A method and device for offshore wind farm power prediction

Abstract

The application provides a kind of offshore wind farm power prediction method and device, it is related to wind power prediction technical field, comprising: obtaining the typhoon path data and wind wave field data of target offshore wind farm;Based on the wind wave field enhancement feature with spatial distance attenuation weight is constructed based on station center coordinates and external wind wave field point coordinates;Based on the prediction error difference between the reference prediction model and the enhanced prediction model, determine the typhoon influence index of a plurality of candidate typhoons, select the target typhoon corresponding to the maximum index;The feature set containing wind wave field enhancement feature is input into wind speed prediction model to obtain predicted wind speed sequence;Determine the power prediction result based on wind power curve model and predicted wind speed sequence.The application improves the accuracy and stability of offshore wind farm power prediction under typhoon working condition.

Description

Technical Field

[0001] This application relates to the field of wind power prediction technology, specifically to a method and apparatus for predicting the power of offshore wind farms. Background Technology

[0002] In offshore wind power forecasting, numerical weather prediction data and historical operating data from wind farms are typically used to build predictive models. However, under typhoon influence, offshore wind farms face complex meteorological conditions such as drastic changes in wind speed and direction, and coupled wind and wave disturbances. Existing technologies often directly extract meteorological grid data from the nearest meteorological grid point as input features, failing to fully quantify the spatial distance differences between external wide-area wind and wave field points and the target wind farm, resulting in insufficient spatial representativeness of meteorological features. Furthermore, when multiple typhoon systems coexist in the sea area, existing technologies lack a unified screening index for target typhoons, easily introducing characteristic noise from weakly correlated typhoons into the model, leading to inaccurate power prediction results. Therefore, there is an urgent need for an offshore wind farm power forecasting method that can accurately characterize the spatial impact of typhoon wind and wave fields and effectively screen dominant typhoons. Summary of the Invention

[0003] To address the problems of insufficient spatial representativeness of external wind and wave fields under the influence of typhoons and inaccurate predictions caused by interference from multiple typhoons in existing technologies, this application proposes a method and device for predicting the power of offshore wind farms. This method and device achieves accurate aggregation of typhoon wind and wave field characteristics and adaptive screening of target typhoons, thereby improving the accuracy and stability of power prediction.

[0004] To achieve the above objectives, this application adopts the following technical solution: A method for predicting the power of an offshore wind farm includes: acquiring typhoon path data and wave field data of a target offshore wind farm; constructing a wave field enhancement feature with spatial distance attenuation weights based on the center coordinates of the target offshore wind farm and the coordinates of external wave field points in the wave field data; determining multiple candidate typhoons based on the typhoon path data, and determining the typhoon impact index of the multiple candidate typhoons based on the prediction error difference between the baseline prediction model and the enhanced prediction model, and selecting the candidate typhoon with the largest typhoon impact index as the target typhoon; inputting the feature set containing the wave field enhancement feature into a pre-trained wind speed prediction model to obtain a predicted wind speed sequence; and determining the power prediction result of the target offshore wind farm based on a pre-fitted wind power curve model and the predicted wind speed sequence.

[0005] As one implementation method, constructing the wind and wave field enhancement feature with spatial distance attenuation weight includes: determining the spherical distance between the external wind and wave field location and the target offshore wind farm based on the coordinates of the external wind and wave field location and the coordinates of the station center; determining the spatial distance attenuation weight based on the spherical distance and a monotonically decreasing function; and weighting and aggregating the wind and wave field features corresponding to the external wind and wave field location based on the spatial distance attenuation weight to obtain the wind and wave field enhancement feature.

[0006] As one implementation method, determining the spherical distance between the external wind and wave field location and the target offshore wind farm based on the coordinates of the external wind and wave field location and the coordinates of the station center includes: calculating the spherical distance using a spherical distance formula based on the latitude and longitude of the external wind and wave field location coordinates and the latitude and longitude of the station center coordinates; and weighting and aggregating the wind and wave field features corresponding to the external wind and wave field location based on the spatial distance attenuation weight to obtain the wind and wave field enhancement features includes: weighting and summing at least one of the wind speed, air pressure, and wave height corresponding to each external wind and wave field location with the corresponding spatial distance attenuation weight, and dividing by the sum of all the spatial distance attenuation weights to obtain the wind and wave field enhancement features.

[0007] As one implementation method, determining the typhoon impact index of multiple candidate typhoons based on the prediction error difference between the baseline prediction model and the enhanced prediction model includes: the baseline prediction model is a prediction model that does not include the enhanced wind and wave field features, and the enhanced prediction model is a prediction model that includes the enhanced wind and wave field features; the prediction error difference is calculated as follows: within a historical typhoon impact window, the difference is obtained by subtracting the second prediction error of the enhanced prediction model from the first prediction error of the baseline prediction model; the typhoon path data and the wind and wave field data of the multiple candidate typhoons are input into a pre-trained typhoon impact index estimation model, and the typhoon impact index corresponding to each of the multiple candidate typhoons is output; wherein, the typhoon impact index estimation model is trained based on the prediction error difference.

