A method for inhibiting wide-frequency oscillation of offshore wind power grid connection by network-structured energy storage

Abstract

This invention discloses a method for suppressing broadband oscillations during offshore wind power grid connection using grid-based energy storage, relating to the field of power control technology. The method includes: using a convolutional neural network (CNN) to extract features from collected grid voltage, grid current, and grid frequency data to identify grid state characteristics; and using generative adversarial networks (GANs) to simulate the behavior patterns of the grid under normal operating conditions. Based on the grid state characteristics, a deep Q-network (DQN) is used to dynamically adjust the charging and discharging strategy of the energy storage device. A machine learning model is used to analyze battery operating data and predict the battery health status to obtain a predicted value. The optimal charging and discharging strategy is selected by combining the charging and discharging strategy and the predicted battery health status. Based on the optimal charging and discharging strategy and grid state characteristics, a feedforward control combined with a feedback control algorithm is applied to quickly respond to grid load changes, adjust the grid voltage and grid current output by the inverter, regulate reactive power to improve the grid power factor, and output the adjusted grid connection parameters.

Description

Technical Field

[0001] This invention relates to the field of power control technology, and in particular to a method for suppressing broadband oscillations during offshore wind power grid connection through grid-connected energy storage. Background Technology

[0002] With the rapid increase in offshore wind power penetration, broadband oscillation problems are becoming increasingly prominent. Traditional suppression methods mainly include adaptive fusion control of grid-connected energy storage and additional damping devices. Positive damping is provided by adjusting converter parameters or introducing external equipment. Although this can alleviate the oscillation risk in specific scenarios, it suffers from problems such as dynamic response lag and insufficient multi-objective coordination. For example, Tsinghua University's fusion control strategy relies on a fixed ratio of grid-connected / grid-connected characteristics, making it difficult to adapt to instantaneous changes in grid conditions. Sifang Co., Ltd.'s additional device has limited suppression effect on time-varying broadband oscillations due to parameter tuning accuracy. In addition, the dual-loop optimized control strategy for energy storage mentioned in Baidu Scholar focuses on low-frequency oscillations but does not fully consider the coupling effect of grid condition characteristics and battery health status, resulting in insufficient robustness of the strategy.

[0003] This invention proposes an innovative method based on convolutional neural networks to extract grid state features and combined with deep reinforcement learning to dynamically adjust energy storage strategies. Compared with existing technologies, this invention achieves adaptive suppression of multi-frequency oscillations by integrating physical model predictive control (MPC) and data-driven algorithms, while taking into account grid stability and battery life management. The multi-objective optimization mechanism improves the universality and dynamic response efficiency of the strategy, making it particularly suitable for the governance of complex oscillation modes in scenarios with a high proportion of renewable energy access. Summary of the Invention

[0004] In view of the aforementioned existing problems, the present invention is proposed.

[0005] Therefore, this invention provides a method for suppressing broadband oscillations of offshore wind power grid connection through grid-connected energy storage to solve the problems of dynamic response lag and insufficient multi-objective coordination.

[0006] To solve the above-mentioned technical problems, the present invention provides the following technical solution:

[0007] In a first aspect, the present invention provides a method for suppressing broadband oscillations during offshore wind power grid connection through grid-connected energy storage, comprising:

[0008] include:

[0009] Convolutional neural networks (CNNs) are used to extract features from the collected grid voltage, grid current, and grid frequency data to identify grid state characteristics. Based on these grid state characteristics, generative adversarial networks (GANs) are used to simulate the behavior patterns of the grid under normal operating conditions and identify potential abnormal oscillation patterns.

[0010] Based on the characteristics of the power grid state, a deep Q-network (DQN) is used to dynamically adjust the charging and discharging strategy of the energy storage device. A machine learning model is used to analyze the battery operation data and predict the battery health status to obtain the battery health status prediction value. The optimal charging and discharging strategy is selected by combining the charging and discharging strategy and the battery health status prediction value.

[0011] Based on the optimal charging and discharging strategy and grid state characteristics, feedforward control combined with feedback control algorithm is applied to quickly respond to grid load changes, adjust the grid voltage and grid current output by the inverter, regulate reactive power to improve the grid power factor, and output the adjusted grid connection parameters.

[0012] The integrated physical model predictive control (MPC) and deep reinforcement learning (DRL) take grid connection parameters and grid state characteristics as inputs, dynamically adjust inverter parameters, suppress oscillations in different grid frequency ranges, and output an oscillation suppression strategy.

