Offshore wind power prediction method based on seasonal adaptive wind condition identification

Abstract

The application discloses a sea wind power prediction method based on seasonal self-adaptive wind condition identification and belongs to the field of wind power prediction.The method solves the problems of frequent alternation of sea wind conditions in four seasons, "power cliff" distortion in gale days, large fluctuation error in light wind days, insufficient prediction accuracy in normal wind conditions, and poor adaptability of general models.The three-stage scheme of seasonal self-adaptive wind condition identification, multi-model special prediction and dynamic wake power aggregation is adopted, different thresholds are set according to seasons to distinguish gale days, light wind days and normal wind conditions, corresponding special models are called for prediction respectively, and the full-field power is corrected through a dynamic Jensen wake model and a spatial aggregation algorithm.The prediction accuracy of the four seasons is significantly improved, the deviation in gale days is less than 7.8%, the MAPE in light wind days is less than 9.2%, the MAPE in normal wind conditions is stably within 7.0%, the power grid consumption capacity is effectively improved, the examination loss is reduced, and support is provided for safe and stable operation of the sea wind power in the four seasons with frequent alternation of wind conditions.

Description

Technical Field

[0001] This invention relates to the field of wind power prediction technology, and in particular to a method for predicting offshore wind power based on seasonal adaptive wind condition identification. Background Technology

[0002] Offshore wind power, as a crucial component of clean energy, relies heavily on accurate power forecasting for grid dispatch and energy management. Precise wind power forecasting not only helps grid operators rationally plan power generation, improving the stability and reliability of the power system, but also effectively reduces reserve capacity requirements caused by wind power fluctuations, thereby lowering operating costs. However, due to the complexity and diversity of coastal climate characteristics, offshore wind power forecasting faces numerous challenges.

[0003] Currently, common wind power forecasting methods mainly rely on unified prediction models. These models are typically constructed based on historical meteorological data and wind farm operation data, using statistical analysis and machine learning algorithms. While these methods can predict wind power to some extent, their accuracy is often significantly affected by region-specific climate patterns. Particularly in coastal areas of my country, such as the Yangjiang sea area, wind conditions vary significantly throughout the four seasons, with frequent alternations of strong winds, light winds, and normal wind conditions, posing considerable challenges to wind power forecasting. Normal wind conditions, as the dominant wind condition during seasonal transitions, exhibit stable wind speeds and small power fluctuations, falling between the thresholds of strong and light winds. However, existing technologies lack specific modeling for this type of wind condition, resulting in insufficient prediction accuracy.

[0004] In windy conditions, when wind speeds exceed the rated wind speed, wind turbine generators enter a power-limited operation state. At this time, power output no longer increases linearly with increasing wind speed, but instead exhibits a "power cliff" phenomenon, meaning a sharp drop in power. Traditional prediction models, such as the power prediction method, device, storage medium, and system for offshore wind farms disclosed in CN115034159A, which mentions a method based on single numerical weather prediction (NWP) and physical modeling, often fail to accurately capture the power decline trend because they do not fully consider this special wind condition, leading to significant deviations in prediction results.

[0005] In light wind conditions, when wind speeds are below or fluctuate around the cut-in wind speed, turbulent wind direction and unstable airflow lead to highly unstable power output from wind turbines. In such cases, traditional prediction models, such as the model based on historical data and simple machine learning algorithms mentioned in CN118296475A's multi-model prediction method for offshore wind power, lack effective fusion of multi-source real-time monitoring data. This makes it difficult to accurately reflect the impact of minute changes in wind speed and direction on power output, resulting in significant prediction errors.

[0006] Under normal wind conditions, wind speed is stable and power fluctuations are small. However, due to the indirect influence of seasonal meteorological factors (such as spring sea breezes, summer subtropical high pressure, and autumn and winter northeast monsoons), there are still problems such as slight fluctuations in wind speed and changes in air density. Traditional unified prediction models are not designed with a specific architecture for this type of wind condition. They either use models for high wind days, resulting in computational redundancy and overly smoothed predictions, or they use models for low wind days, resulting in insufficient capture of stable time series patterns. They cannot meet the requirements of prediction accuracy and stability under normal wind conditions, and prediction biases will also occur, making it difficult to meet the actual needs of wind farm full-condition prediction.

[0007] Furthermore, existing commonly used prediction models, such as the power prediction method, device, storage medium and system for offshore wind farms disclosed in CN115034159A, the multi-model prediction method for offshore wind power disclosed in CN118296475A, and the rapid prediction system and method for high-power offshore wind farms based on refined wind resource assessment and grid integration disclosed in CN121688824A, generally suffer from the following technical defects and shortcomings: 1. The balance between model complexity and computational efficiency: Although CN121688824A achieves high-precision prediction through multi-module collaboration, its deep neural network and complex interactive operations may lead to a heavy computational burden, affecting real-time prediction efficiency. Especially in the scenario of large-scale wind farms in deep-sea areas, it is difficult to meet the real-time requirements of ultra-short-term power prediction.

[0008] 2. Insufficient adaptability to extreme weather conditions: Although the existing schemes take into account conventional meteorological factors, their dynamic response and real-time modeling capabilities for extreme weather such as typhoons and severe convection are limited, which may lead to a significant increase in prediction bias.

[0009] 3. Data Dependence and Data Quality Risks: It is highly dependent on multi-source observation data (such as NWP, buoy observation, and LiDAR (Light Detection and Ranging) wind measurement data), but the data quality is affected by factors such as sensor accuracy and transmission delay. If the data preprocessing is not good, noise or missing values ​​may be introduced, which will affect the accuracy of key parameter calculation and prediction.

[0010] 4. Simplified assumptions in wake effect modeling: While the Jensen wake model can simulate the wake effect of wind turbines, it is based on idealized assumptions and may not fully reflect the actual wake distribution under complex terrain and turbulent conditions. This is especially true in areas with dense wind farm layouts or significant terrain undulations, potentially leading to errors in power generation calculations. Furthermore, existing models do not dynamically adjust the wake attenuation coefficient based on the differences between strong winds, light winds, and normal wind conditions, further exacerbating errors in overall power aggregation.

[0011] 5. Insufficient dynamic adaptability of the integrated power grid model: The integrated power grid module achieves stability analysis through grid-connected dynamic response simulation, but the dynamic feedback on power grid operating characteristics (such as load fluctuations and changes in the output of other power sources) may be insufficient, making it difficult to adjust the predicted power curve in a timely manner to adapt to sudden changes in the power grid state.

[0012] 6. Limitations of model update and iteration mechanism: Existing solutions support quarterly incremental updates and parameter iterations, but the update frequency and triggering conditions may not be flexible enough, making it difficult to adapt to rapidly changing climate patterns or wind farm equipment status, which may lead to long-term operational accuracy degradation, especially in the inability to adapt to the subtle characteristics of normal wind conditions changing with the seasons.

[0013] 7. Lack of multi-model fusion strategy: For example, CN121688824A does not explicitly introduce a multi-model fusion mechanism, which cannot make full use of the advantages of different prediction models (such as physical models, statistical models, and deep learning models), thus limiting the improvement of overall prediction robustness and also failing to achieve dedicated model switching for windy days, light wind days, and normal wind conditions.

[0014] 8. Trade-off between economy and scalability: While high-precision modeling and complex algorithms improve prediction accuracy, they may increase system deployment and maintenance costs, limiting their widespread application in small and medium-sized wind farms or resource-limited scenarios.

[0015] Especially for the unique climatic environment of the Yangjiang sea area, the shortcomings of existing technologies become even more apparent. For example, CN121688824A points out that existing methods either ignore the wake effect of wind farms and turbine response, have insufficient prediction accuracy, are computationally inefficient, or lack coupling with the grid level, failing to meet the needs of grid dispatch and energy management for high-precision wind power prediction. Particularly noteworthy is that existing methods do not consider the specific prediction needs of normal wind conditions, failing to achieve accurate predictions under all operating conditions—strong winds, light winds, and normal wind conditions—and are ill-suited to the complex wind conditions of the sea area during the four seasons.

