A physical constraint-based heterogeneous deep hybrid wind power intelligent prediction system and method

Abstract

This invention belongs to the field of new energy power generation and smart grid technology, and discloses a heterogeneous deep hybrid intelligent wind power prediction system and method based on physical constraints. The system deploys an IoT sensor network (wind speed, wind direction, temperature and humidity, air pressure, and wind turbine status monitoring modules) in the wind farm, combined with a SCADA system to collect environmental and operational parameters in real time. After data quality processing, the CEEMDAN algorithm is used to decompose and denoise the power sequence at multiple scales; a CNN-Transformer heterogeneous fusion model is constructed to collaboratively extract local and long-term time-series features; PINN is introduced to embed aerodynamic constraints, jointly optimizing data and physical residuals; and the SHAP method is used to quantify feature contribution, improving model interpretability and scheduling decision support capabilities.

Description

Technical Field

[0001] This invention belongs to the field of new energy power generation and smart grid technology, specifically relating to a heterogeneous deep hybrid wind power intelligent prediction system and method based on physical constraints. Background Technology

[0002] With the rapid development of new energy technologies and deep learning, intelligent operation and management of wind farms has become crucial for improving renewable energy absorption capacity and grid stability. However, traditional wind power forecasting methods largely rely on empirical physical models or linear statistical models, which struggle to effectively handle the non-stationarity and noise interference of wind power sequences. They also fail to accurately characterize the complex nonlinear relationship between meteorological factors such as wind speed and power output, resulting in limited forecast accuracy and delayed early warning and dispatch responses. Although existing technologies have incorporated data-driven models such as neural networks for wind power forecasting, shortcomings remain in multi-scale feature collaborative modeling, physical consistency constraints, and interpretability analysis. These models also have limitations in cross-wind farm migration, real-time response, and adaptability to complex environments, making it difficult to meet the refined control requirements of smart grids under highly dynamic conditions. Summary of the Invention

[0003] To address the shortcomings of existing wind power prediction technologies in terms of accuracy, real-time performance, and scheduling coordination, this invention provides a heterogeneous deep hybrid intelligent wind power prediction system and method based on physical constraints. The system collects real-time wind farm operation and environmental data through an IoT sensor network and a SCADA system. After anomaly detection, missing data repair, normalization, and feature filtering, the power sequence is decomposed into multiple scales using CEEMDAN to reduce the impact of non-stationarity. A fusion model of CNN and Transformer is constructed to achieve collaborative modeling of local fluctuation features and long-term time-series dependent features. PINN is introduced to embed aerodynamic power conversion relationships as physical constraints to improve the model's physical consistency and generalization ability. Based on the prediction results, the system generates risk warnings and scheduling suggestions, and issues control commands to wind turbines or energy storage devices through a feedback and scheduling decision module, thereby improving prediction accuracy and enhancing grid connection stability. Furthermore, the SHAP method is used to analyze the feature contribution of the prediction results, improving model interpretability and assisting in scheduling decisions and risk management.

[0004] To achieve the above objectives, the present invention provides the following solution: A physical constraint-based intelligent prediction system for heterogeneous deep hybrid wind power, the system comprising: a data acquisition module, a data quality processing module, a multi-scale decomposition module, a heterogeneous deep feature extraction module, a feature fusion module, a physical constraint embedding module, a prediction output module, an interpretability analysis module, and a feedback and scheduling decision module; The data acquisition module is used to collect real-time environmental parameters and equipment operating parameters of the wind farm through the IoT sensor network and SCADA system deployed in the wind farm. The data quality processing module is used to perform anomaly detection, missing value repair, normalization processing, and feature filtering on the collected raw data. The multi-scale decomposition module is used to decompose the selected features using the fully adaptive noise set CEEMDAN algorithm to obtain multiple intrinsic mode function components and residual components. The heterogeneous deep feature extraction module is used to jointly model and predict the input features after data quality processing and the multi-scale components obtained by CEEMDAN decomposition. The physical constraint embedding module is used to embed physical constraints based on the wind energy aerodynamic power conversion relationship during model training, and to construct a joint optimization objective of physical residual loss term and data error loss term; The feature fusion module is used to couple and fuse the features output by the heterogeneous deep feature extraction module to form a comprehensive feature representation and output the prediction result; The prediction output module is used to generate wind power prediction values ​​for multiple future time points and output power change trends and fluctuation risk indicators. The interpretability analysis module is used to decompose the feature contribution of the prediction results using the SHAP method, and to quantify the influence of each input feature on the prediction power. The feedback and scheduling decision module is used to generate scheduling suggestions based on the prediction results and risk indicators, assisting in the optimization of wind farm operating parameters and grid scheduling decisions, and realizing rolling prediction and dynamic updates.

[0005] Preferably, the data acquisition module includes a sampling frequency control unit, which is used to dynamically adjust the sampling frequency of various sensors according to the wind speed change rate or power fluctuation amplitude; the adjustment mechanism is as follows: ; in, This is the original sampling frequency; The adjusted frequency; For adjustment coefficients; This represents the prediction error at the current moment; This is a preset reference threshold; This is the maximum normalized reference value for the error or fluctuation amplitude. This represents the degree of deviation of the current error from the reference threshold. Preferably, the process of performing anomaly detection, missing value repair, normalization, and feature selection on the collected raw data includes: Anomaly detection of raw collected data is performed based on the Isolation Forest algorithm. Multiple random isolation trees are constructed to recursively partition the samples until a sample point is isolated. The path length of a sample within the random isolation trees is used as the anomaly detection criterion. An average path length normalization factor is applied to normalize the path length for different sample sizes. The calculation formula is as follows: ; in, For the sample size The average path length normalization factor at that time; for The harmonic number; The number of samples used to construct the randomized isolation tree; For missing data resulting from anomaly removal and the original missing data, interpolation is used for completion. The completion process is expressed as follows: ; in, Indicates the time to be completed or estimated. Data values; In a known data sequence, time... The corresponding known data value; In a known data sequence, time... The corresponding known data value; This indicates the time point where data needs to be supplemented. After anomaly repair and missing data completion, to eliminate the impact of differences in the dimensions of different feature variables on model training, each feature variable is normalized. The normalization process is expressed as follows: ; in, Normalized data representing various relevant parameters of a wind farm. This represents the raw data of a wind farm. This represents the minimum value in the wind farm dataset. This represents the maximum value in the wind farm dataset.

[0006] Preferably, the process of decomposing the selected features using the fully adaptive noise set CEEMDAN algorithm to obtain multiple intrinsic mode function components and residual components includes: ; in, The reconstructed signal; Extracted by Empirical Mode Decomposition (EMD) eigenmode functions , representing the different frequency components of the signal, It is the residual, which represents the long-term trend of the signal; For reconstruction operations, it means combining multiple IMFs and residuals into a single overall signal. For time.

[0007] Preferably, the process of jointly modeling and predicting the input features after data quality processing and the multi-scale components obtained by CEEMDAN decomposition includes: a CNN branch and a Transformer encoder branch. The CNN branch convolution operation is represented as: ; in, The output of the CNN layer is a weighted sum of the input data; These are the weights of the convolution kernel; Input timing data; For bias terms; The size of the convolution kernel; The calculation formula for the multi-head self-attention mechanism of the Transformer encoder branch is as follows: ; in, The query value is a vector obtained from the input data through a linear transformation; The key vector represents the feature of each input data in the model; It is a value vector, representing the actual information of the input data; This is the scaling factor; The dimensions of the query and key vectors; To query the dot product of the key vector and the query vector.