[0013] As a preferred embodiment of the method for suppressing broadband oscillations of offshore wind power grid connection through grid-type energy storage as described in this invention, the specific steps for identifying grid state characteristics are as follows:

[0014] The collected grid voltage, grid current, and grid frequency are divided into multiple data blocks according to a fixed time window. Each data block is converted into a three-dimensional tensor and used as the input of a convolutional neural network. Local features are extracted from the three-dimensional tensor through multiple convolutional layers. The initial convolutional layer extracts low-level features, and subsequent convolutional layers stack and combine low-level features to generate higher-level grid state features. Each convolutional layer is followed by a max pooling layer to reduce the feature dimension, and the ReLU activation function is used to enhance the nonlinear expressive power. Finally, a feature vector containing the grid operating state is output.

[0015] As a preferred embodiment of the method for suppressing broadband oscillations of offshore wind power grid connection through grid-type energy storage described in this invention, the specific steps for identifying potential abnormal oscillation modes are as follows:

[0016] The generator of the generative adversarial network receives the power grid state features output by the convolutional neural network and generates a feature distribution simulating the normal operation of the power grid through a fully connected layer. The discriminator receives the power grid state features and the feature distribution generated by the generator and determines whether the input power grid state features belong to the normal operation state through a multilayer perceptron. During the adversarial training phase, the parameters of the generator and discriminator are optimized so that the generator generates normal pattern features with high fitting degree. During the real-time detection phase, if the generator cannot generate matching normal pattern features or the discriminator determines an anomaly, it is determined that there is a potential abnormal oscillation mode in the current power grid.

[0017] As a preferred embodiment of the method for suppressing broadband oscillations of offshore wind power grid connection through grid-based energy storage described in this invention, the specific steps of dynamically adjusting the charging and discharging strategy of the energy storage device using a deep Q-network (DQN) based on grid state characteristics are as follows:

[0018] The charging and discharging strategy of the energy storage device is set as a discrete action space. A multi-objective reward function is defined to integrate grid stability reward, battery health reward and economic reward. The sequence correlation is reduced through experience playback mechanism. The main network parameters are periodically copied to the target network for stable training. The charging and discharging strategy with the highest Q value is selected as the output in real time operation.

[0019] As a preferred embodiment of the method for suppressing broadband oscillations of offshore wind power grid connection through grid-type energy storage as described in this invention, the specific steps of using a machine learning model to analyze battery operating data and predict battery health status to obtain a predicted battery health status value are as follows:

[0020] Battery operation data is collected and analyzed using a gated recurrent neural network model. A feature matrix is ​​constructed that includes the slope of the voltage curve, the rate of change of internal resistance, and the amplitude of temperature fluctuation. Wavelet transform is used to denoise and reconstruct the data. The model is trained using historical cycle life data to predict the battery health status value. If the battery health status value is lower than the health threshold, an early warning is triggered.

[0021] As a preferred embodiment of the method for suppressing broadband oscillations of offshore wind power grid connection through grid-type energy storage described in this invention, the specific steps for selecting the optimal charging and discharging strategy by combining charging and discharging strategies and battery health status prediction values ​​are as follows:

[0022] An objective function is constructed based on grid stability, battery health status prediction, and economic incentives. Constraints are set, and the charging and discharging strategy with the highest total score is selected as the optimal solution based on the weighted score of the objective function.

[0023] As a preferred embodiment of the method for suppressing broadband oscillations of offshore wind power grid connection through grid-connected energy storage as described in this invention, the specific steps of adjusting the grid voltage and grid current output by the inverter to regulate reactive power and improve the grid power factor are as follows:

[0024] By continuously monitoring the actual output voltage and current through a PID controller, calculating the error and generating control signals, the inverter switching frequency or duty cycle is dynamically adjusted to make the output parameters approach the target value, while adjusting reactive power to improve the power factor of the power grid.

[0025] As a preferred embodiment of the method for suppressing broadband oscillations of offshore wind power grid connection through grid-connected energy storage as described in this invention, the integrated physical model predictive control (MPC) and deep reinforcement learning (DRL) dynamically adjust inverter parameters to suppress oscillations within different grid frequency ranges. The specific steps are as follows:

[0026] The physical model predicts the trend of the power grid state over multiple time steps based on the power grid dynamic equations. The deep reinforcement learning network generates inverter parameter adjustment strategies based on the prediction results. The strategy is optimized through a multi-objective reward function to dynamically modify the inverter output parameters to suppress oscillations in different frequency ranges.

