Hybrid energy storage assisted secondary frequency modulation method

Abstract

The present application relates to the field of power grid frequency modulation technology, and particularly relates to a hybrid energy storage assisted secondary frequency modulation method, comprising: constructing a multi-dimensional time sequence characteristic vector based on historical operation data; predicting AGC compensation power demand and its uncertainty interval by using a trained quantile neural network prediction model; obtaining energy storage system output power by using a preset switching MPC optimization strategy; performing frequency domain decomposition on the energy storage system output power by using a preset VMD algorithm to obtain high-frequency power components and low-frequency power components, and performing power distribution on the hybrid energy storage to assist secondary frequency modulation. The present application realizes AGC compensation power and uncertainty prediction by using a quantile neural network, improves control adaptability in combination with switching MPC; realizes high and low frequency decoupling by using VMD to decompose power signals; respectively distributes high-frequency and low-frequency power to super capacitor and battery energy storage to realize hybrid energy storage collaborative frequency modulation, improve system response speed, stability and energy storage utilization efficiency, and reduce SOC fluctuation and frequency modulation error.

Description

Technical Field

[0001] This invention relates to the field of power grid frequency regulation technology, and specifically to a hybrid energy storage-assisted secondary frequency regulation method. Background Technology

[0002] With the large-scale grid connection of new energy sources, the output of renewable energy such as wind power and photovoltaics exhibits significant randomness and volatility, leading to increasingly prominent power imbalances in the power grid and a continuously increasing demand for secondary frequency regulation (AGC). Traditional frequency regulation methods, which are mainly based on thermal power units, are limited by slow ramp rates, low regulation accuracy, and response lags, making it difficult to meet the requirements of high-proportion renewable energy power systems for fast and accurate frequency regulation.

[0003] Hybrid energy storage systems, combining the advantages of high power density and high energy density, are gradually becoming an important means of participating in secondary frequency regulation of the power grid. Supercapacitors offer fast response times and are suitable for compensating for high-frequency power fluctuations, while battery storage boasts large energy capacity and is suitable for smoothing medium- and low-frequency power. The synergy between the two can effectively improve the system's frequency regulation performance. However, in actual operation, frequency regulation power demand exhibits strong uncertainty and time-varying characteristics, and AGC commands possess non-stationary features, posing significant challenges to the power allocation and control strategies of energy storage systems.

[0004] Existing methods mostly employ deterministic prediction or fixed control strategies, which make it difficult to accurately characterize the uncertainty range of AGC compensation power. They also lack the ability to effectively decompose and coordinate the power characteristics at different time scales, resulting in insufficient response stability of energy storage systems. Summary of the Invention

[0005] (a) Purpose of the invention The purpose of this invention is to provide a hybrid energy storage-assisted secondary frequency regulation method. This method uses a quantile neural network to achieve AGC compensation power and uncertainty prediction, and combines switching MPC to improve control adaptability. It uses VMD to decompose the power signal to achieve high- and low-frequency decoupling. The high-frequency and low-frequency power are respectively allocated to the supercapacitor and battery energy storage to achieve hybrid energy storage coordinated frequency regulation, thereby improving system response speed, stability and energy storage utilization efficiency, and reducing SOC fluctuation and frequency regulation error.

[0006] (II) Technical Solution To address the above problems, this invention provides a hybrid energy storage-assisted secondary frequency regulation method, comprising: Based on historical operating data, a multi-dimensional time-series feature vector is constructed. The historical operating data includes power grid operating data, traditional generating unit operating data, energy storage operating data, and environmental data. Based on the multidimensional time-series feature vector, a trained quantile neural network prediction model is used to predict the AGC compensation power demand and its uncertainty range. Based on the AGC compensation power requirement and its uncertainty range, the output power of the energy storage system is obtained by using a preset switching MPC optimization strategy. The output power of the energy storage system is decomposed in the frequency domain using a preset VMD algorithm to obtain high-frequency power components and low-frequency power components. Based on the high-frequency power component and the low-frequency power component, power is allocated to the hybrid energy storage, generating a hybrid energy storage response AGC command to assist in secondary frequency regulation.

[0007] In another aspect of the present invention, preferably, The hybrid energy storage includes energy storage batteries and supercapacitors; The power grid operation data includes: power grid frequency deviation, tie-line power deviation, area control deviation, and AGC dispatch instructions; The conventional unit operating data includes: the actual load power output of the conventional unit; The energy storage operation data includes: state of charge, maximum rated charging power, and maximum rated discharging power. The environmental data includes: ambient temperature.

[0008] In another aspect of the present invention, preferably, The quantile neural network prediction model includes: an input layer, an LSTM unit, a TCN unit, a feature fusion layer, and a quantile output layer. The input layer is connected to the LSTM unit and the TCN unit respectively. The LSTM unit and the TCN unit are arranged in parallel. The LSTM unit and the TCN unit are connected to the feature fusion layer respectively. The feature fusion layer is connected to the quantile output layer. The method of predicting the AGC compensation power demand and its uncertainty range for future time series based on the multidimensional time series feature vector using a trained quantile neural network prediction model includes: Based on the multidimensional temporal feature vector, input layer, LSTM unit and TCN unit, the LSTM unit output and TCN unit output are obtained; Based on the LSTM unit output, TCN unit output, and feature fusion layer, comprehensive temporal features are obtained. Based on the comprehensive time series features and the quantile output layer, the AGC compensation power requirements of multiple preset quantiles in future time series are predicted. Based on the AGC compensation power requirements of multiple preset quantiles, the corresponding uncertainty range is obtained.

[0009] In another aspect of the present invention, preferably, The step of obtaining the energy storage system output power based on the AGC compensation power demand and its uncertainty range, using a preset switching MPC optimization strategy, includes: The AGC compensation power requirement and its uncertainty range are divided according to a preset optimization cycle to obtain the AGC operation optimization sequence; For each optimization cycle, based on the preset membership algorithm, the membership degree is calculated for the corresponding feature vector and typical scenario in the AGC running optimization sequence to obtain the membership degree of each scenario. Based on the membership degree of each scenario, determine the multi-scenario adaptive AGC model; Based on the membership degree of each scenario, the cost functions corresponding to different scenarios are weighted to construct a multi-scenario optimization objective function; Based on the multi-scenario adaptive AGC model, the objective function for scenario optimization is solved to obtain the output power of the first energy storage system; Based on the uncertainty range and energy storage operation data, determine the power constraint boundary corresponding to the energy storage system; The output power of the energy storage system is obtained based on the output power of the first energy storage system and the power constraint boundary.