[0016] Therefore, there is an urgent need for a high-precision offshore wind power prediction method that can adapt to strong winds, light winds, and normal wind conditions throughout the four seasons, and possesses adaptive wind condition recognition, multi-model specialized prediction, and dynamic wake aggregation, to improve the prediction stability and accuracy under complex wind conditions. This invention addresses this problem by proposing an offshore wind power prediction method based on seasonal adaptive wind condition recognition. It aims to significantly improve the prediction accuracy for various wind conditions by constructing a multi-mode adaptive prediction system, combining adaptive wind condition recognition, specialized modeling for power limiting in strong winds, multi-source data fusion for light winds, stable prediction under normal wind conditions, dynamic wake correction, and integrated coupling between the power grid and the grid. This provides more reliable technical support for grid dispatching and energy management. Summary of the Invention

[0017] The technical problem this invention aims to solve is to provide a method for predicting offshore wind power based on seasonal adaptive wind condition identification. This method overcomes the shortcomings of existing offshore wind power prediction technologies, which fail to customize modeling for the frequent alternation of strong, light, and normal wind conditions in specific sea areas throughout the year. These shortcomings lead to distorted predictions of the "power cliff" during strong winds, large prediction errors for power fluctuations during light winds, and insufficient accuracy and stability in predicting normal wind conditions. By constructing a three-level collaborative system—seasonal adaptive wind condition identification, multi-model specialized prediction, and dynamic wake power aggregation—and designing dedicated prediction models for each of the three wind conditions, the method significantly improves the accuracy of wind power prediction under different wind conditions throughout the year, thereby enhancing grid dispatch security and the economic efficiency of wind farm operation.

[0018] To achieve the above technical objectives, the present invention adopts the following technical solution: This paper presents a method for predicting offshore wind power based on seasonal adaptive wind condition recognition. The overall process includes three stages: wind condition pattern recognition, calculation of a special prediction model, and dynamic wake power aggregation, which can achieve high-precision power prediction from a single wind turbine to the entire wind farm.

[0019] 1. Seasonal adaptive wind condition pattern recognition Based on nearly ten years of measured meteorological and wind turbine SCADA operation data for the sea area, a threshold system for determining strong and light wind days by season was established. Combined with the dominant seasonal meteorological factors, automatic wind condition classification was completed. The identification rules are as follows: Spring: Wind speed ≥ 10 m / s is considered a windy day, wind speed < 6 m / s is considered a light wind day, and wind speed ≤ 6 m / s < 10 m / s is considered a normal wind condition, affected by cold wave gusts and sea and land breezes; Summer: Wind speed ≥ 9 m / s is considered a windy day, wind speed < 5 m / s is considered a light wind day, and wind speed ≤ 5 m / s < 9 m / s is considered a normal wind condition, influenced by the southwest monsoon and the subtropical high pressure. Autumn: Wind speed ≥ 11 m / s is considered a windy day, wind speed < 7 m / s is considered a light wind day, and wind speed 7 m / s ≤ wind speed < 11 m / s is considered a normal wind condition, affected by the northeast monsoon and sea fog. Winter: Wind speed ≥ 12 m / s is considered a strong wind day, wind speed < 8 m / s is considered a light wind day, and wind speed 8 m / s ≤ wind speed < 12 m / s is considered a normal wind condition, affected by stable northeast monsoon and low temperature sea fog.

[0020] The wind condition identification module takes into account numerical weather prediction (NWP) data, SCADA real-time data, and wind tower monitoring data, and outputs the current wind condition mode (strong wind mode, light wind mode, or normal wind mode) and triggers the corresponding special prediction model.

[0021] 2. Multi-model specialized forecasting It addresses the issues of power distortion due to high winds, unstable output on light winds, and insufficient prediction accuracy under normal wind conditions, thereby achieving high-precision single-unit power prediction under different wind conditions.

[0022] (1) Strong wind prediction model It adopts a physical-data hybrid driving architecture, focusing on capturing the "power cliff" characteristic of power operation after wind speed exceeds rated limit, and adapts to the meteorological characteristics of strong winds in different seasons.

[0023] Input features: NWP wind speed and direction, real-time pitch angle, power limit status flag, wind turbine theoretical power curve, historical 15-day SCADA operation data, with additional gust factor and wind speed change rate added in spring and summer, and additional severe convective weather index added in autumn and winter. Model configuration: Spring / Summer: WT-CNN-LSTM+Gust module (5 wavelet transform WT decomposition layers, CNN convolution kernel size 3×3, 64 channels, 128 hidden units in LSTM, 0.35 weight coefficient of Gust module). Autumn / Winter: GNN-Transformer + Strong Convection Module (GNN node dimension 64, Transformer encoder layer 4 layers, attention head 8, strong convection module weight coefficient 0.40). Output: Initial predicted power sequence of a single wind turbine for the next 0 to 72 hours in 15-minute increments, providing basic data for subsequent wake correction.

[0024] (2) Prediction model for light wind days A multi-source data fusion architecture is adopted to improve the forecast stability under low wind speed, turbulent wind direction, and sea fog interference, and to adapt to the meteorological characteristics of light wind days in different seasons.

[0025] Input features: real-time wind speed / direction from multiple meteorological towers, turbulence intensity, vertical wind shear, sea surface temperature, real-time air density, and sea fog concentration. In spring and summer, the subtropical high pressure index is added, and in autumn and winter, a low temperature correction coefficient is added. Model configuration: Spring and Summer: SARIMA + Density Correction Model (SARIMA model order (p=2, d=1, q=1) (P=1, D=0, Q=1), density correction module weight coefficient 0.35); Autumn and winter: LSTM + sea fog-density coupling model (LSTM hidden layer unit number 256, sea fog-density coupling module weight coefficient 0.40); Output: Initial predicted power sequence of a single wind turbine for 10 minutes every 0 to 24 hours, solving the problem of large power fluctuations caused by low wind speed and sea fog interference on windy days.

[0026] (3) Normal wind condition prediction model It adopts a hybrid architecture of data-driven and physical constraints, which is adapted to the stable wind speed and small power fluctuation during the seasonal transition period, and balances prediction accuracy and stability.

[0027] Input features: NWP wind speed / direction, LiDAR wind measurement data, real-time turbine pitch angle, historical 7-day SCADA operation data, turbulence intensity, vertical wind shear, real-time air density, and theoretical power curve of the turbine; Model configuration: CNN-LSTM hybrid temporal prediction model (CNN convolutional kernel size 3×3, number of channels 64, LSTM hidden layer number of units 128, with Dropout layer introduced to prevent overfitting); Output: Initial predicted power sequence of a single wind turbine for the next 0-48 hours in 15-minute increments. The output is stable and without abnormal jumps, which is consistent with the operating characteristics under normal wind conditions, providing reliable basic data for subsequent wake correction.

[0028] 3. Dynamic wake power aggregation By coupling a dynamic wake model with a spatial aggregation algorithm, wake interference and spatial heterogeneity between wind turbines are eliminated, enabling accurate power synthesis across the entire field.

[0029] (1) Dynamic wake calculation The downstream wind turbine velocity attenuation is calculated using the Jensen wake model or its improved version, as shown in the following formula: (1); In the formula, For downstream fan speed, The incoming air velocity from the upstream fan. It is an axial inducing factor. The wake attenuation coefficient is... The distance between upstream and downstream wind turbines. This refers to the diameter of the fan impeller.

[0030] wake attenuation coefficient The value is dynamically adjusted according to the wind pattern: 0.06~0.09 for strong winds, 0.03~0.05 for light winds, and 0.05~0.06 for normal winds.

