen the penalty effect of the degree of physical constraint violation on the prediction results. This ensures that the global model always meets the physical feasibility requirements of distribution network operation during federated training and transfer applications, and solves the problems of physical non-compliance and insufficient cross-regional generalization ability in the prediction results of existing technologies.

[0083] in, This is the output value of the global model, i.e., the normalized load forecast result. Its value ranges from 0 to 1. The larger the value, the higher the overall load level of the target area during the forecast period. This is the upper limit of the load forecasting time window, in seconds. The Gamma function is used to characterize the higher-order statistical properties of random load fluctuations. The load fluctuation distribution pattern adjustment parameter is obtained from historical load data statistics. It is an exponential function, used to amplify the suppressive effect of physical constraint violations on prediction results. The physical constraint penalty strength coefficient is preset by the system security level. This is a physical consistency filtering function used to measure the degree of physical compliance of the predicted load in the digital twin. This represents the upper limit of the time mapping after fusing virtual and real samples. For a moment The load power input data after normalization processing It is the natural logarithm function, used to compress the influence of load extrema on the model. The Riemann Zeta function is used to describe the multi-scale characteristics of cross-regional load energy distribution. This is a parameter representing the cross-regional load coupling strength, and it is also an independent variable of the Zeta function. These are Bessel functions of the first kind, used to characterize the power oscillation behavior caused by the distribution network topology. This is the order of the Bessel function, corresponding to the topology level of the distribution network tie lines, and is a topology order parameter. These are parameters for adjusting the network structure scale, and are also input coefficients for the Bessel function. This is the error function, used to suppress the impact of abnormal load samples on the model output. Here, represents the anomaly suppression sensitivity parameter, and is the scaling factor for the error function. This represents the effective load characteristic of the k-th region after physical verification, and is the regional load characteristic input. This represents the number of regions participating in federal learning, and is the total number of regions. These are Jacobi elliptic functions used to characterize periodic load changes and fluctuations in new energy output; they are elliptic function operators. This is a load cycle-scale adjustment parameter, and also a time scaling factor. The modulus parameter of the elliptic function reflects the nonlinearity of the load cycle fluctuation and is a periodic nonlinearity coefficient.

[0084] In this embodiment, the global model output value The range of values ​​is ,in A value close to 0 indicates that the target area is under low load or that the load can be completely absorbed by local power sources during the forecast time window. A value close to 1 indicates that the target area is under high load or requires cross-regional collaborative scheduling. This output can be directly used as the input for subsequent digital twin physical verification, reinforcement learning scheduling decisions, and cross-regional load allocation strategies in Example 2, ensuring consistency in the dimensional and physical meaning of the prediction, verification, and optimization processes.

[0085] By introducing the aforementioned unified global model formula, this embodiment can achieve deep fusion modeling of multi-region load data, virtual load samples, and physical constraint information within the federated meta-learning framework. This enables the global model to maintain stable prediction accuracy and strict physical compliance when migrating across regions, thereby significantly improving the practicality and reliability of distribution network load forecasting and optimization methods in complex, multi-regional power system scenarios.

[0086] Example 3 is the third embodiment of the present invention. Based on Example 1 and Example 2, this embodiment further deepens the federated aggregation and convergence process of the global model, the fast model adaptation mechanism for the target region, and the closed-loop correction mechanism for the physical compliance of the prediction results.

[0087] I. Overall working principle of Example 3; This embodiment obtains a global model that satisfies cross-regional generalization capability through a federated average aggregation mechanism under differential privacy protection. Then, the global model is migrated to the target area and quickly fine-tuned using a small amount of local real-world operational data. Subsequently, the prediction results are fed back to the digital twin of the distribution network in the same target area for real-time physical simulation verification. Based on the verification results, a dynamic feedback closed loop is formed. While ensuring prediction accuracy, the secure distribution of scheduling load allocation and power optimization strategies is realized, thereby constructing a closed-loop operation mechanism of "collaborative training - local verification - model optimization - scheduling execution".

