A method for improving transient stability of new energy power grid supported by network-type energy storage

Abstract

The present application relates to the field of coordinated control of grid-forming energy storage and new energy inverters in a multi-source parallel power system, including forming a high-dimensional spatiotemporal dynamic dataset by collecting key node data of the power grid, using the dataset to drive multiple independent simulation channels in parallel, obtaining multi-perspective evaluation results of the system state, and generating system interactive coupling feature descriptions through consistency fusion and confidence calibration; further identifying oscillation energy transmission paths in a digital twin environment through distributed machine learning to generate a risk correlation graph; inputting the above features and graph into a reinforcement learning system to train a coordinated damping control strategy; finally, based on real-time measurement data, multi-channel prediction information and the strategy, using a predictive control method to dynamically adjust the control parameters of the grid-forming energy storage converter, and continuously updating the model and strategy according to the performance index. This method can effectively suppress subsynchronous oscillation caused by implicit interaction of the control loop and enhance the damping characteristics of the power grid.

Description

Technical Field

[0001] This invention relates to the field of collaborative control technology for multi-source parallel grid-connected energy storage and new energy inverters in power systems, and particularly to a method for improving the transient stability of new energy power grids considering the support of grid-connected energy storage. Background Technology

[0002] In the application scenario of multi-source parallel grid-type energy storage collaborative control technology, multi-source parallel refers to the parallel connection of various distributed energy sources such as photovoltaic, wind power and energy storage systems to the power grid through power electronic interfaces to form a complementary power supply network to improve energy utilization efficiency; grid-type energy storage refers to the energy storage converter having active support capability, simulating the characteristics of synchronous generators, and providing voltage and frequency regulation to maintain grid stability; collaborative control technology is based on centralized or distributed strategies, and uses real-time communication and algorithm optimization to coordinate multi-source parallel units to achieve power distribution, dynamic response and fault management, thereby enhancing the overall reliability and operational resilience of the system.

[0003] Existing grid-connected technologies for co-connecting grid-based energy storage and new energy power generation equipment face the following technical challenges: Grid-based energy storage inverters employ control strategies that simulate the external characteristics of synchronous generators, aiming to provide autonomous voltage and frequency support; while large-scale new energy inverters typically adopt a grid-following control mode based on phase-locked loops (PLLs). During the recovery phase after transient disturbances, the inner-loop current regulation, outer-loop power control, and voltage loop dynamic response characteristics of the two types of equipment control systems exhibit bandwidth overlap, and their nonlinear dynamics couple through the point of common coupling under weak grid conditions. This implicit interaction between multiple control loops can excite specific resonant modes. A typical example is in a power station aggregation system containing a high proportion of wind power and grid-based energy storage, where the virtual inertia response of the energy storage interacts adversely with the shaft torsional vibration mode of the wind turbine or the PLL bandwidth of the grid-connected inverter, leading to the excitation and continuous amplification of subsynchronous frequency electrical oscillations. This deteriorates the inherent damping characteristics of the system, and in severe cases, can trigger cascading failures, threatening the safe and stable operation of the power grid. Summary of the Invention

[0004] To address the shortcomings of existing technologies, this invention provides a method for improving the transient stability of new energy power grids by considering grid-based energy storage support. This method solves the technical problem that the implicit interactive coupling between grid-based energy storage and the control loop of new energy inverters causes subsynchronous oscillations during the transient recovery period, which in turn weakens the grid damping characteristics.

[0005] To solve the above-mentioned technical problems, the specific technical solution of the present invention is as follows:

[0006] The present invention provides a method for improving the transient stability of a new energy power grid that considers grid-based energy storage support, comprising:

[0007] Step 1: Collect electrical quantity data and power electronic equipment control parameters from key nodes of the power grid, and form a high-dimensional spatiotemporal dynamic dataset through time synchronization processing;

[0008] Step 2: Using the high-dimensional spatiotemporal dynamic dataset formed in Step 1, run multiple independent simulation channels in parallel, estimate the dynamic characteristics of the system based on different principles, and obtain multi-perspective evaluation results of the system state.

[0009] Step 3: The multi-view evaluation results obtained in Step 2 are subjected to cross-validation and confidence calibration using a consensus fusion algorithm to generate a description of the system's interactive coupling characteristics.

[0010] Step 4: Using the system interaction coupling feature description generated in Step 3, identify the transmission path of oscillation energy between control loops in the digital twin environment through a distributed machine learning architecture, and generate a risk association map.

[0011] Step 5: Input the system interaction coupling feature description generated in Step 3 and the risk correlation map generated in Step 4 into the reinforcement learning system, and generate a collaborative damping control strategy through reinforcement learning training.

[0012] Step 6: Based on the real-time measurement data continuously collected from Step 1, the prediction information generated by multiple independent simulation channels in Step 2, and the cooperative damping control strategy generated from Step 5, the optimal control command is generated using the predictive control method, and the control parameters of the grid-type energy storage converter are dynamically adjusted; based on the performance indicators evaluated in Step 8, the system model and cooperative damping control strategy are updated to improve transient stability.

[0013] Furthermore, the method for improving the transient stability of a new energy power grid considering grid-based energy storage, as described in this invention, includes, in step 1, the collection of electrical quantity data and power electronic equipment control parameters from key nodes of the power grid, comprising:

[0014] By deploying synchronous phasor measurement units at the grid common connection point, grid-connected energy storage converters, and new energy inverter grid-connected nodes, fundamental and harmonic phasor data of three-phase voltage and current are synchronously collected with millisecond-level time accuracy.

[0015] Capture broadband transient waveform data using a fault recording device;

[0016] The inertial time constant and damping coefficient of the virtual synchronous control loop of the grid-type energy storage converter, as well as the bandwidth and phase margin of the phase-locked loop of the new energy inverter, are read through the device local area network communication protocol.

[0017] The phasor data collected by the synchronous phasor measurement unit, the transient waveform data captured by the fault recording device, and the parameters read through the device's local area network are transmitted to the central data processing unit through a time-sensitive network according to a preset scheduling strategy.

[0018] The central data processing unit aligns and interpolates data from different spatial locations based on precise time stamps, removes outliers, and integrates and encapsulates them into a structured high-dimensional spatiotemporal dynamic dataset.

[0019] Furthermore, the transient stability improvement method for new energy power grids considering grid-based energy storage support described in this invention, in step 2, uses a high-dimensional spatiotemporal dynamic dataset to run multiple independent simulation channels in parallel, including:

[0020] The first simulation channel is a high-fidelity digital twin model based on physical mechanisms. The first simulation channel includes detailed switching device models, nonlinear control loop transfer functions, and frequency-related power grid impedance networks. It outputs system state estimates by running electromagnetic transient simulations.

