A hybrid energy storage collaborative control method of adaptive variable frequency TCN-SAC

Abstract

This invention discloses a hybrid energy storage collaborative control method based on adaptive variable-step-frequency TCN-SAC, belonging to the field of intelligent energy management and predictive control technology. The method first classifies scenarios into three categories—normal, abnormal, and fault—based on grid fault characteristics and heterogeneous energy storage properties, and constructs a hybrid energy storage system model. Relevant time-series variables are collected, and the model is trained using an adaptive variable-step-frequency TCN network to output supply and demand variable prediction information. Grid fault state data, historical system operation data, and prediction information are integrated to form the SAC algorithm input vector. A multi-objective reward function covering system revenue, energy storage lifetime, and reliability is constructed. Based on this function, a collaborative control strategy is iteratively output, repeatedly executed until the reward function is maximized. Several types of energy storage systems respectively undertake peak shaving, frequency regulation, voltage regulation, and black start functions, achieving collaborative operation under differentiated scenarios.

Description

Technical Field

[0001] This invention belongs to the field of intelligent energy management and power system control technology, specifically relating to the field of collaborative control and time-series prediction of heterogeneous hybrid energy storage systems in power grid operation scenarios. Background Technology

[0002] With the rapid development of smart grid technology and the widespread integration of distributed energy resources, the power grid operating environment is becoming increasingly complex, and the probability of various differentiated fault scenarios is constantly increasing, posing challenges to the safe and stable operation of the power grid. Hybrid energy storage systems, with their complementary technical characteristics, have played an important role in mitigating the impact of power grid faults and maintaining power quality, and have become an important supporting technology in the field of smart energy management.

[0003] However, current hybrid energy storage collaborative control technologies applied to differentiated grid fault scenarios still have many shortcomings. In terms of algorithm application, the TCN algorithm does not fully consider the continuous dependency characteristics of time-series data when performing single-step prediction, resulting in a large number of runs and long processing time. Furthermore, iterative error accumulation is prone to occur during multi-step prediction, leading to a decrease in long-term prediction accuracy. The SAC algorithm, on the other hand, has the limitation of not being able to predict future states, affecting the foresight and rationality of the control strategy.

[0004] In terms of control strategy design, existing technologies fail to fully leverage the characteristics of heterogeneous energy storage, making it difficult for different types of energy storage systems to fully utilize their capabilities under various fault scenarios and achieve coordinated operation with each system fulfilling its specific function. Furthermore, existing control technologies exhibit poor fault adaptability, struggling to provide precisely tailored control solutions for different types and degrees of grid faults. They also lack multi-objective optimization and decision-making capabilities, failing to effectively balance multiple key objectives such as fault costs, park revenue, and battery cycle life, thus failing to meet the actual needs of safe, economical, and efficient operation of smart grids. Summary of the Invention

[0005] The purpose of this invention is to provide a hybrid energy storage collaborative control method for adaptive variable step frequency TCN-SAC, which solves the problems of insufficient adaptability of existing technologies to different grid fault scenarios, as well as the poor accuracy of algorithm prediction and the unsatisfactory hybrid energy storage collaborative control effect.

[0006] To achieve the above objectives, the present invention provides the following technical solution:

[0007] A hybrid energy storage cooperative control method for adaptive variable step frequency TCN-SAC includes the following steps:

[0008] Based on the characteristics of power grid faults and the properties of heterogeneous energy storage, a power grid fault classification scheme is proposed, and a heterogeneous hybrid energy storage system model for dealing with differentiated power grid fault scenarios is constructed.

[0009] Collect various time-series variables that affect the system's operating state, extract historical time-series data from these variables, and use the historical time-series data to model and train an adaptive variable step frequency TCN network to obtain supply and demand variable prediction information for future moments.

[0010] According to the power grid fault classification scheme, the monitored power grid operating status is identified, and the fault status data corresponding to the current power grid is determined.

[0011] Historical system operation data is extracted from the various time-series variables, and the fault status data, historical system operation data and supply and demand variable prediction information are fused to form the input state vector of the SAC algorithm.

[0012] Taking into account the objectives related to failure cost, park revenue, and battery cycle life cost, a reward function for the SAC algorithm is constructed.

[0013] The reward function is used to iterate the input state vector to obtain a cooperative control strategy for the heterogeneous hybrid energy storage system.

[0014] Repeat the following steps: power grid operation state identification, input state vector construction, reward function application, and strategy iteration, until the value of the reward function is maximized, and then output the final cooperative control strategy.

[0015] In one possible implementation, the construction of the power grid fault classification scheme includes the following steps:

[0016] Based on the real-time frequency and voltage of the power grid, the system operation scenarios are classified into normal scenarios, abnormal scenarios, and fault scenarios.

[0017] Based on the preset ranges corresponding to the deviation value of the real-time frequency of the power grid and the actual value of the voltage, the scenario type is determined, and the scenario in which the system is located is identified.

[0018] In one possible implementation, the modeling and training of the adaptive variable step frequency TCN network includes the following steps:

[0019] The multi-step prediction results and single-step prediction results of the adaptive variable step frequency TCN network are calculated respectively to obtain the corresponding two mean absolute errors.

[0020] Calculate the error ratio based on the two mean absolute errors;

[0021] The step size of the adaptive variable step frequency TCN network is adjusted according to the magnitude of the error ratio.

[0022] The adaptive variable step frequency TCN network with adjusted step size is used to model and train the historical time series data, and output the supply and demand variable prediction information.

[0023] In one possible implementation, the construction of the input state vector of the SAC algorithm includes the following steps:

[0024] Collect grid voltage and grid frequency data to supplement the fault status data;

[0025] Collect renewable energy generation data and grid load data to supplement the system's historical operating data;

[0026] By integrating the supplemented fault state data, the supplemented historical system operation data, and the supply and demand variable prediction information, the input state vector of the SAC algorithm is obtained.

[0027] In one possible implementation, the reward function construction of the SAC algorithm includes the following steps:

[0028] Based on parameters related to electricity sales, electricity purchase, peak shaving services, frequency regulation services, and voltage regulation services, the objective function for system revenue is calculated.

[0029] Based on the power parameters and state of charge parameters of each energy storage system, the objective function for the lifespan of the energy storage system is calculated.

[0030] Based on the discharge energy parameters, real-time load parameters, and load deficit parameters of the energy storage system, the reliability objective function of the energy storage system is calculated.

[0031] The reward function of the SAC algorithm is obtained by weighted summing of the system revenue objective function, the energy storage system lifetime objective function, and the energy storage system reliability objective function.