[0088] II. Federated Aggregation and Model Iterative Convergence Process; In step S4, after completing the local model training in step S3, each region first performs differential privacy processing on the local model parameters to prevent the original load data from being deduced from the model parameters. During the differential privacy processing, controlled random noise is injected into the model parameters to effectively mask the influence of individual samples on the parameter update results.

[0089] The local model parameters, after differential privacy processing, are uploaded to the coordination server. The coordination server then uses a federated averaging algorithm to aggregate the model parameters from multiple regions, generating updated global model parameters. This federated averaging algorithm, by weighting the model parameters from each region, allows the global model to simultaneously incorporate load characteristic information from different regions.

[0090] Subsequently, the coordination server distributes the updated global model parameters to each region. Each region then executes steps S3 and S4 again based on the distributed global model, forming an iterative training process for the model.

[0091] When the loss function value of the global model drops below the preset convergence threshold of 0.01, the global model is determined to have reached convergence, and the federated training process ends.

[0092] Technical effects: Through this federated aggregation and convergence mechanism, collaborative learning of load features in multiple regions can be achieved without sharing the original load data. Overfitting or ineffective iterations can be avoided by controlling the convergence threshold, thereby improving the efficiency and stability of model training.

[0093] III. Target Region Model Transfer and Rapid Fine-tuning Mechanism; In step S5, the globally trained and converged model is transferred to the target region. The target region is typically a distribution network area with a small load data scale or a load structure with obvious regional characteristics.

[0094] After model transfer, a model-independent meta-learning algorithm (MAML) is used to quickly fine-tune the global model. Specifically, the model parameters are updated based only on a small amount of real-world data collected from the target region, where: The number of samples of a small amount of real-world data is less than the preset threshold of 10 groups for a small amount of data. The number of gradient updates for model parameter updates shall not exceed the preset threshold of 10.

[0095] By adjusting model parameters within a limited number of gradient updates, the model can quickly adapt to the load change characteristics of the target region.

[0096] Technical effect: This mechanism avoids the computational overhead of large-scale retraining of the target region, enabling the model to maintain high prediction accuracy and response speed even under data-scarce conditions.

[0097] IV. Simulation verification process for the physical compliance of the prediction results; In step S6, the load forecast results for the target area within the forecast time window are output using the fine-tuned target load forecast model. The load forecast results are fed back to the digital twin of the distribution network in the same target area for real-time simulation.

[0098] During the simulation, the voltage deviation of key nodes and the overall power imbalance of the system are calculated using a digital twin, and an error index is generated based on the above calculation results to characterize the physical feasibility of the prediction results.

[0099] Among them: voltage deviation is used to reflect the degree of deviation of node voltage from rated voltage under predicted load conditions; power imbalance is used to reflect the balance between input power, output power and system losses under predicted load conditions.

[0100] Technical effect: By verifying through physical simulation, the predicted results are aligned with the actual power grid operation constraints, thus avoiding safety risks caused by the model output during the engineering implementation phase.

[0101] V. Dynamic feedback closed-loop mechanism based on digital twins; In step S6, based on the error index, a dynamic feedback closed-loop mechanism with the distribution network digital twin as the core is further formed.

[0102] When the error index exceeds the preset error threshold, the first feedback path is triggered: Specifically, the preset error threshold is defined as a node voltage deviation exceeding ±5% or a system power imbalance exceeding 1%. In this case, online parameter fine-tuning of the target load prediction model is triggered, and the prediction results are corrected by locally adjusting the model parameters without having to re-participate in federated training, thus avoiding the introduction of additional communication overhead and privacy risks.

[0103] Specifically, the online parameter fine-tuning process includes the following steps: First, within the current prediction time window, the digital twin performs real-time physical simulation of the load prediction results output by the target load prediction model, calculates the node voltage deviation and system power imbalance, and maps the error indicators to feedback signals for model fine-tuning. Specifically, when the voltage deviation of any key node exceeds ±5%, or the system power imbalance exceeds 1%, the error indicator is determined to have exceeded the preset error threshold.

