Energy storage frequency modulation instruction prediction method based on AFGRU and related equipment

Abstract

The invention relates to the technical field of energy storage frequency modulation prediction, in particular to an AFGRU-based energy storage frequency modulation instruction prediction method and related equipment. The method comprises the following steps: decomposing an obtained original frequency modulation sequence by using a variational mode decomposition method to obtain a plurality of sub-sequences; fusing the plurality of sub-sequences into a set number of comprehensive components through a multi-scale entropy-dynamic alignment fusion method; and inputting the comprehensive component into the trained AFGRU network model, and predicting a next energy storage frequency modulation instruction to obtain an energy storage frequency modulation instruction prediction result. According to the method, by accurately predicting the energy storage frequency modulation instruction, the stability and reliability of the power grid are improved, and the influence of frequency fluctuation on power equipment and users is reduced.

Description

Technical Field

[0001] This invention relates to the field of energy storage frequency regulation prediction technology, specifically to an energy storage frequency regulation command prediction method and related equipment based on AFGRU. Background Technology

[0002] With the large-scale integration of new energy sources into the power grid, frequency fluctuations in the power system have become increasingly prominent. Against this backdrop, energy storage systems, due to their rapid response capabilities, have become a key means of supporting power system frequency regulation. Frequency regulation command prediction, as a crucial upstream component of the energy storage control system, directly impacts the unit's regulation efficiency and overall economic viability. According to relevant statistics, for every 1 percentage point reduction in prediction error, the revenue of a combined thermal power and energy storage frequency regulation system can increase by 2.3%.

[0003] However, frequency regulation command sequences exhibit significant multi-scale characteristics: on the one hand, they contain second-level random fluctuations, such as sudden load changes; on the other hand, they also contain minute-level trend changes, such as the ramp-up process of renewable energy output. This characteristic makes it difficult for traditional forecasting methods to simultaneously capture dynamic features at different time scales. Especially during the transformation of the new power system, the integration of a high proportion of renewable energy has led to a sharp increase in the complexity of frequency regulation commands. Taking a provincial power grid in 2023 as an example, the peak-to-valley difference in its frequency regulation commands reached 48% of the rated power, which places higher demands on forecasting models.

[0004] Current research on frequency modulation (FM) command prediction mainly falls into two categories. One category is the signal decomposition-prediction framework. This type of method uses techniques such as Variational Mode Decomposition (VMD) and Empirical Mode Decomposition (EMD) to decompose the original sequence into multiple sub-sequences, which are then predicted separately. For example, the VMD-GRU model is used to decompose the sequence into five IMF components for prediction, but the fixed decomposition mode leads to high-frequency noise and low-frequency trends being mixed together. The other category utilizes deep learning, i.e., directly modeling sequence dependencies through TCN, Transformer, etc. For example, existing methods use spatiotemporal Transformer models to predict regional FM demand, but the prediction error for step commands still exceeds 3%.

[0005] In summary, existing frequency modulation methods suffer from the following bottlenecks: First, subsequence redundancy: decomposition methods such as VMD generate a large number of components with similar patterns. For example, IMF1-IMF3 may all represent random noise, which not only increases the computational burden but also introduces error coupling. Second, lack of scale adaptability: Standard recurrent networks (such as GRU) use the same memory mechanism for high-frequency and low-frequency components, resulting in over-smoothing of high-frequency components and underfitting of low-frequency components. Summary of the Invention

[0006] The technical problem to be solved by the present invention is to provide an AFGRU-based method and related equipment for predicting energy storage frequency regulation commands, in order to address the shortcomings of the prior art and solve the technical problem of low prediction accuracy of current energy storage frequency regulation commands.

[0007] The objective of this invention is achieved through the following technical solutions: In a first aspect, the present invention provides a method for predicting energy storage frequency regulation commands based on AFGRU, comprising: The original frequency modulation sequence was decomposed using the variational mode decomposition method to obtain several sub-sequences; Several sequences are fused into a set number of composite components using a multi-scale entropy-dynamic alignment fusion method. The integrated components are input into the trained AFGRU network model to predict the next energy storage frequency regulation command, and the prediction result of the energy storage frequency regulation command is obtained.

[0008] As a further improvement of the present invention, the obtained original frequency-modulated sequence is decomposed using the variational mode decomposition method, including: The original frequency-modulated sequence is decomposed into several modal components. With the goal of minimizing the estimated bandwidth of each modal component, and taking the sum of all modal components as the original frequency-modulated sequence as a constraint, a constrained variational model is constructed. Each modal component corresponds to a center frequency. The constrained variational model is solved iteratively; the steps of the iterative solution are as follows: The optimal spectrum of the modal components and the updated center frequency are calculated using the alternating direction multiplier algorithm in each iteration until the iteration termination condition is met. Based on the order of the center frequencies, several optimal spectra of the corresponding modal components are selected as subsequences.

[0009] As a further improvement of the present invention, the constrained variational model is as follows:

[0010]

[0011] In the formula, For the k-th modal component, The center frequency corresponding to the k-th modal function. For the Dirac function, is the convolution operator; K is the preset modality. The original frequency modulation sequence signal, For averaging, The imaginary unit, It is the square of the L2 norm.

