Control method of distributed wind, light, storage output prediction model with scene self-adaptation capability

Abstract

The present application relates to power system scheduling and new energy output prediction technical field, especially to a kind of control method of distributed wind, light, storage output prediction model with scene adaptive capacity, including empirical mode decomposition, high-frequency IMF component wavelet threshold denoising, signal reconstruction, hybrid prediction modeling, energy storage system power regulation strategy.The present application introduces empirical mode decomposition and wavelet threshold denoising method, and the leading component in wind light power is extracted and high-frequency disturbance is suppressed;In the aspect of prediction modeling, the long short-term memory network LSTM is used to capture time series trend, and the KAN network is introduced to nonlinearly map high-dimensional features, to build a series-parallel double-channel collaborative prediction architecture, to realize full-scale modeling from trend to disturbance;At the system control level, combined with the prediction error feedback mechanism, the energy storage system is guided to carry out dynamic charging and discharging response, effectively suppresses wind and light output fluctuation, improves the controllability and actual availability of prediction results.

Description

Control methods for distributed wind, solar, and energy storage power output prediction models with scene adaptability Technical Field

[0001] This invention relates to the fields of power system dispatching and new energy output prediction technology, and in particular to a control method for a distributed wind, solar and energy storage output prediction model with scene adaptive capability. Background Technology

[0002] With the transformation and upgrading of the energy structure, new energy power generation technologies such as wind power and photovoltaics have developed rapidly, and their share in the global electricity production structure continues to rise. However, wind and solar power output is significantly affected by natural conditions, exhibiting obvious nonlinearity, non-stationarity, and strong randomness, posing unprecedented challenges to power system dispatch and stable operation. Therefore, improving the accuracy, robustness, and scenario adaptability of wind and solar power prediction has become a core challenge in current research on new energy grid integration.

[0003] Currently, wind and solar power forecasting methods can be classified from multiple dimensions, including forecast time scale, modeling method, and result representation. From a time scale perspective, they can be divided into medium-to-long-term, short-term, and ultra-short-term forecasts: medium-to-long-term forecasts support planning and scheduling, short-term forecasts serve the day-ahead market and load tracking, while ultra-short-term forecasts provide real-time input for grid operation control. The required accuracy and response speed vary depending on the scale.

[0004] From the perspective of modeling methods, they mainly include the following categories:

[0005] 1. Physical model method: Based on meteorological input and station parameters, a mechanistic model is constructed. It has good interpretability, but its accuracy is limited when dealing with complex terrain and rapid fluctuations.

[0006] 2. Statistical modeling methods: such as ARIMA and support vector machines, can model some time series patterns, but their performance degrades when facing non-stationary, multivariate scenarios;

[0007] 3. Deep learning methods: including LSTM, GRU, CNN, etc., have been widely used in new energy prediction tasks in recent years. They can learn nonlinear relationships and long-term dependencies, but they have weak model generalization ability, fixed activation functions, and strong training dependence.

[0008] Furthermore, current research often uses raw or simply cleaned data as model input, neglecting the high-frequency noise and modal overlap issues in wind and solar power data, leading to information redundancy and amplified errors. For example, traditional EMD decomposition often introduces non-physical features due to modal aliasing, reducing the reliability of subsequent modeling.

[0009] At the regulation level, energy storage systems are widely regarded as an important technological path to mitigate the fluctuations of new energy sources. However, existing prediction models fail to integrate the feedback regulation characteristics of energy storage systems into the prediction closed loop, only realizing a linear process of "predict first, then operate," making it difficult to form a dynamic feedback system of prediction-regulation-correction.

[0010] To address the aforementioned shortcomings, there is an urgent need for an output prediction model that integrates multi-scale data denoising mechanisms, possesses adaptive nonlinear modeling capabilities, and can be dynamically coupled with energy storage systems. Summary of the Invention

[0011] To address the problem that existing prediction models fail to integrate the feedback regulation characteristics of energy storage systems into the prediction closed loop, this invention provides a control method for a distributed wind, solar, and energy storage power output prediction model with scene adaptability.