[0032] In one possible implementation, the output of the cooperative control strategy includes the following steps:

[0033] Identify the energy storage systems in the heterogeneous hybrid energy storage system that correspond to peak shaving, frequency regulation, voltage regulation, and black start functions, respectively;

[0034] By utilizing the results of the strategy iteration, the charging and discharging actions and parameters of the energy storage system are coordinated and controlled to obtain an energy storage system operation scheme adapted to different fault scenarios.

[0035] In one possible implementation, policy iteration based on the reward function includes the following steps:

[0036] The SAC algorithm generates a parameterized policy through its policy network, and the actions are sampled and mapped to a preset range.

[0037] The value function of the state-action pair is evaluated using the two Q-value networks of the SAC algorithm.

[0038] The SAC algorithm provides a stable TD target through two target Q networks, and the parameters of the target Q networks are adjusted using a soft update method.

[0039] Automatic entropy adjustment is performed using the entropy adjustment coefficient and optimizer of the SAC algorithm.

[0040] Based on the reward function, calculate the target Q value, the Q network loss function, and the policy network loss function;

[0041] Based on the loss function, the parameters of the Q-value network and the policy network are updated respectively, completing one policy iteration.

[0042] In one possible implementation, the energy storage system operation scheme adapted to different fault scenarios includes scenario switching and strategy connection, which includes the following steps: collecting grid frequency data and grid voltage data and supplementing them to the various time-series variables;

[0043] The scenario type is determined based on the supplemented time-series variables according to a preset time interval;

[0044] Based on the determination result, the collaborative control strategy is switched when the corresponding scenario conditions are met.

[0045] When both the grid voltage and grid frequency are abnormal or faulty, first adjust the voltage to the normal range, and then adjust the frequency.

[0046] When switching scenarios, the power change rate of the energy storage system is controlled to not exceed a preset threshold.

[0047] After the abnormal or faulty scenario ends, the energy storage system is controlled to return to the target state of charge within a preset time.

[0048] In one possible implementation, the step size adjustment of the adaptive variable step frequency TCN network includes the following steps:

[0049] Set the initial step size of the adaptive variable step frequency TCN network;

[0050] The error ratio is obtained by calculating the ratio of the mean absolute error of the multi-step prediction results to the mean absolute error of the single-step prediction results.

[0051] When the error ratio is greater than a preset value, the step size of the adaptive variable step frequency TCN network is reduced.

[0052] When the error ratio is equal to a preset value, the step size of the adaptive variable step frequency TCN network remains unchanged;

[0053] When the error ratio is less than a preset value, the step size of the adaptive variable step frequency TCN network is increased;

[0054] The adjusted step size is subject to upper and lower limit constraints to ensure that the step size is within a preset range.

[0055] In one possible implementation, when performing a weighted summation of the system revenue objective function, the energy storage system lifetime objective function, and the energy storage system reliability objective function, the weight coefficients of each objective function are adjusted according to different failure scenarios.

[0056] Compared with existing technologies, the advantages of this invention are as follows: the adaptive variable step frequency TCN network solves the problems of traditional TCN single-step prediction neglecting the time-series data dependency characteristics and multi-step prediction iteration error accumulation by automatically adjusting the step size, thus achieving a balance between prediction accuracy and the rationality of time-series evolution. Compared with traditional prediction methods, this network has higher operating efficiency and the prediction results are more consistent with the actual power grid operation.

[0057] The SAC algorithm's input state vector integrates grid fault state data, historical system operation data, and predictive information from the TCN network. Combined with a dual-Q network architecture, target Q network soft updates, and an automatic entropy adjustment mechanism, the algorithm can comprehensively perceive the grid's operational status. Compared to the limitations of traditional SAC algorithms, which cannot predict future states, this algorithm improves training stability and the adaptability of policy exploration. The generated control strategies are more forward-looking and reasonable, better adapting to dynamic changes in grid operation.

[0058] The power grid fault classification scheme subdivides system states into different scenarios and sub-scenarios, and uses quantitative judgment criteria based on frequency and voltage parameters to make fault identification more objective and operable. Differentiated control strategies are formulated for different scenarios, allowing the four types of energy storage systems in the heterogeneous hybrid energy storage system to perform their respective functions and operate collaboratively. Compared with the shortcomings of existing technologies that fail to fully utilize energy storage characteristics, the effectiveness of various energy storage systems is fully realized, resulting in superior performance in mitigating the impact of power grid faults and maintaining power quality.

[0059] The multi-objective reward function comprehensively considers system revenue, energy storage system lifespan, and reliability. By using weighted summation to adapt to the optimization needs of different scenarios, it achieves multi-objective collaborative optimization compared to the limitations of existing technologies that focus on single-objective or a few-objective optimization. This effectively reduces system operating costs, increases park revenue, extends the cycle life of energy storage systems, improves renewable energy utilization and load matching, and ensures the safe, economical, and efficient operation of the smart grid.

[0060] The scenario switching and strategy integration mechanism avoids grid impact during scenario switching by determining the scenario in real time, prioritizing voltage adjustment, constraining the power change rate, and restoring the target state of charge. This ensures a smooth transition of control strategies under different scenarios and the continuous response capability of the energy storage system, further improving the operational stability and reliability of the entire system. Attached Figure Description

[0061] To more clearly illustrate the technical solutions in the embodiments of the present invention, the drawings used in the embodiments will be briefly introduced below. Obviously, the drawings described below are only some embodiments of this application. For those skilled in the art, other drawings can be obtained based on these drawings without creative effort.

[0062] Figure 1 This is a schematic diagram of the adaptive variable step frequency TCN-SAC hybrid energy storage collaborative control method according to an embodiment of the present invention;

[0063] Figure 2 This is a schematic diagram of the overall architecture and scenario-adaptive control of hybrid energy storage collaborative control in an embodiment of the present invention. Detailed Implementation

[0064] The technical solutions of the embodiments of the present invention will be clearly and completely described below with reference to the accompanying drawings. Obviously, the described embodiments are only some embodiments of this application, and not all embodiments. Based on the embodiments of this application, all other embodiments obtained by those of ordinary skill in the art without creative effort are within the scope of protection of this application.

[0065] Example:

[0066] It should be noted that the terms "comprising" and "having" and any variations thereof in the embodiments of the present invention are intended to cover non-exclusive inclusion. For example, a process, method, system, product, or device that includes a series of steps or units is not necessarily limited to those steps or units that are explicitly listed, but may include other steps or units that are not explicitly listed or that are inherent to such processes, methods, products, or devices.