[0104] Secondly, using the error index as a constraint signal, an online fine-tuning loss function is constructed. This online fine-tuning loss function introduces a penalty term related to voltage deviation and power imbalance on the basis of the original data fitting loss term, which is used to constrain the model output to adjust in the direction that meets the physical operating conditions.

[0105] Subsequently, a small-step gradient update method was used to update some parameters of the target load prediction model online, where: Online fine-tuning only applies to high-level parameters of the model or output mapping layer parameters; The number of gradient updates in a single online fine-tuning session shall not exceed 5. The learning rate for online fine-tuning is lower than that for offline training to avoid model parameter oscillations.

[0106] After completing the online parameter update, the updated target load forecasting model is used to re-output the load forecasting results, which are then input again into the digital twin of the distribution network in the same target area for physical simulation verification. If the error index still exceeds the preset error threshold, the above online fine-tuning process is repeated; if the error index meets the threshold requirements, the online fine-tuning is terminated.

[0107] Technical effect: Through the above-mentioned online parameter fine-tuning mechanism, the target load prediction model can adaptively correct itself based on actual physical feedback during operation, avoiding the continuous accumulation of prediction errors. At the same time, it can maintain the physical compliance and real-time performance of the prediction results without having to re-participate in federated training.

[0108] When the error index does not exceed the preset error threshold, the second feedback path is triggered: Based on the current simulated real-time operating status of the power grid, the digital twin generates load allocation and power optimization strategies by comprehensively considering factors such as line load rate and renewable energy output level, and then distributes the load allocation and power optimization strategies to the actual dispatching system for execution.

[0109] In Example 3, when the error index obtained from the physical compliance simulation verification does not exceed the preset error threshold, the process of generating load allocation and power optimization strategies based on digital twins is triggered.

[0110] Specifically, the process includes the following steps: First, within the current prediction time window, the digital twin acquires real-time power grid operation status parameters for the target area based on the verified load prediction results, including the load rate of each line, the voltage level of each node, the output capacity of distributed renewable energy sources, and the adjustable range of conventional power sources.

[0111] Secondly, using the power grid operating state parameters as the environmental state input, a reinforcement learning optimization environment is constructed, and the Deep Deterministic Policy Gradient Algorithm (DDPG) is adopted as the optimization decision algorithm. Specifically: The state space of reinforcement learning includes line load rate, node voltage deviation, available output of new energy sources, and load demand level. The action space includes the load distribution ratio adjustment of each load node and the output regulation of each distributed power source. The reward function takes into account factors such as reducing line overload risk, suppressing voltage overruns, increasing the proportion of renewable energy consumption, and reducing the magnitude of dispatch adjustments.

[0112] Subsequently, the DDPG algorithm, based on the state space and reward function, generates load allocation and power output optimization strategies that satisfy line current carrying capacity limits and node voltage constraints. Before generation, the optimization strategies are verified through simulation using a digital twin to ensure that no new physical violations occur under predicted load conditions.

[0113] After the simulation verification is successful, the load allocation and power output optimization strategy will be sent to the actual scheduling system for execution, including time-sharing and zone-based adjustment of controllable loads, and coordinated control of distributed power sources and conventional power output.

[0114] The actual results after execution can be collected again by the digital twin and used for subsequent prediction and verification processes, thus forming a continuously running optimization closed loop.

[0115] Technical effects: By generating and executing load allocation and power optimization strategies under the premise that the prediction results are physically feasible, the distribution network can improve the renewable energy absorption capacity and line utilization efficiency while ensuring operational safety, and enhance the system's adaptability to load fluctuations and renewable energy uncertainties.

[0116] In both of the above feedback paths, the corrected prediction results or optimization strategies can be input into the digital twin for verification, thus forming a continuously iterative closed-loop process.

[0117] This closed-loop mechanism enables dynamic coordination between the prediction model and the power grid operation status, forming a unified closed loop for load forecasting, model optimization, and dispatch execution, which significantly improves the safety, flexibility, and operational efficiency of the distribution network in complex operating scenarios.