[0008] Preferably, the process of embedding physical constraints based on wind energy aerodynamic power conversion relationships during model training, and constructing a joint optimization objective of physical residual loss term and data error loss term, includes: Using aerodynamic formulas, the power conversion relationship of wind turbine units is constructed as a physical constraint: ; in, This refers to the power output of the wind turbine generator; air density; The swept area of ​​the wind turbine; This represents the power coefficient of the fan. Wind speed; By combining physical constraints with data error terms, training is performed through joint optimization, ultimately optimizing the objective function. for: ; in, For data error loss, This is the physical residual loss term constructed based on the aerodynamic power relationship. These are the physical constraint weighting coefficients.

[0009] Preferably, the process of coupling and fusing the features output by the heterogeneous deep feature extraction module to form a comprehensive feature representation and output the prediction result includes: ; in, This is the final feature output after weighting; The weighting coefficients represent the importance of the two features and control the contribution of each branch to the final fusion result. Output features for CNN branches; This is the output feature of the Transformer encoder branch.

[0010] Preferably, the process of generating wind power forecasts for multiple future time points and outputting power change trends and fluctuation risk indicators includes: ; in, To predict the power value, Historical power values As a risk index, The trend is predicted to be: ; in, Forecasted wind power output for the future; This is a trend adjustment based on historical data, representing the changing trend of wind power output; The volatility risk index is: ; in, For the Sigmoid function; For feature input; As weight; For bias terms, To predict the probability of risk.

[0011] Preferably, the process of generating scheduling recommendations based on forecast results and risk indicators to assist in optimizing wind farm operating parameters and grid scheduling decisions, and achieving rolling forecasting and dynamic updating, includes: Receive wind power forecasts, power fluctuation risk indices, and confidence intervals generated by the forecast output module; Based on wind power forecast values ​​and power fluctuation risk index, calculate system operation status evaluation indicators and power change trend indicators; When the system operation status evaluation index and power change trend index exceed the preset threshold, a scheduling control command is generated and sent to the wind turbine or energy storage system for execution through the SCADA system. The adjusted operating parameters and scheduling records are stored in the data storage module; and the prediction results and control execution status are displayed in real time through the user interface module to form a closed-loop operation mechanism of prediction-evaluation-control-recording-feedback.

[0012] This invention also provides a physically constrained intelligent prediction method for heterogeneous deep hybrid wind power, which is implemented through the aforementioned system and includes: Real-time collection of environmental and operational parameter data is achieved through an IoT sensor network and SCADA system deployed in the wind farm. The collected data undergoes anomaly detection, missing value repair, normalization, and feature filtering to obtain input feature data for modeling. CEEMDAN was used to perform multi-scale decomposition on the wind power time series, resulting in multiple intrinsic mode function components and residual components. The decomposed components and input feature data are input into a heterogeneous deep fusion model composed of CNN branches and Transformer encoder branches for training and prediction, and the wind power prediction value is output through feature fusion. PINN is introduced during model training to construct the aerodynamic power conversion relationship as a physical residual term and jointly optimize it with the data error term. Based on the prediction results, a power fluctuation risk index is constructed and early warning information and scheduling suggestions are generated. The feedback and scheduling decision module converts the prediction results into operation control commands and sends them to the wind turbine or energy storage system through the SCADA system for execution, so as to realize power balance and closed-loop optimized operation of the wind power system.

[0013] Compared with the prior art, the beneficial effects of the present invention are as follows: This invention achieves multi-timescale prediction and fluctuation risk assessment of wind power through multi-source data acquisition, data quality processing, and wind power prediction modeling. It aims to improve wind farm operating efficiency, enhance grid connection stability, and support intelligent grid dispatch and control. Integrated with a SCADA system, this invention transforms wind power prediction results and risk indices into operational control commands, enabling dynamic management and real-time parameter adjustment of wind turbines and energy storage systems. Simultaneously, it introduces CEEMDAN to decompose wind power time series data at multiple scales, constructing a heterogeneous deep fusion model combining CNN and Transformer encoders to characterize local fluctuation features and long-term temporal dependencies. PINN constraints are embedded during training to improve the physical consistency and generalization ability of the prediction results. Furthermore, the SHAP method is used to perform feature contribution decomposition analysis on the prediction results, quantifying the impact of different meteorological and operational characteristics on the predicted power, thereby improving system transparency and providing support for dispatch decisions and risk management. Attached Figure Description

[0014] To more clearly illustrate the technical solution of the present invention, the drawings used in the embodiments are briefly introduced below. Obviously, the drawings described below are only some embodiments of the present invention. For those skilled in the art, other drawings can be obtained based on these drawings without creative effort.

[0015] Figure 1 This is a schematic diagram of a heterogeneous deep hybrid wind power intelligent prediction system based on physical constraints, according to an embodiment of the present invention. Figure 2 This is a schematic diagram of the overall workflow of the heterogeneous hybrid prediction model proposed in this embodiment of the invention; Figure 3 The diagram shows the importance and interaction of SHAP features in the wind power prediction model according to the embodiment of the present invention. (a) is a summary diagram of SHAP for the WPF model based on meteorological input features, and (b) is a heatmap of SHAP interaction showing the interaction effect between meteorological input features. Detailed Implementation

[0016] The technical solutions of the embodiments of the present invention will be clearly and completely described below with reference to the accompanying drawings. Obviously, the described embodiments are only some embodiments of the present invention, and not all embodiments. Based on the embodiments of the present invention, all other embodiments obtained by those skilled in the art without creative effort are within the scope of protection of the present invention.

[0017] To make the above-mentioned objects, features and advantages of the present invention more apparent and understandable, the present invention will be further described in detail below with reference to the accompanying drawings and specific embodiments.

[0018] Example 1 like Figures 1-3 As shown, this invention provides a heterogeneous deep hybrid wind power intelligent prediction system based on physical constraints. This technology is applicable to various wind power scenarios, including but not limited to onshore wind power, offshore wind power, and microgrid systems. It can effectively improve the intelligent scheduling and management level of new energy access, reduce wind curtailment rate, and optimize the consumption and scheduling strategy of wind power in smart grids. It includes: (1) a data acquisition module, used to collect environmental and operating parameter data in real time through the Internet of Things sensor network and SCADA system deployed in the wind farm. The data includes wind speed, wind direction, temperature, humidity, atmospheric pressure, and historical power data at different altitudes. The data acquisition module includes multiple sensor nodes and data acquisition terminals set in the wind farm. The sensor nodes are connected to the SCADA system through a communication network to realize the real-time acquisition of wind farm environmental parameters and equipment operating parameters. (2) Data quality processing module, used to perform anomaly detection, missing value repair, normalization and feature selection on the collected raw data to improve data reliability and modeling stability; (3) Multi-scale decomposition module, the CEEMDAN algorithm is a technique for adaptive noise set decomposition of signals, suitable for processing non-stationary data such as wind power time series. In this process, CEEMDAN can effectively extract the intrinsic features of the signal by decomposing the signal into multiple intrinsic mode functions (IMF) and residual components. The process is described as follows: Original signal preparation: Assume the wind power time series is as follows This signal contains information about changes in wind power output. The raw data is typically non-stationary and may contain noise and local anomalies.

