Intelligent power grid adaptive control system and method

Abstract

The invention discloses a smart power grid adaptive control system and method, and relates to the technical field of smart power grid operation control, and the system comprises a physical sensing layer, an edge calculation layer, a cloud super-simulation digital twin platform and a communication network. According to the method, through a central coordination agent in a digital twinborn platform, a multi-agent reinforcement learning algorithm is adopted to carry out cooperative training on a local agent. According to the method, a super-simulation digital twinborn model integrating geometry, physics, behaviors and rules in multiple dimensions is constructed, and through a training normal form of deep fusion of a physical information neural network and a generative adversarial network, the model has both a physical rule following capability and an uncertainty deduction capability; and safe, rapid and adaptive optimization of power grid control in a large-scale and uncertain environment is realized.

Description

Technical Field

[0001] This invention relates to the field of smart grid operation control technology, specifically a smart grid adaptive control system and method. Background Technology

[0002] With the high proportion of renewable energy integration and the diversification of load types, power grid operation exhibits a high degree of uncertainty and complexity. Traditional power grid control systems often employ centralized optimization based on precise physical models or simple local control strategies, which have significant limitations: 1. Strong model dependence: Traditional optimal power flow (OPF) and other methods rely heavily on accurate mathematical models of the power grid. When faced with topology changes and the randomness of distributed energy sources (DER), model mismatch can lead to a decline in control performance or even failure.

[0003] 2. Slow response speed: Centralized control requires the collection of data from the entire network and the solving of complex optimization problems, which takes a long time and is difficult to cope with the second-level and millisecond-level dynamic processes of the power grid (such as voltage over-limit and frequency fluctuations).

[0004] 3. Weak adaptive capability: Most of the parameters of the existing control strategies are tuned offline, which cannot adjust and optimize themselves according to the real-time operating status of the power grid, and lacks the ability to "learn" and "evolve".

[0005] 4. Difficulty in coordinating multiple objectives: It is difficult to simultaneously optimize multiple objectives such as economy (minimum network loss), security (voltage and frequency stability), and power quality in real time.

[0006] 5. Insufficient security safeguards: The exploration of data-driven control methods (such as reinforcement learning) in physical systems may lead to security risks.

[0007] While some research has applied reinforcement learning (RL) to power grid control, most studies involve centralized learning by a single agent, which suffers from the "curse of dimensionality" and poor generalization ability. Some distributed control schemes also exist, but the lack of coordination between controllers easily leads to suboptimal solutions or control conflicts. Therefore, there is an urgent need for an adaptive control scheme that can learn independently, respond quickly, ensure safety, and collaboratively optimize multiple objectives. Summary of the Invention

[0008] The purpose of this invention is to provide an intelligent grid adaptive control system and method. This system can construct a virtual simulation environment through digital twin technology and utilize multi-agent reinforcement learning algorithms to achieve distributed autonomous decision-making and collaborative optimization of units such as distributed power sources, energy storage systems, and flexible loads. This enables the power grid control system to have the capabilities of online learning, adaptive adjustment, safety constraints, and multi-objective collaborative optimization, thereby solving the problems mentioned in the background art.

[0009] To achieve the above objectives, the present invention provides the following technical solution: A smart grid adaptive control system includes: Physical sensing layer, edge computing layer, cloud-based hyper-realistic digital twin platform, and communication network; The physical sensing layer is used to collect electrical measurement data and switch status information of each node in the power grid in real time. The edge computing layer is distributed across multiple key nodes in the power grid. Each node deploys a local intelligent agent to load and execute control strategies for distributed real-time control of local power equipment. The key nodes are the connection nodes where the controlled power equipment is located in the power grid. The cloud-based hyper-realistic digital twin platform includes: A high-fidelity digital twin model that is updated synchronously with the physical power grid; the high-fidelity digital twin model is a refined mathematical model in which the error between the dynamic simulation results and the actual operating state of the physical power grid is within ±5%. A central coordinating agent is configured in the virtual environment provided by the digital twin model and uses a centralized training distributed execution framework to conduct multi-agent reinforcement learning collaborative training on the multiple local agents. An experience replay pool is used to store training data generated by digital twin simulations; The communication network connects the physical sensing layer, the edge computing layer, and the cloud-based hyper-realistic digital twin platform, and is used to realize data exchange and policy parameter distribution.