[0012] This invention is achieved through the following technical solution: a control method for a distributed wind, solar, and energy storage power prediction model with scene adaptive capabilities, comprising the following steps:

[0013] S1. Empirical Mode Decomposition:

[0014] Using the EMD method, the original wind and solar power data sequence Decomposed into a series of IMF components and a residual component:

[0015] ;

[0016] In the formula: Indicates the first One IMF component; Indicates the remaining component; The total number of IMF components;

[0017] S2. Wavelet thresholding for high-frequency IMF components:

[0018] Based on the frequency characteristics of the IMF components obtained by the EMD method, the first... Wavelet thresholding noise reduction is performed on each high-frequency IMF component, as follows:

[0019] First, the high-frequency IMF components to be denoised. Perform wavelet transform to obtain wavelet coefficients :

[0020] ;

[0021] In the formula: Represents the discrete wavelet transform;

[0022] Then, a soft thresholding function is selected for the wavelet coefficients. Thresholding is performed, and the thresholding formula is as follows:

[0023] ;

[0024] In the formula: This represents the wavelet coefficients after thresholding. Indicates the threshold;

[0025] in The method for determining it is:

[0026] In the formula: This is an estimate of the standard deviation of the noise level; For data length;

[0027] Then the wavelet coefficients after thresholding Perform inverse wavelet transform to reconstruct the denoised IMF components. :

[0028] ;

[0029] In the formula: Indicates the inverse wavelet transform;

[0030] S3. Signal Reconstruction:

[0031] The noise-reduced IMF components By merging the remaining unprocessed low-frequency IMF components and residual components, the denoised wind and solar power sequence is reconstructed. :

[0032] ;

[0033] S4. Hybrid Prediction Modeling:

[0034] A hybrid neural network structure integrating Long Short-Term Memory (LSTM) and Kaplan-Antibody (KAN) networks is constructed to create an LSTM-KAN hybrid prediction model. Each network extracts information from different dimensions, which are then weighted and combined through a fusion module. The final LSTM-KAN hybrid prediction model outputs the predicted wind and solar power. :

[0035]

[0036] In the formula: Indicates the fusion weight. This represents the predicted wind and solar power output from the Long Short-Term Memory (LSTM) network. This represents the predicted wind and solar power output from the KAN network;

[0037] S5. Energy Storage System Power Regulation Strategy

[0038] When the actual power of wind and solar Higher than the predicted power of wind and solar power At that time, that is:

[0039] ;

[0040] The energy storage system performs a charging operation, storing excess electrical energy; charging power... for:

[0041] ;

[0042] When the actual power of wind and solar Lower than the predicted power of wind and solar power At that time, that is:

[0043] ;

[0044] The energy storage system performs a discharge operation to compensate for the power deficit, and the discharge power... for:

[0045] ;

[0046] In the formula: Indicates the maximum allowable charging and discharging power of the energy storage system; , These represent the maximum and minimum states of charge that the energy storage system can tolerate, respectively.

[0047] As a further improvement to the technical solution of the present invention, in step S5, the energy storage system is provided with operating constraints, including state of charge constraints:

[0048] ;

[0049] Charge and discharge power constraints:

[0050] .

[0051] As a further improvement to the technical solution of the present invention, the energy storage system is a battery energy storage, flywheel energy storage, or supercapacitor.

[0052] As a further improvement to the technical solution of this invention, when the energy storage system is a battery energy storage system, its charge and discharge characteristics can be expressed as follows:

[0053] ;

[0054] ;

[0055] In the formula: and These represent the current and previous charging states of the battery, respectively. and These represent the current and previous battery discharge states, respectively. To improve the charging and discharging efficiency of energy storage; For time step; This refers to the battery's rated capacity.

[0056] The control method for a distributed wind, solar, and energy storage power prediction model with scene adaptability provided by this invention has the following advantages compared with existing technologies:

[0057] 1. This invention discloses a dual-channel hybrid prediction model integrating Long Short-Term Memory (LSTM) networks and Kolmogorov-Arnold Networks (KAN networks), combined with a hierarchical data preprocessing algorithm and an energy storage regulation mechanism based on prediction error feedback, to achieve dynamic output prediction and control of wind power, photovoltaic, and energy storage systems under various operating scenarios. This method aims to improve the adaptability of prediction models to multi-source uncertainties such as climate fluctuations, geographical differences, and load disturbances, and overcomes the expressive limitations of traditional deep network structures by introducing KAN networks to enhance deep nonlinear learning.