[0019] Additive noise addition: In the CEEMDAN algorithm, to avoid mode aliasing, different noise is added to the signal with each decomposition. Adaptive noise is added to the original signal. This generates a noisy signal: in, For the first The signal after adding auxiliary noise, that is, the wind power time series after adding noise; This is the original wind power time series, representing the time at time... Actual wind power data; For the first The amplitude coefficient when noise is added for the first time, also known as the noise amplitude adjustment parameter, is used to control the strength of the added noise. For at any time The auxiliary noise signal added is usually zero-mean Gaussian white noise; It is a time variable.

[0020] Empirical Mode Decomposition (EMD): For each noisy signal The EMD method is used to extract the IMF of a signal. The EMD steps include: first, finding the local maxima and local minima of the signal; constructing the upper and lower envelopes of the signal using these extrema; calculating the average of these two envelopes as the trend component of the signal; and removing the trend component from the signal to obtain the residual. Then, the IMF of the signal is obtained by repeating the above steps. Finally, the signal processed by EMD can be represented as: in, It is the first One eigenmode function It is the residual, which represents the long-term trend of the signal; Noisy signal The results obtained after empirical mode decomposition Number of components; It is a time variable.

[0021] Multiple Noise Addition Decomposition: The key to CEEMDAN lies in its use of multiple noisy EMD decompositions. For each added noise signal... The EMD process described above is repeated. Through multiple decompositions, CEEMDAN can eliminate the interference of noise on the decomposition results. Setting the number of repetitions to N, a new IMF is generated each time, ultimately resulting in N IMF components and one residual component. Each IMF component represents an eigenmode of the signal, and the residual represents the long-term trend of the signal.

[0022] Denoising and reconstruction: Each decomposed IMF component represents a different frequency band of the signal and can be used to analyze the high-frequency and low-frequency components of the signal. The signal reconstruction process is as follows: in, The reconstructed signal; For multiple extracted by EMD , representing the different frequency components of the signal, It is the residual, which represents the long-term trend of the signal; For reconstruction operations, it means combining multiple IMFs and residuals into a single overall signal.

[0023] Adaptive noise adjustment: The amplitude of the noise is adjusted each time noise is added. It is adaptive; as the decomposition proceeds, the noise amplitude is gradually adjusted to ensure that each decomposition yields a relatively independent mode and to minimize the impact of noise on the decomposition.

[0024] (4) Heterogeneous deep feature extraction module, including a convolutional neural network branch and a Transformer encoder branch, wherein the CNN branch is used to extract local temporal fluctuation features, and the Transformer encoder branch is used to model global dependencies over a long time span; (5) Feature fusion module, used to couple and fuse the features output by the convolutional neural network branch and the Transformer encoder branch to form a comprehensive feature representation and output the prediction result; (6) Physical constraint embedding module, used to embed physical constraints based on the wind energy aerodynamic power conversion relationship during model training, and improve the physical consistency and generalization ability of the model prediction result by constructing a joint optimization objective of physical residual loss term and data error loss term; (7) Prediction output module, used to generate wind power prediction values ​​for multiple future time points, and output power change trend and fluctuation risk indicators; (8) Interpretability analysis module, used to decompose the feature contribution of the prediction result using the SHAP method, and quantify the influence of each input feature on the predicted power; (9) Feedback and scheduling decision module, used to generate scheduling suggestions based on the prediction results and risk indicators, assisting in the optimization of wind farm operation parameters and grid scheduling decisions, and realizing rolling prediction and dynamic updating.

[0025] The system forms a closed-loop operation mechanism of prediction-evaluation-control through a prediction output module, a risk assessment module, and a feedback and scheduling decision module.

[0026] In this embodiment, the data acquisition module includes a SCADA system and multiple sensor nodes. These sensors are distributed at different heights and key operating locations within the wind farm. Each sensor node includes hardware devices such as wind speed sensors, wind direction sensors, temperature and humidity sensors, atmospheric pressure sensors, and a wind turbine operation status monitoring module. These devices measure wind speed information at different heights, including but not limited to wind speeds at 50m and 70m from the weather tower and at the hub height, as well as environmental parameters such as wind direction, temperature, humidity, and atmospheric pressure. They also collect operating parameters such as wind turbine output power, current, voltage, speed, pitch angle, yaw angle, nacelle temperature, bearing temperature, vibration signals, and start / stop status. The sensors are installed at key locations such as the weather tower, wind turbine nacelle, and tower, enabling real-time monitoring of the wind power equipment and its environmental status. The sensors transmit data to the SCADA system via a wireless communication module or industrial Ethernet. The sensor nodes form a distributed Internet of Things (IoT) data acquisition network, enabling multi-source real-time acquisition of wind farm environmental parameters and equipment operating parameters, thereby providing stable and reliable data input for wind power prediction models. The SCADA system collects and stores high-frequency operating data at preset time intervals, providing input data support for the multi-scale decomposition module, heterogeneous deep feature extraction module, and physical constraint embedding module. The data acquisition module supports the access of various types of sensors and can dynamically adjust the sampling frequency according to wind speed changes or operating status to adapt to the operating conditions of different wind farms.

[0027] Specifically, the data acquisition module can automatically adjust the sensor sampling frequency based on external dynamic conditions such as wind speed changes and sudden weather changes, thereby improving data granularity during critical periods. The adjustment mechanism can be described as follows: in, This is the original sampling frequency; The adjusted frequency; This is for adjusting the coefficient. This represents the prediction error at the current moment; This is a preset reference threshold; This is the maximum normalized reference value for the error or fluctuation amplitude. This represents the degree of deviation of the current error from the reference threshold.

[0028] In this embodiment, the data quality processing module is used to perform anomaly detection, missing data completion, scale unification, and feature extraction on the collected data to improve the consistency, validity, and prediction accuracy of the input data and subsequent models. Specifically, the data quality processing module first performs anomaly detection on the original collected data based on the isolated forest algorithm. It constructs multiple random isolation trees to recursively partition the samples until a sample point is isolated, and uses the path length of the sample in the random isolation tree as the anomaly detection criterion. To normalize the path length for different sample sizes, an average path length normalization factor is used, the calculation formula of which is: in, For the sample size The average path length normalization factor at that time; for The harmonic number; The number of samples used to construct the randomized isolation tree.

[0029] For missing data resulting from anomaly removal and the original missing data, interpolation is used to complete the data and restore the continuity of the time series. The completion process can be represented as follows: in, Indicates the time to be completed or estimated. Data values; In a known data sequence, time... The corresponding known data value; In a known data sequence, time... The corresponding known data value; This indicates the time point when data needs to be supplemented. and between; This is an earlier time point among the known data points. This represents another later time point among the known data points; The proportion representing the time difference.

[0030] After anomaly repair and missing data completion, to eliminate the impact of differences in the dimensions of different feature variables on model training, each feature variable is normalized. The normalization process can be expressed as: in, Normalized data representing various relevant parameters of a wind farm. This represents the raw data of a wind farm. This represents the minimum value in the wind farm dataset. This represents the maximum value in the wind farm dataset.

[0031] In this embodiment, the heterogeneous deep feature extraction module includes a CNN branch and a Transformer encoder branch. The convolutional neural network branch extracts local temporal fluctuation features through multi-layer convolution and pooling operations. To improve the performance of the CNN branch, feature extraction capabilities can be enhanced by increasing the depth of convolutional layers, using residual connections, and optimizing the convolutional kernel size. By optimizing the convolutional kernel parameters, the CNN can better capture local temporal fluctuation features. Specifically, the convolution operation can be represented as: in, The output of the CNN layer is a weighted sum of the input data. These are the weights of the convolution kernel; Input timing data; For bias terms; This represents the size of the convolution kernel.