[0021] The second simulation channel is a data-driven black box state-space model. The second simulation channel uses a dynamic mode decomposition algorithm to perform singular value decomposition on the snapshot matrix of measurement data organized by time windows to extract the dominant dynamic mode, and constructs a reduced-order state-space model to output the system state estimate.

[0022] The third simulation channel is an equivalent circuit model based on reduced-order analytical equations. The third simulation channel simulates the system behavior by solving a set of differential-algebraic equations that describe the interface characteristics of grid-connected energy storage and grid-connected inverters, and outputs system state estimates.

[0023] The first, second, and third simulation channels run in parallel after initialization, independently calculating and outputting a sequence of estimated values ​​for the system's subsynchronous frequency band impedance, dominant oscillation mode frequency, and damping.

[0024] Furthermore, the transient stability improvement method for new energy power grids considering grid-based energy storage support described in this invention, in step 3, employs a consensus fusion algorithm to perform cross-validation and confidence calibration on the multi-perspective evaluation results, including:

[0025] Real-time confidence index is calculated for the estimated value sequences output by the first simulation channel, the second simulation channel, and the third simulation channel. The real-time confidence index is dynamically generated by comprehensively considering the average deviation between the historical prediction errors of the first simulation channel, the second simulation channel, and the third simulation channel and the consensus value after preliminary filtering, the time delay and completeness of the latest input data, and model complexity factors.

[0026] The estimated values ​​and corresponding confidence levels of the same state quantity from the first, second, and third simulation channels at the current moment are transformed into the basic probability assignment function in the DS evidence theory. The evidence provided by the first, second, and third simulation channels is fused using the Dempster synthesis rule, and the confidence interval of the system being in a specific oscillation risk state is calculated after fusion.

[0027] When the degree of conflict between the evidence provided by any of the first, second, or third simulation channels and the consensus interval after fusion exceeds a preset threshold, it is determined that the simulation channel has a significant deviation.

[0028] For simulated channels with significant deviations, the consensus value formed by the fusion of high-confidence channels is used as a reference benchmark. A recursive least squares algorithm with a forgetting factor is used to adjust the parameters to be identified in the simulated channel model with significant deviations online, so as to minimize the error between the output of the simulated channel with significant deviations and the consensus reference value.

[0029] Furthermore, in the method for improving the transient stability of a new energy power grid considering grid-based energy storage, step 4 of the present invention, which involves identifying the transmission path of oscillating energy between control loops using a distributed machine learning architecture in a digital twin environment, includes:

[0030] In the high-fidelity digital twin environment calibrated in step 3, a local federated learning agent is set up for each grid-type energy storage or new energy power generation unit;

[0031] Each local federated learning agent has a subset of real-time operational data uploaded by the corresponding physical unit through a wide-area measurement system, as well as data generated by simulation under various operating conditions using a digital twin model;

[0032] Each local federated learning agent initializes a graph neural network model using local data. The graph neural network model has a graph structure based on power grid connection relationships and equipment types, and uses electrical quantities and control parameters as node features.

[0033] The coordination server periodically collects the model gradients or weight updates uploaded by each local federated learning agent, performs a weighted average, and then distributes the updated global model parameters to each local federated learning agent. After multiple iterations, the global model converges.

[0034] The real-time data and the system interaction coupling feature description generated in step 3 are input into the trained global graph neural network model. The trained global graph neural network model analyzes the transmission intensity and direction of oscillation energy between device control loop nodes through the graph attention layer and outputs a risk correlation map.

[0035] Furthermore, in the method for improving the transient stability of a new energy power grid considering grid-based energy storage, step 5 of the present invention, which involves generating a cooperative damping control strategy through reinforcement learning training, includes:

[0036] The system interaction coupling feature description generated in step 4 and the risk correlation map generated in step 4 are used as the enhanced state observation space;

[0037] A multi-agent deep reinforcement learning framework is constructed based on the enhanced state observation space. The multi-agent deep reinforcement learning framework includes a master agent and multiple subordinate agents.

[0038] The main agent aims to improve the overall damping ratio of the system, while the subordinate agents focus on optimizing the impedance characteristics of specific frequency bands or the decoupling effect between specific control loops.

[0039] In the high-fidelity digital twin environment calibrated in step 3, a reinforcement learning training environment is constructed. The action space of the master agent and the subordinate agent is defined as the adjustment amount of the virtual damping compensation signal parameters. The virtual damping compensation signal parameters include the notch filter center frequency, notch filter depth, phase compensation angle, and virtual impedance value.

[0040] Design a reward function that gives the main agent a reward that is positively correlated with the improvement of the global damping ratio, and gives the subordinate agents a reward that is related to the attenuation of oscillation energy in the frequency band under the responsibility of the subordinate agent or the decoupling effect between devices. The reward function also includes a penalty term for large fluctuations in control commands.

[0041] The proximal policy optimization algorithm is used to train the master agent and subordinate agents in a multi-agent deep reinforcement learning framework offline.

[0042] Furthermore, in the method for improving the transient stability of a new energy power grid considering grid-based energy storage, step 6 of the present invention, which involves generating optimal control commands using a predictive control method, includes:

[0043] Use the collaborative damping control strategy generated in step 5 as the set value;

[0044] Receive the real-time measurement data continuously collected in step 1;

[0045] Receive the parallel prediction result sequence of the system state for several future sampling periods from the first simulation channel, the second simulation channel and the third simulation channel in step 2;

[0046] Based on the setpoint of the collaborative damping control strategy generated in step 5, the real-time measurement data continuously collected in step 1, and the parallel prediction result sequence of the first, second, and third simulation channels in step 2, a local optimization problem is established with the goal of minimizing the weighted sum of oscillation energy in the future finite prediction time domain.

[0047] The local optimization problem is solved in a distributed manner using the alternating direction multiplier method, during which adjacent controllers exchange boundary state prediction information.

[0048] The first element of the obtained control command sequence is used as the actual control command at the current moment and sent to the grid-type energy storage converter.

[0049] Furthermore, in the method for improving the transient stability of a new energy power grid considering grid-based energy storage support described in this invention, step 6, the continuous evaluation of the control effect, includes:

[0050] The wide-area synchronous measurement system described in step 1 continuously monitors the voltage and current data streams.

[0051] The amplitude attenuation rate and modal damping ratio of the subsynchronous oscillation component are extracted in real time from the voltage and current data streams using signal processing algorithms.

[0052] The extracted amplitude attenuation rate and modal damping ratio are used as performance indicators.