[0008] As one implementation method, the calculation of the prediction error difference further includes: using at least one of mean absolute error, root mean square error, or prediction deviation as an error evaluation index; when training the typhoon impact index estimation model, the typhoon impact index is the ratio of the difference between the first prediction error and the second prediction error to the first prediction error; when the ratio is less than zero, the typhoon impact index is determined to be zero; selecting the candidate typhoon with the largest typhoon impact index as the target typhoon includes: determining the candidate typhoon with the largest typhoon impact index among the multiple candidate typhoons as the target typhoon, and incorporating the typhoon path data and the wind and wave field enhancement features corresponding to the target typhoon into the feature set.

[0009] As one implementation method, the step of inputting a feature set containing the enhanced wind and wave field features into a pre-trained wind speed prediction model to obtain a predicted wind speed sequence includes: the feature set further includes conventional meteorological features, basic features of the wind farm, and historical operational features; inputting the feature set into the pre-trained wind speed prediction model to output the predicted wind speed sequence of the target offshore wind farm during the target prediction period; and determining the power prediction result of the target offshore wind farm based on a pre-fitted wind power curve model and the predicted wind speed sequence includes: inputting the predicted wind speed at each moment in the predicted wind speed sequence into the pre-fitted wind power curve model to map and obtain an initial power prediction value; and post-processing the initial power prediction value to obtain the power prediction result.

[0010] As one implementation method, the training label of the pre-trained wind speed prediction model is the hub height equivalent average wind speed; the hub height equivalent average wind speed is determined by: weighting the measured wind speed, rated power and availability of each wind turbine in the target offshore wind farm to obtain the hub height equivalent average wind speed; or when there are differences in the hub height of different wind turbines, the corresponding wind speed is converted to a unified reference hub height based on the wind shear index; the pre-fitted wind power curve model is a parametric model, which includes at least one of a piecewise model, a polynomial model or a sigmoid model.

[0011] As one implementation, the post-processing of the initial power prediction value to obtain the power prediction result includes: based on the cut-in wind speed constraint, when the predicted wind speed at the corresponding time is lower than the cut-in wind speed, setting the initial power prediction value at the corresponding time to zero or a preset low power value; based on the cut-out wind speed constraint, when the predicted wind speed at the corresponding time is higher than the cut-out wind speed, setting the initial power prediction value at the corresponding time to zero, a limited power value, or a safe power value; based on the upper and lower limits of installed capacity constraints, limiting the initial power prediction value between the lower and upper limits of the prediction; and performing time series smoothing processing on the power prediction value after the above constraint processing using at least one of moving average, convolutional smoothing, exponential smoothing, or median filtering to obtain the power prediction result.

[0012] As one implementation method, after acquiring the typhoon path data and wave field data of the target offshore wind farm, the method further includes: acquiring the station operation data and weather forecast data of the target offshore wind farm; unifying the time of different data sources to the time zone where the target offshore wind farm is located, and sorting and deduplicating by timestamp; interpolating or removing missing values, outliers, negative power, negative wind speed, and wind direction values ​​that are outside the reasonable range; converting the wind direction angle into sine and cosine components; calculating the equivalent average wind speed at the hub height at the station scale based on the station operation data; removing data from power curtailment or shutdown periods that are not suitable for power curve fitting based on the power curtailment value, shutdown status, or outlier identification results of the power curve; aligning the weather forecast data, the station operation data, the typhoon path data, and the wave field data to a unified time granularity to generate training samples.

[0013] Furthermore, this application also provides an offshore wind farm power prediction device, comprising: a data acquisition module for acquiring typhoon path data and wave field data of a target offshore wind farm; a wave field enhancement feature construction module for constructing wave field enhancement features with spatial distance attenuation weights based on the center coordinates of the target offshore wind farm and the coordinates of external wave field points in the wave field data; a typhoon impact index determination module for determining multiple candidate typhoons based on the typhoon path data, and determining the typhoon impact index of the multiple candidate typhoons based on the prediction error difference between the baseline prediction model and the enhanced prediction model, and selecting the candidate typhoon with the largest typhoon impact index as the target typhoon; a wind speed prediction module for inputting the feature set containing the wave field enhancement features into a pre-trained wind speed prediction model to obtain a predicted wind speed sequence; and a power prediction module for determining the power prediction result of the target offshore wind farm based on a pre-fitted wind power curve model and the predicted wind speed sequence.

[0014] Beneficial effects: The offshore wind farm power prediction method and apparatus provided in this application construct wind and wave field enhancement features with spatial distance attenuation weights based on the coordinates of the wind farm center and the coordinates of external wind and wave field points. By utilizing the distance attenuation mechanism in physical space to weight and aggregate widely distributed wind and wave field data, the input features can accurately reflect the actual impact intensity of external meteorological disturbances on the target wind farm, overcoming the spatial representativeness deficiency caused by traditional methods that rely solely on the data of the single nearest grid point. At the same time, the typhoon impact index is determined by comparing the prediction error difference between the baseline prediction model and the enhanced prediction model within the historical typhoon impact window, and the target typhoon is selected accordingly. This mechanism can quantify the marginal contribution of each candidate typhoon to the power prediction accuracy from a data-driven perspective, effectively eliminating redundant feature noise in the scenario of multiple typhoons coexisting, thereby significantly improving the accuracy and robustness of offshore wind farm power prediction under typhoon conditions.