[0067] See Figure 1 An embodiment of the present invention provides a hybrid energy storage cooperative control method for an adaptive variable step frequency TCN-SAC, comprising the following steps:

[0068] Step 101: Based on the characteristics of power grid faults and the properties of heterogeneous energy storage, propose a power grid fault classification scheme and construct a heterogeneous hybrid energy storage system model to cope with differentiated power grid fault scenarios.

[0069] Specifically, the heterogeneous hybrid energy storage system model can be a hybrid system that includes pumped hydro storage, supercapacitors, flywheel energy storage, and lithium-ion batteries.

[0070] The construction of the power grid fault classification scheme includes the following steps: classifying the system operation scenario types according to the real-time frequency and voltage of the power grid to obtain normal scenario, abnormal scenario and fault scenario; determining the scenario type according to the preset range corresponding to the deviation value of the real-time frequency and the actual value of the voltage, and determining the scenario in which the system is located.

[0071] Specifically, the real-time frequency of the power grid can be the instantaneous frequency during power grid operation, in Hz; the power grid voltage can be the real-time voltage at the grid connection point, in kV; the normal scenario can be a scenario where both the frequency and voltage are within the rated range; the abnormal scenario can be a scenario where the frequency or voltage exceeds the rated range but does not reach the fault threshold; the fault scenario can be a scenario where the frequency or voltage deviates significantly from the rated range; the frequency deviation value can be the difference between the real-time frequency and the rated frequency.

[0072] For example, the power grid fault classification scheme divides the system state into normal scenarios, abnormal scenarios, and fault scenarios, with the scenario determination based on the real-time frequency of the power grid. With voltage The specific judgment criteria are as follows:

[0073] ① When the frequency and voltage are within the normal range ( At that time, the system was in a normal operating environment;

[0074] ②When At that time, the system was in an abnormal scenario;

[0075] ③When At that time, the system was in a fault scenario.

[0076] Step 102: Collect various time-series variables that affect the system's operating state, extract historical time-series data from these variables, and use the historical time-series data to model and train the adaptive variable step frequency TCN network to obtain supply and demand variable prediction information for future moments.

[0077] Specifically, various time-series variables can be photovoltaic power generation, load power, ambient temperature, grid frequency, and grid voltage; historical time-series data can be hourly operation data from the past year; the adaptive variable step frequency TCN network can be a time-series convolutional network with dynamically adjustable step size; and supply and demand variable prediction information can be the load forecast and photovoltaic output forecast for the next 24 hours.

[0078] The modeling and training of the adaptive variable step frequency TCN network includes the following steps:

[0079] The multi-step prediction results and single-step prediction results of the adaptive variable step frequency TCN network are calculated respectively to obtain two corresponding mean absolute errors; based on the two mean absolute errors, the error ratio is calculated; according to the magnitude of the error ratio, the step size of the adaptive variable step frequency TCN network is adjusted; using the adaptive variable step frequency TCN network with the adjusted step size, the historical time series data is modeled and trained, and the supply and demand variable prediction information is output.

[0080] Specifically, the multi-step prediction result can be a single prediction of supply and demand data for the next 12 hours; the single-step prediction result can be an hourly rolling prediction of supply and demand data for the next 12 hours; the mean absolute error can be the average absolute deviation between the predicted value and the actual value; the error ratio can be the ratio of the mean absolute error of the multi-step prediction to the mean absolute error of the single-step prediction; the step size can be the sliding step size of the convolution kernel of the TCN network, with the initial step size set to 12; and the supply and demand variable prediction information can be the hourly load prediction values ​​for the next 24 hours.

[0081] Furthermore, the step size adjustment of the adaptive variable step frequency TCN network includes the following steps:

[0082] Set the initial step size of the adaptive variable step frequency TCN network; calculate the ratio of the mean absolute error of the multi-step prediction results to the mean absolute error of the single-step prediction results to obtain the error ratio; when the error ratio is greater than a preset value, decrease the step size of the adaptive variable step frequency TCN network; when the error ratio is equal to the preset value, keep the step size of the adaptive variable step frequency TCN network unchanged; when the error ratio is less than the preset value, increase the step size of the adaptive variable step frequency TCN network; impose upper and lower limit constraints on the adjusted step size to ensure that the step size is within a preset range.

[0083] Specifically, the initial step size can be 12, corresponding to a 12-hour prediction step size; the mean absolute error can be the arithmetic mean of the absolute deviations between the predicted value and the actual value; the error ratio can be the ratio of the mean absolute error of multi-step prediction to the mean absolute error of single-step prediction; the preset value can be 1; the preset range can be 1-24, corresponding to a prediction step size of 1 hour to 24 hours; decreasing the step size can be done by decreasing by 1 each time, and increasing the step size can be done by increasing by 1 each time.

[0084] For example, let the multi-step prediction result at time step t be... The single-step prediction result is .

[0085] initial step size .

[0086] During model execution, the mean absolute error (MAE) of multi-step predictions is calculated. )) and the mean absolute error (MAE) of single-step prediction To eliminate the error differences caused by dimensionality, an error ratio is set. .when When, appropriately reduce the step size; when When, the step size remains unchanged; when At this time, the step size should be increased appropriately.

[0087] The step size update formula is as follows ;

[0088] in, The step size is the step size at the previous time step. The step size update formula includes the error ratio and the step size at the previous time step, and the step size at this time step is calculated. This adaptive variable step frequency strategy can achieve a balance between prediction accuracy and the rationality of temporal evolution, thereby significantly improving the overall prediction performance and applicability of the model.

[0089] Step 103: Based on the power grid fault classification scheme, identify the monitored power grid operating status and determine the fault status data corresponding to the current power grid.

[0090] Specifically, fault status data can be grid voltage anomaly indicators or frequency deviation data.

[0091] Step 104: Extract historical system operation data from the various time series variables, and fuse the fault status data, the historical system operation data, and the supply and demand variable prediction information to form the input state vector of the SAC algorithm.

[0092] Specifically, the SAC algorithm is a soft actor-critic reinforcement learning algorithm.

[0093] The input state vector construction of the SAC algorithm includes the following steps: collecting grid voltage data and grid frequency data to supplement the fault state data; collecting renewable energy generation data and grid load data to supplement the system historical operation data; and fusing the supplemented fault state data, the supplemented system historical operation data, and the supply and demand variable prediction information to obtain the input state vector of the SAC algorithm.

[0094] Specifically, grid voltage data can be hourly voltage values ​​at the grid connection point; grid frequency data can be hourly frequency values; renewable energy generation data can be photovoltaic power generation and wind power generation; grid load data can be hourly power consumption of industrial load and residential load; and the input state vector can be a high-dimensional vector with a dimension of 20, containing feature values ​​of various types of data.