[0012] As a further improvement of the present invention, the optimal spectrum of the modal components and the updated center frequency are calculated in each iteration using the alternating direction multiplier algorithm, including:

[0013]

[0014] In the formula, The spectrum of the k-th modal component in the (n+1)-th iteration. For frequency variables, The spectrum of the original frequency-modulated sequence signal. It is the sum of the Fourier transforms of all modal components except the k-th modal component. As a secondary penalty factor, For Lagrange factors, The Fourier transform of the Lagrange factor Let be the center frequency of the k-th modal component. Let be the center frequency of the k-th modal component in the (n+1)-th iteration. Let be the power spectral density of the k-th modal component.

[0015] As a further improvement of the present invention, a multi-scale entropy-dynamic alignment fusion method is used to fuse several sub-sequences into a set number of comprehensive components, including: Several subsequences are preprocessed to obtain standardized subsequences; The multi-scale dynamic similarity between any two standardized subsequences is calculated using multi-scale entropy and the DTW method, and the similarity score between the two standardized subsequences is obtained. Based on the similarity scores, subsequence pairs with high similarity are selected, and the subsequence pairs are fused using an adaptive energy weighting method to obtain a comprehensive component.

[0016] As a further improvement to the present invention, the AFGRU (Attention-based Gated RecurrentUnit) network model is a network model based on GRU, specifically including: An instantaneous velocity factor for the input characteristics is introduced into the standard reset gate of the GRU, and the instantaneous velocity factor is used to enhance the abrupt change characteristics of the response input synthesis components. An adaptive factor is introduced into the update mechanism of the update gate of the GRU, and the memory strength of the update gate is dynamically adjusted using the adaptive factor.

[0017] As a further improvement of the present invention, the adaptive factor is:

[0018] In the formula, As an adaptive factor, For learnable parameters, To update the gate's output vector at time t, Let be the feature vector of the input composite component at time t. Let be the eigenvector of the input composite component at time t-1.

[0019] Secondly, the present invention provides an AFGRU-based energy storage frequency regulation command prediction system for implementing the aforementioned AFGRU-based energy storage frequency regulation command prediction method, comprising: The original sequence decomposition module is used to decompose the acquired original frequency modulation sequence using the variational mode decomposition method to obtain several sub-sequences; The feature fusion module is used to fuse several sub-sequences into a set number of comprehensive components using the multi-scale entropy-dynamic alignment fusion method. The frequency modulation prediction module is used to input the integrated components into the trained AFGRU network model to predict the next energy storage frequency modulation command and obtain the energy storage frequency modulation command prediction result.

[0020] Thirdly, the present invention provides a computer-readable storage medium for storing one or more programs, the one or more programs including instructions that, when executed by a computing device, cause the computing device to perform the above-described AFGRU-based energy storage frequency modulation instruction prediction method.

[0021] Fourthly, the present invention provides a computing device, comprising: One or more processors, a memory, and one or more programs, wherein the one or more programs are stored in the memory and configured to be executed by the one or more processors, and the one or more programs include steps for performing the above-described AFGRU-based energy storage frequency modulation command prediction method.

[0022] The beneficial effects of this invention are as follows: This invention provides an AFGRU-based method for predicting frequency modulation commands in energy storage. It decomposes the complex original frequency modulation sequence into several subsequences with different center frequencies and finite bandwidths using variational mode decomposition, avoiding mode aliasing, better preserving the original signal characteristics, and improving the accuracy and stability of signal decomposition. Through a multi-scale entropy-dynamic alignment fusion method, based on multi-scale dynamic similarity analysis between sequences, it adaptively weights and fuses subsequences with similar characteristics, effectively reducing the number of redundant subsequences and avoiding error accumulation caused by parallel prediction of multiple models. Simultaneously, it effectively extracts important features from each subsequence and organically fuses these features to generate a predetermined number of comprehensive components. These comprehensive components not only contain the main information of the original frequency modulation sequence but also highlight the intrinsic connections between different features, providing richer and more representative input features for subsequent prediction models, thereby improving the accuracy of subsequent predictions. The integrated components are input into the trained AFGRU network model. An instantaneous velocity factor is introduced into the reset gate to enhance the ability to capture abrupt changes in input features. Simultaneously, an adaptive factor is introduced into the update gate to dynamically adjust the memory strength. This allows the network to adaptively process component features at different time scales, overcoming the shortcomings of traditional GRU networks such as over-smoothing of high-frequency components and underfitting of low-frequency components. These three technical steps form a synergistic mechanism, optimizing the feature extraction process and improving the model's ability to express multi-scale features. This, in turn, more accurately captures the changing patterns and trends of energy storage frequency regulation commands, thereby improving the prediction accuracy of the next energy storage frequency regulation command.

[0023] Furthermore, the original frequency modulation sequence is effectively preprocessed by variational mode decomposition and multi-scale entropy-dynamic alignment fusion method, which removes noise and redundant information from the original frequency modulation sequence, making the integrated component data input into the AFGRU model cleaner and more standardized, thereby reducing the model's sensitivity to noise and outliers and improving the model's robustness.

[0024] Furthermore, a refined decomposition of the frequency-modulated sequence was achieved by constructing a variational model with specific constraints. Using the estimated bandwidth and minimization of modal components as the objective function effectively suppresses redundant components generated during the decomposition process, ensuring that each modal component has a clear frequency characteristic. The integrity of the modal components and the original sequence is used as a constraint to avoid signal energy loss during decomposition. The correspondence between modal components and center frequencies allows the decomposition results to characterize the dynamic properties at different time scales. An iterative solution using an alternating direction multiplier algorithm, through alternating calculations of spectrum optimization and center frequency updates, gradually approaches the optimal solution. This dynamic adjustment mechanism can adapt to the non-stationary characteristics of the frequency-modulated sequence. The setting of the iteration termination condition ensures the convergence of the decomposition process, while the modal component selection strategy based on the center frequency order can eliminate noise interference components and retain physically meaningful subsequences. Attached Figure Description

[0025] To more clearly illustrate the technical solutions in the embodiments of the present invention or the prior art, the drawings used in the description of the embodiments or the prior art will be briefly introduced below. Obviously, the drawings described below are some embodiments of the present invention. For those skilled in the art, other drawings can be obtained based on these drawings without creative effort.