[0058] 2. In terms of data preprocessing, this invention introduces Empirical Mode Decomposition (EMD) and a wavelet threshold denoising method based on signal-to-noise ratio weighting to selectively extract the dominant components in wind and solar power and suppress high-frequency disturbances. In terms of prediction modeling, a Long Short-Term Memory (LSTM) network is used to capture time series trends, and a KAN network is introduced to perform nonlinear mapping on high-dimensional features, constructing a serial-parallel dual-channel collaborative prediction architecture to achieve full-scale modeling from trends to disturbances. At the system control level, a prediction error feedback mechanism is combined to guide the energy storage system to perform dynamic charging and discharging responses, effectively smoothing out fluctuations in wind and solar power output and improving the controllability and practical usability of the prediction results.

[0059] 3. The multi-level structural design and inter-module collaborative optimization mechanism of this invention not only improve the prediction accuracy and generalization ability of the model, but also possess good algorithm interpretability and engineering adaptability. This method can be widely applied to short-term and ultra-short-term power prediction, power dispatch optimization, and smart microgrid operation control of new energy power plants on the distribution network side, and has strong theoretical value and application prospects. Attached Figure Description

[0060] The accompanying drawings, which are incorporated in and form part of this specification, illustrate embodiments consistent with the invention and, together with the description, serve to explain the principles of the invention.

[0061] 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, for those skilled in the art, other drawings can be obtained based on these drawings without creative effort.

[0062] Figure 1 shows a flowchart of the control method for the distributed wind, solar and energy storage power output prediction model with scene adaptation capability described in this invention.

[0063] Figure 2 shows the wind power output curve.

[0064] Figure 3 shows the photovoltaic power curve.

[0065] Figure 4 shows a schematic diagram of the EMD data decomposition results.

[0066] Figure 5 shows the comparison results of wind power predictions from different models.

[0067] Figure 6 shows the comparison results of photovoltaic power predictions from different models. Detailed Implementation

[0068] To better understand the above-mentioned objectives, features, and advantages of the present invention, the solutions of the present invention will be further described below. It should be noted that, unless otherwise specified, the embodiments of the present invention and the features thereof can be combined with each other.

[0069] Many specific details are set forth in the following description in order to provide a full understanding of the invention, but the invention may also be practiced in other ways different from those described herein; obviously, the embodiments in the specification are only some embodiments of the invention, and not all embodiments.

[0070] A control method for a distributed wind, solar, and energy storage power prediction model with scene adaptability includes the following steps:

[0071] S1. Empirical Mode Decomposition:

[0072] Using the EMD method, the original wind and solar power data sequence Decomposed into a series of IMF components and a residual component:

[0073] ;

[0074] In the formula: Indicates the first One IMF component; Indicates the remaining component; The total number of IMF components;

[0075] S2. Wavelet thresholding for high-frequency IMF components:

[0076] Based on the frequency characteristics of the IMF components obtained by the EMD method, the first... Wavelet threshold denoising is typically performed on each high-frequency IMF component (usually). In the following examples The process is as follows:

[0077] First, the high-frequency IMF components to be denoised. Perform wavelet transform to obtain wavelet coefficients :

[0078] ;

[0079] In the formula: Represents the discrete wavelet transform;

[0080] Then, a soft thresholding function is selected for the wavelet coefficients. Thresholding is performed, and the thresholding formula is as follows:

[0081] ;

[0082] In the formula: This represents the wavelet coefficients after thresholding. Indicates the threshold;

[0083] in The method for determining it is:

[0084] In the formula: This is an estimate of the standard deviation of the noise level; For data length;

[0085] Then the wavelet coefficients after thresholding Perform inverse wavelet transform to reconstruct the denoised IMF components. :

[0086] ;

[0087] In the formula: Indicates the inverse wavelet transform;

[0088] S3. Signal Reconstruction:

[0089] The noise-reduced IMF components By merging the remaining unprocessed low-frequency IMF components and residual components, the denoised wind and solar power sequence is reconstructed. :

[0090] .

