An ultra-capacitive energy storage coupled thermal power unit frequency modulation instruction prediction method and system

Abstract

The embodiment of the disclosure provides a kind of super-capacity energy storage coupling thermal power unit frequency modulation instruction prediction method and system, belong to power system frequency modulation technical field.The method comprises: the topology analysis of super-capacity energy storage, thermal power and coupling control subsystem, determine collaborative division mechanism, and analyze instruction multi-scale characteristics;Layered acquisition system parameters and construct standardization matrix;The original instruction data is preprocessed;Based on multi-scale characteristics, the improved variational mode decomposition model is used to decompose instruction sequence by introducing slack variable, to obtain multiple intrinsic mode function subsequences;Based on multi-scale entropy and dynamic time warping, the similarity of subsequence is calculated, and the integrated component is generated by energy proportion weighted fusion;Each integrated component is input into the improved adaptive feature gated recurrent unit network for parallel prediction;The final prediction value is obtained by anti-standardization superposition to the prediction result, and multidimensional evaluation is carried out.The embodiment of the disclosure realizes high-precision, high-engineering adaptability frequency modulation instruction prediction.

Description

Technical Field

[0001] The embodiments disclosed herein belong to the field of power system frequency regulation technology, specifically relating to a method and system for predicting frequency regulation commands of supercapacitor energy storage coupled thermal power units. Background Technology

[0002] As the penetration rate of new energy power generation such as wind power and photovoltaics continues to exceed 50%, the inertia level of regional power grids has significantly decreased, and the sensitivity of frequency to disturbances such as load changes and the ramp-up of new energy output has greatly increased. This places stringent demands on the synergistic capability of frequency regulation services, which combines "high-frequency rapid response + low-frequency continuous support." With its advantages of millisecond-level response speed and ±1% regulation accuracy, supercapacitor energy storage (such as supercapacitors and lithium-ion battery energy storage) forms a coupled frequency regulation system with thermal power units (traditional frequency regulation resources with large capacity and strong continuous regulation capabilities). Through a synergistic model of "energy storage smoothing high-frequency fluctuations + thermal power tracking low-frequency trends," it has become the core solution to address the problem of power grid frequency fluctuations.

[0003] Frequency regulation command prediction is the "decision core" of a coupled system, and its prediction accuracy directly determines the overall system's regulation efficiency and economic benefits. According to industry empirical data, a 1% reduction in prediction error can increase the overall benefits of a combined thermal power storage and frequency regulation system by 2.3%. However, existing prediction methods have key flaws and are difficult to meet actual engineering needs. Current technologies often employ a loose architecture of simple decomposition and single-model prediction, failing to form a complete technology chain of topology analysis, data processing, feature modeling, and prediction evaluation. Furthermore, the core formula design does not consider the dynamic characteristics of the coupled system, resulting in prediction accuracy (MAPE generally exceeding 7%) and engineering adaptability that cannot meet the high-precision requirements of coupled frequency regulation of supercapacity energy storage and thermal power units. Therefore, it is urgent to overcome the limitations of existing technologies. Summary of the Invention

[0004] The embodiments disclosed herein aim to at least solve one of the technical problems existing in the prior art, and provide a method and system for predicting frequency regulation commands of supercapacity energy storage coupled thermal power units.

[0005] One aspect of this disclosure provides a method for predicting frequency regulation commands of a supercapacitive energy storage coupled thermal power unit, the method comprising: Topology analysis was performed on the supercapacity energy storage subsystem, the thermal power frequency regulation subsystem, and the coupled control subsystem to determine the collaborative division of labor mechanism, and multi-scale characteristics were analyzed based on historical frequency regulation command data. Parameters are collected at the component level, subsystem level, and system level in a hierarchical manner to construct a standardized parameter matrix; The raw frequency modulation command data is subjected to outlier handling, noise filtering, and standardization preprocessing to obtain a standardized command sequence. Based on the parameters set according to the multi-scale features, the standardized instruction sequence is decomposed using an improved variational mode decomposition model that introduces slack variables to obtain multiple intrinsic mode function subsequences. The similarity between the intrinsic mode function subsequences is calculated based on multi-scale entropy and dynamic time warping. Similar subsequence pairs are selected and fused according to energy ratio to generate multiple comprehensive components. Each of the aforementioned integrated components is input into multiple improved adaptive feature-gated recurrent unit networks for parallel prediction. The instantaneous rate of change of the input sequence is introduced into the reset gate of the improved adaptive feature-gated recurrent unit network, and an adaptive adjustment factor is embedded in the update gate. The prediction results of each integrated component are denormalized and superimposed to obtain the predicted value of the frequency modulation command. The prediction is then evaluated in multiple dimensions based on the prediction error index, response adaptability index, and engineering practicality index.

[0006] Furthermore, the multi-scale features include high-frequency random fluctuations at the second level and low-frequency trend changes at the minute level; The collaborative division of labor mechanism is as follows: frequency regulation requirements for energy storage response frequencies higher than or equal to 1Hz; frequency regulation requirements for thermal power tracking frequencies lower than or equal to 0.01Hz.

[0007] Furthermore, the constrained variational model of the improved variational mode decomposition model is expressed as follows:

[0008] In the formula, For the first A sequence of intrinsic mode functions For the first The center frequencies of the subsequence of eigenmode functions For the Dirac function, The imaginary unit, Pi For convolution operators, It is the L2 norm. As slack variables, It is a standardized instruction sequence.

[0009] Furthermore, the energy-weighted fusion generates multiple composite components, as shown in the following formula:

[0010]

[0011] In the formula, For the first Each comprehensive component in The value at time, Frequency band identifier for the composite components, , These are the intrinsic mode function subsequences. and The fusion weight, , These are the intrinsic mode function subsequences. and Energy.

[0012] Furthermore, the output vector of the reset gate of the improved adaptive feature-gated recurrent unit network is shown in the following equation:

[0013]

[0014] In the formula, for The gate's output vector is reset at all times. It is the Sigmoid activation function. To reset the weight matrix of the gate, for The input feature vector at time t, for The hidden state vector at time step 1. To reset the gate's bias vector, for The instantaneous rate of change of the input sequence at time t. The weight vector is the instantaneous rate of change. for The input feature vector at time t, The dimension of the input feature vector.