[0010] In another aspect of the present invention, preferably, The preset membership algorithm is expressed by the following formula: in, Let represent the membership degree of the k-th scenario in the t-th optimization cycle. This represents the eigenvector of the t-th optimization cycle; This represents the cluster center of the k-th scene; This represents the cluster center of the j-th scene, where j is different from k. This represents the temperature coefficient.

[0011] In another aspect of the present invention, preferably, The multi-scenario optimization objective function is expressed by the following formula: in, Let Jk represent the membership degree of the k-th scene in the t-th optimization cycle, Jk represent the cost function corresponding to scene k, and u represent the input sequence to be optimized. Let represent the estimated model parameters for the k-th scene at time t, where i represents the step index in the prediction time domain, N represents the length of the prediction time domain, N-1 represents the first N-1 time steps in the prediction time domain, y(t+i|t) represents the prediction of the system output at time t+i at time t, yref represents the setpoint of the controlled variable, and Qk represents the output bias penalty matrix for scene k. Rk represents the change in control input, Rk represents the penalty matrix for control increment in scenario k, y(t+N) represents the prediction time domain endpoint, and Pk represents the terminal deviation penalty matrix in scenario k.

[0012] In another aspect of the present invention, preferably, The step of determining the power constraint boundary corresponding to the energy storage system based on the uncertainty interval and energy storage operation data includes: Based on the uncertainty range, extract the corresponding uplink uncertainty and downlink uncertainty; When the state of charge is less than the preset lower threshold, the upper limit of the rated discharge power is constrained and tightened according to the downlink uncertainty to determine the discharge power constraint boundary. When the state of charge is greater than the preset upper limit threshold, the upper limit of the rated charging power is constrained and tightened according to the uplink uncertainty to determine the charging power constraint boundary. When the state of charge is between a preset lower threshold and a preset upper threshold, the charging and discharging power constraint is processed according to the uplink uncertainty and downlink uncertainty to determine the charging and discharging power constraint boundary; The discharge power constraint boundary, charging power constraint boundary, and charge / discharge power constraint boundary are the power constraint boundaries corresponding to the energy storage system.

[0013] In another aspect of the present invention, preferably, The step of using a preset VMD algorithm to perform frequency domain decomposition on the output power of the energy storage system to obtain high-frequency power components and low-frequency power components includes: The output power of the energy storage system is preprocessed to obtain the power signal to be decomposed; The number of decomposition layers of the preset VMD algorithm is determined based on the decomposition information entropy index and the decomposition energy ratio index. The original cutoff frequency is adjusted based on the scenario with the highest membership degree and energy storage operation data to obtain the preset cutoff frequency of the VMD algorithm. Based on the number of decomposition layers and the cutoff frequency, the power signal to be decomposed is decomposed into high-frequency power components and low-frequency power components.

[0014] In another aspect of the present invention, preferably, The process of allocating power to the hybrid energy storage based on the high-frequency and low-frequency power components, generating a hybrid energy storage response AGC command, and assisting in secondary frequency regulation includes: The high-frequency power component is allocated to the supercapacitor, and the low-frequency power component is allocated to the energy storage battery. When the state of charge of any energy storage unit reaches a preset upper limit threshold or a preset lower limit threshold, the equalization control mode is triggered to obtain the power allocation result. The power allocation result is then subjected to multi-constraint correction to generate a corrected hybrid energy storage response AGC command. Secondary frequency regulation is assisted based on the modified hybrid energy storage response AGC command.

[0015] In another aspect of the present invention, preferably, The method further includes: training the quantile neural network prediction model and updating the switching MPC optimization strategy using an adaptive learning mechanism; The adaptive learning mechanism includes: A scrolling data window is established based on historical and real-time operational data; Based on the scrolling data window, error calculation is performed to determine whether an update should be triggered; If the triggering condition is not met, return to continue error calculation; If the triggering conditions are met, the switching MPC optimization strategy is quickly updated based on meta-learning of the data in the scrolling data window. During the update process, EWC regularization constraints are introduced to restrict the updates of key parameters based on the Fisher information matrix.

[0016] (III) Beneficial Effects The above-described technical solution of the present invention has the following beneficial technical effects: This invention constructs a multi-dimensional time-series feature vector and introduces a quantile neural network prediction model to predict the power demand for AGC compensation and its uncertainty range. This effectively characterizes the randomness and volatility of grid frequency regulation power, improving the accuracy and robustness of the prediction results. Based on the prediction results, a switching MPC optimization strategy is further introduced, enabling the energy storage system to adaptively switch control strategies under different operating conditions, thereby improving the system's response capability and operational stability to complex dynamic changes.

[0017] Meanwhile, the VMD algorithm is used to decompose the output power of the energy storage system in the frequency domain, effectively decomposing the power signal into high-frequency and low-frequency components. The high-frequency power components are allocated to the supercapacitor energy storage unit with fast response speed, and the low-frequency power components are allocated to the battery energy storage unit with large energy capacity. This achieves complementary advantages and coordinated regulation of multiple types of energy storage, improves the system response speed, stability and energy storage utilization efficiency, and reduces SOC fluctuation and frequency regulation error. Attached Figure Description

[0018] Figure 1 This is an overall flowchart of one embodiment of the present invention; Figure 2 This is a schematic diagram of the overall architecture of one embodiment of the present invention; Figure 3This is a schematic diagram of the quantile neural network prediction model structure according to an embodiment of the present invention; Figure 4 This is a flowchart of an adaptive learning mechanism according to an embodiment of the present invention. Detailed Implementation

[0019] To make the objectives, technical solutions, and advantages of this invention clearer, the invention will be further described in detail below with reference to specific embodiments and the accompanying drawings. It should be understood that these descriptions are merely exemplary and not intended to limit the scope of the invention. Furthermore, descriptions of well-known structures and techniques are omitted in the following description to avoid unnecessarily obscuring the concept of the invention.

[0020] The accompanying drawings show structural schematic diagrams according to embodiments of the present invention. These drawings are not drawn to scale, and some details have been enlarged for clarity, and some details may have been omitted. The shapes of the various regions and layers shown in the drawings, as well as their relative sizes and positional relationships, are merely exemplary and may deviate from reality due to manufacturing tolerances or technical limitations. Furthermore, those skilled in the art can design regions / layers with different shapes, sizes, and relative positions as needed.

[0021] Obviously, the described embodiments are only some, not all, of the embodiments of the present invention. All other embodiments obtained by those skilled in the art based on the embodiments of the present invention without inventive effort are within the scope of protection of the present invention.

[0022] In the description of this invention, it should be noted that the terms "first," "second," and "third" are used for descriptive purposes only and should not be construed as indicating or implying relative importance.

[0023] Furthermore, the technical features involved in the different embodiments of the present invention described below can be combined with each other as long as they do not conflict with each other.