[0026] Figure 1 This is a schematic diagram of the prediction method for GRU networks in the prior art; Figure 2 This is a schematic diagram of the VMD-GRU prediction model processing flow in the existing technology; Figure 3 This is a schematic diagram of the original frequency modulation sequence processing flow in an embodiment of the present invention.

[0027] Figure 4 This is a schematic diagram of the GRU network structure in an embodiment of the present invention.

[0028] Figure 5 This is a schematic diagram of the energy storage frequency regulation command prediction method based on AFGRU in an embodiment of the present invention.

[0029] Figure 6 This is a graph of the original frequency modulation sequence in the first example of this invention.

[0030] Figure 7 This is a graph of the original frequency modulation sequence in the second example of this invention.

[0031] Figure 8 This is a schematic diagram comparing the prediction results of three sets of models in the first example of this invention.

[0032] Figure 9 This is a schematic diagram comparing the prediction results of three sets of models in the second example of this invention.

[0033] Figure 10 This is a schematic diagram of the structure of the computer device in an embodiment of the present invention. Detailed Implementation

[0034] To make the objectives and technical solutions of this invention clearer and easier to understand, the invention will be further described in detail below with reference to the accompanying drawings and embodiments. The specific embodiments described herein are for illustrative purposes only and are not intended to limit the invention.

[0035] The technical solution of the present invention will be clearly and completely described below with reference to the accompanying drawings and specific embodiments. The described embodiments are only some embodiments of the present invention, and not all embodiments.

[0036] In existing technologies, power system frequency regulation command prediction faces challenges such as subsequence redundancy and insufficient adaptability to multi-scale features. Traditional methods employ a single GRU network, such as... Figure 1 As shown, modeling multi-scale features leads to over-smoothing of high-frequency components and hysteresis response of low-frequency components. The classic VMD-GRU, as... Figure 2 As shown, although signal decomposition alleviates this problem, its fixed mode number output is prone to generating redundant subsequences (approximately 40% information overlap), and independent prediction ignores the inherent regularity between components. To address these issues, this invention analyzes the potential of variational mode decomposition in eliminating mode aliasing, finding that its constrained variational model can optimize the distribution of mode components. Subsequently, a fusion method of multi-scale entropy and dynamic time warping is explored, revealing that similarity scores can effectively identify redundant components. Finally, improvements are made to the GRU network memory mechanism, proposing to enhance feature capture capabilities through instantaneous velocity factors and adaptive factors.

[0037] Example 1 This embodiment discloses an AFGRU-based method for predicting energy storage frequency regulation commands, aiming to solve the problems of low prediction accuracy and slow response speed of traditional prediction methods for nonlinear and non-stationary energy storage frequency regulation command sequences. Figures 3 to 5 As shown, this method extracts sequence features through variational mode decomposition (VMD), optimizes feature components by combining multi-scale entropy-dynamic alignment fusion, and finally achieves high-precision prediction using an improved AFGRU network.

[0038] like Figure 3 , Figure 5 As shown, the main steps of the AFGRU-based energy storage frequency regulation command prediction method include: decomposing the acquired original frequency regulation sequence using the variational mode decomposition method to obtain several sub-sequences; fusing the several sub-sequences into a set number of comprehensive components using the multi-scale entropy-dynamic alignment fusion method; and inputting the comprehensive components into the trained AFGRU network model to predict the next energy storage frequency regulation command, thereby obtaining the energy storage frequency regulation command prediction result.

[0039] The original frequency-modulated signal is transformed into physically meaningful subsequences through variational mode decomposition. Multi-scale dynamic similarity analysis is used to identify and fuse similar patterns, eliminating redundant components. The fused composite components are then input into an improved AFGRU network. An instantaneous velocity factor captures the abrupt changes in input features, while an adaptive factor dynamically adjusts the memory strength based on these changes. These three processing stages form a closed-loop optimization: the decomposition stage ensures effective feature separation, the fusion stage reduces the feature set, and the network improvement enhances multi-scale feature modeling capabilities.

[0040] Specifically, the original energy storage frequency regulation command sequence exhibits nonlinear and non-stationary characteristics, and directly inputting it into the model will lead to a decrease in prediction accuracy. This embodiment uses variational mode decomposition to decompose the original energy storage frequency regulation command sequence into several regular subsequences, thereby reducing the volatility and nonlinearity of the original frequency regulation sequence and improving prediction accuracy. The specific steps of variational mode decomposition of the original energy storage frequency regulation command sequence are as follows: Acquire historical energy storage frequency regulation command data from the target power grid to form the original frequency regulation sequence signal. .

[0041] A constrained variational model is constructed, aiming to minimize the estimated bandwidth of each modal component. This model is built using the sum of all modal components as constraints on the original frequency-modulated sequence. Variational mode decomposition (VMD) refers to a signal processing method that optimizes the center frequency distribution of modal components using a constrained variational model. Specifically, the alternating direction multiplier algorithm can be used iteratively to solve for the optimal spectrum. This decomposition method avoids the mode aliasing problem of empirical mode decomposition.

