Energy storage dispatching method, system and device based on wind power prediction and medium

Abstract

The application provides a wind power prediction-based energy storage scheduling method, system, device and medium, relates to the technical field of energy storage scheduling, and the method comprises the following steps: backtracking and segmenting original wind power time series data; performing multi-scale decomposition and denoising; pre-establishing a wind power hybrid prediction model; performing sequence prediction to output an ultra-short-term wind power prediction sequence; performing smoothing processing based on low-pass filtering to obtain a benchmark power curve; in a local rolling optimization process of a real-time energy storage scheduling strategy, a related prediction power curve group is called, power coordination multi-objective optimization is performed, linkage scheduling instructions are obtained, and multi-device collaborative energy storage scheduling is performed. The application solves the technical problem in the prior art that wind power prediction mainly depends on traditional statistical models or simple machine learning models, cannot accurately capture the dynamic fluctuations of wind power, thereby affecting the decision of energy storage scheduling, and reducing the wind power utilization efficiency and power grid stability.

Description

Technical Field

[0001] This invention relates to the field of energy storage dispatching technology, specifically to energy storage dispatching methods, systems, equipment, and media based on wind power forecasting. Background Technology

[0002] Wind power generation is highly unstable and volatile, especially when wind speed varies significantly. Wind power output is often difficult to predict, posing a major challenge to grid dispatch and power system stability. To address this issue, energy storage technology is widely used to smooth wind power fluctuations and ensure stable grid operation. However, existing technologies still face many challenges, particularly in wind power prediction accuracy, energy storage dispatch optimization, and grid stability.

[0003] Existing wind power forecasting technologies mainly rely on traditional statistical models, such as autoregressive models, moving average methods, or simple machine learning models, such as linear regression and decision trees. These methods usually cannot accurately capture the dynamic fluctuations of wind power, especially under conditions of drastic wind speed changes or extreme weather, where the prediction error is large. This deviation will directly affect the decision-making of energy storage dispatch, resulting in inaccurate charging and discharging plans for energy storage devices. Consequently, the energy storage system cannot adequately cope with wind power fluctuations, cannot effectively suppress grid power fluctuations, and reduces wind power utilization efficiency and grid stability. Summary of the Invention

[0004] This application provides a wind power forecast-based energy storage dispatch method, system, equipment, and medium, aiming to solve the technical problem that existing wind power forecasting mainly relies on traditional statistical models or simple machine learning models, which cannot accurately capture the dynamic fluctuations of wind power, thereby affecting energy storage dispatch decisions and reducing wind power utilization efficiency and grid stability.

[0005] The first aspect disclosed in this application provides a wind power forecast-based energy storage scheduling method, the method comprising: backtracking and segmenting the original wind power time series data based on a preset forecast time window; performing multi-scale decomposition and denoising on the original wind power time series data to obtain multiple intrinsic mode function component sequences; pre-constructing a wind power hybrid forecasting model, wherein the wind power hybrid forecasting model performs global optimization of key hyperparameters of the long short-term memory network through a particle swarm optimization algorithm to complete the model construction; inputting the multiple intrinsic mode function component sequences into the wind power hybrid forecasting model for sequence prediction, and outputting an ultra-short-term wind power forecasting sequence; performing low-pass filtering-based smoothing on the ultra-short-term wind power forecasting sequence to obtain a baseline power curve; and in the process of local rolling optimization of the real-time energy storage scheduling strategy based on the baseline power curve, calling the associated forecasting power curve group according to the forecast start and end timestamps of the baseline power curve, performing multi-objective optimization of power coordination with the goal of smoothing power fluctuations at the grid connection point, obtaining a linkage scheduling instruction, and performing multi-device collaborative energy storage scheduling.

[0006] The second aspect of this application discloses an energy storage dispatch system based on wind power prediction. The system is used in the aforementioned wind power prediction-based energy storage dispatch method. The system includes: a raw data backtracking and segmentation module for backtracking and segmenting raw wind power time-series data based on a preset prediction time window; a multi-scale decomposition and denoising module for performing multi-scale decomposition and denoising on the raw wind power time-series data to obtain multiple intrinsic mode function component sequences; and a prediction model construction module for pre-constructing a hybrid wind power prediction model, wherein the hybrid wind power prediction model uses a particle swarm optimization algorithm to globally optimize key hyperparameters of a long short-term memory network to complete model construction. The prediction module is used to input the multiple intrinsic mode function component sequences into the wind power hybrid prediction model for sequence prediction and output an ultra-short-term wind power prediction sequence. The smoothing module is used to perform low-pass filtering-based smoothing on the ultra-short-term wind power prediction sequence to obtain a reference power curve. The rolling optimization module is used to call the associated predicted power curve group according to the prediction start and end timestamps of the reference power curve during the local rolling optimization of the real-time energy storage scheduling strategy based on the reference power curve, and perform multi-objective optimization of power coordination with the goal of smoothing power fluctuations at the grid connection point to obtain linkage scheduling instructions and perform multi-device collaborative energy storage scheduling.

[0007] The third aspect disclosed in this application provides a computer device, including a memory and a processor, wherein the memory stores a computer program, and the processor executes the computer program to implement the steps of the energy storage scheduling method based on wind power forecasting in the first aspect.

[0008] The fourth aspect disclosed in this application provides a storage medium having a computer program stored thereon, which, when executed by a processor, implements the steps of the energy storage scheduling method based on wind power forecasting in the first aspect.

[0009] One or more technical solutions provided in this application have at least the following beneficial effects:

[0010] By retrospectively analyzing wind power time-series data and dividing it into fixed time windows, the timeliness and representativeness of data acquisition can be ensured, providing time-series data for subsequent prediction models and enabling accurate predictions. Multi-scale decomposition extracts key signal components from wind power data, removing unnecessary noise and high-frequency interference, making the data clearer and providing cleaner input features for subsequent wind power prediction models, thus improving prediction accuracy. Optimizing the hyperparameters of the Long Short-Term Memory (LSTM) network model using particle swarm optimization (PSO) ensures the adaptability and efficiency of the wind power prediction model across different prediction tasks. PSO globally searches for optimal hyperparameters, avoiding local optima that may result from traditional manual hyperparameter adjustments, thereby constructing a more accurate hybrid wind power prediction model. The LSM network model calculates ultra-short-term wind power prediction sequences through forward propagation. Because LSM networks excel at processing time-series data, they can effectively capture the temporal changes in wind power. The order-dependent ultra-short-term forecast sequence output can provide the grid dispatch with the future short-term wind power change trend, helping dispatch decisions to respond better. Low-pass filtering removes high-frequency noise from the forecast sequence and smooths the forecast curve, which helps reduce short-term wind power fluctuations caused by sudden events, resulting in a more stable final reference power curve. The reference power curve provides a reliable forecast basis for energy storage dispatch, avoiding unnecessary responses caused by noise. Based on the reference power curve, local rolling optimization of the real-time energy storage dispatch strategy is performed, and associated forecast power curve groups are called according to the start and end timestamps of the power forecast. This rolling optimization ensures that the dispatch of the energy storage system can be adjusted according to the latest wind power changes. Multi-objective optimization of power coordination not only ensures the efficient operation of the energy storage system but also smooths power fluctuations at the grid connection point, improving grid stability. Multi-device collaborative energy storage dispatch ensures effective cooperation between various energy storage devices, improving overall energy utilization efficiency.