[0024] The invention will now be described in more detail with reference to the accompanying drawings. In the various drawings, the same elements are indicated by similar reference numerals. For clarity, the various parts in the drawings are not drawn to scale.

[0025] Example 1 A hybrid energy storage-assisted secondary frequency regulation method, Figure 1 An overall flowchart of one embodiment of the present invention is shown. Figure 2 A schematic diagram of the overall architecture of an embodiment of the present invention is shown, as follows: Figure 1 and Figure 2 As shown, it includes: Based on historical operating data, the data is divided according to a preset sampling period, for example, 1 second. The length of the historical sequence after the data division is 127 steps. A multi-dimensional time-series feature vector is constructed. The historical operating data includes grid operating data, traditional generating unit operating data, energy storage operating data, and environmental data. Here, grid operating data includes: grid frequency deviation, tie-line power deviation, regional control deviation, and AGC dispatch instructions; traditional generating unit operating data includes: actual load power output of traditional generating units; energy storage operating data includes: state of charge, rated charging power limit, and rated discharging power limit; environmental data includes: ambient temperature. Hybrid energy storage includes energy storage batteries and supercapacitors. The multi-dimensional time-series feature vector includes: grid frequency deviation and its rate of change, tie-line power deviation and its rate of change, regional control deviation and its rate of change, AGC dispatch instructions and their rate of change, actual load power output of traditional generating units, energy storage battery state of charge, supercapacitor state of charge, ambient temperature, time period encoding, and seasonal encoding.

[0026] Based on the aforementioned multidimensional time-series feature vector, a trained quantile neural network prediction model is used to predict the AGC compensation power demand and its uncertainty range, including: Based on the multidimensional time-series feature vector, input layer, LSTM unit, and TCN unit, the outputs of the LSTM unit and TCN unit are obtained. Parallel feature learning is performed through the LSTM unit (Long Short-Term Memory network unit) and the TCN unit (Time-Series Convolutional Network Unit). The LSTM unit models long-term dependencies in the time series through its gating mechanism, effectively capturing the cumulative effects and trend changes of AGC commands over a longer time scale, such as slow load changes, unit regulation lag effects, and energy storage SOC evolution trends. Meanwhile, the TCN unit, based on one-dimensional causal convolution and dilated convolution structures, extracts local fluctuation features within a short time scale, effectively capturing high-frequency dynamic characteristics such as grid frequency fluctuations, short-term load shocks, and random fluctuations of renewable energy, thereby enhancing the sensitivity of the quantile neural network prediction model to rapidly changing signals.

[0027] Based on the outputs of the LSTM unit, the TCN unit, and the feature fusion layer, comprehensive temporal features are obtained; the outputs of the LSTM unit and the TCN unit are input to the feature fusion layer for information fusion, which can be a splicing process.

[0028] Based on the comprehensive time-series features and the quantile output layer, the AGC compensation power demand at multiple preset quantiles in the future time series is predicted. The quantile output layer, by setting multiple preset quantiles, such as 0.1, 0.5, and 0.9 quantiles, performs conditional quantile regression prediction on the AGC compensation power demand in the future time series, outputting power demand estimates at different confidence levels. In this embodiment, five quantile prediction values ​​are output simultaneously: , , , , These correspond to confidence levels of 10%, 25%, 50%, 75%, and 90%, respectively.

[0029] Based on the AGC compensation power demand at multiple preset quantiles, a corresponding uncertainty range is obtained. For example, the lower quantile prediction value is used as the lower bound of uncertainty, and the higher quantile prediction value is used as the upper bound of uncertainty, thus constructing the uncertainty range of the AGC compensation power demand. This uncertainty range can reflect the fluctuation range and risk boundary of future frequency regulation demand.

[0030] In this embodiment, the upward uncertainty of the uncertainty interval is expressed by the following formula, and the upward uncertainty of the uncertainty interval characterizes the risk of underestimating the prediction: The downside uncertainty of an uncertainty interval is expressed by the following formula, which characterizes the risk of overestimating the forecast: The overall uncertainty is the predicted AGC compensation power requirement: The uplink uncertainty and downside uncertainty The input is to the preset switching MPC optimization strategy to achieve direction-aware constraint tightening. The uncertainty range is used to determine the power constraint boundary corresponding to the energy storage system, so that the constraint boundary has direction-aware capability: tighten the upper limit of discharge when the SOC is close to the lower limit (to prevent downlink risk), and tighten the upper limit of charging when the SOC is close to the upper limit (to prevent uplink risk).

[0031] Specifically, Figure 3 A schematic diagram of a quantile neural network prediction model structure according to an embodiment of the present invention is shown, as follows: Figure 3As shown, the quantile neural network prediction model includes: an input layer, an LSTM unit, a TCN unit, a feature fusion layer, and a quantile output layer. The input layer is connected to both the LSTM unit and the TCN unit, which are arranged in parallel. The LSTM unit and the TCN unit are each connected to the feature fusion layer, which is then connected to the quantile output layer. The quantile output layer simultaneously predicts five quantiles: 10% / 25% / 50% / 75% / 90%. The LSTM unit and the TCN unit employ a parallel dual-path structure. The LSTM unit consists of three... The network is constructed using stacked Long Short-Term Memory (LSTM) units. The first LSTM unit has an input dimension of 12, a hidden layer dimension of 256, and a dropout rate of 0.2. The second and third LSTM units also have an input dimension of 256, a hidden layer dimension of 256, and a dropout rate of 0.2. These LSTM units capture the AGC scheduling instruction sequence, specifically the intraday patterns and long-term trends of AGC compensation power demand, and output a feature vector with a dimension of 256. The TCN unit consists of six residual dilated causal convolutional modules connected sequentially. Each residual dilated causal convolutional module has a kernel size of 3, dilation coefficients of 1, 2, 4, 8, 16, and 32 respectively, 64 channels, a ReLU activation function, and a dropout rate of 0.2. The TCN unit has a receptive field of 127 steps, used to extract multi-scale high-frequency features, and outputs a feature vector with a dimension of 64.

[0032] The 256-dimensional feature vector output from the LSTM branch is concatenated with the 64-dimensional feature vector output from the TCN branch to obtain a 320-dimensional concatenated feature vector. This concatenated feature vector is then passed sequentially through a first fully connected layer, a second fully connected layer, and a quantile output head. The first fully connected layer has an input dimension of 320, an output dimension of 512, a ReLU activation function, and a dropout rate of 0.2. The second fully connected layer has an input dimension of 512, an output dimension of 256, and a ReLU activation function. The quantile output head has an input dimension of 256 and an output dimension of 5, corresponding to the predicted values ​​of the 10%, 25%, 50%, 75%, and 90% quantiles, respectively.