[0042] Specifically, the constrained variational model decomposes the original frequency-modulated sequence signal into K modal components (the value of K is adjusted according to the sequence complexity; in this embodiment, K is 6). Each modal component corresponds to a center frequency. Furthermore, it satisfies that the modal components are amplitude-frequency modulated signals with finite bandwidth, and the sum of all modal components equals the original frequency modulation sequence.

[0043] The constrained variational model is:

[0044]

[0045] In the formula, For the k-th modal component, The center frequency corresponding to the k-th modal function. For the Dirac function, is the convolution operator; K is the preset modality. The original frequency modulation sequence signal, For averaging, The imaginary unit, It is the square of the L2 norm.

[0046] The constrained variational model is solved iteratively by using the alternating direction multiplier algorithm. The optimal spectrum of the modal components and the updated center frequency are calculated in each iteration until the iteration termination condition is met. Several optimal spectra of the corresponding modal components are selected as subsequences according to the order of the center frequencies.

[0047] Specifically, the process of iteratively updating the modal components and center frequencies is as follows:

[0048]

[0049] In the formula, The spectrum of the k-th modal component in the (n+1)-th iteration. For frequency variables, The spectrum of the original frequency-modulated sequence signal. It is the sum of the Fourier transforms of all modal components except the k-th modal component. As a secondary penalty factor, For Lagrange factors, The Fourier transform of the Lagrange factor Let be the center frequency of the k-th modal component. Let be the center frequency of the k-th modal component in the (n+1)-th iteration. Let be the power spectral density of the k-th modal component.

[0050] This embodiment introduces a secondary penalty factor. and Lagrange multipliers Eliminate the constraints of the variational model. Quadratic penalty factor. The constraint strength used to balance the spectral reconstruction accuracy and modal component bandwidth can be dynamically adjusted using the gradient descent method. Its numerical range can be from 0.1 to 1.0, and can be increased... The value can suppress the bandwidth spread of modal components; Lagrange factor Fourier transform Used to compensate for the decomposition residuals in the frequency domain, specifically by updating the residual spectrum through the accumulated error during the iteration process, ensuring that the constraints of the original signal reconstruction are strictly satisfied; center frequency The update calculation is based on power spectral density. Frequency domain energy weighting can be implemented by numerical integration to calculate the integral terms in the numerator and denominator, automatically converging the center frequency to the main energy band of the modal component. During the spectrum update phase, the spectrum of each modal component is adjusted by removing the superposition effects of other modal component spectra from the original signal spectrum and by incorporating a quadratic penalty term to suppress frequency bands deviating from the current center frequency. This calculation method forces the spectral energy of each modal component to concentrate on the frequency band with the dominant energy of the modal component. Within a narrow band centered on the center, spectral overlap between different components is avoided. During the center frequency update phase, the first moment in the frequency domain of the power spectral density is calculated, causing the center frequency to automatically shift to the energy centroid position of the current modal component. The alternating iterative execution of these two calculations ensures that the decomposition process meets the constraint of complete reconstruction of the original signal, and achieves accurate separation of signal components at different time scales through dynamic focusing of frequency domain energy.

[0051] In each iteration, the spectrum and center frequency parameters of the modal components are updated alternately, and convergence of the decomposition process is achieved through dynamic adjustment. The iteration termination condition refers to the criterion for controlling the end of the decomposition process. Specifically, it can be a preset maximum number of iterations or a threshold for the change in the modal component spectrum, for example, terminating the calculation when the change in center frequency between two adjacent iterations is less than 0.1 Hz. The resulting subsequence is the intrinsic mode function (IMF). In this embodiment, the original frequency-modulated sequence is decomposed into 6 subsequences IMF1, IMF2, IMF3, ..., IMF6, each subsequence having a length of N (i.e., sampling once per second, for a total of N points). Let... 'i' represents any value from 1 to 6. The following steps are all based on IMF... i Represents any subsequence.

[0052] The multi-scale entropy-dynamic alignment fusion method is used to fuse several subsequences into a set number of comprehensive components. Specifically, it includes: preprocessing several subsequences to obtain standardized subsequences; calculating the multi-scale dynamic similarity between any two standardized subsequences based on multi-scale entropy and the DTW method to obtain a similarity score between the two standardized subsequences; selecting subsequence pairs with high similarity based on the similarity score, and fusing the subsequence pairs using an adaptive energy weighting method to obtain the comprehensive components.

[0053] Specifically, the multi-scale entropy-dynamic alignment fusion method refers to a sequence fusion technique that combines multi-scale entropy analysis with dynamic time warping. Specifically, it can use an adaptive energy weighting method to merge similar sub-sequences. This fusion process effectively reduces the number of redundant components.

[0054] Standardization preprocessing refers to the normalization of subsequences, where each subsequence is standardized to have zero mean and unit variance, resulting in a standardized subsequence. The standardized formula is:

[0055] In the formula, It is the mean of the i-th subsequence. It is the standard deviation of the i-th subsequence.

[0056] For any two normalized subsequences and Multi-scale dynamic similarity is calculated using multi-scale entropy and DTW (Dynamic Time Warping) method to obtain similarity scores between two standardized subsequences. Multi-scale dynamic similarity combines multi-scale entropy and DTW distance to assess regular similarity at different time scales.

[0057] Multi-scale dynamic similarity is:

[0058] In the formula, The similarity score between two standardized subsequences. The composite distance between subsequences i and j is a combination of multi-scale entropy distance and DTW distance. This is a monotonic constraint factor. The composite distance is:

[0059] In the formula, The multi-scale entropy distance between subsequences This represents the dynamic time-warped distance between subsequences. These are weighting coefficients used to balance the contributions of entropy and dynamic distortion, as shown in this embodiment. It is 0.6.