[0118] VI. Explanation of the overall effects of Example 3; Through Example 3, the model can be rapidly adapted to the target region based on cross-regional collaborative training. The prediction results and load allocation and power optimization strategies can be continuously corrected through a closed-loop feedback mechanism driven by a digital twin. This improves the accuracy of power distribution network load forecasting and the feasibility of scheduling decisions while ensuring data privacy and operational security.

[0119] Example 4 is the fourth embodiment of the present invention. Based on Examples 1 to 3, this embodiment further refines the boundary data security sharing, cross-regional load coupling relationship identification, and scenario-based scheduling strategy generation mechanism in cross-regional collaborative operation scenarios.

[0120] I. Overall working principle of Example 4; This embodiment is based on Embodiments 1 to 3, and further refines the boundary data security sharing, cross-regional load coupling relationship identification, and scenario-based scheduling strategy generation mechanism in cross-regional collaborative operation scenarios.

[0121] Its core working principle is as follows: by constructing an encrypted communication channel that supports long-distance secure transmission, it enables boundary operation data sharing between digital twins of multi-regional distribution networks under the premise of ensuring data security and physical feasibility; then, based on the shared data, it conducts collaborative simulation in digital twins of each region to accurately identify the load coupling relationship between regions, and generates executable load scheduling and energy mutual assistance strategies under different cross-regional operation scenarios, thereby improving the overall operation efficiency and safety level of cross-regional distribution networks.

[0122] II. Secure communication channels and boundary operation data sharing mechanism; (a) Construction of secure communication channels; In step S7, the digital twins of the distribution networks in each region exchange data through an encrypted communication channel that supports long-distance secure transmission. The secure communication channel combines quantum encryption with traditional encryption mechanisms, enabling secure transmission over long distances of 125 kilometers and above to prevent data theft or tampering during cross-regional communication.

[0123] Technical benefits: This secure communication channel provides a reliable data transmission foundation for cross-regional collaborative simulation, ensuring the integrity and confidentiality of boundary operation data during the sharing process.

[0124] (ii) The process of security verification and sharing of boundary operation data; Before sharing boundary operation data, each region verifies the data to be shared according to preset security verification principles. Boundary operation data includes, but is not limited to, physical state quantities such as tie line power and voltage of critical boundary nodes.

[0125] The specific verification process includes: The boundary operation data to be shared by each region is compared according to the safety verification principle of "taking the smaller of the two", that is, the smaller value of the transmittable power calculated by two adjacent regions for the same interconnection channel is taken as the actual shareable and executable boundary power value, so as to ensure that any cross-regional load transfer plan does not exceed the limit capacity of the interconnection channel.

[0126] The limit capacity of the interconnection channel is the maximum safe transmission power allowed by the channel in its design and operation specifications. When the physical quantity corresponding to the boundary operation data exceeds the limit capacity, it is not allowed to enter the subsequent collaborative simulation process.

[0127] Technical effect: By introducing a security verification mechanism before data sharing, the cross-regional scheduling scheme is prevented from exceeding the physical carrying capacity from the source, reducing the risk of tie line overload and system instability.

[0128] III. Cross-regional digital twin collaborative simulation process; After completing the security verification and sharing of boundary operation data, the digital twins of the distribution networks in each region conduct cross-regional collaborative simulations within their respective regions based on the shared boundary operation data.

[0129] Specifically, while maintaining the network topology, load distribution, and power supply structure of their respective regions, the digital twins of each region input the shared boundary power and boundary node voltage as external constraints into the simulation model to simulate and analyze the overall network operation status after cross-regional load transfer.

[0130] During the simulation, by comparing the line load rate, node voltage changes, and system power distribution under different load transfer schemes, the load coupling relationship between regions caused by interconnection channels is accurately identified, including but not limited to the degree of influence of load changes in a certain region on the line load and voltage level of adjacent regions.

[0131] Technical effect: This collaborative simulation mechanism can break through the limitations of single-region simulation, enabling the quantitative identification of cross-regional load interaction relationships, and providing a reliable basis for subsequent scheduling decisions.

[0132] IV. Utilization of cross-regional coupling characteristics and scenario-based scheduling optimization; In Example 4, based on the load coupling relationship identified by cross-regional collaborative simulation, the relevant coupling features are fed back to the federated meta-learning framework to enhance the global model's ability to learn cross-regional interaction characteristics, thereby improving the accuracy of subsequent load forecasting and scheduling decisions in cross-regional scenarios.