[0015] This application determines the spatial distance attenuation weight by using spherical distance combined with a monotonically decreasing function, and performs normalized weighted aggregation on multi-dimensional features such as wind speed, air pressure, and wave height. This ensures the physical consistency and numerical stability of external point features from different geographical locations during the conversion to the field station scale, and avoids feature distortion caused by coordinate system projection deformation or simple linear superposition.

[0016] This application establishes a typhoon impact index with clear physical meaning by defining the typhoon impact index as the ratio of the error difference between the baseline model and the enhanced model to the error of the baseline model, and setting it to zero when the ratio is less than zero. This index not only reflects the positive gain of typhoon wind and wave field characteristics on prediction accuracy, but also automatically filters out negatively correlated or invalid typhoon processes through non-negative constraints. Combined with a pre-trained typhoon impact index estimation model, it enables rapid and automated screening of target typhoons in the future prediction period, ensuring the real-time operating efficiency of the prediction system.

[0017] This application employs a cascaded prediction architecture consisting of a wind speed prediction model, a wind power curve model, and multiple constraint post-processing. This architecture retains the data-driven model's ability to fit complex nonlinear relationships while ensuring that the prediction results conform to the actual operating characteristics of wind turbines through power curve mapping and physical constraints. It effectively suppresses abnormal jumps in model output that may occur under extreme weather conditions, thereby improving the reliability and usability of the prediction results in power grid dispatch applications. Attached Figure Description

[0018] Figure 1 This is a flowchart of the offshore wind farm power prediction method according to an embodiment of this application; Figure 2 This is a scatter plot of the equivalent average wind speed at the hub height and the total power output of the station according to an embodiment of this application. Figure 3 This is a schematic diagram of wind power curve fitting according to an embodiment of this application. Detailed Implementation

[0019] To make the objectives, technical solutions, and advantages of this application clearer, the technical solutions in the embodiments of this application will be clearly and completely described below with reference to the accompanying drawings. Obviously, the described embodiments are only some, not all, of the embodiments of this application.

[0020] Unless otherwise defined, all technical and scientific terms used herein have the same meaning as commonly understood by one of ordinary skill in the art to which this application pertains.

[0021] Example 1 like Figure 1 As shown, this embodiment provides a method for predicting the power of offshore wind farms. This method is applied to operational scenarios where short-term or medium-term power prediction of a target offshore wind farm is required. The offshore wind farm power prediction method includes the following steps: Step S1 involves acquiring typhoon track data and wave field data for the target offshore wind farm. Specifically, typhoon track data typically originates from real-time reports or numerical weather prediction products issued by meteorological departments. This data includes parameters characterizing the macroscopic motion of the typhoon, such as typhoon number, center latitude and longitude, central pressure, maximum wind speed near the center, direction of movement, and speed of movement. Wave field data describes the sea surface conditions within the typhoon-affected area and is usually provided in a gridded format, containing elements such as sea surface wind speed, sea surface wind direction, sea level pressure, significant wave height, wave direction, and wave period at various external wave field points. Acquiring these two types of data provides meteorological disturbance information covering a wide spatial area for subsequent analysis, compensating for the limited sensing range of a single observation point during typhoon weather.

[0022] Step S2 involves constructing wind and wave field enhancement features with spatial distance attenuation weights based on the center coordinates of the target offshore wind farm and the coordinates of external wind and wave field points in the wind and wave field data. In this step, the system first calculates the spatial distance between the center coordinates of the wind farm and the coordinates of each external wind and wave field point, and then determines the corresponding attenuation weight based on this distance. Generally, external points closer to the wind farm have a more significant impact on the local wind conditions and are therefore assigned a larger weight; conversely, points farther away have a smaller weight. Subsequently, these attenuation weights are used to weighted aggregate the wind and wave field features of each external point, thereby transforming the discrete wide-area grid data into a single or combined feature vector that can characterize the overall disturbance level of the wind farm. This processing method effectively establishes the physical connection between the external environmental field and the target wind farm, enabling the input features to accurately reflect the actual intensity of the typhoon wind and wave field's effect on the target offshore wind farm, and avoiding feature noise interference caused by directly splicing distant grid data.

[0023] Step S3: Based on typhoon path data, multiple candidate typhoons are identified. Based on the difference in prediction errors between the baseline prediction model and the enhanced prediction model, the typhoon impact index of each candidate typhoon is determined. The candidate typhoon with the largest typhoon impact index is selected as the target typhoon. In actual sea areas, multiple typhoon systems may exist simultaneously, but not all typhoons will have a substantial impact on a specific wind farm. This embodiment introduces the typhoon impact index as a quantitative indicator to identify the dominant typhoon. The baseline prediction model refers to a model that uses only conventional meteorological data and historical operational data for prediction, while the enhanced prediction model further incorporates enhanced wind and wave field characteristics. The difference in prediction errors between the two models within the historical typhoon impact window objectively reflects the marginal contribution of wind and wave field characteristics to improving prediction accuracy. During the prediction operation phase, based on the evaluation logic established by this error difference, the system directly calculates or infers the typhoon impact index of each candidate typhoon and automatically selects the one with the largest index as the target typhoon. This mechanism is equivalent to an adaptive feature denoiser, which can filter out the truly dominant disturbance sources from the complex multi-typhoon environment, prevent redundant information from weakly correlated typhoons from diluting key features, and thus ensure the focus and accuracy of the prediction model.

