Method for evaluating capability of improving transient stability of power grid through network construction type energy storage

Abstract

The invention relates to the technical field of power system stability control and new energy grid connection, in particular to a method for evaluating the capability of improving the transient stability of a power grid through network construction type energy storage, and the method comprises the steps: collecting data, and obtaining a high-quality data source through time synchronization and quality evaluation; constructing a digital twinborn simulation environment; injecting an optimized excitation signal in a simulation environment, and identifying hidden coupling mode characteristics between virtual inertia control and a power grid oscillation mode by adopting a tensor decomposition method; inputting the features into a deep reinforcement learning agent to explore optimal control parameters, and establishing an agent model by adopting a meta-model optimizer to predict a Pareto frontier; and generating a comprehensive evaluation report containing a transient stability boundary, a risk map, a parameter feasible region and a self-adaptive strategy through high-fidelity verification and calibration of the intelligent agent and the meta-model, and integrating the comprehensive evaluation report to an energy storage control system and a power grid energy management system. According to the method, the problem of stable boundary quantization misalignment caused by implicit coupling is effectively solved.

Description

Technical Field

[0001] This invention relates to the fields of power system stability control and new energy grid connection technology, and in particular to a method for assessing the ability of grid-based energy storage to improve the transient stability of the power grid. Background Technology

[0002] In a power electronic grid containing a high proportion of renewable energy, renewable energy such as wind and solar power is intermittent. A power electronic grid refers to a grid that connects a large number of new energy sources through equipment such as inverters. Transient stability reflects the system's ability to maintain synchronous operation after being disturbed. Simulation technology simulates the dynamic process of the power grid based on mathematical models and analyzes transient behavior. Online evaluation technology uses real-time data to continuously calculate stability indicators to achieve dynamic monitoring and early warning.

[0003] Existing power grid transient stability simulation and online assessment technologies suffer from the following technical challenges: In power electronic grids with a high proportion of renewable energy, the virtual inertial control of grid-connected energy storage forms an implicit frequency domain coupling with the inherent subsynchronous or supersynchronous oscillation modes of the power grid due to the wide bandwidth characteristics of power electronic equipment control; Existing assessment methods, based on linearized models or preset operating conditions, cannot fully capture the nonlinear dynamics of multimodal interactions during transient processes, leading to inaccurate quantification of the stability boundary of energy storage support capabilities; For example, when large-scale wind power disconnection causes frequency drops, the virtual inertial response of energy storage may unexpectedly excite subsynchronous resonance in tie lines, causing the stable operating point identified in the simulation to become unstable in the actual system, thereby affecting the reliability of the assessment results and threatening power grid security. Summary of the Invention

[0004] To address the shortcomings of existing technologies, this invention provides a method for evaluating the ability of grid-based energy storage to improve the transient stability of the power grid. This method solves the technical problem of inaccurate quantification of the stability boundary in evaluating the ability of grid-based energy storage to improve the transient stability of the power grid due to the implicit frequency domain coupling between the virtual inertia of grid-based energy storage and the sub / supersynchronous oscillation modes of the power grid during transient processes.

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

[0006] This invention provides a method for evaluating the ability of grid-based energy storage to improve the transient stability of the power grid, comprising:

[0007] Step 1: Collect synchronous phasor data from the wide-area measurement system and command stream data from the local controller of the grid-type energy storage system. Perform time synchronization processing on the two types of data, perform data quality assessment, and obtain high-quality data sources.

[0008] Step 2: Using the high-quality data source, construct a digital twin simulation environment, which includes an equivalent model of the power grid electromechanical transients and a detailed model of the energy storage controller.

[0009] Step 3: In the digital twin simulation environment, inject optimized excitation signals, collect system response data, and use tensor decomposition method to identify the implicit coupling mode characteristics between virtual inertial control and power grid oscillation mode to obtain the implicit coupling mode characteristics.

[0010] Step 4: Input the implicit coupled modal features into the deep reinforcement learning agent. The deep reinforcement learning agent uses the energy storage control parameters as the action space and the stability index as the reward function. It interacts with the digital twin simulation environment to explore the optimal parameters and obtain the optimal parameter set recommended by the agent. At the same time, a meta-model optimizer is used to establish a surrogate model from the parameter space to the stability index to predict the Pareto front and obtain the Pareto front solution predicted by the meta-model.

[0011] Step 5: Feed the optimal parameter set recommended by the agent and the Pareto front solution predicted by the meta-model back to the digital twin simulation environment for high-fidelity verification to obtain the verification results; compare the consistency of the verification results, the reward value of the agent in the process of exploring the optimal parameters, and the predicted value corresponding to the Pareto front solution predicted by the meta-model; if there is a deviation, calibrate the reward function of the deep reinforcement learning agent according to the verification results, calibrate the surrogate model of the meta-model optimizer, and repeat Step 4 and this step until convergence, to obtain the converged agent policy, the calibrated surrogate model, and the verification data;

[0012] Step 6: Integrate the converged agent strategy, the calibrated surrogate model, the verification data, and the implicit coupling mode features to generate a comprehensive capability assessment report. The comprehensive capability assessment report generates transient stability boundaries and implicit coupling risk maps based on the verification data and the implicit coupling mode features, generates robust parameter feasible regions based on the calibrated surrogate model, and generates adaptive control strategy mapping functions based on the converged agent strategy. The comprehensive capability assessment report is then integrated into the upper-level controller of the energy storage system and into the grid energy management system.

[0013] Furthermore, in the method for assessing the ability of grid-based energy storage to improve the transient stability of the power grid described in this invention, step 1, which involves performing time synchronization processing on two types of data and assessing data quality, includes:

[0014] Based on a precision clock protocol, the timestamps of synchronous phasor measurement unit data from different plants are aligned to a unified reference time axis to obtain the aligned frequency sequence and voltage phasor sequence.

[0015] The communication messages of the grid-type energy storage inverter are analyzed to extract the time series of internal state variables of the virtual synchronous machine control loop. The internal state variables include power reference value, voltage reference value, and phase-locked loop output angle.