[0032] The Transformer encoder branch models global dependencies over long time spans using a multi-head self-attention mechanism. To enhance the modeling effect of time-series data, positional encoding and temporal information can be combined, and an adaptive attention mechanism can be used. The calculation formula for the multi-head self-attention mechanism is as follows: in, The query value is a vector obtained from the input data through a linear transformation; The key vector represents the feature of each input data in the model; It is a value vector, representing the actual information of the input data; This is the scaling factor; The dimensions of the query and key vectors; To query the dot product of the key vector and the query vector.

[0033] The feature fusion module couples and fuses the output features of the two branches to form a comprehensive feature representation, and updates and trains it online based on newly acquired data to improve the accuracy and stability of wind power prediction. The specific process is as follows: First, the features output by the two branches (CNN branch and Transformer branch) are... and The features are coupled and fused. This coupling and fusion can be achieved through weighted summation or weighted concatenation. Specifically, weighted summation can be used to merge two features into a comprehensive feature representation. : in, This is the final feature output after weighting; The weighting coefficients represent the importance of the two features and control the contribution of each branch to the final fusion result. This is the first eigenvalue; This is the second eigenvalue.

[0034] Integrated characteristics after fusion This data will be used as new input for subsequent model training. The model will then be updated and trained online based on the newly collected data, with the goal of minimizing the loss function. The weights are continuously adjusted to improve the accuracy and stability of wind power forecasting. The update process can be represented as: in, The weights before the update; The updated weights; The learning rate; This is the gradient of the loss function.

[0035] The interpretability analysis module uses the SHAP method to perform feature contribution decomposition analysis on the prediction results in order to evaluate the degree of influence of each input feature on the predicted power, thereby enhancing the system transparency and scheduling decision support capabilities.

[0036] The system supports distributed deployment across multiple wind farms. The heterogeneous deep feature extraction module and physical constraint embedding module perform model training locally at each wind farm, and the local model parameters of each wind farm are weighted and aggregated through a federated learning mechanism to form a global model. During parameter transmission, a differential privacy mechanism is used to process the model parameters for noise perturbation to prevent leakage of original operating data, thereby improving data security and privacy protection capabilities while ensuring wind power prediction performance.

[0037] Figure 2The heterogeneous deep ensemble prediction model presented here employs a modeling framework that combines multi-scale decomposition, heterogeneous deep fusion, and physical constraints. It innovatively integrates the CEEMDAN algorithm, a multi-layer deep learning model, and physical constraints, significantly distinguishing it from existing wind power prediction systems that typically rely on a single model or traditional physical modeling methods. Specifically, the model first utilizes the CEEMDAN algorithm to perform multi-scale decomposition of the input features, addressing the non-stationarity of wind power time series data. It extracts multiple intrinsic mode functions (IMFs) and residual components, ensuring effective data analysis across different time scales and significantly improving prediction accuracy and robustness. The core of the model includes a convolutional neural network (CNN) branch and a Transformer encoder branch, a structure rarely seen in existing wind power prediction models. The CNN branch focuses on extracting local temporal fluctuation features and multivariate coupling features. Through multi-layer convolution operations, it accurately captures fluctuations in short-term environmental factors such as wind speed, temperature, and humidity, demonstrating its advantages, especially in scenarios with rapid short-term changes. The Transformer encoder branch, on the other hand, models global dependencies over long time spans, addressing the shortcomings of traditional deep learning models in long-term dependency issues. It effectively captures global dependencies in long time series through a self-attention mechanism, making it suitable for long-term fluctuations in wind power and wind speed changes. Finally, the feature fusion module fuses the output features of the CNN and Transformer branches, ensuring that local temporal patterns and long-term dependency patterns can be co-expressed, greatly improving the accuracy and stability of wind power prediction. Simultaneously, a Physical Information Neural Network (PINN) mechanism is introduced during model training, constructing aerodynamic power conversion relationships as physical constraints, which are jointly optimized with data error terms to improve the physical consistency and generalization ability of the prediction results. The specific process is as follows: First, the power conversion relationship of the wind turbine is constructed as a physical constraint using aerodynamic formulas: in, This refers to the power output of the wind turbine generator; air density; The swept area of ​​the wind turbine; This represents the power coefficient of the fan. This refers to wind speed.

[0038] Next, the physical constraint term and the data error term are combined and trained through joint optimization. The final optimization objective function is... As shown below: in, For data error loss, This is the physical residual loss term constructed based on the aerodynamic power relationship. These are the physical constraint weighting coefficients.

[0039] This invention further employs an interpretability analysis module and the SHAP method to perform feature contribution decomposition analysis on the prediction results, quantifying the impact of each input feature on the wind power prediction results. Specifically, the SHAP method, based on Shapley value theory, calculates the contribution of each input feature across all possible feature combinations, thereby enabling interpretation of the model output. Each input feature... The contribution to the prediction results is calculated using the following formula: in, For the first The contribution value (SHAP value) of each feature indicates the degree of influence of that feature on the model's prediction results; This is the model's predicted output; To remove the first Input data following each feature; This represents the expected value calculation for all possible combinations of features; For the first One eigenvalue; To remove the first The model prediction output after each feature; To remove the first The input vector after each feature.

[0040] Through this calculation process, the SHAP method can quantify the contribution of each input feature to the wind power prediction results, thereby improving the transparency of the model and helping to optimize scheduling decisions.

[0041] The innovation of this invention lies in dynamically updating the SHAP value by combining the operating status and physical constraints of the wind farm, and then comprehensively analyzing these values ​​with the prediction results. Compared with traditional methods, this invention can consider the impact of physical constraints on the prediction results, thereby improving the physical consistency and generalization ability of the prediction results. Specifically, the SHAP method process is as follows: First, the model generates predicted wind power values ​​and calculates the SHAP value of each input feature; then, the specific contribution of each input feature to the prediction result is quantified; next, the contribution of each feature is further analyzed by combining the physical constraints of the wind turbine, such as aerodynamic power conversion relationships; finally, by generating a feature contribution map, interpretable analysis results are provided to help improve the model's decision support capabilities.

[0042] To further enhance the generalization ability of models across different wind farms and reduce data silos, this invention combines federated learning and differential privacy technologies. This allows each wind farm to train its model locally, uploading only model parameter updates, thus preventing data from leaving the site. Simultaneously, noise injection protects data privacy. The formula for aggregating federated learning model parameters is: in, For global models; For the first A local model of a wind farm; These are the weighting coefficients; This represents the total number of sub-models, i.e., the number of local models or wind farm models participating in the fusion.

[0043] The differential privacy protection formula is: in, These are the parameters after adding noise; Standard deviation; For the original first One parameter; It follows a normal distribution.

[0044] The system of this invention also includes a SCADA system module for real-time monitoring of the wind power equipment's operating status and, based on control suggestions from the feedback and scheduling module, automatically executing output adjustment, turbine on / off control, or energy storage strategy optimization. The SCADA system can work in conjunction with the feedback module to proactively control turbine operating parameters based on prediction results, improving system responsiveness and stability.

