A method for predicting output of offshore wind farm under extreme and turning weather

Abstract

The application relates to a kind of offshore wind farm output prediction methods under extreme and turning weather, belong to wind power generation and smart grid field;Including obtaining power meteorological joint data set from multiple source data;Power meteorological joint data includes reanalysis meteorological data and numerical weather prediction data after pretreatment of target time axis alignment historical measured power of wind farm;With reanalysis meteorological data as supervision benchmark, the numerical weather prediction data after pretreatment is revised to obtain revised NWP meteorological element;Based on the revised NWP meteorological element and the historical measured power data after pretreatment of wind farm, the event label corresponding to the target moment is generated;Label is used to distinguish normal period and extreme and turning event period and train model;Based on the revised NWP meteorological element in prediction window, the event label of target prediction moment is generated, corresponding pre-trained model is called respectively, future multi-step output prediction value is obtained, and constraint is applied, to obtain output prediction result.

Description

Technical Field

[0001] This invention relates to the fields of wind power generation and smart grid technology, and in particular to a method for predicting the output of offshore wind farms under extreme and transitional weather conditions. Background Technology

[0002] With the continuous growth of wind power installed capacity, especially the increasing proportion of offshore wind power grid connection, wind farm output forecasting results have become an important basis for power system dispatch and operation planning, reserve capacity configuration, safety verification, and new energy consumption. Under normal weather conditions, existing offshore wind power output assessment and forecasting technologies mostly use meteorological elements such as wind speed, wind direction, and temperature output from numerical weather prediction (NWP) as input, and combine wind turbine power curves, air density correction, altitude extrapolation, or simplified physical models to calculate offshore wind power sequences. This type of method can reflect the basic wind power generation mechanism under given meteorological conditions, has certain engineering applicability under normal weather conditions, and is often used as a baseline scheme for power systems.

[0003] However, the actual operating environment of offshore wind farms is complex, with significant nonlinearity between wind speed and power. Furthermore, it is affected by wake, differences in wind turbine units, control strategies, and switching of operating states. Simply relying on power curves and simplifying a single physical model often fails to capture the systematic deviations caused by the coupling of multiple wind turbine units and the combined effects of multiple factors, resulting in errors between the predicted output of offshore wind farms and the actual output of wind turbine units. As the application of machine learning in new energy forecasting deepens, this method mainly performs fusion learning in the feature space, without establishing a unified deviation correction and quality control framework for input meteorological errors, nor introducing ramp sensitivity and physical boundary constraints to ensure the usability of output prediction during extreme and transitional periods. Therefore, under state transition scenarios such as strong wind shedding and power limiting, the deviation of offshore wind power output prediction is still easily amplified, and unreasonable output prediction results may occur. Summary of the Invention

[0004] Based on the above analysis, the present invention aims to provide a method for predicting the output of offshore wind farms under extreme and transitional weather conditions, in order to solve the technical problem that the existing offshore wind farm output prediction methods have poor usability under extreme and transitional weather conditions.

[0005] The objective of this invention is mainly achieved through the following technical solutions: This invention provides a method for predicting the output of offshore wind farms under extreme and transitional weather conditions, comprising the following steps: Multi-source data were acquired and preprocessed to obtain a power-meteorological joint dataset; the power-meteorological joint dataset includes reanalysis meteorological data aligned to the target time axis, as well as preprocessed historical measured power data and numerical weather prediction data of offshore wind farms; Using the reanalysis meteorological data as a monitoring benchmark, the preprocessed numerical weather prediction data is biased to obtain the corrected NWP meteorological elements; based on the corrected NWP meteorological elements and the preprocessed historical measured power data of wind farms, event labels corresponding to the target time of the corrected NWP meteorological elements are generated; the event labels are used to distinguish between normal time periods and extreme and turning point event periods, and are used to train the segmented prediction model; Based on the event labels generated for each target time of the NWP meteorological elements within the prediction window, the corresponding pre-trained segmented prediction models are called for the normal period and the periods of extreme and turning events to obtain the future multi-step power output prediction values ​​of the offshore wind farm; physical boundary constraints are applied to the future multi-step power output prediction values ​​to obtain the final power output prediction results of the offshore wind farm.

[0006] Furthermore, the multi-source data includes historical measured power data of offshore wind farms, numerical weather prediction data, and reanalysis meteorological data; The preprocessing includes: A target time axis is constructed based on the sampling resolution of the historical measured power data of the offshore wind farm, and the historical measured power data of the offshore wind farm is mapped to the target time axis. Physical boundary quality control is performed on the measured power data of the offshore wind farm based on the installed capacity of the offshore wind farm, and power data points that do not meet the physical boundary quality control conditions are marked as abnormal. For continuous abnormal or missing power data segments, differentiated processing is carried out based on the duration of the abnormality to obtain preprocessed historical measured power data of offshore wind farms. For rolling numerical weather forecast data, there are multiple candidate forecast records with different start times corresponding to the same target effective time. The candidate forecast record with the shortest forecast lead time is selected as the numerical weather forecast data for the target effective time, and the preprocessed numerical weather forecast data is obtained. Based on the reanalysis meteorological data, the preprocessed historical measured power data of offshore wind farms, and the preprocessed numerical weather forecast data, a power-meteorological joint dataset aligned to the target time axis is obtained.

[0007] Furthermore, using the reanalysis meteorological data as a monitoring benchmark, bias corrections are performed on the preprocessed numerical weather prediction data to obtain the corrected NWP meteorological elements, including: The reanalysis meteorological data at the same effective time of the target on the target time axis are paired with the preprocessed numerical weather forecast data; Based on key meteorological elements in preprocessed numerical weather prediction data, a conditional feature vector is constructed; wherein, the key meteorological elements include at least the wind turbine hub height. Wind speed at the location, above ground The temperature and relative humidity of the rice; Using the original forecast values ​​from the numerical weather prediction of the key meteorological elements and the conditional feature vectors as sample data, and the reanalysis meteorological data at the corresponding time as supervision labels, a meteorological element bias correction model is trained. Using a trained meteorological element bias correction model, the preprocessed numerical weather forecast data is mapped to the first corrected NWP meteorological elements. The first corrected NWP meteorological element is constrained to obtain the second corrected NWP meteorological element; The second corrected NWP meteorological element is integrated into the preprocessed numerical weather forecast data at the corresponding time to obtain the corrected NWP meteorological element. The conditional feature vector includes time features, wind direction and flow features, meteorological element mutation rate features, and sliding window statistical features.

[0008] Furthermore, based on the corrected NWP meteorological elements and the preprocessed historical measured power data of the wind farm, event labels for the target time corresponding to the corrected NWP meteorological elements are generated, including: Based on the corrected NWP meteorological elements, the changes in each meteorological element at adjacent times are calculated to obtain the abrupt change rate of the NWP meteorological elements; wherein, the abrupt change rate of the NWP meteorological elements includes the wind turbine hub height. Wind speed change rate at the location, above ground Temperature abrupt change rate and relative humidity abrupt change rate of rice; Based on the preprocessed historical measured power data of the wind farm, the change in adjacent time moments is calculated to obtain the power ramp-up rate. Based on the wind direction in the corrected NWP meteorological elements, the change in wind direction at adjacent times is calculated; the wind direction angle at adjacent times is mapped around to map the wind direction difference to a preset angle range. Regarding the height of the wind turbine hub The rate of change in wind speed and the amount of change in wind direction at the location above ground level The temperature mutation rate, relative humidity mutation rate, and power ramp rate of meters were standardized. Standardized wind turbine hub height The rate of change in wind speed and the amount of change in wind direction at the location above ground level The transition intensity index is obtained by weighting and summing the temperature change rate, relative humidity change rate, and power ramp rate of the meter. When the turning point intensity index exceeds the preset event trigger threshold, it is determined to be an extreme turning point event trigger point; Centered on the extreme and turning point of the event, an event window is formed by extending a preset number of time steps forward and backward by m. The times within the event window are marked as extreme and turning point event periods, and the times outside the event window are marked as regular periods, generating event labels for the corresponding target times.

[0009] Furthermore, the pre-trained segmented prediction model is obtained through the following process: Based on the preprocessed historical measured power data of the offshore wind farm, the historical power lag vector and the corresponding time feature vector are obtained. Based on the corrected wind turbine hub height in the corrected NWP meteorological elements Wind speed at the location, above ground The temperature and relative humidity of the rice were used to obtain the corrected NWP meteorological element vector; Based on the corrected NWP meteorological element vector, the mutation rate vector of the corrected NWP meteorological element vector is obtained. Based on the sliding window length, the mean and standard deviation of power at historical moments within the sliding window, and the mean and standard deviation of wind speed at historical moments within the sliding window, a sliding window statistical feature vector is constructed. Based on the historical power lag vector, the corrected NWP meteorological element vector, the mutation rate vector of the corrected NWP meteorological element vector, the sliding window statistical feature vector, and the time feature vector, the step size related feature vector is calculated. Based on the corresponding event labels, the step-size related feature vectors are used as training sample data to be divided into a regular training subset and an extreme and turning point training subset. The corresponding segmented prediction models are trained using the regular training subset and the extreme and transition training subsets, respectively; wherein, the segmented prediction models include independent regular models and extreme and transition models; The process continues until the regression loss function converges, resulting in a well-trained piecewise prediction model.

[0010] Furthermore, based on the event labels of each prediction target time, the corresponding pre-trained segmented prediction models are invoked for normal periods and extreme and turning point event periods to obtain the future multi-step power output prediction values ​​of the offshore wind farm, including: For each prediction step, based on the event label corresponding to the prediction target time, perform the following steps: If the event label corresponding to the predicted target time is a regular time period, then the regular model is called, and the power prediction value of the predicted step size is output with the step size-related feature vector corresponding to the predicted target time as input. If the event label corresponding to the predicted target time is an extreme and turning point event period, then the extreme and turning point model is invoked, and the power prediction value of the predicted step size is output with the step size-related feature vector corresponding to the target time as input. Based on the power prediction values ​​of each prediction step combined with the target time axis, the future multi-step power output prediction values ​​of the offshore wind farm are obtained.

[0011] Furthermore, physical boundary constraints are applied to the predicted future multi-step power output values ​​to obtain the final offshore wind farm power output prediction results, including: The power prediction value of each prediction step in the future multi-step power output prediction value of the offshore wind farm is compared with 0, and the maximum value of the two is taken to obtain the first constraint result. The first constraint result is compared with the upper limit of the installed capacity of offshore wind farms, and the minimum value of the two is taken to obtain the final power prediction value for the prediction step. Based on the final power prediction values ​​of each prediction step combined with the target time axis, the final offshore wind farm output prediction result is obtained.

[0012] Furthermore, weights are assigned to the training samples based on the event labels and the power ramp rate: Based on the event labels, the training samples during extreme and turning point event periods are assigned a first weight enhancement coefficient; A second weighting enhancement coefficient is assigned to training samples whose absolute value of the power ramp rate exceeds a preset ramp threshold; Based on the first weight enhancement coefficient and the second weight enhancement coefficient, the sample weight of each training sample is determined; The regression loss of the segmented prediction model is weighted using sample weights.

[0013] Furthermore, the weighted regression loss function is as follows: ; in, and The first Optimal parameters for the conventional model and the extreme and inflection models under each step size; and These are the parameters to be optimized for the conventional model and the extreme and inflection models, respectively. and The first The set of regular samples and the set of extreme and turning points under each step size; For the target time Sample weights; The loss function; The measured active power of the wind farm at the target prediction time; For the first The input feature vector corresponding to each step size; For the conventional model In parameters The output below; For extreme and turning point models In parameters The output below.

[0014] Furthermore, the conventional model and the extreme and transition models are respectively constructed based on the LightGBM model.