[0011] The above description is only an overview of the technical solution of this application. In order to better understand the technical means of this application and to implement it in accordance with the contents of the specification, and to make the above and other objects, features and advantages of this application more obvious and understandable, the following are specific embodiments of this application. Attached Figure Description

[0012] Figure 1This is a schematic diagram of the energy storage scheduling method based on wind power forecasting provided in the embodiments of this application.

[0013] Figure 2 A schematic diagram of the structure of an energy storage dispatch system based on wind power forecasting provided in an embodiment of this application.

[0014] Figure 3 This is a schematic diagram of the structure of an exemplary computer device provided in an embodiment of this application.

[0015] Figure labeling: 10 raw data backtracking segmentation module, 20 multi-scale decomposition and denoising module, 30 prediction model construction module, 40 sequence prediction module, 50 smoothing processing module, 60 rolling optimization module, 21 processor, 22 memory, 23 input device, 24 output device. Detailed Implementation

[0016] To further illustrate the technical means and effects of the present invention in achieving its intended purpose, the following detailed description of the specific implementation methods, structures, features and effects of the present invention, in conjunction with the accompanying drawings and preferred embodiments, is provided below.

[0017] Example 1, as Figure 1 As shown in the embodiments of this application, an energy storage dispatch method based on wind power forecasting is provided, the method comprising:

[0018] The original wind power time series data is back-segmented based on a preset prediction time window.

[0019] A preset forecast time window refers to the time range used for backtracking and segmenting time-series data during wind power forecasting. For example, it can be set to 24 hours, 48 ​​hours, etc., depending on the forecasting requirements and system design. Based on the set time window, data is extracted from the historical wind power time-series database. Backtracking means collecting wind power data within a specified time period from the database. For example, if a 24-hour time window is set, wind power data from the past 24 hours is extracted and processed as raw data. This data is then segmented into multiple time-series segments, each segment being a data sequence within a time window. This backtracked wind power data is used in subsequent modeling and forecasting processes. Each data segment represents the wind power change trend within a certain time period, used for model training and prediction.

[0020] The original wind power time series data is decomposed and denoised using a multi-scale method to obtain multiple intrinsic mode function component sequences.

[0021] The raw wind power time series data is decomposed into multiple scales using wavelet transform or other similar decomposition techniques. The aim is to break the data down into signals at multiple different scales, which helps extract variation patterns at different frequencies. Wavelet transform decomposes the original signal into multiple frequency levels, and then the CEEMDAN algorithm (Complete Ensemble Empirical Mode Decomposition) is used to suppress noise and eliminate mode aliasing in these signals, ensuring the removal of unwanted high-frequency noise while retaining low-frequency information useful for prediction. After denoising, the resulting sequences of multiple intrinsic mode functions (IMFs) represent the essential characteristics of the wind power data at different scales and frequencies. These denoised IMF sequences will serve as input features for the model to capture key patterns of wind power variation.

[0022] A pre-constructed hybrid wind power prediction model is used, wherein the hybrid wind power prediction model is constructed by globally optimizing the key hyperparameters of the long short-term memory network through the particle swarm optimization algorithm.

[0023] A hybrid wind power prediction model combines multiple prediction methods to improve prediction accuracy. Long Short-Term Memory (LSTM) networks are the primary neural network architecture, exhibiting excellent performance in time series prediction and capable of handling data with strong time dependencies. Particle Swarm Optimization (PSO) is an optimization method that simulates the foraging behavior of bird flocks, searching for the optimal solution within a given search space. In this step, PSO is used to globally optimize the hyperparameters of the LSM network, such as the number of network layers, learning rate, and number of neurons. Specifically, the initial positions of the particles are first set, with each particle representing a hyperparameter combination of the LSM network. The hyperparameter configuration of each particle is trained, and the prediction error of the model is calculated. Based on the historical and global optimal positions of the particles, the positions of the particles are iteratively updated until the optimal hyperparameter combination is found. Through PSO optimization, the hyperparameters of the LSM network are precisely tuned to find the optimal configuration. The optimized model has stronger predictive capabilities and can better adapt to the needs of wind power prediction.

[0024] The multiple intrinsic mode function component sequences are input into the wind power hybrid prediction model for sequence prediction, and the ultra-short-term wind power prediction sequence is output.

[0025] Multiple intrinsic mode function (EMF) component sequences are fed as input features into a hybrid wind power prediction model. The model is trained and inferred from these input sequences to predict wind power values ​​over a future period. The ability of the Long Short-Term Memory (LSTM) network lies in its ability to capture the temporal dependencies and long-term trends in time-series data, thereby outputting ultra-short-term wind power prediction sequences, such as those ranging from a few minutes to a few hours. These predictions serve as the basis for subsequent energy storage scheduling, providing expected values ​​for future wind power trends.

[0026] The ultra-short-term wind power prediction sequence is smoothed using a low-pass filter to obtain a baseline power curve.

[0027] Ultra-short-term wind power forecast sequences contain high-frequency noise or short-term fluctuations. These fluctuations do not represent the long-term trend of wind power. Therefore, low-pass filtering is used to remove these high-frequency components while retaining the long-term trend of the low-frequency components. The basic principle of low-pass filtering is to allow low-frequency signals to pass through while attenuating high-frequency signals to eliminate noise caused by sudden events or short-term fluctuations. After low-pass filtering, a smooth wind power forecast curve is obtained, called the baseline power curve. This baseline curve better reflects the overall trend of wind power, rather than simply short-term fluctuations. This baseline power curve is one of the key inputs for energy storage dispatch decisions. It represents the expected wind power output over a future period and serves as a benchmark for formulating dispatch strategies.