[0010] As a further preferred embodiment of the present invention: the reward function used by the central coordinating agent during the training process is a dynamic multi-objective weighted function, which has the form: R(t)=α(t)*R_economic(t) +β(t)*R_security(t)+γ(t)*R_quality(t); where R_economic(t) is the economic sub-reward function, whose value is negatively correlated with grid loss and control cost; R_security(t) is the security sub-reward function, whose value is negatively correlated with node voltage deviation and line current over-limit degree; R_quality(t) is the power quality sub-reward function, whose value is negatively correlated with voltage harmonic distortion rate; the weight coefficients α(t), β(t), and γ(t) are generated online in real time by the fuzzy inference system, and the input variables of the fuzzy inference system are the normalized load index characterizing the overall operating pressure of the system and the normalized voltage deviation index characterizing the operating safety risk of the system.

[0011] As a further preferred embodiment of the present invention: each local agent deployed in the edge computing layer integrates a security constraint layer module; the security constraint layer module is configured to: during execution, receive the original control action a_raw output by the local agent policy network, and perform online verification and correction of the original control action based on the optimization model constructed based on Lyapunov stability theory, to ensure that the final executed action a_safe satisfies the Lyapunov stability condition ΔV(x, a) = V(x_{t+1})-V(x_t)≤-ηV(x_t), where V(x) is a Lyapunov function characterizing the stability of the power grid, and its value is a function of the system state x, and η is a small positive constant used to control the convergence rate.

[0012] As a further preferred embodiment of the present invention: the optimization model executed by the security constraint layer module is: min||a_safe-a_raw||² stΔV(x,a_safe)≤-ηV(x_t) Where a_raw is the original control action, a_safe is the modified safety action, and the optimization model solves for a safety action that is closest to the original action and satisfies the stability constraints.

[0013] As a further preferred embodiment of the present invention: the commentator network in the central coordinating agent adopts a graph attention network structure; the graph attention network models the power grid topology as a graph G=(N, E), where node N represents a local agent and edge E represents the electrical connection relationship between nodes; the graph attention network dynamically aggregates the observation information h_j of neighboring agents by calculating the attention weight α_ij between nodes i and j, so as to accurately evaluate the global value function; the attention weight α_ij is calculated by the formula α_ij = exp(e_ij) / Σ_k exp(e_ik), where e_ij =LeakyReLU( a^T [ W h_i || W h_j ] ), and W and a are learnable parameter matrices.

[0014] Furthermore, the present invention also provides an adaptive control method for smart grids, comprising the following steps: Initialize the power grid digital twin model and the neural network parameters of each intelligent agent in a cloud-based hyper-realistic digital twin platform; The physical perception layer continuously collects and uploads data, driving the digital twin model to remain synchronized with the physical power grid. In the digital twin virtual environment, the central coordinating agent organizes local agents to conduct multi-agent reinforcement learning collaborative training. After training, the policy network parameters are distributed to each edge computing layer through the communication network. Based on the distributed policy network and local real-time observation data, each local agent generates control commands, which are then verified by safety constraints before being executed to control local power equipment. Based on the feedback performance data of the physical power grid, the online adaptive update cycle of model parameters is triggered.

[0015] As a further preferred embodiment of the present invention: in the multi-agent reinforcement learning collaborative training step, the central coordinating agent dynamically adjusts the exploration rate ε using a simulated annealing algorithm, and its annealing plan is related to the intensity of power grid disturbance simulated in the digital twin model: when simulating a high disturbance scenario, the exploration rate is reduced to utilize the current optimal strategy, and when simulating a stable scenario, the exploration rate is increased to explore a better strategy.

[0016] As a further preferred embodiment of the present invention: the condition for triggering the online adaptive update of the model parameters is: detecting a trend of deterioration in the overall system performance indicators within a continuous time window, or detecting a permanent change event in the power grid topology; the trend of deterioration refers to the continuous decline of the moving average of the global reward function characterizing system performance for more than 3 training cycles within a continuous 24-hour time window; the permanent change event refers to a signal issued by the power grid energy management system (EMS) or the dispatch automation system indicating an irreversible change in the power grid topology, and the duration exceeds 5 minutes.