[0033] The multiquantile output head outputs a predicted sequence of AGC compensation power demand for the next 60 steps, or 60 seconds, as follows: , , , , , where k=1,2,...,60.

[0034] In the uncertainty interval calculation process, based on the predicted values ​​of the five quantiles, the predicted AGC compensation power demand and its uncertainty interval are calculated using the following formula: Furthermore, in this embodiment, the quantile neural network prediction model is trained offline, and the offline training process includes: The training dataset was obtained from 24 months of measured AGC command sequences from a 660MW thermal power unit. The sampling period was 1 second, and the total number of samples was approximately 6.3 × 10⁻⁶. 7 The dataset contains 10 data points. The dataset is divided into a training set, a validation set, and a test set in a 7:1.5:1.5 ratio.

[0035] The quantile loss function is used as the optimization objective during training. in, The quantile loss function, ∈{0.1,0.25,0.5,0.75,0.9} represents the quantile levels, and y represents the measured value. Let be the predicted value of the i-th quantile.

[0036] The optimizer used is the AdamW optimizer, with an initial learning rate of 3×10. -4 The learning rate scheduling strategy is cosine annealing, and the weight decay coefficient is 1×10. -4 The batch size is 256, and the training rounds are 120. An early stopping strategy is adopted, and training is terminated when the validation set loss does not decrease for 10 consecutive rounds.

[0037] Tests showed that the quantile neural network prediction model reduced the RMSE by 23.7% and the MAE by 19.8% on the test set compared to the single LSTM point prediction model, and increased the 90% quantile coverage from 82.3% to 91.6%.

[0038] The quantile neural network prediction model is deployed on an industrial server equipped with an NVIDIA RTX 3090 GPU. During online inference, it receives the latest multidimensional time-series feature vector every 1 second, updates the 127-step sliding window, performs one forward propagation calculation, and outputs the 60-step predicted AGC compensation power requirement and its uncertainty range. The latency of a single inference is no more than 120ms, meeting the real-time requirement of a 1-second control cycle.

[0039] Based on the AGC compensation power demand and its uncertainty range, a preset switching MPC optimization strategy is used to obtain the energy storage system output power. This preset switching MPC optimization strategy replaces the traditional single fixed model MPC, providing a foundation for the accuracy of subsequent scheduling. In this embodiment, the process includes: dividing the AGC compensation power demand and its uncertainty range according to a preset optimization period to obtain an AGC operation optimization sequence; here, the preset optimization period can be a control period. The continuous predicted power demand is divided into discrete optimization time steps, such as one cycle per second, one minute, or five minutes, forming an MPC rolling optimization input sequence that can be used for the preset switching MPC optimization strategy.

[0040] For each optimization cycle, based on the preset membership algorithm, the membership degree of the corresponding feature vector and typical scenario in the AGC running optimization sequence is calculated to obtain the membership degree of each scenario. The specific content of the typical scenario is not limited here. Optionally, it can include four types of typical scenarios: normal frequency adjustment, depth adjustment, ramp transition, and AGC standby. The typical scenario can use the multi-dimensional time-series feature vector constructed based on the above historical running data to pre-train four typical scenario cluster centers offline using the K-means clustering algorithm.

[0041] Membership degree calculation is used to identify the scenario corresponding to the optimization cycle. The preset membership degree algorithm is expressed by the following formula: in, Let represent the membership degree of the k-th scenario in the t-th optimization cycle. This represents the eigenvector of the t-th optimization cycle; This represents the cluster center of the k-th scene; This represents the cluster center of the j-th scene, where j is different from k. This represents the temperature coefficient. Here, k = 1, 2, 3, 4, corresponding to four typical scenarios. This includes regional control deviations and their rates of change, AGC scheduling instructions, actual load power output of traditional generating units, state of charge of energy storage batteries, state of charge of supercapacitors, time-period coding, and seasonal coding. The computational complexity of membership is O(K·d), with a single calculation time of <5ms. Membership calculation determines the corresponding scenario, improving the robustness of hybrid energy storage scheduling under various possible operating conditions.

[0042] Based on the membership degree of each scenario, a multi-scenario adaptive AGC model is determined; specifically, the scenario with the highest membership degree can be used as the determined scenario, and a multi-scenario adaptive AGC model can be determined based on the determined scenario. In this embodiment, a set of dedicated discrete state equations for the AGC system is maintained for each scenario. in, This is a state vector, including grid frequency deviation and tie-line power deviation; For controlling the output of the energy storage system; Let be the state transition matrix for the k-th scene; Let be the input matrix for the k-th scene; For the offset term of the k-th scene, the matrix... Continuous updates are achieved through online recursive least squares identification.

[0043] In the updating of the multi-scenario adaptive AGC model, based on the variable forgetting factor strategy, the recursive least squares (RLS) algorithm is used to identify and update the model parameters for each scenario online. in, Let φ be the parameter vector to be identified, y be the regression vector, and λ be the observed value. k Let λ be the forgetting factor for the k-th scenario, ∈ [0.95, 0.99]. A smaller value is used for scenarios involving deep adjustment to quickly track dynamic changes, while a larger value is used for scenarios involving normal frequency adjustment to ensure stability. In this embodiment, the variable forgetting factor strategy uses λ=0.95 for scenarios involving deep adjustment to quickly track dynamic changes, λ=0.99 for scenarios involving normal frequency adjustment to ensure stability, λ=0.97 for scenarios involving ramp-up transitions to balance response speed and stability, and λ=0.96 for scenarios involving AGC standby to retain a moderate forgetting rate, ensuring that invalid data accumulated during standby does not excessively pollute the model.

[0044] Based on the membership degree of each scenario, the cost functions corresponding to different scenarios are weighted to construct a multi-scenario optimization objective function; in each control cycle, a scenario-weighted convex quadratic programming optimization problem is constructed, and the multi-scenario optimization objective function is expressed by the following formula: Where represents the membership degree of the k-th scene in the t-th optimization cycle, Jk represents the cost function corresponding to scene k, and u represents the input sequence to be optimized. Let represent the estimated model parameters for the k-th scene at time t, where i represents the step index in the prediction time domain, N represents the length of the prediction time domain, N-1 represents the first N-1 time steps in the prediction time domain, y(t+i|t) represents the prediction of the system output at time t+i at time t, yref represents the setpoint of the controlled variable, and Qk represents the output bias penalty matrix for scene k. Rk represents the change in control input, Rk represents the penalty matrix for control increment in scenario k, y(t+N) represents the prediction time domain endpoint, and Pk represents the terminal deviation penalty matrix in scenario k.

[0045] Based on the multi-scenario adaptive AGC model, the objective function of the scenario optimization is solved to obtain the output power of the first energy storage system; the energy storage power that meets the frequency regulation requirements of AGC is generated, while taking into account the constraints of different scenarios.