[0045] The main functions of the SCADA system include: (1) real-time acquisition of operating parameters such as wind speed, wind turbine status, voltage, current, and meteorological data; the SCADA system communicates with wind speed sensors, wind direction sensors, temperature and humidity sensors and wind turbine operation status monitoring modules to realize centralized acquisition and management of wind farm environmental parameters and equipment operation status data. (2) receiving the prediction results output by the prediction module, generating scheduling suggestions in combination with the scheduling decision module, and sending the scheduling suggestions to wind turbines and energy storage systems. (3) monitoring and recording the operating status to form historical data trajectories. (4) providing a visual interface for dispatchers to view system status, alarms and execution status.

[0046] Integrating the SCADA system into the wind power prediction and scheduling control platform, which includes a data acquisition module, a data quality processing module, a prediction module, a scheduling decision module, an execution control module, and a visualization monitoring module, can bring the following advantages: (1) prediction-driven control to achieve automated scheduling; (2) anomaly detection closed-loop control to improve system reliability; (3) real-time feedback of prediction results and control execution status to enhance operation and maintenance transparency; and (4) reduced manual intervention to improve the system's intelligence level.

[0047] In this embodiment, the prediction output module is used to construct a power fluctuation risk index based on the deviation between the predicted power value and the historical power series. The risk index is then compared with a prediction threshold. When the risk index... Exceeding the preset threshold When this happens, the system will generate an alarm signal, prompting the wind farm to adjust its operating strategy. The formula for calculating the risk index is as follows: in, To predict power values; This is the historical power value.

[0048] like Exceeding the preset threshold An alarm signal is then generated, and the formula for generating this signal is: in, This is an alarm signal used to indicate whether an alarm needs to be triggered. The power fluctuation risk index represents the deviation between predicted power and actual power. A preset risk threshold is used to determine whether an alarm should be triggered.

[0049] Furthermore, the prediction output module also combines historical operating data to generate a prediction of wind power output trends over a certain future time range, thereby improving the wind farm's dispatching capabilities and operational stability. The formula for trend prediction is: in, Forecasted wind power output for the future; This is a trend adjustment based on historical data, representing the changing trend of wind power output.

[0050] Through the above process, the prediction output module can effectively assess power fluctuation risks, generate corresponding scheduling suggestions, and help wind farms optimize operation, ensuring grid stability and efficient energy utilization.

[0051] Specifically, the prediction output module generates predicted wind power output values ​​for future periods based on the output of the integrated model and calculates a volatility risk index. When the risk value exceeds a threshold, the system issues an early warning and triggers the SCADA system to intervene in scheduling. The volatility risk index can be calculated using the following formula: in, For the Sigmoid function; For feature input; As weight; This is a bias term.

[0052] In this embodiment, the prediction output module predicts the power output trend at a future point in time based on the historical operating data of the wind farm, and generates a Remaining Usable Capacity (RUL) calculation that includes risk assessment and scheduling recommendations, to assist in maintenance plan formulation. The specific process is as follows: First, meteorological data such as wind speed, temperature, humidity, and air pressure are collected and input into the prediction model for training. The prediction model learns from historical data and the wind farm's operating patterns to generate a predicted power value for a future point in time. The model's calculation formula is as follows: in, For the predicted power output; The input feature data; These are the weights of the model.

[0053] Then, based on the predicted power output, combined with the current status and historical operating data of the wind turbine, the model generates the wind turbine's Remaining Lifetime (RUL). RUL represents the remaining lifespan of the equipment and is used to assess the equipment's health status and support maintenance decisions. The RUL calculation formula is as follows: in, These are parameters related to the health status of wind turbine units; It is a mapping function learned from historical data.

[0054] Finally, the system generates scheduling recommendations based on the prediction results and RUL (Regular Utility) to optimize the wind farm's operation strategy. These recommendations include adjusting turbine start-up and shutdown strategies, optimizing target output, or regulating the operating power of the energy storage system to reduce no-load operating losses and mechanical shocks caused by frequent start-ups and shutdowns, thereby improving the efficiency and stability of the wind farm. The generation of scheduling recommendations can be expressed by the following formula: in, yes RUL and risk assessment A function that generates scheduling suggestions.

[0055] In this embodiment, the feedback and scheduling decision module is used to calculate the deviation between the target output and the current actual output based on the wind power forecast and power fluctuation risk index generated by the prediction output module, and to generate operation parameter optimization control instructions based on the deviation. The specific process is as follows: First, calculate the deviation between the target output and the current actual output. This deviation reflects the difference between the predicted and actual wind power output. The calculation formula is: in, This is the output deviation. Contribute to the goal; To make practical contributions at present.

[0056] Then, based on the calculated deviation value, optimized control commands for wind turbine adjustment or energy storage system are generated. The specific process includes adjusting the wind turbine start / stop angle, regulating its speed, or adjusting the charging / discharging power of the energy storage system. The formula for generating the control commands is: in, The optimized control commands; As a weighting factor, it controls the deviation term. The extent of influence in the final control commands; Power deviation; Controlling the adjustment parameters The extent of influence in the final control commands; For parameter adjustment, it represents certain parameters used to adjust control commands.

[0057] The feedback and scheduling decision module is also used to automatically generate scheduling suggestion information and push it to the SCADA system, realizing a closed-loop operation mechanism of prediction-control-feedback. The feedback and scheduling decision module supports interface with the enterprise energy management system (EMS), and can directly synchronize control suggestions to the upper-level system. This process involves real-time collection of wind farm operating data, combined with predicted wind power output. Power fluctuation risk index The system performs a state assessment and generates scheduling recommendations. The formula for the state assessment is as follows: in, This is a system status assessment value, representing the health status of the wind turbine under given conditions; This refers to the rated power of the fan; It serves as a risk index for power fluctuations, measuring the stability and magnitude of system fluctuations.

[0058] Through feedback loops, the system can adjust the operating strategy of wind turbines in real time to optimize the overall power output and equipment lifespan of the wind farm, ensuring grid stability and efficient energy utilization.

[0059] Specifically, the feedback and scheduling decision module dynamically generates control commands based on the forecast results and adjusts the wind power output. For example, when excessive power may lead to wind curtailment risk, the system can adjust the output parameters: in, These are the adjusted fan output parameters; The fan output parameters before adjustment; This is a volatility risk index; To preset risk thresholds; This is the maximum normalized reference value for the risk index; This is the adjustment coefficient.

[0060] In this embodiment, the feedback and scheduling decision module is used to calculate the predicted power change rate when the future power prediction value generated by the prediction output module shows a downward trend and the power fluctuation risk index exceeds a preset threshold, and to generate an output adjustment control command based on the change rate. The output adjustment control command includes reducing the target output power of the wind turbine or controlling the energy storage system to perform discharge compensation, so as to maintain the power balance of the grid and reduce the impact of power drop on grid stability. The adjustment formula is as follows: in, This is the adjusted output power; This is the original predicted power; This is an adjustment factor (e.g., 0.1). This is a risk index.

[0061] In this embodiment, the feedback and scheduling decision module communicates with the SCADA system to convert the wind power forecast and power fluctuation risk index generated by the forecast output module into control execution commands, which are then sent to the SCADA system via a communication interface. The control execution commands include wind turbine target output adjustment commands or energy storage system charging and discharging power adjustment commands. When the predicted power exceeds a preset operating limit or the grid load pressure index exceeds a safety threshold, the feedback and scheduling decision module generates a power limiting or energy storage regulation control strategy to achieve smooth control of wind farm output and improve grid operation stability. The specific regulation control strategy can be expressed as follows: in, This refers to the adjusted wind power output. To predict power; The adjustment coefficient controls the adjustment range of the fan power output; This is the power fluctuation risk value. When the risk value exceeds the set threshold, the system will automatically adjust the wind turbine output or energy storage charging and discharging power to smooth wind power fluctuations and reduce the impact on grid stability.