[0028] During the local rolling optimization of the real-time energy storage scheduling strategy based on the benchmark power curve, the associated predicted power curve group is called according to the predicted start and end timestamps of the benchmark power curve. The power coordination multi-objective optimization with the goal of smoothing power fluctuations at the grid connection point is executed to obtain the linkage scheduling instruction and carry out multi-device collaborative energy storage scheduling.

[0029] Local rolling optimization is a dynamic optimization method that adjusts the system based on the latest forecast results at each time point, ensuring the system can respond flexibly to actual conditions. In this process, the baseline power curve serves as one of the main inputs, and adjustments are made within each forecast window based on this curve to ensure optimal balance between wind power output and energy storage dispatch. Based on the start and end timestamps of the baseline power curve's forecast, multiple associated sets of forecast power curves are invoked. These sets contain power forecast information for different time periods, providing the energy storage system with reference at various points in time. For example, a suitable set of forecast curves can be selected based on the wind power forecast sequence for the day to address potential power fluctuations in the future.

[0030] In the process of energy storage dispatch, in addition to ensuring power balance, it is also necessary to consider the suppression of power fluctuations at the grid connection point. At this time, multi-objective optimization is performed, one of the objectives of which is to reduce the power fluctuations that may occur when wind power is connected to the grid. The optimization objectives include not only the charging and discharging scheduling of energy storage devices, but also the stability and power balance of the grid, and the coordination of cooperation between different energy storage devices to make the overall system power output stable and reliable.

[0031] After multi-objective optimization calculations, coordinated dispatch commands are output to control the charging or discharging processes of different energy storage devices, ensuring that all energy storage devices work collaboratively under the same objective to complete the energy storage dispatch task. These coordinated dispatch commands facilitate the collaborative work of multiple energy storage devices to minimize power fluctuations and maximize grid stability. The ultimate goal is to achieve multi-device coordinated energy storage dispatch, meaning that multiple energy storage devices can work collaboratively according to dispatch commands to minimize power fluctuations at the grid connection point during wind power fluctuations. This coordinated dispatch not only improves the stability of wind power grid connection but also optimizes the utilization efficiency of energy storage devices, ensuring the overall effectiveness of the grid and energy storage system.

[0032] Furthermore, the method involves backtracking and segmenting the original wind power time-series data based on a preset prediction time window, and includes:

[0033] The forward inference time consumption test of the wind power hybrid prediction model is performed to obtain the model calculation delay parameters; the prediction time window is set by fusing the grid dispatch response delay and the model calculation delay parameters based on the preset weighted rules; the prediction time window is used as the sliding window length to collect back-tracking wind power time series data in the historical wind power time series database to segment and obtain the original wind power time series data.

[0034] Forward inference refers to feeding input data into a pre-trained wind power hybrid prediction model, performing one inference or prediction, and obtaining the model's output. In this process, the latest wind power data is used for prediction, outputting the predicted value of ultra-short-term wind power. During forward inference, the time required for the inference process is recorded. The main purpose of this process is to evaluate the computational latency of the model in practical applications, i.e., the time required for the model to complete the prediction task. The measured latency parameter reflects the model's computational performance and is used to evaluate whether it can meet real-time prediction requirements. Latency is a key factor, especially in power system dispatching. After multiple forward inference time tests, the average time required for each inference is calculated, and the model's computational latency parameter is derived. This parameter is used for system optimization and dispatching strategy formulation.

[0035] In power system dispatching, setting the prediction time window is a crucial factor. To reasonably set the prediction time window, both grid dispatching response delay and model calculation delay need to be considered. Grid dispatching response delay refers to the time required from receiving the instruction from the grid to executing the corresponding dispatch; model calculation delay refers to the time required for the wind power prediction model to complete its prediction task. The grid dispatching response delay and model calculation delay are fused using a preset weighting rule. This rule weights these two delay parameters based on different factors, such as grid response speed, model calculation performance, and system requirements, resulting in a comprehensive delay parameter. For example, weights can be set based on the urgency of grid dispatching or the stability of calculation delay, making the fused delay parameter more consistent with actual needs. The length of the prediction time window is then set based on the fused delay parameter. The length of the prediction time window determines the data acquisition cycle and the lead time of the prediction model. For example, if the grid dispatching response is slow, a longer prediction time window is set to ensure the grid has sufficient time to respond to energy storage dispatching instructions; conversely, if the grid response is fast and the model calculation delay is short, a shorter prediction time window is used.

[0036] Using the prediction time window length as the sliding window length, the sliding window method means that historical wind power time-series data is sampled in units of this time window, with each sample representing a new time window, and wind power data is extracted from it. Based on the set sliding window, wind power time-series data is collected retrospectively; that is, according to the start time of the sliding window, wind power data for the corresponding time period is extracted from historical wind power data. By retrospectively collecting data, wind power variation data for different time periods can be obtained as input to the prediction model. Each collected wind power time-series data segment represents the wind power value for a specific time period. These segments are divided into independent time-series data for subsequent processing and analysis.

[0037] Furthermore, the method for pre-constructing a hybrid wind power prediction model includes:

[0038] After using the prediction time window to slide and segment multiple consecutive training sample sequences in the historical wind power time series database, multi-scale decomposition and denoising are performed to obtain multiple sets of sample mode function component sequences. Based on the end timestamps of the multiple consecutive training sample sequences, measured power is extracted from the historical wind power time series database to obtain multiple target power sequences. The multiple sets of sample mode function component sequences are used as input features, and the corresponding multiple target power sequences are used as prediction labels to form a training dataset. On the prediction model architecture built based on a long short-term memory network, the training dataset is used to train the model, and the particle swarm optimization algorithm is used to perform global search optimization of the key hyperparameters of the model to complete the construction of the wind power hybrid prediction model.

[0039] Using a set prediction time window, the historical wind power time series data is segmented by a sliding window. Each window corresponds to an independent time period. The wind power data for each time period is extracted from the historical data and used as a training sample. These continuous training sample sequences will be used as the subsequent training dataset for the construction of the wind power hybrid prediction model.

[0040] After sliding segmentation, multi-scale decomposition denoising is performed on each training sample sequence. The main purpose is to remove high-frequency noise from the original time-series data and extract more meaningful low-frequency trends. Multi-scale decomposition methods, such as wavelet transform or empirical mode decomposition, are used to decompose the data into multiple sub-signals (intrinsic mode functions, IMFs) of different frequencies. These IMFs represent the variation characteristics of the data at different scales. The purpose of denoising is to remove irrelevant noise and retain signal components with predictive significance. The denoised signal can be more effectively used as input features for the model, thereby improving prediction accuracy. After multi-scale decomposition denoising, multiple intrinsic mode function component sequences are obtained, which represent the variations of wind power data at different frequencies and time scales.