[0015] Furthermore, the output vector of the update gate of the improved adaptive feature-gated recurrent unit network is shown in the following equation:

[0016]

[0017] In the formula, for Update the gate's output vector continuously. It is the Sigmoid activation function. To update the gate weight matrix, for The input feature vector at time t, for The hidden state vector at time step 1. To update the bias vector of the gate, for The adaptive adjustment factor at time. The weight vector of the adaptive adjustment factor. To update the initial output of the gate, It is a function with maximum value. for The input feature vector at time t, The dimension of the input feature vector.

[0018] Furthermore, the response adaptability indicators include high-frequency response deviation rate and low-frequency lag time; the engineering practicality indicators include computational efficiency and robustness coefficient.

[0019] Another aspect of this disclosure provides a frequency regulation command prediction system for supercapacitive energy storage coupled thermal power units, the system comprising: The feature analysis module is used to perform topology analysis on the supercapacity energy storage subsystem, thermal power frequency regulation subsystem and coupled control subsystem, determine the collaborative division of labor mechanism, and analyze multi-scale features based on historical frequency regulation command data; The parameter acquisition module is used to acquire parameters at the component level, subsystem level, and system level in a hierarchical manner, and to construct a standardized parameter matrix. The preprocessing module is used to perform outlier handling, noise filtering, and standardization preprocessing on the raw frequency modulation command data to obtain a standardized command sequence. The mode decomposition module is used to set parameters based on the multi-scale features and decompose the standardized instruction sequence using an improved variational mode decomposition model that introduces relaxation variables to obtain multiple intrinsic mode function subsequences. The weighted fusion module is used to calculate the similarity between the intrinsic mode function subsequences based on multi-scale entropy and dynamic time warping, filter similar subsequence pairs and fuse them in a weighted manner according to energy ratio to generate multiple comprehensive components. The parallel prediction module is used to input each of the comprehensive components into multiple improved adaptive feature-gated recurrent unit networks for parallel prediction. The instantaneous rate of change of the input sequence is introduced into the reset gate of the improved adaptive feature-gated recurrent unit network, and an adaptive adjustment factor is embedded in the update gate. The multidimensional evaluation module is used to de-standardize and superimpose the prediction results of each integrated component to obtain the frequency modulation command prediction value, and to perform multidimensional evaluation based on prediction error index, response adaptability index and engineering practicality index.

[0020] Another aspect of this disclosure provides an electronic device, comprising: At least one processor; and a memory communicatively connected to the at least one processor for storing one or more programs that, when executed by the at least one processor, enable the at least one processor to implement the above-described method for predicting frequency regulation commands for supercapacity energy storage coupled thermal power units.

[0021] Another aspect of this disclosure provides a computer-readable storage medium storing a computer program that, when executed by a processor, implements the above-described method for predicting frequency regulation commands for supercapacity energy storage coupled thermal power units.

[0022] This disclosure discloses a method and system for predicting frequency regulation commands of supercapacitor energy storage coupled with thermal power units. Starting from the topology analysis of the coupled system, it clarifies the collaborative division of labor mechanism of "energy storage responding to high frequencies and thermal power tracking low frequencies." Based on this, it constructs a complete technology chain including hierarchical data acquisition, data preprocessing, improved VMD decomposition, MSEDAF fusion, improved AFGRU parallel prediction, and multi-dimensional evaluation. This achieves accurate capture and separation prediction of high-frequency mutations and low-frequency trends in frequency regulation commands, significantly reducing prediction errors and meeting ultra-high precision engineering requirements. Furthermore, by introducing dedicated evaluation indicators such as response adaptability, it ensures that the prediction results can directly and effectively support the actual coordinated regulation of energy storage and thermal power units, fundamentally solving the problems of low prediction accuracy and poor engineering adaptability. This provides highly reliable and practical decision support for supercapacitor energy storage-thermal power unit coupled frequency regulation. Attached Figure Description

[0023] Figure 1 This is a flowchart illustrating a method for predicting frequency regulation commands for a supercapacity energy storage coupled thermal power unit according to an embodiment of the present disclosure. Figure 2 This is a schematic diagram of the structure of a frequency regulation command prediction system for a supercapacity energy storage coupled thermal power unit according to another embodiment of the present disclosure; Figure 3 This is a schematic diagram of the structure of an electronic device according to another embodiment of the present disclosure. Detailed Implementation

[0024] The technical solutions of the embodiments of this disclosure will be clearly and completely described below with reference to the accompanying drawings. Obviously, the described embodiments are only a part of the embodiments of this disclosure, and not all of them. Based on the embodiments of this disclosure, all other embodiments obtained by those of ordinary skill in the art without creative effort are within the scope of protection of this disclosure.

[0025] Furthermore, the described features, structures, or characteristics can be combined in any suitable manner in one or more embodiments. Numerous specific details are provided in the following description to give a thorough understanding of embodiments of this disclosure. However, those skilled in the art will recognize that the technical solutions of this disclosure can be practiced without one or more of the specific details, or other methods, components, apparatuses, steps, etc., can be employed. In other instances, well-known methods, apparatuses, implementations, or operations are not shown or described in detail to avoid obscuring various aspects of this disclosure.

[0026] The flowcharts shown in the accompanying drawings are merely illustrative and do not necessarily include all content and operations / steps, nor do they necessarily have to be performed in the described order. For example, some operations / steps can be broken down, while others can be combined or partially combined; therefore, the actual execution order may change depending on the specific circumstances.

[0027] It should be understood that although the terms first, second, third, etc., may be used in this disclosure to describe various components, these components should not be limited by these terms. These terms are used to distinguish one component from another. Therefore, the first component discussed below may be referred to as the second component without departing from the teachings of this disclosure. As used in this disclosure, the term "and / or" includes all combinations of any one and more of the associated listed items.

[0028] Those skilled in the art will understand that the accompanying drawings are merely schematic diagrams of exemplary embodiments, and the modules or processes in the drawings are not necessarily necessary for implementing this disclosure, and therefore cannot be used to limit the scope of protection of this disclosure.

[0029] like Figure 1 As shown, one embodiment of this disclosure provides a method for predicting frequency regulation commands of supercapacity energy storage coupled thermal power units, the method comprising: Step S1: Perform topology analysis on the supercapacity energy storage subsystem, thermal power frequency regulation subsystem, and coupled control subsystem to determine the collaborative division of labor mechanism, and analyze multi-scale characteristics based on historical frequency regulation command data.