[0015] Compared with the prior art, the present invention can achieve at least one of the following beneficial effects: 1. This invention establishes a meteorological bias correction mechanism for non-stationary errors, suppressing the amplification of input bias at its source. Addressing the shortcomings of existing methods that directly input NWP numerical weather prediction data into the prediction model and lack a unified bias correction framework for extreme weather, leading to the amplification of meteorological errors to power prediction errors, this invention utilizes ERA5 reanalysis data as a high-reliability reference benchmark to construct a condition-feature-driven NWP meteorological element bias correction model. This model extracts multi-dimensional conditional features such as time period, wind direction and flow, meteorological abrupt change rate, and sliding window statistics, enabling the meteorological element correction process to perceive error variation patterns under different meteorological scenarios. It achieves adaptive correction of non-stationary errors in key elements such as wind speed, temperature, and humidity, and is further constrained by both physical and statistical boundaries to ensure the rationality of the meteorological element correction results. Through this mechanism, the systematic bias and error drift of NWP numerical weather prediction under extreme and transitional weather conditions are effectively reduced, decreasing the amplification effect of meteorological element bias to power prediction errors from the data source, thus solving the fundamental problem of "large input bias" in existing methods. 2. This invention constructs a transition intensity index and an event-driven segmented prediction mechanism to address the shortcomings of single models in adapting to extreme and transitional weather. Addressing the deficiencies of existing methods that use a single model to cover all weather types and lack explicit identification and mechanism switching capabilities for extreme and transitional weather processes, leading to significantly amplified prediction errors in transition windows, this invention constructs a transition intensity index based on the mutation rate of meteorological elements. This enables automatic and quantitative identification of extreme and transitional weather events, and forms a continuous sequence of event labels through event window expansion. Based on this, an event-driven segmented direct multi-step prediction architecture is established, modeling regular periods and extreme and transitional periods separately: a regular model and an extreme / transitional model are independently trained for each prediction step. During prediction, the system automatically switches to the corresponding regular model or extreme / transitional model based on the event label at the target time. Simultaneously, a direct multi-step prediction strategy avoids error accumulation in recursive methods. Through the mechanism of "identification first, segmentation then prediction," the segmented prediction model can automatically adopt an adaptation strategy when weather mechanisms undergo abrupt changes, maintaining accuracy during regular periods and capturing abrupt changes during extreme and transitional periods, effectively solving the problem of "insufficient model adaptability" in existing technologies. 3. This invention introduces ramp-sensitive weighting and hard physical boundary constraints to ensure the accuracy and engineering usability of critical window predictions. Addressing the shortcomings of existing methods in the background, which lack targeted constraints on critical ramp and descent windows and lack physical boundary verification of output results, leading to misalignment of critical window amplitude and timing, and unreasonable outputs such as negative power or over-installation, this invention introduces a ramp-sensitive sample weighting mechanism during the segmented prediction model training phase. Training samples in extreme and transitional event windows, and those experiencing high-power ramps, are assigned higher training weights than regular samples. This allows the segmented prediction model to prioritize the rapid power change process most important to grid dispatch during optimization, enhancing its ability to fit the consistency of critical window amplitude and timing. In the prediction output phase, non-negative hard physical boundary constraints and the upper limit of installed capacity are applied, strictly limiting the power value of each prediction step to a physically feasible range. Furthermore, the upper limit of installed capacity can be dynamically adjusted based on actual operating conditions such as wind turbine maintenance and power limiting commands, ensuring the physical feasibility of the offshore wind farm output timing prediction results. By combining a "weighted training + hard constraint output" strategy, we ensure that the power output time series prediction results of offshore wind farms can both "capture" extreme changes and "do not violate" physical laws, directly meeting the engineering application requirements of high-reliability prediction results for grid dispatching operation and safety verification.

[0016] In this invention, the above-described technical solutions can be combined with each other to achieve more preferred combinations. Other features and advantages of this invention will be set forth in the following description, and some advantages may become apparent from the description or be learned by practicing the invention. The objects and other advantages of this invention can be realized and obtained from what is particularly pointed out in the description and drawings. Attached Figure Description

[0017] The accompanying drawings are for illustrative purposes only and are not intended to limit the invention. Throughout the drawings, the same reference numerals denote the same parts.

[0018] Figure 1 This is a flowchart of a method for predicting the output of an offshore wind farm under extreme and transitional weather conditions, as described in an embodiment of the present invention. Figure 2 This is a schematic diagram of the rolling prediction time axis as exemplified in an embodiment of the present invention; Figure 3 This is a schematic diagram of the functional modules of an offshore wind farm output prediction system under extreme and transitional weather conditions, as described in an embodiment of the present invention. Detailed Implementation

[0019] Preferred embodiments of the present invention will now be described in detail with reference to the accompanying drawings, which form part of this application and are used together with the embodiments of the present invention to illustrate the principles of the present invention, but are not intended to limit the scope of the present invention.

[0020] This invention addresses three major technical problems in offshore wind power output prediction under extreme and transitional weather conditions: large input deviations of meteorological elements, insufficient adaptability of prediction models, and lack of constraints on output prediction results. It establishes an offshore wind farm output prediction method that integrates meteorological deviation suppression, adaptive identification of extreme and transitional events, and physical boundary constraints of the ramping process, in order to improve the stability, robustness, and engineering usability of offshore wind farm output prediction results.

[0021] To address the problems in offshore wind power output prediction under extreme and transitional weather conditions, such as significant input bias of meteorological elements, insufficient identification of regular and transitional weather processes, and lack of constraints on predicted output results during critical ramp and sudden drop windows, this invention establishes a time series prediction method that integrates meteorological data bias suppression, identification of extreme and transitional weather events and condition-driven modeling, and physical consistency constraints on output changes. This method aims to reduce the propagation and amplification of meteorological bias on offshore wind farm power prediction errors and improve the stability, robustness, and engineering usability of the model's predicted output during extreme and transitional weather periods.

[0022] The offshore wind farm output prediction method of this invention is applicable to multi-step output prediction and evaluation of wind farms under normal weather conditions, as well as extreme and transitional weather conditions such as typhoons and cold fronts.

[0023] Extreme weather refers to rare meteorological events with low probability of occurrence, significant deviation from the climate average in intensity, and a high risk of causing disasters, such as typhoons, severe convection, and extreme high temperatures. Turning point weather refers to a weather process in which meteorological elements (such as wind speed, wind direction, and temperature) undergo rapid, significant, and directional changes in a short period of time, resulting in a change in weather conditions from one form (such as sunny and stable weather) to another form (such as strong winds).

[0024] This invention's method is based on continuously updated NWP numerical weather prediction data and historical measured power data of offshore wind farms. First, it selects data corresponding to the shortest effective forecast lead time from the multi-source NWP numerical weather prediction data according to the matching relationship between forecast start time and forecast lead time, and splices them in time series. Then, it uses ERA5 reanalysis data as supervision labels to construct a meteorological element bias correction model, improving the consistency and stability of key meteorological input features. Based on this, it constructs extreme and transition intensity indicators and generates weather event labels, dividing the training sample data into two categories: normal weather and extreme / transitional weather. It establishes an event-driven segmented prediction model based on LightGBM to perform multi-step predictions, achieving independent fitting of future power with different lead times. Simultaneously, it introduces a sample enhancement mechanism based on power ramp rate to improve the segmented prediction model's ability to characterize rapidly changing processes. Finally, it applies physical boundary constraints such as non-negativity and upper limits of installed capacity to the offshore wind farm output of the segmented prediction model, thereby obtaining more stable, reasonable, and engineering-approving offshore wind farm output prediction results under both normal and extreme / transitional weather conditions.

[0025] Example 1: A specific embodiment of the present invention discloses a method for predicting the power output of offshore wind farms under extreme and transitional weather conditions, such as... Figure 1 As shown, it includes the following steps: Step S1: Acquire multi-source data and preprocess it to obtain a power-meteorological joint dataset; wherein, the power-meteorological joint dataset includes reanalysis meteorological data aligned to the target time axis and preprocessed historical measured power data and numerical weather prediction data of offshore wind farms; Step S2: Using the reanalysis meteorological data as a supervision benchmark, the preprocessed numerical weather forecast data is biased to obtain the corrected NWP meteorological elements; based on the corrected NWP meteorological elements and the preprocessed historical measured power data of wind farms, event labels corresponding to the target time of the corrected NWP meteorological elements are generated; the event labels are used to distinguish between normal time periods and extreme and turning point event periods, and are used to train the segmented prediction model; Step S3: Based on the event labels generated for each target time of the NWP meteorological elements within the prediction window after correction, call the corresponding pre-trained segmented prediction models for the normal period and the extreme and turning point event periods to obtain the future multi-step power output prediction values ​​of the offshore wind farm; apply physical boundary constraints to the future multi-step power output prediction values ​​to obtain the final offshore wind farm power output prediction results.

[0026] Step S1 is to acquire and preprocess basic data for training the segmented prediction model; Step S2 is the offline training stage of the segmented prediction model. After the offline training is completed, the pre-trained segmented prediction model is obtained. Step S3 is the online prediction stage using the pre-trained segmented prediction model.

[0027] Step S1, specifically.

[0028] The core technical problem addressed in this step is: how to integrate the measured power data of offshore wind farms and the NWP numerical weather prediction data, which have different sources, time resolutions, and time reference systems, into a joint dataset of power and meteorological elements that is aligned on a unified time axis and has controllable quality.

[0029] Specific objectives: ①Unify the temporal resolution and time axis of measured power data from offshore wind farms with meteorological element data in NWP numerical weather prediction; ②Remove or repair outliers and missing values ​​in the measured power data of offshore wind farms; ③Resolve the redundancy and conflict issues in the rolling forecasts of NWP numerical weather prediction data, where "multiple forecast versions correspond to the same target time"; ④ Establish a joint power-meteorological dataset with the target time axis fully aligned to provide a standardized data foundation for subsequent steps.

[0030] The multi-source data includes historical measured power data of offshore wind farms, numerical weather prediction data, and reanalysis meteorological data. The preprocessing includes: A target time axis is constructed based on the sampling resolution of the historical measured power data of the offshore wind farm, and the historical measured power data of the offshore wind farm is mapped to the target time axis. Physical boundary quality control is performed on the measured power data of the offshore wind farm based on the installed capacity of the offshore wind farm, and power data points that do not meet the physical boundary quality control conditions are marked as abnormal. For continuous abnormal or missing power data segments, differentiated processing is carried out based on the duration of the abnormality to obtain preprocessed historical measured power data of offshore wind farms. For rolling numerical weather forecast data, there are multiple candidate forecast records with different start times corresponding to the same target effective time. The candidate forecast record with the shortest forecast lead time is selected as the numerical weather forecast data for the target effective time, and the preprocessed numerical weather forecast data is obtained. Based on the reanalysis meteorological data, the preprocessed historical measured power data of offshore wind farms, and the preprocessed numerical weather forecast data, a power-meteorological joint dataset aligned to the target time axis is obtained.

[0031] (1) Data source and characteristics of raw data: The accuracy of offshore wind power sequence forecasting is highly dependent on the consistency of the quality and time alignment of the input meteorological data.

[0032] To ensure the stability of inputs for subsequent bias correction, event identification, and training of the segmented prediction model for offshore wind farm output, this invention first acquires historical measured power data of offshore wind farms, continuously updated NWP numerical weather prediction data, and ERA5 (ECMWF Reanalysis v5, the fifth-generation global atmospheric reanalysis dataset from the European Centre for Medium-Range Weather Forecasts). The ERA5 reanalysis meteorological data is the fifth-generation global atmospheric reanalysis dataset released by the European Centre for Medium-Range Weather Forecasts (ECMWF). By deeply fusing numerical weather prediction models with historical observation data (including satellite, weather station, and radar data) from around the world and performing physical consistency back-calculation, a set of high-precision historical meteorological records that is temporally continuous, spatially global, and physically consistent is generated. This dataset has been continuously updated since 1940, providing hundreds of atmospheric, land, and ocean variables (including wind speed, temperature, humidity, and air pressure) at an hourly horizontal resolution of approximately 31 kilometers, and is freely available.

[0033] ① The measured power data of offshore wind farms comes from sensors on the wind turbines or the SCADA (Supervisory Control And Data Acquisition) system of the wind farm's central control center.

[0034] The measured active power data of offshore wind farms, including timestamps and the measured active power at the corresponding time, often have quality problems such as missing measurements, spikes, and out-of-bounds errors. The time resolution is usually 15 minutes or 1 hour. ②NWP numerical weather forecast data, including but not limited to forecast values ​​of meteorological elements such as wind speed, wind direction, temperature, humidity, and air pressure, are usually released in a rolling manner, according to a fixed cycle (such as 6 hours); in the future, the same valid time will correspond to multiple forecast versions with different start times and forecast lead times.

[0035] For example, The NWP numerical weather forecast, which begins at 8:00 AM today, will provide a forecast for 2:00 PM tomorrow (with a lead time of +30 hours). The NWP numerical weather forecast, which will be issued starting at 12:00 noon today, will also provide a forecast for 2:00 pm tomorrow (with a lead time of +26 hours).

[0036] NWP numerical weather prediction data is the target data that needs to be corrected. It is obtained free of charge from major meteorological centers and is a recognized public data resource in the meteorological field.

[0037] ③ ERA5 reanalysis meteorological data, including but not limited to reanalysis values ​​of meteorological elements such as wind speed, wind direction, temperature, humidity, and air pressure; the time resolution is usually 1 hour; global coverage, physical consistency, and high accuracy serve as a monitoring benchmark for subsequent bias correction.