[0062] In this embodiment, the feedback and scheduling decision module is used to calculate the equipment operating load index based on the wind power prediction value and power change trend generated by the prediction output module, and to generate operating parameter optimization control instructions according to the load index. The operating parameter optimization control instructions include adjusting the wind turbine start-up and shutdown strategy, optimizing the target output, or adjusting the operating power of the energy storage system to reduce no-load operation losses and reduce mechanical shocks caused by frequent start-ups and shutdowns, thereby reducing the overall energy consumption of the wind farm and extending the service life of the equipment. The specific calculation formulas for the equipment operating load index and the operating parameter optimization control instructions are as follows; in, This is an indicator of equipment operating load, reflecting the difference between the predicted and actual wind power output. This is the predicted value of wind power output; This represents the actual wind power output. This is an adjustment coefficient used to adjust the impact of power variation trends on load indicators; The power change rate represents the fluctuation range of wind power. The optimized wind power control command; This is a control factor used to adjust the power output of the wind turbine based on load indicators.

[0063] In this embodiment, the scheduling strategy module combines historical operating data with predictive input. First, it establishes a prediction model based on historical power prediction data of the wind farm, and generates predicted wind power values ​​and risk indices for future periods based on historical wind speed changes and power fluctuations. Next, based on data such as wind speed change trends and power fluctuation risk indices, it optimizes and adjusts the scheduling control strategies of wind turbines or energy storage systems to reduce the impact of power fluctuations on grid load.

[0064] In this embodiment, the feedback and scheduling decision module and the user interface module are connected for data interaction, used to display real-time wind power forecast values, power fluctuation risk index, equipment operating status, and execution status of scheduling control commands; the user interface module is used to receive manual confirmation or intervention commands and feed back the manual intervention commands to the feedback and scheduling decision module, so as to realize an operation mode that combines automatic control and manual decision-making. Specifically, the scheduling strategy can be optimized and calculated using the following formula: in, This represents the current output power of the wind turbine or energy storage system. This is the adjusted output power; This is the output adjustment coefficient; This represents the volatility risk index at the current moment. To preset risk thresholds; This is the maximum normalized reference value for the risk index; This represents the degree of deviation of the current volatility risk index from a preset threshold.

[0065] In this embodiment, the operation process of the feedback and scheduling decision module includes: receiving wind power prediction values, power fluctuation risk index, and confidence interval information generated by the prediction output module; calculating system operation status evaluation indicators and power change trend indicators based on the prediction values ​​and risk index; generating scheduling control instructions when the risk index exceeds a preset threshold and sending them to the wind turbine or energy storage system for execution through the SCADA system; storing the adjusted operating parameters and scheduling records in the data storage module; and displaying the prediction results and control execution status in real time through the user interface module to form a closed-loop operation mechanism of prediction-evaluation-control-recording-feedback.

[0066] In this embodiment, the SCADA system is communicatively connected to the feedback and scheduling decision module to receive the operation control commands generated by the feedback and scheduling decision module and to send the operation control commands to the wind turbine or energy storage system for execution. The operation control commands include wind turbine output adjustment commands, pitch angle adjustment commands, or energy storage charging and discharging power adjustment commands to achieve smooth control of wind power and optimization of equipment operating status.

[0067] In this embodiment, the system further includes a data storage module. This data storage module is interconnected with the data acquisition module, prediction output module, and feedback and scheduling decision module to store wind farm operation data, preprocessed feature data, wind power prediction results, operation control commands, and scheduling records. It also provides historical data support to the heterogeneous deep feature extraction module and the physical constraint embedding module to enable model training updates, historical backtracking analysis, and model performance evaluation. Data storage supports local hard drives, NAS, and cloud platform storage structures to meet flexible deployment requirements.

[0068] In this embodiment, the system further includes a user interface module. This user interface module is connected to the prediction output module and the feedback and scheduling decision module for data interaction. It is used to graphically display wind power prediction trends, model confidence intervals, power fluctuation risk indices, historical scheduling records, and the current control command execution status. It is also used to receive manual confirmation or intervention commands. The user interface supports web and mobile access, adapting to various terminal devices and improving ease of operation and maintenance.

[0069] In this embodiment, the system further includes a communication module. This communication module is connected to the data acquisition module, the prediction output module, and the feedback and scheduling decision module. It transmits wind farm operation data, wind power prediction results, and operation control commands to a cloud platform or scheduling control center via wired or wireless networks. It also receives configuration parameters or control strategies from the scheduling control center to achieve coordinated control of multiple wind farms. The communication module supports mainstream industrial protocols such as OPC UA, Modbus TCP, and IEC 61850, facilitating system integration.

[0070] In this embodiment, the heterogeneous deep feature extraction module and / or physical constraint embedding module are used to adaptively adjust the model configuration parameters according to the wind farm type and operating environment information. The wind farm type includes onshore wind farms or offshore wind farms, and the operating environment information includes altitude, ambient temperature, and ambient humidity parameters. The model configuration parameters include input feature combinations, model training hyperparameters, or physical constraint weight coefficients to improve the wind power prediction performance and generalization ability under different scenarios. The heterogeneous deep fusion model supports mainstream deep learning platforms such as TensorFlow, PyTorch, and XGBoost, facilitating rapid deployment by developers.

[0071] In this embodiment, the data acquisition module includes a sampling frequency control unit, which dynamically adjusts the sampling frequency of various sensors based on the wind speed change rate or power fluctuation amplitude. When the wind speed change rate exceeds a preset threshold, the sampling frequency of the corresponding sensor is increased; when the wind speed change rate is below the preset threshold, the sampling frequency is decreased, thereby achieving a balance between data acquisition accuracy and system resource utilization. The data acquisition module supports user-defined configuration, allowing for the addition or removal of sensor types, setting of data acquisition cycles, and channel grouping.

[0072] In this embodiment, the prediction output module is used to predict the trend of wind power output change within a preset time range based on the historical operation data of the wind farm, generate scheduling strategy or maintenance plan information in combination with the current operation status, and push the scheduling strategy or maintenance plan information to the scheduling management system.

[0073] In this embodiment, the feedback and scheduling decision module is used to calculate the operational health assessment index and energy consumption index based on the current operating status parameters of the system and the wind power prediction value generated by the prediction output module, and to generate maintenance suggestion information or maintenance plan information based on the operational health assessment index, and to generate scheduling plan information based on the energy consumption index; when the power fluctuation risk index exceeds a preset threshold or the energy consumption index exceeds a preset upper limit, the feedback and scheduling decision module generates an operational parameter adjustment instruction and sends it to the wind turbine or energy storage system for execution, and pushes the maintenance suggestion information or scheduling plan information to the scheduling management system.

[0074] In this embodiment, the system further includes a data calibration module, which is connected to the data acquisition module for data interaction. The data calibration module is used to calibrate the sensor acquisition data according to a preset time period or under the condition of sensor self-test abnormality triggering. The calibration process includes zero-point offset correction and scale coefficient correction of the acquisition data based on reference sensor data or based on historical statistical benchmarks, and outputting the calibrated data to the data quality processing module to improve the consistency and reliability of the model input data.