[0060] The multi-scale entropy distance between subsequences is:

[0061] In the formula, Let be the sample entropy of the i-th subsequence at scale s. Let S be the sample entropy of the j-th subsequence at scale s, and let the set of scale factors be S=[1,2,..,5].

[0062] The dynamic time-warped distance between subsequences is:

[0063] In the formula, This represents the set of all possible DTW paths.

[0064] The monotonicity constraint factor is: .

[0065] By calculating multi-scale dynamic similarity, standardized subsequence pairs are obtained. similarity score Sort the sequences by similarity score from highest to lowest, and select the standardized subsequence pairs with the highest similarity (a set number) for fusion.

[0066] In this embodiment, the fusion employs adaptive energy weighting to ensure that high-energy subsequences (representing stronger signal components) dominate the fusion process.

[0067] For comprehensive components, The weights corresponding to subsequence i* are... Let be the weight corresponding to subsequence j*. The weights are derived based on energy ratios, and the weights of each subsequence are:

[0068]

[0069]

[0070] in, Let i be the energy of the i*th subsequence. Let be the energy of the j*th subsequence. After fusion, the composite component Fused1 is obtained. Fused1 replaces the original IMF. i and IMF j In this embodiment, the six subsequences are fused to obtain three composite components: Fused1, Fused2, and Fused3.

[0071] After standardization and preprocessing, the subsequences undergo multi-scale entropy calculation to generate coarse-grained sequences at multiple time scales, extracting complexity features at each scale. Dynamic time warping (DTW) is used to dynamically align non-uniform-length subsequences and calculate the matching degree of local fluctuation patterns. The global complexity features of multi-scale entropy are weighted and fused with the local alignment path distances from DTW to generate a comprehensive similarity score. Highly correlated subsequence pairs are selected based on a preset similarity threshold, and their energy proportions are calculated. Adaptive weight coefficients are generated by combining the similarity scores, and finally, effective fusion of subsequences is achieved through weighted superposition. This process, through dynamic selection and feature integration, avoids ineffective calculations of irrelevant components while retaining key feature information.

[0072] The integrated components are input into the trained AFGRU network model to predict the next energy storage frequency regulation command, thus obtaining the predicted energy storage frequency regulation command result.

[0073] In this embodiment, the AFGRU network model is an improved version of the GRU network model. By introducing an instantaneous velocity factor and an adaptive factor into the gated recurrent unit, specifically by adding an input characteristic change rate calculation module to the reset gate and setting dynamically adjustable parameters in the update gate, this structure enhances the network's ability to express multi-scale features.

[0074] Existing GRU (Gated Recurrent Unit) networks effectively capture temporal dependencies through gating mechanisms, such as... Figure 4 As shown, the calculation process is as follows: The degree of forgetting of historical states is determined by resetting the gate:

[0075] in, The output vector of the gate is reset at time t. It is the sigmoid activation function. Reset the trainable weight matrix of the gate. It is the hidden state vector at time t-1. It is the input feature vector at time t.

[0076] Adjust the merging ratio of old and new states using the update gate:

[0077] in, The gate's output vector is updated at time t. It is the trainable weight matrix of the update gate.

[0078] The generated candidate hidden states are:

[0079] in, It is the candidate hidden state vector at time t. It is the hyperbolic tangent activation function. The trainable weight matrix for candidate states, Element-wise multiplication (Hadamard product) Reset the historical state filter of the gate control.

[0080] The final hidden state is updated as follows:

[0081] However, existing GRU networks, due to their fixed gating mechanism, struggle to adapt to the multi-scale characteristics of frequency modulation commands: high-frequency component prediction is susceptible to random noise interference, and low-frequency trend response exhibits lag. To address this, this embodiment proposes an Adaptive Feature Gating Unit (AFGRU), which achieves collaborative optimization prediction of multi-scale features through feature importance weighting and a time-varying memory update mechanism.

[0082] Specifically, the instantaneous gradient of the input feature is introduced into the standard reset gate to enhance the ability to perceive abrupt changes.

[0083]

[0084] in: It is the time-domain gradient factor. This is the gradient weight matrix. When the frequency modulation command changes abruptly... Increase Approaching zero forced forgetting of historical states avoids the inertial influence of low-frequency components.

[0085] An adaptive factor is introduced into the update mechanism of the GRU update gate to dynamically adjust the memory strength of the update gate, enabling it to automatically adapt to different time series variation characteristics.

[0086] in, This is the adaptive factor. The adaptive factor is:

[0087] In the formula, As an adaptive factor, For learnable parameters, To update the gate's output vector at time t, Let be the feature vector of the input composite component at time t. This is the feature vector of the input composite component at time t-1. In this embodiment... When the input changes drastically (the overall component exhibits weak regularity), increase... Accelerate updates. When input changes are gradual (strong regularity in the composite components), reduce... Maintain long-term memory.

