Renewable energy power grid voltage control method, device and system based on machine learning and medium

Abstract

The invention discloses a renewable energy power grid voltage control method, device and system based on machine learning and a medium, and aims to solve the problems of power grid voltage fluctuation and out-of-limit caused by high-permeability renewable energy access. The method comprises the following steps: firstly, collecting power grid operation parameters, denoising, complementing and standardizing through Kalman filtering; constructing and training a GNN model in combination with power grid topology to realize real-time prediction of voltage deviation; then, an MARL decision framework with stable voltage, low consumption and high efficiency as targets is built, and an intelligent agent is trained to complete decentralized collaborative decision; driving a voltage regulating device to execute regulation based on a GNN prediction result and an MARL optimal strategy, and adapting to the dynamic change of the power grid through an interactive feedback mechanism; and finally, continuously monitoring the state of the power grid, and triggering iterative optimization of the model and the strategy. Through multi-module cooperation and dynamic iterative optimization, the power grid voltage stability is effectively maintained, the voltage fluctuation is suppressed, and the operation reliability and the electric energy quality of the smart power grid are improved while the efficient consumption of renewable energy sources is ensured.

Description

Technical Field

[0001] This invention relates to the field of distribution network voltage management, specifically a method, device, system, and medium for controlling the voltage of renewable energy power grids based on machine learning. Background Technology

[0002] With the acceleration of the global energy transition, the penetration rate of renewable energy sources (RES) such as wind and solar power in the power system continues to increase, gradually becoming an important part of the power supply. As the penetration rate of renewable energy in the power grid increases, the intermittency of solar and wind power, combined with load fluctuations, leads to frequent over-limits in grid voltage, threatening power quality and system stability.

[0003] Traditional voltage control methods mainly include Volt / VAR control (VVC), model predictive control (MPC), particle swarm optimization (PSO), genetic algorithm (GA), and fuzzy logic control (FLC). While these methods have played a role in traditional deterministic power grids, they have significant limitations in power grids with high penetration of renewable energy: First, the computational complexity of heuristic optimization methods (such as PSO and GA) increases exponentially with the scale of the power grid and the number of control variables, making it difficult to meet real-time control requirements; second, model predictive control relies on accurate mathematical models of the power grid, but the stochastic nature of renewable energy causes dynamic changes in the grid model parameters, which can easily lead to prediction bias and reduce control accuracy; third, traditional methods often adopt centralized control architectures, which pose communication delays and single-point failure risks in distributed power grids, and are difficult to adapt to dynamic adjustments in the power grid topology.

[0004] In recent years, the application of artificial intelligence technology in the power system field has provided a new direction for voltage control. Among them, deep reinforcement learning (DRL) has shown potential in voltage regulation by optimizing control strategies through the interaction between agents and the environment. However, single-agent DRL suffers from insufficient scalability in large-scale power grids and difficulty in handling multi-node collaborative control. At the same time, existing DRL methods often ignore the impact of power grid topology on voltage propagation and cannot accurately capture the spatial dependencies between nodes, resulting in a lack of global optimization capabilities in the control strategy.

[0005] Graph Neural Networks (GNNs), as deep learning models for graph-structured data, possess inherent topological feature learning capabilities, effectively capturing spatial relationships between grid nodes and providing technical support for voltage fluctuation prediction. Multi-Agent Reinforcement Learning (MARL), through distributed collaborative decision-making by multiple agents, can adapt to the control needs of distributed grids and overcome the limitations of centralized control. However, currently, no technical solution deeply integrates the topological perception and prediction capabilities of GNNs with the distributed collaborative decision-making capabilities of MARLs to construct an intelligent voltage control framework covering the entire process from data preprocessing to voltage prediction, decision optimization, action execution, and feedback iteration. This fails to fully meet the precise, real-time, and adaptive voltage control requirements of high-penetration renewable energy grids. Therefore, this invention proposes a machine learning-based voltage control method for renewable energy grids, addressing the shortcomings of traditional methods through the synergistic effect of both approaches, and providing a technical solution for the stable and efficient operation of modern power grids. Summary of the Invention

[0006] The core objective of this invention is to overcome the problems of insufficient adaptability, low control accuracy, poor real-time performance, and high power loss of existing voltage control methods in renewable energy grids with high penetration rates. It provides a voltage control method, device, system, and medium for renewable energy grids based on machine learning, which realizes accurate prediction of grid voltage, distributed collaborative decision-making, dynamic adjustment, and closed-loop optimization, ensuring the long-term stable operation of the grid under conditions of renewable energy fluctuations and load changes, while reducing system power loss and improving grid operating efficiency.

[0007] The technical solution of the present invention is as follows:

[0008] A machine learning-based method for voltage control of renewable energy power grids includes the following steps:

[0009] Step 1: Real-time acquisition of target power grid operating parameters containing renewable energy (RES), and the use of Kalman filtering algorithm to suppress data noise and fill in missing values ​​to obtain a standardized power grid status dataset;

[0010] Step 2: Based on the power grid topology and the standardized state dataset obtained in Step 1, construct a graph neural network (GNN) model to learn the spatial dependencies between power grid nodes and the voltage variation patterns, and train the model to output the voltage deviation prediction results for each node.