[0031] (2) Total wake loss superposition No. Typhoon total wake loss factor All upstream wind turbines within the same wind farm are connected to the first The cumulative effect of the typhoon's wake is as follows: (3); In the formula, For the first The total wake loss factor of the typhoon turbine. For the first Typhoon machine to the first The weighting factor of typhoon generators For the first Typhoon machine to the first The wake effect coefficient of a single typhoon turbine.

[0032] (3) Full-field power aggregation The total output power of the wind farm is obtained by combining the corrected power of a single wind turbine with the wake loss and summing them over the entire field. (2); In the formula, This represents the total predicted power of the wind farm. For the first Typhoon generator power after correction For the first Total wake loss coefficient of typhoon generator.

[0033] The offshore wind power prediction method based on seasonal adaptive wind condition identification provided by this invention has the following beneficial effects: 1. This invention designs a multi-mode adaptive prediction system. This system comprehensively considers the complex climate characteristics of a specific sea area, covering normal wind conditions during windy days, light wind days, and the transition period between seasons. Through the deep integration of intelligent algorithms and meteorological models, it effectively solves the problem that the accuracy of offshore wind power prediction is limited by variable climate conditions, and significantly improves the accuracy and reliability of prediction.

[0034] 2. This invention constructs different specialized prediction models for strong winds, light winds, and normal wind conditions. The strong wind model focuses on dealing with the "power cliff" phenomenon, the light wind model focuses on capturing weak signals, and the normal wind condition model is adapted to the conditions of stable wind speed and small power fluctuations, taking into account both prediction accuracy and stability. Through refined modeling and optimized algorithms, the prediction accuracy of wind power under various wind conditions is significantly improved. The prediction deviation for strong winds is significantly reduced to below 7.8%, the mean absolute percentage error (MAPE) for light winds is controlled within 9.2%, and the MAPE for normal wind conditions is stably controlled within 7.0%.

[0035] 3. This invention establishes a seasonally adaptive wind condition identification standard. Based on years of meteorological data from the sea area, the standard sets differentiated wind condition thresholds according to the seasons, clearly defines the judgment ranges of strong wind days, light wind days, and normal wind conditions, and combines real-time meteorological monitoring data to successfully overcome the limitations of traditional models that cannot accurately identify wind conditions in different seasons and switch to the corresponding models. This achieves intelligent and accurate wind condition identification, and can accurately identify various wind conditions in all four seasons and trigger the corresponding exclusive models.

[0036] 4. This invention adopts a method of automatically identifying wind conditions and switching models according to the seasonal climate characteristics of the sea area. It can automatically identify windy days, light wind days and normal wind conditions. Through the built-in intelligent decision-making module, the system can analyze the current climate conditions in real time and automatically select the most suitable prediction model for calculation. This achieves strong self-adaptation of the system to seasonal wind conditions and can perfectly adapt to complex scenarios such as cold waves, monsoons, sea fog, strong convection and normal wind conditions during the transition period of the four seasons.

[0037] 5. The system constructed by this invention has a high degree of automation and intelligence. It can automatically adapt to the different climate characteristics of the sea area in all four seasons. Through continuous meteorological monitoring and data analysis, it can automatically identify windy days, light wind days and normal wind conditions and switch the corresponding prediction model. It can achieve full coverage prediction of all working conditions throughout the year without human intervention, which greatly improves the prediction efficiency and accuracy.

[0038] 6. This invention introduces relevant technologies to enhance the grid's wind power absorption capacity. By optimizing the wind power prediction algorithm, it covers predictions for all scenarios, including strong winds, light winds, and normal wind conditions, reducing prediction deviations under various wind conditions. This increases the grid's ability to accept wind power by 5% to 8%, effectively reduces the wind curtailment rate, improves the utilization efficiency of new energy sources, and makes a positive contribution to the development of green energy.

[0039] 7. This invention adopts a method to reduce assessment costs due to prediction deviations. By improving the accuracy of wind power prediction and covering predictions under all operating conditions, including normal wind conditions, it reduces grid assessment costs caused by prediction deviations under various wind conditions, directly reducing assessment costs by 15% to 20%, while improving the economic benefits of wind farms and enhancing their competitiveness.

[0040] 8. This invention designs a scheme to provide a basis for power grid dispatch and wind farm optimization. The scheme provides reliable decision support for power grid dispatch by predicting wind power in real time and accurately, covering windy days, light wind days and normal wind conditions. It also provides a scientific basis for the optimization of wind farm operation. It has outstanding engineering application value and broad market prospects.

[0041] 9. This invention integrates physical constraints and data-driven methods to establish a model that balances theoretical accuracy and field adaptability. Corresponding physical constraints are incorporated for strong winds, light winds, and normal wind conditions. This model not only considers the physical characteristics of offshore wind power but also makes full use of the advantages of big data and artificial intelligence technologies, thereby improving the robustness of the model and enabling it to operate stably for a long time, thus providing a strong guarantee for the sustainable development of offshore wind power.

[0042] 11. This invention employs wind-specific models to address the issues of power limitation under strong winds, weak signals under light winds, and stable operation under normal wind conditions. By designing specific prediction models for different wind conditions, the relevance and accuracy of the models are improved, while also enhancing their robustness, making the prediction results more reliable and stable under various wind conditions.

[0043] 12. This invention introduces dynamic wake correction technology, which calculates the wind speed attenuation between wind turbines in real time and dynamically adjusts the wake attenuation coefficient according to windy days, light wind days and normal wind conditions. This effectively solves the problem of power aggregation error caused by spatial heterogeneity and improves the accuracy of power prediction.

[0044] 13. This invention designs a scheme that can output a long-term forecast of 0~72 hours + an ultra-short-term forecast of 0~6 hours. The forecast is output for 0~72 hours in windy weather, for 0~24 hours in light wind weather, and for 0~48 hours in normal wind conditions. This satisfies both the grid dispatching requirements for long-term forecasts and the wind farm optimization operation requirements for ultra-short-term forecasts, providing strong support for the refined management and optimized operation of offshore wind power.

[0045] 14. This invention adopts technology that supports seamless integration with SCADA, NWP, and dispatch master station. Through standardized data interfaces and communication protocols, it realizes fully automated prediction of wind conditions, including strong winds, light winds, and normal wind conditions. This reduces manual intervention and operation and maintenance costs, improves prediction efficiency and accuracy, and lays the foundation for intelligent management of offshore wind power.

[0046] 15. The design of this invention fully considers the typical meteorological characteristics of the sea area, such as cold waves, monsoons, sea fog, and strong convection, while also taking into account the stable characteristics of normal wind conditions during the transition period between the four seasons. Through targeted modeling and optimization, the regional adaptability and all-condition adaptability are improved, making the prediction results more consistent with the actual situation, with high engineering feasibility and broad application prospects.

[0047] 16. This invention adopts a method that supports quarterly incremental updates and parameter iterations. By continuously collecting and analyzing real-time meteorological data and wind power data, it continuously optimizes the prediction model and parameters, achieving long-term operational accuracy without attenuation. This meets the usage requirements of the wind farm throughout its entire life cycle and provides a strong guarantee for the long-term stable operation of offshore wind power.