[0088] The adaptive factor constructs a dynamic adjustment mechanism by fusing learnable parameters with changes in input features. During the gate update operation, the learnable parameters... With update gate output The product term inherits the network's original memory characteristics, maintaining a stable memory ability for low-frequency trend components. Simultaneously, the absolute value of the difference between adjacent time steps of the input feature vector undergoes a non-linear transformation using the arg function, enhancing sensitivity to high-frequency abrupt changes. When the input features change drastically, the difference term increases significantly, prompting the adaptive factor to... Increasing the attention given to current input features by the update gate effectively prevents high-frequency components from being overly smoothed by historical states. For steadily changing low-frequency components, the difference term is kept at a small value. In this case, the adaptive factor mainly relies on the learnable parameters for adjustment, preserving the network's ability to remember long-term trends. This composite adjustment mechanism enables the model to automatically adjust the weights of the memory units' dependence on current information and historical states based on the dynamic characteristics of the input features.

[0089] Example 2 To verify the frequency regulation command prediction capability of the prediction model in Example 1, this invention conducts an experimental study on frequency regulation command data prediction using a power plant in Inner Mongolia as an example. To further verify the prediction capability in Example 1, a comparative study is conducted using frequency regulation data from two different time periods.

[0090] like Figure 6As shown, the first example of this study uses frequency modulation command data acquired every 2 seconds on January 10, 2025, for a total of 751 data points. The frequency modulation sequences are divided into training and testing sets according to a set ratio. 671 data samples are selected to train the AFGRU network model, and the remaining 80 data points are used to test the performance of the proposed model.

[0091] like Figure 7 As shown, the second example in this study uses frequency modulation (FM) command data acquired every 2 seconds starting at 4:00 AM on January 11, 2025, totaling 287 data points. The FM sequences are divided into training and testing sets according to a predetermined ratio. 228 data samples are selected to train the AFGRU network model, and the remaining 59 data points are used to test the performance of the proposed model.

[0092] In this embodiment, GRU network and AFGRU network are compared as prediction models to better highlight the accuracy of the prediction method in Embodiment 1. Figure 8 and Figure 9 The prediction results of the two examples are compared. Figure 8 and Figure 9 It can be clearly seen that the prediction method of this invention has the best fit with the original data. Figure 8 , Figure 9 In the diagram, the blue line represents the original frequency modulation command sequence, the orange part represents the frequency modulation command result predicted by the GRU, the red part represents the frequency modulation command result predicted by a single AFGRU, and the green part represents the frequency modulation command result obtained by the prediction method of this invention.

[0093] Each prediction method was trained and tested using frequency modulation command data from two examples. Root mean square error (RMSE), mean absolute error (MAE), and mean absolute percentage error (MAPE) were used as evaluation metrics. Smaller evaluation metrics indicate smaller prediction errors and better performance of the prediction model.

[0094] Specifically, the evaluation indicators are as follows:

[0095]

[0096]

[0097] In the formula, Root mean square error (RMSE), mean absolute error (MAE), and mean absolute percentage error (MAPE) are all expressed as values. These are predicted values.

[0098] Table 1 shows the prediction error coefficients for different models.

[0099] Table 1. Prediction error coefficients of the three models

[0100] Table 1 clearly shows that the error coefficient decreases with the improvement of the model structure, meaning that the established model has the smallest error coefficient, while the single GRU model has the largest error coefficient. As shown in Table 1, when compared with all other benchmark models, the prediction in this embodiment has the smallest error coefficient, indicating that the established model has the best prediction results. For example, the above conclusions can be drawn from Table 1, where the first instance's predicted RMSE decreased by 96.80% and 92.31%, MAE decreased by 91.67% and 88.52%, and MAPE decreased by 96.91% and 93.84%, respectively. The second instance's predicted RMSE decreased by 96.74% and 94.12%, MAE decreased by 88.61% and 82.69%, and MAPE decreased by 96.40% and 92.57%, respectively. This shows that compared with the GRU network and the AFGRU network, the integrated components after MSEDAF fusion better reflect the inherent laws of the original sequence, thereby improving prediction accuracy. Compared with a single GRU model, AFGRU has satisfactory prediction performance. As shown in Table 1, the MAPE value of the first instance is reduced by 49.84%, and the MAPE value of the second instance is reduced by 51.54%. This shows that AFGRU effectively solves the shortcomings of traditional GRU in complex time series prediction through feature importance weighting and time-varying memory update mechanism, and significantly improves prediction accuracy.

[0101] The experimental results based on frequency regulation data from power plants in Inner Mongolia are summarized as follows: The prediction method in this embodiment exhibits the best performance compared to the benchmark models (GRU, AFGRU), demonstrating that the model can accurately capture the multi-scale dynamic features of frequency regulation commands. The variational mode decomposition (VMD) fusion mechanism in this embodiment effectively solves the subsequence redundancy problem. Through the synergy of multi-scale entropy regularity measurement and dynamic time warping, the six VMD subsequences are condensed into three comprehensive components, eliminating information overlap and significantly improving feature expression efficiency. The AFGRU network innovatively optimizes prediction capabilities through dual improvements: the reset gate introduces an instantaneous gradient factor to enhance the response to sudden commands; the update gate embeds an adaptive factor to dynamically adjust the memory strength.

[0102] In summary, the MSEDAF-AFGRU model provides high-precision decision support for energy storage frequency regulation control systems through feature fusion coupling adaptive gating mechanism optimization guided by physical laws, and can be extended to stability optimization scenarios of high-proportion new energy power grids.

[0103] Example 3 This embodiment provides an AFGRU-based energy storage frequency regulation command prediction system to implement the AFGRU-based energy storage frequency regulation command prediction method in Embodiment 1. The system mainly includes an interactive original sequence decomposition module, a feature fusion module, and a frequency regulation prediction module.