[0011] Step 3: With the goals of voltage stability and minimum power loss, construct a multi-agent reinforcement learning (MARL) decision framework, define agent types, state space, action space, and multi-objective reward function, and train the agents to output the optimal strategy for decentralized collaborative control based on the voltage deviation prediction results from Step 2.

[0012] Step 4: Based on the voltage deviation prediction results from Step 2 and the optimal control strategy output from Step 3, drive the voltage control equipment in the power grid to perform regulation actions. At the same time, establish an interactive feedback mechanism between GNN and MARL to feed back the power grid operating status data and the effect of control actions to GNN and MARL in real time, and dynamically update the prediction parameters and decision-making strategies.

[0013] Step 5: Continuously monitor the grid voltage level, power flow status, and equipment operating parameters. If voltage exceeds limits or system instability occurs, trigger the updating of the GNN model parameters in Step 2 and the MARL model parameters in Step 3, and continuously iterate and improve the control strategy to ensure long-term stable operation of the grid.

[0014] Furthermore, in step 1, the target power grid operating parameters collected include: voltage at each node. Active power between nodes reactive power Line resistance Line reactance Active and reactive power of node loads , Renewable energy injection of active and reactive power , Phase angle at nodes and line admittance The state update formula for the Kalman filter algorithm is:

[0015] ;

[0016] in, For Kalman gain, Let H be the prediction error covariance from time t-1 to time t, H be the observation matrix, and R be the measurement noise covariance. This formula enables the suppression of high-frequency noise in power grid operation data and the accurate completion of missing observations.

[0017] Furthermore, in step 2, the GNN model adopts a Graph Convolutional Network (GCN) architecture. The model input is the feature vector of the power grid nodes, including node voltage, active power, reactive power, admittance, and phase angle. The model output is the predicted voltage deviation value of each node. During model training, the Adam optimizer is used with a learning rate of 0.001. The batch size is adapted to the power grid scale from 32 to 128, and the training epochs are 100 to 500 to ensure that the model fully learns the topological features of the power grid and the voltage change pattern. At the same time, the GNN model adapts to the potential changes in the power grid topology in real time through the dynamic adjustment of the edge weights between nodes, thereby improving the model's adaptability to complex power grid scenarios.

[0018] Furthermore, in step 3, the MARL decision framework employs the Multi-Agent Deep Deterministic Policy Gradient (MADDPG) algorithm. The agents include: a capacitor bank control agent, an on-load tap-changing transformer control agent, a distributed energy (DER) inverter control agent, and a battery energy storage system (BESS) control agent. The state space is a set of real-time grid operating states, including node voltage, power flow, renewable energy output, and load demand. The action space is a set of adjustment quantities for each control device, including capacitor bank reactive power compensation, transformer tap adjustment, BESS charging and discharging power, and DER inverter reactive power output. The multi-objective reward function is defined as:

[0019] ;

[0020] in, Let be the actual voltage at the node at time t. The rated voltage of the power grid is 1.0 pu. Let t be the instantaneous power loss of the system. Let t be the total power consumed by the system at time t. Let be the change in node voltage between time t and time t-1. , , The weighting coefficients are used to balance the three objectives of voltage stability, power loss optimization, and voltage fluctuation suppression; the optimal policy function of the MARL framework is:

[0021]

[0022] in, The optimal control strategy for agent i. The output of the action of agent i at grid node j. Let j be the state of node j perceived by agent i. This is a discount factor used to balance short-term and long-term control returns. The instantaneous reward for agent i at time t.

[0023] Furthermore, in step 4, the voltage control equipment includes: a capacitor bank, an on-load tap-changing transformer, a distributed energy source (DER) inverter, and a battery energy storage system (BESS); during the execution of the control strategy, dynamic voltage adjustment is achieved based on the following mathematical model:

[0024] A node voltage deviation calculation model is used to quantify the impact of renewable energy injection and load changes on node voltage. The formula is as follows:

[0025] ;

[0026] in, Let be the voltage at node i. Let j be the voltage of node i. This is the voltage reference value. , Let be the resistance and reactance of the line between nodes i and j, respectively. , These represent the active and reactive power demands of the load at node j, respectively. , These represent the active and reactive power injections of renewable energy at node j, respectively.

[0027] A reactive power compensation calculation model is used to determine the amount of reactive power compensation required from capacitor banks and BESS (Block Energy Saving System) to balance power imbalance and stabilize voltage. The formula is as follows:

[0028] ;

[0029] in, This represents the total reactive power compensation required by the system. Let i be the equivalent reactance of node i;

[0030] A voltage sensitivity analysis model is used to assess the impact of power changes on voltage deviation, guiding the precise adjustment of the control strategy. The formula is:

[0031] ;

[0032] in, , Let be the voltage deviation and phase angle deviation at node i, respectively. , These represent the changes in active power and reactive power at node i, respectively.

[0033] The voltage update model is used to predict the node voltage at the next time step, thereby adjusting the control action. The formula is as follows:

[0034] ;

[0035] in, , Let the voltage at node i be at time t+1 and time t. Let be the line admittance between nodes i and j. Let i be the set of neighboring nodes of node i. This refers to the transformer tap adjustment amount. The coefficient representing the effect of tap adjustment on voltage. This is the reactive power regulation quantity. The coefficient representing the impact of reactive power regulation on voltage. Let be the amount of active power injected by BESS at node i. The effect coefficient of BESS active power injection on voltage. Let be the reactive power injection amount of the capacitor bank at node i. The coefficient representing the impact of reactive power injection from the capacitor bank on voltage.