[0091] By employing the aforementioned EMD (Empirical Mode Decomposition) method and a multi-level signal processing approach using wavelet thresholding denoising, this invention effectively reduces high-frequency noise interference in raw wind and solar power data while fully preserving the low-frequency trends and important features of the original signal, aiming to improve the predictability and modeling efficiency of raw wind and solar power data. EMD decomposes the raw power sequence into several IMF (Intrinsic Mode Function) components and residual components. To address the potential mode aliasing problem in EMD, a screening mechanism based on the modal energy entropy index is designed to eliminate invalid perturbation modes, retaining only the dominant feature components. Subsequently, wavelet thresholding is used to denoise the selected high-frequency IMF components, and finally, the processed effective modes are reconstructed to obtain high-quality input data with significantly suppressed noise and a clear information structure. This preprocessing process not only effectively suppresses high-frequency fluctuations and measurement errors in wind and solar power data but also provides more discriminative input features for subsequent deep learning models, significantly improving the accuracy and stability of the overall prediction system. Furthermore, this invention improves the predictive model's adaptability and generalization ability to data, significantly enhancing the accurate prediction capability of wind and solar power changes under different operating scenarios.

[0092] S4. Hybrid Prediction Modeling:

[0093] A hybrid neural network structure integrating Long Short-Term Memory (LSTM) and Kaplan-Antibody (KAN) networks is constructed to create an LSTM-KAN hybrid prediction model. Each network extracts information from different dimensions, which are then weighted and combined through a fusion module. The final LSTM-KAN hybrid prediction model outputs the predicted wind and solar power. :

[0094]

[0095] In the formula: Indicates the fusion weight. This represents the predicted wind and solar power output from the Long Short-Term Memory (LSTM) network. This represents the wind and solar power predicted output by the KAN network; in this invention... It is dynamically optimized during training and supports adaptive model switching in different scenarios.

[0096] S5. Energy Storage System Power Regulation Strategy

[0097] When the actual power of wind and solar Higher than the predicted power of wind and solar power At that time, that is:

[0098] ;

[0099] The energy storage system performs a charging operation, storing excess electrical energy; charging power... for:

[0100] ;

[0101] When the actual power of wind and solar Lower than the predicted power of wind and solar power At that time, that is:

[0102] ;

[0103] The energy storage system performs a discharge operation to compensate for the power deficit, and the discharge power... for:

[0104] ;

[0105] In the formula: Indicates the maximum allowable charging and discharging power of the energy storage system; , These represent the maximum and minimum permissible states of charge of the energy storage system, respectively.

[0106] In step S5, the energy storage system is configured with operational constraints, including state of charge constraints:

[0107] ;

[0108] Charge and discharge power constraints:

[0109] .

[0110] In steps S4 and S5, this invention designs an LSTM-KAN hybrid prediction model, constructing a dual-channel collaborative prediction model that combines temporal memory and nonlinear expression capabilities. The Long Short-Term Memory (LSTM) network serves as the main channel to extract long-term dependency features from wind and solar power data, making it particularly suitable for modeling fluctuation trends across multiple time scales. The KAN network acts as an auxiliary channel, dynamically adjusting the nonlinear mapping structure to enhance the model's adaptability to high-dimensional and complex features. The two networks are fused through a series-parallel structure, and the final prediction result is a weighted synthesis of the dual-channel outputs, with the fusion weights dynamically optimized based on scene characteristics and training errors. Compared to traditional single-network models or static activation function structures, the hybrid modeling method proposed in this invention exhibits stronger generalization ability and modeling stability when facing different climate, geographical, and load scene variations, effectively improving the accuracy and practicality of wind and solar power prediction.

[0111] Long Short-Term Memory (LSTM) is a special type of Recurrent Neural Network (RNN). Its basic structure includes four important parts: an input gate, a forget gate, an output gate, and a memory unit. This structure effectively solves the gradient vanishing and exploding problems of traditional RNNs and has excellent long-sequence memory capabilities. This invention employs a two-layer stacked LSTM network to capture short-term fluctuations and long-term trends, respectively.

[0112] The computation process of an LSTM cell can be represented by the following formula:

[0113] Input Gate:

[0114] Forgotten Gate:

[0115] Output gate: (8)

[0116] Memory cell state update: (9)

[0117] Memory cell state transfer: (10)

[0118] Hidden layer output: (11)

[0119] In the formula: This is the input for the current moment; This is the output of the hidden layer at the current moment; This represents the current state of the memory cell. These are the input gate, forget gate, and output gate, respectively. This is the weight matrix. For bias terms; This represents the Sigmoid activation function, with a value range of [0,1]. The hyperbolic tangent function has a range of [-1, 1]; the symbol " "" indicates element-wise multiplication.