[0104] The original sequence decomposition module is used to decompose the acquired original frequency modulation sequence using the variational mode decomposition method to obtain several sub-sequences; The feature fusion module is used to fuse several sub-sequences into a set number of comprehensive components using the multi-scale entropy-dynamic alignment fusion method. The frequency modulation prediction module is used to input the integrated components into the trained AFGRU network model to predict the next energy storage frequency modulation command and obtain the energy storage frequency modulation command prediction result.

[0105] The original sequence decomposition module decomposes the original signal into modal components with definite center frequencies using a constrained variational model. It iteratively optimizes the spectral distribution using an alternating direction multiplier algorithm, generating subsequences with clear physical meaning. The feature fusion module standardizes the subsequences, calculates sequence complexity features using multi-scale entropy, evaluates cross-scale similarity using dynamic time warping, and finally performs energy-weighted fusion on highly similar subsequences. The frequency modulation prediction module introduces instantaneous velocity features of the input sequence into the reset gate operation of the AFGRU network to enhance its response to abrupt signals. In the update gate, an adaptive factor based on the input rate of change is designed to dynamically adjust the fusion ratio of historical states and the current input.

[0106] Example 4 In another embodiment of the present invention, a computer-readable storage medium is provided as a storage component within a terminal device, the function of which is to store programs and data. It should be noted that the computer-readable storage medium here encompasses not only the built-in storage components of the terminal device but also extended storage components supported by the device. Essentially, it is a tangible medium capable of containing or storing programs that can be invoked by or in conjunction with an instruction execution system, device, or apparatus. This storage medium provides storage areas for the terminal's operating system and stores one or more instructions suitable for processor loading and execution, which can constitute one or more computer programs containing program code.

[0107] Specifically, examples of computer-readable storage media (a non-exclusive list) include: electrical connections with one or more wires, portable disks, hard disks, random access memory, read-only memory, erasable programmable read-only memory, optical fibers, portable optical disc read-only memory, optical storage devices, magnetic storage devices, or any reasonable combination of the above types.

[0108] The storage medium may also include data signals propagated as part of a baseband portion or a carrier wave, carrying readable program code. Such propagated data signals may take various forms, including but not limited to electromagnetic signals, optical signals, or any reasonable combination of both. Furthermore, computer-readable storage medium may also refer to other readable media besides conventional readable storage media, capable of sending, propagating, or transmitting programs for use or operation by an instruction execution system, apparatus, or device. Program code on the storage medium can be transmitted via any suitable medium, including but not limited to wireless, wired, optical fiber, or any reasonable combination thereof.

[0109] The program code used to implement the operations of this invention can be written in any combination of one or more programming languages, including object-oriented programming languages ​​such as Java and C++, as well as conventional procedural programming languages ​​such as C. The execution modes of the program code include: running entirely on the user's computing device, running partially on the user's device as a standalone software package, running partially in a distributed manner on both the user's device and a remote computing device, or running entirely on a remote computing device or server. When a remote computing device is involved, the device can be connected to the user's computing device via any type of network such as a local area network (LAN) or a wide area network (WAN), or connected to an external computing device via the Internet through an Internet service provider.

[0110] The processor is capable of loading and executing one or more instructions stored in a computer-readable storage medium to implement the corresponding steps of the AFGRU-based energy storage frequency modulation instruction prediction method described in Example 1.

[0111] Example 5 Figure 10 This is a schematic diagram of a computer device provided according to an embodiment of the present invention.

[0112] Please see Figure 10 The terminal device is a computer device. In this embodiment, the computer device 60 includes a processor 61, a memory 62, and a computer program 63 stored in the memory 62 and executable on the processor 61. When executed by the processor 61, the computer program 63 implements the AFGRU-based energy storage frequency modulation command prediction method in Embodiment 1. To avoid repetition, this will not be described in detail here. Alternatively, when executed by the processor 61, the computer program 63 implements the functions of each model / unit in the AFGRU-based energy storage frequency modulation command prediction component computing system of the embodiment. To avoid repetition, this will not be described in detail here.

[0113] Computer device 60 can be a desktop computer, laptop, handheld computer, cloud server, or other computing device. Computer device 60 may include, but is not limited to, a processor 61 and a memory 62. Those skilled in the art will understand that... Figure 10This is merely an example of computer device 60 and does not constitute a limitation on computer device 60. It may include more or fewer components than shown, or combine certain components, or different components. For example, computer device may also include input / output devices, network access devices, buses, etc.

[0114] The processor 61 may be a central processing unit (CPU), or other general-purpose processors, CPUs, graphics processing units (GPUs), digital signal processors (DSPs), application-specific integrated circuits (ASICs), field-programmable gate arrays (FPGAs), or other programmable logic devices, discrete gate or transistor logic devices, quantum computing-based data processing logic units, discrete hardware components, etc. A general-purpose processor may be a microprocessor or any conventional processor.

[0115] The memory 62 can be an internal storage unit of the computer device 60, such as a hard disk or RAM of the computer device 60. The memory 62 can also be an external storage device of the computer device 60, such as a plug-in hard disk, smart media card (SMC), secure digital (SD) card, flash card, etc. equipped on the computer device 60.

[0116] Furthermore, the memory 62 may include both internal storage units of the computer device 60 and external storage devices. The memory 62 is used to store computer programs and other programs and data required by the computer device. The memory 62 can also be used to temporarily store data that has been output or will be output.