[0053] Furthermore, in the method for improving the transient stability of a new energy power grid considering grid-based energy storage support described in this invention, step 6, which involves updating the system model and the cooperative damping control strategy, includes:

[0054] When the performance metrics obtained from step 8 fail to meet the expected threshold or a significant change in the system's operating mode is detected, the joint evolution process is triggered.

[0055] The model evolution path is as follows: the latest collected running data and the current system state are used as training data and input into the first simulation channel, the second simulation channel and the third simulation channel in step 2. The internal parameters of the first simulation channel, the second simulation channel and the third simulation channel are fine-tuned in real time using adaptive filtering or online parameter identification methods.

[0056] The strategy evolution path is as follows: using the latest running data and the updated digital twin model as the new training environment, start the reinforcement learning training process described in step 5, and use the Bayesian optimization algorithm to adjust the hyperparameters and reward function weights of the reinforcement learning system's neural network.

[0057] The internal parameters of the evolved first, second, and third simulation channels, along with the updated parameters of the collaborative damping control strategy, are synchronously updated to the online system.

[0058] Furthermore, the method for improving the transient stability of a new energy power grid considering grid-based energy storage, as described in this invention, includes the following specific data processing steps:

[0059] When collecting electrical quantity data in step 1, harmonic phasor data in the frequency band near the characteristic frequency of torsional vibration of the wind turbine shaft system are collected;

[0060] When running the simulation channel in parallel in step 2, a differential-algebraic equation characterizing the coupling effect between the virtual inertial response of energy storage and the torsional vibration mode of the wind turbine shaft system is established;

[0061] When generating the risk correlation map in step 4, the energy transfer path between the virtual synchronous control loop of the grid-type energy storage converter and the torsional vibration mode of the wind turbine shaft system is analyzed.

[0062] When generating the collaborative damping control strategy in step 5, the subordinate agent is set to optimize the decoupling effect between the energy storage virtual inertia response and the torsional vibration mode of the wind turbine shaft system;

[0063] When generating the optimal control command in step 6, the virtual impedance parameters of the grid-type energy storage converter are adjusted to reshape the impedance characteristics of the subsynchronous frequency band.

[0064] The beneficial effects of this invention are:

[0065] This invention constructs a high-dimensional spatiotemporal dynamic dataset through synchronous acquisition and fusion of multi-source data, providing a unified and high-quality input foundation for system state assessment. It utilizes parallel operation of multi-principle simulation channels to generate redundant and complementary multi-perspective assessment results, combined with a consistency fusion algorithm for cross-validation and confidence calibration, improving the reliability of the system's interactive coupling characteristics description. Based on a digital twin environment and distributed machine learning architecture, it identifies oscillatory energy transfer paths between control loops, refining the risk correlation map. It generates an adaptive collaborative damping control strategy through reinforcement learning training and employs predictive control methods for rolling optimization and real-time adjustment. Finally, it continuously updates the model and strategy through a closed-loop evolution mechanism, enabling the system to dynamically adapt to changes in grid operating conditions. This effectively suppresses subsynchronous oscillations caused by implicit interactions between grid-connected energy storage and new energy inverter control loops, significantly enhancing grid damping characteristics and improving transient stability. Attached Figure Description

[0066] To more clearly illustrate the technical solution of the present invention, the drawings used in the embodiments will be briefly introduced below. Obviously, for those skilled in the art, other drawings can be obtained based on the drawings without creative effort.

[0067] Figure 1 This is a schematic diagram of the process of the present invention. Detailed Implementation

[0068] To make the objectives, technical solutions, and advantages of this invention clearer, the technical solutions of this invention will be clearly and completely described below in conjunction with specific embodiments and corresponding drawings. Obviously, the described embodiments are only a part of the embodiments of this invention, and not all of them. Based on the embodiments of this invention, all other embodiments obtained by those skilled in the art without creative effort are within the scope of protection of this invention. The technical solutions provided by various embodiments of this invention will be described in detail below with reference to the accompanying drawings.

[0069] To better understand the purpose of this invention, the invention will now be described in further detail.

[0070] The present invention provides a method for improving the transient stability of a new energy power grid that considers grid-based energy storage support, comprising:

[0071] Step 1: Collect electrical quantity data and power electronic equipment control parameters from key nodes of the power grid, and form a high-dimensional spatiotemporal dynamic dataset through time synchronization processing;

[0072] Step 2: Using the high-dimensional spatiotemporal dynamic dataset formed in Step 1, run multiple independent simulation channels in parallel, estimate the dynamic characteristics of the system based on different principles, and obtain multi-perspective evaluation results of the system state.

[0073] Step 3: The multi-view evaluation results obtained in Step 2 are subjected to cross-validation and confidence calibration using a consensus fusion algorithm to generate a description of the system's interactive coupling characteristics.

[0074] Step 4: Using the system interaction coupling feature description generated in Step 3, identify the transmission path of oscillation energy between control loops in the digital twin environment through a distributed machine learning architecture, and generate a risk association map;

[0075] Step 5: Input the system interaction coupling feature description generated in Step 3 and the risk correlation map generated in Step 4 into the reinforcement learning system, and generate a collaborative damping control strategy through reinforcement learning training.

[0076] Step 6: Based on the real-time measurement data continuously collected from Step 1, the prediction information generated by multiple independent simulation channels in Step 2, and the cooperative damping control strategy generated from Step 5, the optimal control command is generated using the predictive control method, and the control parameters of the grid-type energy storage converter are dynamically adjusted; based on the performance indicators evaluated in Step 8, the system model and cooperative damping control strategy are updated to improve transient stability.

[0077] Step 1 involves deploying synchronous phasor measurement units and fault recording devices at the grid common connection point, grid-connected energy storage converters, and new energy inverter grid-connected nodes. This allows for the synchronous acquisition of fundamental and harmonic phasor data of three-phase voltage and current with millisecond-level time accuracy. Simultaneously, real-time dynamic parameters such as the inertial time constant and damping coefficient of the grid-connected energy storage virtual synchronous control loop, as well as the bandwidth and phase margin of the new energy inverter's phase-locked loop, are read via the equipment's local area network communication protocol. The acquired multidimensional heterogeneous data stream is transmitted to the central data processing unit via a time-sensitive network. Data alignment, interpolation, and outlier removal are performed based on precise time scales, integrating the data to form a structured high-dimensional spatiotemporal dynamic dataset. This step serves as the data foundation for this invention, acquiring grid physical quantities and equipment internal control parameters through wide-area synchronous measurement technology. This ensures a unified, high-quality data source for subsequent analysis, providing spatiotemporally consistent input for system state estimation and risk identification.