[0016] Integrity checks are performed on the aligned frequency sequence, voltage phasor sequence, and extracted internal state variable time sequence. Anomaly pattern detection based on dynamic time warping algorithm is used to identify and mark missing data segments and data transition segments.

[0017] A pre-trained spatiotemporal generative adversarial network is invoked to repair the marked missing data segments and data jump segments. The generator generates repaired data based on the spatiotemporal correlation of adjacent normal data points, and the discriminator judges the consistency between the repaired data and the original normal data distribution, thereby outputting the repaired continuous data sequence to form the high-quality data source.

[0018] Furthermore, in the method for evaluating the ability of grid-based energy storage to improve the transient stability of the power grid, step 2 of the present invention, which involves constructing a digital twin simulation environment, includes:

[0019] Using the power grid topology connections and measurement data from the high-quality data source, the Ward equivalent method is used to dynamically equivalence the remote power grid, retaining all generator nodes and key load nodes in the study area to form the electromechanical transient equivalent model.

[0020] The instruction stream data of the grid-type energy storage local controller is compiled into model description language code supported by a real-time digital simulator to obtain executable controller model code.

[0021] By using the power hardware-in-the-loop simulation interface, the executable controller model code is injected into the simulation loop corresponding to the electromechanical transient equivalent model. The calculation step size, communication delay parameters, and output limiting circuit are configured to establish a real-time data exchange channel between the electromechanical transient simulation subsystem and the electromagnetic transient simulation subsystem, thereby constructing the digital twin simulation environment.

[0022] Furthermore, in the method for assessing the ability of grid-based energy storage to improve the transient stability of the power grid, step 3 of the present invention, which involves injecting an optimized excitation signal, collecting system response data, and identifying implicitly coupled mode characteristics using tensor decomposition, includes:

[0023] Set the frequency band range to be detected, and use the quantum annealing algorithm to search for a set of broadband excitation signal waveforms with the maximum energy concentration and mutual orthogonality within the frequency band range to obtain the optimized excitation signal;

[0024] The optimized excitation signal is injected into the voltage reference point of the energy storage device in the digital twin simulation environment. At the same time, the active power signal of the energy storage output port, the power angle signal of the selected generator in the system, and the power signal of the key tie line are collected as multi-channel system response data.

[0025] For each channel of the multi-channel system response data, a short-time Fourier transform is performed to obtain the time-spectrum matrix of each channel;

[0026] Stack the time-frequency matrix of all channels along the spatial channel dimension to construct a three-dimensional response tensor of time, frequency, and space;

[0027] Perform a higher-order singular value decomposition on the response tensor to extract the decomposed core tensor, time factor matrix, frequency factor matrix, and spatial factor matrix;

[0028] By analyzing the frequency factor matrix and spatial factor matrix, oscillation modes with significant spatial factor weights in the energy storage control channel and frequency factors distributed near the virtual inertial control frequency band are selected. These selected oscillation modes are used as the implicit coupling mode features.

[0029] Furthermore, in the method for evaluating the ability of grid-based energy storage to improve the transient stability of the power grid, step 4 of the present invention, the deep reinforcement learning agent interacts with the digital twin simulation environment to explore the optimal parameters, using energy storage control parameters as the action space and stability indicators as the reward function.

[0030] Construct the state vector of the agent, the state vector including the amplitude, frequency, and participation factor corresponding to each generator of the implicit coupled mode feature;

[0031] The action vector of the intelligent agent is defined as the adjustment amount of the time constant and damping coefficient of the virtual inertial element of the grid-type energy storage;

[0032] The action vector is executed in the digital twin simulation environment to obtain the system's dynamic response;

[0033] Based on the system's dynamic response, calculate the damping ratio of the system's dominant low-frequency oscillation mode and the amplitude attenuation rate of the implicitly coupled mode characteristics.

[0034] Based on the increase in the damping ratio and the increase in the amplitude decay rate, the current reward value is calculated using a preset reward function, and the policy network of the deep reinforcement learning agent is updated using the current reward value.

[0035] Furthermore, in the method for assessing the ability of grid-based energy storage to improve the transient stability of the power grid described in this invention, step 4, which involves establishing a surrogate model using a meta-model optimizer to predict the Pareto front, includes:

[0036] Within the domain of the energy storage control parameters, an initial sample point set is generated using the Latin hypercube sampling method;

[0037] The control parameters corresponding to each sample point in the initial sample point set are input into the digital twin simulation environment to obtain the stability index output corresponding to each sample point. The stability index includes transient stability limit and coupled mode amplitude.

[0038] Using all sample points and their corresponding stability index outputs, a Gaussian process regression model based on the Kriging method is trained to obtain the initial surrogate model.

[0039] The expected improved acquisition function is adopted to guide the selection of new sample points in the parameter space. The newly selected sample points are input into the digital twin simulation environment to obtain the corresponding stability index output. The initial surrogate model is iteratively updated using the new sample points and their stability index output until the Pareto front predicted by the model converges, thereby obtaining the surrogate model of the meta-model optimizer.

[0040] Furthermore, in the method for evaluating the ability of grid-based energy storage to improve the transient stability of the power grid described in this invention, the calibration agent reward function in step 5, and the calibration meta-model include:

[0041] The actual stability index obtained from the high-fidelity verification is compared with the estimated reward value during the interaction process of the intelligent agent to calculate the prediction deviation.

[0042] Based on the magnitude and direction of the predicted deviation, the weight coefficients of the damping ratio and modal amplitude in the reward function are dynamically adjusted to calibrate the reward function of the deep reinforcement learning agent.

[0043] The new parameter-performance data pairs generated by the high-fidelity verification are added to the training sample library of the meta-model optimizer to obtain the expanded training sample library.

[0044] Using the expanded training sample library, the hyperparameter optimization of the Gaussian process regression model is re-executed to update the surrogate model and calibrate the surrogate model of the meta-model optimizer.