[0036] The control action execution model, used to correlate GNN prediction results with MARL decisions, outputs the voltage regulation amount, as shown in the formula:

[0037] ;

[0038] ;

[0039] in, for Voltage regulation at any given time. The voltage control prediction function of the GNN. , , , Let be the node i voltage, active power, reactive power, and admittance at time t, which are related to voltage control. Let be the target control voltage of node i at time t+1. Let i be the phase angle of node i at time t+1. , where is the weight coefficient of the agent's policy, and N is the total number of power grid nodes;

[0040] The interactive feedback mechanism between GNN and MARL dynamically updates the prediction parameters of GNN and the agent strategy of MARL by feeding back the power grid operation status data and the execution effect of control actions to the GNN model and MARL framework in real time, thus ensuring that the control method adapts to changes in the power grid operation status in real time.

[0041] Furthermore, in step 5, the indicators for monitoring whether the system is unstable include: node voltage deviation ≤ ±0.05 pu, system power loss rate ≤ preset threshold, voltage recovery time ≤ preset response time threshold, and whether the equipment operating parameters are within a safe range. The equipment operating parameters include transformer tap position, capacitor bank switching status, and BESS charging and discharging power. If the monitored indicators exceed the threshold, the model parameters are triggered to update, including GNN model parameters such as training iteration count and learning rate, and MARL model parameters such as reward function weight coefficient, discount factor, and exploration strategy parameters. At the same time, the equipment adjustment priority and action step size in the control strategy are optimized to achieve continuous iterative improvement of the control effect.

[0042] Furthermore, the renewable energy sources include wind energy, solar energy, and biomass energy; the target power grid includes distributed power grids, smart grids, and microgrids; the method uses tools such as MATLAB and Python to build a simulation verification environment, and relies on power grid real-time monitoring systems and Supervisory Control and Data Acquisition (SCADA) systems to achieve engineering applications.

[0043] A machine learning-based renewable energy grid voltage control device includes:

[0044] The data acquisition and preprocessing module is used to collect target power grid operating parameters containing renewable energy (RES) in real time, and to use the Kalman filter algorithm to suppress data noise and fill in missing values ​​to obtain a standardized power grid status dataset.

[0045] The Graph Neural Network (GNN) model construction and training module is used to build a Graph Neural Network (GNN) model based on the power grid topology and standardized state dataset, learn the spatial dependence between power grid nodes and the voltage change law, and train the model to output the voltage deviation prediction results of each node.

[0046] The module for constructing and training a multi-agent reinforcement learning (MARL) decision framework is used to construct a multi-agent reinforcement learning (MARL) decision framework with the goals of voltage stability and minimum power loss. It defines agent types, state space, action space and multi-objective reward function, and trains agents to output the optimal strategy for decentralized collaborative control based on voltage deviation prediction results.

[0047] The strategy execution and feedback module is used to drive the voltage control equipment in the power grid to perform adjustment actions based on the voltage deviation prediction results and the optimal control strategy. At the same time, it establishes an interactive feedback mechanism between GNN and MARL to feed back the power grid operation status data and the effect of control action execution to GNN and MARL in real time, and dynamically update the prediction parameters and decision-making strategies.

[0048] The model update and strategy iteration module is used to continuously monitor the grid voltage level, power flow status and equipment operating parameters. If voltage exceeds the limit or system instability occurs, it triggers the updating of GNN model parameters and MARL model parameters, as well as the continuous iteration and improvement of control strategies to ensure the long-term stable operation of the grid.

[0049] A machine learning-based renewable energy grid voltage control system includes: a computer-readable storage medium and a processor;

[0050] The computer-readable storage medium is used to store executable instructions;

[0051] The processor is used to read executable instructions stored in the computer-readable storage medium and execute the machine learning-based renewable energy grid voltage control method.

[0052] A non-transitory computer-readable storage medium having a computer program stored thereon, which, when executed by a processor, implements the aforementioned machine learning-based renewable energy grid voltage control method.

[0053] Compared with the prior art, the present invention has the following beneficial effects:

[0054] This invention deeply integrates Graph Neural Networks (GNNs) and Multi-Agent Reinforcement Learning (MARL) to form an intelligent voltage control framework adapted to renewable energy grids. GNNs can accurately capture the spatiotemporal dependencies of the power grid, achieving more precise voltage state perception and prediction based on the grid topology. MARLs, on the other hand, enable decentralized collaborative decision-making, dynamically adjusting control strategies to cope with voltage fluctuations. The combination of these two approaches effectively overcomes the limitations of traditional optimization methods in nonlinear and dynamic grids. This framework significantly improves the real-time performance and adaptability of voltage control, optimizes grid voltage stability, and reduces system power loss. It provides an efficient solution for grid voltage management in scenarios with high renewable energy penetration, and possesses good scalability and practical application potential. Its adaptability to actual grids can be verified using existing simulation tools. Attached Figure Description

[0055] Figure 1 This is a flowchart of a machine learning-based renewable energy grid voltage control method according to the present invention;

[0056] Figure 2 Topology diagram of a 14-node system;

[0057] Figure 3 For voltage control actions of a 14-node system based on GNN-MARL;

[0058] Figure 4 The power loss diagram of the 14-node system provided by this invention. Detailed Implementation

[0059] To make the objectives, technical solutions, and advantages of the embodiments of the present invention clearer, the technical solutions of the embodiments of the present invention will be clearly and completely described below with reference to the accompanying drawings. Obviously, the described embodiments are only some embodiments of the present invention, 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.