[0024] Step S4 involves inputting the feature set containing enhanced wind and wave field features into the pre-trained wind speed prediction model to obtain the predicted wind speed sequence. This feature set, in addition to the enhanced wind and wave field features generated in the previous steps, can also include multi-dimensional information such as conventional numerical weather prediction features, basic station attribute features, and historical operational statistics. The wind speed prediction model is a model that has undergone parameter learning and solidification using a large number of historical samples in the offline phase. In the online prediction phase, the system only needs to input the real-time constructed feature set into the model to quickly output the predicted wind speed sequence at the hub height of the target offshore wind farm for a future period. Using a pre-trained model for inference not only meets the stringent requirements of power dispatch for prediction timeliness but also ensures the stability of the technical solution in product applications.

[0025] Step S5: Based on the pre-fitted wind power curve model and predicted wind speed sequence, determine the power prediction results for the target offshore wind farm. The wind power curve model characterizes the nonlinear mapping relationship between wind speed and wind turbine output, and is also obtained by fitting historical measured data in the offline stage.

[0026] like Figure 2As shown in the scatter plot of the equivalent average wind speed at hub height versus total power output of the wind turbine, the horizontal axis represents wind speed and the vertical axis represents power. The data points generally show a trend of power increasing with wind speed and then saturating after reaching the rated wind speed. However, the original scatter plot often contains various abnormal operating condition samples, which need to be identified and removed before fitting the power curve. Specifically, three typical abnormal regions are marked with red circles in the figure: First, sparse outliers are discrete points that deviate from the main trend band, usually caused by transient sensor disturbances or abnormal data transmission; second, wind curtailment reflects active power reduction operation caused by grid dispatch instructions; third, shutdown indicates power close to zero, meaning the wind turbine automatically shuts down due to wind speed exceeding the cut-off wind speed. These three types of abnormal samples correspond to data quality issues, operational strategy interventions, and equipment protection states, respectively. Including them in the power curve fitting will cause the fitting results to deviate significantly from the actual aerodynamic conversion characteristics of the wind turbine under normal operating conditions. Therefore, before fitting, this application will retain only the data under normal operating conditions for parameter identification based on power rationing records, shutdown logs, and statistical outlier detection results.

[0027] like Figure 3 As shown, the wind power curve obtained after removing outliers is plotted, with wind speed on the x-axis and power on the y-axis. The blue scatter points represent actual observed data under normal operating conditions, exhibiting an overall S-shaped distribution; the orange curve represents the fitted result, with a smooth shape that runs through the dense data area. This curve begins to rise at approximately 3 m / s, corresponding to the wind turbine's cut-in wind speed; it tends to saturate in the 12-15 m / s range, corresponding to a rated power of approximately 390 MW. This S-shaped curve accurately depicts the climb from the cut-in wind speed to the rated wind speed and the saturation phase after reaching the rated power, better reflecting the actual aerodynamic conversion characteristics of wind turbines compared to piecewise linear models.

[0028] After obtaining the predicted wind speed sequence, the system inputs the predicted wind speed at each moment into the model, converts it into the corresponding power value, and then summarizes the power prediction results for the entire wind farm. This step-by-step cascaded architecture of "wind speed prediction + power mapping" retains clear physical interpretability compared to the black-box mode of directly predicting power, making it easier for operation and maintenance personnel to understand the causes of the prediction results, and also facilitating subsequent rationality verification and correction in conjunction with wind turbine operating parameters.

[0029] By executing steps S1 to S5 above, this embodiment constructs a complete offshore wind farm power prediction process. This process strictly distinguishes the logical boundaries between feature construction, typhoon screening, and model inference during operation. It fully utilizes the spatiotemporal information in typhoon path and wave field data, and ensures the real-time response capability of the prediction service through a pre-fixed model. This effectively solves the problem of prediction inaccuracy caused by insufficient spatial representativeness of meteorological features and interference from multiple typhoons under typhoon conditions.

[0030] Example 2 This embodiment further refines the specific implementation of step S2 based on embodiment 1. In some embodiments, the process of constructing wind and wave field enhancement features with spatial distance attenuation weights specifically includes determining the spherical distance between the external wind and wave field points and the target offshore wind farm based on the coordinates of the external wind and wave field points and the center coordinates of the wind farm. Specifically, since offshore wind farms are usually located in open sea areas, their coverage and typhoon-affected areas are large, and using a spherical distance that considers the curvature of the Earth can more accurately reflect the actual propagation path of meteorological elements in geographic space than a planar Euclidean distance. Specifically, the Haversine formula is used to calculate this spherical distance, and the calculation formula is as follows: in, This represents the spherical distance between the external wind and wave field location and the target offshore wind farm. This represents the Earth's average radius. (Intermediate variable) The calculation method is as follows In the formula and These represent the latitude and longitude of the center point of the target offshore wind farm, respectively. and These represent the latitude and longitude of the external wind and wave field location, respectively. This is a two-parameter arctangent function. Through the above calculations, discrete grid point coordinates can be transformed into spatial distance scalars with clear physical meaning, providing a precise geometric reference for subsequent feature weighting.