[0041] Based on the end timestamp of each consecutive training sample sequence, measured power data corresponding to these timestamps are extracted from the historical wind power time-series database. This extracted measured power data serves as the target power sequence, representing the actual power output of real wind farms over past time periods, and is used as the target variable for model prediction. These target power sequences will be used to train the wind power prediction model, enabling the model to learn how to accurately predict wind power output from the input intrinsic mode function component sequence.

[0042] Multiple sets of sample mode function component sequences are used as input features of the model. These component sequences represent the variations of wind power time series data at different frequencies and time scales. The target power sequence is the actual wind power output corresponding to these input features. They serve as prediction labels for the model, guiding it to learn wind power prediction. Each set of intrinsic mode function component sequences corresponds to a target power sequence. Therefore, each input feature (mode component sequence) and prediction label (target power sequence) forms a paired data pair. These paired data are provided as training datasets to the model for learning. Through these paired data, the model can learn the relationship between the input features and the wind power output.

[0043] Long Short-Term Memory (LSTM) networks are recurrent neural network architectures suitable for time-series data. They possess strong time-dependency modeling capabilities, capturing patterns and trends in long-term series. In wind power forecasting, LTM networks are well-suited for handling the temporal nature of wind power data. In this step, a wind power forecasting model architecture is built upon the LTM network. This architecture will be trained using a training dataset to learn how to predict future wind power from the modal function component sequence. The model training process is a crucial step in the LTM network's learning of wind power forecasting. Through training, the LTM network gradually adjusts its parameters, learning how to predict the target power sequence from the input modal component sequence. During training, the model continuously optimizes its internal weights until the prediction error is minimized and the model can accurately predict wind power.

[0044] Particle Swarm Optimization (PSO) is a global optimization method that finds the optimal solution to a problem by simulating swarm behavior. In this step, PSO is used to optimize key hyperparameters of Long Short-Term Memory (LSTM) networks, such as the number of layers, neurons, and learning rate. Specifically, PSO first initializes a swarm of particles, with each particle representing a set of possible hyperparameter configurations. For each hyperparameter configuration, it is applied to the LSM network, and the model's prediction error is calculated, for example, by cross-validation. Based on the particle's historical best position and global best position, PSO iteratively updates the particle's hyperparameter configuration, searching the entire search space for the optimal combination of hyperparameters to maximize the LSM network's predictive power. Once PSO finds the optimal hyperparameter combination, the LSM network uses these hyperparameters for final training and adjustment, resulting in a model with better predictive capabilities and the ability to accurately predict wind power.

[0045] Furthermore, based on the prediction model architecture constructed using a long short-term memory network, the model is trained using the aforementioned training dataset, and a particle swarm optimization algorithm is employed to globally search and optimize the key hyperparameters of the model, thereby completing the construction of the wind power hybrid prediction model. The method includes:

[0046] A particle swarm is initialized in a pre-constructed key hyperparameter search space, wherein the key hyperparameter search space is subject to hyperparameter value constraints and iteration number constraints. K sets of candidate hyperparameter values ​​corresponding to K initial space particles in the initialized particle swarm are configured into the prediction model architecture, and the model is trained using the training dataset to obtain K prediction errors. Using the hyperparameter value constraints and iteration number constraints as boundary conditions, the K prediction errors are used as the fitness values ​​of the K initial space particles. Based on the individual particle's historical best position and the swarm's global best position, the particle positions are iteratively updated within the key hyperparameter search space until the fitness values ​​satisfy a preset convergence condition, thus obtaining the target key hyperparameter combination. The target key hyperparameter combination is then used to construct the wind power hybrid prediction model.

[0047] When performing particle swarm optimization (PSO), a key hyperparameter search space is first defined. This search space encompasses all possible hyperparameter values ​​and is used to optimize the performance of the Long Short-Term Memory (LSTM) network. Hyperparameters include the learning rate, the number of LSM layers, the number of neurons, the batch size, the number of training epochs, and the regularization coefficient. The definition of the search space is set according to actual needs and computational resource constraints. Hyperparameter value constraints limit the range of each hyperparameter value. For example, the learning rate is usually set between 0 and 0.1, because an excessively large learning rate can lead to gradient explosion, while an excessively small learning rate can slow down training. The number of LSM layers and the number of neurons per layer also have certain range limitations; exceeding these ranges may lead to model overfitting or wasted computational resources. In the PSO algorithm, the iteration count constraint refers to the maximum number of iterations in the PSO optimization. For example, setting the maximum number of iterations to 1000 limits the algorithm's runtime during optimization, ensuring that the search does not proceed indefinitely when computational resources are limited. Based on the predefined hyperparameter search space and iteration number constraints, a particle swarm is initialized. The particle swarm consists of multiple particles, each representing a hyperparameter combination. The initial position of each particle (i.e., its hyperparameter value) is random, but must be within the constraints of the search space. Each particle also has a fitness value, which measures the performance of the hyperparameter combination it represents in the current model training.

[0048] After initializing the particle swarm, K particles are selected from it, where K represents the number of particles. Each particle corresponds to a set of candidate hyperparameter values. For each particle, its corresponding hyperparameters are configured into the architecture of the prediction model, effectively constructing a Long Short-Term Memory (LSTM) network model and using these configurations for training. The training dataset consists of the previously mentioned sample mode function component sequences and target power sequences, used to train each LSM network model based on different hyperparameter configurations. Each particle's hyperparameter configuration corresponds to an independent model training process. The training process optimizes the parameters of the LSM network model using the backpropagation algorithm until a specified number of training epochs is reached, indicating training is complete. The goal of training is to enable each model to predict wind power based on the input mode function component sequences.

[0049] After training for each particle, the model's performance is evaluated by calculating the prediction error to measure the prediction accuracy of each model. Error metrics include mean squared error (MSE) and root mean square error (RMSE). Each particle's corresponding Long Short-Term Memory (LSTM) network model will receive a prediction error value; the smaller the prediction error value, the stronger the model's predictive ability.