[0060] Please see Figure 1 The first aspect of this invention provides a renewable energy grid voltage control method based on machine learning, the specific steps of which are as follows:

[0061] Step 1: Real-time acquisition of target power grid operating parameters containing renewable energy (RES), and the use of Kalman filtering algorithm to suppress data noise and fill in missing values ​​to obtain a standardized power grid status dataset;

[0062] The collected power grid operating parameters include: voltage at each node. Active power between nodes reactive power Line resistance Line reactance Active and reactive power of node loads , Renewable energy injection of active and reactive power , Phase angle at nodes and line admittance The Kalman filter state update formula is:

[0063]

[0064] in, For Kalman gain, Let H be the prediction error covariance from time t-1 to time t, H be the observation matrix, and R be the measurement noise covariance. This formula enables the suppression of high-frequency noise in power grid operation data and the accurate completion of missing observations.

[0065] Step 2: Based on the power grid topology and preprocessed state data (standardized state dataset obtained in Step 1), construct a graph neural network (GNN) model to learn the spatial dependencies between power grid nodes and the voltage change patterns, and train the model to achieve real-time prediction of voltage deviations at each node of the power grid.

[0066] The GNN model employs a Graph Convolutional Network (GCN) architecture. The model input consists of feature vectors of power grid nodes, including node voltage, active power, reactive power, admittance, and phase angle. The model output is the predicted voltage deviation for each node. During training, the Adam optimizer is used with a learning rate of 0.001. The batch size is adapted to the power grid scale (32-128), and the training epochs are 100-500 to ensure the model fully learns the power grid topology features and voltage variation patterns. Simultaneously, the GNN model dynamically adjusts the edge weights between nodes (proportional to the line admittance value) to adapt to potential changes in the power grid topology in real time, improving the model's adaptability to complex power grid scenarios.

[0067] Step 3: With the goals of voltage stability and minimum power loss, construct a multi-agent reinforcement learning (MARL) decision framework, define agent types, state space, action space and multi-objective reward function, and combine the voltage deviation prediction results from Step 2 to train the agents to output the optimal strategy for decentralized collaborative control.

[0068] The MARL framework employs the Multi-Agent Deep Deterministic Policy Gradient (MADDPG) algorithm. The agents include: a capacitor bank control agent, an on-load tap-changing transformer control agent, a distributed energy (DER) inverter control agent, and a battery energy storage system (BESS) control agent. The state space is a set of real-time grid operating states, including node voltage, power flow, renewable energy output, and load demand. The action space is a set of adjustment parameters for each control device, including capacitor bank reactive power compensation, transformer tap adjustment, BESS charging and discharging power, and DER inverter reactive power output. The multi-objective reward function is defined as follows:

[0069]

[0070] in, Let be the actual voltage at the node at time t. The rated voltage of the power grid is 1.0 pu. Let t be the instantaneous power loss of the system. Let t be the total power consumed by the system at time t. Let be the change in node voltage between time t and time t-1. , , These are weighting coefficients (ranging from 0 to 1, dynamically adjustable based on grid operation priorities), used to balance the three objectives of voltage stability, power loss optimization, and voltage fluctuation suppression; the optimal strategy function of the MARL framework is:

[0071]

[0072] in, The optimal control strategy for agent i. The output of the action of agent i at grid node j. Let j be the state of node j perceived by agent i. This is a discount factor used to balance short-term and long-term control returns. The instantaneous reward for agent i at time t.

[0073] Step 4: Based on the voltage deviation prediction results of the GNN model and the optimal control strategy output by the MARL framework, drive the voltage control equipment in the power grid to perform regulation actions. At the same time, establish an interactive feedback mechanism between GNN and MARL to feed back the power grid operation status data and the effect of control actions to GNN and MARL in real time, dynamically update prediction parameters and decision strategies, and adapt to changes in power grid operation status in real time.

[0074] Voltage control equipment includes: capacitor banks, on-load tap-changing transformers, distributed energy resources (DER) inverters, and battery energy storage systems (BESS); during the execution of the control strategy, dynamic voltage regulation is achieved based on the following mathematical model:

[0075] A node voltage deviation calculation model is used to quantify the impact of renewable energy injection and load changes on node voltage. The formula is as follows:

[0076]

[0077] in, Let be the voltage at node i. Let j be the voltage of node i. This is the voltage reference value (1.0 pu). , Let be the resistance and reactance of the line between nodes i and j, respectively. , These represent the active and reactive power demands of the load at node j, respectively. , These represent the active and reactive power injections of renewable energy at node j, respectively.