[0030] Specifically, a topology analysis is first performed to clarify the composition and connection relationships of the supercapacity energy storage subsystem, the thermal power frequency regulation subsystem, and the coupled control subsystem. The supercapacity energy storage subsystem includes energy storage units, a bidirectional PCS (Power Conversion System), and a BMS (Battery Management System), with its core function being rapid response to high-frequency frequency regulation demands. The thermal power frequency regulation subsystem includes thermal power units, a DEH (Digital Electric Hydraulic) speed governor, and a CCS (Coordinated Control System), with its core function being continuous support for low-frequency frequency regulation demands. The coupled control subsystem includes a command receiver, a power distributor, and a synchronization device, with its core function being the precise allocation of frequency regulation commands and the coordinated operation of the two devices. Based on the functions of each subsystem, a collaborative division of labor mechanism is determined: "energy storage responds to high frequencies (≤1Hz), and thermal power tracks low frequencies (≥0.01Hz)."

[0031] Next, command characteristic analysis was performed. Historical FM command data from the past year was collected (0.1s sampling, data volume not less than 1.5 × 10⁻⁶).6 (Points), through statistical analysis, identify the multi-scale characteristics of the commands: second-level high-frequency random fluctuations (caused by sudden load changes, amplitude range ±50kW) and minute-level low-frequency trend changes (caused by the ramp-up of renewable energy output, amplitude range ±200kW). Combined with the regulation capability of the coupled system, define the prediction accuracy indicators (MAE≤0.1kW, MAPE≤1%) and time scale (prediction step size 2s, i.e., predicting the frequency regulation command for the next 2 seconds based on historical 20s data).

[0032] The collaborative division of labor mechanism output in this step directly determines the data acquisition range of step S2 (which needs to cover the regulation characteristic parameters of energy storage / thermal power), the decomposition scale of step S4 (dividing the number of modes according to high frequency, medium frequency, and low frequency), and the model improvement direction of step S6 (designing differentiated gating mechanisms for high frequency and low frequency characteristics), providing a foundation for the design of the entire technical solution.

[0033] Step S2: Collect parameters at the component level, subsystem level, and system level in a hierarchical manner to construct a standardized parameter matrix.

[0034] Specifically, the first step is to conduct tiered data collection, systematically collecting various types of data to support the prediction model, ensuring that the data is comprehensive and traceable. This includes: At the component level, the focus is on the core component parameters of each subsystem, including the regulation rate of the energy storage unit. (Unit: MW / s), Cycle life (unit: cycles), Response delay of thermal power DEH speed governor (Unit: s), adjustment dead zone (unit: Hz), synchronization accuracy of the synchronization device in the coupling control subsystem (unit: Hz), etc., the data can be obtained from the equipment's factory technical documents; At the subsystem level, these parameters reflect the overall performance of each subsystem, including the rated power of the energy storage subsystem. (Unit: MW) Rated Capacity (Unit: MWh), SOC upper and lower limits (unit: %), maximum adjustment range of fire-electrode system (Unit: MW), Adjustment coefficient (Unit: %), data can be sourced from on-site operation and maintenance records (statistical period not less than 1 year); At the system level, parameters reflecting the cooperative characteristics of the coupled system include power allocation coefficients. (i.e., the ratio of energy storage to the rated frequency regulation power of thermal power), grid frequency response coefficient (Unit: MW / Hz), Rated Frequency (Unit: Hz), data can be sourced from the operation statistics of the power grid dispatch center.

[0035] Then, a parameter matrix is ​​established according to the standardized format of "hierarchy-subsystem-parameter name-value-unit-data source-collection time" to ensure that all data is traceable and unambiguous.

[0036] The parameter matrix established in this step provides the core data input for all subsequent steps. The preprocessing in step S3 requires calling the original frequency modulation command data, the VMD decomposition in step S4 requires determining the decomposition parameters by referring to the frequency response coefficient, and the AFGRU model improvement in step S6 requires designing a gating mechanism based on the regulation characteristic parameters of energy storage / thermal power. The completeness and accuracy of the data directly determine the performance of the entire prediction method.

[0037] Step S3: Perform outlier processing, noise filtering, and standardization preprocessing on the original frequency modulation command data to obtain a standardized command sequence.

[0038] Specifically, outlier handling: Box plot method is used to identify outliers in the original frequency modulation command data. First, the lower quartiles of the data are calculated. Upper quartiles and quartiles It will exceed Data within the specified range is identified as outliers (such as sudden command jumps caused by sensor malfunctions or missing values ​​due to data transmission interruptions). To avoid outliers distorting the true characteristics of the commands, linear interpolation of data from adjacent time points is used for replacement; that is, for outliers... ,pass Calculate the replacement value (where (The time corresponding to the outlier).

[0039] Noise Filtering: For sensor noise in the 0.1s-level sampling data, Kalman filtering is used for denoising, which removes irrelevant noise while fully preserving the true fluctuation characteristics of the command. The state equation and observation equation of the Kalman filter are as follows:

[0040] In the formula, for The prior state estimate at time t represents the value based on Time-state pair Prediction of the command value at any given time; for The posterior state estimate at time t, i.e. The actual value of the instruction after filtering and correction at any given moment; This is the state transition matrix (it can be set to 1 here to indicate the temporal continuity of the instruction data). To control the input matrix (which can be set to 0.1 here to match the rate of change of the instruction). for Control input at any time (can be 0 here, focusing only on the timing changes of the instruction itself); for The observed value at a given time, i.e., the raw command data collected by the sensor; This is the observation matrix (it can be set to 1 here, indicating that the observed values ​​are directly related to the state values). The observed noise (follows a Gaussian distribution with a mean of 0 and a variance of 0.01).

[0041] Standardization preprocessing: To eliminate the differences in units between different data (e.g., the unit of frequency modulation commands is kW, while the unit of energy storage power is MW), and to ensure the fairness of subsequent similarity calculations and model training, the denoised command data is normalized to [specific unit]. The interval, the formula is as follows:

[0042] in, for The frequency modulation command value after time standardization (dimensionless). for The original frequency modulation command value at the time (unit: kW); The minimum value in the historical instruction data (unit: kW); This is the maximum value (unit: kW) in the historical instruction data.

[0043] The normalized sequence output in this step It is the only input for step S4 to improve VMD (Variational Mode Decomposition). The preprocessed high signal-to-noise ratio data can significantly improve the decomposition accuracy, reduce the information redundancy of subsequences, and lay a high-quality data foundation for subsequent fusion and prediction.

[0044] Step S4: Based on the multi-scale features, set parameters and use an improved variational mode decomposition model that introduces slack variables to decompose the standardized instruction sequence to obtain multiple intrinsic mode function subsequences.