[0095] Step 105: Taking into account the relevant objectives of fault cost, park revenue, and battery cycle life cost, construct the reward function of the SAC algorithm.

[0096] Specifically, the reward function can be a weighted summation function that integrates multiple objectives.

[0097] The reward function construction of the SAC algorithm includes the following steps: calculating the system revenue objective function based on electricity sales parameters, electricity purchase parameters, peak shaving service parameters, frequency regulation service parameters, and voltage regulation service parameters; calculating the energy storage system lifetime objective function based on the power parameters and state of charge parameters of each energy storage system; calculating the energy storage system reliability objective function based on the energy storage system discharge energy parameters, real-time load parameters, and load deficit parameters; and performing a weighted summation of the system revenue objective function, the energy storage system lifetime objective function, and the energy storage system reliability objective function to obtain the reward function of the SAC algorithm.

[0098] Specifically, the parameters related to electricity sales can be electricity price and electricity volume; the parameters related to peak shaving services can be peak shaving active power and peak shaving service unit price; the power parameters can be the charging and discharging power of the energy storage system; the parameters related to the state of charge can be the actual state of charge and the target state of charge; and the discharge energy parameters can be the energy released by the energy storage system under fault scenarios.

[0099] For example, heterogeneous hybrid energy storage systems such as Figure 2 As shown, it includes energy storage systems 1 to 4, wherein,

[0100] ① Based on parameters related to electricity sales, electricity purchase, peak shaving services, frequency regulation services, and voltage regulation services, the system revenue objective function is calculated as follows:

[0101] ;

[0102] in, The objective function is the system's revenue. The price at which the industrial park sells electricity to the power grid; The price at which the industrial park purchases electricity from the power grid; Let i be the electricity sold by the i-th energy storage system; The amount of electricity purchased for the i-th energy storage system; This refers to the peak-shaving active power of energy storage system 1; Price for peak shaving services, unit: yuan / kWh; This is the auxiliary peak-shaving active power for energy storage system 2; The active power of energy storage system 2 participating in frequency regulation; Price per kWh for frequency regulation service; The reactive power of energy storage system 3 participating in voltage regulation; Price per kVarh is the reactive power service unit price.

[0103] The system revenue objective function includes the electricity sales price, electricity purchase price, electricity sales volume, electricity purchase volume, peak-shaving active power of energy storage system 1, peak-shaving service unit price, auxiliary peak-shaving active power of energy storage system 2, active power of energy storage system 2 participating in frequency regulation, frequency regulation service unit price, reactive power of energy storage system 3 participating in voltage regulation, and reactive power service unit price. The system revenue objective function is obtained through calculation, and the system revenue is evaluated.

[0104] ②Based on the power parameters and state-of-charge parameters of each energy storage system, the objective function for the lifespan of the energy storage system is calculated:

[0105] ;

[0106] in, The target state of charge for the i-th energy storage system; ;

[0107] The objective function for the lifespan of the energy storage system; Power of each energy storage system; , , The values ​​are 0.1, 0.5, and 1, which are the correlation coefficients for the lifespan of the energy storage system.

[0108] The lifespan objective function of an energy storage system includes the lifespan correlation coefficient of the energy storage system, the power of each energy storage system, and the target state of charge. The lifespan objective function of the energy storage system is obtained by calculation, and the lifespan of the system is evaluated.

[0109] ③ Based on the discharge energy parameters, real-time load parameters, and load deficit parameters of the energy storage system, the reliability objective function of the energy storage system is calculated:

[0110] ;

[0111] in, Let be the electrical energy released by the i-th energy storage system during discharge. For real-time load; For real-time load deficit; The objective function for the reliability of the energy storage system is denoted as .

[0112] The reliability objective function of an energy storage system includes the released electrical energy, real-time load, and real-time load deficit. The reliability objective function of the energy storage system is calculated to evaluate the reliability of the energy storage system.

[0113] ④ The reward function of the SAC algorithm is obtained by weighted summing of the system revenue objective function, the energy storage system lifetime objective function, and the energy storage system reliability objective function. :

[0114] .

[0115] Furthermore, when performing a weighted summation of the system revenue objective function, the energy storage system lifetime objective function, and the energy storage system reliability objective function, the weight coefficients of each objective function are adjusted according to different failure scenarios.

[0116] Specifically, the weighting coefficient can be a numerical value used to assign importance to each objective function; different fault scenarios can be normal scenarios, abnormal scenarios, and fault scenarios.

[0117] For example, , , The weights for the three objective functions mentioned above are adjusted according to different scenarios, as shown in the table below.

[0118]

[0119] Step 106: Use the reward function to iterate the input state vector to obtain a cooperative control strategy for the heterogeneous hybrid energy storage system.

[0120] Specifically, the coordinated control strategy can be the charging and discharging power commands of each energy storage system.

[0121] The policy iteration based on the reward function includes the following steps: generating a parameterized policy through the policy network of the SAC algorithm, sampling actions and mapping them to a preset interval; evaluating the value function of the state-action pair using the two Q-value networks of the SAC algorithm; providing a stable TD objective through the two target Q-networks of the SAC algorithm, and adjusting the parameters of the target Q-network using a soft update method; performing automatic entropy adjustment using the entropy adjustment coefficient and its optimizer of the SAC algorithm; calculating the target Q-value, the Q-network loss function, and the policy network loss function based on the reward function; and updating the parameters of the Q-value network and the policy network according to the loss functions to complete one policy iteration.

[0122] Specifically, the policy network can be a neural network with two fully connected layers; the preset interval can be [-1,1]; the Q-value network can be a three-layer fully connected network that evaluates the state-action value; the soft update method can be parameter-weighted update; the entropy adjustment coefficient can be a parameter that controls the degree of policy exploration; and the loss function can be the mean squared error loss function.

[0123] For example, the SAC algorithm includes several main components: a policy network for generating parameterized policies, two Q-value networks for evaluating the value functions of state-action pairs, two objective Q-networks for providing stable TD objectives, entropy adjustment coefficients and their optimizers for achieving automatic entropy adjustment, and a model saving and loading mechanism for facilitating experimental reproduction and breakpoint retraining.

[0124] Entropy adjustment coefficient This coefficient, used to balance strategy diversity and reward, controls the degree of randomness in the strategy, thereby adjusting the exploration level. To improve generalization ability and algorithm adaptability, this algorithm employs an automatic adjustment mechanism to learn ln(α), with the optimization objective being... ,in This means sampling an action from policy π based on the state. It is the log probability of the action in the state. The preset target entropy is used to control the randomness of the strategy. This adjusts the algorithm's exploration level; when the strategy entropy falls below the target entropy, Increase it to encourage more exploration; conversely, if the policy entropy is too high, then... Reduce to increase the certainty of the strategy.