[0038] ERA5 reanalysis data is used as a "monitoring baseline" to correct meteorological elements in NWP numerical weather prediction data.

[0039] ERA5 reanalysis data is generated by fusing historical observations and numerical models using advanced data assimilation techniques, resulting in a high-precision historical meteorological record set that is time-continuous and physically consistent, covering the period from 1940 to the present.

[0040] The ERA5 reanalysis data for subsequent training of the "meteorological element bias correction model" and historical data (such as 2020-2025) can be stably and freely obtained from CDS (Climate Data Store, European Centre for Medium-Range Weather Forecasts). This step can also utilize ERA5T data with a 5-day delay. ERA5T stands for "ERA5-Temporary," referring to reanalysis data no more than three months after the real-time point. ERA5T data for the latest time period is continuously supplemented at a frequency of "daily updates with a lag of approximately 5 days." For this invention, this means it can provide the necessary supervisory benchmark after supervised correction in near real-time, meeting the timeliness requirements for power output forecasting in online offshore wind farms.

[0041] Continuously acquire ERA5 or ERA5T data for subsequent full or incremental training of meteorological element bias correction models and to supplement the monitoring baseline data.

[0042] (2) Construct a unified target timeline: Based on the sampling time resolution of the measured active power data from offshore wind farms and the requirements of the prediction task, a unified target time axis is constructed, as shown below: Formula (1); in, For the target timeline, For the 1st, 2nd, ..., The effective time for each target; , For the first , No. The effective time for each target , This represents the sampling time resolution.

[0043] The principle for selecting the target time axis is: based on the sampling time resolution of the measured active power data of offshore wind farms, while also taking into account the requirements of the forecasting task.

[0044] Example 1: If the measured power data of the SCADA system in the wind farm control center has a sampling time resolution of 15 minutes, then take... Minutes are used to construct a target timeline with a 15-minute sampling time resolution; Example 2: If the forecast task is a day-ahead hourly forecast (the next 24 hours), the measured active power data and meteorological data of offshore wind farms can also be uniformly resampled to a 1-hour resolution. Hour.

[0045] Once the target timeline is determined, all subsequent data is mapped to the target timeline to strictly ensure time consistency.

[0046] (3) Physical boundary quality control of measured power data from offshore wind farms: The measured power data of offshore wind farms are mapped to the target time axis as shown below: Formula (2); in, For offshore wind farms at all times The measured active power.

[0047] Based on the installed capacity of offshore wind farms Physical boundary quality control is performed on the measured power data of offshore wind farms at any given time. If the condition in formula (3) is not met, then the measured power data point is marked as abnormal. See below: Formula (3); For segments with continuous anomalies or missing data, a differentiated output strategy is adopted based on the duration of the anomaly: Suppose a certain consecutive missing or abnormal segment contains If there are consecutive abnormal time points, then the duration is... As shown below: Formula (4); Based on duration and preset short-term missing detection threshold It performs short-time segment processing and long-time segment processing.

[0048] ① Short segment processing: when If so, the segment is defined as a short-term missing test or a short-term abnormal segment; For short-term missing or short-term abnormal segments, linear interpolation reconstruction is used, which utilizes the interpolation method between adjacent valid time points for reconstruction.

[0049] If the measurement time is missing The most recent valid times before and after are respectively and The interpolation result is as follows: Formula (5); in, Effective time respectively and The measured power value at the location.

[0050] ② Long segment processing: when When this occurs, the segment is defined as a long-term missing segment or a long-term abnormal segment; For segments with long-term missing data or long-term abnormal data, interpolation reconstruction is not recommended. These segments should be marked as unusable and skipped directly in the sample window containing these segments during subsequent training sample construction.

[0051] For example, when A short-term missing threshold can be preset every minute. Anomalies defined as consecutive missing data or anomalous segments lasting no more than 4 hours are considered short-term anomalous segments and are repaired using linear interpolation. Segments exceeding 4 hours are defined as long-term anomalous segments and are removed. In practical applications, adjustments can be made flexibly based on data quality and prediction task requirements. .

[0052] Short-term missing data can be restored by linear interpolation to restore data continuity, while long-term missing data results in severe information loss. Forced interpolation will introduce false information, so it is better to remove them to avoid contaminating the training sample data.

[0053] This section obtains preprocessed historical measured power data of offshore wind farms.

[0054] (4) NWP numerical weather forecasts are rolled out and the data preprocessing is carried out using the "shortest forecast lead time priority" splicing strategy.

[0055] NWP numerical weather forecasts are published in a rolling format, for example: The NWP numerical weather forecast, which starts reporting at 08:00, provides a forecast for 14:00 (forecast lead time + 6 hours). The NWP numerical weather forecast, which starts reporting at 12:00, also provides a forecast for 14:00 (forecast lead time + 2 hours).

[0056] For NWP numerical weather forecast data that is updated on a rolling basis, at the same target effective time... There may be several candidate forecast records from different reporting times.

[0057] Set the effective time The corresponding candidate forecast set for NWP numerical weather prediction is shown below: Formula (6); in, For the first The start time of each NWP candidate forecast record For the first Candidate forecast records at the effective time The corresponding meteorological element vectors (including but not limited to wind speed, wind direction, temperature, humidity, and air pressure); In order to be effective at the time A set of numbered indexes of all available NWP candidate numerical weather forecast records that have passed quality control.

[0058] For example, for 14:00, there may be both a forecast that started at 08:00 (e.g., forecast number 1) and a forecast that started at 12:00 (e.g., forecast number 2). .

[0059] "Shortest forecast lead time priority" stitching strategy: To obtain a continuous and consistent NWP numerical weather forecast sequence, this invention defines the first... The forecast lead time for each NWP numerical weather prediction candidate record at time t is shown below: Formula (7); in, For the first NWP numerical weather prediction candidate records at the target effective time Lead time is the time between the start time of the forecast and the target effective time, i.e., how far in advance the forecast is made.

[0060] From the candidate set controlled by physical boundary quality, the forecast with the shorter lead time is selected as the NWP numerical weather forecast for the effective time of the target, as shown below: Formula (8); in, In order to make The smallest NWP numerical weather prediction candidate forecast record index is the number of the "best" NWP numerical weather prediction record selected from the candidate set; if the forecast lead time starting at 12:00 is 2 hours, and the forecast lead time starting at 08:00 is 6 hours, then... Pointing to the one that starts reporting at 12:00; To splice and match the target timeline Aligned NWP numerical weather forecast meteorological element vectors.

[0061] Indicates the time when the target is valid From all available NWP numerical weather forecasts, select the one with the shortest forecast lead time; This means using the meteorological element values ​​from the NWP numerical weather prediction record with the shortest forecast lead time as the target effective time. The final NWP weather elements.

[0062] NWP numerical weather prediction candidate forecast records refer to the effective times for the same target under different start times and different forecast lead times. The provided meteorological element forecast records. Each candidate forecast record includes the effective time for that target. The meteorological element vector.

[0063] The shorter the forecast lead time, the smaller the initial long-term disturbance of the NWP numerical weather prediction, and the higher the forecast accuracy is generally. Therefore, NWP numerical weather prediction records that are closer to the effective time of the target should be preferred.

[0064] The physical meaning of formula (8): For example, among all candidate NWP numerical weather forecast records that pass through physical boundary quality control, the one with the shortest forecast lead time is selected as the final input for the target effective time.

[0065] For example, for the target effective time 14:00, if there are both a +6h forecast (lead time of 6 hours) starting from 08:00 and a +2h forecast (lead time of 2 hours) starting from 12:00, and both of them pass quality control, then the +2h forecast starting from 12:00 is selected as the meteorological element of the NWP numerical weather prediction at time 14:00.

[0066] Fault protection is applied to the meteorological element sequences of NWP numerical weather prediction data at the target effective time. when = When all candidate forecast records fail the physical boundary quality control, the forecast is generated using persistent completion from the previous time step. To avoid breaks in the meteorological element sequence of NWP numerical weather prediction data, as shown below: Formula (9); When the target is valid If no NWP forecast records are available, use the previous time. The NWP meteorological element values ​​are used as substitute values ​​for the current moment, assuming that the meteorological elements remain unchanged in a short period of time. This is to avoid breaks in the meteorological element sequence in NWP numerical weather prediction. Although the accuracy of persistent completion is limited, it is better than data missing when there are short-term omissions, and can ensure the continuity of the offshore wind farm output prediction process.

[0067] This section obtains preprocessed numerical weather forecast data.

[0068] (5) Time alignment of ERA5 reanalysis data ERA5 reanalysis data is historical backtesting data; its timeline is known and continuous, requiring no stitching. Simply select the target timeline. Extract the ERA5 meteorological element vector at the corresponding time. The data, along with the preprocessed historical measured power data of offshore wind farms and the preprocessed numerical weather forecast data, form a power-meteorological joint dataset aligned to the target time axis.

[0069] This results in a power-meteorological joint dataset that is fully aligned with the target timeline, as shown below: Formula (10); in, This is the preprocessed historical measured power data of offshore wind farms; This refers to the preprocessed numerical weather forecast data after splicing. A vector of meteorological elements in ERA5 reanalysis data aligned with the target time axis.

[0070] This power-meteorological joint dataset is used for subsequent meteorological element bias correction, weather event identification, and training of offshore wind power segmentation prediction models.

[0071] This power-meteorological joint dataset will be used in subsequent steps, specifically: Step S2: Meteorological element deviation correction, to As the object to be corrected, As a monitoring benchmark; identification of routine, extreme, and turning point events, based on the revised... and Calculate the transition intensity index; Step S3 Power prediction modeling, based on the corrected and historical measured power Construct features and train a segmented prediction model (including the regular model and extreme and inflection models) based on the LightGBM model.

[0072] Step S1 aims to acquire historical measured power data of offshore wind farms, continuously updated NWP numerical weather forecast data, and ERA5 reanalysis data. Through unified time axis construction, quality control, missing data handling, and prioritizing the shortest forecast lead time for stitching together, a power-meteorological joint dataset aligned with the target time axis is formed, providing a stable data foundation for subsequent bias correction, weather event identification, and power prediction modeling.

[0073] Step S2 includes steps S21-S22.

[0074] Step S21: Using the reanalysis meteorological data as a monitoring benchmark, the key meteorological elements in the preprocessed numerical weather forecast data are corrected for deviations to obtain the corrected NWP meteorological elements.

[0075] Using the reanalysis meteorological data as a monitoring benchmark, bias corrections are performed on the preprocessed numerical weather prediction data to obtain the corrected NWP meteorological elements, including: The reanalysis meteorological data at the same effective time of the target on the target time axis are paired with the preprocessed numerical weather forecast data; Based on key meteorological elements in preprocessed numerical weather prediction data, a conditional feature vector is constructed; wherein, the key meteorological elements include at least the wind turbine hub height. Wind speed at the location, above ground The temperature and relative humidity of the rice; Using the original forecast values ​​from the numerical weather prediction of the key meteorological elements and the conditional feature vectors as sample data, and the reanalysis meteorological data at the corresponding time as supervision labels, a meteorological element bias correction model is trained. Using a trained meteorological element bias correction model, the preprocessed numerical weather forecast data is mapped to the first corrected NWP meteorological elements. The first corrected NWP meteorological element is constrained to obtain the second corrected NWP meteorological element; The corrected NWP meteorological elements are then integrated into the preprocessed numerical weather forecast data at the corresponding time to obtain the corrected NWP meteorological elements. The conditional feature vector includes time features, wind direction and flow features, meteorological element mutation rate features, and sliding window statistical features.

[0076] For example, Take a height of 100 meters; in actual needs, the wind speed at the corresponding height can be obtained based on the actual height of the wind turbine hub.

[0077] Above ground The temperature of the meter affects the air density, which in turn affects the fan output (air density correction). Temperature deviations will affect the applicability of the power curve; for example, It is 2 meters above ground level; however, this can be changed according to actual needs in specific applications.

[0078] In NWP numerical weather prediction and ERA5 reanalysis data, temperature variables are output using "T2m" (Temperature at 2 meters above ground) as the standard. The 2-meter value is chosen to maintain consistency with global meteorological observation data. In practical applications, this can be flexibly adjusted according to specific needs.

[0079] Preprocessed NWP numerical weather forecast data Systematic forecast biases exist, especially under extreme and transitional weather conditions where errors increase significantly and become non-stationary. Under extreme and transitional weather conditions, the systematic biases of NWP numerical weather predictions on key elements such as wind speed, temperature, and humidity are significantly amplified, and the errors exhibit non-stationary characteristics as the weather process evolves, directly affecting forecasts. Inputting subsequent segmented prediction models for offshore wind power can easily lead to the amplification of "input meteorological element errors" into "power prediction errors" during critical windows.