[0045] Furthermore, in the method for assessing the ability of grid-based energy storage to improve the transient stability of the power grid, step 6 of the present invention, generating a comprehensive capability assessment report, includes:

[0046] System sampling is performed on a two-dimensional grid consisting of a set of typical power grid operating modes and an energy storage control parameter range to obtain a grid sampling point set;

[0047] For each sampling point in the grid sampling point set, the implicit coupled mode feature identification process is invoked to calculate the coupled mode amplitude corresponding to that point, thereby obtaining the coupled mode amplitude data of all sampling points;

[0048] The coupled modal amplitude data of all the sampling points are visualized as the risk map using color mapping.

[0049] From the calibrated meta-model proxy model, extract all parameter points that satisfy the preset stability threshold constraint to form a feasible solution set;

[0050] A support vector machine algorithm is used to learn the feasible solution set to obtain the classification hyperplane equation describing the boundary of the robust parameter feasible region, and the envelope surface of the robust parameter feasible region is obtained.

[0051] The policy network parameters after the deep reinforcement learning agent training converges are solidified and encapsulated into the adaptive control policy mapping function that receives real-time modal features and outputs parameter adjustment amounts, thus obtaining the adaptive control policy mapping function.

[0052] The comprehensive capability assessment report is generated by combining the risk map, the robustness parameter feasible domain envelope, the adaptive control strategy mapping function, and the transient stability boundary obtained in step 5.

[0053] Furthermore, in the method for assessing the ability of grid-based energy storage to improve the transient stability of the power grid described in this invention, the transient stability boundary included in step 6 of generating the comprehensive capability assessment report includes:

[0054] In the digital twin simulation environment, three-phase short-circuit faults and single-phase ground faults are set, and the faults are specified to occur at the midpoint of the main transmission line and the substation bus.

[0055] A binary search algorithm is used to adjust the fault duration, with the upper limit of the initial search interval set to the system protection action time limit and the lower limit set to zero.

[0056] Simulations were run for each fault duration, and the relative power angle difference of all generator rotors in the system was monitored. The fault duration corresponding to the power angle difference exceeding the critical threshold of 180 degrees was recorded as the transient power angle stability boundary.

[0057] By changing the system's operating mode and repeating the binary search process, the set of transient stable boundary points under different operating modes can be obtained.

[0058] Using the transient stability boundary point set under the different operating modes, a stable boundary surface is constructed with the critical line transmission power and fault clearing time as coordinate axes, which serves as the transient stability boundary.

[0059] Furthermore, in the method for assessing the ability of grid-based energy storage to improve the transient stability of the power grid, step 6, which involves integrating the assessment results into the upper-level controller of the energy storage system and into the power grid energy management system, includes:

[0060] The mathematical description of the feasible domain envelope of the robustness parameter is written into the parameter database of the upper controller. The parameter inspection logic is configured to periodically read the parameter description in the database and compare the relative position of the current operating point with the boundary of the feasible domain. When the operating point approaches the boundary, an early warning is triggered and a safe parameter combination within the feasible domain is recommended, thus completing the integration of the evaluation results into the upper controller of the energy storage system.

[0061] The adaptive control strategy mapping function is deployed in the safety constraint logic of the power grid energy management system, and the wide-area measurement system data interface is configured to receive the frequency, damping, and participation factor data streams output by the modal feature recognition logic in real time.

[0062] The adaptive control strategy mapping function performs normalization processing on the received data stream, generates control parameter adjustment amounts through forward propagation calculation, and sends them to the local controller of the designated energy storage power station after security verification, thus completing the integration of the evaluation results into the grid energy management system.

[0063] The beneficial effects of this invention are:

[0064] This invention effectively solves the problem of inaccurate stability boundary quantization caused by the inability of existing methods to capture the implicit frequency domain coupling between the virtual inertial control of grid-connected energy storage and the grid oscillation modes due to the inability of linearized models or preset operating conditions to capture such implicit frequency domain coupling. It employs multi-source data synchronization processing and quality assessment techniques to obtain high-quality data sources, combines tensor decomposition methods to accurately identify the characteristics of implicit coupling modes, utilizes deep reinforcement learning agents and meta-model optimizers to collaboratively explore optimal parameters, and ensures the reliability of the assessment results through high-fidelity verification and calibration mechanisms. The generated comprehensive capability assessment report integrates transient stability boundaries, implicit coupling risk maps, robust parameter feasible regions, and adaptive control strategy mapping functions, directly supporting the dynamic decision-making of the upper-level controller of the energy storage system and the grid energy management system, thus improving the accuracy and engineering practicality of transient stability assessment for high-proportion renewable energy grids. Attached Figure Description

[0065] 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.

[0066] Figure 1 This is a flowchart illustrating the method for evaluating the ability of grid-based energy storage to improve the transient stability of the power grid according to the present invention. Detailed Implementation

[0067] 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.

[0068] To better understand the purpose of this invention, the invention will now be described in further detail.

[0069] This invention provides a method for evaluating the ability of grid-based energy storage to improve the transient stability of the power grid, comprising:

[0070] Step 1: Collect synchronous phasor data from the wide-area measurement system and command stream data from the local controller of the grid-type energy storage system. Perform time synchronization processing on the two types of data, perform data quality assessment, and obtain high-quality data sources.

[0071] Step 2: Using the high-quality data source, construct a digital twin simulation environment, which includes an equivalent model of the power grid electromechanical transients and a detailed model of the energy storage controller.

[0072] Step 3: In the digital twin simulation environment, inject optimized excitation signals, collect system response data, and use tensor decomposition method to identify the implicit coupling mode characteristics between virtual inertial control and power grid oscillation mode to obtain the implicit coupling mode characteristics.

[0073] Step 4: Input the implicit coupled modal features into the deep reinforcement learning agent. The deep reinforcement learning agent uses the energy storage control parameters as the action space and the stability index as the reward function. It interacts with the digital twin simulation environment to explore the optimal parameters and obtain the optimal parameter set recommended by the agent. At the same time, a meta-model optimizer is used to establish a surrogate model from the parameter space to the stability index to predict the Pareto front and obtain the Pareto front solution predicted by the meta-model.