[0125] The policy network is responsible for outputting a differentiable random action sampling process from a given state. Its specific implementation is as follows: The network first receives the current state vector `state` as input. The state input is passed through two fully connected neural network layers, each equipped with a ReLU non-linear activation function to extract state features. Subsequently, the network outputs the mean μ and standard deviation σ of the action distribution. σ is processed using the Softplus function to ensure it is positive. Then, Gaussian distribution is used for reparameterization sampling to obtain a random sample with a gradient. The tanh transform is applied to this sampled value to map the action value to the interval [-1, 1], and then it is scaled according to the environmental action boundary. The logarithmic probability ln(...) is then calculated. When considering the irreversibility and compression effect of the tanh transform, we introduce... The correction term ensures correct gradient propagation. Finally, two parallel linear mapping branches output the mean μ and log-standard deviation ln(σ) of the action distribution, respectively. The optimization objective of the policy network is to maximize the soft Q-value under the current policy, which is to minimize the loss function. =E[α*ln( [)-Q(s,a)], thereby achieving a trade-off between expected return and policy entropy.

[0126] Two independent Q-value networks, Q1 and Q2, have essentially the same structure, both consisting of three fully connected layers. The input to each Q-value network is a concatenation of the state vector and the current action, and the output is the corresponding state-action value estimate, Q(st,at). Each Q-network corresponds to a target network, whose structure is identical to the main network, but whose parameters are gradually approximated to the main network parameters through soft updates. The target network does not participate in backpropagation; it is only used to calculate the target value. The parameters of the main network are denoted as... It updates itself continuously through interaction with the environment and backpropagation. The parameters of the target network are adjusted using a soft update method, and these parameters are denoted as... The updated formula is as follows ,in For soft update rate, a smaller value is generally taken. In this invention... The value is set to 0.005 to ensure the stability of the update process. The dual-Q network structure can alleviate the overestimation problem. When calculating the target Q value, the smaller of the two Q network predictions is taken, thus obtaining a more conservative estimate.

[0127] In each training iteration, the algorithm samples a batch of state transition samples from the experience replay pool and executes them step by step according to the following process. First, the target Q-value is calculated, using the current policy network... Upsampling action The action is input into two target Q-networks, and the results are calculated. and The formula for calculating the target Q value is: .in γ represents the immediate reward under the current state and action, and γ is a discount factor used to control the weight of future rewards, typically in the range [0,1]. This is a termination flag, indicating whether the current state is a termination state. If it is a termination state, then... =1, otherwise =0, The smaller of the two Q-values ​​output by the target Q-network is used to represent an estimate of future returns. This is the entropy adjustment coefficient, used to adjust the exploratory nature of the strategy. Actions output by the policy network In state The logarithmic probability below.

[0128] After calculating the target Q-value, the Q-network is updated, and the loss function of both Q-networks is calculated. The formula for the loss function is as follows: After calculating the loss function, the gradient descent algorithm is used to minimize the loss function, and then the Q network parameters are updated.

[0129] Then, the policy network is updated for each sampled state. Actions are generated through a policy network. And its logarithmic probability, the formula for calculating the policy loss is as follows: The policy network parameters are updated with the goal of minimizing this loss.

[0130] Next, the entropy adjustment coefficient is performed. Update, its update formula is: By introducing the entropy term into the loss, policy exploration is encouraged.

[0131] Finally, a soft update is performed on the target network, updating the parameters of the two target Q-networks. The entire training process continues iteratively until the policy stabilizes and converges or the preset number of training epochs is reached.

[0132] The SAC algorithm employs a comprehensive save and load mechanism, facilitating model reproduction, breakpoint recovery, and performance comparison under different experimental conditions. The saved content includes the policy network and its optimizer state, the two Q-networks and their optimizer states, and the entropy adjustment parameters and their optimizer states. When loading the model, the same network structure should be used to ensure consistency with the training phase. After loading, to avoid performance degradation due to lagging target network parameters, the current Q-network parameters need to be assigned to the target Q-network via a deep copy. A deep copy copies the entire object and all its internal contents, not just object references, ensuring that the target Q-network parameters remain consistent with the current Q-network on each update, unaffected by other factors. The saved file includes the model path, training epochs, and identifiers, facilitating the management of multiple experimental versions and enabling rapid location and reproduction.

[0133] The output of the collaborative control strategy includes the following steps:

[0134] Identify the energy storage systems in the heterogeneous hybrid energy storage system that correspond to peak shaving, frequency regulation, voltage regulation, and black start functions, respectively; and use the results of the strategy iteration to coordinate the charging and discharging actions and charging and discharging parameters of the energy storage systems to obtain an energy storage system operation scheme adapted to different fault scenarios.

[0135] Specifically, peak-shaving energy storage systems can be pumped hydro storage; frequency regulation energy storage systems can be supercapacitors; voltage regulation energy storage systems can be flywheel energy storage; black-start energy storage systems can be lithium-ion batteries; charging and discharging actions can be charging, discharging, and standby; charging and discharging parameters can be charging and discharging power and charging and discharging duration; and operating schemes can be arbitrage peak-shaving schemes under normal scenarios and emergency power supply schemes under fault scenarios.

[0136] For example, the following parameters are defined:

[0137] Real-time frequency Real-time operating frequency of the power grid, in Hz; frequency deviation. : This reflects the degree to which the grid frequency deviates from its rated value; real-time voltage at the grid connection point. Real-time voltage at the grid connection point of the energy storage system, in kV; rated voltage. Rated operating voltage of the power grid, in kV; real-time load power. System real-time load demand, in MW;

[0138] Load reference value The historical average load for the same period is used as the threshold for determining peak shaving and valley filling, in MW. Real-time electricity price. Real-time electricity trading price in the system, unit: yuan / kWh;

[0139] Peak electricity price Electricity price at its highest point, in yuan / kWh; Off-peak electricity price. The lowest electricity price in the system, expressed in yuan / kWh; apparent power. : No. Apparent power of an energy storage system, in MW; rated charge / discharge power of the energy storage system. : No. The maximum charging / discharging power of an energy storage system, in MW;

[0140] Rated capacity of energy storage system : No. The rated capacity of an energy storage system, in MWh;

[0141] State of charge of energy storage system : No. The percentage of remaining electricity in each energy storage system reflects the available capacity of the energy storage system.

[0142] (1) Normal scenario: The power grid is operating stably without significant disturbances. The triggering conditions are as follows:

[0143] .