[0046] During the solution process, the adaptive weight configuration within the multi-scenario adaptive AGC model corresponding to each scenario is as follows: In deep adjustment scenarios, the ACE deviation penalty weight Q is increased by 30% compared to the benchmark value to enhance tracking accuracy; in ramp transition scenarios, the power change rate penalty weight R is increased by 40% compared to the benchmark value to suppress battery high-frequency charging and discharging losses; in AGC standby scenarios, the ACE deviation penalty weight Q is decreased by 50% compared to the benchmark value to avoid ineffective responses to minor deviations. These weights are dynamically linked to the battery health status (SOH): when SOH is below 80%, the power change rate penalty weight R for all scenarios is uniformly increased to 1.5 times the benchmark value to protect aging batteries.

[0047] Based on the uncertainty interval and energy storage operation data, the power constraint boundary corresponding to the energy storage system is determined to achieve dynamic constraint with direction awareness; in this embodiment, it includes: Based on the uncertainty range, extract the corresponding uplink uncertainty and downlink uncertainty; When the state of charge is less than a preset lower threshold, the upper limit of the rated discharge power is tightened based on the downlink uncertainty to determine the discharge power constraint boundary; specifically, this may include: when the battery state of charge is close to the lower limit, tightening the upper limit of the discharge power to prevent downlink prediction risks. Charging power lower limit maintained: in, Where λ represents the upper limit of the rated discharge power, and λ∈[0.1,0.3] is the safety margin coefficient. This represents the downside uncertainty.

[0048] When the state of charge (SOC) exceeds a preset upper limit threshold, the upper limit of the rated charging power is tightened based on the uplink uncertainty to determine the charging power constraint boundary. Specifically, this may include: when the battery SOC approaches the upper limit, tightening the upper limit of the charging power to prevent uplink prediction risks. Discharge power lower limit maintained: in, This is the upper limit of the rated charging power. This represents the upward uncertainty.

[0049] When the state of charge is between a preset lower threshold and a preset upper threshold, the charging and discharging power constraint is processed based on the uplink and downlink uncertainties to determine the charging and discharging power constraint boundary; specifically, this may include: when the battery state of charge is in the intermediate safe region, dynamically relaxing the power constraint based on the comprehensive uncertainty. Where γ∈[0.05,0.1] is the relaxation coefficient. This represents the highest level of uncertainty in history.

[0050] The discharge power constraint boundary, charging power constraint boundary, and charge / discharge power constraint boundary are the power constraint boundaries corresponding to the energy storage system. This enables MPC to adaptively balance the full utilization of energy storage capacity when prediction is reliable and the reserved safety margin when prediction is uncertain, and the uplink and downlink asymmetric processing avoids the capacity waste caused by a one-size-fits-all approach.

[0051] Based on the output power of the first energy storage system and the power constraint boundary, the output power of the energy storage system is obtained. The convex quadratic programming problem is solved using the OSQP solver. Configuration parameters include: absolute convergence tolerance 10. -3 Relative convergence tolerance 10 -3 The maximum number of iterations is 1000. A warm-start strategy is adopted, using the optimal solution of the previous optimization cycle as the starting point of the current cycle to accelerate convergence. The single solution time is 15-50ms, meeting the real-time requirement of a 1s optimization cycle.

[0052] A preset VMD algorithm is used to perform frequency domain decomposition on the output power of the energy storage system, obtaining high-frequency and low-frequency power components. This decouples power fluctuation characteristics at different time scales, providing a basis for subsequent hierarchical power allocation in the hybrid energy storage system. By separating the output power of the energy storage system in the frequency domain, rapidly fluctuating power and slowly changing power can be allocated to different types of energy storage units, thereby improving the response performance and operational lifespan of the hybrid energy storage-assisted secondary frequency regulation. This embodiment includes: The output power of the energy storage system is preprocessed to obtain the power signal to be decomposed. The preprocessing includes time synchronization, missing value repair, outlier removal, and normalization of the output power to eliminate the influence of sampling errors, random noise, and abnormal disturbances on the subsequent frequency domain decomposition results. Simultaneously, to enhance the VMD algorithm's ability to identify power fluctuation characteristics, the power signal can be subjected to moving average filtering or detrending processing to highlight the dynamic changes in power fluctuations, thus forming a power signal suitable for frequency domain decomposition.

[0053] The number of decomposition layers for the preset VMD algorithm is determined based on the decomposition information entropy index and the decomposition energy ratio index. Specifically, for different candidate decomposition layers, multiple VMD decompositions are performed on the power signal to be decomposed, and the information entropy value and energy ratio corresponding to each modal component are calculated. The information entropy index characterizes the information complexity of each modal component after decomposition. When the number of decomposition layers is insufficient, different frequency features may overlap in the same mode, leading to high information entropy. When the number of decomposition layers is too large, over-decomposition may occur, resulting in redundancy between modes. The decomposition energy ratio index measures the contribution of each modal component to the energy of the original power signal. By analyzing the cumulative energy ratio of each modal component, it is determined whether the current number of decomposition layers can effectively cover the main frequency features of the original signal. Combining the information entropy index and the decomposition energy ratio index, the number of decomposition layers that results in low modal information complexity and reasonable energy distribution is selected as the optimal number of decomposition layers for the VMD algorithm, thereby improving the accuracy and stability of frequency domain decomposition. For example, the comprehensive evaluation index G = 0.6·IE + 0.4·SE is used to adaptively determine the number of VMD decomposition layers K ∈ [2, 6].

[0054] The original cutoff frequency is adjusted based on the scenario with the highest membership degree and energy storage operation data to obtain the preset cutoff frequency of the VMD algorithm; specifically, the cutoff frequency is calculated using the following formula: Follow Adaptive adjustment, further, cutoff frequency Adaptive adjustment based on the scenario: Slope transition scenario Increased capacity allows for higher frequency component response in AGC standby scenarios. Reduce and avoid over-responding to low-frequency disturbances.

[0055] Based on the number of decomposition layers and the cutoff frequency, the power signal to be decomposed is decomposed into high-frequency power components and low-frequency power components. Specifically, the VMD algorithm iteratively decomposes the original power signal into several intrinsic mode components with finite bandwidth, and classifies them according to the center frequency and the cutoff frequency of each mode component. Mode components with a center frequency higher than the cutoff frequency are classified as high-frequency power components, used to characterize rapid power fluctuations over short time scales; mode components with a center frequency lower than the cutoff frequency are classified as low-frequency power components, used to characterize slow power change trends over long time scales. Subsequently, mode components belonging to the same category are superimposed and reconstructed to obtain the final high-frequency power components and low-frequency power components, respectively.