[0074] Step 5: Feed the optimal parameter set recommended by the agent and the Pareto front solution predicted by the meta-model back to the digital twin simulation environment for high-fidelity verification to obtain the verification results; compare the consistency of the verification results, the reward value of the agent in the process of exploring the optimal parameters, and the predicted value corresponding to the Pareto front solution predicted by the meta-model; if there is a deviation, calibrate the reward function of the deep reinforcement learning agent according to the verification results, calibrate the surrogate model of the meta-model optimizer, and repeat Step 4 and this step until convergence, to obtain the converged agent policy, the calibrated surrogate model, and the verification data;

[0075] Step 6: Integrate the converged agent strategy, the calibrated surrogate model, the verification data, and the implicit coupling mode features to generate a comprehensive capability assessment report. The comprehensive capability assessment report generates transient stability boundaries and implicit coupling risk maps based on the verification data and the implicit coupling mode features, generates robust parameter feasible regions based on the calibrated surrogate model, and generates adaptive control strategy mapping functions based on the converged agent strategy. The comprehensive capability assessment report is then integrated into the upper-level controller of the energy storage system and into the grid energy management system.

[0076] The assessment method for enhancing the transient stability of the power grid through grid-based energy storage begins with the collaborative acquisition and processing of multi-source data. Synchronous phasor data from grid nodes is acquired via a wide-area measurement system, while simultaneously collecting low-level command stream data from the local controller of the grid-based energy storage system. These two types of data are synchronized and aligned using a precision clock protocol, forming frequency sequences, voltage phasor sequences, and control command sequences under a unified time axis. Subsequently, a dynamic time warping algorithm is used to verify the integrity of the data sequences. After identifying abnormal data segments, a pre-trained spatiotemporal generative adversarial network is invoked for data repair. The generator generates repaired values ​​based on the spatiotemporal correlation characteristics of normal data, while the discriminator verifies the consistency of data distribution, ultimately outputting a continuous, high-quality data source. This step establishes the data foundation for subsequent analysis and solves the problems of data asynchrony and poor quality in existing assessments.

[0077] A digital twin simulation environment is constructed based on high-quality data sources. Utilizing power grid topology connections and real-time measurement data, the Ward equivalent method is employed to dynamically equivalence the remote power grid, preserving all generator nodes and critical load nodes within the study area to form an electromechanical transient equivalent model. Simultaneously, the energy storage controller command stream is compiled into executable model code for a real-time digital simulator and injected into the simulation loop via a power hardware-in-the-loop interface. By configuring the computational step size, communication delay parameters, and output limiting, a real-time data exchange channel is established between the electromechanical transient simulation and electromagnetic transient simulation subsystems. This dual-timescale simulation architecture ensures both the computational efficiency of the power grid-level simulation and achieves accurate simulation of equipment-level dynamics.

[0078] Active detection and mode identification are implemented in a digital twin simulation environment. A quantum annealing algorithm is used to optimize and generate a wideband excitation signal combination within a preset frequency band, with signal energy concentration and orthogonality as optimization objectives. The optimized excitation signal is injected into the voltage reference point of the energy storage device, and multi-channel response data such as energy storage output power, generator power angle, and tie-line power are simultaneously acquired. After performing a short-time Fourier transform on the response data, a three-dimensional time-frequency-space response tensor is constructed. Core modal components in the tensor are extracted through high-order singular value decomposition. Based on the frequency factor distribution and spatial factor weights, oscillation modes strongly correlated with virtual inertial control are selected, accurately identifying implicit coupling characteristics.

[0079] Implicitly coupled modal features serve as input parameters to drive the optimization process. A deep reinforcement learning agent constructs a state vector using modal amplitude, frequency, and participation factors, and uses virtual inertial parameter adjustments as its action space to explore policies within a digital twin environment. After each action, the reward function is updated by calculating the damping ratio of the dominant mode and the modal amplitude decay rate, thereby optimizing the policy network. A parallel-running meta-model optimizer uses Latin hypercube sampling to generate parameter samples, establishes a surrogate model of parameters and stability indices based on the Kriging method, and guides the Pareto front search through the expected improvement criterion. This dual-path optimization mechanism balances global search efficiency with local fine-tuning.

[0080] A closed-loop verification and calibration mechanism is established to ensure the reliability of the evaluation. The agent's recommended parameters and the meta-model's predicted solutions are injected into a digital twin environment for high-fidelity verification, comparing the deviations between the actual stability indicators and the predicted values. When a systematic deviation is detected, the reward function coefficients are reweighted based on the verification results, and new data is added to the training set to update the agent model's hyperparameters. Through multiple iterations of calibration, the agent's policy and meta-model predictions converge towards physical reality, forming a self-correcting evaluation mechanism.

[0081] The final comprehensive capability assessment report for engineering applications is generated. By performing gridded sampling in the operating mode-parameter space, a latent coupling risk map is generated using the modal identification process. The feasible region of parameters satisfying stability constraints is extracted from the calibrated surrogate model, and the envelope boundary is fitted using a support vector machine algorithm. The converged agent policy network is encapsulated as an adaptive control mapping function, which can output parameter adjustment amounts based on real-time modal characteristics. The assessment report is integrated into the energy storage upper-level controller and the grid energy management system, where the feasible region of parameters is used for offline tuning, and the adaptive mapping function supports online adjustment, realizing a complete technology chain from static assessment to dynamic control.

[0082] In the data synchronization processing stage, a precise clock protocol is used to align the time stamps of multi-source data from the wide-area measurement system. Addressing the clock drift issue in the synchronization phasor measurement units of different plants, transmission delay compensation values ​​are calculated to map the data of each node to a unified time axis. When parsing communication messages from the energy storage inverter, application layer data frames are decoded layer by layer according to the protocol document provided by the manufacturer, extracting the time-series change trajectories of key parameters such as power reference values ​​and voltage reference values ​​in the virtual synchronous machine control loop. In the data quality assessment stage, a dynamic time warping algorithm is introduced to calculate the similarity between sequences. An abnormal data segment is identified through a sliding window detection mechanism, and a spatiotemporal generative adversarial network is used to repair the identified missing data segments. The generator network learns the spatiotemporal correlation features of normal data through an encoder-decoder structure, while the discriminator guides the training process by comparing the KL divergence between the generated data and the real data distribution, ultimately outputting a repaired sequence with spatiotemporal consistency.