[0048] 17. The power aggregation module constructed in this invention adopts a combination of dynamic wake model and spatial aggregation algorithm. By comprehensively considering the wake influence between wind turbines and spatial heterogeneity, the wake attenuation coefficient is dynamically adjusted according to the wind condition mode, thereby realizing accurate aggregation of power across the entire field and improving the accuracy of power prediction. Attached Figure Description

[0049] The present invention will be further described below with reference to the accompanying drawings and embodiments: Figure 1 This is a flowchart illustrating the prediction method of the present invention. Figure 2 This is a flowchart illustrating the operation of the power aggregation module of the present invention. Detailed Implementation

[0050] The technical solution of the present invention will be further described below with reference to the accompanying drawings and embodiments: Example 1 like Figure 1 As shown, this embodiment provides a method for predicting offshore wind power based on seasonal adaptive wind condition identification. It addresses the high-wind conditions in the Yangjiang sea area caused by spring cold waves and alternating sea and land breezes, as well as the high-wind conditions caused by the summer southwest monsoon, subtropical high pressure, and frequent strong convection. The method combines... Figure 1 , Figure 2 The specific implementation process of this invention is described in detail below, with the specific steps as follows: Step 1: Automatic determination of seasonal wind condition mode Based on nearly 10 years of measured meteorological data for spring and summer in the Yangjiang sea area, and wind turbine SCADA (Supervisory Control and Data Acquisition) operation data, specific wind condition identification standards for spring and summer were established: Spring: wind speed ≥ 10 m / s is determined to be a strong wind day, and wind speed < 6 m / s is determined to be a light wind day; Summer: wind speed ≥ 9 m / s is determined to be a strong wind day, and wind speed < 5 m / s is determined to be a light wind day.

[0051] In this embodiment, real-time NWP (Numerical Weather Prediction) data and real-time wind speed data at a height of 100m from a wind tower were collected from an offshore wind farm in Yangjiang during spring and summer. The calculated 10-minute average wind speed is 12.3m / s, which meets the threshold for strong winds in spring and summer. The wind condition is automatically identified as a strong wind mode, thus completing the wind condition mode classification.

[0052] Step 2: Calling up the special prediction model for strong winds and predicting the power of individual units Based on the wind pattern recognition results from step 1, the WT-CNN-LSTM+gust module-specific physical-data hybrid prediction model for spring and summer windy days is invoked. This prediction model uses wavelet transform (WT), convolutional neural network (CNN), and long short-term memory (LSTM) as its core architecture, and additionally introduces a gust physical correction module. On the basis of data-driven fitting, it incorporates physical prior constraints such as wind turbine power limit, pitch angle action, wind speed jump, and power cliff, to achieve high-precision prediction of wind power during spring and summer windy days. Input features: NWP wind speed / direction, real-time turbine pitch angle, power limit status flag, theoretical power curve of the turbine, and historical SCADA operation data for the past 15 days; Model parameter settings: Wavelet transform (WT) decomposition layer is 5 layers, CNN (convolutional neural network) convolution kernel size is 3×3 and number of channels is 64, LSTM (long short-term memory network) hidden layer unit number is 128, gust module weight coefficient is 0.35, the optimizer is Adam (Adaptive Moment Estimation), the initial learning rate is 0.001, and the training batch size is 32; Model output: Initial predicted power sequence of a single wind turbine every 15 minutes for the next 0 to 72 hours, completing the power prediction of a single turbine and providing basic data for subsequent wake correction.

[0053] Step 3: Dynamic wake correction and full-field power aggregation Combination Figure 2 The dynamic wake calculation and power aggregation process performs the following operations: Input for dynamic wake calculation: Input the single-unit power prediction sequence, real-time wind direction and speed, wind turbine coordinates and impeller parameters (impeller diameter D=220m), and environmental parameters (air density, turbulence intensity) output from step 2 into the dynamic wake calculation module; Secondary wind condition determination: Based on real-time wind speed, the windy weather conditions are reconfirmed, and the high-speed wake model (improved Jensen model) is invoked. The calculation formula is as follows: (1); In the formula, For downstream fan speed, The incoming air velocity from the upstream fan. The axial induction factor is set to 0.25. The wake attenuation coefficient is 0.075 (taken as 0.075 on windy days). The distance between upstream and downstream wind turbines. This refers to the diameter of the fan impeller.

[0054] Effective wind speed calculation: Based on the high-wind-speed wake model, the wake effects of all upstream wind turbines on downstream wind turbines within the wind farm are superimposed to calculate the effective wind speed at each turbine location. The formula for calculating the total wake loss coefficient is as follows: (3); In the formula, For the first The total wake loss factor of the typhoon turbine. For the first Typhoon machine to the first The weighting factor of typhoon generators For the first Typhoon machine to the first The wake effect coefficient of a single typhoon turbine.

[0055] Spatial power aggregation: The initial predicted power of a single unit is corrected by combining the wake loss coefficient, and then the power of the entire field is superimposed using a spatial aggregation algorithm. The calculation formula is as follows: (2); In the formula, This represents the total predicted power of the wind farm. For the first Typhoon generator power after correction For the first Total wake loss coefficient of typhoon generator.

[0056] Output results: The total power prediction value of the wind farm in the next 0-72 hours is output. According to actual measurement, the prediction deviation of this embodiment on windy days is 7.2%, which meets the accuracy requirement of ≤7.8%.

[0057] Example 2 In another preferred embodiment, based on Embodiment 1 above, this embodiment provides a method for predicting offshore wind power based on seasonal adaptive wind condition identification. This method addresses the high wind conditions in the Yangjiang sea area caused by the combined effects of the autumn and winter northeast monsoon, strong convection, and low-temperature sea fog. Figure 1 , Figure 2 The specific implementation process of this invention is described in detail below, with the specific steps as follows: Step 1: Automatic determination of seasonal wind condition mode Based on nearly 10 years of measured meteorological data and wind turbine SCADA operation data from the Yangjiang sea area during autumn and winter, a wind condition identification standard specifically for autumn and winter was established: wind speeds ≥11 m / s in autumn are classified as gale-force winds, and wind speeds ≥12 m / s in winter are classified as gale-force winds. In this embodiment, real-time NWP data and real-time wind speed data at a height of 100m from a wind tower were collected from an offshore wind farm in Yangjiang during autumn and winter. The calculated 10-minute average wind speed was 12.8 m / s, which meets the gale-force wind threshold for autumn and winter, and was automatically identified as a gale-force winds, thus completing the wind condition mode classification.

[0058] Step 2: Calling up the special prediction model for strong winds and predicting the power of individual units Based on the strong wind pattern recognition results from step 1, a physical-data hybrid prediction model specifically designed for strong winds in autumn and winter is invoked, utilizing the GNN-Transformer + strong convection module. This prediction model consists of three coupled parts: a Graph Neural Network (GNN) spatial feature extraction layer, a Transformer temporal feature learning layer, and a strong convection physical constraint correction layer. It incorporates prior physical knowledge such as wind turbine power limits, pitch angle action, sudden gust changes, and strong convective disturbances on a data-driven basis, achieving accurate fitting of the "power cliff" and strong convective wind conditions. Input features: NWP wind speed / direction, real-time turbine pitch angle, power limit status flag, theoretical power curve of turbine, historical 15-day SCADA operation data, severe convective weather index; Model parameter settings: GNN (Graph Neural Network) node dimension 64, Transformer encoder layer 4 layers, attention head 8, strong convection module weight coefficient 0.40, optimizer Adam, initial learning rate 0.001, training batch size 32; Model output: Initial predicted power sequence of a single wind turbine every 15 minutes for the next 0 to 72 hours, completing the power prediction of a single unit and accurately capturing the "power cliff" characteristics of power limitation on windy days.

[0059] Step 3: Dynamic wake correction and full-field power aggregation Combination Figure 2 The dynamic wake calculation and power aggregation process performs the following operations: Input for dynamic wake calculation: Input the single-unit power prediction sequence, real-time wind direction and speed, wind turbine coordinates and impeller parameters (impeller diameter D=220m), and environmental parameters (air density, turbulence intensity) output from step 2 into the dynamic wake calculation module; Secondary wind condition determination: Based on real-time wind speed, the windy weather conditions are reconfirmed, and the high-speed wake model (improved Jensen model) is invoked. The calculation formula is the same as in Example 1, and the wake attenuation coefficient is used. Take 0.075; Effective wind speed calculation: Based on the high wind speed wake model, the wake effects of all upstream wind turbines on downstream wind turbines in the wind farm are superimposed to calculate the effective wind speed at the location of each wind turbine. The total wake loss coefficient is calculated in the same way as in Example 1. Spatial power aggregation: The initial predicted power of a single unit is corrected by combining the wake loss coefficient, and then the power of the entire field is superimposed by the spatial aggregation algorithm. The calculation formula is the same as in Example 1. Output results: The total power prediction value of the wind farm in the next 0-72 hours is output. According to actual measurement, the prediction deviation of this embodiment on windy days is 7.5%, which meets the accuracy requirement of ≤7.8%.