[0031] After determining the spatial distance, the spatial distance decay weight is determined based on the spherical distance and a monotonically decreasing function. As a preferred implementation, this monotonically decreasing function can take the form of exponential decay, i.e. .in, Indicates distance as Spatial distance decay weight over time This is the distance attenuation scale parameter. The effective radius of influence or attenuation rate of the external wind and wave field on the target station, when When the value is large, distant points still retain a higher weight, which is suitable for scenarios affected by large-scale weather systems; when When the values ​​are small, the weights decrease rapidly with distance, and the model focuses more on local disturbances around the station. In practical applications, The calibration can be performed based on the statistical analysis results of historical typhoon data, for example, by taking a specific value between 50 km and 200 km, in order to balance the information contribution of the wide-area background field and the local fine structure.

[0032] Subsequently, based on the spatial distance attenuation weight, the wind and wave field features corresponding to the external wind and wave field locations are weighted and aggregated to obtain the wind and wave field enhancement features. Specifically, at least one feature of wind speed, air pressure, and wave height corresponding to each external wind and wave field location is weighted and summed with the corresponding spatial distance attenuation weight, and then divided by the sum of all spatial distance attenuation weights to obtain the wind and wave field enhancement features. This process can be expressed by the following formula: in, express Enhancement characteristics of wind and wave fields at the scale of a target offshore wind farm. This represents the total number of external wind and wave field points involved in the calculation. Indicates the first Spatial distance attenuation weights corresponding to each external wind and wave field location. Indicates the first External wind and wave field locations The original characteristic values ​​such as wind speed, air pressure, or wave height at any given time are used. The numerator of the formula implements distance-based weighted superposition of features, reflecting the physical law that features are larger when closer and smaller when farther away; the denominator normalizes all weights, eliminating the problem of unstable feature amplitude caused by changes in the number of selected points or fluctuations in absolute weight values, thus ensuring the robustness of the generated wind and wave field enhancement features in terms of numerical distribution.

[0033] It should be understood that although the above embodiments use Haversine spherical distance and exponential decay function as examples for detailed description, this is not a limitation on the scope of protection of this application. In other feasible implementations, spherical distance can also be replaced by planar Euclidean distance based on map projection or more precise geodesic distance, as long as it can characterize the spatial interval between two points. Similarly, the monotonically decreasing function is not limited to the exponential form, but can also be in the form of Gaussian function, inverse distance function, or power-law decay function, etc., to adapt to the spatial correlation structure of meteorological elements in different sea areas. In addition, in addition to weighted average, the weighted aggregation method can also select robust statistics such as weighted median, weighted mode, or truncated weighted mean according to the data distribution characteristics to reduce the interference of extreme outliers on the aggregation results. These alternative solutions are all within the scope of the technical concept of this application, aiming to achieve the optimal characterization of the spatial impact of typhoon wind and wave fields through flexible distance measurement and aggregation strategies.

[0034] Example 3 This embodiment, based on Embodiment 1, further refines the specific implementation of step S3, focusing on the typhoon screening mechanism based on error improvement. The core of this mechanism lies in quantifying the actual contribution of each candidate typhoon to the power prediction accuracy, and automatically selecting the dominant typhoon during the prediction operation phase. Specifically, based on the difference in prediction errors between the baseline prediction model and the enhanced prediction model, the typhoon impact index of multiple candidate typhoons is determined. The baseline prediction model is a prediction model that does not include enhanced wind and wave field features; its input only includes information such as conventional weather forecasts, historical power output of power stations, and basic attributes. The enhanced prediction model, on the other hand, is a prediction model that includes enhanced wind and wave field features, that is, it further integrates the enhanced wind and wave field features constructed in Embodiment 2 on the basis of the baseline model's input. This design of input difference allows the difference in the output results to be directly attributed to the introduction of wind and wave field information, thus providing a controllable benchmark for assessing typhoon impact.

[0035] The specific calculation method for determining the difference in prediction errors is as follows: within the historical typhoon impact window, the difference is obtained by subtracting the second prediction error of the enhanced prediction model from the first prediction error of the baseline prediction model. This difference reflects the absolute improvement in prediction accuracy after introducing wind and wave field characteristics. To eliminate the dimensional influence caused by the difference in prediction difficulty of different typhoon processes, this application uses relative improvement as an evaluation index. Specifically, when training the typhoon impact index estimation model, the typhoon impact index is the ratio of the difference between the first and second prediction errors to the first prediction error. This ratio characterizes the marginal contribution rate of wind and wave field characteristics to reducing prediction errors. It is particularly important to note that when this ratio is less than zero, the typhoon impact index is set to zero. This treatment has a clear physical meaning and technical effect: if the ratio is negative, it indicates that introducing the wind and wave field characteristics of the typhoon actually leads to an increase in prediction error, that is, the typhoon characteristics are negatively correlated with the current power station output change or have introduced invalid noise. By setting the values ​​to zero, the system can automatically filter out typhoon processes with negative or weak correlations, preventing them from misleading subsequent predictions. This is equivalent to building an adaptive noise reduction filter at the feature level.

[0036] Regarding the selection of error evaluation indicators, this application supports multiple measurement methods to adapt to different business assessment needs. For example, mean absolute error (MAE) can be used to reflect the average level of prediction deviation, root mean square error (RMSE) can be used to penalize larger extreme deviations, or the predicted deviation in electricity volume can be used to directly align with grid dispatch assessment standards. In practical applications, at least one of the above indicators can be flexibly selected as the calculation benchmark based on the specific grid connection agreement or historical data distribution characteristics of the offshore wind farm to ensure that the construction of the typhoon impact index is consistent with the evaluation system of the actual application scenario.