[0050] K particles acquire fitness values ​​based on their corresponding prediction errors. The smaller the prediction error, the larger the particle's fitness value, indicating that the hyperparameter combination is more effective for the prediction task. The individual particle's historical best position refers to the best-performing hyperparameter combination among all positions the particle has experienced during the optimization process; the swarm's global best position refers to the best-performing hyperparameter combination across the entire particle swarm during the optimization process. During particle swarm optimization, particles adjust their positions (hyperparameter configurations) based on these two pieces of information. Even if the prediction error of the current particle position is large, the particle will adjust towards a better direction. Particles are iteratively updated using the particle swarm optimization algorithm. In each iteration, the particle's position (hyperparameter configuration) is adjusted based on its individual historical best position and the global best position. The particle's position update formula combines the particle's current velocity, individual best position, and swarm best position to determine the particle's new position in the next iteration. The updated position represents the new hyperparameter combination.

[0051] In the particle swarm optimization process, iterations continue until a preset convergence condition is reached. The convergence condition includes: an error threshold (if the fitness value of the particle swarm (i.e., the prediction error) meets the preset threshold, the optimization process is considered converged); and a maximum number of iterations (if the convergence condition is not met after a certain number of iterations, the optimization process stops). Ultimately, the particle swarm optimization algorithm will converge to an optimal combination of hyperparameters. This combination is the target key hyperparameter combination, which enables the Long Short-Term Memory (LSTM) network model to achieve the best performance in wind power prediction tasks, i.e., the lowest prediction error.

[0052] The final hybrid wind power prediction model is constructed by applying the key hyperparameter combination of this objective. This means setting the various hyperparameters of the Long Short-Term Memory (LSTM) network model, such as the learning rate, the number of LSM layers, and the number of neurons, to the optimal values ​​obtained through particle swarm optimization. The final model can accurately predict future wind power based on wind power time-series data, i.e., the sequence of input modal function components. The optimized hyperparameters enable the model to better fit the data, reduce prediction errors, and improve the accuracy of wind power prediction.

[0053] Furthermore, the method involves inputting the multiple intrinsic mode function component sequences into the wind power hybrid prediction model for sequence prediction, and outputting an ultra-short-term wind power prediction sequence.

[0054] The multiple intrinsic mode function component sequences are organized into an input feature tensor according to a preset time step; the input feature tensor is loaded into the wind power hybrid prediction model, and forward propagation calculation is performed through a long short-term memory network to obtain a preliminary power prediction value; the preliminary power prediction value is inversely normalized to obtain the ultra-short-term wind power prediction sequence.

[0055] Multiple intrinsic mode function (IMF) component sequences are organized according to a preset time step. Each IMF component sequence represents the variation of wind power data at different frequencies and time scales. These sequences will be used as input features of the model. The time step refers to the length of the multiple IMF component sequences into several time segments. For example, assuming the time step is set to 1 hour, the multiple IMF component sequences are divided into multiple 1-hour data windows, and these slices are combined into an input feature tensor.

[0056] The organized input feature tensor is fed into the wind power hybrid prediction model. This input feature tensor contains data from each time step; therefore, the Long Short-Term Memory (LSTM) network can leverage its powerful time-dependency modeling capabilities to identify trends and patterns in the time series. Forward propagation refers to processing the input data through the various layers of the LTM network to obtain the model's output. During forward propagation, the LTM network effectively retains long-term memory and learns short-term dependencies in the time series through its gating mechanisms, including forget gates, input gates, and output gates, to obtain preliminary power prediction values.

[0057] During model training, the input data is normalized to facilitate model learning and accelerate convergence. Normalization scales the data within a certain range, such as 0 to 1, making the data more suitable for model training. However, the model's output is the normalized result. Therefore, the initial power prediction values ​​need to be denormalized to restore the predicted values ​​from the normalized range to the original actual range of wind power. The denormalization process uses the minimum and maximum values ​​or the standardized mean and standard deviation used during normalization to restore the normalized prediction values ​​to the original wind power units. The result of denormalization is the final wind power prediction sequence, called the ultra-short-term wind power prediction sequence. This prediction sequence represents the predicted wind power changes over a future period. Because of the denormalization, the final prediction values ​​can be directly used for power system scheduling and optimization.

[0058] Furthermore, during the local rolling optimization of the real-time energy storage scheduling strategy based on the benchmark power curve, the associated predicted power curve group is invoked according to the predicted start and end timestamps of the benchmark power curve. Multi-objective optimization for power coordination, aimed at smoothing power fluctuations at the grid connection point, is performed to obtain a coordinated scheduling instruction for multi-device collaborative energy storage scheduling. The method includes:

[0059] By comparing the measured power curve with the reference power curve, a power prediction deviation sequence is obtained; the real-time state of charge of the energy storage module is obtained interactively; the power prediction deviation sequence is constrained and corrected using the real-time state of charge to obtain a real-time charge and discharge power command; and the real-time energy storage scheduling strategy is iterated with low-latency feedback based on the real-time charge and discharge power command.

[0060] The measured power curve is obtained by measuring the actual power output of the wind farm in real time. It reflects the actual changes in wind power and is acquired through real-time data collected from the power grid by monitoring equipment. The reference power curve is a wind power prediction sequence generated by a prediction model, representing the system's expectation of future wind power changes. The power prediction deviation sequence is calculated by comparing the measured power curve with the reference power curve, determining the deviation at each time point. The deviation represents the difference between the actual and predicted power and reflects the accuracy of the prediction model and the need for real-time adjustments.

[0061] Real-time state of charge (SOC) indicates the current charge level of an energy storage module, expressed as a percentage, ranging from 0% (fully discharged) to 100% (fully charged). This information is crucial for energy storage scheduling, as the charging and discharging capabilities of energy storage devices are limited by the SOC. This information is obtained by interacting with the energy storage module through monitoring devices, sensors, or communication protocols.

[0062] The power prediction deviation sequence is constrained and corrected using real-time state of charge (SOC). Specifically, if there is a large deviation in the prediction—that is, a significant difference between the measured power and the reference power—appropriate charging or discharging operations are required based on the current SOC to compensate for the power deviation. For example, if the measured power is higher than the predicted power, a discharging operation is needed to provide additional power; if the measured power is lower than the predicted power, charging is needed to prepare for future power fluctuations. Real-time charging and discharging power commands are generated based on the corrected power prediction deviation sequence. These commands instruct the energy storage module how much charging or discharging operation is required to eliminate power prediction deviations and ensure power balance in the grid.