[0045] Specifically, based on the instruction characteristic analysis (high-frequency, mid-frequency, and low-frequency characteristics) in step S1, the core parameter for VMD decomposition is set as: the number of modes. =6 (corresponding to 2 high-frequency, 2 mid-frequency, and 2 low-frequency features, ensuring that features at each scale can be fully decomposed), secondary penalty factor. =2500 (Balance between decomposition accuracy and computational efficiency), iterative convergence threshold (Ensure the stability of the decomposition results).

[0046] To address the issue of overlapping subsequence features in traditional VMD decomposition, slack variables are introduced. To reduce constraint stiffness, an improved constraint variational model is constructed to ensure that each IMF (Intrinsic Mode Function) subsequence after decomposition is feature-independent and without redundancy. The model expression is shown below:

[0047] in, For the first A sequence of intrinsic mode functions ( =1,2,…,6), each subsequence corresponds to a single scale of instruction features; For the first The center frequencies (in Hz) of each intrinsic mode function subsequence are used to distinguish features at different scales. This is the Dirac function, used to characterize the impulse response; The imaginary unit ( = 1); Pi (value 3.1416). This is the convolution operator, used to implement signal filtering. This is the time derivative operator, used to calculate the rate of change of a signal; It is a complex exponential function used to convert signals to the frequency domain; It is the L2 norm, used to calculate the bandwidth of the signal; These are slack variables (dimensionless) used to reduce the rigidity of constraints and avoid overlapping features of subsequences; This is the standardized instruction sequence output in step S3.

[0048] The modal components are then iteratively updated using the Alternating Direction Method of Multipliers (ADMM) algorithm. With center frequency The specific iterative logic is as follows: first, fix the center frequency. Update modal components To minimize the objective function Then fix the modal components. Update center frequency To optimize the frequency domain distribution; repeat the above process until the modal components of two adjacent iterations are optimized. Change in quantity and center frequency The changes were all less than the convergence threshold. The iteration stops, and finally six feature-independent IMF subsequences are output.

[0049] The six IMF subsequences output in this step are the core inputs of MSEDAF (Multi-Scale Entropy-Dynamic Alignment Fusion) in step S5. The low-redundancy features after decomposition can significantly improve the fusion efficiency and quality. At the same time, the subsequences divided by scale provide clear modeling objects for multi-scale accurate prediction in step S6.

[0050] Step S5: Calculate the similarity between the intrinsic mode function subsequences based on multi-scale entropy and dynamic time warping, filter similar subsequence pairs and fuse them by weighting according to energy ratio to generate multiple comprehensive components.

[0051] Specifically, to accurately select IMF subsequences with similar features, a composite similarity index is constructed by combining two metrics: multi-scale entropy and dynamic time warping (DTW), to comprehensively evaluate the regularity and temporal phase matching degree of IMF subsequences: Multi-scale entropy measurement: This involves calculating the sample entropy (embedding dimension) of IMF subsequences at different scales. =2, similarity tolerance ,in (The standard deviation of the IMF subsequence) quantifies the regularity and similarity of the IMF subsequences. The closer the sample entropy is, the more similar the fluctuation patterns of the IMF subsequences are. DTW metric: By using dynamic programming to find the optimal matching path between two IMF subsequences, the sum of the distances along the path is calculated to quantify the temporal phase matching degree of the subsequences. The smaller the DTW distance, the more consistent the temporal change trend of the IMF subsequences.

[0052] Based on the combined results of the two metrics, subsequence pairs with a similarity of ≥0.7 were selected (the similarity threshold was obtained through extensive experimental verification to ensure that the feature overlap of the selected IMF subsequences was ≥80%).

[0053] To ensure that the fused composite component retains strong feature information, fusion weights are assigned based on the energy ratio of the subsequences; the higher the energy of the subsequence, the greater its contribution to the composite component. The fusion formula is as follows:

[0054]

[0055] In the formula, For the first Each comprehensive component in The value at time; =1,2,3 are the frequency band identifiers of the composite components (corresponding to the high-frequency, mid-frequency, and low-frequency composite components, respectively). , These are the selected similar IMF subsequence pairs; , They are respectively and The fusion weights (dimensionless) satisfy the following conditions: + =1; , They are respectively and Energy (unit: kW) 2 ),energy The calculation is shown in the following formula:

[0056] in, The length of the IMF subsequence (compared to the normalized instruction sequence) (The lengths are consistent).

[0057] Following the principle of "high frequency-high frequency", "medium frequency-medium frequency", and "low frequency-low frequency", the six IMF subsequences are fused sequentially: First, two high-frequency IMF subsequences are fused to obtain... (High-frequency composite components), then fused with two mid-frequency IMF subsequences to obtain (Intermediate frequency composite components), and finally fused with two low-frequency IMF subsequences to obtain (Low-frequency composite component) Finally, three low-redundancy, highly representative composite components are output.

[0058] This step uses MSEDAF to fuse the 6 IMF subsequences into 3 integrated components, which reduces the computational complexity of the prediction model in step S6 (reducing the input dimension by 50%), and retains the core features at each scale through energy weighting, providing optimized input for subsequent multi-scale accurate prediction.

[0059] Step S6: Input each of the comprehensive components into multiple improved adaptive feature-gated recurrent unit networks for parallel prediction. The instantaneous rate of change of the input sequence is introduced into the reset gate of the improved adaptive feature-gated recurrent unit network, and an adaptive adjustment factor is embedded in the update gate.

[0060] Specifically, to address the scale differences among the three integrated components, an Adaptive Feature Gate Recurrent Unit (AFGRU) network is constructed, consisting of an input layer, a feature enhancement layer, an improved gating layer, and an output layer. The specific structural parameters are as follows: input layer dimension = 10 (corresponding to an input window length of 10, i.e., 20 seconds of historical data), the feature enhancement layer contains 128 ReLU activation units (used to extract more complex nonlinear features), the improved gating layer contains 64 hidden units (used to adapt to multi-scale temporal features), and output layer dimension = 1 (corresponding to a prediction step size of 1, i.e., the instruction value after 2 seconds).