[0060] Example 3 In another preferred embodiment, based on Embodiment 1 above, this embodiment provides a method for predicting offshore wind power based on seasonal adaptive wind condition identification. This method addresses the low wind speeds and turbulent wind conditions caused by spring cold waves and alternating sea and land breezes in the Yangjiang sea area, as well as the frequent occurrence of low wind speeds and turbulent wind directions due to the southwest monsoon, subtropical high pressure, and strong convection in summer. Figure 1 , Figure 2 The specific implementation process of this invention is described in detail below, with the specific steps as follows: Step 1: Automatic determination of seasonal wind condition mode Based on nearly 10 years of measured meteorological data and wind turbine SCADA operation data from the Yangjiang sea area during spring and summer, a wind condition identification standard specifically for spring and summer was established: wind speeds <6 m / s in spring are classified as light wind days, and wind speeds <5 m / s in summer are also classified as light wind days. In this embodiment, real-time NWP data and real-time wind speed data from multiple meteorological towers were collected from a certain offshore wind farm in Yangjiang during spring and summer. The calculated 10-minute average wind speed was 4.6 m / s, which meets the light wind day threshold for spring and summer, and was automatically identified as a light wind day, thus completing the wind condition mode classification.

[0061] Step 2: Calling up the special forecasting model for light wind days and predicting single-unit power Based on the light wind pattern recognition results from step 1, the SARIMA+density correction multi-source data fusion prediction model specifically designed for light winds in spring and summer is invoked. This prediction model uses the Seasonal Auto-Regressive Integrated Moving Average (SARIMA) model as the time-series prediction backbone, coupled with an air density physical correction module, and integrates multi-source monitoring information from multiple wind towers, turbulence, wind shear, sea surface temperature, and sea fog. It maintains high stability and prediction accuracy even in low wind speed and weak power signal scenarios. Input features: real-time wind speed / direction from multiple meteorological towers, turbulence intensity, vertical wind shear, sea surface temperature, real-time air density, sea fog concentration, and subtropical high pressure index; Model parameter settings: SARIMA model order (p=2, d=1, q=1) (P=1, D=0, Q=1), density correction module weight coefficient 0.35, optimizer used is Adam, initial learning rate is 0.001, training batch size is 64, physical constraints are introduced to prevent outliers; Model output: Initial predicted power sequence of a single wind turbine every 10 minutes for the next 0-24 hours, completing the power prediction of a single unit and solving the problem of large power fluctuations caused by low wind speed and turbulent wind direction on windy days.

[0062] Step 3: Dynamic wake correction and full-field power aggregation Combination Figure 2 The dynamic wake calculation and power aggregation process performs the following operations: Input for dynamic wake calculation: Input the single-unit power prediction sequence, real-time wind direction and speed, wind turbine coordinates and impeller parameters (impeller diameter D=220m), and environmental parameters (air density, turbulence intensity, sea fog concentration) output from step 2 into the dynamic wake calculation module; Secondary wind condition determination: Based on real-time wind speed, the light wind condition is reconfirmed, and the high-end wake model (improved Jensen model) is invoked, with the wake attenuation coefficient... The value is dynamically adjusted to 0.04 (range 0.03~0.05 on light windy days), and the other parameters are the same as in Example 1; Effective wind speed calculation: Based on the high-end wake model, the wake effects of all upstream wind turbines on downstream wind turbines in the wind farm are superimposed to calculate the effective wind speed at the location of each wind turbine. The total wake loss coefficient is calculated in the same way as in Example 1. Spatial power aggregation: The initial predicted power of a single unit is combined with the wake loss coefficient and air density for secondary physical constraint correction, and then the power of the whole field is superimposed through the spatial aggregation algorithm. The calculation formula is the same as in Example 1. Output results: The output shows the predicted total power of the wind farm for the next 0-24 hours. According to actual measurement, the MAPE (Mean Absolute Percentage Error) of this embodiment is 8.8% for light wind days, which meets the accuracy requirement of ≤9.2%.

[0063] Example 4 In another preferred embodiment, based on Embodiment 1 above, this embodiment provides a method for predicting offshore wind power based on seasonal adaptive wind condition identification. This method addresses the low-wind conditions caused by frequent northeast monsoons and sea fog in the Yangjiang sea area during autumn and winter, and combines... Figure 1 , Figure 2 The specific implementation process of this invention is described in detail below, with the specific steps as follows: Step 1: Automatic determination of seasonal wind condition mode Based on nearly 10 years of measured meteorological data and wind turbine SCADA operation data in the Yangjiang sea area during autumn and winter, a special wind condition identification standard for autumn and winter was established: Autumn: wind speed ≥11m / s is determined to be a strong wind day, and wind speed <7m / s is determined to be a light wind day; Winter: wind speed ≥12m / s is determined to be a strong wind day, and wind speed <8m / s is determined to be a light wind day.

[0064] In this embodiment, real-time NWP data and real-time wind speed data from multiple meteorological towers were collected from an offshore wind farm in Yangjiang during autumn. The calculated 10-minute average wind speed is 5.8 m / s, which meets the threshold for light wind days in autumn and winter. The wind condition is automatically identified as a light wind day, thus completing the wind condition mode classification.

[0065] Step 2: Calling up the special forecasting model for light wind days and predicting single-unit power Based on the light wind pattern recognition results from step 1, the autumn / winter light wind-dependent LSTM + sea fog-density coupled multi-source data fusion prediction model is invoked. This prediction model uses a Long Short-Term Memory (LSTM) network as the backbone network, superimposed with a sea fog-density coupled physical correction module, forming a multi-source data fusion-specific prediction model for light winds. By introducing marine meteorological and physical prior constraints on the basis of pure data-driven learning, the stability and accuracy of power prediction under low wind speed, high sea fog, and low temperature environments are significantly improved. Input features: real-time wind speed / direction from multiple meteorological towers, turbulence intensity, vertical wind shear, sea surface temperature, real-time air density, sea fog concentration, and subtropical high pressure index; Model parameter settings: LSTM hidden layer units 256, fog-density coupling module weight coefficient 0.4, optimizer used is AdamW (Adaptive Moment Estimation with Weight Decay), initial learning rate 0.0008, training batch size 64, early stopping mechanism is introduced to prevent overfitting; Model output: Initial predicted power sequence of a single wind turbine every 10 minutes for the next 0-24 hours, completing the power prediction of a single unit and solving the problem of large power fluctuations caused by low wind speed and sea fog interference on windy days.

[0066] Step 3: Dynamic wake correction and full-field power aggregation Combination Figure 2 The dynamic wake calculation and power aggregation process performs the following operations: Input for dynamic wake calculation: Input the single-unit power prediction sequence, real-time wind direction and speed, wind turbine coordinates and impeller parameters (impeller diameter D=220m), and environmental parameters (sea fog concentration, air density, turbulence intensity) output from step 2 into the dynamic wake calculation module; Secondary wind condition determination: Based on real-time wind speed, the light wind condition is reconfirmed, and the high-end wake model (improved Jensen model) is invoked, with the wake attenuation coefficient... The value is dynamically adjusted to 0.04 (range 0.03~0.05 on light windy days), and the other parameters are the same as in Example 1; Effective wind speed calculation: Based on the high-end wake model, the wake effects of all upstream wind turbines on downstream wind turbines in the wind farm are superimposed to calculate the effective wind speed at the location of each wind turbine. The total wake loss coefficient is calculated in the same way as in Example 1. Spatial power aggregation: The initial predicted power of a single unit is combined with the wake loss coefficient, sea fog concentration, and air density for secondary physical constraint correction, and then the power superposition of the entire field is completed by the spatial aggregation algorithm. The calculation formula is the same as in Example 1. Output results: The total power prediction value of the wind farm in the next 0-24 hours is output. After actual measurement and verification, the MAPE (mean absolute percentage error) of this embodiment in light wind days is 8.7%, which meets the accuracy requirement of ≤9.2%.