[0133] At the same time, scheduling strategies are designed in a scenario-based manner according to different cross-regional operation scenarios, including but not limited to: Large-scale DC power transmission projects: In this scenario, collaborative simulation prioritizes the basic load demand of the receiving area, and, without exceeding the limit capacity of the DC channel, transmits the surplus power from the sending area across regions.

[0134] Interconnected network engineering mutual assistance scenarios: In this scenario, each region prioritizes meeting its own load demand. When there is surplus capacity, cross-regional collaborative simulation is used to evaluate the impact of different mutual assistance schemes on the overall network operation. Under the premise of meeting safety constraints, the surplus capacity is opened to other regions.

[0135] Taking the Fujian-Guangdong Interconnection Project as an example, while ensuring the transaction needs of the province, the allocation of remaining channel capacity is optimized by using collaborative simulation results to achieve efficient utilization of cross-regional resources.

[0136] Technical effects: Through scenario-based scheduling optimization, cross-regional load scheduling strategies are made more aligned with actual operational needs, improving cross-regional energy sharing capabilities and overall system operating efficiency.

[0137] V. Explanation of the overall effects of Example 4; The method described in Example 4 enables efficient collaborative simulation and scheduling decision generation between digital twins of multi-regional distribution networks, while ensuring cross-regional communication security and physical feasibility. This effectively solves the problem of difficulty in accurately identifying cross-regional load coupling relationships and improves the security, flexibility, and resource allocation efficiency of cross-regional distribution networks in complex operating scenarios.

[0138] The above are merely preferred embodiments of the present invention. The scope of protection of the present invention is not limited to the above embodiments. All technical solutions falling within the scope of the present invention's concept are within the scope of protection of the present invention. It should be noted that for those skilled in the art, any improvements and modifications made without departing from the principle of the present invention should also be considered within the scope of protection of the present invention.

Claims

1. A method for load forecasting and optimization of distribution networks based on big data, characterized in that, Includes the following steps: Step S1: In at least two regional power distribution networks, construct digital twins of the power distribution network for each region based on real-time operational data collected by sensors in the corresponding regions. Step S2: Based on the digital twin of the distribution network, generate virtual load samples for each region; Step S3: Construct a physical information neural network as the global model for federated learning; each region uses local real-world data and virtual load samples to train the global model locally and obtain local model parameters. Step S4: After privacy processing of the local model parameters of each region, the parameters are uploaded to the coordination server. The coordination server performs federated aggregation of the privacy-processed local model parameters, generates updated global model parameters, and distributes them to each region to iteratively converge the global model. Step S5: The global model that has been converged through iteration in step S4 is transferred to the target area, and the global model is quickly fine-tuned using a small amount of real operating data in the target area through a meta-learning algorithm to obtain the target load prediction model. Step S6: Input the load forecasting results output by the target load forecasting model into the digital twin of the distribution network in the target area for physical compliance simulation verification; if the error index obtained from the simulation verification exceeds the preset error threshold, online parameter fine-tuning of the target load forecasting model is triggered. If the error index does not exceed the preset error threshold, the distribution network digital twin will generate load allocation and power optimization strategies based on the current power grid operating status and issue them for execution. In step S7, digital twins of distribution networks in multiple regions share boundary operation data through a secure communication channel, perform cross-regional collaborative simulation, and optimize and generate cross-regional load dispatch strategies based on the collaborative simulation results and execute them.

2. The method for distribution network load forecasting and optimization based on big data according to claim 1, characterized in that, In step S1, the process of constructing a digital twin of the distribution network also includes: Based on real-time operational data, the structural similarity and physical constraint verification of the constructed digital twin of the distribution network are performed. Among them, the structural similarity verification index is not lower than the preset similarity threshold, and the power balance error of the physical constraint verification does not exceed the preset accuracy threshold. The digital twin of the distribution network constructed in step S1 includes a sensing layer, an analysis layer, and a decision-making layer; The perception layer is used to collect real-time operational data via sensors at millisecond-level frequencies; The analysis layer is used for real-time streaming processing of the collected data; The decision layer is used to run the reinforcement learning decision engine.