[0078] Step 2 utilizes a high-dimensional spatiotemporal dynamic dataset to construct three independent simulation channels in parallel: The first channel is based on a high-fidelity digital twin model of physical mechanisms, including detailed switching device models, nonlinear control loop transfer functions, and frequency-dependent grid impedance networks, running electromagnetic transient simulations to output system state estimates; the second channel is based on a data-driven black-box state-space model, employing a dynamic mode decomposition algorithm to perform singular value decomposition on the measurement data snapshot matrix, extracting the dominant dynamic modes and constructing a reduced-order state-space model; the third channel is based on an equivalent circuit model of reduced-order analytical equations, simulating system behavior by solving a system of differential-algebraic equations describing the interface characteristics of grid-connected energy storage and grid-connected inverters. The three channels run in parallel, outputting estimated sequences of the system's subsynchronous frequency band impedance, dominant oscillation mode frequency, and damping, respectively. This step provides a multi-perspective evaluation of the system's dynamic characteristics through multi-principle simulation channels, overcoming the limitations of a single model and laying the foundation for cross-validation.

[0079] Step 3 employs a consensus fusion algorithm to perform cross-validation and confidence calibration on the multi-perspective evaluation results. The system outputs a real-time confidence index for each simulated channel, dynamically generated by considering historical prediction errors, data freshness, and model complexity. Using Dempster evidence theory, channel estimates are transformed into basic probability assignment functions. The evidence body is fused using Dempster's synthesis rules to calculate the confidence interval for a specific oscillation risk state. When the conflict between a channel's result and the fused consensus interval exceeds a threshold, an online calibration process is triggered. Using a high-confidence consensus value as a benchmark, a recursive least squares algorithm with a forgetting factor is employed to adjust the parameters to be identified that deviate from the channel. This step, as the core of information credibility governance, improves the reliability of state estimation through dynamic fusion and calibration, outputting a highly reliable system interaction coupling index matrix.

[0080] Step 4 involves deploying a federated learning architecture within the calibrated digital twin environment. A local agent is established for each grid-connected energy storage or renewable energy generation unit. This agent collaboratively trains a global graph neural network model using real-time data from the corresponding physical unit and digital twin simulation data. The model uses grid connectivity and equipment type as its graph structure, and electrical quantities and control parameters as node features. A graph attention layer analyzes the intensity and direction of oscillating energy transmission between control loop nodes, generating a risk correlation graph. This step, as a deep risk analysis, uncovers the networked transmission paths of oscillating energy between control loops while protecting data privacy, refining interactive coupling characteristics and providing causal insights for control decisions.

[0081] Step 5 uses the system interaction coupling index matrix and risk correlation map as the enhanced state observation space, inputting them into the multi-agent deep reinforcement learning system. The system includes a master agent and multiple subordinate agents. The master agent aims to improve the system's global damping ratio, while the subordinate agents focus on optimizing impedance characteristics in specific frequency bands or the decoupling effect between control loops. A training environment is constructed in a digital twin environment. The agent action space is defined as the adjustment amount of virtual damping compensation signal parameters, and the reward function integrates oscillation suppression and control action smoothness. Offline training is performed using a proximal policy optimization algorithm to generate a collaborative virtual damping compensation strategy. This step serves as the intelligent decision center, adaptively generating a comprehensive compensation strategy through multi-objective reinforcement learning to adapt to complex coupling scenarios.

[0082] Step 6 uses the coordinated damping control strategy as the setpoint, combining the real-time measurement data from Step 1 and the multi-channel prediction results from Step 2, to generate the optimal control command using a distributed model predictive control method. Each local controller establishes an optimization problem with the objective of minimizing the weighted sum of oscillation energy in the future prediction time domain, and solves it using the alternating direction multiplier method in a distributed manner. Adjacent controllers exchange boundary state prediction information to achieve coordination. The first element of the control command sequence is sent to the grid-type energy storage converter in real time, dynamically adjusting the virtual synchronous control loop parameters. Simultaneously, the subsynchronous oscillation amplitude decay rate and modal damping ratio are continuously monitored through a wide-area measurement system as performance indicators. When the indicators do not meet expectations or the system operating mode changes, a joint evolution process is triggered: the model evolution path uses new data to fine-tune the simulation channel parameters, and the strategy evolution path uses Bayesian optimization to adjust the reinforcement learning hyperparameters and reward function weights. This step achieves control execution and closed-loop optimization, improves control accuracy through rolling time-domain prediction and multi-channel feedforward compensation, and enables the system to adapt to grid changes through model and strategy joint evolution.

[0083] The data acquisition process involves deploying synchronous phasor measurement units and fault recording devices at the grid connection points, grid-connected energy storage converters, and new energy inverter grid-connected nodes. The synchronous phasor measurement units synchronously acquire fundamental and harmonic phasor data of three-phase voltage and current with millisecond-level time accuracy, while the fault recording devices capture broadband transient waveform data. Simultaneously, the inertial time constant and damping coefficient of the virtual synchronous control loop of the grid-connected energy storage converter, as well as real-time dynamic parameters such as the bandwidth and phase margin of the phase-locked loop of the new energy inverter, are read via the device's local area network communication protocol. Multi-source heterogeneous data streams are transmitted to the central data processing unit via a time-sensitive network, and data transmission is performed according to a preset scheduling strategy. The central data processing unit performs time alignment and spatial interpolation processing on data from different spatial locations based on precise time stamps, removes outlier data, and integrates and encapsulates it into a structured high-dimensional spatiotemporal dynamic dataset. This multi-dimensional data acquisition and fusion method provides a spatiotemporally consistent input foundation for subsequent analysis.

[0084] The three independent simulation channels employ differentiated modeling principles. The first simulation channel establishes a high-fidelity digital twin model based on physical mechanisms, including detailed switching device models, nonlinear control loop transfer functions, and frequency-dependent grid impedance networks. System state estimates are calculated through electromagnetic transient simulation. The second simulation channel uses a data-driven approach to construct a black-box state-space model. A dynamic mode decomposition algorithm is applied to perform singular value decomposition on the time-window-organized measurement data snapshot matrix, extracting the dominant dynamic modes and establishing a reduced-order state-space model. The third simulation channel establishes an equivalent circuit model based on reduced-order analytical equations. The dynamic behavior of the system is simulated by solving a set of differential-algebraic equations describing the interface characteristics of grid-connected energy storage and grid-connected inverters. After initialization, the three simulation channels run in parallel, independently calculating and outputting estimated values ​​for the system's subsynchronous frequency band impedance, dominant oscillation mode frequency, and damping.