[0117] Any references to memory, databases, or other media used in the embodiments provided in this application may include at least one of non-volatile and volatile memory. Non-volatile memory may include read-only memory (ROM), magnetic tape, floppy disk, flash memory, optical memory, high-density embedded non-volatile memory, resistive random access memory (ReRAM), magnetic random access memory (MRAM), ferroelectric random access memory (FRAM), phase change memory (PCM), graphene memory, etc. Volatile memory may include random access memory (RAM) or external cache memory, etc. By way of illustration and not limitation, RAM may be in various forms, such as static random access memory (SRAM) or dynamic random access memory (DRAM), etc.

[0118] The databases involved in the embodiments provided in this application may include at least one type of relational database and non-relational database. Non-relational databases may include, but are not limited to, blockchain-based distributed databases. The processors involved in the embodiments provided in this application may be general-purpose processors, central processing units, graphics processing units, digital signal processors, programmable logic devices, quantum computing-based data processing logic devices, etc., and are not limited to these.

Claims

1. A method for predicting frequency regulation commands for energy storage based on AFGRU, characterized in that, include: The original frequency modulation sequence was decomposed using the variational mode decomposition method to obtain several sub-sequences; Several sequences are fused into a set number of composite components using a multi-scale entropy-dynamic alignment fusion method. The integrated components are input into the trained AFGRU network model to predict the next energy storage frequency regulation command, and the prediction result of the energy storage frequency regulation command is obtained.

2. The energy storage frequency regulation command prediction method based on AFGRU according to claim 1, characterized in that, The obtained original frequency-modulated sequence is decomposed using variational mode decomposition, including: The original frequency-modulated sequence is decomposed into several modal components. With the goal of minimizing the estimated bandwidth of each modal component, and taking the sum of all modal components as the original frequency-modulated sequence as a constraint, a constrained variational model is constructed. Each modal component corresponds to a center frequency. The constrained variational model is solved iteratively; the steps of the iterative solution are as follows: The optimal spectrum of the modal components and the updated center frequency are calculated using the alternating direction multiplier algorithm in each iteration until the iteration termination condition is met. Based on the order of the center frequencies, several optimal spectra of the corresponding modal components are selected as subsequences.

3. The energy storage frequency regulation command prediction method based on AFGRU according to claim 2, characterized in that, The constrained variational model is: In the formula, For the k-th modal component, The center frequency corresponding to the k-th mode function. For the Dirac function, is the convolution operator; K is the preset modality. The original frequency modulation sequence signal, For averaging, The imaginary unit, It is the square of the L2 norm.

4. The energy storage frequency regulation command prediction method based on AFGRU according to claim 2, characterized in that, The optimal spectrum of the modal components and the updated center frequency are calculated in each iteration using the alternating direction multiplier algorithm, including: In the formula, The spectrum of the k-th modal component in the (n+1)-th iteration. For frequency variables, The spectrum of the original frequency-modulated sequence signal. It is the sum of the Fourier transforms of all modal components except the k-th modal component. As a secondary penalty factor, For Lagrange factors, The Fourier transform of the Lagrange factor Let be the center frequency of the k-th modal component. Let be the center frequency of the k-th modal component in the (n+1)-th iteration. Let be the power spectral density of the k-th modal component.

5. The energy storage frequency regulation command prediction method based on AFGRU according to claim 1, characterized in that, The multi-scale entropy-dynamic alignment fusion method is used to fuse several subsequences into a set number of comprehensive components, including: Several subsequences are preprocessed to obtain standardized subsequences; The multi-scale dynamic similarity between any two standardized subsequences is calculated using multi-scale entropy and the DTW method, and the similarity score between the two standardized subsequences is obtained. Based on the similarity scores, subsequence pairs with high similarity are selected, and the subsequence pairs are fused using an adaptive energy weighting method to obtain a comprehensive component.

6. The energy storage frequency regulation command prediction method based on AFGRU according to claim 1, characterized in that, The AFGRU network model is a network model based on an improved version of GRU, specifically including: An instantaneous velocity factor for the input characteristics is introduced into the standard reset gate of the GRU, and the instantaneous velocity factor is used to enhance the abrupt change characteristics of the response input synthesis components. An adaptive factor is introduced into the update mechanism of the update gate of the GRU, and the memory strength of the update gate is dynamically adjusted using the adaptive factor.

7. The energy storage frequency regulation command prediction method based on AFGRU according to claim 6, characterized in that, The adaptive factor is: In the formula, As an adaptive factor, For learnable parameters, To update the gate's output vector at time t, Let be the feature vector of the input composite component at time t. Let be the eigenvector of the input composite component at time t-1.

8. An AFGRU-based energy storage frequency regulation command prediction system, used to implement the AFGRU-based energy storage frequency regulation command prediction method according to any one of claims 1 to 7, characterized in that, include: The original sequence decomposition module is used to decompose the acquired original frequency modulation sequence using the variational mode decomposition method to obtain several sub-sequences; The feature fusion module is used to fuse several sub-sequences into a set number of comprehensive components using the multi-scale entropy-dynamic alignment fusion method. The frequency modulation prediction module is used to input the integrated components into the trained AFGRU network model to predict the next energy storage frequency modulation command and obtain the energy storage frequency modulation command prediction result.

9. A computer-readable storage medium for storing one or more programs, characterized in that, The one or more programs include instructions that, when executed by a computing device, cause the computing device to perform the AFGRU-based energy storage frequency modulation instruction prediction method according to any one of claims 1 to 7.

10. A computing device, characterized in that, include: One or more processors, a memory, and one or more programs, wherein the one or more programs are stored in the memory and configured to be executed by the one or more processors, the one or more programs including steps for performing the energy storage frequency modulation command prediction method based on any one of claims 1 to 7.

Images