[0027] The beneficial effects of this invention are as follows: This invention uses a convolutional neural network (CNN) to extract three-dimensional tensor features from grid voltage, current, and frequency data, improving the accuracy of grid state modeling. It combines generative adversarial networks (GANs) to simulate normal operating modes and identify abnormal oscillations, achieving real-time detection of abnormal modes. Through a deep Q-network (DQN) combined with a multi-objective reward function, it dynamically optimizes charging and discharging strategies, establishing a quantitative balance between grid stability, battery health, and economic efficiency, extending battery life and reducing operation and maintenance costs. It employs a gated recursive unit (GRU) recurrent neural network to predict battery health status and constructs a degradation trend model based on historical lifespan data, ensuring the safety boundary of the energy storage system. It integrates physical model predictive control (MPC) and deep reinforcement learning (DRL) to construct a hybrid decision-making system for cross-frequency band oscillation suppression, enabling inverter parameter adjustments to have both immediate response and global convergence characteristics. Ultimately, it achieves hierarchical suppression of full-frequency domain oscillations, significantly improving the stability margin of the new power system. Attached Figure Description

[0028] To more clearly illustrate the technical solutions of the embodiments of the present invention, the drawings used in the following description of the embodiments will be briefly introduced. Obviously, the drawings described below are only some embodiments of the present invention. For those skilled in the art, other drawings can be obtained based on these drawings without creative effort.

[0029] Fig. 1 This is a flowchart illustrating a method for suppressing broadband oscillations during offshore wind power grid connection through grid-connected energy storage.

[0030] Fig. 2 This is a structural diagram for three-dimensional spatiotemporal feature extraction.

[0031] Fig. 3 A scoring chart for multi-objective charge / discharge strategies.

[0032] Fig. 4 This is a schematic diagram of the joint control for cross-frequency band oscillation suppression. Detailed Implementation

[0033] To make the above-mentioned objects, features and advantages of the present invention more apparent and understandable, the specific embodiments of the present invention will be described in detail below with reference to the accompanying drawings.

[0034] Many specific details are set forth in the following description in order to provide a full understanding of the invention. However, the invention may also be practiced in other ways different from those described herein, and those skilled in the art can make similar extensions without departing from the spirit of the invention. Therefore, the invention is not limited to the specific embodiments disclosed below.

[0035] Secondly, the term "one embodiment" or "embodiment" as used herein refers to a specific feature, structure, or characteristic that may be included in at least one implementation of the present invention. The phrase "in one embodiment" appearing in different places in this specification does not necessarily refer to the same embodiment, nor is it a single or selective embodiment that is mutually exclusive with other embodiments.

[0036] Reference Figs. 1 to 4 This is one embodiment of the present invention, which provides a method for suppressing broadband oscillations during offshore wind power grid connection through grid-connected energy storage, comprising the following steps:

[0037] S1. Use a convolutional neural network (CNN) to extract features from the collected grid voltage, grid current, and grid frequency to identify grid state characteristics. Based on the grid state characteristics, use generative adversarial networks (GANs) to simulate the behavior patterns of the grid under normal operation and identify potential abnormal oscillation patterns.

[0038] The collected grid voltage, grid current, and grid frequency are divided into multiple data blocks according to a fixed time window. Each data block is converted into a three-dimensional tensor, which includes the time step, feature dimension, and number of samples. This three-dimensional tensor is used as input to a convolutional neural network (CNN). The CNN extracts local features from the three-dimensional tensor through multiple convolutional layers. The initial convolutional layer extracts low-level features of grid voltage, grid current, and grid frequency. Low-level features refer to waveform edges, periodic change patterns, and grid frequency fluctuations. Subsequent convolutional layers combine low-level features to generate higher-level grid state features. Each convolutional layer is followed by a max-pooling layer to reduce the feature dimension, and the ReLU activation function is used to enhance the nonlinear expressive power. Finally, the output is a feature vector containing the grid operating state, i.e., the grid state features.

[0039] The generator of the generative adversarial network (GAN) receives the power grid state features output by the convolutional neural network (CNN). It generates a feature distribution simulating the normal operation of the power grid through a fully connected layer. The generator's goal is to make the generated feature distribution approximate the real normal state feature distribution. The discriminator of the GAN receives the power grid state features from the CNN and outputs the power grid state features and the power grid normal operation feature distribution generated by the generator. A multilayer perceptron determines whether the input power grid state features belong to the normal operation state. During the adversarial training phase, the generator and discriminator optimize the generator and discriminator parameters by minimizing the maximum game, so that the generator generates normal pattern features with high fitting degree, and the discriminator can accurately distinguish between real and generated features. During the real-time detection phase, the real-time power grid state feature vector is input into the trained GAN. If the generator cannot generate normal pattern features that match the input power grid state features, or if the discriminator determines that the input power grid state features do not belong to the normal distribution, then it is determined that the current power grid has a potential abnormal oscillation mode.