[0085] The cross-validation mechanism calculates a real-time confidence index for the estimated value sequence output by each simulated channel. The confidence index is dynamically generated, taking into account factors such as the average deviation between each channel's historical prediction error and the consensus value after initial filtering, the latency and completeness of the latest input data, and model complexity. The estimated values ​​and confidence levels of the same state variable from the three simulated channels are transformed into the basic probability assignment function in Dempster's evidence theory. Evidence fusion is performed using Dempster's synthesis rule to calculate the confidence interval for a specific oscillation risk state of the system. When the conflict between the evidence provided by a simulated channel and the fused consensus interval exceeds a preset threshold, the channel is deemed to have a significant deviation. For deviating channels, the consensus value formed by fusing high-confidence channels is used as a reference benchmark, and a recursive least squares algorithm with a forgetting factor is employed to adjust the parameters to be identified in the deviating channel model online.

[0086] A federated learning architecture is deployed in a calibrated high-fidelity digital twin environment. A local federated learning agent is established for each grid-connected energy storage or renewable energy generation unit. Each agent possesses a subset of real-time operational data uploaded by the corresponding physical unit through a wide-area measurement system, as well as data simulated under various operating conditions using the digital twin model. The local agent initializes a graph neural network model using local data. This model uses grid connection relationships and equipment types as the graph structure, and electrical quantities and control parameters as node features. The coordination server periodically collects model gradients or weight updates uploaded by each agent, performs a weighted average, and then distributes the updated global model parameters to each agent. After multiple rounds of iterative training, the real-time data and system interaction coupling feature descriptions are input into the trained global graph neural network model. The graph attention layer analyzes the intensity and direction of oscillatory energy transmission between nodes in the equipment control loop.

[0087] The system's interactive coupling characteristics and risk correlation map are constructed as an enhanced state observation space. Based on this space, a multi-agent deep reinforcement learning framework is built, comprising one master agent and multiple subordinate agents. The master agent aims to improve the system's global damping ratio, while the subordinate agents focus on optimizing impedance characteristics in specific frequency bands or decoupling effects between specific control loops. A reinforcement learning training environment is constructed within a digital twin environment. The agent's action space is defined as the adjustment amount of virtual damping compensation signal parameters, including the notch filter center frequency, notch filter depth, phase compensation angle, and virtual impedance value. The reward function is designed so that the master agent receives a reward positively correlated with the improvement of the global damping ratio, while the subordinate agents receive rewards related to the attenuation of oscillation energy in their assigned frequency band or the decoupling effect between devices. A penalty term is also introduced for large fluctuations in control commands.

[0088] The predictive control method uses the cooperative damping control strategy as the setpoint, and simultaneously receives the real-time measurement data continuously acquired in step 1 and the parallel prediction result sequence of the system state for several future sampling periods from the three simulation channels in step 2. Based on the input, a local optimization problem is established with the objective of minimizing the weighted sum of oscillation energy within the finite future prediction time domain. The alternating direction multiplier method is used to solve each local optimization problem in a distributed manner. During the iterative solution process, adjacent controllers exchange boundary state prediction information to achieve global coordination. The first element of the obtained control command sequence is used as the actual control command at the current moment and issued to the grid-type energy storage converter.

[0089] A wide-area synchronous measurement system continuously monitors voltage and current data streams, and signal processing algorithms are applied to extract the amplitude decay rate and modal damping ratio of the subsynchronous oscillation components in real time as performance indicators. When the performance indicators fail to reach the expected threshold or a significant change in the system operating mode is detected, a joint evolutionary process is triggered. The model evolution path uses the latest acquired operating data and the current system state as training data, inputs them into three analog channels, and uses adaptive filtering or online parameter identification methods to fine-tune the internal parameters in real time. The policy evolution path uses the latest operating data and the updated digital twin model as the new training environment, initiates the reinforcement learning training process, and uses Bayesian optimization algorithms to adjust the neural network hyperparameters and reward function weights.

[0090] For power station aggregation systems containing a high proportion of wind power and grid-connected energy storage, the data acquisition phase focuses on harmonic phasor data acquisition in the frequency band near the characteristic frequency of the wind turbine shaft torsional vibration. A differential-algebraic equation characterizing the coupling effect between the virtual inertial response of the energy storage and the torsional vibration modes of the wind turbine shaft is established in the simulation channel. During the risk identification phase, the energy transfer path between the virtual synchronization control loop of the grid-connected energy storage converter and the torsional vibration modes of the wind turbine shaft is analyzed. In the control strategy generation phase, a dedicated subordinate agent is set up to optimize the decoupling effect between the virtual inertial response of the energy storage and the torsional vibration modes of the wind turbine shaft. Finally, the impedance characteristics of the subsynchronous frequency band are reshaped by adjusting the virtual impedance parameters of the grid-connected energy storage converter.

[0091] This layered and progressive technical architecture forms a complete transient stability improvement solution through multi-source data fusion, multi-model collaborative evaluation, intelligent decision generation, and closed-loop optimization control. Tight coupling is established between each step through data flow and state transmission, ensuring both the system's real-time response capability and adapting to changes in grid operating conditions through a continuous evolution mechanism. Particularly in grid environments with a high proportion of renewable energy sources, this method can effectively suppress subsynchronous oscillations caused by implicit interactions in the control loop and improve the system's damping characteristics.

[0092] In power grid environments with a high proportion of renewable energy sources, the control interaction between grid-connected energy storage converters and renewable energy inverters can easily lead to subsynchronous oscillations. This invention addresses this issue by deploying synchronous phasor measurement units at the grid's point of common coupling, the grid-connected energy storage converter, and the renewable energy inverter's grid-connected nodes. These units acquire fundamental and harmonic phasor data of the three-phase voltage and current with millisecond-level time accuracy, while simultaneously capturing broadband transient waveform data using a fault recording device. Dynamic parameters such as the inertial time constant and damping coefficient of the grid-connected energy storage virtual synchronization control loop, as well as the bandwidth and phase margin of the renewable energy inverter's phase-locked loop, are read in real-time via the device's local area network communication protocol. Multi-source heterogeneous data is transmitted to a central data processing unit via a time-sensitive network, where data alignment and interpolation are performed based on precise time scales to form a structured, high-dimensional spatiotemporal dynamic dataset.

[0093] Based on a high-dimensional spatiotemporal dynamic dataset, the system operates three independent simulation channels in parallel. The first simulation channel uses electromagnetic transient simulation to construct a high-fidelity digital twin model, including switching device models, nonlinear control loop transfer functions, and frequency-dependent grid impedance networks. The second simulation channel employs a dynamic mode decomposition algorithm to perform singular value decomposition on the measurement data snapshot matrix, extracting the dominant dynamic modes to construct a data-driven state-space model. The third simulation channel establishes an equivalent circuit model by solving a system of differential-algebraic equations describing the interface characteristics of grid-connected energy storage and grid-connected inverters. The three channels output estimated sequences of the system's subsynchronous frequency band impedance, dominant oscillation mode frequency, and damping, respectively, forming a multi-perspective evaluation result.