[0078] A reactive power compensation calculation model is used to determine the amount of reactive power compensation required from capacitor banks and BESS (Block Energy Saving System) to balance power imbalance and stabilize voltage. The formula is as follows:

[0079]

[0080] in, This represents the total reactive power compensation required by the system. Let i be the equivalent reactance of node i;

[0081] Voltage sensitivity analysis model. Used to assess the impact of power changes on voltage deviation, guiding precise adjustment of control strategies. The formula is:

[0082]

[0083] in, , Let be the voltage deviation and phase angle deviation at node i, respectively. , These represent the changes in active power and reactive power at node i, respectively.

[0084] The voltage update model is used to predict the node voltage at the next time step, thereby adjusting the control action. The formula is as follows:

[0085]

[0086] in, , Let the voltage at node i be at time t+1 and time t. Let be the line admittance between nodes i and j. Let i be the set of neighboring nodes of node i. This refers to the transformer tap adjustment amount. The coefficient representing the effect of tap adjustment on voltage. This is the reactive power regulation quantity. The coefficient representing the impact of reactive power regulation on voltage. Let be the amount of active power injected by BESS at node i. The effect coefficient of BESS active power injection on voltage. Let be the reactive power injection amount of the capacitor bank at node i. The coefficient representing the impact of reactive power injection from the capacitor bank on voltage.

[0087] The control action execution model, used to correlate GNN prediction results with MARL decisions, outputs the voltage regulation amount, as shown in the formula:

[0088]

[0089]

[0090] in, for Voltage regulation at any given time. The voltage control prediction function of the GNN. , , , Let be the node i voltage, active power, reactive power, and admittance at time t, which are related to voltage control. Let be the target control voltage of node i at time t+1. Let i be the phase angle of node i at time t+1. , where is the weight coefficient of the agent's policy, and N is the total number of power grid nodes;

[0091] The interactive feedback mechanism between GNN and MARL dynamically updates the prediction parameters of GNN and the agent strategy of MARL by feeding back the power grid operation status data and the execution effect of control actions to the GNN model and MARL framework in real time, thus ensuring that the control method adapts to changes in the power grid operation status in real time.

[0092] Step 5: Continuously monitor the grid voltage level, power flow status, and equipment operating parameters. If voltage exceeds limits or system instability occurs, trigger model parameter updates and control strategy iterations to ensure long-term stable operation of the grid.

[0093] The system stability monitoring indicators include: node voltage deviation ≤ ±0.05 pu, system power loss rate ≤ preset threshold, voltage recovery time ≤ preset response time threshold, and equipment operating parameters, such as transformer tap position, capacitor bank switching status, and BESS charging and discharging power must be within the safe range. If the monitored indicators exceed the threshold, the model parameters are updated, including adjusting the number of training iterations and learning rate of GNN, as well as the reward function weight coefficient, discount factor, and exploration strategy parameters of MARL. At the same time, the equipment adjustment priority and action step size in the control strategy are optimized to achieve continuous iterative improvement of control performance.

[0094] The proposed machine learning-based voltage control method for renewable energy power grids is simulated and analyzed in a 33kV radial distribution network system with a 14-node system. (14-node system topology) Figure 2 As shown, a 2MW solar photovoltaic power station is installed at node 1, a 3MW wind farm is installed at node 12, and a 1.5MW battery energy storage system is installed at node 5.

[0095] from Figure 3 As can be seen, the agent maintains the voltage level within the target range by dynamically adjusting reactive power injection. Compared with traditional control methods, this method reduces voltage exceedances by 30%, verifying the applicability of multi-agent reinforcement learning in complex power grid scenarios. Voltage deviations are well controlled within the safe operating range of 0.95-1.05 per unit. The observed control actions ensure overall grid voltage stability and reduce overvoltage and undervoltage situations.

[0096] from Figure 4 It can be seen that the method of the present invention has an advantage of about 18% in terms of loss reduction compared with the traditional control method. By intelligently allocating reactive power and dynamically adjusting voltage, it improves the efficiency and stability of the power grid.

[0097] Table 1. Comparative Analysis of the Method Proposed in the Invention and Existing Methods

[0098]

[0099] As shown in Table 1, in real-time simulations under various dynamic grid conditions, the GNN-MARL method outperforms the baseline method and the GNN control method. The voltage deviation of the GNN-MARL method is reduced to 0.35, significantly lower than the baseline method's 0.85 and the GNN control method's 0.5, indicating a substantial improvement in voltage stability under dynamic grid conditions. The power loss reduction effect is also significantly enhanced, reaching 0.35 for the GNN-MARL method, compared to 0.2 for the GNN control method and 0.1 for the baseline method, reflecting higher energy efficiency. The response time of the GNN-MARL method is 30% faster than the GNN control method, with a response time of 0.25, compared to 0.35 for the GNN control method and 0.5 for the baseline method, indicating that the algorithm can recover voltage levels more quickly and improve the quality of real-time voltage control. The GNN-MARL method achieves faster voltage stabilization, reduces power loss, and maintains grid stability under dynamic conditions, making it the most effective method for achieving voltage stabilization and power loss reduction, and enabling faster control under dynamic grid conditions.