[0083] The equivalent model of the power grid's electromechanical transients is constructed using the Ward-PV equivalent method, which equates the external system to equivalent generators and equivalent impedances connected to the boundary bus based on the real-time topology. During the equivalence process, the dynamic characteristics of all generator nodes within the study area must be preserved, while considering the static characteristics of key load nodes. The detailed model of the energy storage controller reconstructs the control algorithm logic by parsing command stream data, transforming discrete command sequences into a continuous state-space model. The model compilation phase requires converting the control algorithm into C language or FPGA code supported by a real-time digital simulator, generating an executable module that can run on the simulation hardware using a cross-compilation toolchain. The power hardware-in-the-loop interface configuration needs to coordinate the electromechanical transient simulation step size with the electromagnetic transient simulation step size, establishing a data exchange buffer management mechanism to ensure normal interaction of simulation data at different time scales.

[0084] The quantum annealing algorithm constructs the Ising model in the excitation signal optimization, mapping the problem of maximizing the band energy concentration to a ground state search problem. By adjusting the quantum tunneling effect parameters, a global search is performed in the solution space to obtain the combination of excitation signals with optimal orthogonality. In the signal injection stage, the optimized excitation signal is superimposed onto the voltage reference value of the energy storage device, while multi-channel response data is acquired. Response data processing uses short-time Fourier transform to obtain the time-spectrum matrix, and a windowing function is used to control the spectral leakage effect. In the tensor construction stage, the time-spectrum matrices of each channel are stacked along the spatial dimension to form a three-dimensional tensor data structure. Higher-order singular value decomposition solves the tensor decomposition problem using alternating least squares, extracting a frequency factor matrix that reflects the modal spectral characteristics, and a spatial factor matrix that characterizes the mode shape distribution.

[0085] The deep reinforcement learning agent employs a deterministic policy gradient algorithm, with an Actor-Critic architecture for the policy network. The Actor network outputs continuous action values, while the Critic network evaluates the state-action function. State vector construction requires normalization of implicitly coupled modal features to eliminate dimensionality effects. The reward function design comprehensively considers the damping ratio increase and modal amplitude decay rate, generating a scalar reward value through weighted summation. The meta-model optimizer uses the Kriging method to construct a Gaussian process regression model, utilizing a variogram to describe the correlation in the parameter space. The desired improvement acquisition function selects the most promising sample points in the parameter space by balancing exploration and utilization. Co-optimization between the agent and the meta-model is achieved through an experience replay mechanism, adding sample points recommended by the meta-model to the agent's training data pool.

[0086] During calibration, statistical hypothesis testing methods are used to assess the significance of the predicted deviation, and t-tests are used to compare the difference between the agent's reward value and the actual index value. Reward function weights are adjusted using gradient descent, with weight coefficients dynamically updated based on the deviation direction. The meta-model training set is expanded using an incremental learning strategy, merging new validation data with historical data to retrain the surrogate model. Gaussian process regression hyperparameter optimization employs maximum likelihood estimation, solving for the optimal hyperparameter combination using the conjugate gradient method. A convergence criterion is set for the calibration loop; the calibration process terminates when the deviation value is less than a threshold for three consecutive iterations.

[0087] Risk map generation employs inverse distance weighted interpolation to transform discrete grid sampling point data into a continuous spatial distribution map. The support vector machine classifier uses radial basis function kernels and solves for the classification hyperplane using a sequential minimum optimization algorithm. The adaptive control policy mapping function constructs a forward propagation computation graph by storing the weight parameters of the agent policy network. The parameter feasible region envelope surface extraction uses the alpha-shape algorithm to reconstruct the geometric boundary from point cloud data. During system integration, a standardized data interface is designed to convert the evaluation results into a data format recognizable by the energy management system. The real-time control parameter adjustment calculation module is deployed on an edge computing device and communicates with the upper-layer control system via the OPC-UA protocol.

[0088] The core challenge of grid-based energy storage in actual power grids lies in the implicit coupling between virtual inertial control and the inherent oscillation modes of the power grid. This coupling may induce subsynchronous oscillations during transient processes such as large-scale wind power disconnection, leading to inaccurate stability boundary judgments in existing linear model-based evaluation methods. This solution achieves accurate quantification of energy storage support capabilities by constructing an evaluation system that combines a digital twin simulation environment with intelligent optimization.

[0089] The first challenge in implementation is the fusion of multi-source heterogeneous data. The synchronous phasor data acquired by the wide-area measurement system differs in time scale from the command stream data of the energy storage local controller, requiring millisecond-level alignment using a precise clock protocol. Specifically, to address clock drift in the synchronous phasor measurement units of different plants, a unified time axis is established by calculating transmission delay compensation values. Parsing the communication messages of the energy storage controller requires decoding data frames layer by layer according to the protocol documentation provided by the manufacturer to extract the dynamic parameters of the virtual synchronous machine control loop. Data quality assessment employs a dynamic time warping algorithm to detect abnormal segments, while a spatiotemporal generative adversarial network (GAN) outputs a continuous and reliable data source by reconstructing missing data using a generator and verifying data distribution consistency using a discriminator.

[0090] The construction of a digital twin simulation environment requires a balance between simulation accuracy and computational efficiency. The Ward equivalent method is used to represent the dynamics of the remote power grid as equivalent generators and impedances on the boundary bus, preserving the dynamic characteristics of all generator nodes within the study area. The energy storage controller command stream is compiled into executable code for the real-time digital simulator through model compilation and injected into the simulation loop via the power hardware-in-the-loop interface. Key configurations include setting the data exchange cycle between electromechanical transient simulations and electromagnetic transient simulations, coordinating simulation step sizes at different time scales, and ensuring accurate interaction between the rapid switching dynamics of the energy storage device and the slow electromechanical dynamics of the power grid.