[0017] As a further preferred embodiment of the present invention: in the training step, the experience replay pool adopts a priority experience replay mechanism, the priority p of which is determined by the formula p=|δ|+λ|ΔV|, where δ is the temporal difference error, ΔV is the change of the Lyapunov function, and λ is the weight coefficient; experience samples with higher priority p are assigned higher sampling priority.

[0018] Compared with the prior art, the beneficial effects of the present invention are: 1. Innovative System Architecture: It pioneered a collaborative architecture of "cloud training - edge secure execution", which perfectly balances the contradiction between global optimization and real-time local control.

[0019] 2. Deep Algorithm Integration and Innovation: The Hyper-Realistic Digital Twin Model (IG-HRDTM) is constructed by deeply integrating the Physical Information Neural Network (PINN) and the Generative Adversarial Network (GAN) at the training paradigm level. This model has both physical consistency and data generalization capabilities, solving the problem of low accuracy and inability to perform forward inference in traditional models in areas with scarce data.

[0020] Dynamic intelligent reward mechanism: By dynamically adjusting the weights of multiple objectives through fuzzy reasoning, the control system can possess decision-making intelligence similar to that of an expert.

[0021] Rigid safety boundary: A safety constraint layer based on Lyapunov stability theory, providing unprecedented safety guarantees for data-driven control.

[0022] Generalization and Collaboration Capability: The use of graph attention networks enables the system to understand and model the complex relationships between agents, allowing it to generalize to different power grid topologies.

[0023] Efficient Exploration and Learning: Combining simulated annealing exploration strategies with safety- and value-based priority experience replay greatly improves training efficiency and safety.

[0024] 3. High practicality and engineering value: All high-risk training is completed in a hyper-realistic virtual environment, requiring minimal modification to existing infrastructure. Powerful functions can be achieved through software upgrades, making it highly applicable and secure. Detailed Implementation

[0025] The technical solutions in the embodiments of the present invention will be clearly and completely described below. Obviously, the described embodiments are only some embodiments of the present invention, and not all embodiments. Based on the embodiments of the present invention, all other embodiments obtained by those skilled in the art without creative effort are within the scope of protection of the present invention.

[0026] In this embodiment of the invention, a smart grid adaptive control system includes: 1. Physical sensing layer: Deployed in the actual power grid, including various sensors (PMU, smart meters) and smart terminals, used to collect voltage, current, power, frequency status quantities and circuit breaker and switch status information of each node of the power grid in real time.

[0027] 2. Edge Computing Layer: This layer consists of edge computing devices deployed in various substations, distribution feeders, or important control objects (such as microgrid gateways). Each edge computing device embeds a local agent and a local twin.

[0028] 3. Cloud-based Hyper-realistic Digital Twin Platform: Constructs a virtual power grid model synchronously mapped to the physical power grid, receives real-time data uploaded from the physical sensing layer, and drives the digital twin model to perform rolling updates. This platform includes: Simulation engine: capable of performing high-precision, faster-than-real-time power grid simulation calculations.

[0029] Central Agent: Responsible for training, updating, and distributing policies to local agents.

[0030] Experience replay pool: Used to store massive amounts of state-action-reward data generated by digital twin model simulation.

[0031] 4. Communication Network: Connects the physical sensing layer, edge computing layer, and cloud digital twin platform to achieve secure, reliable, and low-latency data transmission.

[0032] As a preferred embodiment of the present invention, the present invention also provides a smart grid adaptive control method based on the above system, comprising the following steps: Step S1: System initialization: Construct a high-fidelity power grid model in the cloud-based digital twin platform and initialize the neural network parameters of the central coordinating agent and each local agent.

[0033] Step S2: Data Synchronization and Status Awareness: The physical sensing layer collects power grid operation data in real time and uploads it to the cloud-based digital twin platform through the communication network, driving the digital twin model to keep synchronized with the physical power grid.

[0034] Step S3: Cloud-based collaborative training and policy generation: S301: The central coordinating agent organizes local agents to conduct multi-agent reinforcement learning based on the current digital twin state of the power grid.

[0035] S302: In a digital twin environment, each local agent makes control decisions based on its own observed state and interacts to form joint actions.

[0036] S303: The simulation engine calculates the next state and global reward value after the joint action.

[0037] S304: Store the (state, joint action, reward, next state) data as an experience sample in the experience replay pool.