[0120] KAN (Kolmogorov-Arnold) network is an effective alternative to traditional MLP (Multi-Level Processing) networks, overcoming the gradient vanishing or exploding problems common in traditional models and improving prediction accuracy and generalization performance. Based on the Kolmogorov-Arnold representation theorem, the core idea of ​​KAN network is to approximate multivariable continuous functions using a combination of univariate functions, thereby significantly reducing the number of model parameters and effectively alleviating the optimization difficulties of traditional neural networks under high-dimensional data conditions. Considering the high degree of randomness and nonlinearity in wind and solar power prediction data, traditional neural networks inevitably get stuck in local optima when extracting these features, limiting prediction performance. The unique structure and optimization method of KAN network effectively solve these problems. Therefore, this invention adopts KAN network as the core prediction architecture to improve the model's ability to capture complex features in wind and solar power data, thereby further improving prediction accuracy and generalization performance.

[0121] An energy storage system is a device that can store and release energy. Through technologies such as battery energy storage, flywheel energy storage, and supercapacitors, it can quickly respond to system demands and provide functions such as smoothing power fluctuations, improving power quality, and grid stability.

[0122] Step S5 of this invention can effectively mitigate the grid stability risks caused by power prediction deviations of renewable energy sources such as wind and solar power. Through step S5, this invention can effectively alleviate the negative impact of wind and solar power prediction errors on grid stability, further enhance the application value of wind and solar power prediction results in actual operation, and reduce the reserve capacity demand caused by wind and solar power fluctuations, thereby lowering system dispatch costs. In summary, the method described in this invention effectively improves the power system's adaptability to wind and solar power fluctuations, supporting the efficient and safe consumption of new energy sources.

[0123] When the energy storage system uses batteries for energy storage, its charge and discharge characteristics can be expressed as:

[0124] ;

[0125] ;

[0126] In the formula: and These represent the current and previous charging states of the battery, respectively. and These represent the current and previous battery discharge states, respectively. To improve the charging and discharging efficiency of energy storage; For time step; This refers to the battery's rated capacity.

[0127] The specific embodiments of the present invention will be described in detail below.

[0128] Example

[0129] 1) Dataset selection

[0130] To verify the effectiveness of the above LSTM-KAN hybrid prediction model, data collected from a wind farm and photovoltaic power station of a certain factory from January to December 2023 were used for testing. Figures 2 and 3 show the daily power curve changes of wind power and photovoltaic power.

[0131] 2) Experimental parameter settings

[0132] In this embodiment, the Long Short-Term Memory (LSTM) network is configured as follows: Adam is used as the optimizer, and the learning rate of the LSTM is set to 0.0005. To verify the superiority of the proposed model, six models—GRU, GRU-KAN, TCN, TCN-KAN, LSTM, and LSTM-KAN—were also constructed for comparison of prediction results (the construction process of each model is the same as that of the LSTM-KAN model in this invention, except for the model used). The GRU and LSTM models have 2 layers, 32 hidden layers, and a dropout rate of 0.2. The KAN model has a grid size of 200. The TCN model has [25, 50, 25] channels, and the TCN model has a kernel size of 3.

[0133] 3) Evaluation Indicators

[0134] This embodiment selects the root mean square error (RMSE) and the coefficient of determination (R²). 2 These two indicators are used as evaluation metrics. The definitions of these two metrics are as follows:

[0135] ;

[0136] ;

[0137] In the formula: It is the average power.

[0138] 4) Data preprocessing

[0139] To effectively reduce noise and interference components in the raw wind and solar power data, all models adopted the data processing method of steps S1-3. Figure 4 shows the decomposition results of the raw wind power by empirical mode decomposition.

[0140] 5) Comparison of prediction results

[0141] After data preprocessing, predictions are made using the six prediction methods mentioned above. This paper selects a typical day from the annual test set as a sample to compare and demonstrate the changing trends of actual and predicted values. The prediction results for this typical day are representative and can reflect the model's predictive ability in the annual scenario without loss of generality. Figures 5 and 6 show the prediction comparison of each model for wind power and photovoltaic power on the typical day. Table 1 shows the performance of each model in the annual prediction of wind power and photovoltaic power.