[0080] To reduce the propagation impact of meteorological biases on offshore wind farm power forecasting, this invention, after preprocessing the NWP numerical weather prediction data, uses ERA5 reanalysis data as a monitoring benchmark to process the preprocessed NWP numerical weather prediction data. The key meteorological elements in the data are corrected for biases to obtain the corrected meteorological element vectors of the NWP numerical weather prediction data. This will improve the credibility of meteorological data.

[0081] For example, the set of key meteorological elements in the preprocessed NWP numerical weather prediction data is shown below: Formula (11); in, This is the set of elements to be corrected. These are the heights of the wind turbine hubs. Wind speed at the location, above ground The temperature of the rice, and the relative humidity.

[0082] Wind turbine hub height Wind speed is the most critical input variable for wind power prediction. Wind speed deviation directly leads to power deviation, and the wind speed error is amplified more significantly under extreme weather conditions. The actual hub height of offshore wind turbines may be 80m, 90m, 100m, 110m, 120m, or 150m. It affects air density and is related to weather processes such as fronts and precipitation. Humidity changes reflect the attributes of weather systems and help depict transition processes.

[0083] Only correcting the preprocessed NWP numerical weather forecast data The three core meteorological elements that are directly related to the physical properties of power prediction are corrected, rather than all meteorological elements, to avoid overfitting and computational redundancy.

[0084] In different business scenarios It can be replaced with the wind turbine hub height corresponding to the business scenario. Wind speed at the location and above ground Temperature at a height of meters.

[0085] Reanalysis meteorological data at the same effective time point on the target time axis are paired with preprocessed numerical weather prediction data; for any meteorological element On the target timeline The above constructs paired samples as follows: Formula (12); in, For meteorological elements in preprocessed numerical weather prediction data The meteorological element values ​​at time t are for the NWP numerical weather prediction of the first time. Each meteorological element at time Forecast values ​​(for meteorological elements to be corrected); For ERA5 reanalysis data at time The values ​​for the same element (as the monitoring baseline), and compared with... Perform strict timestamp alignment on the target timeline.

[0086] Pairing significance: Preprocessed NWP numerical weather forecast data As input features, the ERA5 supervision benchmark value As a supervised training label (target output), the meteorological element bias correction model learns the mapping relationship from "biased NWP" to "more realistic ERA5".

[0087] The forecast error of NWP numerical weather prediction is not random noise, but is closely related to the time period, wind direction and flow direction, as well as the intensity of sudden changes in weather processes and recent trends.

[0088] Based on meteorological elements in the preprocessed NWP numerical weather prediction data, a conditional feature vector is constructed. It is used as a meteorological element bias correction model to perceive the current meteorological situation, as shown below: Formula (13); in, For a moment Conditional eigenvectors, These represent time characteristics, wind direction and flow characteristics, meteorological abrupt change rate characteristics, and sliding window statistical characteristics, respectively; superscript This indicates transpose.

[0089] Conditional eigenvectors These are derived features of NWP numerical weather prediction data.

[0090] ①Time characteristics As shown below: Formula (14); in, For a moment The hour varies because the NWP numerical weather prediction error exhibits a diurnal variation pattern; for example, the boundary layer turbulence intensifies during the day, resulting in different prediction error characteristics. For a moment The month in question varies depending on the seasonal patterns of errors in NWP numerical weather predictions. For example, the seasonal error pattern for typhoons differs from that of the winter monsoon season.

[0091] The errors of the same NWP numerical weather prediction exhibit periodic variations across different time scales. Introducing time characteristics helps meteorological element bias correction models capture this periodicity.

[0092] ② Characteristics of wind direction and flow As shown below: Formula (15); in, The wind direction angle for NWP numerical weather forecasts is 0°-360°; Wind direction is represented by a circumference; 0° and 360° represent the same direction; using sine and cosine forms to represent wind direction avoids the problem of discontinuity in wind direction angles around 0°-360°; ③ Characteristics of meteorological element mutation rate As shown below: Formula (16); in, Let NWP be the difference between the current time t and the previous time. high wind speed The change in; Let NWP be the difference between the current time t and the previous time. High temperature The change This represents the humidity of NWP at the current time t compared to the previous time. The change in; The mutation rate characteristics are shown below: Formula (17); ④ Statistical characteristics of sliding window As shown below: Formula (18); in, The preset sliding window length; They are respectively of length The mean and standard deviation are calculated within the sliding window.

[0093] For example, when the target timeline resolution is 1 hour, The value can be 3 or 6, representing the sliding statistical window of the past 3 hours or 6 hours, respectively.

[0094] Meteorological element bias correction model With regard to training objectives: For each preprocessed NWP numerical weather forecast, the meteorological elements to be corrected ,in, Indicates an index of meteorological elements to be corrected. For the set of meteorological elements to be corrected, establish the deviation correction function for each meteorological element to be corrected. , will the The forecast values ​​of the NWP numerical weather prediction for each meteorological element and the conditional eigenvector are used as input to output the corrected meteorological elements of the NWP numerical weather prediction: Formula (19); in, For the first The corrected meteorological element values ​​for each meteorological element; The parameters for the meteorological element deviation correction model.

[0095] During the training phase of the meteorological element bias correction model, ERA5 reanalysis data was used. As a monitoring baseline label, the optimal parameters of the meteorological element bias correction model are obtained by minimizing the error between the correction output of the meteorological element bias correction model and the ERA5 monitoring baseline label value, as shown below: Formula (20); in, To correct the optimal parameters of the model for meteorological elements; This is a historical time sample set used to train the meteorological element bias correction model; The loss function can be represented by squared error, absolute error, or Huber loss function.

[0096] The physical meaning of formula (20) is to minimize the difference between the output of the meteorological element deviation correction model and the ERA5 supervision benchmark value on the training sample time set, so as to learn the nonlinear variation law of NWP meteorological forecast error under different time, wind direction and weather change conditions.

[0097] Historical time samples are shown below: Formula (21); Sample labels are .

[0098] Historical time sample set The complete structure is shown below: Formula (22); Typically, this includes tens of thousands of paired samples, such as one sample every 15 minutes or hour over the past 2-3 years. The sample data and sample label timestamps are strictly aligned.

[0099] For example, a meteorological element bias correction model It can be implemented using XGBoost, LightGBM, random forest, neural network or a combination thereof; in a preferred embodiment, an XGBoost regression model is used for training.

[0100] After training, the combination of correction results for key meteorological elements in the preprocessed NWP numerical weather prediction is the first corrected NWP meteorological element vector. As shown below: Formula (23); in, The corrected versions are as follows: Wind speed at high altitudes, above ground Temperature and relative humidity.

[0101] To avoid unreasonable extrapolation by the meteorological element bias correction model under extreme sample scarcity conditions, the first corrected NWP meteorological element vector was modified. The correction results were further constrained to obtain the second corrected NWP meteorological elements. As shown below: Formula (24); in, Range clipping operator; For the first Reasonable value ranges for each meteorological element to be corrected. Different meteorological elements are subject to corresponding reasonable range constraints; for example, wind speed, air temperature, and relative humidity each have different upper and lower limits.

[0102] Range clipping operator Definition: If the output value is lower than the lower limit, the lower limit value is taken; if it is higher than the upper limit, the upper limit value is taken; otherwise, it remains unchanged.

[0103] The reasonable range of values ​​is determined jointly by the physical boundaries of the meteorological elements and the statistical range of historical samples. Specifically, it can be determined as follows: Formula (25); in, and The first The physical reasonable lower and upper limits of each meteorological element; and These are the first in the ERA5 reanalysis sample. The lower and upper quantiles of each meteorological element; α is a preset quantile threshold, which is 0.01-0.05. For example, α is 0.01. In order to target the The margin coefficients set for each meteorological element are used to avoid over-cropping of extreme weather samples.

[0104] For example, ① Regarding the height of the wind turbine hub The wind speed at the location can be set to a lower limit of 0, and the upper limit can be determined by the high quantile of the ERA5 historical samples and in combination with the extreme wind speed range of the sea area where the wind farm is located. ②For above ground The upper and lower limits of the temperature in meters can be determined by the low and high quantiles of historical ERA5 samples and by adding a preset margin. ③ Regarding relative humidity Its physical range can be set to [0, 100]%.

[0105] like Then output ;like Then output Otherwise, keep constant.

[0106] The corrected NWP weather elements By incorporating the preprocessed NWP numerical weather forecast for the corresponding time point, the corrected NWP meteorological element vector is obtained. Replace the preprocessed numerical weather forecast data This serves as input for subsequent identification of extreme and transitional weather events, and for event-driven segmented prediction modeling.

[0107] Step S21 uses ERA5 reanalysis data as a monitoring benchmark and a meteorological element bias correction model built on XGBoost to correct non-stationary errors in key meteorological elements such as wind speed, temperature, and humidity in the preprocessed NWP numerical weather forecast data. This reduces the amplification of meteorological input bias to power prediction errors under extreme and transitional weather conditions and improves the reliability of the basic data.

[0108] Step S22: Based on the corrected NWP meteorological elements and the preprocessed historical measured power data of the wind farm, generate event tags for the corrected NWP meteorological elements; the event tags are used to distinguish between normal periods and extreme and turning point events.

[0109] The core problem addressed in this step is: how to automatically and quantitatively determine whether the current weather process is in an extreme / transitional state.

[0110] Based on the corrected NWP meteorological elements and the preprocessed historical measured power data of the wind farm, event tags corresponding to the target time for the corrected NWP meteorological elements are generated, including: Based on the corrected NWP meteorological elements, the changes in each meteorological element at adjacent times are calculated to obtain the abrupt change rate of the NWP meteorological elements; wherein, the abrupt change rate of the NWP meteorological elements includes the wind turbine hub height. Wind speed change rate at the location, above ground Temperature abrupt change rate and relative humidity abrupt change rate of rice; Based on the preprocessed historical measured power data of the wind farm, the change in adjacent time moments is calculated to obtain the power ramp-up rate. Based on the wind direction in the corrected NWP meteorological elements, the change in wind direction at adjacent times is calculated; the wind direction angle at adjacent times is mapped around to map the wind direction difference to a preset angle range. Regarding the height of the wind turbine hub The rate of change in wind speed and the amount of change in wind direction at the location above ground level The temperature mutation rate, relative humidity mutation rate, and power ramp rate of meters were standardized. Standardized wind turbine hub height The rate of change in wind speed and the amount of change in wind direction at the location above ground level The transition intensity index is obtained by weighting and summing the temperature change rate, relative humidity change rate, and power ramp rate of the meter. When the turning point intensity index exceeds the preset event trigger threshold, it is determined to be an extreme turning point event trigger point; Centered on the extreme and turning point of the event, an event window is formed by extending a preset number of time steps forward and backward by m. The times within the event window are marked as extreme and turning point event periods, and the times outside the event window are marked as regular periods, and event labels corresponding to the target times are generated.

[0111] During extreme and transitional weather events, meteorological elements such as wind speed, wind direction, and air pressure often undergo sudden changes in a short period of time. As a result, offshore wind turbines experience switching between rated operation, power limitation, and shutdown, leading to a rapid rise, sharp drop, and increased fluctuations in the power sequence.

[0112] If a single prediction strategy is adopted, the segmented prediction model is easily dominated by conventional samples and it is difficult to accurately characterize the changes in the output mechanism at extreme and turning points.

[0113] After completing the NWP meteorological element bias correction in step S21, a more reliable corrected NWP meteorological element vector is obtained. However, the correction itself cannot solve the problem of which strategy to adopt in the segmented prediction model of this invention. Under normal weather conditions, the power changes gradually, while under extreme and transitional weather conditions, the power rises or falls rapidly, and the output mechanisms of the two are completely different.

[0114] If a single model is used to cover all samples, the majority of the samples will dominate the model training, leading to inaccurate predictions at critical extremes and turning points.

[0115] After obtaining the corrected NWP weather vector Subsequently, a method for identifying and labeling extreme and turning events is further constructed to generate an event label sequence corresponding to the target time. and event window collection This is used for strategy selection and sample enhancement in subsequent segmented prediction models.

[0116] Construct a physics-statistics hybrid driven transition intensity index By integrating abrupt changes in meteorological elements (rates of change in wind speed, wind direction, and humidity) with the power output response of offshore wind farms (power ramp-up rate), the extreme and dramatic nature of weather processes can be quantitatively characterized. Exceeding the event trigger threshold At that time, extreme and turning events are automatically marked, and an event window is generated by expanding forward and backward to form event labels corresponding to the target time. This serves as a gating switch for subsequent segmented prediction models; where 0 represents the normal period and 1 represents the period of extreme and turning events.