[0040] Periodic variation patterns refer to the repetitive fluctuation characteristics of grid voltage and grid current in a time series, especially the periodic fluctuations related to the basic grid frequency (50Hz). Periodic variation patterns include, but are not limited to, periodic shifts in grid frequency caused by generator speed fluctuations, and periodic fluctuations in grid voltage amplitude caused by load switching.

[0041] The behavior pattern feature distribution under normal operating conditions refers to the discriminator receiving the input power grid state features through a multilayer perceptron (MLP) and comparing them with the known feature distribution under normal operating conditions of the power grid. If the power grid voltage deviation and power grid frequency deviation in the input power grid state features both fall within the preset normal range, then the power grid state features are considered to be in normal operating mode; otherwise, they are considered abnormal.

[0042] S2. Based on the characteristics of the power grid state, the deep Q-network (DQN) is used to dynamically adjust the charging and discharging strategy of the energy storage device. The machine learning model is used to analyze the battery operation data, predict the battery health status, and select the best charging and discharging strategy by combining the charging and discharging strategy and the battery health status.

[0043] The grid state features obtained from the convolutional neural network (CNN) are used as input to the deep Q-network (DQN). These grid state features include grid voltage, grid current, and grid frequency, reflecting the current state of the grid. Based on the received grid state features, the DQN evaluates different charging and discharging strategies for energy storage devices through its internal mechanism, predicts the long-term return value under each strategy, and selects the strategy that maximizes the expected cumulative reward as the output. Specifically:

[0044] The charging and discharging strategy of the energy storage device is set as a discrete action space, which includes charging, discharging, and standby. The power range for charging and discharging is 0% to 100% of the rated power. A multi-objective reward function is defined, which integrates grid stability and battery health. The multi-objective reward function includes grid stability reward, battery health reward, and economic reward. The grid stability reward is given as follows: if the grid frequency deviation is ±0.2Hz and the grid voltage deviation is ±5%, it is considered to be within the allowable range, and a positive reward is given if the grid frequency deviation and grid voltage deviation exceed the allowable range; if the broadband oscillation amplitude decreases, it indicates successful oscillation suppression, and an additional reward is given. Battery health is determined based on the predicted value of the battery health status. (Based on battery industry experience, the typical cycle life of lithium batteries is 2000~5000 cycles, corresponding to a battery health status prediction rate of 0.1%~0.5% per cycle.) If the charging and discharging strategy causes the battery health status prediction rate of 0.5% / cycle to decrease, a positive reward is given; if the battery health status prediction rate of 0.5% / cycle decreases, a negative reward is given. (Based on the battery end-of-life standard, 20%~80% is set as a safe range.) If the battery health status prediction value remains within the 20%~80% range, an additional reward is given. The economic reward refers to a positive reward given based on the dynamic electricity market price, during off-peak hours for charging and during peak hours for discharging.

[0045] A Deep Q-Network (DQN) architecture is defined, consisting of an input layer, hidden layers, and an output layer. The input layer receives state space features, including grid state characteristics and battery health status. The hidden layers contain multiple fully connected layers, using the ReLU activation function to enhance nonlinear expression. The output layer outputs the Q-values ​​of each action, corresponding to the expected long-term returns of different charge / discharge strategies. A training set is constructed from historical data, using grid state characteristics and the number of charging cycles. An empirical replay mechanism is used to randomly sample the training set to reduce sequence correlation. Mean squared error is selected as the loss function, and the network parameters are updated by minimizing the Q-value prediction error. The parameters of the main network are periodically copied to the target network to stabilize the training process and complete the training. In real-time operation, the DQN selects the action with the highest Q-value as the final output charge / discharge strategy based on the current state input.

[0046] Collect battery operating data, including but not limited to temperature, number of charging cycles, and charge / discharge grid current. Analyze this data using a gated recurrent neural network (RNN) model to predict the battery's health status. The predicted battery health status provides crucial information about the battery's remaining lifespan and current health condition. The specific steps are as follows:

[0047] Battery operating data is normalized, and wavelet transform is used for denoising and missing value imputation. A feature matrix is ​​constructed from the preprocessed battery operating data, which includes the slope of the voltage curve, the rate of change of internal resistance, and the temperature fluctuation amplitude. Based on the feature matrix, the features are compressed by an encoder, and the data is reconstructed by a decoder to remove noise. The denoised battery operating data is input into a gated recurrent unit recurrent neural network model to capture the time series pattern of battery capacity degradation. The gated recurrent unit recurrent neural network model is trained using historical cycle life data (capacity decay curve with the number of cycles) to predict the current battery health status. Hyperparameters (learning rate) are adjusted through cross-validation to ensure that the prediction error is less than 5%. The current battery operating data is input into the gated recurrent unit recurrent neural network model in real time, and the predicted battery health status is output. If the predicted battery health status is lower than the health threshold, a battery health warning is triggered. In this embodiment, the health threshold is set to 80%.