[0142]

[0143] Note: In the above prediction method, (no noise reduction) means that the model has not performed the data preprocessing process of step S2. wavelet threshold noise reduction of high frequency IMF components and step S3. signal reconstruction.

[0144] Based on the analysis of Figures 5 and 6 and Table 1, the results are as follows:

[0145] Analysis of Data Preprocessing Effects: The performance changes of the same model before and after noise reduction show that data preprocessing improved performance in all models. Taking GRU as an example, the RMSE for wind power decreased from 27.75 to 26.87, and the RMSE for photovoltaics decreased from 3.61 to 3.60; while for the LSTM-KAN model, after preprocessing, the RMSE for wind power decreased to 22.80, and the RMSE for photovoltaics was 3.31, both reaching optimal levels. This demonstrates that the data preprocessing method provided in this invention can effectively improve prediction accuracy.

[0146] Analysis of the effect of introducing the KAN module: Comparing models with and without the KAN module, it can be found that the introduction of KAN has a significant effect on improving prediction performance. For example, in wind power prediction, TCN-KAN reduced the RMSE from 26.80 to 24.82, and R... 2 The RSE was increased to 0.79; LSTM-KAN further reduced the RSE to 23.45. 2 The efficiency reached 0.81. The introduction of the KAN module significantly enhances the model's ability to model complex nonlinear laws in wind and solar power, making it particularly suitable for time-series renewable energy data with strong nonstationary characteristics.

[0147] The LSTM-KAN hybrid prediction model demonstrates superior overall performance: Among all compared models, the LSTM-KAN hybrid prediction model proposed in this invention achieves the best results in both wind power and solar power predictions, with a wind power RMSE of 22.80 and R... 2 It is 0.83; the photovoltaic RMSE is 3.31, R 2The model achieves a performance score of 0.94, which is superior to baseline models such as GRU-KAN, TCN-KAN, and LSTM. This model integrates the temporal memory capability of LSTM, the high-order nonlinear expression capability of KAN, and the noise suppression capability of EMD, resulting in stronger stability and generalization, and enabling it to more accurately characterize the evolutionary characteristics of new energy power.

[0148] In summary, the LSTM-KAN hybrid prediction model constructed in this invention demonstrates significant advantages in both wind and solar power prediction tasks. The model structure is rationally designed, exhibits superior overall performance, and shows promising application prospects. Furthermore, this invention proposes a control method for a distributed wind, solar, and energy storage power prediction model that integrates multi-level data processing, adaptive nonlinear modeling, and error feedback control mechanisms. It constructs a scenario-adaptive prediction system composed of three parts: "EMD data augmentation + LSTM-KAN hybrid network + dynamic energy storage adjustment," demonstrating significant advantages in prediction accuracy, robustness, and engineering feasibility.

[0149] At the data preprocessing level, this invention addresses the characteristics of strong non-stationarity and large noise disturbance in wind and solar power data by proposing a data preprocessing method based on modal energy entropy screening and wavelet thresholding. This method can effectively separate invalid high-frequency components and adaptively complete noise reduction, providing clear and high-information-density input features for the model and improving the discernibility and stability of subsequent modeling.

[0150] At the modeling level, the constructed LSTM-KAN hybrid network fully combines time series modeling capabilities with nonlinear expression capabilities: LSTM extracts temporal dependencies, and the KAN network achieves dynamic adaptation and high-order mapping of complex input features. The two are synergistically integrated in a series-parallel structure, and finally, through a trainable fusion factor, the weights are dynamically optimized in multiple scenarios, significantly enhancing the model's generalization ability.

[0151] At the output control level, an innovative energy storage power regulation mechanism based on prediction error feedback is introduced, breaking the traditional linear process of "prediction-output-regulation" and constructing a closed-loop architecture of "prediction-error-regulation-correction". This mechanism can dynamically adjust the charging and discharging strategy of the energy storage system according to the prediction error, realizing rolling correction and smoothing of the prediction results, thereby improving the overall system's regulation capability and practical value.

[0152] Experimental comparisons were conducted on real-world wind and solar power plant annual forecast datasets, and the results show that the method described in this invention achieves better performance in terms of accuracy, RMSE, and R-value. 2 It outperforms traditional models and physical statistics methods in key metrics, with prediction accuracy and robustness improving simultaneously.