[0038] S305: The central coordinating agent samples data from the replay pool and collaboratively trains the policy network and value network of all local agents.

[0039] S306: Periodically distribute the trained policy network parameters to the corresponding local agent in the edge computing layer.

[0040] Step S4: Distributed Real-Time Control at the Edge. Local agents in each edge computing layer independently and in parallel generate control commands based on the distributed policy network and their locally observed real-time state information. Before outputting the commands, they must be verified and corrected by a security constraint layer. Finally, the secure control commands are sent to the local device for execution.

[0041] Step S5: Online Adaptive Update. The actual operating data of the physical power grid is continuously fed back to the digital twin platform. The central coordinating agent can periodically or trigger a new round of training (return to step S3) when a significant change in the power grid operating mode is detected, update the strategy, and realize online adaptive optimization of the control system.

[0042] Example: Demonstration application of power distribution network in an industrial park 1. System Deployment Physical sensing layer: 15 sets of synchronous phasor measurement units (PMUs) compatible with the IEC 61850 standard are installed at key nodes of the 10kV distribution lines in the park to collect data such as voltage, current, phase, and power at a sampling rate of 100 frames / second.

[0043] Edge computing layer: NVIDIA Jetson AGX Orin edge computing devices are deployed in the gateway rooms of 4 photovoltaic power plants, 1 energy storage power plant and 2 important load nodes.

[0044] Cloud-based hyper-realistic digital twin platform: Built using an Amazon AWS EC2 G4dn instance. CloudPSS is used as part of the simulation engine.

[0045] Communication network: adopts a hybrid solution of "fiber optic + 5G".

[0046] 2. Construction and Optimization of the Ultra-Realistic Digital Twin Model of Smart Grid (IG-HRDTM) The construction and optimization of the IG-HRDTM model is a multi-stage iterative process. Its core lies in the deep integration of the physical law constraints of the Physical Information Neural Network (PINN) with the distributed learning capabilities of the Generative Adversarial Network (GAN) at the training paradigm level. The specific steps are as follows: High-fidelity geometric and physical model construction: Based on the CAD drawings, BIM models, and GIS data of the power grid, and utilizing laser point cloud scanning technology, centimeter-level precision 3D geometric models of substations, lines, and equipment are constructed. On this basis, a multiphysics simulation model is built, which describes physical processes such as electromagnetic transients, electromechanical transients, heat conduction, and mechanical stress based on partial differential equations (PDEs).

[0047] Fusion Architecture Design: An embedded constraint-based generative architecture is adopted as the core training paradigm. Generative Adversarial Networks (GANs) serve as the basic framework. The generator G aims to generate state data sequences that conform to the dynamics of the physical power grid; the discriminator D aims to distinguish the generated data from real historical data. Physical Information Neural Networks (PINNs) are embedded as strong regularization terms in the generator's training process. The core of PINN is a feedforward neural network U(t,x;θ) with parameter θ, which learns the solution u(t,x) of an approximate partial differential equation.

[0048] Training and optimization process: The optimization objective of the model is a composite loss function that integrates data fitting loss, physical law loss, and adversarial loss.

[0049] Physical law loss (L_physics): Substituting the output of generator G into the PINN network U, the residual of the physical equation is calculated. For the governing equation F, placement points {ti, xi} are collected within the domain Ω. The physical loss is defined as: L_physics(θ) = 1 / N_p*Σ|F(t_i, x_i,U(t_i,x_i;θ),∂U / ∂t,∇U,...)|². This term forces the generated data to satisfy physical laws.

[0050] Data distribution loss (L_adv): using the adversarial training mechanism of standard GAN: min_G max_D V(D, G) = E{x~p_data}[log D(x)]+E{z~p_z}[log(1-D(G(z)))].

[0051] Composite Loss Function and Co-optimization: The overall optimization objective L_total of the generator G is a weighted sum: L_total = λ_adv * L_adv + λ_physics * L_physics + λ_recon * L_recon. All parameters are jointly optimized using the backpropagation algorithm. The training process is essentially solving the optimization problem: θ,ω = arg min_θmax_ω L_total.

[0052] Through the above-mentioned fusion training, the final IG-HRDTM model has the ability to strictly follow the physical laws of the power grid and the ability to generate and extrapolate highly generalized uncertainties from multiple sources.