[0048] The formula for calculating the slope of the voltage curve is:

[0049] ;

[0050] in, Indicates time interval, It represents the amount of voltage change. and Representing two points in time, satisfying > , Indicates time voltage, Indicates time The voltage;

[0051] The formula for calculating the rate of change of internal resistance is:

[0052] ;

[0053] in, This represents the change in internal resistance. This represents the internal resistance value at the current moment. This represents the internal resistance value at the previous moment;

[0054] The formula for calculating the temperature fluctuation amplitude is:

[0055] ;

[0056] in, The standard deviation represents the amplitude of temperature fluctuations. Indicates the first Temperature at each point in time, Represents the average temperature. Indicates the number of samples within the time window;

[0057] The charging and discharging strategy derived from the Deep Q-Network (DQN) is combined with the battery health state prediction value obtained from the gated recurrent neural network (RNN) model. The changing trends of battery health state under different charging and discharging strategies are compared to select the optimal charging and discharging strategy that meets grid demands while maximizing battery lifespan. This optimal strategy is then output. The optimal charging and discharging strategy considers not only immediate grid state characteristics and demands but also long-term battery health management, ensuring that battery efficiency and lifespan are optimized while maintaining grid stability. The specific steps for combining the charging and discharging strategy with the battery health state prediction value are as follows:

[0058] An objective function is constructed based on grid stability, predicted battery health status, and economic incentives. Grid stability includes minimizing grid frequency deviation and grid voltage deviation. Predicted battery health status refers to maximizing the predicted battery health status. Constraints are set, which require that the predicted battery health status must be kept within a safe range of 20% to 80%, the charging and discharging power must not exceed the rated capacity of the energy storage device, and the predicted battery health status must be higher than the critical threshold of 70%. When the predicted battery health status drops below 70%, it means that the maximum capacity is significantly reduced, the internal resistance increases, resulting in longer charging time, reduced discharge efficiency, and reduced energy density.

[0059] The remaining charge / discharge strategies are weighted and scored according to the objective function. In this embodiment, grid stability is assigned a weight of 40%, battery health status prediction is assigned a weight of 30%, and economic incentive is assigned a weight of 30%. The calculation formula is as follows:

[0060] ;

[0061] The charging and discharging strategy with the highest total score is selected as the best charging and discharging strategy. If there are multiple charging and discharging strategies with similar scores, the scheme with the higher grid stability score is given priority.

[0062] S3. Based on the optimal charging and discharging strategy and grid state characteristics, the feedforward control combined with feedback control algorithm is applied to quickly respond to grid load changes, adjust the grid voltage and grid current output by the inverter, regulate reactive power to improve the grid power factor, and output the adjusted grid connection parameters.

[0063] Based on the obtained optimal charging and discharging strategy and grid state characteristics, preliminary inverter output parameter settings are determined. Specifically, based on the current grid voltage, grid current, and grid frequency, and combined with the power range recommendation in the optimal charging and discharging strategy (both charging and discharging power ranges from 0% to 100% of rated power), the target values ​​for the inverter output grid voltage and grid current are calculated. The inverter's voltage regulation function sets the target grid voltage value as the inverter's output voltage reference value. The inverter's current output limit is adjusted according to the target grid current value to ensure that it does not exceed the rated current. Based on the target grid voltage and grid current values, feedforward control is then implemented... The control mechanism directly adjusts the inverter's switching signals to quickly reach the target output value, thus rapidly responding to changes in grid load. It primarily relies on a pre-set multi-objective reward function and constraints to rapidly adjust the inverter's output, aiming to reduce latency and improve response speed. After feedforward control adjustment, the actual grid voltage and current outputs of the inverter are collected and compared with the target grid voltage and current values ​​to obtain voltage and current errors. When deviations exist, the inverter's output parameters are fine-tuned through a feedback control algorithm. A PID controller continuously monitors and adjusts the inverter output until the error is minimized. The specific steps for adjusting the inverter output are as follows:

[0064] Based on the current voltage error, a control signal is directly generated using the proportional gain parameter. A higher proportional gain results in a larger adjustment range for the control signal. The accumulated historical voltage error refers to all error values ​​from startup to the current time, and an integral control term is generated using the integral gain parameter. The integral term gradually eliminates long-term steady-state errors. Based on the rate of change of the voltage error (i.e., the trend of error change over time), a derivative control term is generated using the derivative gain parameter. The derivative term predicts the future direction of error change and adjusts the control signal in advance to avoid oscillations. The outputs of the proportional, integral, and derivative control terms are added to form the final PID control signal. This final PID control signal directly acts on the inverter's output parameters (such as switching frequency or duty cycle), dynamically adjusting the inverter's voltage and current outputs to quickly approach the target values. The controller continuously monitors the inverter's actual output voltage and current, compares the actual values ​​with the target values, recalculates the error, and repeatedly adjusts to generate new control signals.