[0100] This invention addresses the problems of poor voltage stability and high power loss caused by the intermittent power output fluctuations of solar and wind power sources after high penetration of renewable energy (RES) into the power grid. It deeply analyzes the core mechanism of frequent voltage exceedances and points out that traditional voltage control methods suffer from computational complexity, insufficient real-time performance, and reliance on fixed models in nonlinear dynamic power grids. Furthermore, existing single intelligent algorithms struggle to capture grid topology relationships and achieve multi-device collaborative control. Therefore, this invention proposes an intelligent voltage control framework based on graph convolutional networks (GCNs) and multi-agent deep deterministic policy gradients. It preprocesses grid operation data using Kalman filtering, utilizes GCNs to learn grid topology and node spatial dependencies to achieve accurate voltage deviation prediction, and leverages MARL to treat voltage regulation devices such as capacitor banks, OLTCs, BESS, and DER inverters as independent agents. Through a multi-objective reward function, it coordinates decentralized decision-making across multiple devices, and through a closed-loop regulation process of data acquisition, prediction, decision-making, execution, and feedback, accurately maintains the voltage within a safe range and reduces power loss. This framework possesses good scalability and practical applicability, providing an innovative approach to voltage control in high RES-penetrating power grids and offering key technical support for improving renewable energy absorption capacity.

[0101] A second aspect of the present invention provides a machine learning-based renewable energy grid voltage control device, comprising:

[0102] The data acquisition and preprocessing module is used to collect target power grid operating parameters containing renewable energy (RES) in real time, and to use the Kalman filter algorithm to suppress data noise and fill in missing values ​​to obtain a standardized power grid status dataset.

[0103] The Graph Neural Network (GNN) model construction and training module is used to build a Graph Neural Network (GNN) model based on the power grid topology and standardized state dataset, learn the spatial dependence between power grid nodes and the voltage change law, and train the model to output the voltage deviation prediction results of each node.

[0104] The module for constructing and training a multi-agent reinforcement learning (MARL) decision framework is used to construct a multi-agent reinforcement learning (MARL) decision framework with the goals of voltage stability and minimum power loss. It defines agent types, state space, action space and multi-objective reward function, and trains agents to output the optimal strategy for decentralized collaborative control based on voltage deviation prediction results.

[0105] The strategy execution and feedback module is used to drive the voltage control equipment in the power grid to perform adjustment actions based on the voltage deviation prediction results and the optimal control strategy. At the same time, it establishes an interactive feedback mechanism between GNN and MARL to feed back the power grid operation status data and the effect of control action execution to GNN and MARL in real time, and dynamically update the prediction parameters and decision-making strategies.

[0106] The model update and strategy iteration module is used to continuously monitor the grid voltage level, power flow status and equipment operating parameters. If voltage exceeds the limit or system instability occurs, it triggers the updating of GNN model parameters and MARL model parameters, as well as the continuous iteration and improvement of control strategies to ensure the long-term stable operation of the grid.

[0107] Another aspect of the present invention provides a machine learning-based renewable energy grid voltage control system, comprising: a computer-readable storage medium and a processor;

[0108] The computer-readable storage medium is used to store executable instructions;

[0109] The processor is used to read executable instructions stored in the computer-readable storage medium and execute the machine learning-based renewable energy grid voltage control method described in the first aspect.

[0110] In another aspect, the present invention provides a non-transitory computer-readable storage medium having a computer program stored thereon, which, when executed by a processor, implements the machine learning-based renewable energy grid voltage control method described in the first aspect.

[0111] Those skilled in the art will understand that embodiments of the present invention can be provided as methods, systems, or computer program products. Therefore, the present invention can take the form of a completely hardware embodiment, a completely software embodiment, or an embodiment combining software and hardware aspects. Furthermore, the present invention can take the form of a computer program product embodied on one or more computer-usable storage media (including, but not limited to, disk storage, CD-ROM, optical storage, etc.) containing computer-usable program code.

[0112] This invention is described with reference to flowchart illustrations and / or block diagrams of methods, apparatus (systems), and computer program products according to embodiments of the invention. It will be understood that each block of the flowchart illustrations and / or block diagrams, and combinations of blocks in the flowchart illustrations and / or block diagrams, can be implemented by computer program instructions. These computer program instructions can be provided to a processor of a general-purpose computer, special-purpose computer, embedded processor, or other programmable data processing apparatus to produce a machine, such that the instructions, which execute via the processor of the computer or other programmable data processing apparatus, generate instructions for implementing the flowchart illustrations and / or block diagrams. Figure 1 One or more processes and / or boxes Figure 1 A device that provides the functions specified in one or more boxes.

[0113] These computer program instructions may also be stored in a computer-readable storage medium that can direct a computer or other programmable data processing device to function in a particular manner, such that the instructions stored in the computer-readable storage medium produce an article of manufacture including instruction means, which are implemented in a process Figure 1 One or more processes and / or boxes Figure 1 The function specified in one or more boxes.

[0114] These computer program instructions may also be loaded onto a computer or other programmable data processing equipment to cause a series of operational steps to be performed on the computer or other programmable equipment to produce a computer-implemented process, thereby providing instructions that execute on the computer or other programmable equipment for implementing the process. Figure 1 One or more processes and / or boxes Figure 1 The steps of the function specified in one or more boxes.