[0053] Specifically, the implementation method of the ultra-realistic digital twin model of smart grid (IG-HRDTM) is as follows: Physical Information Neural Network (PINN) Implementation: Network structure: A fully connected deep neural network (DNN) is used, with spatial coordinates x and time t as inputs and system state u (such as voltage or current) as output. The network contains 8-10 hidden layers, each with 256-512 neurons, and uses the Swish activation function.

[0054] Physical loss term L_physics: For electromagnetic transient simulations, the governing equations are a system of partial differential algebraic equations (PDAEs). Taking the nodal voltage equations as an example, the physical residuals can be constructed as follows: F_i = I_i - Σ Y_ij V_j = 0; Where I_i is the node injection current (output from PINN or calculated), Y_ij is an element of the network admittance matrix, and V_j is the node voltage (output from PINN). The physical loss is the sum of the mean squares of the residuals at all placement points. L_physics = 1 / N_p * Σ_i Σ_k |F_i(t_k, x_k)|²; Training: An adaptive weighting strategy is adopted to balance data loss and physical loss. The initial weights λ_physics are set to 0.5 ~ 1.5 and are dynamically adjusted during training based on the gradient magnitude of each loss term.

[0055] Generative Adversarial Network (GAN) Implementation: Network structure: Generator G: Inputs are a random noise vector z and a condition vector c (such as the current system state), and the output is the generated power grid data sequence. A one-dimensional transpose convolutional network is used.

[0056] Discriminator D: Input is either a real data sequence or a generated data sequence, and output is the probability that it is true. A one-dimensional convolutional network is used.

[0057] Adversarial loss L_adv: The loss function of Wasserstein GAN with Gradient Penalty (WGAN-GP) is adopted to improve training stability. ; in These are sampling points for both real and generated data.

[0058] Integration training: Composite loss function: The total loss of generator G is: L_total=L_adv+λ_physics *L_physics+λ_recon* L_recon λ_physics and λ_recon are hyperparameters that are set through grid search, with typical values ​​ranging from [0.5, 2.0] to [10, 100].

[0059] Optimizer: The Adam optimizer is used, with the generator learning rate set to 1e-4 and the discriminator learning rate set to 4e-4.

[0060] Implementation of Fuzzy Inference System (FIS) Input / output variable definitions: Input 1: Normalized load index, universe of discourse [0, 1.2], fuzzy set defined as {low, medium, high}.

[0061] Input 2: Normalized voltage deviation index, with a universe of discourse of [0, 1.5], and the fuzzy set is defined as {low, medium, high}.

[0062] Output: Economic weight adjustment factor Δβ, universe of discourse [-0.3, 0.3], fuzzy set defined as {negative large, negative small, zero, positive small, positive large}.

[0063] Membership function: All fuzzy sets use the triangular membership function.

[0064] Fuzzy rule base: Rule 1: If the load index IS is high and the voltage deviation IS is high, then ΔβIS is positive. (Significantly increases the safety weight) Rule 2: If the load index (IS) is low and the voltage deviation (IS) is low, then Δβ (IS) is negative and small. (Slightly increase the weighting of economic efficiency) Defuzzification: The centroid method is used to calculate the precise output value Δβ and to adjust the weights: β(t) = β_0 + Δβ.

[0065] Implementation of the security constraint layer Design of the Lyapunov function V(x): V(x) = 1 / 2 * Σ_i ( (V_i - V_ref) / ΔV_max )^2 + 1 / 2 * ( (f - f_ref) / Δf_max )^2 Where V_i is the node voltage, f is the system frequency, V_ref and f_ref are reference values, and ΔV_max and Δf_max are the maximum permissible deviations.

[0066] Optimized Solver: The quadratic programming (QP) problem in the security constraint layer is solved in real time using an embedded QP solver (such as OSQP or CVXGEN). This solver is deployed in C code on an edge computing device, ensuring a solution latency of less than 1ms.

[0067] Parameter settings: The stability constraint constant η is set according to the system inertia, with a typical value of 0.01 ~ 0.05.

[0068] Implementation of Graph Attention Network (GAT) Network structure: Input: Local observations h_i of each agent.

[0069] Graph structure construction: An adjacency matrix is ​​generated based on the real-time topology of the power grid, which serves as the edges of the graph.

[0070] Graph attention layer: Two GAT layers are used. The first layer has 8 attention heads and an output dimension of 64; the second layer has 1 attention head and an output dimension of 32. The ELU activation function is used.