[0037] It is important to emphasize that although the construction logic of the aforementioned typhoon impact index is based on statistical analysis of historical data, the actual power prediction operation does not use two models—a real-time benchmark and an enhanced model—to calculate the error. To meet real-time requirements and reflect the execution attributes of the technical solution, this application employs model inference to achieve rapid screening. Specifically, typhoon path data and wave field data of multiple candidate typhoons are input into a pre-trained typhoon impact index estimation model, which outputs the typhoon impact index corresponding to each candidate typhoon. This typhoon impact index estimation model is a lightweight regression model, such as a gradient boosting tree or neural network, trained based on the aforementioned prediction error differences. It has internalized the complex error comparison logic into a direct mapping relationship from meteorological features to the impact index. At the prediction time, the system only needs to input the real-time acquired candidate typhoon features into this model to obtain the impact index estimate of each typhoon in milliseconds, avoiding the computational overhead and latency caused by dual model inference and ensuring efficient response of the prediction service.

[0038] After obtaining the typhoon impact index (TPI) of each candidate typhoon, the candidate typhoon with the highest TPI is selected as the target typhoon. This step achieves a focus from a multi-typhoon coexistence environment to a single dominant disturbance source. Subsequently, the candidate typhoon with the highest TPI among the multiple candidate typhoons is determined as the target typhoon, and the typhoon track data and wind and wave field enhancement features corresponding to the target typhoon are included in the feature set. This means that only typhoon information deemed to have the greatest positive contribution to prediction accuracy will enter the subsequent wind speed prediction model, while other typhoon features with low or zero indices are eliminated. This data-driven gain-based screening mechanism, compared to traditional geographical distance screening or manual experience judgment, can more objectively and dynamically adapt to the complex and ever-changing marine meteorological environment, fundamentally ensuring the effectiveness and purity of the input features, and laying a solid foundation for improving the accuracy of power prediction.

[0039] Example 4 This embodiment provides an offshore wind farm power prediction device, which is configured to execute the offshore wind farm power prediction method in any of the aforementioned method embodiments to achieve a complete operational closed loop from meteorological data perception to scheduling available power output. The offshore wind farm power prediction device includes a data acquisition module, a wind and wave field enhancement feature construction module, a typhoon impact index determination module, a wind speed prediction module, and a power prediction module.

[0040] The data acquisition module is used to acquire typhoon path data and wave field data for the target offshore wind farm. Specifically, this module serves as the information entry point for the entire forecasting system. It is configured to retrieve typhoon path reports and gridded wave field reanalysis data covering the target sea area in real time or at regular intervals via standard meteorological data interfaces, satellite remote sensing data links, or numerical weather prediction product servers. Simultaneously, this module can be further configured to connect to the site's SCADA system to obtain wind turbine operating status and measured output data, and to interface with a basic information database to read static parameters such as site coordinates and installed capacity. Through unified acquisition and caching of multi-source heterogeneous data, a complete data foundation is provided for subsequent feature engineering. It should be understood that the specific implementation form of the data acquisition module can be an independent data acquisition server, an API gateway service, or an embedded communication unit, as long as it can meet the requirement of stable acquisition of typhoon path and wave field data.

[0041] The wind and wave field enhancement feature construction module is used to construct wind and wave field enhancement features with spatial distance attenuation weights based on the coordinates of the target offshore wind farm's center and the coordinates of external wind and wave field points in the wind and wave field data. This module is configured to execute the feature construction logic described in Example 2, i.e., it integrates a distance calculation unit and a weighted aggregation unit. In actual operation, the module first calculates the spherical distance or projected plane distance based on the latitude and longitude information of the center and each external grid point. Then, it generates corresponding spatial weights using a preset monotonically decreasing function. Finally, it performs normalized weighted summation on the wind speed, air pressure, wave height, and other elements at each point, outputting an enhanced feature vector representing the overall disturbance level of the wind farm. This modular design encapsulates the complex spatial feature extraction process into an independent processing unit, facilitating computational optimization or parallel acceleration on different hardware platforms. It also supports flexible replacement of distance metrics or aggregation algorithms without changing the main prediction process.

[0042] The typhoon impact index determination module is used to identify multiple candidate typhoons based on typhoon path data. Based on the difference in prediction error between the baseline prediction model and the enhanced prediction model, it determines the typhoon impact index of these candidate typhoons and selects the candidate typhoon with the highest impact index as the target typhoon. This module is configured to execute the typhoon screening mechanism described in Example 3. During the prediction operation phase, this module does not perform on-site model training. Instead, it loads a pre-trained typhoon impact index estimation model parameter file, inputs the real-time acquired path features and wave field features of each candidate typhoon into the model, and quickly infers the estimated impact index of each typhoon. Subsequently, the comparator inside the module automatically identifies the typhoon ID corresponding to the highest index and marks it as the target typhoon, while simultaneously routing the associated feature data to downstream modules. This configuration ensures that in scenarios with multiple typhoons coexisting, the device can complete the adaptive screening of the dominant disturbance source with a millisecond-level response speed, effectively avoiding the interference of redundant features on subsequent prediction accuracy, demonstrating the device's intelligent level at the real-time decision-making level.