[0153] In summary, the wind-solar-storage joint prediction method proposed in this invention not only integrates advanced deep learning structures and signal processing algorithms in theoretical modeling, but also has good adaptability and deployment capabilities in engineering practice. It is suitable for regional power grid dispatching scenarios with a high proportion of new energy sources and severe power fluctuations, providing solid technical support for the safe, low-carbon, and efficient operation of the power system.

[0154] The above description is merely a specific embodiment of the present invention, enabling those skilled in the art to understand or implement the present invention. Although detailed descriptions have been provided with reference to the foregoing embodiments, those skilled in the art should understand that modifications can still be made to the technical solutions described in the foregoing embodiments, or equivalent substitutions can be made to some or all of the technical features therein; and these modifications or substitutions do not cause the essence of the corresponding technical solutions to deviate from the scope of the technical solutions of the embodiments, and they should all be covered within the protection scope of the claims.

Claims

1. A control method for a distributed wind, solar, and energy storage power output prediction model with scene adaptive capability, characterized in that, Includes the following steps: S1. Empirical Mode Decomposition: Using the EMD method, the original wind and solar power data sequence is decomposed. Decomposed into a series of IMF components and a residual component: In the formula: Indicates the first One IMF component; Indicates the remaining component; The total number of IMF components; S2. Wavelet thresholding for high-frequency IMF components: Based on the frequency characteristics of the IMF components obtained by the EMD method, select the... Wavelet threshold denoising is performed on the high-frequency IMF components as follows: First, the high-frequency IMF components to be denoised are... Perform wavelet transform to obtain wavelet coefficients ： In the formula: Represents the discrete wavelet transform; Then, a soft thresholding function is selected for the wavelet coefficients. Thresholding is performed, and the thresholding formula is as follows: In the formula: This represents the wavelet coefficients after thresholding. Indicates the threshold; in The method for determining it is: In the formula: This is an estimate of the standard deviation of the noise level; The data length is then used; the wavelet coefficients after thresholding are then... Perform inverse wavelet transform to reconstruct the denoised IMF components. ： In the formula: S3. Signal Reconstruction: Reconstructing the denoised IMF components... (The rest of the text appears to be incomplete and requires further context.) By merging the remaining unprocessed low-frequency IMF components and residual components, the denoised wind and solar power sequence is reconstructed. ： ； S4. Hybrid Prediction Modeling: A hybrid neural network structure integrating Long Short-Term Memory (LSTM) and KAN networks is constructed to build an LSTM-KAN hybrid prediction model. LSTM and KAN networks each extract information from different dimensions, which are then weighted and combined through a fusion module. Finally, the LSTM-KAN hybrid prediction model outputs the predicted wind and solar power. ： In the formula: Indicates the fusion weight. This represents the predicted wind and solar power output from the Long Short-Term Memory (LSTM) network. S5. Energy storage system power regulation strategy when the actual wind and solar power output is... (This refers to the predicted wind and solar power output from the KAN network.) Higher than the predicted power of wind and solar power At that time, that is: ； The energy storage system performs a charging operation, storing excess electrical energy; charging power... for: ； When the actual power of wind and solar Lower than the predicted power of wind and solar power At that time, that is: The energy storage system performs a discharge operation to compensate for the power deficit, and the discharge power... for: In the formula: Indicates the maximum allowable charging and discharging power of the energy storage system; 、 These represent the maximum and minimum states of charge that the energy storage system can tolerate, respectively.

2. The control method for the distributed wind, solar, and energy storage power output prediction model with scene adaptive capability according to claim 1, characterized in that, In step S5, the energy storage system is configured with operational constraints, including state of charge constraints: Charge and discharge power constraints: 。 3. The control method for the distributed wind, solar, and energy storage power output prediction model with scene adaptive capability according to claim 1, characterized in that, The energy storage system is a battery energy storage system, a flywheel energy storage system, or a supercapacitor.

4. The control method for the distributed wind, solar, and energy storage power output prediction model with scene adaptive capability according to claim 3, characterized in that, When the energy storage system uses batteries for energy storage, its charge and discharge characteristics can be expressed as: ； In the formula: and These represent the current and previous charging states of the battery, respectively. and These represent the current and previous battery discharge states, respectively. To improve the charging and discharging efficiency of energy storage; For time step; This refers to the battery's rated capacity.

Images