[0071] Output: Embedded nodes that aggregate global information, which are then fed into the subsequent Critic value network.

[0072] Attention coefficient calculation: W is the learnable linear transformation weight matrix, and a is the learnable attention vector.

[0073] Training parameters for Multi-Agent Reinforcement Learning (MARL) Algorithm basis: The MADDPG (Multi-Agent Deep Deterministic Policy Gradient) algorithm framework is adopted.

[0074] Network structure: Actor Network (Policy Network): A 2-layer fully connected network with 256 neurons per layer. It uses the ReLU activation function, and the output layer uses the Tanh activation function to constrain actions within the range of [-1, 1].

[0075] Critic Network (Value Network): A two-layer fully connected network with 256 neurons in each layer, using the ReLU activation function. The central Critic receives action and observation information from all agents.

[0076] Experience replay: Priority Experience Replay (PER) is used, with priority p = |δ| + 0.1 * |ΔV|, where δ is the timing difference error. The replay pool capacity is 10^6.

[0077] Exploration strategy: Simulated annealing is used to explore noise. The initial exploration rate ε_init = 0.5, the minimum exploration rate ε_min = 0.05, and the annealing cycle is 50,000 training steps.

[0078] Hyperparameters: discount factor γ = 0.99, policy network learning rate 1e-4, value network learning rate 1e-3, soft update parameter τ = 0.01, batch size = 512.

[0079] It will be apparent to those skilled in the art that the present invention is not limited to the details of the exemplary embodiments described above, and that the invention can be implemented in other specific forms without departing from the spirit or essential characteristics of the invention. Therefore, the embodiments should be considered in all respects as exemplary and non-limiting, and the scope of the invention is defined by the appended claims rather than the foregoing description. Thus, it is intended that all variations falling within the meaning and scope of equivalents of the claims be included within the present invention.

[0080] Furthermore, it should be understood that although this specification describes embodiments, not every embodiment contains only one independent technical solution. This narrative style is merely for clarity. Those skilled in the art should consider the specification as a whole, and the technical solutions in each embodiment can also be appropriately combined to form other embodiments that can be understood by those skilled in the art.

Claims

1. A smart grid adaptive control system, characterized in that, include: Physical sensing layer, edge computing layer, cloud-based hyper-realistic digital twin platform, and communication network; The physical sensing layer is used to collect electrical measurement data and switch status information of each node in the power grid in real time. The edge computing layer is distributed across multiple key nodes in the power grid. These key nodes include connection nodes or power grid hub nodes where the controlled power equipment is located. Each node deploys a local agent and a local twin. The local agent is used to load and execute control strategies to perform distributed real-time control of the local power equipment. The local twin has a built-in linearized or reduced-order physical equation calculation engine, which is used to perform millisecond-level fast calculation and deduction on local events such as voltage deviation, frequency deviation, and load mutation, providing the local agent with fast control command suggestions without cloud intervention. The cloud-based hyper-realistic digital twin platform includes: A highly realistic digital twin model of a smart grid is constructed and optimized by integrating the physical law constraints of the Physical Information Neural Network (PINN) with the distributed learning capabilities of the Generative Adversarial Network (GAN) through a training paradigm. The construction and optimization process includes: adopting a generative architecture with embedded constraints, using the GAN as the basic framework, and embedding the PINN as a strong regularization term into the generator's training process; and co-training by optimizing the composite loss function L_total = λ_adv * L_adv + λ_physics * L_physics, where L_adv is the adversarial loss and L_physics is the physical law loss. A central coordinating agent is configured in the virtual environment provided by the digital twin model and uses a centralized training distributed execution framework to conduct multi-agent reinforcement learning collaborative training on the multiple local agents. An experience replay pool is used to store training data generated by digital twin simulations; The communication network connects the physical sensing layer, the edge computing layer, and the cloud-based hyper-realistic digital twin platform, and is used to realize data exchange and policy parameter distribution.