[0063] The real-time energy storage dispatch strategy is based on low-latency feedback iteration using real-time charging and discharging power commands. This means that whenever a new power prediction deviation or energy storage state of charge update is received, the dispatch strategy is rapidly adjusted to ensure a quick response to dispatch commands. Low latency refers to the response speed to power fluctuations; it is necessary to minimize power fluctuations caused by prediction deviations and make rapid charging and discharging adjustments to ensure stable grid operation. In each control cycle, such as every minute or hour, new power prediction deviations and state of charge information are continuously received, and the energy storage dispatch strategy is updated based on this information. Through continuous feedback and adjustment, the charging and discharging operation of energy storage devices can be optimized in real time to maintain stable grid power.

[0064] Furthermore, the power prediction deviation sequence, obtained by comparing the measured power curve and the baseline power curve, represents the difference between the measured and predicted power. This sequence reflects the prediction error of the model at different time points and contains signals of potential wind turbine faults or abnormal fluctuations. Multidimensional feature extraction is performed on the power prediction deviation sequence, extracting information from multiple dimensions, such as statistical features like mean, variance, standard deviation, kurtosis, and skewness, which reflect the distribution of the deviation sequence; frequency domain features, using methods like Fourier transform or wavelet transform to convert the time series into frequency domain features to capture periodic or sudden changes in the deviation sequence; time domain features, using methods like sliding window to extract trends, periodicity, and local fluctuations; and higher-order features, such as autocorrelation, mutual information, and nonlinear relationships, which reveal deeper patterns in the sequence. These extracted features are combined into a deviation feature vector, representing various types of information in the deviation sequence, used for wind turbine fault detection.

[0065] A pre-built wind power fault classification model is used. This model is a trained machine learning or deep learning model specifically designed to identify potential fault types in wind power equipment based on input bias feature vectors. Classification models include Support Vector Machines, Random Forests, Convolutional Neural Networks, or Long Short-Term Memory Networks. The bias feature vectors are input into the pre-built model. Based on patterns and rules learned during training, the model analyzes the current bias features and outputs a prediction result for each possible fault type. Each prediction result is accompanied by a confidence value, representing the model's confidence in predicting that fault type. For example, the model outputs confidence scores for three fault types: one for equipment faults and its confidence score, one for system faults and its confidence score, and so on.

[0066] A two-dimensional retrieval and matching process using fault type prediction results and fault confidence levels is employed to determine whether a fault warning should be triggered. Specifically: the fault type indicates the predicted type of fault, and this result is used to determine whether this type of fault requires attention; the confidence level represents the model's confidence in the prediction result. A high confidence level indicates that the model has a strong grasp of the reliability of the prediction result, and a more urgent warning can be triggered. Matching rules determine whether a wind power equipment fault warning should be triggered. For example, if the fault type is equipment fault and its confidence level exceeds a preset threshold, such as 80%, an equipment fault warning is triggered; if the fault type is system fault and its confidence level is below a preset threshold, such as 50%, no warning is triggered, or a lower-priority alarm is triggered. Fault warnings include specific operational instructions, such as scheduling energy storage equipment for charging or discharging, or adjusting the operating mode of wind power equipment, to ensure timely response when potential equipment faults occur and to reduce the impact of the fault.

[0067] Furthermore, the original wind power time series data is subjected to multi-scale decomposition and denoising to obtain multiple intrinsic mode function component sequences. The method includes:

[0068] Wavelet transform is used to perform coarse-grained multi-scale decomposition on the cleaned original wind power time series data to obtain multiple heterogeneous frequency feature signal sequences; the CEEMDAN algorithm is used to perform fine-grained decomposition and reconstruction iteration of the multiple heterogeneous frequency feature signal sequences for mode aliasing and noise suppression, and output the multiple intrinsic mode function component sequences.

[0069] Wavelet transform is a signal analysis tool widely used for multi-scale analysis of time-series data. It decomposes signals into different frequency components and provides local features at different scales, making it suitable for processing non-stationary signals. A key advantage of wavelet transform is its ability to perform simultaneous time and frequency analysis, making it particularly suitable for processing time-varying data such as wind power. In this step, coarse-grained multi-scale decomposition means using wavelet transform to decompose the original wind power time-series data into signals at multiple scales (i.e., frequencies), each scale representing a different frequency component. Through multi-scale decomposition, the data is divided into low-frequency (coarse-grained) and high-frequency (fine-grained) components. The low-frequency components represent the signal's trend and overall fluctuations, while the high-frequency components reflect rapid changes and noise. The result of the wavelet transform is multiple heterogeneous frequency feature signal sequences, each corresponding to a specific frequency component, representing the changes in wind power data at different time scales. These heterogeneous frequency signal sequences provide multi-dimensional data for subsequent processing, helping to more comprehensively capture the time-series characteristics of wind power data.

[0070] CEEMDAN (Complete Ensemble Empirical Mode Decomposition) is an improvement on EMD (Empirical Mode Decomposition), employing noise suppression and mode separation techniques to avoid mode aliasing caused by noise in traditional EMD algorithms. It can decompose signals more accurately, aiming to extract multiple intrinsic mode functions (EMFs) from a signal while overcoming the mode aliasing problem in EMD. Mode aliasing refers to the incorrect mixing of signal components of different frequencies during the decomposition process, making it difficult to analyze the signal correctly. CEEMDAN avoids this problem by introducing adaptive noise and performing multiple iterations to gradually separate different signal components. Noise suppression refers to removing the influence of noise during the signal decomposition process, thereby ensuring that the decomposed EMFs more accurately reflect the true characteristics of the signal.

[0071] The CEEMDAN algorithm employs fine-grained decomposition, extracting different frequency components of the signal through more refined iterative decomposition. Each round of decomposition outputs a set of intrinsic mode functions (EMFs), each representing a local fluctuation pattern in the signal. These iterative processes continue until all major components of the signal have been extracted. The output of each iteration step is a sequence of EMF components, representing variations in the signal at different frequencies and time scales. Ultimately, the CEEMDAN algorithm's decomposition process yields multiple EMF component sequences. These sequences reflect the local characteristics of wind power data at different frequencies and time scales, providing useful information for subsequent wind power prediction.