[0067] Example 5 In another preferred embodiment, based on Embodiment 1 above, this embodiment provides a method for predicting offshore wind power based on seasonal adaptive wind condition identification. This method is designed for normal wind conditions in the Yangjiang sea area during the seasonal transition period, where wind speeds fall between the thresholds for strong and light wind days. Figure 1 , Figure 2 The specific implementation process of this invention is described in detail below, with the specific steps as follows: Step 1: Automatic determination of seasonal wind condition mode Based on nearly 10 years of measured meteorological data and wind turbine SCADA operation data from the Yangjiang sea area, a seasonal wind condition identification standard was established: spring wind speed 6 m / s ≤ v < 10 m / s, summer wind speed 5 m / s ≤ v < 9 m / s, autumn wind speed 7 m / s ≤ v < 11 m / s, and winter wind speed 8 m / s ≤ v < 12 m / s, all of which are considered normal wind conditions. In this embodiment, real-time NWP data, LiDAR wind measurement data, and real-time wind speed data at a height of 100 m from a wind measurement tower were collected from an offshore wind farm in Yangjiang. The calculated 10-minute average wind speed was 8.2 m / s, meeting the normal wind condition criteria, and was automatically identified as a normal wind condition, thus completing the wind condition mode classification.

[0068] Step 2: Calling up the special prediction model for normal wind conditions and predicting the power of individual units Based on the normal wind condition pattern recognition results from step 1, the CNN-LSTM hybrid temporal prediction model specifically designed for normal wind conditions is invoked. This model is a data-driven + physically constrained hybrid prediction model adapted to the normal wind conditions in the Yangjiang sea area throughout the four seasons. It uses a convolutional neural network (CNN) and a long short-term memory network (LSTM) as its core architecture, integrates multi-source meteorological and wind turbine operation data, and balances prediction accuracy with the stable operation characteristics under normal wind conditions. The details are as follows: (1) Model architecture and module functions The model consists of a three-layer serially coupled structure: CNN Convolutional Neural Network Layers: The structure consists of 2 convolutional layers and 1 pooling layer. The kernel size is 3×3 and the number of channels is 64. The activation function is ReLU (Rectified Linear Unit), which is used to extract the local nonlinear correlation and spatial features of multi-source input features, providing high-quality feature input for subsequent time series prediction. LSTM Long Short-Term Memory Network Layer: A two-layer unidirectional LSTM structure with 128 hidden units is adopted. A Dropout layer (dropout rate of 0.2) is introduced to prevent overfitting. It is used to learn the dependence of wind speed and power in long time series under normal wind conditions, capture slow-changing patterns such as day-night cycle and sea-land wind alternation, and ensure the stability of long time series prediction. Physical constraint correction layer: Based on the physical laws of wind turbine operation and marine meteorological parameters, the pure data-driven output is corrected to ensure that the predicted power is strictly limited within the range of 0 to rated power of the wind turbine. Linear correction is completed in combination with real-time air density to ensure the physical consistency between pitch angle action and power change. At the same time, the prediction sequence is smoothed to suppress abnormal fluctuations caused by random noise.

[0069] (2) Input features NWP wind speed / direction, LiDAR wind measurement data, real-time turbine pitch angle, historical 7-day SCADA operation data, turbulence intensity, vertical wind shear, real-time air density, and theoretical power curve of the turbine.

[0070] (3) Model parameter settings The CNN convolution kernel size is 3×3 with 64 channels, the LSTM hidden layer has 128 units, the optimizer is Adam, the initial learning rate is 0.001, the cosine annealing learning rate decay strategy is adopted, the training batch size is 32, the loss function is root mean square error (RMSE) + physical constraint penalty term, and an early stopping mechanism is introduced to prevent overfitting.

[0071] (4) Model output The initial predicted power sequence of a single wind turbine for the next 0 to 48 hours is generated in 15-minute increments. The output is stable, without abnormal jumps, and conforms to physical constraints, thus completing the power prediction of a single turbine and providing basic data for subsequent wake correction.

[0072] Step 3: Dynamic wake correction and full-field power aggregation Combined with appendix Figure 2 The dynamic wake calculation and power aggregation process performs the following operations: Input for dynamic wake calculation: Input the single-unit power prediction sequence, real-time wind direction and speed, wind turbine coordinates and impeller parameters (impeller diameter D=220m), and environmental parameters (air density, turbulence intensity) output from step 2 into the dynamic wake calculation module; Secondary wind condition determination: Based on real-time wind speed, normal wind conditions are reconfirmed, and the standard wake model (Jensen wake model) is invoked. The calculation formula is as follows: (1); In the formula, For downstream fan speed, The incoming air velocity from the upstream fan. The axial induction factor is set to 0.25. This is the wake attenuation coefficient (0.055 under normal wind conditions, with a value range of 0.05~0.06). The distance between upstream and downstream wind turbines. This refers to the diameter of the fan impeller.

[0073] Effective wind speed calculation: Based on the standard wake model, the wake effects of all upstream wind turbines on downstream wind turbines within the wind farm are superimposed to calculate the effective wind speed at each turbine location. The formula for calculating the total wake loss coefficient is as follows: (3); In the formula, For the first The total wake loss factor of the typhoon turbine. For the first Typhoon machine to the first The weighting factor of typhoon generators For the first Typhoon machine to the first The wake effect coefficient of a single typhoon turbine.

[0074] Spatial power aggregation: The initial predicted power of a single unit is combined with the wake loss coefficient and air density for physical constraint correction, and then the power of the entire field is superimposed using a spatial aggregation algorithm. The calculation formula is as follows: (2); In the formula, This represents the total predicted power of the wind farm. For the first Typhoon generator power after correction For the first Total wake loss coefficient of typhoon generator.

[0075] Output results: The total power prediction value of the wind farm in the next 0-48 hours is output. According to actual measurement, the MAPE under normal wind conditions in this embodiment is 6.8%, which meets the accuracy requirement of ≤7.0%.

[0076] The five embodiments of this invention address typical windy conditions in the Yangjiang sea area, including strong winds in spring, light winds in autumn, strong winds in autumn and winter, light winds in spring and summer, and normal wind conditions. They comprehensively cover the core technical solutions of this invention: seasonal adaptive wind condition identification, wind-specific model prediction, and dynamic wake power aggregation. Accuracy verification results from multiple embodiments demonstrate that this invention can fully adapt to the complex wind conditions of the sea area throughout the four seasons, significantly improving the accuracy of wind power prediction under strong winds, light winds, and normal wind conditions, demonstrating significant engineering application value.

[0077] In the preferred embodiment, the seasonal adaptive wind condition identification standard in step 1 is to set differentiated wind speed thresholds according to the season. The above settings, by accurately classifying the wind speed characteristics of the four seasons, effectively distinguish the wind conditions of different seasons, provide accurate wind condition classification for the subsequent prediction model, avoid prediction deviations caused by wind condition misjudgment, ensure that the prediction model can be optimized for different wind conditions, and improve the overall prediction accuracy.