[0115] Finally, it should be noted that the above embodiments are only used to illustrate the technical solutions of the present invention and not to limit it. Although the present invention has been described in detail with reference to the above embodiments, those skilled in the art should understand that modifications or equivalent substitutions can still be made to the specific implementation of the present invention. Any modifications or equivalent substitutions that do not depart from the spirit and scope of the present invention should be covered within the scope of protection of the claims of the present invention.

Claims

1. A renewable energy grid voltage control method based on machine learning, characterized in that, Includes the following steps: Step 1: Real-time acquisition of target power grid operating parameters containing renewable energy (RES), and the use of Kalman filtering algorithm to suppress data noise and fill in missing values ​​to obtain a standardized power grid status dataset; Step 2: Based on the power grid topology and the standardized state dataset obtained in Step 1, construct a graph neural network (GNN) model to learn the spatial dependencies between power grid nodes and the voltage variation patterns, and train the model to output the voltage deviation prediction results for each node. Step 3: With the goals of voltage stability and minimum power loss, construct a multi-agent reinforcement learning (MARL) decision framework, define agent types, state space, action space, and multi-objective reward function, and train the agents to output the optimal strategy for decentralized collaborative control based on the voltage deviation prediction results from Step 2. Step 4: Based on the voltage deviation prediction results from Step 2 and the optimal control strategy output from Step 3, drive the voltage control equipment in the power grid to perform regulation actions. At the same time, establish an interactive feedback mechanism between GNN and MARL to feed back the power grid operating status data and the effect of control actions to GNN and MARL in real time, and dynamically update the prediction parameters and decision-making strategies. Step 5: Continuously monitor the grid voltage level, power flow status, and equipment operating parameters. If voltage exceeds limits or system instability occurs, trigger the updating of the GNN model parameters in Step 2 and the MARL model parameters in Step 3, and continuously iterate and improve the control strategy to ensure long-term stable operation of the grid.

2. The renewable energy grid voltage control method based on machine learning according to claim 1, characterized in that, In step 1, the target power grid operating parameters collected include: voltage at each node. Active power between nodes reactive power Line resistance Line reactance Active and reactive power of node loads , Renewable energy injection of active and reactive power , Phase angle at nodes and line admittance The state update formula for the Kalman filter algorithm is: ； in, For Kalman gain, Let H be the prediction error covariance from time t-1 to time t, H be the observation matrix, and R be the measurement noise covariance. This formula enables the suppression of high-frequency noise in power grid operation data and the accurate completion of missing observations.

3. The renewable energy grid voltage control method based on machine learning according to claim 1, characterized in that, In step 2, the GNN model adopts a Graph Convolutional Network (GCN) architecture. The model input is the feature vector of the power grid nodes, including node voltage, active power, reactive power, admittance, and phase angle. The model output is the predicted voltage deviation value of each node. During model training, the Adam optimizer is used with a learning rate of 0.

001. The batch size is adapted to the power grid scale from 32 to 128, and the number of training rounds is 100 to 500 to ensure that the model fully learns the topological features of the power grid and the voltage change pattern. At the same time, the GNN model adapts to the potential changes in the power grid topology in real time by dynamically adjusting the edge weights between nodes, thereby improving the model's adaptability to complex power grid scenarios.

4. The renewable energy grid voltage control method based on machine learning according to claim 1, characterized in that, In step 3, the MARL decision framework employs the Multi-Agent Deep Deterministic Policy Gradient (MADDPG) algorithm. The agents include: a capacitor bank control agent, an on-load tap-changing transformer control agent, a distributed energy (DER) inverter control agent, and a battery energy storage system (BESS) control agent. The state space is a set of real-time grid operating states, including node voltage, power flow, renewable energy output, and load demand. The action space is a set of adjustment quantities for each control device, including capacitor bank reactive power compensation, transformer tap adjustment, BESS charging and discharging power, and DER inverter reactive power output. The multi-objective reward function is defined as: ； in, Let be the actual voltage at the node at time t. The rated voltage of the power grid is 1.0 pu. Let t be the instantaneous power loss of the system. Let t be the total power consumed by the system at time t. Let be the change in node voltage between time t and time t-1. , , The weighting coefficients are used to balance the three objectives of voltage stability, power loss optimization, and voltage fluctuation suppression; the optimal policy function of the MARL framework is: ； in, The optimal control strategy for agent i. The output of the action of agent i at grid node j. Let j be the state of node j perceived by agent i. This is a discount factor used to balance short-term and long-term control returns. The instantaneous reward for agent i at time t.