[0072] Example 2 is based on the same inventive concept as the wind power forecast-based energy storage dispatch method in the previous examples, such as... Figure 2As shown in the embodiment of this application, an energy storage dispatch system based on wind power forecasting is provided, the system comprising:

[0073] The original data backtracking and segmentation module 10 is used to backtrack and segment the original wind power time series data based on a preset prediction time window; the multi-scale decomposition and denoising module 20 is used to perform multi-scale decomposition and denoising on the original wind power time series data to obtain multiple intrinsic mode function component sequences; the prediction model construction module 30 is used to pre-build a wind power hybrid prediction model, wherein the wind power hybrid prediction model uses a particle swarm optimization algorithm to perform global optimization of key hyperparameters of a long short-term memory network to complete the model construction; the sequence prediction module 40 is used to input the multiple intrinsic mode function component sequences into the... A hybrid wind power prediction model performs sequence prediction and outputs an ultra-short-term wind power prediction sequence. A smoothing processing module 50 is used to smooth the ultra-short-term wind power prediction sequence based on low-pass filtering to obtain a baseline power curve. A rolling optimization module 60 is used to call the associated predicted power curve group according to the prediction start and end timestamps of the baseline power curve during the local rolling optimization of the real-time energy storage scheduling strategy, and perform multi-objective optimization of power coordination with the goal of smoothing power fluctuations at the grid connection point to obtain linkage scheduling instructions and perform multi-device collaborative energy storage scheduling.

[0074] Furthermore, the original data backtracking and segmentation module 10 is used to perform the following operation steps:

[0075] The forward inference time consumption test of the wind power hybrid prediction model is performed to obtain the model calculation delay parameters; the prediction time window is set by fusing the grid dispatch response delay and the model calculation delay parameters based on the preset weighted rules; the prediction time window is used as the sliding window length to collect back-tracking wind power time series data in the historical wind power time series database to segment and obtain the original wind power time series data.

[0076] Furthermore, the prediction model construction module 30 is used to perform the following operation steps:

[0077] After using the prediction time window to slide and segment multiple consecutive training sample sequences in the historical wind power time series database, multi-scale decomposition and denoising are performed to obtain multiple sets of sample mode function component sequences. Based on the end timestamps of the multiple consecutive training sample sequences, measured power is extracted from the historical wind power time series database to obtain multiple target power sequences. The multiple sets of sample mode function component sequences are used as input features, and the corresponding multiple target power sequences are used as prediction labels to form a training dataset. On the prediction model architecture built based on a long short-term memory network, the training dataset is used to train the model, and the particle swarm optimization algorithm is used to perform global search optimization of the key hyperparameters of the model to complete the construction of the wind power hybrid prediction model.

[0078] Furthermore, the prediction model construction module 30 is used to perform the following operation steps:

[0079] A particle swarm is initialized in a pre-constructed key hyperparameter search space, wherein the key hyperparameter search space is subject to hyperparameter value constraints and iteration number constraints. K sets of candidate hyperparameter values ​​corresponding to K initial space particles in the initialized particle swarm are configured into the prediction model architecture, and the model is trained using the training dataset to obtain K prediction errors. Using the hyperparameter value constraints and iteration number constraints as boundary conditions, the K prediction errors are used as the fitness values ​​of the K initial space particles. Based on the individual particle's historical best position and the swarm's global best position, the particle positions are iteratively updated within the key hyperparameter search space until the fitness values ​​satisfy a preset convergence condition, thus obtaining the target key hyperparameter combination. The target key hyperparameter combination is then used to construct the wind power hybrid prediction model.

[0080] Furthermore, the sequence prediction module 40 is used to perform the following operation steps:

[0081] The multiple intrinsic mode function component sequences are organized into an input feature tensor according to a preset time step; the input feature tensor is loaded into the wind power hybrid prediction model, and forward propagation calculation is performed through a long short-term memory network to obtain a preliminary power prediction value; the preliminary power prediction value is inversely normalized to obtain the ultra-short-term wind power prediction sequence.

[0082] Furthermore, the rolling optimization module 60 is used to perform the following operation steps:

[0083] By comparing the measured power curve with the reference power curve, a power prediction deviation sequence is obtained; the real-time state of charge of the energy storage module is obtained interactively; the power prediction deviation sequence is constrained and corrected using the real-time state of charge to obtain a real-time charge and discharge power command; and the real-time energy storage scheduling strategy is iterated with low-latency feedback based on the real-time charge and discharge power command.

[0084] Furthermore, the multi-scale decomposition and denoising module 20 is used to perform the following operation steps:

[0085] Wavelet transform is used to perform coarse-grained multi-scale decomposition on the cleaned original wind power time series data to obtain multiple heterogeneous frequency feature signal sequences; the CEEMDAN algorithm is used to perform fine-grained decomposition and reconstruction iteration of the multiple heterogeneous frequency feature signal sequences for mode aliasing and noise suppression, and output the multiple intrinsic mode function component sequences.

[0086] Through the foregoing detailed description of the wind power forecast-based energy storage dispatch method, those skilled in the art can clearly understand the wind power forecast-based energy storage dispatch system in this embodiment. Since it corresponds to the method disclosed in the embodiment, the description is relatively simple, and relevant parts can be referred to the method section.

[0087] Example 3, as Figure 3 The diagram shown is a schematic structural diagram of an exemplary computer device of this application, illustrating a block diagram of an exemplary electronic device suitable for implementing embodiments of the present invention. Figure 3 The computer device shown is merely an example and should not be construed as limiting the functionality or scope of the embodiments of the present invention. Figure 3 As shown, the computer device includes a processor 21, a memory 22, an input device 23, and an output device 24; the number of processors 21 in the electronic device can be one or more. Figure 3 Taking a processor 21 as an example, the processor 21, memory 22, input device 23, and output device 24 in an electronic device can be connected via a bus or other means. Figure 3 Taking the example of a connection between China and Israel via a bus.

[0088] Example 4 provides a storage medium on which a computer program is stored, which, when executed by a processor, implements any step of Example 1.

[0089] The above description is merely a preferred embodiment of the present invention and is not intended to limit the present invention in any way. Although the present invention has been disclosed above with reference to preferred embodiments, it is not intended to limit the present invention. Any person skilled in the art can make some modifications or alterations to the above-disclosed technical content to create equivalent embodiments without departing from the scope of the present invention. Any modifications, equivalent changes, and alterations made to the above embodiments based on the technical essence of the present invention without departing from the scope of the present invention shall still fall within the scope of the present invention.