[0043] The wind speed prediction module is used to input a feature set containing enhanced wind and wave field features into a pre-trained wind speed prediction model to obtain a predicted wind speed sequence. This module is configured to perform the wind speed prediction task as described in step S4 of Example 1. Internally, it stores a wind speed prediction model with completed parameter learning and maintains a multi-dimensional feature buffer containing conventional meteorological data, station foundation data, historical operational data, and enhanced wind and wave field features. Upon receiving the target typhoon features and other synchronously aligned feature data output from the upstream module, this module assembles them into a standard feature tensor and sends it to the model inference engine, outputting a predicted wind speed sequence at the hub height of the target offshore wind farm within the target time period. As an optional implementation, this module can also integrate a feature importance evaluation subunit to dynamically monitor the contribution of each input feature during inference, providing feedback for model iterative optimization.

[0044] The power prediction module is used to determine the power prediction results of the target offshore wind farm based on a pre-fitted wind power curve model and a predicted wind speed sequence. This module is configured to execute the power mapping and post-processing cascade logic described in step S5 of Example 1. Specifically, the module internally stores power curve parameters in the form of sigmoid, piecewise, or polynomial functions, enabling it to map the predicted wind speeds to initial power values ​​one by one. Subsequently, the module sequentially performs post-processing operations such as cut-in / cut-out wind speed constraints, upper and lower limit clamping of installed capacity, and time series smoothing, ultimately outputting a power prediction curve that conforms to grid dispatch specifications. The module's design fully considers the integration of physical constraints and data-driven results. Through its built-in rule engine, it ensures that even in extreme weather or model anomalies, the output power will not violate the basic operating characteristics of the wind turbine, thereby improving the safety and reliability of the device in practical engineering applications.

[0045] It should be understood that although this embodiment describes the logical division of the above five modules, this is only an exemplary architecture of this application. In actual product deployment, each module can be implemented as an independent software component, microservice, FPGA logic circuit, or dedicated AI accelerator card, or multiple modules can be merged and deployed in the same processor core according to computing resources. In addition, data interaction between modules can be achieved through various methods such as shared memory, message queues, RPC calls, or hardware buses. As long as the device as a whole has the functional configuration of acquiring data, building enhanced features, screening target typhoons, predicting wind speed, and mapping power, it falls within the scope of the technical solution of this application.

[0046] The above description is merely a specific embodiment of this application, but the scope of protection of this application is not limited thereto. Any variations or substitutions that can be easily conceived by those skilled in the art within the technical scope disclosed in this application should be included within the scope of protection of this application. Therefore, the scope of protection of this application should be determined by the scope of the claims.

Claims

1. A method for predicting the power output of an offshore wind farm, characterized in that, include: Acquire typhoon path data and wave field data for the target offshore wind farm; Based on the center coordinates of the target offshore wind farm and the coordinates of the external wind and wave field points in the wind and wave field data, a wind and wave field enhancement feature with spatial distance attenuation weight is constructed. Based on the typhoon path data, multiple candidate typhoons are identified, and based on the difference in prediction error between the baseline prediction model and the enhanced prediction model, the typhoon impact index of the multiple candidate typhoons is determined. The candidate typhoon with the largest typhoon impact index is selected as the target typhoon. The feature set containing the enhanced features of the wind and wave field is input into the pre-trained wind speed prediction model to obtain the predicted wind speed sequence. Based on the pre-fitted wind power curve model and the predicted wind speed sequence, the power prediction results of the target offshore wind farm are determined.

2. The offshore wind farm power prediction method according to claim 1, characterized in that, The construction of wind and wave field enhancement features with spatial distance attenuation weights includes: Based on the coordinates of the external wind and wave field location and the coordinates of the station center, the spherical distance between the external wind and wave field location and the target offshore wind farm is determined. The spatial distance attenuation weight is determined based on the spherical distance and the monotonically decreasing function. Based on the spatial distance attenuation weight, the wind and wave field features corresponding to the external wind and wave field locations are weighted and aggregated to obtain the wind and wave field enhancement features.

3. The offshore wind farm power prediction method according to claim 2, characterized in that, Determining the spherical distance between the external wind and wave field location and the target offshore wind farm based on the coordinates of the external wind and wave field location and the center coordinates of the wind farm includes: Based on the latitude and longitude of the coordinates of the external wind and wave field point and the latitude and longitude of the coordinates of the station center, the spherical distance is calculated using the spherical distance formula. The process of weighted aggregation of wind and wave field features corresponding to the external wind and wave field locations based on the spatial distance attenuation weight, to obtain the enhanced wind and wave field features, includes: The wind and wave field enhancement feature is obtained by weighting and summing at least one of the wind speed, air pressure, and wave height corresponding to each external wind and wave field location with the corresponding spatial distance attenuation weight, and then dividing by the sum of all the spatial distance attenuation weights.

