Wind power daily electricity quantity prediction method and system based on multi-model fusion

Abstract

This invention discloses a method and system for predicting daily wind power generation based on multi-model fusion. The method includes: acquiring meteorological data of the wind power station for the prediction day; and adopting an adaptive prediction strategy based on wind speed segments: when the wind speed is in the low-speed segment, prediction is made by fusing historical similar daily power generation data with a dedicated low-speed prediction model; when the wind speed is in the high-speed segment, preliminary prediction is made using the main prediction model, followed by correction based on the physical operating constraints of the wind turbine; when the wind speed is in the normal-speed segment, prediction is directly made using the main prediction model. This invention constructs a dual-model collaborative prediction architecture, breaking the limitations of traditional single-model prediction; it integrates historical experience and a dedicated model in the low-speed segment to effectively overcome the problem of data sparsity and fluctuation; and it introduces physical constraint correction in the high-speed segment to ensure that the prediction results strictly adhere to the unit's power curtailment capacity and physical limits, significantly improving the rationality, safety, and engineering practical value of the prediction results.

Description

Technical Field

[0001] This invention relates to the field of wind power forecasting technology, and in particular to a method and system for forecasting daily wind power generation based on multi-model fusion. Background Technology

[0002] Daily power generation forecasting for wind power plants is a crucial foundation for renewable energy generation dispatch, electricity market trading, and wind farm operation and maintenance management. Accurate forecasts can effectively enhance the grid's capacity to absorb wind power, ensuring the safety and economy of grid operation.

[0003] In existing technologies, daily power generation forecasting for wind power plants typically employs a single machine learning model (such as linear regression, random forest, neural networks, etc.), or selects a fixed master model after comparing multiple models to perform forecasting under all operating conditions. However, in practical engineering applications, existing forecasting methods suffer from the following technical shortcomings: First, the generalization ability of a single model is significantly limited under complex and variable weather conditions. Wind power generation is highly nonlinear and stochastic, making it difficult for a single model to simultaneously ensure prediction accuracy across all wind speed ranges. Especially at the boundaries of wind speed segments (such as near rated wind speeds and in extremely low wind speed ranges) and under extreme weather conditions, the prediction error of a single model will increase significantly, failing to meet the needs of refined scheduling.

[0004] Second, existing technologies lack reliability in low-wind-speed predictions. Under low-wind-speed conditions, wind power generation data often exhibits sparseness and high volatility, making purely data-driven (machine learning) models unreliable in this range. Existing methods fail to effectively combine expert experience based on historical similar daily power generation data with machine learning models, leading to predictions in low-wind-speed conditions that are prone to deviating from reality.

[0005] Third, existing prediction models lack constraints from physical laws and site experience at the output layer. In high-wind-speed conditions or extreme operating conditions (such as low-temperature environments), wind turbines are typically constrained by physical operating limits such as the upper limit of high-wind-speed power curtailment and efficiency degradation due to low temperatures. Existing purely data-driven models often fail to perceive these physical boundaries when making predictions in high-wind-speed conditions, easily producing predictions that contradict the power curtailment capabilities of wind turbines and historical operating experience. This results in predictions under extreme conditions lacking rationality and safety.

[0006] In summary, existing technologies lack a segmented processing mechanism based on meteorological conditions during the forecasting stage, fail to apply different forecasting strategies in coordination for different wind speed ranges, and lack a systematic post-processing scheme for dynamic fusion of multiple models and physical constraints. Summary of the Invention

[0007] The purpose of this invention is to provide a wind power daily power prediction method and system based on multi-model fusion, which organically combines multi-model collaboration, segmented modeling and physical constraint post-processing to solve the above-mentioned technical problems.

[0008] To achieve the above objectives, this invention provides a wind power daily electricity generation prediction method based on multi-model fusion, comprising: Acquire meteorological data for the wind power station on the forecast day, including wind speed; When the wind speed falls into the preset low wind speed range, a fusion prediction is performed based on historical similar daily power generation data and a low wind speed-specific prediction model to obtain the first predicted power value under low wind speed conditions. When the wind speed falls into the preset high wind speed range, a preliminary prediction is made based on the main prediction model, and the preliminary prediction result is corrected in combination with the physical operation constraints of the wind turbine to obtain the second predicted power value under high wind speed conditions. When the wind speed falls within the preset normal wind speed range, the main prediction model is used to make a prediction to obtain the third predicted power value under normal wind speed conditions.

[0009] Preferably, the method for fusing predictions to obtain the first predicted energy value includes: The system searches for matching historical similar day power generation data from historical data using multi-level similarity conditions, and calculates rule-based predicted values ​​based on the historical similar day power generation data. Obtain the model prediction value output by the low wind speed dedicated prediction model; Calculate the coefficient of variation of the electricity values ​​corresponding to the historical similar daily power generation data, and the relative difference between the rule-predicted value and the model-predicted value; Based on the coefficient of variation and the relative difference, the first weight of the rule prediction value and the second weight of the model prediction value are dynamically allocated; Using the first weight and the second weight, the rule-predicted value and the model-predicted value are weighted and fused to obtain the first predicted power value.

[0010] Preferably, the method for calculating the rule prediction value includes: The meteorological data also includes wind direction; based on the preset wind direction sector division rules, the wind direction is mapped to wind direction efficiency; A historical similar daily power generation sample set is obtained by searching based on at least two of the multi-level similarity conditions, including wind speed, wind direction efficiency, temperature, and month. The power generation of the historical similar daily power generation sample set is calculated using a weighted average or median, and the predicted value is obtained after fine-tuning with wind speed and temperature.

[0011] Preferably, when the first predicted power value is obtained, the weighted fusion result is pruned and constrained using the historical quantile interval of the historical similar daily power generation data.