Claims

1. A wind power forecasting-based energy storage dispatch method, characterized in that, The method includes: The original wind power time series data was back-segmented based on a preset prediction time window; The original wind power time series data is decomposed and denoised using a multi-scale method to obtain multiple intrinsic mode function component sequences. A pre-constructed wind power hybrid prediction model is used, wherein the wind power hybrid prediction model is constructed by globally optimizing the key hyperparameters of the long short-term memory network through the particle swarm optimization algorithm. The multiple intrinsic mode function component sequences are input into the wind power hybrid prediction model for sequence prediction, and the ultra-short-term wind power prediction sequence is output. The ultra-short-term wind power prediction sequence is smoothed using a low-pass filter to obtain a baseline power curve; During the local rolling optimization of the real-time energy storage scheduling strategy based on the benchmark power curve, the associated predicted power curve group is called according to the predicted start and end timestamp of the benchmark power curve, and a multi-objective optimization of power coordination with the goal of smoothing power fluctuations at the grid connection point is performed to obtain a linkage scheduling instruction and carry out multi-device collaborative energy storage scheduling. The method involves backtracking and segmenting the original wind power time-series data based on a preset prediction time window, and includes: The forward inference time consumption of the wind power hybrid prediction model was tested to obtain the model calculation delay parameters; The prediction time window is set based on the preset weighted rule fusion of the power grid dispatch response delay and the model calculation delay parameters; Using the predicted time window as the sliding window length, wind power time series data is collected back from the historical wind power time series database to segment and obtain the original wind power time series data; The method involves inputting the multiple intrinsic mode function component sequences into the wind power hybrid prediction model for sequence prediction, and outputting an ultra-short-term wind power prediction sequence. The multiple intrinsic mode function component sequences are organized into an input feature tensor according to a preset time step; The input feature tensor is loaded into the wind power hybrid prediction model, and forward propagation calculation is performed through a long short-term memory network to obtain the preliminary power prediction value. The preliminary power prediction values ​​are inversely normalized to obtain the ultra-short-term wind power prediction sequence.

2. The energy storage dispatch method based on wind power forecasting as described in claim 1, characterized in that, The method for pre-constructing a hybrid wind power prediction model includes: After using the prediction time window to slide and segment multiple consecutive training sample sequences in the historical wind power time series database, multi-scale decomposition and denoising are performed to obtain multiple sets of sample mode function component sequences. Based on the end timestamps of the multiple consecutive training sample sequences, measured power is extracted from the historical wind power time series database to obtain multiple target power sequences; The training dataset is constructed by taking the sequence of multiple sample mode function components as input features and the corresponding multiple target power sequences as prediction labels. Based on the prediction model architecture built on the Long Short-Term Memory network, the model is trained using the training dataset, and the particle swarm optimization algorithm is used to perform global search optimization of the key hyperparameters of the model to complete the construction of the wind power hybrid prediction model.

3. The energy storage dispatch method based on wind power forecasting as described in claim 2, characterized in that, Based on the prediction model architecture built on a long short-term memory network, the model is trained using the training dataset, and the particle swarm optimization algorithm is used to globally search and optimize the key hyperparameters of the model to complete the construction of the wind power hybrid prediction model. The method includes: The particle swarm is initialized in a pre-constructed key hyperparameter search space, wherein the key hyperparameter search space is subject to hyperparameter value constraints and iteration number constraints. After configuring the K sets of hyperparameter candidate values ​​corresponding to the K initial spatial particles in the initial particle swarm into the prediction model architecture, the model is trained using the training dataset to obtain K prediction errors. Using the hyperparameter value constraints and iteration number constraints as boundary conditions, the K prediction errors are used as the fitness values ​​of the K initial space particles. Based on the historical best position of the individual particles and the global best position of the group, the particle positions are iteratively updated in the key hyperparameter search space until the fitness values ​​meet the preset convergence conditions, thus obtaining the target key hyperparameter combination. The wind power hybrid prediction model is constructed using the combination of the target key hyperparameters.

4. The energy storage dispatch method based on wind power forecasting as described in claim 1, characterized in that, During the local rolling optimization of the real-time energy storage scheduling strategy based on the benchmark power curve, the associated predicted power curve group is called according to the predicted start and end timestamps of the benchmark power curve. Multi-objective optimization for power coordination, aimed at smoothing power fluctuations at the grid connection point, is performed to obtain a coordinated scheduling instruction. Multi-device collaborative energy storage scheduling is then carried out. The method includes: By comparing the measured power curve with the reference power curve, a power prediction deviation sequence is obtained; Interactively obtain the real-time state of charge of the energy storage module; The real-time state of charge is used to perform a constraint correction on the power prediction deviation sequence to obtain a real-time charge and discharge power command; The real-time energy storage scheduling strategy is iterated with low latency based on the real-time charging and discharging power command.

5. The energy storage dispatch method based on wind power forecasting as described in claim 1, characterized in that, The original wind power time series data is subjected to multi-scale decomposition and denoising to obtain multiple intrinsic mode function component sequences. The method includes: Wavelet transform is used to perform coarse-grained multi-scale decomposition on the cleaned original wind power time series data to obtain multiple heterogeneous frequency feature signal sequences. The CEEMDAN algorithm is used to perform fine-grained decomposition and reconstruction iteration of the multiple heterogeneous frequency feature signal sequences for mode aliasing and noise suppression, and outputs the multiple intrinsic mode function component sequences.

6. An energy storage dispatch system based on wind power forecasting, characterized in that, The system is used to implement the wind power forecast-based energy storage dispatch method according to any one of claims 1-5, the system comprising: The raw data backtracking and segmentation module is used to backtrack and segment the raw wind power time series data based on a preset prediction time window. The multi-scale decomposition and denoising module is used to perform multi-scale decomposition and denoising on the original wind power time series data to obtain multiple intrinsic mode function component sequences. The prediction model building module is used to pre-build a hybrid wind power prediction model, wherein the hybrid wind power prediction model is built by globally optimizing the key hyperparameters of the long short-term memory network through the particle swarm optimization algorithm. The sequence prediction module is used to input the multiple intrinsic mode function component sequences into the wind power hybrid prediction model for sequence prediction and output the ultra-short-term wind power prediction sequence. A smoothing module is used to perform low-pass filtering-based smoothing on the ultra-short-term wind power prediction sequence to obtain a reference power curve. The rolling optimization module is used to perform local rolling optimization of the real-time energy storage scheduling strategy based on the benchmark power curve. During this process, it calls the associated predicted power curve group according to the predicted start and end timestamps of the benchmark power curve, performs multi-objective optimization for power coordination with the goal of smoothing power fluctuations at the grid connection point, obtains linkage scheduling instructions, and performs multi-device collaborative energy storage scheduling.

7. A computer device comprising a memory and a processor, wherein the memory stores a computer program, characterized in that, The processor executes the computer program to implement the steps of the energy storage scheduling method based on wind power forecasting as described in any one of claims 1 to 5.

8. A storage medium having a computer program stored thereon, characterized in that, When the computer program is executed by the processor, it implements the steps of the energy storage scheduling method based on wind power forecasting as described in any one of claims 1 to 5.

Images