[0078] In the preferred embodiment, the wind condition pattern recognition input data in step 1 includes at least one of historical meteorological data of the sea area, NWP data, SCADA data, and real-time data from the wind measurement tower. The above settings, by fusing multi-source data, comprehensively capture the spatiotemporal variation characteristics of sea area wind conditions, improve the accuracy and real-time performance of wind condition recognition, provide a rich and reliable information foundation for the prediction model, and enhance the model's adaptability to complex wind conditions.

[0079] In the preferred embodiment, the windy weather prediction model in step 2 is a physics-data hybrid driven model. The input features include at least one of the following: wind turbine power limitation status flag, real-time pitch angle, theoretical power curve, wind speed, and wind direction. The above settings combine the advantages of physical mechanisms and data-driven approaches, make full use of wind turbine operating status and meteorological information, accurately characterize the power change features of windy weather, effectively address the prediction challenges under wind turbine power limitation conditions, and improve the accuracy of windy weather prediction.

[0080] In the preferred embodiment, the windy weather prediction model includes: using WT-CNN-LSTM + gust module in spring and summer, and using GNN-Transformer + strong convection module in autumn and winter; the above settings are customized for the wind characteristics of different seasons, the ability to capture gusts is enhanced in spring and summer, and the simulation accuracy of strong convection processes is improved in autumn and winter, so that the prediction model can better adapt to the complex and ever-changing windy weather environment and improve the reliability of prediction.

[0081] In the preferred embodiment, the light wind prediction model in step 2 is a multi-source data fusion model. The input features include at least one of real-time data from multiple meteorological towers, turbulence intensity, vertical wind shear, sea surface temperature, air density, and subtropical high index. By integrating multi-dimensional meteorological information, the above settings comprehensively reflect the turbulent characteristics of light wind conditions, effectively reduce prediction errors caused by wind direction fluctuations and unstable wind speeds, and improve the stability and accuracy of light wind prediction.

[0082] In the preferred embodiment, the windless weather prediction model includes: a SARIMA+density correction model in spring and summer, and an LSTM+sea fog-density coupling model in autumn and winter. The above settings adjust the model structure according to seasonal changes. In spring and summer, time series analysis is used to capture periodic changes, and in autumn and winter, the influence factor of sea fog is introduced to enhance the model's adaptability to complex environmental factors, further improve the accuracy of windless weather prediction, and meet the prediction needs of different seasons.

[0083] In the preferred embodiment, the normal wind condition model in step 2 is a CNN-LSTM hybrid time-series prediction model. The input features include at least one of the following: NWP wind speed / direction, LiDAR wind measurement data, real-time turbine pitch angle, SCADA operation data, turbulence intensity, vertical wind shear, real-time air density, and theoretical power curve of the turbine. By fusing multi-source heterogeneous data and using a deep learning model to capture time-series features, the temporal correlation and environmental adaptability of power prediction under normal wind conditions are effectively improved.

[0084] In the preferred embodiment, the dynamic wake model in step 3 adopts the Jensen model or its improved version to calculate the wake influence coefficients and wind speed attenuation of upstream and downstream wind turbines. The above settings, by quantifying the impact of wake effects on downstream wind turbines, accurately correct the power prediction value, effectively solve the prediction deviation problem caused by wake interference between wind turbines, improve the overall prediction accuracy of wind farms, and provide more reliable power prediction results for grid dispatch.

[0085] In the preferred embodiment, the wake attenuation coefficient is dynamically adjusted according to the wind condition pattern: 0.06~0.09 for strong winds, 0.03~0.05 for light winds, and 0.05~0.06 for normal winds. These settings, by differentiating the wake attenuation coefficient, enable the model to more accurately reflect the changing patterns of the wake effect under different wind conditions, further improving the accuracy and stability of power prediction and adapting to the varied wind conditions of the sea area throughout the four seasons.

[0086] In the preferred embodiment, the spatial aggregation algorithm in step 3 adopts the full-field power superposition formula; the above settings, by integrating the spatial distribution characteristics of each wind turbine location, achieve accurate aggregation of the overall power of the wind farm, avoiding the prediction deviation caused by spatial heterogeneity in traditional methods.

[0087] In the preferred embodiment, the corrected power of the single wind turbine The model outputs data from a specialized prediction model, and incorporates physical constraints and corrections based on real-time air density, turbulence intensity, and sea fog concentration. Under normal wind conditions, the prediction sequence is also smoothed to suppress abnormal fluctuations caused by random noise. These settings, through dynamic correction of physical parameters and noise filtering, enable the model to more accurately reflect the actual operating conditions of wind turbines, effectively overcome prediction biases caused by changes in environmental factors, enhance prediction robustness under complex weather conditions, ensure the smoothness and physical rationality of the output results, significantly improve the power prediction accuracy of a single wind turbine, and provide support for the refined operation of wind farms.

[0088] In the preferred embodiment, the first Typhoon total wake loss factor All upstream wind turbines within the same wind farm are connected to the first The wake effects of typhoon turbines are superimposed; the above settings, by comprehensively considering the wake effects of all upstream wind turbines within the wind farm, accurately quantify the [missing information]. The power loss caused by the wake effect of typhoon turbines provides a reliable basis for predicting the overall power of wind farms, optimizing wind farm layout and operation strategies, and improving wind farm power generation efficiency.

[0089] In the preferred embodiment, the offshore wind power prediction method based on seasonal adaptive wind condition identification further includes a model adaptive update step, which incrementally trains and iterates the parameters of the special prediction model quarterly using the latest measured data of the sea area. The above settings, by continuously updating the model parameters, ensure that the prediction model can adapt to long-term changes in sea area wind conditions, maintain prediction accuracy and stability, extend the effective service life of the model, reduce model maintenance costs, and improve the economic benefits of wind farms.

[0090] In the preferred embodiment, the offshore wind power prediction method based on seasonal adaptive wind condition identification outputs power prediction results for the next 0-72 hours, wherein the prediction deviation is ≤7.8% for strong wind days, MAPE is ≤9.2% for light wind days, and MAPE is ≤7.0% for normal wind conditions. The above settings, through refined modeling and multi-objective optimization for different wind conditions, achieve a balanced improvement in prediction accuracy across the entire wind condition range, meeting the grid dispatching requirements for predictions of different time scales and wind conditions.

[0091] In the preferred embodiment, the offshore wind power prediction method based on seasonal adaptive wind condition recognition can be directly connected to the wind farm SCADA system and dispatch master station system, realizing automatic data collection, automatic model invocation, and automatic power reporting, covering all operating conditions including strong wind days, light wind days, and normal wind conditions. The above settings significantly improve the automation level and practicality of the prediction method, reduce manual intervention and data transmission delays, improve the timeliness and accuracy of power prediction, provide strong support for the intelligent operation of wind farms, and promote the digital transformation of the wind power industry.

[0092] In summary, the offshore wind power prediction method based on seasonal adaptive wind condition identification provided by this invention effectively solves the problems of existing general prediction methods failing to adapt to the frequent alternation of wind conditions in the sea area throughout the four seasons. These methods result in distorted capture of the "power cliff" on windy days, large prediction errors on light wind days, and insufficient accuracy in predicting normal wind conditions. Furthermore, they do not fully consider the dominant seasonal meteorological factors and the dynamic influence of wind turbine wakes, thus failing to meet the actual operational needs of wind farms and the power grid. This method establishes a seasonal adaptive wind condition identification system to address the differentiated wind conditions of the four seasons, distinguishing between windy days, light wind days, and normal wind conditions and calling upon a dedicated prediction model. Combined with dynamic wake power aggregation technology, it achieves high-precision wind power prediction under all operating conditions, filling the gap in specialized prediction technology for complex wind conditions in the sea area throughout the four seasons.