[0094] A consensus fusion algorithm is used for cross-validation of the multi-perspective evaluation results. The system calculates a real-time confidence index for each simulated channel, with the index value dynamically generated by integrating historical prediction errors, data freshness, and model complexity. Dempster's evidence theory is used to transform the estimates of the three channels into basic probability assignment functions, and evidence fusion is performed using Dempster's synthesis rules. When the conflict between a channel's result and the fusion consensus interval exceeds a threshold, the deviation channel parameters are adjusted online using a recursive least squares algorithm with a forgetting factor, based on the high-confidence consensus value, outputting a highly reliable description of the system's interactive coupling characteristics.

[0095] A federated learning architecture is deployed in a calibrated digital twin environment, with a local agent established for each power generation unit. The local agent trains a graph neural network model using real-time data from the physical unit and digital twin simulation data. The model uses grid connections and equipment types as the graph structure, and electrical quantities and control parameters as node features. A coordination server aggregates the model gradients from each agent using a federated averaging algorithm, iterating through multiple rounds to bring the global model to convergence. The trained graph neural network then uses an attention layer to analyze the propagation path of oscillating energy between control loop nodes, generating a risk correlation graph.

[0096] The system's interactive coupling characteristics and risk correlation map are constructed as an enhanced state observation space, which is then input into a multi-agent deep reinforcement learning system. The master agent aims to improve the system's global damping ratio as a macroscopic objective, while subordinate agents optimize impedance characteristics in specific frequency bands or the decoupling effect between control loops. In a digital twin environment, the adjustment amount of virtual damping compensation signal parameters is defined as the action space, and the reward function integrates oscillation suppression and control action smoothness. Offline training is performed using a proximal policy optimization algorithm to generate a collaborative virtual damping compensation strategy.

[0097] The coordinated damping control strategy serves as the setpoint, and combined with real-time measurement data and multi-channel prediction results, a distributed model predictive control method is used to generate optimal control commands. Each local controller establishes an optimization problem with the objective of minimizing the weighted sum of oscillation energy in the future prediction time domain, and solves it in a distributed manner using the alternating direction multiplier method. Adjacent controllers exchange boundary state prediction information to achieve coordinated control. The first element of the control command sequence is sent to the grid-type energy storage converter in real time, dynamically adjusting the parameters of the virtual synchronization control loop.

[0098] A wide-area measurement system continuously monitors voltage and current data streams, and signal processing algorithms are applied to extract the amplitude decay rate and modal damping ratio of the subsynchronous oscillation component as performance indicators. When the indicators fail to meet expectations or the system operating mode changes, a joint evolution process is triggered. The model evolution path uses new data to fine-tune the simulation channel parameters, while the policy evolution path adjusts the reinforcement learning hyperparameters and reward function weights through Bayesian optimization, achieving the co-evolution of the system model and the control strategy.

[0099] For wind power energy storage aggregation systems, the data acquisition phase focuses on collecting harmonic phasor data near the characteristic frequency of the wind turbine shaft torsional vibration. Differential-algebraic equations characterizing the coupling effect between the virtual inertial response of the energy storage and the torsional vibration modes of the wind turbine shaft are established in the simulation channel. The risk identification phase focuses on analyzing the energy transfer path between the virtual synchronization control loop of the grid-type energy storage converter and the torsional vibration modes of the wind turbine shaft. In the control strategy generation phase, a dedicated subordinate agent is set up to optimize the decoupling effect between the two types of modes, and finally, the impedance characteristics of the subsynchronous frequency band are reshaped by adjusting the virtual impedance parameters.

[0100] This invention forms a complete transient stability improvement solution through multi-source data fusion, multi-model collaborative evaluation, and intelligent decision control. Tight coupling is established between each component through data flow and state transmission. While ensuring the system's real-time response capability, a continuous evolution mechanism adapts to changes in power grid operating conditions, effectively suppressing subsynchronous oscillations caused by implicit interactions in the control loop.

[0101] Embodiment 1 of this invention: In the implementation case of a wind power energy storage farm aggregation system, the grid-connected energy storage converter and the wind turbine inverter operate in parallel through a common connection point. The wind turbine adopts a grid-following control strategy, with its phase-locked loop bandwidth set in the range of 10 Hz to 20 Hz, while the inertial time constant of the virtual synchronization control loop of the energy storage converter is set to 5 to 15 seconds, and the damping coefficient is adjusted between 0.05 and 0.2. Synchronous phasor measurement units deployed at each grid-connected node collect voltage and current waveforms at a sampling rate of 1000 points per second. The fault recording device synchronously captures transient harmonic components with frequencies from 5 Hz to 40 Hz, with particular focus on monitoring the frequency band near the torsional vibration characteristic frequency of the wind turbine shaft system around 25 Hz. The central data processing unit receives data through a time-sensitive network, uses the IEEE 1588 precision clock protocol for time alignment, and generates a high-dimensional dataset including spatiotemporal correlation features.

[0102] Three simulation channels process data in parallel. The first simulation channel constructs an electromagnetic transient model based on actual power grid topology parameters, simulating the behavior of power devices with a switching frequency of 2 kHz, and outputs the subsynchronous frequency band impedance curve. The second simulation channel applies dynamic mode decomposition to the sliding window of the measurement data to extract the modal component with a dominant oscillation mode frequency of 18.5 Hz. The third simulation channel calculates the estimated system damping ratio of 0.12 by solving the interface differential equations. A consensus fusion algorithm assigns dynamic weights to each channel: the first channel receives a weight of 0.5 due to model completeness, the second channel receives a weight of 0.3 based on data freshness, and the third channel receives a weight of 0.2 based on computational efficiency. When the output of the second channel deviates from that of other channels by more than 15%, a parameter calibration process is triggered.

[0103] In the federated learning architecture, each wind turbine agent trains a graph neural network using 200 sets of local operating data, and the coordination server aggregates the model parameters every 30 minutes. The trained global model identifies the transmission path of oscillating energy from the 18.5 Hz torsional vibration mode of the wind turbine to the control loop of the energy storage converter through an attention mechanism, and the risk correlation graph shows that key coupling nodes are located at common connection points. The multi-agent reinforcement learning system sets the reward function of the master agent to the damping ratio improvement, and sets a dedicated reward item for the subordinate agents for the 18.5 Hz frequency band. After 100,000 offline training iterations, a virtual impedance adjustment strategy is generated, optimizing the virtual impedance value of the energy storage converter from 0.5 ohms to 0.8 ohms.