3. The method for distribution network load forecasting and optimization based on big data according to claim 1, characterized in that, In step S2, the step of generating virtual load samples based on the digital twin of the distribution network specifically includes: Virtual load samples are generated using a physical constraint generative adversarial network, where the generator of the physical constraint generative adversarial network uses the core physical laws of the power system as the mapping function. The quality of virtual load samples is optimized collaboratively by an authenticity discriminator and a physical compliance discriminator until it meets a preset quality threshold.

4. The method for distribution network load forecasting and optimization based on big data according to claim 1, characterized in that, In step S3, the loss function of the physical information neural network is composed of a weighted average of the data fitting loss term and the physical constraint regularization loss term. The physical constraint regularization loss term includes at least the nodal power balance equation constraints and the line current carrying capacity constraints. When each region trains the global model locally, the weight of the data fitting loss term in the loss function is the first weight coefficient, and the weight of the physical constraint regularization loss term is the second weight coefficient. The first weight coefficient is greater than the second weight coefficient.

5. The method for distribution network load forecasting and optimization based on big data according to claim 1 or 4, characterized in that, In step S3, the local training data used by each region consists of a mixture of the region's real operating data and virtual load samples generated based on the region's digital twin. Among them, virtual load samples are dedicated to expanding the local training set under the federated learning framework to optimize local model parameters; When each region trains the global model locally, the total number of virtual load samples used does not exceed a preset proportion threshold of the total number of local real training samples.

6. The method for distribution network load forecasting and optimization based on big data according to claim 1, characterized in that, Step S4 specifically includes: After differential privacy processing of the local model parameters for each region, they are uploaded to the coordination server; The coordination server uses a federated averaging algorithm to aggregate the received model parameters from each region, generate updated global model parameters, and distribute them to each region. Repeat step S3 up to this step until the loss function value of the global model drops below the preset convergence threshold, thus completing the iterative convergence.

7. The method for distribution network load forecasting and optimization based on big data according to claim 1, characterized in that, In step S5, the global model is rapidly fine-tuned using a meta-learning algorithm, specifically as follows: The model-independent meta-learning algorithm is adopted. Based on only a small amount of real running data in the target region, the parameters of the global model are updated within a preset threshold of gradient update times to adapt to the load characteristics of the target region. The number of samples of the small amount of real running data is lower than the preset threshold of a small amount of samples.

8. The method for distribution network load forecasting and optimization based on big data according to claim 1, characterized in that, In step S6, the physical compliance simulation verification specifically includes: The load prediction results output by the target load prediction model are fed back to the digital twin of the distribution network in the same target area for real-time simulation. By calculating the voltage deviation and system power imbalance of key nodes in the digital twin, error indicators are generated to characterize the physical feasibility of the prediction results.

9. The method for distribution network load forecasting and optimization based on big data according to claim 8, characterized in that, In step S6, the physical compliance simulation verification further includes: Based on error metrics, a dynamic feedback closed loop is formed with the digital twin as its core. When the error index exceeds the preset error threshold, the feedback path for online parameter fine-tuning of the target load prediction model is triggered. When the error index does not exceed the preset error threshold, the feedback path is triggered based on the real-time operation status of the power grid simulated by the digital twin, and load allocation and power optimization strategies are generated and distributed through the optimization decision algorithm.

10. The method for distribution network load forecasting and optimization based on big data according to claim 1, characterized in that, In step S7: The secure communication channel is an encrypted channel that supports secure transmission over long distances; The process of sharing boundary operation data includes: Based on the preset security verification principles, the boundary operation data to be shared in each area is verified to ensure that the physical quantity corresponding to the data does not exceed the limit capacity of the interconnection channel; The process of cross-regional collaborative simulation includes: Based on the shared boundary operation data, the overall network operation status after load transfer is simulated in the digital twins of each region to accurately identify the load coupling relationship between regions.

Images