[0091] The identification of implicitly coupled modes employs a combination of active detection and tensor decomposition. The quantum annealing algorithm maps the excitation signal optimization problem to a ground state search of the Ising model, generating orthogonal signal combinations with maximum energy concentration within a preset frequency band. After injecting the optimized signal, the time-frequency spectrum matrix of the multi-channel response is obtained through short-time Fourier transform, constructing a three-dimensional time-frequency-space tensor. Higher-order singular value decomposition (SVD) decouples the tensor into core modal components. By analyzing the spectral distribution of the frequency factor matrix and the weight allocation of the spatial factor matrix, oscillation modes strongly correlated with virtual inertial control are selected.

[0092] A deep reinforcement learning agent and a meta-model optimizer constitute a collaborative optimization system. The agent constructs its state space using implicitly coupled modal features and explores the optimal adjustment strategy for virtual inertia parameters through a deterministic policy gradient algorithm. The meta-model uses the Kriging method to establish a surrogate model of parameters and stability indices, and uses the expected improvement criterion to guide the selection of sampling points. Data generated by agent interactions updates the meta-model in real time, and the Pareto front predicted by the meta-model in turn guides the agent's exploration direction, forming a closed-loop mechanism of bidirectional optimization.

[0093] Verification and calibration mechanisms are crucial for ensuring the reliability of the evaluation. The agent's recommended parameters and the meta-model's predicted solutions are injected into a high-fidelity simulation environment, and statistical hypothesis testing methods are used to compare the deviations between the actual and predicted values. When a systematic deviation is detected, the reward function's indicator weights are reweighted based on the simulation results, and the agent model's training set is updated through incremental learning. The iterative calibration process continues until the agent's policy, meta-model predictions, and simulation results converge within a preset tolerance range.

[0094] The comprehensive capability assessment report generation phase requires transforming multidimensional data into engineering-usable decision support tools. Grid sampling is performed in the operation mode-parameter space, and an implicit coupling risk map is generated through inverse distance weight interpolation. A support vector machine classifier performs boundary fitting on the feasible solution set output by the meta-model, generating a robust parameter feasible domain envelope. An adaptive control policy mapping function encapsulates the agent policy network parameters, achieving end-to-end mapping from real-time modal features to control parameter adjustment amounts.

[0095] The final integrated design employs a hierarchical control architecture. The upper-level energy storage controller has a pre-configured feasible domain database of parameters. Through periodic inspections, it compares the operating point with boundary positions, triggering an early warning mechanism. The grid energy management system deploys a strategy mapping function module, which receives real-time modal characteristics through a wide-area measurement system interface. These characteristics are then normalized and calculated using forward propagation to generate control commands. All data interactions utilize standard communication protocols, ensuring seamless integration from evaluation results to control execution. This implementation effectively addresses the shortcomings of existing evaluation methods in characterizing implicit coupling, providing technical support for the safe and stable operation of high-proportion renewable energy grids.

[0096] Embodiment 1 of this invention: In areas with concentrated wind power grid integration, when a large-scale wind turbine disconnection fault occurs, the system frequency drops rapidly. Existing assessment methods struggle to capture the interaction between the virtual inertial response of energy storage and the subsynchronous oscillation modes of the tie line. In this embodiment, frequency and voltage phasor data provided by a wide-area measurement system are first acquired, along with command stream data from the local energy storage controller. After aligning timestamps using a precision clock protocol, a dynamic time warping algorithm is employed to detect abnormal data segments, and a spatiotemporal generative adversarial network is used to repair missing data, forming a high-quality data source. Next, a digital twin simulation environment is constructed. Based on real-time topology data, the Ward equivalence method is used to retain key nodes such as the wind farm grid connection point. The energy storage controller command stream is compiled into real-time simulation code and injected into the simulation loop via a power hardware-in-the-loop interface. In the simulation environment, a quantum annealing algorithm is used to generate an optimized excitation signal for the subsynchronous frequency band. After injecting this signal into the energy storage voltage reference point, response data such as generator power angle and line power are acquired. A three-dimensional response tensor is constructed using short-time Fourier transform, and high-order singular value decomposition is used to extract oscillation mode features strongly correlated with virtual inertial control. The deep reinforcement learning agent takes modal amplitude and participation factor as state inputs and explores the optimal strategy by adjusting virtual inertial parameters. Simultaneously, a meta-model optimizer establishes a surrogate model of parameters and stability. After high-fidelity verification and calibration, the generated risk map shows that energy storage parameter settings can induce subsynchronous oscillations under specific operating conditions. The adaptive control strategy mapping function can recommend damping coefficient adjustments in real time. Ultimately, the feasible domain of parameters is integrated into the wind farm control system, effectively avoiding resonance risks during frequency drop.

[0097] Embodiment 2 of this invention: In distribution networks with high photovoltaic power generation penetration, sudden changes in solar irradiance can cause drastic fluctuations in photovoltaic output, leading to local voltage instability. This embodiment addresses this scenario by first synchronously collecting voltage and current data from the distribution automation system and control command streams from the energy storage converter. During data alignment, the focus is on addressing the timestamp differences between the photovoltaic inverter and the energy storage controller, employing a sliding window detection algorithm to identify data jumps caused by communication interruptions. The repaired data is used to construct a digital twin model of the distribution network, where the photovoltaic power station uses an equivalent impedance model, and the energy storage system reproduces its voltage control logic through detailed controller code. The excitation signal optimization stage focuses on the supersynchronous frequency band, generating orthogonal signal sequences through a quantum annealing algorithm and injecting them into the energy storage control loop, while simultaneously collecting response data such as node voltage and reactive power. The implicit coupling modes discovered after tensor decomposition show a strong correlation between virtual inertial parameters and voltage oscillation modes. During the agent exploration process, voltage deviation and mode amplitude are used as reward function indicators, and the meta-model rapidly explores the parameter space through Latin hypercube sampling. The verification process revealed that when photovoltaic output suddenly drops, the existing parameter settings can cause voltage collapse, while the calibrated agent-recommended parameters can suppress voltage fluctuations within a safe range. The resulting robust parameter feasible region is integrated into the distribution management system in the form of a two-dimensional graph. When a sudden change in photovoltaic power is detected, the adaptive mapping function automatically adjusts the energy storage reactive power compensation parameters to maintain the stability of the point of common coupling voltage.