4. The offshore wind farm power prediction method according to claim 1, characterized in that, The determination of the typhoon impact index of the multiple candidate typhoons based on the difference in prediction error between the baseline prediction model and the enhanced prediction model includes: The baseline prediction model is a prediction model that does not include the wind and wave field enhancement features, and the enhancement prediction model is a prediction model that includes the wind and wave field enhancement features. The calculation method for the difference in prediction error is as follows: within the historical typhoon impact window, the difference is obtained by subtracting the second prediction error of the enhanced prediction model from the first prediction error of the baseline prediction model. The typhoon path data and wave field data of the multiple candidate typhoons are input into a pre-trained typhoon impact index estimation model, and the typhoon impact index corresponding to each of the multiple candidate typhoons is output; wherein, the typhoon impact index estimation model is trained based on the prediction error difference.

5. The offshore wind farm power prediction method according to claim 4, characterized in that, The method for calculating the prediction error difference also includes: At least one of the following is used as the error evaluation index: mean absolute error, root mean square error, or predicted deviation of electricity. When training the typhoon impact index estimation model, the typhoon impact index is the ratio of the difference between the first prediction error and the second prediction error to the first prediction error. When the ratio is less than zero, the typhoon impact index is set to zero; The selection of the candidate typhoon with the largest typhoon impact index as the target typhoon includes: The candidate typhoon with the largest typhoon impact index among the multiple candidate typhoons is determined as the target typhoon, and the typhoon path data and the wind and wave field enhancement features corresponding to the target typhoon are included in the feature set.

6. The offshore wind farm power prediction method according to claim 1, characterized in that, The step of inputting the feature set containing the enhanced wind and wave field features into a pre-trained wind speed prediction model to obtain a predicted wind speed sequence includes: The feature set also includes conventional meteorological features, basic station features, and historical operational features; The feature set is input into the pre-trained wind speed prediction model, and the predicted wind speed sequence of the target offshore wind farm during the target prediction period is output. The method for determining the power prediction results of the target offshore wind farm based on the pre-fitted wind power curve model and the predicted wind speed sequence includes: The predicted wind speed at each moment in the predicted wind speed sequence is input into the pre-fitted wind power curve model to obtain the initial power prediction value. The initial power prediction value is post-processed to obtain the power prediction result.

7. The offshore wind farm power prediction method according to claim 6, characterized in that, The training label for the pre-trained wind speed prediction model is the hub height equivalent average wind speed. The methods for determining the equivalent average wind speed at the hub height include: The equivalent average wind speed at the hub height is obtained by weighting the measured wind speed, rated power, and availability of each wind turbine in the target offshore wind farm. Or, when there are differences in the hub height of different wind turbines, the corresponding wind speed is converted to a unified reference hub height based on the wind shear index; The pre-fitted wind power curve model is a parametric model, which includes at least one of a piecewise model, a polynomial model, or a Sigmoid model.

8. The offshore wind farm power prediction method according to claim 6, characterized in that, The post-processing of the initial power prediction value to obtain the power prediction result includes: Based on the cut-in wind speed constraint, when the predicted wind speed at the corresponding time is lower than the cut-in wind speed, the initial power prediction value at the corresponding time is set to zero or a preset low power value. Based on the cut-out wind speed constraint, when the predicted wind speed at the corresponding time is higher than the cut-out wind speed, the initial power prediction value at the corresponding time is set to zero, the power limit value, or the safe power value. Based on the upper and lower limits of installed capacity, the initial power prediction value is restricted to between the lower and upper limits of the prediction. The power prediction value after the above constraint processing is subjected to time series smoothing processing using at least one of moving average, convolutional smoothing, exponential smoothing or median filtering to obtain the power prediction result.

9. The offshore wind farm power prediction method according to claim 1, characterized in that, After acquiring the typhoon path data and wave field data of the target offshore wind farm, the method further includes: acquiring the station operation data and meteorological forecast data of the target offshore wind farm. The timestamps from different data sources are unified to the time zone where the target offshore wind farm is located, and then sorted and deduplicated by timestamp. Missing values, outliers, negative power, negative wind speed, and wind direction values ​​that are outside the reasonable range are corrected by interpolation or removed. Convert the wind direction angle into sine and cosine components; The equivalent average wind speed at the hub height at the station scale is calculated based on the station's operational data. Based on the power rationing value, shutdown status, or outlier identification results of the power curve, data from power rationing or shutdown periods that are not suitable for power curve fitting are removed. The meteorological forecast data, the station operation data, the typhoon path data, and the wind and wave field data are aligned to a unified time granularity to generate training samples.

10. A power prediction device for offshore wind farms, characterized in that, include: The data acquisition module is used to acquire typhoon path data and wave field data of the target offshore wind farm; The wind and wave field enhancement feature construction module is used to construct wind and wave field enhancement features with spatial distance attenuation weights based on the center coordinates of the target offshore wind farm and the coordinates of external wind and wave field points in the wind and wave field data. The typhoon impact index determination module is used to determine multiple candidate typhoons based on the typhoon path data, and to determine the typhoon impact index of the multiple candidate typhoons based on the difference in prediction error between the baseline prediction model and the enhanced prediction model, and to select the candidate typhoon with the largest typhoon impact index as the target typhoon. The wind speed prediction module is used to input the feature set containing the wind and wave field enhancement features into the pre-trained wind speed prediction model to obtain the predicted wind speed sequence. The power prediction module is used to determine the power prediction result of the target offshore wind farm based on a pre-fitted wind power curve model and the predicted wind speed sequence.

Images