[0065] Target value of grid voltage The calculation formula is:

[0066] ;

[0067] ;

[0068] in, This represents the actual measured voltage value of the current power grid. This indicates the amount of compensation for the grid voltage. This represents the proportionality coefficient, used to adjust the magnitude of the compensation. This indicates the grid voltage setpoint. Grid stability bonus requirements stipulate that voltage deviation must be controlled within ±5%. The grid voltage setpoint is the grid standard voltage.

[0069] Target value of grid current The calculation formula is:

[0070] ;

[0071] in, This represents the power requirement determined by the optimal charging and discharging strategy, derived from the discrete action space including the power values ​​of charging, discharging, and standby (0% to 100% of rated power).

[0072] The power factor of the current grid is calculated based on the requirement that the grid voltage deviation be controlled within ±5%. The closer the power factor is to 1, the higher the proportion of active power and the lower the reactive power demand. The required reactive power is calculated based on the actual measured values ​​of grid voltage and grid current and the current grid power factor. Based on the final inverter output parameters, the inverter output parameters are adjusted again, especially the reactive power part is precisely adjusted. After integrating all adjustment measures, the final grid connection parameters are generated. The grid connection parameters include the finely adjusted inverter output voltage, current and reactive power settings. They work together to ensure the stable operation of the grid and meet the requirements of the optimal charging and discharging strategy.

[0073] The final grid connection parameters are applied to the actual operating environment, and the steps of repeatedly adjusting the inverter output parameters according to new grid state characteristics are prepared to continuously optimize and adapt to the ever-changing grid conditions, so as to maintain efficient grid management and battery health management even in dynamic environments.

[0074] S4 integrates physical model predictive control (MPC) and deep reinforcement learning (DRL), taking grid connection parameters and grid state characteristics as inputs, dynamically adjusting inverter parameters, suppressing oscillations within different grid frequency ranges, and outputting an oscillation suppression strategy.

[0075] The inverter's final grid connection parameters and grid state characteristics are input into the joint control framework of Physical Model Predictive Control (MPC) and Deep Reinforcement Learning (DRL). Based on the grid's power dynamics equations and combined with the current input grid connection parameters and grid state characteristics, the PPC predicts the changing trends of grid voltage, grid current, and grid frequency over multiple future time steps. The prediction process considers oscillation modes in different frequency ranges of the grid, including low-frequency oscillations and wide-frequency oscillations, and generates a sequence of predicted values ​​for the future grid state.

[0076] The deep reinforcement learning (DRL) network receives the battery health state prediction from the physical model predictive control (MPC) and the current grid state characteristics. It then generates an adjustment strategy for the inverter parameters using the trained DRL network. The specific steps are as follows:

[0077] Based on the fluctuation trends of grid voltage, grid current, and grid frequency in the grid state prediction values, the adjustment range of inverter output parameters (such as voltage regulation, current limit, and reactive power compensation) is calculated. Combined with the constraints of grid stability, it is ensured that the adjusted parameters can suppress oscillations in different frequency ranges. Through a multi-objective reward function, the grid stability reward, battery health reward, and economic reward are used to optimize the adjustment strategy, and the scheme with higher grid stability score is selected first.

[0078] The inverter's output parameters are dynamically modified based on the adjustment strategy generated by deep reinforcement learning (DRL). Specific adjustments include: adjusting the inverter's voltage reference value to achieve fast voltage tracking via a PID controller (such as the PID control logic described in S3); adjusting the inverter's current output limit to ensure that power demand (derived from the optimal charging and discharging strategy) matches grid load changes; and adjusting reactive power output by calculating reactive power demand based on actual measurements of grid voltage and current.

[0079] The adjusted inverter output parameters are integrated into an oscillation suppression strategy, which is then output to the inverter control unit for execution. The oscillation suppression strategy targets and suppresses oscillation modes within different grid frequency ranges, ensuring that grid voltage and frequency deviations are within allowable ranges while also meeting the safe range of battery health status. The actual grid voltage, grid current, and grid frequency after inverter execution are re-input into the convolutional neural network (CNN) to generate new grid state features, forming a closed-loop control. The grid state is continuously monitored and the oscillation suppression strategy is dynamically updated to ensure stable operation under dynamic grid conditions.