[0098] The above embodiments demonstrate that the present invention, through a closed-loop framework of data-driven modeling, intelligent optimization, and real-time calibration, can effectively address the stability assessment challenges caused by implicit coupling in different scenarios. Embodiment one focuses on the frequency stability problem of large-scale power grids, while Embodiment two focuses on the voltage control problem of distribution networks, jointly verifying the universality and engineering applicability of the method.

Claims

1. A method for evaluating the ability of grid-based energy storage to improve the transient stability of the power grid, characterized in that, include: Step 1: Collect synchronous phasor data from the wide-area measurement system and command stream data from the local controller of the grid-type energy storage system. Perform time synchronization processing on the two types of data, perform data quality assessment, and obtain high-quality data sources. Step 2: Using the high-quality data source, construct a digital twin simulation environment, which includes an equivalent model of the power grid electromechanical transients and a detailed model of the energy storage controller. Step 3: In the digital twin simulation environment, inject optimized excitation signals, collect system response data, and use tensor decomposition method to identify the implicit coupling mode characteristics between virtual inertial control and power grid oscillation mode to obtain the implicit coupling mode characteristics. Step 4: Input the implicit coupled modal features into the deep reinforcement learning agent. The deep reinforcement learning agent uses the energy storage control parameters as the action space and the stability index as the reward function. It interacts with the digital twin simulation environment to explore the optimal parameters and obtain the optimal parameter set recommended by the agent. At the same time, a meta-model optimizer is used to establish a surrogate model from the parameter space to the stability index to predict the Pareto front and obtain the Pareto front solution predicted by the meta-model. Step 5: Feed the optimal parameter set recommended by the agent and the Pareto front solution predicted by the meta-model back to the digital twin simulation environment for high-fidelity verification to obtain the verification results; compare the consistency of the verification results, the reward value of the agent in the process of exploring the optimal parameters, and the predicted value corresponding to the Pareto front solution predicted by the meta-model; if there is a deviation, calibrate the reward function of the deep reinforcement learning agent according to the verification results, calibrate the surrogate model of the meta-model optimizer, and repeat Step 4 and this step until convergence, to obtain the converged agent policy, the calibrated surrogate model, and the verification data; Step 6: Integrate the converged agent strategy, the calibrated surrogate model, the verification data, and the implicit coupling mode features to generate a comprehensive capability assessment report. The comprehensive capability assessment report generates transient stability boundaries and implicit coupling risk maps based on the verification data and the implicit coupling mode features, generates robust parameter feasible regions based on the calibrated surrogate model, and generates adaptive control strategy mapping functions based on the converged agent strategy. The comprehensive capability assessment report is then integrated into the upper-level controller of the energy storage system and into the grid energy management system.

2. The method for evaluating the ability of grid-based energy storage to improve the transient stability of the power grid according to claim 1, characterized in that, Step 1 involves performing time synchronization processing on the two types of data and conducting data quality assessment, including: Based on a precision clock protocol, the timestamps of synchronous phasor measurement unit data from different plants are aligned to a unified reference time axis to obtain the aligned frequency sequence and voltage phasor sequence. The communication messages of the grid-type energy storage inverter are analyzed to extract the time series of internal state variables of the virtual synchronous machine control loop. The internal state variables include power reference value, voltage reference value, and phase-locked loop output angle. Integrity checks are performed on the aligned frequency sequence, voltage phasor sequence, and extracted internal state variable time sequence. Anomaly pattern detection based on dynamic time warping algorithm is used to identify and mark missing data segments and data transition segments. A pre-trained spatiotemporal generative adversarial network is invoked to repair the marked missing data segments and data jump segments. The generator generates repaired data based on the spatiotemporal correlation of adjacent normal data points, and the discriminator judges the consistency between the repaired data and the original normal data distribution, thereby outputting the repaired continuous data sequence to form the high-quality data source.

3. The method for evaluating the ability of grid-based energy storage to improve the transient stability of the power grid according to claim 2, characterized in that, Step 2, which involves constructing a digital twin simulation environment, includes: Using the power grid topology connections and measurement data from the high-quality data source, the Ward equivalent method is used to dynamically equivalence the remote power grid, retaining all generator nodes and key load nodes in the study area to form the electromechanical transient equivalent model. The instruction stream data of the grid-type energy storage local controller is compiled into model description language code supported by a real-time digital simulator to obtain executable controller model code. By using the power hardware-in-the-loop simulation interface, the executable controller model code is injected into the simulation loop corresponding to the electromechanical transient equivalent model. The calculation step size, communication delay parameters, and output limiting circuit are configured to establish a real-time data exchange channel between the electromechanical transient simulation subsystem and the electromagnetic transient simulation subsystem, thereby constructing the digital twin simulation environment.

4. The method for evaluating the ability of grid-based energy storage to improve the transient stability of the power grid according to claim 3, characterized in that, Step 3, which involves injecting optimized excitation signals, collecting system response data, and identifying latent coupled modal characteristics using tensor decomposition, includes: Set the frequency band range to be detected, and use the quantum annealing algorithm to search for a set of broadband excitation signal waveforms with the maximum energy concentration and mutual orthogonality within the frequency band range to obtain the optimized excitation signal; The optimized excitation signal is injected into the voltage reference point of the energy storage device in the digital twin simulation environment. At the same time, the active power signal of the energy storage output port, the power angle signal of the selected generator in the system, and the power signal of the key tie line are collected as multi-channel system response data. For each channel of the multi-channel system response data, a short-time Fourier transform is performed to obtain the time-spectrum matrix of each channel; Stack the time-frequency matrix of all channels along the spatial channel dimension to construct a three-dimensional response tensor of time, frequency, and space; Perform a higher-order singular value decomposition on the response tensor to extract the decomposed core tensor, time factor matrix, frequency factor matrix, and spatial factor matrix; By analyzing the frequency factor matrix and spatial factor matrix, oscillation modes with significant spatial factor weights in the energy storage control channel and frequency factors distributed near the virtual inertial control frequency band are selected. These selected oscillation modes are used as the implicit coupling mode features.