[0117] For any given moment The changes in each meteorological element at adjacent time points are calculated to obtain the abrupt change rate (first difference, changes at adjacent time points) characteristics of the NWP meteorological elements, as shown below: Formula (26); in, The abrupt change rate of the corrected NWP meteorological elements at time t; The corrected NWP meteorological element values ​​were obtained at the previous time.

[0118] The abrupt change rate of meteorological elements directly reflects the speed of change in weather processes. The essential characteristic of extreme and transitional weather is the rapid and significant directional abrupt change of meteorological elements within a short period, and this abrupt change is primarily reflected in the magnitude of the change in each meteorological element rather than its absolute value. For example, during the passage of a cold front, wind speed may jump from 5 m / s to 15 m / s within one hour. (positive), while the temperature dropped sharply ( (Negative numbers).

[0119] Wind direction is an angular measure, ranging from 0° to 360°, and has a circular periodicity.

[0120] Considering wind direction as a circular variable, to avoid the discontinuity between 360° and 1°, a wind direction difference operator is defined as follows: Formula (27); in, Let be the wind direction angle at time t; Indicates the wind direction angle at the previous moment. This indicates the change in wind direction between adjacent moments; An angle-around mapping operator is used to map wind direction difference to... or Within the range.

[0121] The wind farm power ramp rate is calculated as follows: Formula (28); in, The wind farm power ramp rate is used to characterize the intensity of changes in wind power output on a short time scale. It is an important quantity for identifying ramp and drop processes during dispatch and operation. This represents the measured active power of the wind farm at time t. This represents the measured active power of the wind farm at the previous time step.

[0122] Power ramp-up rate directly measures the intensity of changes in wind power output over a short timescale. From the perspective of grid dispatch, power ramp-up and sudden drops are the most critical concerns under extreme and transitional weather conditions. ① Climbing a slope means a rapid increase in wind power output, which may lead to wind curtailment; ②A sudden drop means a rapid decrease in output, requiring the rapid activation of backup capacity.

[0123] Therefore, incorporating the power ramp rate into the transition intensity index ensures that the identification of extreme and transitional events is not only based on meteorological conditions, but also directly coupled with changes in wind power output that are of concern to grid dispatch.

[0124] To comprehensively characterize the synergistic features of meteorological abrupt changes and power output response, this invention constructs a transition intensity index. Considering the differences in the dimensions of different characteristics, it first considers the wind turbine hub height. The rate of change in wind speed and the amount of change in wind direction at the location above ground level The temperature mutation rate, relative humidity mutation rate, and power ramp rate of the meters were standardized.

[0125] Wind turbine hub height Wind speed change rate (m / s), wind direction change (°), above ground The temperature abrupt change rate (°C), relative humidity abrupt change rate (%), and power ramp rate (MW / min) have completely different physical dimensions and numerical scales. For example, the change in wind direction may be between -180° and 180°, while the power ramp rate may be... The magnitude changes by MW. If these variables with different dimensions are directly added together, the larger variable will dominate the smaller variable, causing the exponent to lose physical equilibrium.

[0126] Regarding the height of the wind turbine hub The rate of change in wind speed and the amount of change in wind direction at the location above ground level The temperature mutation rate, relative humidity mutation rate, and power ramp rate of rice were standardized and converted into a standard normal distribution with a mean of 0 and a standard deviation of 1, as shown below: Formula (29); in, For variables The mean on the power-meteorology joint dataset, The standard deviation is given; the standardized variable is dimensionless.

[0127] Based on the standardized wind turbine hub height The rate of change in wind speed and the amount of change in wind direction at the location above ground level The transition intensity index is calculated based on the temperature abruptness rate, relative humidity abruptness rate, and power ramp rate of meters, as shown below:

[0128] Formula (30); in, The transition intensity index at time t; the larger the value, the more drastic the transition. Indicates the standardized height of the wind turbine hub. The rate of change in wind speed and the amount of change in wind direction at the location above ground level Temperature abrupt change rate, relative humidity abrupt change rate, and power ramp rate at a distance of meters; These are the standardized wind turbine hub heights. The rate of change in wind speed and the amount of change in wind direction at the location above ground level Weighting coefficients for temperature abrupt change rate, relative humidity abrupt change rate, and power ramp rate at a distance of meters; Standardizing the hub height of the wind turbine The rate of change in wind speed at the location; This is the standardized rate of change of wind direction; For the standardized above-ground Temperature abrupt change rate at a distance of meters; The standardized relative humidity abrupt change rate; This refers to the standardized power ramp rate; The weighting coefficients can be determined using the equal weighting method, the expert experience weighting method, or the normalized weighting method based on the correlation of historical samples.

[0129] Weighting coefficient The conditions to be met are as follows: Formula (31); For example, the equal weight method, taking ; The expert experience-based weighting method, where meteorological / wind power experts assign values ​​based on their domain knowledge, can improve the sensitivity of offshore wind power output to wind speed and power ramp-up. and The value of ; in order to ensure the objectivity of the prediction, the expert experience weighting method should be avoided as much as possible; Based on the normalized weighting method of historical sample correlation, the correlation (such as correlation coefficient) between each variable and power ramp-up and sudden drop is calculated, and values ​​are assigned according to the correlation ratio.

[0130] When the turning point intensity index Exceeding the preset event trigger threshold At that time, The event fragment corresponding to the target time is identified as an extreme or turning point trigger point.

[0131] For each target moment in the historical time period of the power-meteorology joint dataset, a turning point intensity index is calculated based on the corresponding wind speed change rate, wind direction change rate, temperature change rate, humidity change rate, and power ramp rate.

[0132] The calculated inflection point intensity index values ​​are arranged in chronological order according to the target time, forming an array. As shown below: Formula (32); in, Given the total number of historical moments, if the sampling rate is 15 minutes, then there are approximately 70,000 sample points over 2 years. These are the inflection intensity coefficients for the corresponding target time.

[0133] Event trigger threshold It can be derived from the time series of the turning point intensity index. The quantiles are determined as follows: Formula (33); in, This represents the quantile function, where β is the preset quantile level. For example, β can take values ​​of 0.90, 0.9, or 0.98. Take the time series of the turning point intensity index The front with a high turning point intensity index The moment is used as the trigger point for extreme and turning point events; for example, when β=0.95, it means that the moment with the higher index in the time series of turning point intensity index is determined as the trigger point for extreme and turning point events.

[0134] ① More sensitive, marking more extreme and turning point samples, suitable for sea areas with frequent extreme and turning points; ② It is more stringent, only marking the most intense extreme and turning events, and is suitable for regular and calm sea areas; ③ Recommended value, To balance sensitivity and specificity.

[0135] Centered on the trigger points of extreme and turning events, expand forward and backward by a preset number of [numbers]. The time step is used to form an event window; the event window is then expanded into a time-domain augmentation.

[0136] Extreme and turning point events are not isolated points in time, but rather a continuous evolutionary process. Before the trigger point of an extreme or turning point event, meteorological elements begin to change (precursors); after the trigger point, power continues to respond (aftereffects). This invention takes the trigger point of extreme and turning point events as its center and expands forward and backward. Each time step forms a complete event window.

[0137] To cover the evolution process before and after extreme and turning point events, this invention expands the event window for the trigger points of extreme and turning point events: like Then the interval Record these as extreme and turning point events in the window; in, The number of steps to expand the window indicates the number of steps to expand forward and backward from the event trigger point. At a time step, for example, when When you want to cover the period of one hour before and after the trigger point, you can take minutes. Total window length (Including the trigger points of extreme and turning events themselves). Hour.

[0138] The reason why window expansion is needed is because: ① Before extreme and turning point events are triggered, segmented prediction models need to prepare for (predict) the upcoming power changes. ② Even after extreme and turning point events are triggered, the power ramp-up and descent are still in progress (process integrity). ③ The expanded window provides sufficient "turning point context" samples for subsequent segmented prediction models.

[0139] and the expanded window Incorporate into the transition event window collection As shown below: Formula (34); Based on window collections Define the event label sequence As shown below: Formula (35); in, For event label sequences; Indicates the regular time period. It indicates a period of extreme or turning events.

[0140] To improve the learning sufficiency of the segmented prediction model for key periods with few samples, this invention will use event labels. As a conditional input to the subsequent segmented prediction model, and during the training phase... Samples with a value of 1 are given higher weights to enhance the fitting ability and prediction stability of the segmented prediction model for extreme and transitional window features of climbing and sudden drops.

[0141] Step S2 generates an event label sequence by constructing a transition intensity index and identifying and expanding extreme and transition event windows. This provides conditional input and sample weighting basis for the subsequent segmented prediction model, enhancing the segmented prediction model's ability to characterize the uphill and sudden drop processes under extreme and transition weather and the stability of offshore wind farm output prediction.

[0142] Step S3 includes steps S31-S32.

[0143] Step S31, online prediction stage: Based on the corrected NWP meteorological elements within the prediction window, generate prediction event labels for each target prediction time. Based on the event labels, call the corresponding pre-trained segmented prediction models for normal periods and extreme and turning point event periods to obtain the future multi-step power output prediction values ​​of offshore wind farms.

[0144] In offshore wind power forecasting, there are significant differences in the output mechanism corresponding to normal weather and extreme and transitional weather: during normal periods, the power mainly changes slowly with wind speed and is affected by the difference between the wake and the turbine, while during extreme and transitional periods, the power is often accompanied by rapid changes in wind speed and wind direction, as well as the switching of turbine power limitation, shutdown and other states, resulting in a strong non-stationary characteristic of the power sequence.

[0145] If a single model is used to train all samples uniformly, the model is easily dominated by regular samples, resulting in amplified errors and misaligned climbing at extremes and turning points.

[0146] The core problem addressed in this step is: after completing the correction of meteorological element deviations and the identification of turning events, how to construct a segmented forecasting model that can "sense weather conditions and adaptively switch strategies" so that it maintains high accuracy during normal periods and robustness during extreme and turning periods.

[0147] Current technologies typically train a single model to cover all weather types, but this "one-size-fits-all" modeling approach has a fundamental flaw: ① The sample is unbalanced, with extreme and transitional weather samples accounting for very small proportions in historical data (usually <5%), and the single model is dominated by the conventional sample. ② The mechanisms are mismatched. Under normal weather conditions, the power changes slowly (gradually with wind speed), while under changing weather conditions, the power fluctuates drastically (power limiting / cut-off / climbing). The mapping relationship between the two is completely different. ③ Inaccurate extreme and turning point windows: The model tends to output a smoothed average value in extreme and turning point windows, causing temporal misalignment and amplitude deviation during ramp-up and sudden drop.

[0148] The core innovative idea behind this step is "divide and conquer," based on event tags. The training samples are divided into a regular subset and extreme and transition subsets to construct an event-driven power segmentation prediction framework: at power prediction time t, for each prediction step... (e.g., the 1st hour, 2nd hour... Hth hour in the future), use the corresponding valid future time. Corrected NWP weather vector And its derived features, combined with historical power lag features, to construct step-size related feature vectors Two independent LightGBM models were trained: a regular model and a standard model. and transition model From the conventional model With the transition model It outputs a multi-step power prediction sequence for the future, thereby improving the forecast robustness and stability during extreme and transitional weather periods.

[0149] Obtain NWP numerical weather forecast data for the target time, and use the trained meteorological element bias correction model to correct the bias, thus obtaining the first corrected NWP meteorological elements for the forecast. By applying range constraints to the first revised NWP meteorological elements, the second revised NWP meteorological elements are obtained. The predicted second-corrected NWP meteorological elements are integrated into the preprocessed numerical weather forecast data at the corresponding time to obtain the predicted corrected NWP meteorological elements; and event labels for each target time are generated.

[0150] Based on the event labels corresponding to each prediction target time, the pre-trained segmented prediction models are invoked for normal periods and extreme and transitional event periods to obtain multi-step power output predictions for offshore wind farms, including: For each prediction step, based on the event label corresponding to the prediction target time, perform the following steps: If the event label corresponding to the predicted target time is a regular time period, then the regular model is called, and the power prediction value of the predicted step size is output with the step size-related feature vector corresponding to the predicted target time as input. If the event label corresponding to the predicted target time is an extreme and turning point event period, then the extreme and turning point model is invoked, and the power prediction value of the predicted step size is output with the step size-related feature vector corresponding to the predicted target time as input. Based on the power prediction values ​​of each prediction step combined with the target time axis, the future multi-step power output prediction values ​​of the offshore wind farm are obtained.