[0056] Based on the aforementioned high-frequency and low-frequency power components, power allocation is performed on the hybrid energy storage to generate a hybrid energy storage response AGC command, which assists in secondary frequency modulation, including: The high-frequency power component is allocated to the supercapacitor within ≤200 ms to compensate for the power gap during the unit's ramp-up and start-up phase; the low-frequency power component is allocated to the energy storage battery, continuously tracking AGC commands on a second to minute scale. When the state of charge of any energy storage unit reaches a preset upper or lower threshold, the equalization control mode is triggered to obtain the power allocation result. The power allocation result is subjected to multi-constraint correction to generate a corrected hybrid energy storage response AGC command; the multi-constraint correction may include: SOC over-limit correction, power ramp-up limiting, temperature derating protection and standby correction within the ACE dead zone; Temperature derating protection can be expressed using the following formula: in, Indicates the actual maximum usable power. Indicates the rated maximum power. This represents the temperature derating factor. Indicates ambient temperature. This indicates the temperature protection threshold.

[0057] Secondary frequency regulation is assisted based on the modified hybrid energy storage response AGC command.

[0058] This embodiment constructs a multi-dimensional time-series feature vector and introduces a quantile neural network prediction model to predict the AGC compensation power demand and its uncertainty range. This effectively characterizes the randomness and volatility of grid frequency regulation power, improving the accuracy and robustness of the prediction results. Based on the prediction results, a switching MPC optimization strategy is further introduced, enabling the energy storage system to adaptively switch control strategies under different operating conditions, thereby improving the system's response capability and operational stability to complex dynamic changes.

[0059] Meanwhile, the VMD algorithm is used to decompose the output power of the energy storage system in the frequency domain, effectively decomposing the power signal into high-frequency and low-frequency components. The high-frequency power components are allocated to the supercapacitor energy storage unit with fast response speed, and the low-frequency power components are allocated to the battery energy storage unit with large energy capacity. This achieves complementary advantages and coordinated regulation of multiple types of energy storage, improves the system response speed, stability and energy storage utilization efficiency, and reduces SOC fluctuation and frequency regulation error.

[0060] Furthermore, in this embodiment, the method further includes: training the quantile neural network prediction model and updating the switching MPC optimization strategy using an adaptive learning mechanism; wherein, training the quantile neural network prediction model is an offline meta-learning pre-training stage, and during the training of the quantile neural network prediction model, the 24-month historical data is divided into M meta-tasks {T1,T2,...,T...} according to season × time period × load rate. M The seasonal dimension includes four categories: spring, summer, autumn, and winter; the time period dimension includes five categories: morning peak, noon peak, evening peak, off-peak, and flat period; and the load rate dimension includes three categories: high load, medium load, and low load. M=20 representative meta-tasks are sampled from the theoretical combination, and a new task-sensitive initialization parameter θ is pre-trained using the Model Independent Meta-Learning (MAML) approach. meta : Among them, T m Let α be the learning rate of the inner loop for the m-th meta-task. The gradient of the loss function with respect to the parameters is given. The MAML pre-training has an inner loop learning rate of α=0.01 and an outer loop learning rate of β=0.001, with 5000 pre-training epochs. After pre-training, the quantile neural network prediction model has the ability to quickly adapt to new working conditions; when facing new working conditions, it only requires a small number of gradient steps to converge to good performance.

[0061] Figure 4 A flowchart of an adaptive learning mechanism according to an embodiment of the present invention is shown, as follows: Figure 4 As shown, the adaptive learning mechanism is a two-layer adaptive mechanism, including: A scrolling data window is established based on historical and real-time operational data; Based on the scrolling data window, error calculation is performed to determine whether an update should be triggered; If the triggering condition is not met, return to continue error calculation; If the triggering conditions are met, the switching MPC optimization strategy is quickly updated based on meta-learning of the data in the scrolling data window. During the update process, EWC regularization constraints are introduced to restrict the updates of key parameters based on the Fisher information matrix.

[0062] Specifically, for example, using 24-hour operating data as a rolling data window, when a shift in operating conditions is detected, a few steps of gradient update are performed; in, The online adaptive learning rate is approximately 1 / 10 of the offline pre-training learning rate; the update steps do not exceed 3 steps, and the time taken does not exceed 2 minutes.

[0063] While enabling rapid adaptation, an Elastic Weight Consolidation (EWC) regularization term is introduced to constrain important parameters from deviating from their historical optimum values: in, Let the loss function be the current task. F is the EWC regularization coefficient. i Let i be the i-th diagonal element of the Fisher information matrix, quantifying the importance of the i-th parameter to the historical task. This represents the historically optimal parameter value. For highly important parameters (F... i Parameters exceeding the threshold (approximately 15% of the total parameters) are frozen during the update process and do not participate in gradient updates.

[0064] The Fisher information matrix is ​​updated online using an exponential moving average strategy. Where η=0.9 is the decay coefficient. After each two-layer adaptive update, the expected value of the squared gradient is calculated using the latest data and then exponentially averaged with the historical values ​​to smooth out historical fluctuations.

[0065] Meta-learning initialization enables rapid learning capabilities, while EWC regularization ensures that updates do not deviate too far from historical optimalities. The combination of these two features allows for both rapid adaptation to new operating conditions and effective retention of old knowledge.

[0066] Furthermore, this embodiment also includes: an uncertainty interval feedback closed loop, whereby the updated quantile neural network prediction model is immediately put into online inference, and its output quantile prediction and uplink / downlink asymmetric uncertainty are updated accordingly, directly affecting the determination of the power constraint boundary corresponding to the energy storage system.

[0067] Improved prediction accuracy after model adaptation leads to increased uplink uncertainty. and downside uncertainty As the MPC constraint decreases, the energy storage capacity utilization rate improves; however, the prediction error increases during sudden changes in operating conditions, leading to uplink uncertainty. and downside uncertainty As the number of learners increases, the MPC constraints automatically tighten, and the safety margin is reserved. The feedback path enables meta-learning adaptation, EWC anti-forgetting, and MPC optimization to form an organic closed loop.

[0068] Furthermore, this embodiment also includes a triggering strategy for the adaptive learning mechanism, employing a dual-trigger strategy to initiate online adaptive updates, including: Timed triggering, i.e., performing meta-learning rapid adaptation every 1 hour, only the first layer, no more than 3 steps; Abnormal triggering means that when the real-time RMSE exceeds a preset threshold, such as 2.5MW, a complete two-layer update is immediately triggered, consisting of the first layer and the second layer, in no more than 8 steps. Emergency expansion means that if the RMSE does not decrease in the next cycle after an anomaly is triggered, the data window is expanded to 72 hours for reassessment.