[0144] Based on the optimization results of the SAC algorithm, the following control strategy is adopted:

[0145] The energy storage system 1 dynamically adjusts its charging and discharging power according to the load level and electricity price period to achieve "peak-load discharge for peak shaving and off-peak charging for arbitrage".

[0146] .

[0147] The power calculation formula of the energy storage system 1 includes real-time load power, load benchmark value, maximum charging / discharging power of the energy storage system 1, real-time electricity price, peak electricity price and valley electricity price. The power of the energy storage system 1 is obtained by calculation, realizing "peak discharge for peak shaving and charging arbitrage during off-peak hours".

[0148] Energy storage system 2 reserves basic frequency regulation capacity to cope with potential frequency disturbances. If the load suddenly increases (but does not reach the abnormal threshold), it will assist energy storage system 1 in peak regulation.

[0149] .

[0150] The power calculation formula of the energy storage system 2 includes real-time load power, load reference value, maximum charging / discharging power of the energy storage system 2, real-time electricity price, peak electricity price and valley electricity price. The power of the energy storage system 2 is obtained by calculation, which can effectively cope with potential frequency disturbances.

[0151] The energy storage system outputs / absorbs reactive power to maintain voltage stability. If there is no voltage deviation, it can charge and discharge slightly to store idle electrical energy.

[0152] .

[0153] This is the voltage regulation coefficient for normal scenarios, balancing voltage regulation accuracy and reactive power loss.

[0154] This indicates that there is no reactive power output. This indicates the absorption of reactive power.

[0155] The power calculation formula of the energy storage system 3 includes the voltage regulation coefficient, the maximum charging / discharging power of the energy storage system 3, the real-time voltage at the grid connection point, and the rated voltage. The power of the energy storage system 3 is obtained by calculation to maintain voltage stability.

[0156] The energy storage system 4 maintains a high charge state to cope with fault scenarios, and only charges slightly during periods of low electricity prices.

[0157] ; For time step, For charging efficiency.

[0158] The power calculation formula of the energy storage system 4 includes the rated apparent power of the fourth energy storage system, the voltage regulation coefficient, the rated capacity of the energy storage system 4, the time step, and the charging efficiency. The power of the energy storage system 4 is obtained by calculation to cope with fault scenarios.

[0159] (2) Abnormal scenario: The power grid experiences a slight disturbance, requiring local regulation to be initiated. The triggering conditions are as follows:

[0160] ;

[0161] Based on the optimization results of the SAC algorithm, the following control strategy is adopted:

[0162] Sub-scenario 1: Voltage Anomaly .

[0163] Energy storage system 1 continues to perform peak shaving tasks, but if charging is required, the charging power is limited to avoid exacerbating the active power deficit of the grid.

[0164] .

[0165] The power calculation formula of the energy storage system 1 includes real-time load power, load benchmark value, maximum charging / discharging power of the energy storage system 1, real-time electricity price, peak electricity price and valley electricity price. The power of the energy storage system 1 is obtained by calculation, realizing "peak discharge for peak shaving and charging arbitrage during off-peak hours".

[0166] Energy storage system 2 recovers auxiliary peak-shaving power and reserves full frequency regulation backup to avoid the superposition of frequency and voltage disturbances;

[0167] .

[0168] The power calculation formula of the energy storage system 2 includes the maximum charging / discharging power of the energy storage system 2. The power of the energy storage system 2 is obtained by calculation, which can effectively cope with potential frequency disturbances.

[0169] Energy storage system 3 increases reactive power output, stops light-load energy storage, and ensures voltage regulation response priority;

[0170] ;

[0171] This is the voltage regulation coefficient for abnormal voltage conditions.

[0172] The power calculation formula of the energy storage system 3 includes the voltage abnormality regulation coefficient, the maximum charging / discharging power of the energy storage system 3, the real-time voltage at the grid connection point, and the rated voltage. The power of the energy storage system 3 is obtained by calculation to maintain voltage stability.

[0173] Energy storage system 4 stops charging during off-peak hours to avoid increasing the active power load on the power grid and maintains a SOC of ≥80% for standby.

[0174] .

[0175] The power and state of charge formulas of the energy storage system 4 limit the power and state of charge of the system so that it can serve as a backup.

[0176] Sub-scenario 2: Frequency anomaly ;

[0177] Energy storage system 1 performs normal peak shaving, but reduces the rate of change of charging and discharging power to avoid aggravating frequency fluctuations;

[0178] .

[0179] The power calculation formula of the energy storage system 1 includes the power of the energy storage system 1 in the previous time step, the real-time load power, the load reference value, and the maximum charging / discharging power of the energy storage system 1. The power of the energy storage system 1 is obtained by calculation to achieve peak regulation.

[0180] Energy storage system 2 utilizes all reserved backup power to output / absorb active power to correct frequency deviations;

[0181] ;

[0182] (Frequency anomaly modulation coefficient).

[0183] The power calculation formula of the energy storage system 2 includes the frequency abnormality modulation coefficient, frequency deviation and the maximum charging / discharging power of the energy storage system 2. The power of the energy storage system 2 is obtained by calculation, which effectively corrects the frequency deviation.

[0184] Energy storage system 3 regulates voltage normally, but reduces the reactive power change rate to avoid affecting frequency regulation;

[0185] ;

[0186] This is the voltage regulation coefficient for normal scenarios, balancing voltage regulation accuracy and reactive power loss.

[0187] The reactive power calculation formula of the energy storage system 3 includes the voltage regulation coefficient in normal scenarios, the apparent power of the energy storage system 3, the reactive power of the energy storage system 3 at the previous moment, the real-time voltage at the grid connection point, and the rated voltage. The reactive power of the energy storage system 3 is obtained by calculation to maintain voltage stability.

[0188] Energy storage system 4 stops charging to prevent the energy storage system from absorbing active power during charging, which would exacerbate the low frequency.

[0189] .

[0190] The power and state of charge formulas of the energy storage system 4 limit the power and state of charge of the system so that it can serve as a backup.

[0191] (3) Fault scenario: The power grid experiences a severe disturbance, requiring the activation of emergency control. The triggering conditions are as follows:

[0192] .

[0193] Based on the optimization results of the SAC algorithm, the following control strategy is adopted:

[0194] Sub-scenario 1: Voltage Failure ;

[0195] Energy storage system 1 stops daily peak shaving, charges and discharges at full power, and the auxiliary voltage is restored;

[0196] ;

[0197] For real-time photovoltaic power, This represents real-time wind power output.