[0151] The pre-trained segmented prediction model is obtained through the following process: Based on the preprocessed historical measured power data of the offshore wind farm, the historical power lag vector and the corresponding time feature vector are obtained. Based on the corrected wind turbine hub height in the corrected NWP meteorological elements Wind speed at the location, above ground The temperature and relative humidity of the rice were used to obtain the corrected NWP meteorological element vector; Based on the corrected NWP meteorological element vector, the mutation rate vector of the corrected NWP meteorological element vector is obtained. Based on the sliding window length, the mean and standard deviation of power at historical moments within the sliding window, and the mean and standard deviation of wind speed at historical moments within the sliding window, a sliding window statistical feature vector is constructed. Based on the historical power lag vector, the corrected NWP meteorological element vector, the mutation rate vector of the corrected NWP meteorological element vector, the sliding window statistical feature vector, and the time feature vector, the step size related feature vector is calculated. Based on the corresponding event labels, the step-size related feature vectors are used as training sample data to be divided into a regular training subset and an extreme and turning point training subset. The corresponding segmented prediction models are trained using the regular training subset and the extreme and transition training subsets, respectively; wherein, the segmented prediction models include independent regular models and extreme and transition models; The process continues until the regression loss function converges, resulting in a well-trained piecewise prediction model.

[0152] During online prediction, the event labels corresponding to the target time are used. Automatically switch to the corresponding model to achieve an adaptive power segmentation prediction mechanism that uses the conventional model during normal periods and the extreme and transitional model during extreme and transitional periods.

[0153] First, the target timeline resolution is The set of prediction step sizes is as follows: Formula (36); in, This represents the maximum number of prediction steps. Determined by the needs of the forecasting task, such as the current forecast. , Hour.

[0154] For online prediction of target time The future multi-step wind power forecast targets are as follows: Formula (37); in, For online prediction of target time The output future multi-step power prediction sequence includes A column vector of predicted values; Indicates the future number The predicted wind power value corresponding to each prediction step size. ,when This is the first step in prediction; This is the final step in the prediction process; The transpose symbol indicates that the sequence is a column vector.

[0155] For example, Recent segmented online predictions: When =1 hour =24 indicates the predicted hourly wind power output for the next 24 hours; Intraday ultra-short-term segmented online prediction: When =15 minutes =16 indicates the predicted wind power output for the next 4 hours, calculated in 15-minute increments.

[0156] Multi-step segmented online forecasting provides grid dispatch with a complete view of the "future power change trajectory," rather than a single moment, making it easier for dispatchers to formulate generation plans, reserve capacity configurations, and safety checks.

[0157] For each segment, the online prediction step size The present invention provides an online prediction step size for each segment. Construct the corresponding step-size related feature vector independently As shown below: Formula (38); in, This is the historical power lag vector, including the current time and past times. The measured power at each moment; For the future The corrected NWP meteorological element vector for each predicted target time; This represents the corrected changes in NWP meteorological elements between adjacent forecast steps; For the sliding window statistical feature vector, The time period characteristics of the target prediction time.

[0158] ①Historical power lag vector As shown below: Formula (39); in, This represents the historical power lag step; For the current moment The measured active power comes from the SCADA system of the wind power control center; The power sequence is the measured active power at the previous moment, with a time lag of 1 step. For the first Measured active power at each moment, time lag of the power sequence step.

[0159] The value is determined based on the temporal resolution of the prediction task: when the target time axis resolution... =15 minutes, The value ranges from 16 to 24, corresponding to the historical power sequence of the past 4-6 hours; when =1 hour, The value ranges from 12 to 24, corresponding to the historical power sequence of the past 12-24 hours.

[0160] With prediction step size Relationship: Can be with At the same level, or slightly larger ,For example: Very short-term forecast: ; Recent forecast: .

[0161] Offshore wind power exhibits strong autocorrelation, meaning the current active power is highly correlated with the power at several past moments. Introducing a historical power lag vector allows the model to "remember" the trajectory of recent power changes, providing a temporal context for prediction. For example, if power has been rising continuously over the past hour, it is highly likely that power will continue to rise in the next moment (inertia).

[0162] ② Corrected NWP meteorological element vector As shown below: Formula (40); in, For the future Step-corrected wind turbine hub height The wind speed at the location is the most crucial input for power prediction; For the future The corrected ground level The temperature at a height of meters affects the air density, which in turn affects the output of the fan; For the future The corrected relative humidity affects air density and helps to characterize weather process attributes.

[0163] This directly correlates the corrected NWP meteorological elements with the target forecast time. Traditional methods use NWP as an "input feature of the current moment" to predict multiple future steps, essentially using current weather forecasts to predict future wind power. The key difference in this step is that, for the future... The prediction of the step, using NWP for the effective moment of the future. The weather forecast values ​​are used as input features. For example, when predicting wind power at 14:00 tomorrow, the wind speed forecast for 14:00 tomorrow is directly used as input by the NWP, instead of using the wind speed at the current time to extrapolate. This method avoids the accumulation of errors in the extrapolation prediction.

[0164] ③ The mutation rate vector of the corrected NWP meteorological element vector This characterizes the trend of weather changes at adjacent forecast times, as shown below: Formula (41); in, For the first Step-corrected meteorological vector; For the first Step-by-step correction of meteorological vectors.

[0165] The changing trends of meteorological elements (whether wind speed is increasing or decreasing, and how quickly it is increasing) are a better indicator of future power changes than their absolute values. For example, a wind speed increasing from 12 m / s to 15 m / s (positive abrupt change rate) and a wind speed stabilizing at 12 m / s (zero abrupt change rate) indicate that power is climbing, while the latter indicates that power is stabilizing. Introducing abrupt change rate feature allows models to perceive the direction and speed of meteorological changes, thereby improving their ability to predict power increases or decreases.

[0166] ④ Sliding window statistical feature vector As shown below: Formula (42); in, The length of the sliding window; For the past The average power at each moment represents the recent average power level; For the past The standard deviation of power at a given moment indicates the degree of recent power fluctuation. For the past The average wind speed at any given moment; For the past The standard deviation of wind speed at any given moment.

[0167] For example, A value of 3 indicates the past 3 time steps.

[0168] ⑤ Time feature vector As shown below: Formula (43); in, The target prediction time is set to the hour of the day, 0-23, capturing diurnal variations such as day / night wind power differences; The target prediction time is set to the month, 1-12, capturing seasonal variations such as the difference between the typhoon season and the winter monsoon.

[0169] Wind power exhibits a clear periodic pattern. The output characteristics of the same wind farm show systematic differences at different times of day (e.g., daytime vs. nighttime) and in different seasons of the year (e.g., typhoon season vs. winter). Introducing time characteristics allows the model to utilize this periodic prior knowledge, thereby improving prediction accuracy.

[0170] Training sample segmentation: Based on the corresponding event tag The training samples are divided into two mutually exclusive subsets: the normal training subset and the extreme and transition training subset; Key timing logic: Predicting the target time For the target time Whether a prediction falls within a turning point window depends on the target time of the prediction. The event label, rather than the predicted target time. The event label. This is because what is being predicted is the power at a future time, while the weather conditions at that time are determined by the current weather conditions. Decide.

[0171] The standard training subset and the extreme and transition training subsets are shown below: Formula (44); in, Indicates the first A typical sample set under a target prediction step size. This represents the set of extreme and turning point samples at the h-th target prediction step.

[0172] For each online prediction step The sample set is divided independently. This means that at the same prediction time... Possibly at different step sizes They were divided into different sets because Follow Changes. For example, the predicted time. The target is the next 4 hours, with the first 2 hours likely to be the regular period. The last two hours may be the turning point. Therefore, the sample enters the regular set in the first two steps. and In the last two steps, we enter the turning point set. and .

[0173] This method of independently partitioning the target time makes conventional models and transition model Each is trained on its own dedicated samples, focusing on learning the mapping relationship under the corresponding weather mechanism.

[0174] Offline training and online prediction of segmented prediction models: For each prediction step Train the conventional model separately and extreme and turning point models The differences between the two independent models, the conventional model and the extreme and inflection models, are shown in Table 1.

[0175] Table 1 - Differences between the conventional model and the extreme and inflection models:

[0176] During offline training, As input samples, the measured power of offshore wind power is used. To supervise the sample labels, the optimal parameters of the piecewise prediction model are obtained by minimizing the error between the piecewise prediction model output and the measured power. During the online prediction phase, the power prediction value of the corresponding model output is selected based on the event label at the target prediction time, as shown below: Formula (45); in, The final output is the predicted value of offshore wind power, which is the prediction result after selecting the segmented prediction model; For the first The conventional model corresponding to each prediction step size; This is the output of the standard model, used when the target time is within a standard time period; No. Extreme and inflection models corresponding to each prediction step size; This is the output of the extreme and transition models, used when the target time is an extreme or transitional period.

[0177] For example, the segmented prediction model can be implemented using LightGBM, XGBoost, random forest, neural network, or a combination thereof.

[0178] In this invention, the conventional model and the extreme and transition models are respectively constructed based on the LightGBM model.

[0179] For any prediction step size The standard model and the extreme and transition models are trained on their respective sample sets by minimizing the regression loss. The regression loss function is shown below: Formula (46); in, and The first The optimal parameters for the conventional model and the extreme and transition models under each prediction step size are the optimal parameters for the piecewise prediction model after training and convergence. and These are the parameters to be optimized for the corresponding conventional model and the extreme and inflection models, respectively; and The first The normal sample set and the extreme and inflection sample sets under each prediction step size; The loss function can be the squared error, absolute error, or Huber loss function; is the measured active power of the wind farm at the target prediction time, and is the target value for training the piecewise prediction model; For the first The input feature vector corresponding to each prediction step size; For the conventional model in parameters The output below; For extreme and transition models in parameters The output below.

[0180] Formula (46) indicates that the conventional model is trained offline on the conventional sample set to learn the mapping relationship between meteorological input and wind power under normal weather conditions; the extreme and turning point model is trained on the extreme and turning point sample set to learn the non-stationary change characteristics of wind power such as rapid rise and sudden drop within the extreme and turning point weather window, thereby improving the prediction stability and robustness of the segmented prediction model under key weather processes.

[0181] In the preferred embodiment, LightGBM is used as the base model, which has the following advantages: ① High efficiency: Histogram-based decision tree algorithm with fast training speed and low memory usage; ② It excels at nonlinearity and is suitable for handling complex nonlinear relationships between meteorological elements and power; ③ The importance of features is interpretable, and the contribution of each feature to the prediction can be output, which facilitates model analysis and debugging; ④ It has good robustness and is not sensitive to missing or outlier values, making it suitable for real-world engineering data.

[0182] Alternatives include XGBoost, Random Forest, Neural Networks, or combinations thereof, but LightGBM offers better training efficiency and lower memory usage.

[0183] During the offline training of the segmented prediction model, weights are assigned to the training samples based on the event labels and the power ramp rate: Based on the event labels, the training samples during extreme and turning point event periods are assigned a first weight enhancement coefficient; A second weighting enhancement coefficient is assigned to training samples whose absolute value of the power ramp rate exceeds a preset ramp threshold; Based on the first weight enhancement coefficient and the second weight enhancement coefficient, the sample weight of each training sample is determined; The regression loss of the segmented prediction model is weighted using sample weights.

[0184] In extreme and transitional weather conditions, power grid dispatching is most concerned with the rapid ramp-up and sudden drop processes of wind power. If the segmented prediction model only uses the average error of the entire sample as the optimization objective, it is easy to have good overall performance but large deviations in key windows. To improve the learning ability of the segmented prediction model for key ramp-up and sudden drop processes under extreme and transitional weather conditions, and to avoid prediction results exceeding the physical feasibility range of wind farms, this invention introduces a ramp-sensitive sample weighted training mechanism on the basis of segmented prediction modeling, and imposes non-negative constraints and an upper limit of installed capacity on the output power prediction results of the segmented prediction model.

[0185] Based on power ramp rate With event tags Define weights for each training sample This allows the turning point event window and the large climbing samples to contribute more significantly during training. The preferred weighting form is as follows: Formula (47); in, For a moment The weights corresponding to the training samples; This is an event label indicator function; it takes the value 1 when the event is a transitional event, and 0 otherwise. Power ramp rate; The slope rate threshold; Let be the power ramp rate indication function, if If true, the value is 1; otherwise, it is 0. , are hyperparameters, and are the weight enhancement coefficients for the turning point event samples and the large slope samples, respectively, used to increase the contribution of the turning point event samples and the large slope samples to the training loss. For example, Available training samples The 95th percentile, , The settings can be adjusted based on the prediction performance on the validation set.