[0075] Specifically, the data calibration module is used to calibrate data collected by the wind speed sensor, wind direction sensor, temperature and humidity sensor, and atmospheric pressure sensor. The calibration formula is as follows: in, This is the original data; For calibration deviation; This is the data after calibration.

[0076] Example 2 This invention also provides a heterogeneous deep hybrid wind power intelligent prediction method based on physical constraints, comprising the following steps: (1) collecting wind farm environment and operation parameter data through the IoT sensor network and SCADA system deployed in the wind farm, the data including wind speed, wind direction, temperature, humidity, atmospheric pressure and historical power data at different heights; (2) performing data quality processing on the collected data, including anomaly detection, missing value repair, normalization processing and feature filtering, to obtain input feature data for modeling; (3) using CEEMDAN to perform multi-scale decomposition of wind power time series to obtain multiple intrinsic mode function components and residual components; (4) inputting the decomposed components and input feature data into a heterogeneous deep fusion model for prediction modeling, the heterogeneous deep fusion model including a CNN branch and a Transformer encoder branch, wherein the CNN branch is used to extract local temporal fluctuation features, and the Transformer encoder branch is used to extract local temporal fluctuation features. The r encoder branch is used to model global dependencies over a long period of time and obtain wind power prediction values ​​through feature fusion; (5) PINN is introduced during model training to optimize the physical constraints based on aerodynamic power conversion relationship as physical residual terms and data error terms to improve the physical consistency and generalization ability of the prediction results; (6) A power fluctuation risk index is constructed based on the wind power prediction value. When the risk index exceeds the preset threshold, an early warning message or scheduling suggestion is generated and output to the scheduling management system; (7) The feedback and scheduling decision module generates operating parameter adjustment instructions based on the wind power prediction value and the power fluctuation risk index, and sends them to the wind turbine or energy storage system through the SCADA system to achieve rolling prediction and closed-loop scheduling control; (8) The SHAP method is used to decompose and analyze the feature contribution of the prediction results to quantify the influence of each input feature on the predicted power in order to provide interpretable analysis results.

[0077] In specific embodiments, the present invention can be applied to short-term power prediction and emergency dispatch of offshore wind farms. By analyzing wind speed change trends, the energy storage system can be dispatched in advance to reduce grid connection fluctuation risks.

[0078] The heterogeneous deep ensemble prediction model is a heterogeneous deep fusion model composed of a convolutional neural network (CNN) branch and a Transformer encoder branch. The CNN branch is used to extract local fluctuation features from the wind power time series and multi-source input features, while the Transformer encoder branch is used to model long-term dependencies in the time series. The features extracted by the two branches are fused and input into the prediction output layer to generate wind power prediction results for future periods.

[0079] Furthermore, to improve the adaptability of the heterogeneous deep ensemble prediction model to changes in the real-time operating conditions of wind farms, the heterogeneous deep ensemble prediction model supports an online parameter update mechanism. That is, upon receiving new wind power operating data, the model parameters in the CNN branch, Transformer encoder branch, and feature fusion layer are iteratively updated using the following formula: in, For the first The parameter set of the model at each iteration, including CNN branch parameters, Transformer encoder branch parameters, and feature fusion layer parameters; For the first ; The learning rate; The gradient of the loss function; The loss function; The gradient of the loss function with respect to the model parameters.

[0080] Through the above-mentioned online parameter update mechanism, the heterogeneous deep integrated prediction model can continuously adjust the model parameters according to the changes in real-time wind farm operation data, thereby improving the prediction accuracy and model stability under different operating conditions.

[0081] This invention enables high-precision prediction and intelligent control of wind power, improves the absorption capacity of new energy sources, reduces wind curtailment rate, and is applicable to intelligent operation and optimized scheduling scenarios of various wind farms and regional power grids.

[0082] Example 3 Wind power forecasting and risk warning 1. Various industrial-grade sensor hardware is installed at key locations in the wind farm, including wind speed sensors, wind direction sensors, temperature and humidity sensors, atmospheric pressure sensors, and a wind turbine operation status monitoring module. The wind speed and wind direction sensors are installed at different heights on the weather tower, while the temperature and humidity sensors and pressure sensors are installed in the wind turbine nacelle and tower. The wind turbine operation status monitoring module is used to collect operating parameters such as wind turbine output power, speed, and current. The above sensors are connected to the SCADA system through an IoT communication module to realize real-time acquisition of wind farm environmental and equipment operation data.

[0083] 2. The system cleans and normalizes the collected data to remove outliers and noise points, ensuring the quality and consistency of the input data.

[0084] 3. Extract key influencing factors related to wind power, such as wind speed change rate, wind direction change trend, temperature, humidity, and historical output fluctuation of wind turbines.

[0085] 4. A heterogeneous deep ensemble model based on physical constraints is used to model and analyze the preprocessed data. CEEMDAN is used for multi-scale decomposition of wind power sequences, combined with CNN and Transformer fusion modeling. PINN embedding of aerodynamic power conversion relationships is introduced as physical constraints during training. Simultaneously, the power fluctuation risk value is calculated and the prediction results are output. On the test dataset, the model exhibits excellent prediction performance: the mean absolute error (MAE) is 2.4568, the root mean square error (RMSE) is 4.3424, and the coefficient of determination (R²) is 0.9930. The results show that the method of this invention has high accuracy and robustness, and can effectively characterize complex nonlinear power output changes. Furthermore, SHAP analysis quantifies the contribution of features such as wind speed, wind direction, temperature, and humidity, enhancing prediction transparency and assisting in optimizing grid dispatching and wind farm operation strategies. When the reconstruction error exceeds a preset threshold (e.g., 0.1) or the risk index, the system generates a fluctuation warning signal and triggers the SCADA system to issue control commands. The reconstruction error calculation formula is as follows: in, This represents the magnitude of the difference between the original data and the reconstructed data. Input data representing wind power data, This represents the difference between the original wind power data and the wind power reconfiguration data.

[0086] 5. The SCADA system automatically adjusts the wind turbine's output power or links it with the energy storage system based on power prediction and risk assessment results. The formula for adjusting operating parameters is as follows: in, This is the adjusted output power; This is the original predicted power; Adjustment coefficient (e.g., 0.1); This is the current volatility risk index; The benchmark risk value; This is the maximum risk limit.

[0087] 6. By predicting and sensing wind power output, the system can coordinate dispatch and control in advance to avoid the impact of large power fluctuations on the power grid, improve the grid-friendliness of wind power and dispatch response capability, and at the same time reduce the risk of wind curtailment and improve the efficiency of new energy consumption.

[0088] The embodiments described above are merely preferred embodiments of the present invention and are not intended to limit the scope of the present invention. Various modifications and improvements made by those skilled in the art to the technical solutions of the present invention without departing from the spirit of the present invention should fall within the protection scope defined by the claims of the present invention.