[0104] The model predictive controller uses the next 3 seconds as the prediction time domain and solves the optimization problem every 100 milliseconds. The first element of the control command sequence adjusts the parameters of the energy storage converter in real time, increasing the system damping ratio to 0.25. The wide-area measurement system continuously monitors the oscillation amplitude decay rate. When a sudden change in wind speed is detected that causes performance fluctuations to exceed 20%, the joint evolution mechanism initiates model parameter updates, thus forming a closed-loop autonomous system.

[0105] Embodiment 2 of this invention: In a photovoltaic-energy storage hybrid power station application scenario, the photovoltaic inverter adopts maximum power point tracking control and operates in conjunction with grid-based energy storage. The synchronous phasor measurement unit focuses on monitoring the power fluctuation characteristics caused by changes in irradiance, collecting harmonic data at frequencies from 2 Hz to 15 Hz. The virtual inertia of the energy storage converter is set to 8 seconds, the damping coefficient is set to 0.15, and the phase-locked loop bandwidth of the photovoltaic inverter is adjusted to 12 Hz. After data fusion, the time correlation characteristics between rapid fluctuations in photovoltaic power and energy storage response are highlighted.

[0106] Simulated channel analysis shows that photovoltaic power fluctuations mainly induce oscillation modes around 10 Hz. The federated learning model identifies the energy transfer path through capacitive coupling of the DC link. The adaptive control strategy generated after reinforcement learning training enables the system to remain stable under conditions where the solar irradiance changes by 20% per minute, maintaining a damping ratio above 0.18. The model predictive controller adjusts the energy storage output 100 milliseconds in advance based on the photovoltaic power prediction curve, effectively mitigating the oscillation risk caused by power fluctuations.

Claims

1. A method for improving the transient stability of a new energy power grid considering grid-based energy storage support, characterized in that, include: Step 1: Collect electrical quantity data and power electronic equipment control parameters from key nodes of the power grid, and form a high-dimensional spatiotemporal dynamic dataset through time synchronization processing; Step 2: Using the high-dimensional spatiotemporal dynamic dataset formed in Step 1, run multiple independent simulation channels in parallel, estimate the dynamic characteristics of the system based on different principles, and obtain multi-perspective evaluation results of the system state. Step 3: The multi-view evaluation results obtained in Step 2 are subjected to cross-validation and confidence calibration using a consensus fusion algorithm to generate a description of the system's interactive coupling characteristics. Step 4: Using the system interaction coupling feature description generated in Step 3, identify the transmission path of oscillation energy between control loops in the digital twin environment through a distributed machine learning architecture, and generate a risk association map; Step 5: Input the system interaction coupling feature description generated in Step 3 and the risk correlation map generated in Step 4 into the reinforcement learning system, and generate a collaborative damping control strategy through reinforcement learning training. Step 6: Based on the real-time measurement data continuously collected from Step 1, the prediction information generated by multiple independent simulation channels in Step 2, and the cooperative damping control strategy generated from Step 5, the optimal control command is generated using the predictive control method, and the control parameters of the grid-type energy storage converter are dynamically adjusted; based on the performance indicators obtained from Step 2, the system model and cooperative damping control strategy are updated to improve transient stability.

2. The method for improving the transient stability of a new energy power grid considering grid-based energy storage support as described in claim 1, characterized in that, Step 1, which involves collecting electrical quantity data and power electronic equipment control parameters from key nodes of the power grid, includes: By deploying synchronous phasor measurement units at the grid common connection point, grid-connected energy storage converters, and new energy inverter grid-connected nodes, fundamental and harmonic phasor data of three-phase voltage and current are synchronously collected with millisecond-level time accuracy. Capture broadband transient waveform data using a fault recording device; The inertial time constant and damping coefficient of the virtual synchronous control loop of the grid-type energy storage converter, as well as the bandwidth and phase margin of the phase-locked loop of the new energy inverter, are read through the device local area network communication protocol. The phasor data collected by the synchronous phasor measurement unit, the transient waveform data captured by the fault recording device, and the parameters read through the device's local area network are transmitted to the central data processing unit through a time-sensitive network according to a preset scheduling strategy. The central data processing unit aligns and interpolates data from different spatial locations based on precise time stamps, removes outliers, and integrates and encapsulates them into a structured high-dimensional spatiotemporal dynamic dataset.

3. The method for improving the transient stability of a new energy power grid considering grid-based energy storage support as described in claim 2, characterized in that, Step 2 uses a high-dimensional spatiotemporal dynamic dataset to run multiple independent simulation channels in parallel, including: The first simulation channel is a high-fidelity digital twin model based on physical mechanisms. The first simulation channel includes detailed switching device models, nonlinear control loop transfer functions, and frequency-related power grid impedance networks. It outputs system state estimates by running electromagnetic transient simulations. The second simulation channel is a data-driven black box state-space model. The second simulation channel uses a dynamic mode decomposition algorithm to perform singular value decomposition on the snapshot matrix of measurement data organized by time windows to extract the dominant dynamic mode, and constructs a reduced-order state-space model to output the system state estimate. The third simulation channel is an equivalent circuit model based on reduced-order analytical equations. The third simulation channel simulates the system behavior by solving a set of differential-algebraic equations that describe the interface characteristics of grid-connected energy storage and grid-connected inverters, and outputs system state estimates. The first, second, and third simulation channels run in parallel after initialization, independently calculating and outputting a sequence of estimated values ​​for the system's subsynchronous frequency band impedance, dominant oscillation mode frequency, and damping.

4. The method for improving the transient stability of a new energy power grid considering grid-based energy storage support as described in claim 3, characterized in that, Step 3 involves using a consensus fusion algorithm to perform cross-validation and confidence calibration on the multi-view evaluation results, including: Real-time confidence index is calculated for the estimated value sequences output by the first simulation channel, the second simulation channel, and the third simulation channel. The real-time confidence index is dynamically generated by comprehensively considering the average deviation between the historical prediction errors of the first simulation channel, the second simulation channel, and the third simulation channel and the consensus value after preliminary filtering, the time delay and completeness of the latest input data, and model complexity factors. The estimated values ​​and corresponding confidence levels of the same state quantity from the first, second, and third simulation channels at the current moment are transformed into the basic probability assignment function in the DS evidence theory. The evidence provided by the first, second, and third simulation channels is fused using the Dempster synthesis rule, and the confidence interval of the system being in a specific oscillation risk state is calculated after fusion. When the degree of conflict between the evidence provided by any of the first, second, or third simulation channels and the consensus interval after fusion exceeds a preset threshold, it is determined that the simulation channel has a significant deviation. For simulated channels with significant deviations, the consensus value formed by the fusion of high-confidence channels is used as a reference benchmark. A recursive least squares algorithm with a forgetting factor is used to adjust the parameters to be identified in the simulated channel model with significant deviations online, so as to minimize the error between the output of the simulated channel with significant deviations and the consensus reference value.