[0186] For example, , If the value is too large, the model may overfit to a few extreme samples; if the value is too small, the weighting effect will be insignificant.

[0187] Sample weights By introducing the training objective and regression loss function, for any prediction step size h, we obtain the weighted loss functions for the conventional model and the extreme and inflection models.

[0188] The weighted regression loss function is as follows:

[0189] Formula (48); in, and The first Optimal parameters for the conventional model and the extreme and inflection models under each step size; and These are the parameters to be optimized for the conventional model and the extreme and inflection models, respectively. and The first The set of regular samples and the set of extreme and turning points under each step size; For the target time Sample weights; The loss function; The measured active power of the wind farm at the target prediction time; For the first The input feature vector corresponding to each step size; For the conventional model In parameters The output below; For extreme and turning point models In parameters The output below.

[0190] In weighted training, the loss for each sample is weighted by its weights. Zoom in or out. When a sample is at an extreme, a turning point, or experiencing a large climb, its weight is >1, meaning that the prediction error of that sample accounts for a higher proportion of the total loss. During model optimization, the error of this type of sample will be reduced more preferentially.

[0191] Through the weighted training described above, the segmented prediction model pays more attention to turning point event windows and samples with rapid climbs and sudden drops during the optimization process, thereby improving the prediction stability of key processes.

[0192] Training continues until the weighted regression loss function converges, at which point the training ends and a pre-trained piecewise prediction model is obtained.

[0193] Step S32: During the online prediction process, physical boundary constraints are applied to the predicted output values ​​of the future multi-step wind farm to obtain the final predicted output results of the offshore wind farm.

[0194] Physical boundary constraints are applied to the predicted future multi-step power output values ​​to obtain the final power output prediction results for the offshore wind farm, including: The power prediction value of each prediction step in the future multi-step power output prediction value of the offshore wind farm is compared with 0, and the maximum value of the two is taken to obtain the first constraint result. The first constraint result is compared with the upper limit of the installed capacity of offshore wind farms, and the minimum value of the two is taken to obtain the final power prediction value for the prediction step. Based on the final power prediction values ​​of each prediction step combined with the target time axis, the final offshore wind farm output prediction result is obtained.

[0195] To ensure that the predicted power meets the basic physical boundaries of offshore wind farms, the power prediction value output by the piecewise prediction model of this invention... Apply nonnegativity and upper limit constraints on the installed capacity, as follows: Formula (49); in, The output of the segmented prediction model is the predicted value of the future multi-step output. This represents the final power prediction value for the offshore wind farm at the h-th prediction step in the future. This refers to the installed capacity of an offshore wind farm or the upper limit of available capacity at the predicted time. This means taking the larger of 0 and the original predicted value to ensure that the power value is ≥ 0; Take the capacity of the machine and The smaller of the results shall be used to ensure that the power is less than or equal to the installed capacity of the wind turbine.

[0196] In actual business scenarios, wind turbine installed capacity It is not necessarily a fixed constant; it can be dynamically adjusted according to the situation in Table 2.

[0197] Table 2 - Examples of Rated Installed Capacity Adjustments:

[0198] The current available capacity is taken as the upper limit, which is determined in real time by the operating status of the wind farm.

[0199] The purpose of this step is to limit the power prediction results to [0, P]. cap Within the specified range, avoid unreasonable predictions that result in negative power or exceed the installed capacity.

[0200] By assigning higher training weights to extreme and turning point event windows and large ramp samples, and by imposing non-negative constraints and upper limits on installed capacity on the model output, the segmented prediction model's ability to predict key ramp and descent processes is improved, while ensuring that the offshore wind power prediction results meet the physical feasibility requirements.

[0201] Step S3 serves to drive the segmented model to make adaptive predictions based on event labels, and through slope-sensitive weighted training and hard constraints on physical boundaries, to achieve accurate prediction of future multi-step output of offshore wind power under extreme and transitional weather conditions and ensure engineering availability.

[0202] The complete set of offline trained models and data processing procedures built in steps S1 to S3 are deployed to the actual wind farm business system to achieve online, real-time, continuous and stable multi-step power prediction.

[0203] In real-world power grid dispatching scenarios, there are real-world constraints as shown in Table 3.

[0204] Table 3 - Examples of constraints in real power grid dispatching business scenarios:

[0205] The complete chain of methods formed by steps S1 to S3 is deployed as an automated online rolling forecasting process for offshore wind farm output.

[0206] Because NWP is continuously updated on a rolling basis, wind farm power forecasts need to be updated at each operational time. Based on the latest available power observation and NWP numerical weather forecast data, the system outputs real-time multi-step offshore wind farm output prediction sequences to support intraday scheduling and safety verification.

[0207] This invention is an online rolling forecasting process. One of the core features of online rolling forecasting is "rolling," meaning that at each running time... Based on the latest available measured power data for offshore wind power, predictions are made regarding the future development of offshore wind farms. A power prediction sequence with a step size; firstly, the latest rolling NWP data is spliced ​​together according to the shortest forecast lead time priority strategy to obtain the effective future time. The weather forecast input; the forecast window slides backward as a whole.

[0208] The following is a specific example to illustrate this: Business scenario setting: Forecasting task: Ultra-short-term power forecast for the next 4 hours; Event resolution: minute; Prediction steps: This means predicting 16 points in the future, totaling 4 hours; NWP numerical weather forecast data update frequency: released every 6 hours; Predicted service operation frequency: Wind power prediction is triggered every 15 minutes.

[0209] A rolling forecast timeline illustration, as shown below. Figure 2 As shown.

[0210] ① Within the same NWP numerical weather forecast data reporting period (e.g., NWP data starting at 8:00 covers 8:00-14:00), multiple forecasts will be triggered (8:15, 8:30, 8:45, ...), each using different valid times of the same batch of NWP data; ② The output of each prediction is a sequence of lengths... A sequence (e.g., 16 time points), rather than a single point; ③ When new NWP numerical weather forecast data arrives (e.g., at 12:00), switch to the latest NWP numerical weather forecast data source.

[0211] Triggering process for online rolling prediction: At each running time The following processes are executed sequentially to form a complete prediction link for the future multi-step power output sequence of an offshore wind farm.

[0212] Latest data retrieval and preprocessing: Get: ① Latest SCADA measured power data: ; ② The latest NWP numerical weather prediction rolling forecast data: including meteorological element forecasts with multiple start times and multiple forecast lead times; ③ Current installed capacity of wind turbines in the wind farm: .

[0213] Read the current time from the SCADA system of the wind power control center. Latest power observations; Pull NWP numerical weather forecast files from the NWP numerical weather forecast data server and identify all available start times and forecast lead times; Obtain the current available capacity from the wind farm's energy management system (such as whether there are turbine maintenance or power limiting instructions).

[0214] The system obtains the original power sequence, the latest NWP numerical weather forecast data, and the current available capacity. Among these, the ERA5 or ERA5T reanalysis data are mainly used for offline training, periodic updates, or retraining of meteorological element bias correction models, and are not used as necessary inputs for each online rolling forecast.

[0215] The differences between the online prediction phase and the offline model training phase are shown in Table 4.

[0216] Table 4 - Differences between the online prediction phase and the offline model training phase:

[0217] For each target effective time

[0218] From the latest NWP numerical weather prediction data, identify all candidate records that have forecasts for time t: Select the best record according to the principle of prioritizing the shortest forecast lead time; If no NWP numerical weather forecast data is available at a certain time t, persistent completion will be used; Output and Prediction Window Aligned NWP input sequence .

[0219] NWP meteorological element bias correction, resulting in forecast window Aligned and corrected NWP meteorological element vectors .

[0220] The subsequently invoked meteorological element deviation correction model yields the corrected NWP meteorological element vector. ; Calculation of Turning Point Intensity Index and Generation of Event Tags: The inflection point intensity coefficient is calculated based on the corrected NWP meteorological elements and their variation characteristics corresponding to the future target prediction time, and event labels for the target prediction time are generated. ; Online prediction using a segmented prediction model: Finally, based on the event label corresponding to the target prediction time, the corresponding segmented prediction model is selected to output the predicted values ​​for each step. .

[0221] Physical boundary constraint clipping: The final offshore wind farm power output prediction results are obtained by applying non-negative constraints and the upper limit of the installed capacity of the offshore wind farm. This yields the final multi-step wind power prediction sequence that satisfies the physical boundary constraints, thus forming a multi-step prediction sequence that can be directly used for engineering applications.

[0222] Update strategies for meteorological element bias correction models and piecewise prediction models: Online meteorological element bias correction models and segmented prediction models need to be updated regularly to adapt to drift in meteorological models, wind farm operating characteristics, and the NWP numerical weather prediction data itself. Three training strategies are employed simultaneously: regular full retraining, incremental learning, and performance monitoring-triggered updates, balancing model accuracy and computational efficiency.

[0223] (1) Regular full-weight training, recommended: Retraining cycles, once a month or quarter; Training data, accumulated up to the latest data at the current moment, such as the past 2 years; Triggering method: Timed triggering; Advantages: The model always reflects the latest data releases; Disadvantages: High computational cost, requiring a long retraining time.

[0224] (2) Incremental learning and updating: Update frequency: Daily or weekly; Update method: Incremental training is performed using newly arrived data based on the existing model; Triggering method: Timed triggering or data accumulation triggering; Advantages: Low computational cost and high update frequency; Disadvantage: It may gradually deviate from the optimal state.

[0225] (3) Performance monitoring-triggered updates, adaptive: Monitoring metrics: Rolling forecast error (e.g., MAE, RMSE); Triggering condition: The prediction error exceeds a preset threshold for a number of consecutive times; Triggering method: Event-driven; Advantages: Updates only when necessary, resulting in the highest computational efficiency; Disadvantage: A reasonable performance benchmark needs to be set.

[0226] Example 2: A specific embodiment of the present invention discloses an offshore wind farm output prediction system under extreme and transitional weather conditions, thereby implementing the offshore wind farm output prediction method under extreme and transitional weather conditions described in Embodiment 1. The specific implementation methods of each module are as described in the corresponding descriptions in Embodiment 1.

[0227] like Figure 3 As shown, the system includes a data acquisition and preprocessing module (M1), an event recognition and label generation module (M2), and a segmented prediction and physical constraint module (M3). The data acquisition and preprocessing module (M1) is used to acquire multi-source data and preprocess it to obtain a power-meteorological joint dataset. The power-meteorological joint dataset includes reanalysis meteorological data aligned to the target time axis, as well as preprocessed historical measured power data and numerical weather prediction data of offshore wind farms. The event identification and tag generation module (M2) is used to correct the bias of the preprocessed numerical weather forecast data based on the reanalysis meteorological data to obtain the corrected NWP meteorological elements; based on the corrected NWP meteorological elements and the preprocessed historical measured power data of the wind farm, it generates event tags for the corrected NWP meteorological elements; the event tags are used to distinguish between normal periods and extreme and turning point event periods; The segmented prediction and physical constraint module (M3) is used to call the corresponding pre-trained segmented prediction model for the normal period and the extreme and turning point event periods based on the event label to obtain the future multi-step power output prediction value of the offshore wind farm; and to apply physical boundary constraints to the future multi-step power output prediction value to obtain the final power output prediction result of the offshore wind farm.

[0228] Since the system in this embodiment and the method in Embodiment 1 are related and can be referenced from each other, this description is redundant and will not be repeated here. Because this system embodiment shares the same principle as the above method embodiment, it also possesses the corresponding technical effects of the above method embodiment.