[0080] This embodiment also provides a computer device applicable to the method of suppressing broadband oscillations of offshore wind power grid connection through grid-type energy storage, including: a memory and a processor; the memory is used to store computer-executable instructions, and the processor is used to execute the computer-executable instructions to realize the method of suppressing broadband oscillations of offshore wind power grid connection through grid-type energy storage as proposed in the above embodiment.

[0081] The computer device can be a terminal, comprising a processor, memory, communication interface, display screen, and input devices connected via a system bus. The processor provides computing and control capabilities. The memory includes non-volatile storage media and internal memory. The non-volatile storage media stores the operating system and computer programs. The internal memory provides an environment for the operation of the operating system and computer programs stored in the non-volatile storage media. The communication interface is used for wired or wireless communication with external terminals; wireless communication can be achieved through Wi-Fi, carrier networks, NFC (Near Field Communication), or other technologies. The display screen can be an LCD screen or an e-ink screen. The input devices can be a touch layer covering the display screen, buttons, a trackball, or a touchpad on the computer device's casing, or an external keyboard, touchpad, or mouse.

[0082] This embodiment also provides a storage medium storing a computer program that, when executed by a processor, implements the method for suppressing broadband oscillations of offshore wind power grid connection through grid-based energy storage as proposed in the above embodiments. The storage medium can be implemented by any type of volatile or non-volatile storage device or a combination thereof, such as Static Random Access Memory (SRAM), Electrically Erasable Programmable Read-Only Memory (EEPROM), Erasable Programmable Read Only Memory (EPROM), Programmable Red-Only Memory (PROM), Read-Only Memory (ROM), magnetic storage, flash memory, magnetic disk, or optical disk.

[0083] In summary, this invention improves the accuracy of grid state modeling by: extracting three-dimensional tensor features from grid voltage, current, and frequency data using a Convolutional Neural Network (CNN); simulating normal operation modes and identifying abnormal oscillations using Generative Adversarial Networks (GANs) to achieve real-time detection of abnormal modes; dynamically optimizing charging and discharging strategies using a Deep Q-Network (DQN) combined with a multi-objective reward function to establish a quantitative balance between grid stability, battery health, and economic efficiency, extending battery life and reducing operation and maintenance costs; predicting battery health status using a Gated Recursive Unit (GRU) recurrent neural network; constructing a degradation trend model based on historical lifespan data to ensure the safety boundary of the energy storage system; and integrating Physical Model Predictive Control (MPC) and Deep Reinforcement Learning (DRL) to construct a hybrid decision-making system for cross-frequency band oscillation suppression, enabling inverter parameter adjustments to have both immediate response and global convergence characteristics. Ultimately, this achieves hierarchical suppression of full-frequency domain oscillations, significantly improving the stability margin of the new power system.

[0084] It should be noted that the above embodiments are only used to illustrate the technical solutions of the present invention and are not intended to limit it. Although the present invention has been described in detail with reference to preferred embodiments, those skilled in the art should understand that modifications or equivalent substitutions can be made to the technical solutions of the present invention without departing from the spirit and scope of the technical solutions of the present invention, and all such modifications or substitutions should be covered within the scope of the claims of the present invention.

Claims

1. A method for suppressing broadband oscillations during offshore wind power grid connection using grid-connected energy storage, characterized in that: include: Convolutional neural networks (CNNs) are used to extract features from the collected grid voltage, grid current, and grid frequency data to identify grid state characteristics. Based on these grid state characteristics, generative adversarial networks (GANs) are used to simulate the behavior patterns of the grid under normal operating conditions and identify potential abnormal oscillation patterns. Based on the characteristics of the power grid state, a deep Q-network (DQN) is used to dynamically adjust the charging and discharging strategy of the energy storage device. A machine learning model is used to analyze the battery operation data and predict the battery health status to obtain the battery health status prediction value. The optimal charging and discharging strategy is selected by combining the charging and discharging strategy and the battery health status prediction value. Based on the optimal charging and discharging strategy and grid state characteristics, feedforward control combined with feedback control algorithm is applied to quickly respond to grid load changes, adjust the grid voltage and grid current output by the inverter, regulate reactive power to improve the grid power factor, and output the adjusted grid connection parameters. The integrated physical model predictive control (MPC) and deep reinforcement learning (DRL) take grid connection parameters and grid state characteristics as inputs, dynamically adjust inverter parameters, suppress oscillations in different grid frequency ranges, and output an oscillation suppression strategy.