5. The method for improving the transient stability of a new energy power grid considering grid-based energy storage support as described in claim 4, characterized in that, Step 4, which describes identifying the transmission path of oscillatory energy between control loops in a digital twin environment using a distributed machine learning architecture, includes: In the high-fidelity digital twin environment calibrated in step 3, a local federated learning agent is set up for each grid-type energy storage or new energy power generation unit; Each local federated learning agent has a subset of real-time operational data uploaded by the corresponding physical unit through a wide-area synchronous measurement system, as well as data generated by simulation under various operating conditions using a digital twin model; Each local federated learning agent initializes a graph neural network model using local data. The graph neural network model has a graph structure based on power grid connection relationships and equipment types, and uses electrical quantities and control parameters as node features. The coordination server periodically collects the model gradients or weight updates uploaded by each local federated learning agent, performs a weighted average, and then distributes the updated global model parameters to each local federated learning agent. After multiple iterations, the global model converges. The real-time data and the system interaction coupling feature description generated in step 3 are input into the trained global graph neural network model. The trained global graph neural network model analyzes the transmission intensity and direction of oscillation energy between device control loop nodes through the graph attention layer and outputs a risk correlation map.

6. The method for improving the transient stability of a new energy power grid considering grid-based energy storage support as described in claim 5, characterized in that, Step 5, which describes generating a collaborative damping control strategy through reinforcement learning training, includes: The system interaction coupling feature description generated in step 3 and the risk correlation map generated in step 4 are used as the enhanced state observation space; A multi-agent deep reinforcement learning framework is constructed based on the enhanced state observation space. The multi-agent deep reinforcement learning framework includes a master agent and multiple subordinate agents. The main agent aims to improve the overall damping ratio of the system, while the subordinate agents focus on optimizing the impedance characteristics of specific frequency bands or the decoupling effect between specific control loops. In the high-fidelity digital twin environment calibrated in step 3, a reinforcement learning training environment is constructed. The action space of the master agent and the subordinate agent is defined as the adjustment amount of the virtual damping compensation signal parameters. The virtual damping compensation signal parameters include the notch filter center frequency, notch filter depth, phase compensation angle, and virtual impedance value. Design a reward function that gives the main agent a reward that is positively correlated with the improvement of the global damping ratio, and gives the subordinate agents a reward that is related to the attenuation of oscillation energy in the frequency band under the responsibility of the subordinate agent or the decoupling effect between devices. The reward function also includes a penalty term for large fluctuations in control commands. The proximal policy optimization algorithm is used to train the master agent and subordinate agents in a multi-agent deep reinforcement learning framework offline.

7. The method for improving the transient stability of a new energy power grid considering grid-based energy storage support as described in claim 6, characterized in that, Step 6, which involves generating the optimal control command using a predictive control method, includes: Use the collaborative damping control strategy generated in step 5 as the set value; Receive the real-time measurement data continuously collected in step 1; Receive the parallel prediction result sequence of the system state for several future sampling periods from the first simulation channel, the second simulation channel and the third simulation channel in step 2; Based on the setpoint of the collaborative damping control strategy generated in step 5, the real-time measurement data continuously collected in step 1, and the parallel prediction result sequence of the first, second, and third simulation channels in step 2, a local optimization problem is established with the goal of minimizing the weighted sum of oscillation energy in the future finite prediction time domain. The local optimization problem is solved in a distributed manner using the alternating direction multiplier method, during which adjacent controllers exchange boundary state prediction information. The first element of the obtained control command sequence is used as the actual control command at the current moment and sent to the grid-type energy storage converter.

8. The method for improving the transient stability of a new energy power grid considering grid-based energy storage support as described in claim 7, characterized in that, Step 2 involves continuously evaluating the effectiveness of the control measures, including: The wide-area synchronous measurement system described in step 1 continuously monitors the voltage and current data streams. The amplitude attenuation rate and modal damping ratio of the subsynchronous oscillation component are extracted in real time from the voltage and current data streams using signal processing algorithms. The extracted amplitude attenuation rate and modal damping ratio are used as performance indicators.

9. The method for improving the transient stability of a new energy power grid considering grid-based energy storage support as described in claim 8, characterized in that, Step 6, which involves updating the system model and the cooperative damping control strategy, includes: When the performance indicators obtained from step 2 fail to meet the expected threshold or a significant change in the system's operating mode is detected, the joint evolution process is triggered. The model evolution path is as follows: the latest collected running data and the current system state are used as training data and input into the first simulation channel, the second simulation channel and the third simulation channel in step 2. The internal parameters of the first simulation channel, the second simulation channel and the third simulation channel are fine-tuned in real time using adaptive filtering or online parameter identification methods. The strategy evolution path is as follows: using the latest running data and the updated digital twin model as the new training environment, start the reinforcement learning training process described in step 5, and use the Bayesian optimization algorithm to adjust the hyperparameters and reward function weights of the reinforcement learning system's neural network. The internal parameters of the evolved first, second, and third simulation channels, along with the updated parameters of the collaborative damping control strategy, are synchronously updated to the online system.

10. The method for improving the transient stability of a new energy power grid considering grid-based energy storage support according to claim 9, characterized in that, The method includes the following specific data processing steps: When collecting electrical quantity data in step 1, harmonic phasor data in the frequency band near the characteristic frequency of torsional vibration of the wind turbine shaft system are collected; When running the simulation channel in parallel in step 2, a differential-algebraic equation characterizing the coupling effect between the virtual inertial response of energy storage and the torsional vibration mode of the wind turbine shaft system is established; When generating the risk correlation map in step 4, the energy transfer path between the virtual synchronous control loop of the grid-type energy storage converter and the torsional vibration mode of the wind turbine shaft system is analyzed. When generating the collaborative damping control strategy in step 5, the subordinate agent is set to optimize the decoupling effect between the energy storage virtual inertia response and the torsional vibration mode of the wind turbine shaft system; When generating the optimal control command in step 6, the virtual impedance parameters of the grid-type energy storage converter are adjusted to reshape the impedance characteristics of the subsynchronous frequency band.

Images