5. The method for evaluating the ability of grid-based energy storage to improve the transient stability of the power grid according to claim 4, characterized in that, The deep reinforcement learning agent described in step 4 uses energy storage control parameters as its action space and a stability index as its reward function to interact with the digital twin simulation environment to explore optimal parameters, including: Construct the state vector of the agent, the state vector including the amplitude, frequency, and participation factor corresponding to each generator of the implicit coupled mode feature; The action vector of the intelligent agent is defined as the adjustment amount of the time constant and damping coefficient of the virtual inertial element of the grid-type energy storage; The action vector is executed in the digital twin simulation environment to obtain the system's dynamic response; Based on the system's dynamic response, calculate the damping ratio of the system's dominant low-frequency oscillation mode and the amplitude attenuation rate of the implicitly coupled mode characteristics. Based on the increase in the damping ratio and the increase in the amplitude decay rate, the current reward value is calculated using a preset reward function, and the policy network of the deep reinforcement learning agent is updated using the current reward value.

6. The method for evaluating the ability of grid-based energy storage to improve the transient stability of the power grid according to claim 5, characterized in that, Step 4, which involves using a meta-model optimizer to build a surrogate model and predict the Pareto front, includes: Within the domain of the energy storage control parameters, an initial sample point set is generated using the Latin hypercube sampling method; The control parameters corresponding to each sample point in the initial sample point set are input into the digital twin simulation environment to obtain the stability index output corresponding to each sample point. The stability index includes transient stability limit and coupled mode amplitude. Using all sample points and their corresponding stability index outputs, a Gaussian process regression model based on the Kriging method is trained to obtain the initial surrogate model. The expected improved acquisition function is adopted to guide the selection of new sample points in the parameter space. The newly selected sample points are input into the digital twin simulation environment to obtain the corresponding stability index output. The initial surrogate model is iteratively updated using the new sample points and their stability index output until the Pareto front predicted by the model converges, thereby obtaining the surrogate model of the meta-model optimizer.

7. The method for evaluating the ability of grid-based energy storage to improve the transient stability of the power grid according to claim 6, characterized in that, The calibration meta-model mentioned in step 5, which calibrates the agent's reward function, includes: The actual stability index obtained from the high-fidelity verification is compared with the estimated reward value during the interaction process of the intelligent agent to calculate the prediction deviation. Based on the magnitude and direction of the predicted deviation, the weight coefficients of the damping ratio and modal amplitude in the reward function are dynamically adjusted to calibrate the reward function of the deep reinforcement learning agent. The new parameter-performance data pairs generated by the high-fidelity verification are added to the training sample library of the meta-model optimizer to obtain the expanded training sample library. Using the expanded training sample library, the hyperparameter optimization of the Gaussian process regression model is re-executed to update the surrogate model and calibrate the surrogate model of the meta-model optimizer.

8. The method for evaluating the ability of grid-based energy storage to improve the transient stability of the power grid according to claim 7, characterized in that, Step 6, generating the comprehensive capability assessment report, includes: System sampling is performed on a two-dimensional grid consisting of a set of typical power grid operating modes and an energy storage control parameter range to obtain a grid sampling point set; For each sampling point in the grid sampling point set, the implicit coupled mode feature identification process is invoked to calculate the coupled mode amplitude corresponding to that point, thereby obtaining the coupled mode amplitude data of all sampling points; The coupled modal amplitude data of all the sampling points are visualized as the risk map using color mapping. From the calibrated meta-model proxy model, extract all parameter points that satisfy the preset stability threshold constraint to form a feasible solution set; A support vector machine algorithm is used to learn the feasible solution set to obtain the classification hyperplane equation describing the boundary of the robust parameter feasible region, and the envelope surface of the robust parameter feasible region is obtained. The policy network parameters after the deep reinforcement learning agent training converges are solidified and encapsulated into the adaptive control policy mapping function that receives real-time modal features and outputs parameter adjustment amounts, thus obtaining the adaptive control policy mapping function. The comprehensive capability assessment report is generated by combining the risk map, the robustness parameter feasible domain envelope, the adaptive control strategy mapping function, and the transient stability boundary obtained in step 5.

9. The method for evaluating the ability of grid-based energy storage to improve the transient stability of the power grid according to claim 8, characterized in that, The transient stability boundary included in generating the comprehensive capability assessment report in step 6 includes: In the digital twin simulation environment, three-phase short-circuit faults and single-phase ground faults are set, and the faults are specified to occur at the midpoint of the main transmission line and the substation bus. A binary search algorithm is used to adjust the fault duration, with the upper limit of the initial search interval set to the system protection action time limit and the lower limit set to zero. Simulations were run for each fault duration, and the relative power angle difference of all generator rotors in the system was monitored. The fault duration corresponding to the power angle difference exceeding the critical threshold of 180 degrees was recorded as the transient power angle stability boundary. By changing the system's operating mode and repeating the binary search process, the set of transient stable boundary points under different operating modes can be obtained. Using the transient stability boundary point set under the different operating modes, a stable boundary surface is constructed with the critical line transmission power and fault clearing time as coordinate axes, which serves as the transient stability boundary.

10. The method for evaluating the ability of grid-based energy storage to improve the transient stability of the power grid according to claim 9, characterized in that, Step 6, which involves integrating the evaluation results into the upper-level controller of the energy storage system and into the grid energy management system, includes: The mathematical description of the feasible domain envelope of the robustness parameter is written into the parameter database of the upper controller. The parameter inspection logic is configured to periodically read the parameter description in the database and compare the relative position of the current operating point with the boundary of the feasible domain. When the operating point approaches the boundary, an early warning is triggered and a safe parameter combination within the feasible domain is recommended, thus completing the integration of the evaluation results into the upper controller of the energy storage system. The adaptive control strategy mapping function is deployed in the safety constraint logic of the power grid energy management system, and the wide-area measurement system data interface is configured to receive the frequency, damping, and participation factor data streams output by the modal feature recognition logic in real time. The adaptive control strategy mapping function performs normalization processing on the received data stream, generates control parameter adjustment amounts through forward propagation calculation, and sends them to the local controller of the designated energy storage power station after security verification, thus completing the integration of the evaluation results into the grid energy management system.

Images