Claims

1. A heterogeneous deep hybrid wind power intelligent prediction system based on physical constraints, characterized in that, The system includes: a data acquisition module, a data quality processing module, a multi-scale decomposition module, a heterogeneous deep feature extraction module, a feature fusion module, a physical constraint embedding module, a prediction output module, an interpretability analysis module, and a feedback and scheduling decision module. The data acquisition module is used to collect real-time environmental parameters and equipment operating parameters of the wind farm through the IoT sensor network and SCADA system deployed in the wind farm. The data quality processing module is used to perform anomaly detection, missing value repair, normalization processing, and feature filtering on the collected raw data. The multi-scale decomposition module is used to decompose the selected features using the fully adaptive noise set CEEMDAN algorithm to obtain multiple intrinsic mode function components and residual components. The heterogeneous deep feature extraction module is used to jointly model and predict the input features after data quality processing and the multi-scale components obtained by CEEMDAN decomposition. The physical constraint embedding module is used to embed physical constraints based on the wind energy aerodynamic power conversion relationship during model training, and to construct a joint optimization objective of physical residual loss term and data error loss term; The feature fusion module is used to couple and fuse the features output by the heterogeneous deep feature extraction module to form a comprehensive feature representation and output the prediction result; The prediction output module is used to generate wind power prediction values ​​for multiple future time points and output power change trends and fluctuation risk indicators. The interpretability analysis module is used to decompose the feature contribution of the prediction results using the SHAP method, and to quantify the influence of each input feature on the prediction power. The feedback and scheduling decision module is used to generate scheduling suggestions based on the prediction results and risk indicators, assisting in the optimization of wind farm operating parameters and grid scheduling decisions, and realizing rolling prediction and dynamic updates.

2. The system according to claim 1, characterized in that, The data acquisition module includes a sampling frequency control unit, which is used to dynamically adjust the sampling frequency of various sensors according to the wind speed change rate or power fluctuation amplitude; the adjustment mechanism is as follows: ； in, This is the original sampling frequency; The adjusted frequency; For adjustment coefficients; This represents the prediction error at the current moment; This is a preset reference threshold; This is the maximum normalized reference value for the error or fluctuation amplitude. This represents the degree of deviation of the current error from the reference threshold.

3. The system according to claim 1, characterized in that, The process of performing anomaly detection, missing value repair, normalization, and feature selection on the collected raw data includes: Anomaly detection of raw collected data is performed based on the Isolation Forest algorithm. Multiple random isolation trees are constructed to recursively partition the samples until a sample point is isolated. The path length of a sample within the random isolation trees is used as the anomaly detection criterion. An average path length normalization factor is applied to normalize the path length for different sample sizes. The calculation formula is as follows: ； in, For the sample size The average path length normalization factor at that time; for The harmonic number; The number of samples used to construct the randomized isolation tree; For missing data resulting from anomaly removal and the original missing data, interpolation is used for completion. The completion process is expressed as follows: ； in, Indicates the time to be completed or estimated. Data values; In a known data sequence, time... The corresponding known data value; In a known data sequence, time... The corresponding known data value; This indicates the time point where data needs to be supplemented. After anomaly repair and missing data completion, to eliminate the impact of differences in the dimensions of different feature variables on model training, each feature variable is normalized. The normalization process is expressed as follows: ； in, Normalized data representing various relevant parameters of a wind farm. This represents the raw data of a wind farm. This represents the minimum value in the wind farm dataset. This represents the maximum value in the wind farm dataset.

4. The system according to claim 1, characterized in that, The process of decomposing the selected features using the fully adaptive noise set CEEMDAN algorithm to obtain multiple intrinsic mode function components and residual components includes: ； in, The reconstructed signal; Extracted by Empirical Mode Decomposition (EMD) eigenmode functions , representing the different frequency components of a signal, It is the residual, which represents the long-term trend of the signal; For reconstruction operations, it means combining multiple IMFs and residuals into a single overall signal. For time.

5. The system according to claim 1, characterized in that, The process of jointly modeling and predicting the input features after data quality processing and the multi-scale components obtained by CEEMDAN decomposition includes: the CNN branch and the Transformer encoder branch. The CNN branch convolution operation is represented as: ； in, The output of the CNN layer is a weighted sum of the input data; These are the weights of the convolution kernel; Input timing data; For bias terms; The size of the convolution kernel; The calculation formula for the multi-head self-attention mechanism of the Transformer encoder branch is as follows: ； in, The query value is a vector obtained from the input data through a linear transformation; The key vector represents the feature of each input data in the model; It is a value vector, representing the actual information of the input data; This is the scaling factor; The dimensions of the query and key vectors; To query the dot product of the key vector and the key vector.

6. The system according to claim 1, characterized in that, The process of embedding physical constraints based on wind energy aerodynamic power conversion relationships during model training, and constructing a joint optimization objective by combining physical residual loss terms and data error loss terms, includes: Using aerodynamic formulas, the power conversion relationship of wind turbine units is constructed as a physical constraint: ； in, This refers to the power output of the wind turbine generator; air density; The swept area of ​​the wind turbine; This represents the power coefficient of the fan. Wind speed; By combining physical constraints with data error terms, training is performed through joint optimization, ultimately optimizing the objective function. for: ； in, For data error loss, This is the physical residual loss term constructed based on the aerodynamic power relationship. These are the physical constraint weighting coefficients.

7. The system according to claim 1, characterized in that, The process of coupling and fusing the features output by the heterogeneous deep feature extraction module to form a comprehensive feature representation and output the prediction result includes: ； in, This is the final feature output after weighting; The weighting coefficients represent the importance of the two features and control the contribution of each branch to the final fusion result. Output features for CNN branches; This is the output feature of the Transformer encoder branch.

8. The system according to claim 1, characterized in that, The process of generating wind power forecasts for multiple future time points and outputting power change trends and fluctuation risk indicators includes: ； in, To predict the power value, Historical power values As a risk index, The trend is predicted to be: ； in, Forecasted wind power output for the future; This is a trend adjustment based on historical data, representing the changing trend of wind power output; The volatility risk index is: ； in, For the Sigmoid function; For feature input; As weight; For bias terms, To predict the probability of risk.

9. The system according to claim 1, characterized in that, Based on the forecast results and risk indicators, scheduling suggestions are generated to assist in the optimization of wind farm operating parameters and grid scheduling decisions. The process of achieving rolling forecasting and dynamic updating includes: Receive wind power forecasts, power fluctuation risk indices, and confidence intervals generated by the forecast output module; Based on wind power forecast values ​​and power fluctuation risk index, calculate system operation status evaluation indicators and power change trend indicators; When the system operation status evaluation index and power change trend index exceed the preset threshold, a scheduling control command is generated and sent to the wind turbine or energy storage system for execution through the SCADA system. The adjusted operating parameters and scheduling records are stored in the data storage module; and the prediction results and control execution status are displayed in real time through the user interface module to form a closed-loop operation mechanism of prediction-evaluation-control-recording-feedback.

10. A method for intelligent prediction of heterogeneous deep hybrid wind power based on physical constraints, wherein the method is implemented by the system described in any one of claims 1-9, characterized in that, The method includes: Real-time collection of environmental and operational parameter data is achieved through an IoT sensor network and SCADA system deployed in the wind farm. The collected data undergoes anomaly detection, missing value repair, normalization, and feature filtering to obtain input feature data for modeling. CEEMDAN was used to perform multi-scale decomposition on the wind power time series, resulting in multiple intrinsic mode function components and residual components. The decomposed components and input feature data are input into a heterogeneous deep fusion model composed of CNN branches and Transformer encoder branches for training and prediction, and the wind power prediction value is output through feature fusion. PINN is introduced during model training to construct the aerodynamic power conversion relationship as a physical residual term and jointly optimize it with the data error term. Based on the prediction results, a power fluctuation risk index is constructed and early warning information and scheduling suggestions are generated. The feedback and scheduling decision module converts the prediction results into operation control commands and sends them to the wind turbine or energy storage system through the SCADA system for execution, so as to realize power balance and closed-loop optimized operation of the wind power system.

Images