[0198] The power calculation formula of the energy storage system 1 includes rated voltage, real-time voltage at the grid connection point, real-time load power, real-time photovoltaic power, real-time wind power, load reference value, and maximum charging / discharging power of the energy storage system 1. The power of the energy storage system 1 is obtained by calculation, so that its auxiliary voltage can be restored.

[0199] Energy storage system 2 stops frequency regulation, charges and discharges at full power, and prioritizes voltage recovery;

[0200] .

[0201] The power calculation formula of the energy storage system 2 includes the rated voltage, the real-time voltage at the grid connection point, and the maximum charging / discharging power of the energy storage system 2. The power of the energy storage system 2 is obtained by calculation, and voltage recovery is given priority.

[0202] Energy storage system 3 outputs reactive power at full capacity based on voltage deviation, while avoiding long-term overload damage to equipment;

[0203] ;

[0204] (Fault voltage regulation coefficient).

[0205] The reactive power calculation formula of the energy storage system 3 includes the apparent power of the energy storage system 3, the reactive power of the energy storage system 3 at the previous moment, the real-time voltage at the grid connection point, the rated voltage, and the fault voltage regulation coefficient. The reactive power of the energy storage system 3 is obtained by calculation, while allowing 1.2 times overload operation (within 10 minutes) to ensure that the voltage regulation intensity remains stable under fault scenarios.

[0206] The energy storage system discharges at full power to ensure power supply to critical loads, and stops when the SOC drops to 20% to prevent over-discharge of the energy storage system.

[0207] ;

[0208] This represents the discharge power of the fourth energy storage system.

[0209] The power calculation formula of the energy storage system 4 includes the real-time voltage at the grid connection point, the rated voltage, the real-time load power, and the load reference value. The power of the energy storage system 4 is obtained by calculation to ensure the power supply of critical loads.

[0210] Sub-scenario 2: Frequency Failure .

[0211] Energy storage system 1 stops peak shaving and charges and discharges according to the direction of frequency deviation to assist energy storage system 2 in frequency regulation;

[0212] ;

[0213] (Fault frequency regulation coefficient of energy storage system 1).

[0214] The power calculation formula of the energy storage system 1 includes the fault frequency regulation coefficient, frequency deviation, and the maximum charging / discharging power of the energy storage system 1. The power of the energy storage system 1 is obtained by calculation to assist in frequency regulation.

[0215] Energy storage system 2 operates under overload conditions to maximize frequency regulation and quickly correct frequency deviations.

[0216] ;

[0217] (Fault frequency regulation coefficient of energy storage system 2).

[0218] This is the charging power for the second energy storage system.

[0219] Allows operation at 1.2 times overload for no more than 10 minutes; after this period, restores rated power.

[0220] The power calculation formula of the energy storage system 2 includes the fault frequency regulation coefficient, frequency deviation, and the charging power of the second energy storage system. The power of the energy storage system 2 is obtained by calculation, and the frequency deviation is quickly corrected.

[0221] Energy storage system 3 regulates voltage normally and does not start overload, avoiding conflict between reactive power regulation and frequency regulation;

[0222] ;

[0223] (Fault regulation coefficient of energy storage system).

[0224] The reactive power calculation formula of the energy storage system 3 includes the apparent power of the energy storage system 3, the real-time voltage at the grid connection point, the rated voltage, and the fault frequency regulation coefficient. The reactive power of the energy storage system 3 is obtained by calculation to ensure normal voltage regulation of the energy storage system 3.

[0225] Energy storage system 4 does not participate in frequency regulation, maintains SOC ≥ 80%, and is prepared to cope with possible voltage faults.

[0226] .

[0227] The power and state of charge formulas of the energy storage system 4 limit the power and state of charge of the system so that it can serve as a backup.

[0228] Furthermore, the energy storage system operation scheme adapted to different fault scenarios includes scenario switching and strategy integration, which includes the following steps:

[0229] Collect grid frequency and voltage data and supplement them to the various time-series variables; determine the scenario type based on the supplemented time-series variables at preset time intervals; switch the collaborative control strategy when the corresponding scenario conditions are met according to the determination results; when both grid voltage and grid frequency are abnormal or faulty, adjust the voltage to the normal range first, and then adjust the frequency; when switching scenarios, control the power change rate of the energy storage system to not exceed a preset threshold; after the abnormal or faulty scenario ends, control the energy storage system to recover to the target state of charge within a preset time.

[0230] For example, during scene switching, the system collects the power grid frequency in real time. and voltage The system assesses the current operating scenario every 100ms. Once the system detects that the current state meets the switching conditions for the corresponding scenario, it immediately executes the control strategy switching operation. When both the voltage and frequency of the power grid experience abnormalities or faults, considering that voltage is a fundamental indicator for ensuring the safe operation of electrical equipment, the system will prioritize adjusting the voltage. Once the voltage returns to the normal range, the frequency will be adjusted accordingly.

[0231] Specific constraints were set during the strategy transition process. At the instant of scenario switching, the power change rate of each energy storage system in the heterogeneous hybrid energy storage system must not exceed 0.5 times the rated charging and discharging power of the energy storage system. This constraint effectively avoids the impact of sudden power changes on the stable operation of the power grid, ensuring a smooth transition during the strategy switching process.

[0232] The SOC recovery mechanism ensures the continuous and reliable operation of the system. After a fault or abnormal scenario ends, each energy storage system in the heterogeneous hybrid energy storage system needs to recover to the preset target state of charge within 30 minutes. This mechanism ensures that the energy storage system can maintain good responsiveness and play its corresponding role in a timely manner when encountering various grid scenarios again.

[0233] Step 107: Repeat the following operations: power grid operation status identification, input state vector construction, reward function application, and strategy iteration, until the value of the reward function is maximized, and output the final cooperative control strategy.

[0234] In the description of this specification, the references to terms such as "one embodiment," "some embodiments," "example," "specific example," or "some examples," etc., indicate that a specific feature, structure, material, or characteristic described in connection with that embodiment or example is included in at least one embodiment or example of the present invention. In this specification, the illustrative expressions of the above terms do not necessarily refer to the same embodiment or example. Furthermore, the specific features, structures, materials, or characteristics described may be combined in any suitable manner in one or more embodiments or examples. Moreover, without contradiction, those skilled in the art can combine and integrate the different embodiments or examples described in this specification, as well as the features of different embodiments or examples.

[0235] The above embodiments are merely illustrative of the technical concept and features of the present invention, and are intended to enable those skilled in the art to understand the content of the present invention and implement it accordingly. They should not be construed as limiting the scope of protection of the present invention. All equivalent changes or modifications made based on the essence of the content of the present invention should be covered within the scope of protection of the present invention.