[0061] To address the poor adaptability of the standard GRU to multi-scale features, the gating mechanism is improved to suit the different characteristics of high-frequency and low-frequency components. Specifically, this includes: 1. Improved reset gate (adapted to high-frequency mutation features): Introducing the instantaneous rate of change of the input sequence. This enhances the network's ability to perceive high-frequency abrupt changes in signals. When the value is large (indicating high-frequency mutations), the reset gate output approaches 0, forcing the network to forget historical redundant information and focus on the current mutation feature. The relevant formulas are shown below:

[0062]

[0063] in, for The output vector of the gate is reset at any time (dimension=64); The Sigmoid activation function (output range [0,1]); The weight matrix for the reset gate (dimension = 64 × (10 + 64)) is used to fuse input features and historical hidden states; for Input feature vector at time step (dimension = 10); for The hidden state vector at time step (dimension = 64); To reset the gate's bias vector (dimension = 64); for The instantaneous rate of change (dimensionless) of the input sequence at each time step is obtained by calculating the mean absolute difference between data at adjacent time steps within the input window; A weight vector (dimension = 64) representing the instantaneous rate of change, used for adjustment. The intensity of the impact on the reset door; for The input feature vector at time step; The dimension of the input feature vector ( =1,2,…,10).

[0064] 2. Update gate improvement (adapting to low-frequency trend features): embedding adaptive adjustment factor Dynamically adjust the memory update intensity of the network when When the value is small (indicating a low-frequency input trend), the update gate output approaches 0, allowing the network to retain more historical state information and improving its ability to track low-frequency trends. The relevant formulas are shown below:

[0065]

[0066]

[0067]

[0068] in, for Update the gate's output vector (dimension = 64) at all times. Use the Sigmoid activation function; To update the weight matrix of the gate (dimension = 64 × (10 + 64)); for Input feature vector at time step (dimension = 10); for The hidden state vector at time step (dimension = 64); To update the bias vector of the gate (dimension = 64); for The adaptive adjustment factor at time (dimensionless). The weight vector of the adaptive adjustment factor (dimension = 64). To update the initial output of the gate (not yet introduced) The calculation result at that time (dimension = 64). This is the maximum value function, used to extract the maximum change within the input window; for The input feature vector at time step; The dimension of the input feature vector ( =1,2,…,10); for Candidate hidden state vector at time step (dimension = 64); For hyperbolic tangent activation function (output range [ 1,1]); The weight matrix for the candidate hidden states (dimension = 64 × (10 + 64)). The bias vector for the candidate hidden state (dimension = 64); This is the element-wise multiplication operator; for The final hidden state vector at time step (dimension=64) integrates information from historical states and the current candidate state.

[0069] The three composite components ( , , Input three independent AFGRU branches respectively, and use the Adam optimizer (learning rate = 0.001, decay rate = ... e 5 Training is performed using mean squared error as the loss function. Training stops when the validation set loss does not decrease for 10 consecutive rounds. After training converges, each branch outputs the prediction result for the corresponding integrated component. , , .

[0070] This step, based on the comprehensive components of step S5, improves the accurate modeling of multi-scale features through a gating mechanism. The prediction results of each component output are the core input of the superposition and reconstruction in step S7, and are the key link to achieve high-precision prediction of the whole method.

[0071] Step S7: Perform denormalization on the prediction results of each integrated component and superimpose them to obtain the predicted value of the frequency modulation command, and conduct multi-dimensional evaluation based on prediction error index, response adaptability index and engineering practicality index.

[0072] Specifically, since the prediction result in step S6 is based on the standardized composite components, the prediction results of each component need to be de-standardized first to restore the original amplitude range of the command. Then, the final frequency modulation command prediction value is obtained by superposition, as shown in the formula:

[0073] in, for The predicted value of the final frequency modulation command at time (unit: kW); For the first Predicted values ​​of each composite component (dimensionless) =1,2,3); , These are the maximum and minimum values ​​(unit: kW) of the historical instruction data determined in step S3, used for anti-standardization calculation.

[0074] To comprehensively verify the overall performance of the frequency modulation command prediction method and ensure that it meets engineering requirements, three types of indicators are used for quantitative evaluation: 1. Prediction error indicators: These quantify the degree of deviation between predicted and actual values, including root mean square error (RMSE), average absolute error (MAE), and average relative percentage error (MAPE). The calculation logic is as follows: RMSE reflects the overall distribution of error, MAE reflects the average level of error, and MAPE reflects the relative proportion of error. The smaller all three indicators are, the better. 2. Response Adaptability Indicators: These assess the degree of matching between the predicted results and the regulatory division of labor in the coupled system, including the high-frequency response deviation rate. With low frequency lag time . The mean relative error between the predicted and actual values ​​of the high-frequency components (required to be ≤1%). The average lag time of the predicted value of the low-frequency component relative to the actual value (required to be ≤0.3s). Meeting the requirement indicates that the prediction results can adapt to the high-frequency response of energy storage and the low-frequency tracking requirements of thermal power. 3. Engineering Practicality Indicators: Verify the practical deployment feasibility of the method, including computational efficiency (unit: samples / second, requirement ≥1000 samples / second to ensure real-time prediction) and robustness coefficient. (Requirement ≥ 0.85). The robustness coefficient is calculated by assessing the changes in MAPE under scenarios such as adding ±5% noise, ±10% missing data, and parameter perturbations to the data. The closer the value is to 1, the stronger the method's anti-interference ability.

[0075] This step is the final verification of the preceding steps S1-S6. By superimposing and reconstructing, the final predicted value usable in engineering is obtained. The performance of the comprehensive method is verified through multi-dimensional evaluation. Simultaneously, the evaluation results can guide the parameter optimization of the preceding steps. For example, if the low-frequency lag time does not meet the requirements, the adaptive adjustment factor in step S6 can be adjusted. The weighting coefficients form a closed-loop improvement mechanism.

[0076] This disclosure discloses a method for predicting frequency regulation commands of supercapacitor energy storage coupled with thermal power units. It combines the collaborative frequency regulation division mechanism of supercapacitor energy storage and thermal power, adapts to the characteristics of the coupled topology, and establishes a prediction model highly matched to actual regulation scenarios, improving the engineering practicality of the prediction results. Through layered data acquisition, standardized preprocessing, and unified management, it eliminates data anomalies and noise interference, constructs a standardized data support system, and provides high-quality, highly reliable input data for the prediction model. It optimizes the decomposition-fusion strategy, eliminates sub-sequence redundancy, and improves the gating mechanism of the GRU network to enhance scale adaptability, achieving accurate capture of high-frequency abrupt signals and low-frequency trend signals, solving the multi-scale modeling problem. A comprehensive evaluation system is established, covering three dimensions: prediction error, response adaptability, and engineering practicality, ensuring that the prediction method not only has high prediction accuracy but can also directly support engineering deployment and actual operation. Finally, it provides a frequency regulation command prediction method with coordinated steps, excellent accuracy, and strong robustness, providing high-precision decision support for frequency regulation coupled with supercapacitor energy storage, and contributing to the improvement of grid frequency stability.