2. The system according to claim 1, characterized in that, The reward function used by the central coordinating agent during training is a dynamic multi-objective weighted function, in the form: R(t) = α(t) * R_economic(t). +β(t)*R_security(t)+γ(t)*R_quality(t); where R_economic(t) is the economic reward function, whose value is negatively correlated with grid losses and control costs; R_security(t) is the security reward function, whose value is negatively correlated with node voltage deviation and line current over-limit; R_quality(t) is the power quality reward function, whose value is negatively correlated with voltage harmonic distortion rate; the weight coefficients α(t), β(t), and γ(t) are generated online in real time by the fuzzy inference system, which adopts a dual-input single-output structure. The input variables are the normalized load index, which represents the overall operating pressure of the system, and the normalized voltage deviation index, which represents the operating safety risk of the system. Its preset fuzzy rules include: if the normalized load index is high and the normalized voltage deviation index is high, then the output adjustment factor is used to significantly increase the weight coefficient β(t) of the security reward function.

3. The system according to claim 1, characterized in that, Each local agent deployed in the edge computing layer integrates a security constraint layer module. The security constraint layer module is configured to: receive the original control action a_raw output by the local agent policy network during execution, and perform online verification and correction of the original control action based on the optimization model constructed based on Lyapunov stability theory, to ensure that the final executed action a_safe satisfies the Lyapunov stability condition ΔV(x, a) = V(x_{t+1})-V(x_t)≤-ηV(x_t), where V(x) is the Lyapunov function characterizing the stability of the power grid, and its value is a function of the system state x, and η is a small positive constant used to control the convergence rate.

4. The system according to claim 3, characterized in that, The optimization model executed by the security constraint layer module is as follows: min||a_safe-a_raw||² stΔV(x,a_safe)≤-ηV(x_t) Where a_raw is the original control action, a_safe is the modified safety action, and the optimization model solves for a safety action that is closest to the original action and satisfies the stability constraints.

5. The system according to claim 1, characterized in that, The commentator network in the central coordinating agent adopts a graph attention network structure. The graph attention network models the power grid topology as a graph G=(N, E), where node N represents a local agent and edge E represents the electrical connection between nodes. The graph attention network dynamically aggregates the observation information h_j of neighboring agents by calculating the attention weight α_ij between nodes i and j to accurately evaluate the global value function. The attention weight α_ij is calculated by the formula α_ij = exp(e_ij) / Σ_k exp(e_ik), where e_ij = LeakyReLU( a^T [ W h_i || W h_j ] ), and W and a are learnable parameter matrices.

6. A smart grid adaptive control method based on the system described in any one of claims 1-5, characterized in that, Includes the following steps: Initialize the power grid digital twin model and the neural network parameters of each intelligent agent in a cloud-based hyper-realistic digital twin platform; The physical perception layer continuously collects and uploads data, driving the digital twin model to remain synchronized with the physical power grid. In the digital twin virtual environment, the central coordinating agent organizes local agents to conduct multi-agent reinforcement learning collaborative training. After training, the policy network parameters are distributed to each edge computing layer through the communication network. Based on the distributed policy network and local real-time observation data, each local agent generates control commands, which are then verified by safety constraints before being executed to control local power equipment. Based on the feedback performance data of the physical power grid, the online adaptive update cycle of model parameters is triggered.

7. The method according to claim 6, characterized in that, In the multi-agent reinforcement learning collaborative training step, the central coordinating agent dynamically adjusts the exploration rate ε using a simulated annealing algorithm. Its annealing plan is related to the intensity of power grid disturbances simulated in the digital twin model: reducing the exploration rate to utilize the current optimal strategy when simulating high disturbance scenarios, and increasing the exploration rate to explore better strategies when simulating stable scenarios.

8. The method according to claim 6, characterized in that, The conditions for triggering online adaptive updates of the model parameters are: detecting a trend of overall system performance degradation within a continuous time window, or detecting a permanent change event in the power grid topology; the trend of degradation refers to a continuous decrease in the moving average of the global reward function characterizing system performance for more than 3 training cycles within a continuous 24-hour time window; the permanent change event refers to a signal issued by the power grid energy management system (EMS) or dispatch automation system indicating an irreversible change in the power grid topology, and lasting for more than 5 minutes.

9. The method according to claim 6, characterized in that, In the training step, the experience replay pool adopts a priority experience replay mechanism, and its priority p is determined by the formula p=|δ|+λ|ΔV|, where δ is the temporal difference error, ΔV is the change of Lyapunov function, and λ is the weight coefficient; experience samples with higher priority p values ​​are assigned higher sampling priority.