[0229] In summary, the method and system for predicting the power output time series of offshore wind farms under extreme weather conditions according to embodiments of the present invention have the following beneficial effects: 1. This invention establishes a meteorological bias correction mechanism for non-stationary errors, suppressing the amplification of input bias at its source. Addressing the shortcomings of existing methods that directly input NWP numerical weather prediction data into the prediction model and lack a unified bias correction framework for extreme weather, leading to the amplification of meteorological errors to power prediction errors, this invention utilizes ERA5 reanalysis data as a high-reliability reference benchmark to construct a condition-feature-driven NWP meteorological element bias correction model. This model extracts multi-dimensional conditional features such as time period, wind direction and flow, meteorological abrupt change rate, and sliding window statistics, enabling the meteorological element correction process to perceive error variation patterns under different meteorological scenarios. It achieves adaptive correction of non-stationary errors in key elements such as wind speed, temperature, and humidity, and is further constrained by both physical and statistical boundaries to ensure the rationality of the meteorological element correction results. Through this mechanism, the systematic bias and error drift of NWP numerical weather prediction under extreme and transitional weather conditions are effectively reduced, decreasing the amplification effect of meteorological element bias to power prediction errors from the data source, thus solving the fundamental problem of "large input bias" in existing methods. 2. This invention constructs a transition intensity index and an event-driven segmented prediction mechanism to address the shortcomings of single models in adapting to extreme and transitional weather. Addressing the deficiencies of existing methods that use a single model to cover all weather types and lack explicit identification and mechanism switching capabilities for extreme and transitional weather processes, leading to significantly amplified prediction errors in transition windows, this invention constructs a transition intensity index based on the mutation rate of meteorological elements. This enables automatic and quantitative identification of extreme and transitional weather events, and forms a continuous sequence of event labels through event window expansion. Based on this, an event-driven segmented direct multi-step prediction architecture is established, modeling regular periods and extreme and transitional periods separately: a regular model and an extreme / transitional model are independently trained for each prediction step. During prediction, the system automatically switches to the corresponding regular model or extreme / transitional model based on the event label at the target time. Simultaneously, a direct multi-step prediction strategy avoids error accumulation in recursive methods. Through the mechanism of "identification first, segmentation then prediction," the segmented prediction model can automatically adopt an adaptation strategy when weather mechanisms undergo abrupt changes, maintaining accuracy during regular periods and capturing abrupt changes during extreme and transitional periods, effectively solving the problem of "insufficient model adaptability" in existing technologies. 3. This invention introduces ramp-sensitive weighting and hard physical boundary constraints to ensure the accuracy and engineering usability of critical window predictions. Addressing the shortcomings of existing methods in the background, which lack targeted constraints on critical ramp and descent windows and lack physical boundary verification of output results, leading to misalignment of critical window amplitude and timing, and unreasonable outputs such as negative power or over-installation, this invention introduces a ramp-sensitive sample weighting mechanism during the segmented prediction model training phase. Training samples in extreme and transitional event windows, and those experiencing high-power ramps, are assigned higher training weights than regular samples. This allows the segmented prediction model to prioritize the rapid power change process most important to grid dispatch during optimization, enhancing its ability to fit the consistency of critical window amplitude and timing. In the prediction output phase, non-negative hard physical boundary constraints and the upper limit of installed capacity are applied, strictly limiting the power value of each prediction step to a physically feasible range. Furthermore, the upper limit of installed capacity can be dynamically adjusted based on actual operating conditions such as wind turbine maintenance and power limiting commands, ensuring the physical feasibility of the offshore wind farm output timing prediction results. By combining a "weighted training + hard constraint output" strategy, we ensure that the power output time series prediction results of offshore wind farms can both "capture" extreme changes and "do not violate" physical laws, directly meeting the engineering application requirements of high-reliability prediction results for grid dispatching operation and safety verification.

[0230] Those skilled in the art will understand that all or part of the processes of the methods described in the above embodiments can be implemented by a computer program instructing related hardware, and the program can be stored in a computer-readable storage medium. The computer-readable storage medium may be a disk, optical disk, read-only memory, or random access memory, etc.

[0231] The above description is only a preferred embodiment of the present invention, but the scope of protection of the present invention is not limited thereto. Any changes or substitutions that can be easily conceived by those skilled in the art within the scope of the technology disclosed in the present invention should be included within the scope of protection of the present invention.

Claims

1. A method for predicting the power output of offshore wind farms under extreme and transitional weather conditions, characterized in that, include: Multi-source data were acquired and preprocessed to obtain a power-meteorological joint dataset; the power-meteorological joint dataset includes reanalysis meteorological data aligned to the target time axis, as well as preprocessed historical measured power data and numerical weather prediction data of offshore wind farms; Using the reanalysis meteorological data as a monitoring benchmark, the preprocessed numerical weather prediction data is biased to obtain the corrected NWP meteorological elements; based on the corrected NWP meteorological elements and the preprocessed historical measured power data of wind farms, event labels corresponding to the target time of the corrected NWP meteorological elements are generated; the event labels are used to distinguish between normal time periods and extreme and turning point event periods, and are used to train the segmented prediction model; Based on the event labels generated for each target time of the NWP meteorological elements within the prediction window, the corresponding pre-trained segmented prediction models are called for the normal period and the periods of extreme and turning events to obtain the future multi-step power output prediction values ​​of the offshore wind farm; physical boundary constraints are applied to the future multi-step power output prediction values ​​to obtain the final power output prediction results of the offshore wind farm.

2. The method for predicting the output of offshore wind farms under extreme and transitional weather conditions according to claim 1, characterized in that, The multi-source data includes historical measured power data of offshore wind farms, numerical weather prediction data, and reanalysis meteorological data. The preprocessing includes: A target time axis is constructed based on the sampling resolution of the historical measured power data of the offshore wind farm, and the historical measured power data of the offshore wind farm is mapped to the target time axis. Physical boundary quality control is performed on the measured power data of the offshore wind farm based on the installed capacity of the offshore wind farm, and power data points that do not meet the physical boundary quality control conditions are marked as abnormal. For continuous abnormal or missing power data segments, differentiated processing is carried out based on the duration of the abnormality to obtain preprocessed historical measured power data of offshore wind farms. For rolling numerical weather forecast data, there are multiple candidate forecast records with different start times corresponding to the same target effective time. The candidate forecast record with the shortest forecast lead time is selected as the numerical weather forecast data for the target effective time, and the preprocessed numerical weather forecast data is obtained. Based on the reanalysis meteorological data, the preprocessed historical measured power data of offshore wind farms, and the preprocessed numerical weather forecast data, a power-meteorological joint dataset aligned to the target time axis is obtained.

3. The method for predicting the output of offshore wind farms under extreme and transitional weather conditions according to claim 2, characterized in that, Using the reanalysis meteorological data as a monitoring benchmark, bias corrections are performed on the preprocessed numerical weather prediction data to obtain the corrected NWP meteorological elements, including: The reanalysis meteorological data at the same effective time of the target on the target time axis are paired with the preprocessed numerical weather forecast data; Based on key meteorological elements in preprocessed numerical weather prediction data, a conditional feature vector is constructed; wherein, the key meteorological elements include at least the wind turbine hub height. Wind speed at the location, above ground The temperature and relative humidity of the rice; Using the original forecast values ​​from the numerical weather prediction of the key meteorological elements and the conditional feature vectors as sample data, and the reanalysis meteorological data at the corresponding time as supervision labels, a meteorological element bias correction model is trained. Using a trained meteorological element bias correction model, the preprocessed numerical weather forecast data is mapped to the first corrected NWP meteorological elements. The first corrected NWP meteorological element is constrained to obtain the second corrected NWP meteorological element; The corrected NWP meteorological elements are then integrated into the preprocessed numerical weather forecast data at the corresponding time to obtain the corrected NWP meteorological elements. The conditional feature vector includes time features, wind direction and flow features, meteorological element mutation rate features, and sliding window statistical features.

4. The method for predicting the output of offshore wind farms under extreme and transitional weather conditions according to claim 1, characterized in that, Based on the corrected NWP meteorological elements and the preprocessed historical measured power data of the wind farm, event tags corresponding to the target time for the corrected NWP meteorological elements are generated, including: Based on the corrected NWP meteorological elements, the changes in each meteorological element at adjacent times are calculated to obtain the abrupt change rate of the NWP meteorological elements; wherein, the abrupt change rate of the NWP meteorological elements includes the wind turbine hub height. Wind speed change rate at the location, above ground Temperature abrupt change rate and relative humidity abrupt change rate of rice; Based on the preprocessed historical measured power data of the wind farm, the change in adjacent time moments is calculated to obtain the power ramp-up rate. Based on the wind direction in the corrected NWP meteorological elements, the change in wind direction at adjacent times is calculated; the wind direction angle at adjacent times is mapped around to map the wind direction difference to a preset angle range. Regarding the height of the wind turbine hub The rate of change in wind speed and the amount of change in wind direction at the location above ground level The temperature mutation rate, relative humidity mutation rate, and power ramp rate of meters were standardized. Standardized wind turbine hub height The rate of change in wind speed and the amount of change in wind direction at the location above ground level The transition intensity index is obtained by weighting and summing the temperature change rate, relative humidity change rate, and power ramp rate of the meter. When the turning point intensity index exceeds the preset event trigger threshold, it is determined to be an extreme turning point event trigger point; Centered on the aforementioned extreme and turning point trigger point, extend forward and backward by a predetermined number of [number] [units]. The time step forms the event window; The times within the event window are marked as extreme and turning point event periods, and the times outside the event window are marked as regular periods, and event labels corresponding to the target times are generated.

5. The method for predicting the output of offshore wind farms under extreme and transitional weather conditions according to claim 4, characterized in that, The pre-trained segmented prediction model is obtained through the following process: Based on the preprocessed historical measured power data of the offshore wind farm, the historical power lag vector and the corresponding time feature vector are obtained. Based on the corrected wind turbine hub height in the corrected NWP meteorological elements Wind speed at the location, above ground The temperature and relative humidity of the rice were used to obtain the corrected NWP meteorological element vector; Based on the corrected NWP meteorological element vector, the mutation rate vector of the corrected NWP meteorological element vector is obtained. Based on the sliding window length, the mean and standard deviation of power at historical moments within the sliding window, and the mean and standard deviation of wind speed at historical moments within the sliding window, a sliding window statistical feature vector is constructed. Based on the historical power lag vector, the corrected NWP meteorological element vector, the mutation rate vector of the corrected NWP meteorological element vector, the sliding window statistical feature vector, and the time feature vector, the step size related feature vector is calculated. Based on the event labels of the corresponding training samples, the step-size related feature vectors are used as training sample data to divide into a regular training subset and an extreme and turning point training subset. The corresponding segmented prediction models are trained using the regular training subset and the extreme and transition training subsets, respectively; wherein, the segmented prediction models include independent regular models and extreme and transition models; The process continues until the regression loss function converges, resulting in a well-trained piecewise prediction model.

6. The method for predicting the output of offshore wind farms under extreme and transitional weather conditions according to claim 5, characterized in that, Based on the event labels for each prediction target time, the corresponding pre-trained segmented prediction models are invoked for normal periods and periods of extreme and turning events to obtain multi-step power output predictions for offshore wind farms, including: For each prediction step, based on the event label corresponding to the prediction target time, perform the following steps: If the event label corresponding to the predicted target time is a regular time period, then the regular model is called, and the power prediction value of the predicted step size is output with the step size-related feature vector corresponding to the predicted target time as input. If the event label corresponding to the predicted target time is an extreme and turning point event period, then the extreme and turning point model is invoked, and the power prediction value of the predicted step size is output with the step size-related feature vector corresponding to the target time as input. Based on the power prediction values ​​of each prediction step combined with the target time axis, the future multi-step power output prediction values ​​of the offshore wind farm are obtained.

7. The method for predicting the output of offshore wind farms under extreme and transitional weather conditions according to claim 6, characterized in that, Physical boundary constraints are applied to the predicted future multi-step power output values ​​to obtain the final power output prediction results for the offshore wind farm, including: The power prediction value of each prediction step in the future multi-step power output prediction value of the offshore wind farm is compared with 0, and the maximum value of the two is taken to obtain the first constraint result. The first constraint result is compared with the upper limit of the installed capacity of offshore wind farms, and the minimum value of the two is taken to obtain the final power prediction value for the prediction step. Based on the final power prediction values ​​of each prediction step combined with the target time axis, the final offshore wind farm output prediction result is obtained.

8. The method for predicting the output of offshore wind farms under extreme and transitional weather conditions according to claim 5, characterized in that, Weights are assigned to the training samples based on the event labels and the power ramp rate: Based on the event labels, the training samples during extreme and turning point event periods are assigned a first weight enhancement coefficient; A second weighting enhancement coefficient is assigned to training samples whose absolute value of the power ramp rate exceeds a preset ramp threshold; Based on the first weight enhancement coefficient and the second weight enhancement coefficient, the sample weight of each training sample is determined; The regression loss of the segmented prediction model is weighted using sample weights.

9. The method for predicting the output of offshore wind farms under extreme and transitional weather conditions according to claim 8, characterized in that, The weighted regression loss function is as follows: in, and The first Optimal parameters for the conventional model and the extreme and inflection models under each step size; and These are the parameters to be optimized for the conventional model and the extreme and inflection models, respectively. and The first The set of regular samples and the set of extreme and turning points under each step size; For the target time Sample weights; The loss function; The measured active power of the wind farm at the target prediction time; For the first The input feature vector corresponding to each step size; For the conventional model In parameters The output below; For extreme and turning point models In parameters The output below.

10. The method for predicting the output of offshore wind farms under extreme and transitional weather conditions according to any one of claims 5-9, characterized in that, The conventional model and the extreme and inflection models are respectively built based on the LightGBM model.

Images