[0077] To verify the effectiveness, engineering feasibility, and universality of this disclosure, the following is a case study of an overcapacity energy storage-thermal power unit coupled frequency regulation system in a provincial regional power grid. This power grid has a renewable energy penetration rate of 50% (30% wind power + 20% photovoltaic). The specific implementation process is as follows: I. Parameters of the Empirical Object Power grid parameters: rated frequency =50Hz, frequency response coefficient =200MW / Hz, rated load 5000MW, frequency stability requirements are frequency deviation ≤±0.2Hz and convergence time ≤30s; Supercapacity energy storage subsystem: 200MW / 200MWh lithium battery energy storage power station, regulation rate =10MW / s, SOC range 20%~80%, charge and discharge efficiency 95%, includes 6 independent energy storage units (each with a capacity of 33.3MW / 33.3MWh), BMS sampling frequency 10Hz; Thermal power frequency regulation subsystem: 2 600MW thermal power units (total frequency regulation capacity 1200MW), regulation rate =3MW / s, DEH governor response delay =8s, adjustment coefficient =5%, CCS coordination system adjustment dead zone ±0.03Hz; Coupled control subsystem: power allocation coefficient 0.5 (energy storage and thermal power each bear 50% of the rated frequency regulation power), command transmission adopts IEC61850 communication protocol, command transmission delay ≤50ms, synchronization device synchronization accuracy ≤0.01Hz.

[0078] II. Scene Setting Three typical operating conditions were selected, covering common power grid operation scenarios, to comprehensively verify the universality and robustness of the method: Three typical scenarios were set up, covering different frequency disturbance intensities and renewable energy fluctuation levels (all of which are key scenarios affecting frequency stability), as shown in Table 1 below: Table 1 Classification of Power Grid Operation Scenarios

[0079] III. Implementation Steps Step 1: Topology and Instruction Characteristics Analysis The topology is clear: the supercapacity energy storage subsystem (6 lithium battery units + bidirectional PCS + BMS) is connected to the coupling control subsystem through the combiner cabinet, and the thermal power frequency regulation subsystem (2 600MW units + DEH + CCS) is connected to the coupling control subsystem through the coordination controller. After receiving the grid dispatch command, the coupling control subsystem allocates and regulates the power according to the division of labor of "energy storage response ≤ 1Hz high frequency, thermal power tracking ≥ 0.01Hz low frequency". Command characteristics: Frequency modulation command data collected from January 1, 2025 to December 31, 2025 (0.1s sampling, totaling 1.5768 × 10⁻⁶). 6 (Data points), statistical analysis shows that: high-frequency fluctuation amplitude range is ±48kW (average ±32kW), low-frequency trend change rate is ±18kW / min (average ±10kW / min), prediction step size is set at 2s, and the accuracy target is MAE≤0.1kW and MAPE≤1%.

[0080] Step 2: Data Acquisition and Matrix Establishment Layered data collection: Component level: Energy storage unit regulation rate =10MW / s (factory documentation), thermal power DEH response delay =8s (maintenance record), synchronization device accuracy 0.01Hz (factory documentation); Subsystem level: Rated energy storage power =200MW (Operation and Maintenance Record), Maximum Adjustment Range for Thermal Power Plants =300MW (O&M records), energy storage SOC upper limit 80% (factory documentation); System level: Power allocation factor 0.5 (scheduling data), frequency response factor =200MW / Hz (dispatch data), rated frequency =50Hz (scheduling data); Establish a standardized parameter matrix that covers all collected data to ensure data traceability and unambiguity.

[0081] Step 3: Data Preprocessing Outlier handling: Twelve outliers were identified using box plots (such as the command at 10:23:45 on March 15, 2025, which suddenly jumped to 800kW, exceeding the normal range). Linear interpolation was used to replace them, and the continuity of the data after replacement met the requirements. Noise filtering: Kalman filtering was applied to denoise the data. The observation noise variance was set to 0.01. After denoising, the signal-to-noise ratio of the data increased from 25dB to 38dB. The sudden change characteristics and trend characteristics of the instructions were not lost. Standardization: Historical instruction data =600kW =100kW, normalize the data to the [0,1] interval, and output the standardized sequence. .

[0082] Step 4: Improve VMD decomposition Parameter settings: =6、 =2500、 =10 6 ; Iterative decomposition: After 42 iterations using the ADMM algorithm, the convergence condition is met (IMF component change = 8.2 × 10⁻⁶). 7 < The change in center frequency = 1.5 × 10 8 < Stop iteration; Output: Of the 6 IMF subsequences, It is a high-frequency component (center frequency 0.2Hz~0.5Hz). This is the intermediate frequency component (center frequency 0.05Hz~0.2Hz). For low-frequency components (center frequency 0.01Hz~0.05Hz), through feature correlation analysis, the subsequence overlap is reduced to 15%, which is 25 percentage points lower than that of traditional VMD decomposition.

[0083] Step 5: MSEDAF fusion Similarity calculation: Calculated using a combination of multi-scale entropy and DTW. and The similarity is 0.804. and The similarity is 0.762. and The similarity is 0.738, which meets the screening threshold of ≥0.7. Energy calculation: energy , energy Calculate weights =0.545、 =0.455; Fusion generation: Fusion yields high-frequency composite components. Similarly, integration Obtaining intermediate frequency , Obtain low frequency The feature recognition of the three integrated components is improved by 40% compared with the original IMF subsequence.

[0084] Step 6: Improve AFGRU parallel prediction Sample construction: Input window length = 10 (20s of historical data), prediction step size = 1 (instruction after 2s), original data length of scenario 1 = 751 points, number of training set samples = 661 (88%), number of test set samples = 90 (12%). Network training: The Adam optimizer was used, with a batch size of 32. After 50 training rounds, the validation set loss decreased from the initial 0.85 to 0.02, satisfying the convergence condition. Parallel prediction: The prediction results of the three AFGRU branches are: high frequency Predicted MAE = 0.03kW, intermediate frequency Predicted MAE = 0.05kW, low frequency The predicted MAE is 0.04kW, and the prediction accuracy of each component meets the requirements of the sub-target.