[0069] In this embodiment, the meta-learning and EWC two-layer adaptive learning mechanism endows the model with rapid adaptation capabilities (≤3 steps, ≤2 minutes), while EWC regularization protects historical knowledge; the adaptation time for operating condition drift is shortened from hours in traditional methods to minutes; and the uplink uncertainty after update is reduced. and downside uncertainty The power constraint boundary corresponding to the feedback energy storage system forms an organic closed loop of prediction, control, and learning.

[0070] Taking a 660MW coal-fired power unit as an example, the method of this embodiment is verified through a combination of digital simulation and field measurement. Verification environment: The simulation platform uses MATLAB / Simulink R2023b; the unit model is a dynamic model of a 660MW supercritical coal-fired power unit; the power grid model is an extended IEEE 14-bus system; and the AGC signal uses an actual AGC command sequence from a provincial dispatch center (72 hours). The hardware platform includes a multi-core industrial-grade server (CPU clock speed not less than 3.0GHz, memory not less than 128GB), a GPU with deep learning acceleration capabilities (video memory not less than 24GB), and a real-time power system simulation platform.

[0071] Four comparison schemes were set up: Scheme A was a pure thermal power unit (without energy storage); Scheme B was a single lithium battery energy storage (15MW) combined with PI control; Scheme C was a hybrid energy storage combined with traditional MPC (single fixed model); and Scheme D was the method of this embodiment. Table 1 shows the core performance indicators of the four comparison schemes. Table 1 Core performance indicators of the four schemes Typical scenario verification: In the normal frequency regulation scenario (accounting for approximately 65%), the K value of Scheme D is 0.945, the ACE elimination time is 38s, and the multi-condition soft switching MPC mainly operates in the normal frequency regulation scenario, with a scenario membership degree of approximately 0.85. In the deep regulation scenario (accounting for approximately 15%), the K value of Scheme D is 0.938, and the ACE elimination time is 42s. When the SOC is below 20%, the discharge power limit is automatically tightened by approximately 15%; when the SOC is above 80%, the charging power limit is automatically tightened by approximately 12%. There were 0 instances of SOC exceeding the limit during the test period. The multi-condition soft switching MPC smoothly transitions to the deep regulation scenario, with the scenario membership degree increasing from 0 to 0.9 in approximately 3s. In the ramp-up transition scenario (accounting for approximately 12%), the ramp-up transition power satisfaction rate of Scheme D is 97.5%, and the high-frequency response of the supercapacitor (not exceeding 200ms) effectively compensates for the ramp-up delay. The penalty weight for battery power change rate is increased by 40%, and the number of daily high-frequency charge and discharge cycles is reduced by about 38% compared to Scheme C. Under AGC standby conditions (accounting for about 8%), the standby rate of the energy storage system in Scheme D is about 92%, and the number of invalid responses does not exceed 2 times / day, which is significantly better than the approximately 15 times / day in Scheme B.

[0072] Online adaptive learning validation: Simulating a sudden change in grid dispatch strategy (a new AGC regulation mode is introduced in the 48th hour). In Scheme C, the K value plummeted from 0.915 to 0.82 after the change and continued to deteriorate, requiring manual intervention to retrain the model. In Scheme D, the K value only slightly decreased to 0.935 after the change, and through meta-learning, it quickly adapted, recovering to 0.94 within 2 hours and stabilizing at 0.945 after 5 hours. After the model update, the average uncertainty σ decreased by approximately 25%, corresponding to automatic relaxation of MPC constraints, and an improvement in energy storage capacity utilization of approximately 4.5%.

[0073] Long-term operational stability verification: Simulation operation tests were conducted for six consecutive months, covering load characteristic changes from spring to summer, accumulating approximately 1.6 × 10⁻⁶ days. 7 One control cycle. During the test, the average K value of scheme D remained stable at 0.944, the average RMSE was 1.79 MW, and the average uncertainty σ was 2.06 MW. Meta-learning was triggered 4320 times, and EWC full update was triggered 56 times. The system achieved fully automated long-term operation without manual intervention.

[0074] Economic benefit analysis: Taking a 660MW unit and the Fujian Province AGC frequency regulation auxiliary service market (mileage compensation of RMB10 / MWh) as an example, the annual AGC frequency regulation revenue of Scheme D is about RMB15.2 million, the battery replacement cost over 10 years is about RMB1.5 million, the operation and maintenance cost over 10 years is about RMB1.2 million, the net profit over 10 years is about RMB146 million, and the static payback period is about 3.8 years.

[0075] It should be understood that the specific embodiments described above are merely illustrative or explanatory of the principles of the invention and do not constitute a limitation thereof. Therefore, any modifications, equivalent substitutions, improvements, etc., made without departing from the spirit and scope of the invention should be included within the protection scope of the invention. Furthermore, the appended claims are intended to cover all variations and modifications falling within the scope and boundaries of the appended claims, or equivalent forms of such scope and boundaries.

[0076] The present invention has been described above with reference to embodiments thereof. However, these embodiments are merely illustrative and not intended to limit the scope of the invention. The scope of the invention is defined by the appended claims and their equivalents. Various substitutions and modifications can be made by those skilled in the art without departing from the scope of the invention, and all such substitutions and modifications should fall within the scope of the invention.

[0077] Although embodiments of the present invention have been described in detail, it should be understood that various changes, substitutions, and modifications can be made to the embodiments of the present invention without departing from the spirit and scope of the invention.

[0078] Obviously, the above embodiments are merely illustrative examples for clear explanation and are not intended to limit the implementation. Those skilled in the art will recognize that other variations or modifications can be made based on the above description. It is neither necessary nor possible to exhaustively list all possible implementations here. However, obvious variations or modifications derived therefrom are still within the scope of protection of this invention.

Claims

1. A hybrid energy storage-assisted secondary frequency regulation method, characterized in that, include: Based on historical operating data, a multi-dimensional time-series feature vector is constructed. The historical operating data includes power grid operating data, traditional generating unit operating data, energy storage operating data, and environmental data. Based on the multidimensional time-series feature vector, a trained quantile neural network prediction model is used to predict the AGC compensation power demand and its uncertainty range. Based on the AGC compensation power requirement and its uncertainty range, the output power of the energy storage system is obtained by using a preset switching MPC optimization strategy. The output power of the energy storage system is decomposed in the frequency domain using a preset VMD algorithm to obtain high-frequency power components and low-frequency power components. Based on the high-frequency power component and the low-frequency power component, power is allocated to the hybrid energy storage, generating a hybrid energy storage response AGC command to assist in secondary frequency regulation.