2. The method for suppressing broadband oscillations of offshore wind power grid connection through grid-type energy storage as described in claim 1, characterized in that: The specific steps for identifying the power grid state characteristics are as follows: The collected grid voltage, grid current, and grid frequency are divided into multiple data blocks according to a fixed time window. Each data block is converted into a three-dimensional tensor and used as the input of a convolutional neural network. Local features are extracted from the three-dimensional tensor through multiple convolutional layers. The initial convolutional layer extracts low-level features, and subsequent convolutional layers stack and combine low-level features to generate higher-level grid state features. Each convolutional layer is followed by a max pooling layer to reduce the feature dimension, and the ReLU activation function is used to enhance the nonlinear expressive power. Finally, a feature vector containing the grid operating state is output.

3. The method for suppressing broadband oscillations of offshore wind power grid connection through grid-type energy storage as described in claim 2, characterized in that: The specific steps for identifying potential abnormal oscillation patterns are as follows: The generator of the generative adversarial network receives the power grid state features output by the convolutional neural network and generates a feature distribution simulating the normal operation of the power grid through a fully connected layer. The discriminator receives the power grid state features and the feature distribution generated by the generator and determines whether the input power grid state features belong to the normal operation state through a multilayer perceptron. During the adversarial training phase, the parameters of the generator and discriminator are optimized so that the generator generates normal pattern features with high fitting degree. During the real-time detection phase, if the generator cannot generate matching normal pattern features or the discriminator determines an anomaly, it is determined that there is a potential abnormal oscillation mode in the current power grid.

4. The method for suppressing broadband oscillations of offshore wind power grid connection through grid-type energy storage as described in claim 3, characterized in that: The specific steps for dynamically adjusting the charging and discharging strategy of energy storage devices using a deep Q-network (DQN) based on grid state characteristics are as follows: The charging and discharging strategy of the energy storage device is set as a discrete action space. A multi-objective reward function is defined to integrate grid stability reward, battery health reward and economic reward. The sequence correlation is reduced through experience playback mechanism. The main network parameters are periodically copied to the target network for stable training. The charging and discharging strategy with the highest Q value is selected as the output in real time operation.

5. The method for suppressing broadband oscillations of offshore wind power grid connection through grid-type energy storage as described in claim 4, characterized in that: The steps for using a machine learning model to analyze battery operating data and predict battery health status to obtain a predicted battery health status value are as follows: Battery operation data is collected and analyzed using a gated recurrent neural network model. A feature matrix is ​​constructed that includes the slope of the voltage curve, the rate of change of internal resistance, and the amplitude of temperature fluctuation. Wavelet transform is used to denoise and reconstruct the data. The model is trained using historical cycle life data to predict the battery health status value. If the battery health status value is lower than the health threshold, an early warning is triggered.

6. The method for suppressing broadband oscillations of offshore wind power grid connection through grid-type energy storage as described in claim 5, characterized in that: The specific steps for selecting the optimal charging and discharging strategy by combining the charging and discharging strategy and the battery health status prediction value are as follows: An objective function is constructed based on grid stability, battery health status prediction, and economic incentives. Constraints are set, and the charging and discharging strategy with the highest total score is selected as the optimal solution based on the weighted score of the objective function.

7. The method for suppressing broadband oscillations of offshore wind power grid connection through grid-type energy storage as described in claim 6, characterized in that: The specific steps for adjusting the grid voltage and grid current output by the inverter to regulate reactive power and improve the grid power factor are as follows: By continuously monitoring the actual output voltage and current through a PID controller, calculating the error and generating control signals, the inverter switching frequency or duty cycle is dynamically adjusted to make the output parameters approach the target value, while adjusting reactive power to improve the power factor of the power grid.

8. The method for suppressing broadband oscillations of offshore wind power grid connection through grid-type energy storage as described in claim 7, characterized in that: The integrated physical model predictive control (MPC) and deep reinforcement learning (DRL) dynamically adjust inverter parameters to suppress oscillations within different grid frequency ranges. The specific steps are as follows: The physical model predicts the trend of the power grid state over multiple time steps based on the power grid dynamic equations. The deep reinforcement learning network generates inverter parameter adjustment strategies based on the prediction results. The strategy is optimized through a multi-objective reward function to dynamically modify the inverter output parameters to suppress oscillations in different frequency ranges.

9. A computer device comprising a memory and a processor, wherein the memory stores a computer program, characterized in that: When the processor executes the computer program, it implements the steps of the method for suppressing broadband oscillations of offshore wind power grid connection by means of grid-type energy storage as described in any one of claims 1 to 8.

10. A computer-readable storage medium having a computer program stored thereon, characterized in that: When the computer program is executed by the processor, it implements the steps of the method for suppressing broadband oscillations of offshore wind power grid connection by means of grid-type energy storage as described in any one of claims 1 to 8.

Images