[0085] Step 7: Overlay Reconstruction and Evaluation Superposition and reconstruction: The component prediction results are destandardized and then superimposed to obtain the final predicted value. ; Multi-dimensional evaluation results (taking scenario 2 as an example): Prediction errors: RMSE=0.04kW, MAE=0.07kW, MAPE=0.49%, all of which meet the preset accuracy targets (MAE≤0.1kW, MAPE≤1%). Response adaptability: High-frequency response deviation rate =0.6% (≤1%), low-frequency lag time =0.2s (≤0.3s), perfectly adapting to the adjustment division of labor in the coupled system; Engineering practicality: Computational efficiency = 1200 samples / second (≥1000 samples / second), robustness coefficient =0.92 (≥0.85), which meets the requirements for project deployment; Comparative verification: Compared with the standard GRU model, the MAE of this method is reduced by 90.48% and the MAPE is reduced by 96.91%; compared with the traditional VMD-GRU model, the MAE is reduced by 88.52% and the MAPE is reduced by 93.84%, which is significantly better than the existing technology.

[0086] Optimization and Validation: To verify the optimizability of the method, the slack variables in step 4 were adjusted. Iteration step size From 0.1 to 0.15, the decomposition-fusion-prediction process was re-executed, and the evaluation result for scenario 2 was: MAPE = 0.42%, robustness coefficient... =0.94, which is a further improvement over the previous method, verifying that the method can be adapted to the flexible needs of different power grid scenarios through parameter fine-tuning.

[0087] Conclusion: The method disclosed in this paper achieves high-precision prediction in three typical scenarios: conventional load fluctuations, load abrupt changes, and high renewable energy fluctuations. All performance indicators meet engineering requirements. The method not only possesses excellent prediction accuracy and response adaptability but also outstanding engineering practicality and robustness. It can be directly deployed in the pre-decision stage of an overcapacity energy storage-thermal power unit coupled frequency regulation system, providing precise support for energy storage charging and discharging strategies and thermal power unit regulation and scheduling. This effectively improves grid frequency stability and frequency regulation economy, and has broad engineering application prospects.

[0088] like Figure 2 As shown, another embodiment of this disclosure provides a frequency regulation command prediction system for supercapacity energy storage coupled to a thermal power unit, the system comprising: The feature analysis module 210 is used to perform topology analysis on the supercapacity energy storage subsystem, the thermal power frequency regulation subsystem and the coupled control subsystem, determine the collaborative division of labor mechanism, and analyze multi-scale features based on historical frequency regulation command data. The parameter acquisition module 220 is used to acquire parameters at the component level, subsystem level and system level in a hierarchical manner, and to construct a standardized parameter matrix. Preprocessing module 230 is used to perform outlier processing, noise filtering and standardization preprocessing on the raw frequency modulation command data to obtain a standardized command sequence; The mode decomposition module 240 is used to set parameters based on the multi-scale features and decompose the standardized instruction sequence using an improved variational mode decomposition model that introduces slack variables to obtain multiple intrinsic mode function subsequences. The weighted fusion module 250 is used to calculate the similarity between the intrinsic mode function subsequences based on multi-scale entropy and dynamic time warping, filter similar subsequence pairs and fuse them in a weighted manner according to energy ratio to generate multiple comprehensive components. The parallel prediction module 260 is used to input each of the comprehensive components into multiple improved adaptive feature-gated recurrent unit networks for parallel prediction. The instantaneous rate of change of the input sequence is introduced into the reset gate of the improved adaptive feature-gated recurrent unit network, and an adaptive adjustment factor is embedded in the update gate. The multidimensional evaluation module 270 is used to de-standardize and superimpose the prediction results of each comprehensive component to obtain the frequency modulation command prediction value, and to perform multidimensional evaluation based on prediction error index, response adaptability index and engineering practicality index.

[0089] Specifically, the frequency regulation command prediction system for supercapacity energy storage coupled thermal power units in this embodiment is used to implement the frequency regulation command prediction method for supercapacity energy storage coupled thermal power units described in the above embodiments. The specific implementation process has been described in detail in the above embodiments and will not be repeated here.

[0090] like Figure 3 As shown, another embodiment of this disclosure provides an electronic device, including: At least one processor 301; and a memory 302 communicatively connected to the at least one processor 301 for storing one or more programs that, when executed by the at least one processor 301, enable the at least one processor 301 to implement the above-described method for predicting frequency regulation commands for supercapacity energy storage coupled thermal power units.

[0091] The memory 302 and processor 301 are connected via a bus, which can include any number of interconnecting buses and bridges. The bus connects various circuits of one or more processors 301 and memory 302 together. The bus can also connect various other circuits, such as peripheral devices, voltage regulators, and power management circuits, which are well known in the art and therefore will not be described further herein. A bus interface provides an interface between the bus and the transceiver. The transceiver can be a single element or multiple elements, such as multiple receivers and transmitters, providing a unit for communicating with various other devices over a transmission medium. Data processed by processor 301 is transmitted over a wireless medium via an antenna, which further receives data and transmits it to processor 301.

[0092] Processor 301 is responsible for managing the bus and general processing, and can also provide various functions, including timing, peripheral interfaces, voltage regulation, power management, and other control functions. Memory 302 can be used to store data used by processor 301 during operation.

[0093] Another embodiment of this disclosure provides a computer-readable storage medium storing a computer program that, when executed by a processor, implements the above-described method for predicting frequency regulation commands for supercapacity energy storage coupled thermal power units.

[0094] The computer-readable storage medium may be included in the systems or electronic devices disclosed herein, or it may exist independently.

[0095] Computer-readable storage media can be any tangible medium that contains or stores a program, and can be an electrical, magnetic, optical, electromagnetic, infrared, or semiconductor system, apparatus, or device. More specific examples include, but are not limited to: electrical connections having one or more wires, portable computer disks, hard disks, optical fibers, random access memory (RAM), read-only memory (ROM), erasable programmable read-only memory (EPROM or flash memory), portable compact disk read-only memory (CD-ROM), optical storage devices, magnetic storage devices, or any suitable combination thereof.

[0096] Computer-readable storage media may also include data signals propagated in baseband or as part of a carrier wave, carrying computer-readable program code, specific examples of which include, but are not limited to, electromagnetic signals, optical signals, or any suitable combination thereof.

[0097] It is understood that the above embodiments are merely exemplary embodiments used to illustrate the principles of this disclosure, and this disclosure is not limited thereto. For those skilled in the art, various modifications and improvements can be made without departing from the spirit and substance of this disclosure, and these modifications and improvements are also considered to be within the scope of protection of this disclosure.