5. The renewable energy grid voltage control method based on machine learning according to claim 1, characterized in that, In step 4, the voltage control equipment includes: a capacitor bank, an on-load tap-changing transformer, a distributed energy source (DER) inverter, and a battery energy storage system (BESS); during the execution of the control strategy, dynamic voltage adjustment is achieved based on the following mathematical model: A node voltage deviation calculation model is used to quantify the impact of renewable energy injection and load changes on node voltage. The formula is as follows: ； in, Let be the voltage at node i. Let j be the voltage of node i. This is the voltage reference value. , Let be the resistance and reactance of the line between nodes i and j, respectively. , These represent the active and reactive power demands of the load at node j, respectively. , These represent the active and reactive power injections of renewable energy at node j, respectively. A reactive power compensation calculation model is used to determine the amount of reactive power compensation required from capacitor banks and BESS (Block Energy Saving System) to balance power imbalance and stabilize voltage. The formula is as follows: ； in, This represents the total reactive power compensation required by the system. Let i be the equivalent reactance of node i; A voltage sensitivity analysis model is used to assess the impact of power changes on voltage deviation, guiding the precise adjustment of the control strategy. The formula is: ； in, , Let be the voltage deviation and phase angle deviation at node i, respectively. , These represent the changes in active power and reactive power at node i, respectively. The voltage update model is used to predict the node voltage at the next time step, thereby adjusting the control action. The formula is as follows: ； in, , Let the voltage at node i be at time t+1 and time t. Let be the line admittance between nodes i and j. Let i be the set of neighboring nodes of node i. This refers to the transformer tap adjustment amount. The coefficient representing the effect of tap adjustment on voltage. This is the reactive power regulation quantity. The coefficient representing the impact of reactive power regulation on voltage. Let be the amount of active power injected by BESS at node i. The effect coefficient of BESS active power injection on voltage. Let be the reactive power injection amount of the capacitor bank at node i. The coefficient representing the impact of reactive power injection from the capacitor bank on voltage. The control action execution model, used to correlate GNN prediction results with MARL decisions, outputs the voltage regulation amount, as shown in the formula: ； ； in, for Voltage regulation at any given time. The voltage control prediction function of the GNN. , , , Let be the node i voltage, active power, reactive power, and admittance at time t, which are related to voltage control. Let be the target control voltage of node i at time t+1. Let i be the phase angle of node i at time t+1. , where is the weight coefficient of the agent's policy, and N is the total number of power grid nodes; The interactive feedback mechanism between GNN and MARL dynamically updates the prediction parameters of GNN and the agent strategy of MARL by feeding back the power grid operation status data and the execution effect of control actions to the GNN model and MARL framework in real time, thus ensuring that the control method adapts to changes in the power grid operation status in real time.

6. The renewable energy grid voltage control method based on machine learning according to claim 1, characterized in that, In step 5, the indicators for monitoring whether the system is unstable include: node voltage deviation ≤ ±0.05 pu, system power loss rate ≤ preset threshold, voltage recovery time ≤ preset response time threshold, and whether the equipment operating parameters are within the safe range. The equipment operating parameters include transformer tap position, capacitor bank switching status, and BESS charging and discharging power. If the monitored indicators exceed the threshold, the model parameters are triggered to update, including GNN model parameters such as training iteration count and learning rate, and MARL model parameters such as reward function weight coefficient, discount factor, and exploration strategy parameters. At the same time, the equipment adjustment priority and action step size in the control strategy are optimized to achieve continuous iterative improvement of the control effect.

7. The renewable energy grid voltage control method based on machine learning according to claim 1, characterized in that, The renewable energy sources include wind energy, solar energy, and biomass energy; the target power grid includes distributed power grids, smart grids, and microgrids; the method uses tools such as MATLAB and Python to build a simulation verification environment and relies on power grid real-time monitoring systems and Supervisory Control and Data Acquisition (SCADA) systems to achieve engineering applications.

8. A renewable energy grid voltage control device based on machine learning, characterized in that, include: The data acquisition and preprocessing module is used to collect target power grid operating parameters containing renewable energy (RES) in real time, and to use the Kalman filter algorithm to suppress data noise and fill in missing values ​​to obtain a standardized power grid status dataset. The Graph Neural Network (GNN) model construction and training module is used to build a Graph Neural Network (GNN) model based on the power grid topology and standardized state dataset, learn the spatial dependence between power grid nodes and the voltage change law, and train the model to output the voltage deviation prediction results of each node. The module for constructing and training a multi-agent reinforcement learning (MARL) decision framework is used to construct a multi-agent reinforcement learning (MARL) decision framework with the goals of voltage stability and minimum power loss. It defines agent types, state space, action space and multi-objective reward function, and trains agents to output the optimal strategy for decentralized collaborative control based on voltage deviation prediction results. The strategy execution and feedback module is used to drive the voltage control equipment in the power grid to perform adjustment actions based on the voltage deviation prediction results and the optimal control strategy. At the same time, it establishes an interactive feedback mechanism between GNN and MARL to feed back the power grid operation status data and the effect of control action execution to GNN and MARL in real time, and dynamically update the prediction parameters and decision-making strategies. The model update and strategy iteration module is used to continuously monitor the grid voltage level, power flow status and equipment operating parameters. If voltage exceeds the limit or system instability occurs, it triggers the updating of GNN model parameters and MARL model parameters, as well as the continuous iteration and improvement of control strategies to ensure the long-term stable operation of the grid.

9. A renewable energy grid voltage control system based on machine learning, characterized in that, include: Computer-readable storage media and processors; The computer-readable storage medium is used to store executable instructions; The processor is used to read executable instructions stored in the computer-readable storage medium and execute the machine learning-based renewable energy grid voltage control method according to any one of claims 1-7.

10. A non-transitory computer-readable storage medium, characterized in that, It stores a computer program that, when executed by a processor, implements the machine learning-based renewable energy grid voltage control method according to any one of claims 1-7.

Images