2. The hybrid energy storage-assisted secondary frequency regulation method according to claim 1, characterized in that, The hybrid energy storage includes energy storage batteries and supercapacitors; The power grid operation data includes: power grid frequency deviation, tie-line power deviation, area control deviation, and AGC dispatch instructions; The conventional unit operating data includes: the actual load power output of the conventional unit; The energy storage operation data includes: state of charge, maximum rated charging power, and maximum rated discharging power. The environmental data includes: ambient temperature.

3. The hybrid energy storage-assisted secondary frequency regulation method according to claim 2, characterized in that, The quantile neural network prediction model includes: an input layer, an LSTM unit, a TCN unit, a feature fusion layer, and a quantile output layer. The input layer is connected to the LSTM unit and the TCN unit respectively. The LSTM unit and the TCN unit are arranged in parallel. The LSTM unit and the TCN unit are connected to the feature fusion layer respectively. The feature fusion layer is connected to the quantile output layer. The method of predicting the AGC compensation power demand and its uncertainty range for future time series based on the multidimensional time series feature vector using a trained quantile neural network prediction model includes: Based on the multidimensional temporal feature vector, input layer, LSTM unit and TCN unit, the LSTM unit output and TCN unit output are obtained; Based on the LSTM unit output, TCN unit output, and feature fusion layer, comprehensive temporal features are obtained. Based on the comprehensive time series features and the quantile output layer, the AGC compensation power requirements of multiple preset quantiles in future time series are predicted. Based on the AGC compensation power requirements of multiple preset quantiles, the corresponding uncertainty range is obtained.

4. The hybrid energy storage-assisted secondary frequency regulation method according to claim 3, characterized in that, The step of obtaining the energy storage system output power based on the AGC compensation power demand and its uncertainty range, using a preset switching MPC optimization strategy, includes: The AGC compensation power requirement and its uncertainty range are divided according to a preset optimization cycle to obtain the AGC operation optimization sequence; For each optimization cycle, based on the preset membership algorithm, the membership degree is calculated for the corresponding feature vector and typical scenario in the AGC running optimization sequence to obtain the membership degree of each scenario. Based on the membership degree of each scenario, determine the multi-scenario adaptive AGC model; Based on the membership degree of each scenario, the cost functions corresponding to different scenarios are weighted to construct a multi-scenario optimization objective function; Based on the multi-scenario adaptive AGC model, the objective function for scenario optimization is solved to obtain the output power of the first energy storage system; Based on the uncertainty range and energy storage operation data, determine the power constraint boundary corresponding to the energy storage system; The output power of the energy storage system is obtained based on the output power of the first energy storage system and the power constraint boundary.

5. The hybrid energy storage-assisted secondary frequency regulation method according to claim 4, characterized in that, The preset membership algorithm is expressed by the following formula: in, Let represent the membership degree of the k-th scenario in the t-th optimization cycle. This represents the eigenvector of the t-th optimization cycle; This represents the cluster center of the k-th scene; This represents the cluster center of the j-th scene, where j is different from k. This represents the temperature coefficient.

6. The hybrid energy storage-assisted secondary frequency regulation method according to claim 4, characterized in that, The multi-scenario optimization objective function is expressed by the following formula: in, Let Jk represent the membership degree of the k-th scene in the t-th optimization cycle, Jk represent the cost function corresponding to scene k, and u represent the input sequence to be optimized. Let represent the estimated model parameters for the k-th scene at time t, where i represents the step index in the prediction time domain, N represents the length of the prediction time domain, N-1 represents the first N-1 time steps in the prediction time domain, y(t+i|t) represents the prediction of the system output at time t+i at time t, yref represents the setpoint of the controlled variable, and Qk represents the output bias penalty matrix for scene k. Rk represents the change in control input, Rk represents the penalty matrix for control increment in scenario k, y(t+N) represents the prediction time domain endpoint, and Pk represents the terminal deviation penalty matrix in scenario k.

7. The hybrid energy storage-assisted secondary frequency regulation method according to claim 4, characterized in that, The step of determining the power constraint boundary corresponding to the energy storage system based on the uncertainty interval and energy storage operation data includes: Based on the uncertainty range, extract the corresponding uplink uncertainty and downlink uncertainty; When the state of charge is less than the preset lower threshold, the upper limit of the rated discharge power is constrained and tightened according to the downlink uncertainty to determine the discharge power constraint boundary. When the state of charge is greater than the preset upper limit threshold, the upper limit of the rated charging power is constrained and tightened according to the uplink uncertainty to determine the charging power constraint boundary. When the state of charge is between a preset lower threshold and a preset upper threshold, the charging and discharging power constraint is processed according to the uplink uncertainty and downlink uncertainty to determine the charging and discharging power constraint boundary; The discharge power constraint boundary, charging power constraint boundary, and charge / discharge power constraint boundary are the power constraint boundaries corresponding to the energy storage system.

8. The hybrid energy storage-assisted secondary frequency regulation method according to claim 4, characterized in that, The step of using a preset VMD algorithm to perform frequency domain decomposition on the output power of the energy storage system to obtain high-frequency power components and low-frequency power components includes: The output power of the energy storage system is preprocessed to obtain the power signal to be decomposed; The number of decomposition layers of the preset VMD algorithm is determined based on the decomposition information entropy index and the decomposition energy ratio index. The original cutoff frequency is adjusted based on the scenario with the highest membership degree and energy storage operation data to obtain the preset cutoff frequency of the VMD algorithm. Based on the number of decomposition layers and the cutoff frequency, the power signal to be decomposed is decomposed into high-frequency power components and low-frequency power components.

9. The hybrid energy storage-assisted secondary frequency regulation method according to claim 2, characterized in that, The process of allocating power to the hybrid energy storage based on the high-frequency and low-frequency power components, generating a hybrid energy storage response AGC command, and assisting in secondary frequency regulation includes: The high-frequency power component is allocated to the supercapacitor, and the low-frequency power component is allocated to the energy storage battery. When the state of charge of any energy storage unit reaches a preset upper limit threshold or a preset lower limit threshold, the equalization control mode is triggered to obtain the power allocation result. The power allocation result is then subjected to multi-constraint correction to generate a corrected hybrid energy storage response AGC command. Secondary frequency regulation is assisted based on the modified hybrid energy storage response AGC command.

10. The hybrid energy storage-assisted secondary frequency regulation method according to claim 1, characterized in that, The method further includes: training the quantile neural network prediction model and updating the switching MPC optimization strategy using an adaptive learning mechanism; The adaptive learning mechanism includes: A scrolling data window is established based on historical and real-time operational data; Based on the scrolling data window, error calculation is performed to determine whether an update should be triggered; If the triggering condition is not met, return to continue error calculation; If the triggering conditions are met, the switching MPC optimization strategy is quickly updated based on meta-learning of the data in the scrolling data window. During the update process, EWC regularization constraints are introduced to restrict the updates of key parameters based on the Fisher information matrix.

Images