Claims

1. A hybrid energy storage cooperative control method for adaptive variable step frequency TCN-SAC, characterized in that, Includes the following steps: Based on the characteristics of power grid faults and the properties of heterogeneous energy storage, a power grid fault classification scheme is proposed, and a heterogeneous hybrid energy storage system model for dealing with differentiated power grid fault scenarios is constructed. Collect various time-series variables that affect the system's operating state, extract historical time-series data from these variables, and use the historical time-series data to model and train an adaptive variable step frequency TCN network to obtain supply and demand variable prediction information for future moments. According to the power grid fault classification scheme, the monitored power grid operating status is identified, and the fault status data corresponding to the current power grid is determined. Historical system operation data is extracted from the various time-series variables, and the fault status data, historical system operation data and supply and demand variable prediction information are fused to form the input state vector of the SAC algorithm. Taking into account the objectives related to failure cost, park revenue, and battery cycle life cost, a reward function for the SAC algorithm is constructed. The reward function is used to iterate the input state vector to obtain a cooperative control strategy for the heterogeneous hybrid energy storage system. Repeat the following steps: power grid operation state identification, input state vector construction, reward function application, and strategy iteration, until the value of the reward function is maximized, and then output the final cooperative control strategy.

2. The hybrid energy storage cooperative control method for adaptive variable step frequency TCN-SAC according to claim 1, characterized in that, The construction of the power grid fault classification scheme includes the following steps: Based on the real-time frequency and voltage of the power grid, the system operation scenarios are classified into normal scenarios, abnormal scenarios, and fault scenarios. Based on the preset ranges corresponding to the deviation value of the real-time frequency of the power grid and the actual value of the voltage, the scenario type is determined, and the scenario in which the system is located is identified.

3. The hybrid energy storage cooperative control method for adaptive variable step frequency TCN-SAC according to claim 1, characterized in that, The modeling and training of the adaptive variable step frequency TCN network includes the following steps: The multi-step prediction results and single-step prediction results of the adaptive variable step frequency TCN network are calculated respectively to obtain the corresponding two mean absolute errors. Calculate the error ratio based on the two mean absolute errors; The step size of the adaptive variable step frequency TCN network is adjusted according to the magnitude of the error ratio. The adaptive variable step frequency TCN network with adjusted step size is used to model and train the historical time series data, and output the supply and demand variable prediction information.

4. The hybrid energy storage cooperative control method of adaptive variable step frequency TCN-SAC according to claim 1, characterized in that, The construction of the input state vector of the SAC algorithm includes the following steps: Collect grid voltage and grid frequency data to supplement the fault status data; Collect renewable energy generation data and grid load data to supplement the system's historical operating data; By integrating the supplemented fault state data, the supplemented historical system operation data, and the supply and demand variable prediction information, the input state vector of the SAC algorithm is obtained.

5. The hybrid energy storage cooperative control method for adaptive variable step frequency TCN-SAC according to claim 1, characterized in that, The reward function construction of the SAC algorithm includes the following steps: Based on parameters related to electricity sales, electricity purchase, peak shaving services, frequency regulation services, and voltage regulation services, the objective function for system revenue is calculated. Based on the power parameters and state of charge parameters of each energy storage system, the objective function for the lifespan of the energy storage system is calculated. Based on the discharge energy parameters, real-time load parameters, and load deficit parameters of the energy storage system, the reliability objective function of the energy storage system is calculated. The reward function of the SAC algorithm is obtained by weighted summing of the system revenue objective function, the energy storage system lifetime objective function, and the energy storage system reliability objective function.

6. The hybrid energy storage cooperative control method for adaptive variable step frequency TCN-SAC according to claim 2, characterized in that, The output of the cooperative control strategy includes the following steps: Identify the energy storage systems in the heterogeneous hybrid energy storage system that correspond to peak shaving, frequency regulation, voltage regulation, and black start functions, respectively; By utilizing the results of the strategy iteration, the charging and discharging actions and parameters of various energy storage systems are coordinated and controlled to obtain an energy storage system operation scheme adapted to different fault scenarios.

7. The hybrid energy storage cooperative control method for adaptive variable step frequency TCN-SAC according to claim 1, characterized in that, The policy iteration based on the reward function includes the following steps: The SAC algorithm generates a parameterized policy through its policy network, and the actions are sampled and mapped to a preset range. The value function of the state-action pair is evaluated using the two Q-value networks of the SAC algorithm. The SAC algorithm provides a stable TD target through two target Q networks, and the parameters of the target Q networks are adjusted using a soft update method. Automatic entropy adjustment is performed using the entropy adjustment coefficient and optimizer of the SAC algorithm. Based on the reward function, calculate the target Q value, the Q network loss function, and the policy network loss function; Based on the loss function, the parameters of the Q-value network and the policy network are updated respectively, completing one policy iteration.

8. The hybrid energy storage cooperative control method for adaptive variable step frequency TCN-SAC according to claim 6, characterized in that, The energy storage system operation scheme adapted to different fault scenarios includes scenario switching and strategy integration, which includes the following steps: Collect power grid frequency data and power grid voltage data, and supplement them to the various time-series variables; The scenario type is determined based on the supplemented time-series variables according to a preset time interval; Based on the determination result, the collaborative control strategy is switched when the corresponding scenario conditions are met. When both the grid voltage and grid frequency are abnormal, first adjust the voltage to the normal range, and then adjust the frequency. When switching scenarios, the power change rate of the energy storage system is controlled to not exceed a preset threshold. After the abnormal scenario ends, the energy storage system is controlled to restore to the target state of charge within a preset time.

9. The hybrid energy storage cooperative control method for adaptive variable step frequency TCN-SAC according to claim 3, characterized in that, The step size adjustment of the adaptive variable step frequency TCN network includes the following steps: Set the initial step size of the adaptive variable step frequency TCN network; The error ratio is obtained by calculating the ratio of the mean absolute error of the multi-step prediction results to the mean absolute error of the single-step prediction results. When the error ratio is greater than a preset value, the step size of the adaptive variable step frequency TCN network is reduced. When the error ratio is equal to a preset value, the step size of the adaptive variable step frequency TCN network remains unchanged; When the error ratio is less than a preset value, the step size of the adaptive variable step frequency TCN network is increased; The adjusted step size is subject to upper and lower limit constraints to ensure that the step size is within a preset range.

10. The hybrid energy storage cooperative control method for adaptive variable step frequency TCN-SAC according to claim 5, characterized in that, When performing a weighted summation of the system revenue objective function, the energy storage system lifetime objective function, and the energy storage system reliability objective function, the weight coefficients of each objective function are adjusted according to different failure scenarios.

Images