[0093] This invention constructs a three-level collaborative architecture of "seasonally adaptive wind condition identification - specialized model prediction - dynamic wake aggregation". Differentiated wind condition thresholds are set according to the season, and dedicated prediction models are designed for strong winds, light winds, and normal wind conditions in spring and summer and autumn and winter. Wavelet transform, graph neural network, long short-term memory network, CNN-LSTM hybrid model and other technologies are deeply integrated with marine meteorological and physical laws to form a dedicated prediction scheme adapted to regional climate. It is different from the homogeneous design of existing general prediction methods and has distinct regional adaptability and all-condition targeting.

[0094] This invention leverages the unique meteorological characteristics of the sea area, including the northeast monsoon and sea fog in autumn and winter, and the sea-land breeze and subtropical high pressure in spring and summer, combining physical constraints with data-driven approaches. On windy days, it introduces specialized modules for gusts and strong convection to accurately capture "power cliffs." On windless days, it designs sea fog-density coupling and air density correction modules to address low-wind-speed fluctuations. Under normal wind conditions, it employs a CNN-LSTM hybrid model to balance prediction accuracy and stability under stable operating conditions. Simultaneously, it uses a dynamic wake model to correct wake interference between wind turbines in real time, dynamically adjusting the wake attenuation coefficient according to different wind conditions. This achieves accurate aggregation of power from single turbines to the entire wind farm, significantly improving prediction accuracy under complex wind conditions. Not only is the prediction stability stronger, but it can also directly interface with existing SCADA and NWP systems in wind farms, demonstrating significant engineering application value.

Claims

1. A method for predicting offshore wind power based on seasonal adaptive wind condition identification, characterized in that, Includes the following steps: Step 1: Establish seasonal adaptive wind condition identification standards based on the four seasons climate characteristics of the sea area, and complete the automatic identification of wind condition patterns, including strong wind day mode, light wind day mode and normal wind condition mode. Step 2: Based on the identification results, call the corresponding wind condition-specific prediction model to obtain the initial predicted power of a single wind turbine. The specific prediction models include the strong wind day prediction model, the light wind day prediction model, and the normal wind condition model. Step 3: Introduce the dynamic wake model and spatial aggregation algorithm to complete the full-field power correction and output.

2. The offshore wind power prediction method based on seasonal adaptive wind condition identification according to claim 1, characterized in that, The seasonal adaptive wind condition identification standard mentioned in step 1 is to set differentiated wind speed thresholds according to the season, specifically: Spring: The threshold for strong winds is ≥10m / s, the threshold for light winds is <6m / s, and the threshold for normal wind conditions is 6m / s ≤ wind speed <10m / s; Summer: Threshold for strong winds ≥ 9 m / s, threshold for light winds < 5 m / s, and for normal wind conditions 5 m / s ≤ wind speed < 9 m / s; Autumn: The threshold for strong winds is ≥11m / s, the threshold for light winds is <7m / s, and the threshold for normal wind conditions is 7m / s ≤ wind speed <11m / s; Winter: Threshold for strong winds ≥ 12 m / s, threshold for light winds < 8 m / s, and for normal wind conditions 8 m / s ≤ wind speed < 12 m / s.

3. The offshore wind power prediction method based on seasonal adaptive wind condition identification according to claim 1, characterized in that: The wind pattern recognition input data in step 1 includes at least one of the following: historical meteorological data of the sea area, NWP data, SCADA data, and real-time data from the wind measurement tower.

4. The offshore wind power prediction method based on seasonal adaptive wind condition identification according to claim 1, characterized in that: The windy weather prediction model described in step 2 is a physics-data hybrid driven model. The input features include at least one of the following: wind turbine power limit status flag, real-time pitch angle, theoretical power curve, wind speed, and wind direction.

5. The offshore wind power prediction method based on seasonal adaptive wind condition identification according to claim 4, characterized in that, The windy weather prediction model includes: WT-CNN-LSTM + gust module in spring and summer, and GNN-Transformer + strong convection module in autumn and winter.

6. The offshore wind power prediction method based on seasonal adaptive wind condition identification according to claim 1, characterized in that: The windless weather prediction model described in step 2 is a multi-source data fusion model. The input features include at least one of the following: real-time data from multiple meteorological towers, turbulence intensity, vertical wind shear, sea surface temperature, air density, and subtropical high index.

7. The offshore wind power prediction method based on seasonal adaptive wind condition identification according to claim 6, characterized in that, The windy weather prediction model includes: a SARIMA+density correction model for spring and summer, and an LSTM+sea fog-density coupling model for autumn and winter.

8. The offshore wind power prediction method based on seasonal adaptive wind condition identification according to claim 1, characterized in that: The normal wind condition model mentioned in step 2 is a CNN-LSTM hybrid time series prediction model. The input features include at least one of the following: NWP wind speed / direction, LiDAR wind measurement data, real-time turbine pitch angle, SCADA operation data, turbulence intensity, vertical wind shear, real-time air density, and theoretical power curve of the turbine.

9. The offshore wind power prediction method based on seasonal adaptive wind condition identification according to claim 1, characterized in that, The dynamic wake model described in step 3 uses the Jensen model or its improved version to calculate the wake influence coefficients and wind speed attenuation of the upstream and downstream wind turbines. The calculation formula is as follows: （1）； In the formula, For downstream fan speed, The incoming air velocity from the upstream fan. It is an axial inducing factor. The wake attenuation coefficient is... The distance between upstream and downstream wind turbines. This refers to the diameter of the fan impeller.

10. The offshore wind power prediction method based on seasonal adaptive wind condition identification according to claim 9, characterized in that: The wake attenuation coefficient The value is dynamically adjusted according to the wind pattern: 0.06~0.09 for strong wind mode, 0.03~0.05 for light wind mode, and 0.05~0.06 for normal wind mode.

11. The offshore wind power prediction method based on seasonal adaptive wind condition identification according to claim 1, characterized in that, The spatial aggregation algorithm described in step 3 uses the full-field power superposition formula: （2）； In the formula, This represents the total predicted power of the wind farm. For the first Typhoon generator power after correction For the first Total wake loss coefficient of typhoon generator.

12. The offshore wind power prediction method based on seasonal adaptive wind condition identification according to claim 11, characterized in that: The corrected power of the single wind turbine The prediction sequence is output by a specialized prediction model and is physically constrained and corrected by real-time air density, turbulence intensity, and sea fog concentration. Under normal wind conditions, the prediction sequence also needs to be smoothed to suppress abnormal fluctuations caused by random noise.

13. The offshore wind power prediction method based on seasonal adaptive wind condition identification according to claim 11, characterized in that, The first Typhoon total wake loss factor All upstream wind turbines within the same wind farm are connected to the first The effect of the typhoon's wake is superimposed, and the calculation formula is: （3）； In the formula, For the first Typhoon machine to the first The weighting factor of typhoon generators For the first Typhoon machine to the first The wake effect coefficient of a single typhoon turbine.

14. The offshore wind power prediction method based on seasonal adaptive wind condition identification according to claim 1, characterized in that: The offshore wind power prediction method based on seasonal adaptive wind condition identification also includes a model adaptive update step, which uses the latest measured data of the sea area quarterly to incrementally train and iterate the parameters of the special prediction model.

15. The offshore wind power prediction method based on seasonal adaptive wind condition identification according to claim 1, characterized in that: The offshore wind power prediction method based on seasonal adaptive wind condition identification outputs power prediction results for the next 0 to 72 hours, where the prediction deviation is ≤7.8% for strong wind days, MAPE is ≤9.2% for light wind days, and MAPE is ≤7.0% for normal wind conditions.

16. The offshore wind power prediction method based on seasonal adaptive wind condition identification according to claim 1, characterized in that: The offshore wind power prediction method based on seasonal adaptive wind condition identification can be directly connected to the wind farm SCADA system and dispatch master station system to realize automatic data collection, automatic model calling and automatic power reporting, covering all operating conditions including strong wind days, light wind days and normal wind conditions.

Images