Claims

1. A method for predicting frequency regulation commands of a supercapacity energy storage coupled thermal power unit, characterized in that, The method includes: Topology analysis was performed on the supercapacity energy storage subsystem, the thermal power frequency regulation subsystem, and the coupled control subsystem to determine the collaborative division of labor mechanism. Multi-scale characteristics were analyzed based on historical frequency regulation command data. The multi-scale characteristics include second-level high-frequency random fluctuations and minute-level low-frequency trend changes. The collaborative division of labor mechanism is as follows: frequency regulation requirements with energy storage response frequency higher than or equal to 1Hz; frequency regulation requirements with thermal power tracking frequency lower than or equal to 0.01Hz. Parameters are collected at the component level, subsystem level, and system level in a hierarchical manner to construct a standardized parameter matrix; The raw frequency modulation command data is subjected to outlier handling, noise filtering, and standardization preprocessing to obtain a standardized command sequence. Based on the parameters set according to the multi-scale features, the standardized instruction sequence is decomposed using an improved variational mode decomposition model that introduces slack variables to obtain multiple intrinsic mode function subsequences. The similarity between the intrinsic mode function subsequences is calculated based on multi-scale entropy and dynamic time warping. Similar subsequence pairs are selected and fused according to energy ratio to generate multiple comprehensive components. Each of the aforementioned integrated components is input into multiple improved adaptive feature-gated recurrent unit networks for parallel prediction. The instantaneous rate of change of the input sequence is introduced into the reset gate of the improved adaptive feature-gated recurrent unit network, and an adaptive adjustment factor is embedded in the update gate. The prediction results of each integrated component are denormalized and superimposed to obtain the predicted value of the frequency modulation command. The predicted value is then evaluated in multiple dimensions based on prediction error index, response adaptability index, and engineering practicality index. The response adaptability index includes high-frequency response deviation rate and low-frequency lag time. The engineering practicality index includes computational efficiency and robustness coefficient.

2. The method for predicting frequency regulation commands of supercapacity energy storage coupled thermal power units according to claim 1, characterized in that, The constrained variational model of the improved variational mode decomposition model is expressed as follows: In the formula, For the first A sequence of intrinsic mode functions For the first The center frequencies of the subsequence of eigenmode functions For the Dirac function, The imaginary unit, Pi For convolution operators, It is the L2 norm. As slack variables, It is a standardized instruction sequence.

3. The method for predicting frequency regulation commands of supercapacity energy storage coupled thermal power units according to claim 1, characterized in that, The energy-weighted fusion generates multiple composite components, as shown in the following formula: In the formula, For the first Each comprehensive component in The value at time, Frequency band identifier for the composite components, , These are the intrinsic mode function subsequences. and The fusion weight, , These are the intrinsic mode function subsequences. and Energy.

4. The method for predicting frequency regulation commands of supercapacity energy storage coupled thermal power units according to claim 1, characterized in that, The output vector of the reset gate of the improved adaptive feature-gated recurrent unit network is shown in the following equation: In the formula, for The gate's output vector is reset at all times. It is the Sigmoid activation function. To reset the weight matrix of the gate, for The input feature vector at time t, for The hidden state vector at time step 1. To reset the gate's bias vector, for The instantaneous rate of change of the input sequence at time t. The weight vector is the instantaneous rate of change. for The input feature vector at time t, The dimension of the input feature vector.

5. The method for predicting frequency regulation commands of supercapacity energy storage coupled thermal power units according to claim 1, characterized in that, The output vector of the update gate of the improved adaptive feature-gated recurrent unit network is shown in the following equation: In the formula, for Update the gate's output vector continuously. It is the Sigmoid activation function. To update the gate weight matrix, for The input feature vector at time t, for The hidden state vector at time step 1. To update the bias vector of the gate, for The adaptive adjustment factor at time. The weight vector of the adaptive adjustment factor. To update the initial output of the gate, It is a function with maximum value. for The input feature vector at time t, The dimension of the input feature vector.

6. A frequency regulation command prediction system for a supercapacity energy storage coupled thermal power unit, characterized in that, The system includes: The feature analysis module is used to perform topology analysis on the supercapacity energy storage subsystem, the thermal power frequency regulation subsystem, and the coupled control subsystem, determine the collaborative division of labor mechanism, and analyze multi-scale features based on historical frequency regulation command data. The multi-scale features include second-level high-frequency random fluctuations and minute-level low-frequency trend changes. The collaborative division of labor mechanism is: frequency regulation requirements with energy storage response frequency higher than or equal to 1Hz; and frequency regulation requirements with thermal power tracking frequency lower than or equal to 0.01Hz. The parameter acquisition module is used to acquire parameters at the component level, subsystem level, and system level in a hierarchical manner, and to construct a standardized parameter matrix. The preprocessing module is used to perform outlier handling, noise filtering, and standardization preprocessing on the raw frequency modulation command data to obtain a standardized command sequence. The mode decomposition module is used to set parameters based on the multi-scale features and decompose the standardized instruction sequence using an improved variational mode decomposition model that introduces relaxation variables to obtain multiple intrinsic mode function subsequences. The weighted fusion module is used to calculate the similarity between the intrinsic mode function subsequences based on multi-scale entropy and dynamic time warping, filter similar subsequence pairs and fuse them in a weighted manner according to energy ratio to generate multiple comprehensive components. The parallel prediction module is used to input each of the comprehensive components into multiple improved adaptive feature-gated recurrent unit networks for parallel prediction. The instantaneous rate of change of the input sequence is introduced into the reset gate of the improved adaptive feature-gated recurrent unit network, and an adaptive adjustment factor is embedded in the update gate. The multidimensional evaluation module is used to de-standardize and superimpose the prediction results of each comprehensive component to obtain the frequency modulation command prediction value, and to perform multidimensional evaluation based on prediction error index, response adaptability index and engineering practicality index; the response adaptability index includes high frequency response deviation rate and low frequency lag time; the engineering practicality index includes computational efficiency and robustness coefficient.

7. An electronic device, characterized in that, include: At least one processor; as well as, A memory communicatively connected to the at least one processor is used to store one or more programs, which, when executed by the at least one processor, enable the at least one processor to implement the frequency regulation command prediction method for supercapacity energy storage coupled thermal power units as described in any one of claims 1 to 5.

8. A computer-readable storage medium storing a computer program, characterized in that, When the computer program is executed by the processor, it implements the method for predicting frequency regulation commands of supercapacity energy storage coupled thermal power units as described in any one of claims 1 to 5.

Images