[0012] Preferably, in the process of predicting the second predicted power value, the method for correcting the preliminary prediction result in conjunction with the physical operating constraints of the wind turbine includes: The meteorological data also includes wind direction and temperature; based on the preset wind direction sector division rules, the wind direction is mapped to wind direction efficiency; Based on the wind direction efficiency, the temperature, and the wind speed, a set of similar condition operation statistics is retrieved from the historical power generation dataset. Select historical quantiles or maximum values ​​from the set of similar condition running statistics as dynamic upper bounds, and restrict the preliminary prediction results within the dynamic upper bounds; Based on the preset correction rules for the combined high wind speed and low temperature scenario, logical adjustments and protective lower bound constraints are applied to the preliminary prediction results.

[0013] Preferably, the method for correcting the preliminary prediction results in conjunction with the physical operating constraints of the wind turbine further includes: A temperature effect table is constructed, which includes several efficiency reduction factors corresponding to different wind speeds and temperatures. Based on the wind speed and the temperature, the temperature effect table is queried to obtain the corresponding efficiency reduction factor; Based on the obtained efficiency reduction factor and the preset correction rules for the combined high wind speed and low temperature scenario, logical adjustments and protective lower bound constraints are applied to the preliminary prediction results.

[0014] Preferably, the training methods for the main prediction model and the low-wind-speed-specific prediction model include: Obtain historical daily electricity consumption sequences and corresponding historical meteorological data to form a training sample set; The training sample set is weighted in segments based on the time difference between the sample date and the training deadline, so that the weight of recent samples is higher than that of distant samples. The model is trained in the candidate model pool based on the weighted full training sample set, and the best model is selected based on the evaluation metrics.

[0015] The present invention also provides a wind power daily power generation prediction system, which predicts daily power generation based on the wind power daily power generation prediction method described above.

[0016] The present invention also provides a wind power daily electricity generation prediction system, which includes: One or more processors; Memory; And one or more programs, wherein the one or more programs are stored in the memory and configured to be executed by the one or more processors, the programs including instructions for performing the wind power daily electricity forecasting method as described above.

[0017] The present invention also provides a computer-readable storage medium comprising a computer program that can be executed by a processor to perform the wind power daily power prediction method as described above.

[0018] Compared with existing technologies, the wind power daily power generation prediction method provided by the above technical solution firstly breaks through the generalization limitations of traditional single models and constructs a collaborative prediction architecture of dual models, significantly improving the overall prediction accuracy across the entire wind speed range. Secondly, in the low wind speed range, it integrates historical experience of similar days with a dedicated model, effectively overcoming prediction distortion caused by data sparsity and large fluctuations in this range, while balancing the stability of expert rules with the sensitivity of data models. Finally, in the high wind speed range, it introduces physical operating constraints of wind turbines to correct the initial values ​​of the main model, ensuring that the prediction results strictly follow the unit's power curtailment capacity and physical limits. This completely solves the pain point of pure algorithm models outputting results that contradict actual operating experience under extreme conditions, significantly enhancing the rationality, safety, and engineering practical value of the prediction results in real power grid dispatch. Attached Figure Description

[0019] Figure 1 This is a schematic diagram of the wind power daily electricity prediction method in an embodiment of the present invention. Detailed Implementation

[0020] To illustrate the technical content, structural features, objectives, and effects of the present invention in detail, the following description is provided in conjunction with the embodiments and accompanying drawings.

[0021] This embodiment discloses a wind power daily power generation prediction method based on multi-model fusion, applicable to wind power station day-ahead planning and operation management scenarios, to predict the daily power generation of wind power stations on the predicted day. The prediction method consists of data input, wind speed segment discrimination, segmented prediction, and result output. The segmented prediction adopts a three-layer prediction architecture: for the conventional wind speed segment, the main prediction model outputs the predicted value; for the low wind speed segment, the predicted value is obtained by fusing historical similar day power generation data with a low wind speed-specific prediction model; for the high wind speed segment, the wind turbine physical operation constraints are introduced to correct the preliminary prediction based on the main prediction model.

[0022] like Figure 1 The specific process of this prediction method is as follows: S1. Obtain meteorological data for the wind power station on the forecast date. The meteorological data must include at least wind speed. The wind speed can be the daily average wind speed on the forecast date, or a representative wind speed index aggregated according to a predetermined time window, as long as it can characterize the wind speed level on the forecast date.

[0023] S2. Match the wind speed with preset segmentation rules for low wind speed, high wind speed, and normal wind speed segments. The segmentation rules can be determined based on wind power plant operating experience or historical statistics. In some optional implementations, the thresholds for low wind speed and high wind speed segments can be adjusted according to the unit's start-up characteristics, error-sensitive areas near rated operating conditions, or power curtailment characteristics of the wind farm.

[0024] S3. When the wind speed falls into the low wind speed range, perform fusion prediction, that is: extract the power generation data of similar historical days with wind speed conditions similar to the predicted day from the historical database to form a reference sample for low wind speed range estimation; and output the model prediction value for the predicted day by the low wind speed dedicated prediction model; and synthesize the empirical estimate formed by the power generation data of similar historical days with the model prediction value according to the preset fusion rules to obtain the first predicted power value under the low wind speed conditions.

[0025] The fusion rules can be weighted superposition, weighted average, or other operational forms that can achieve the synthesis of two types of information.

[0026] S4. When the wind speed falls within the normal wind speed range, the main prediction model is used to make predictions and output the third predicted power value under normal wind speed conditions.

[0027] The main prediction model can be a regression model or an ensemble model trained based on historical daily electricity consumption and meteorological data. The input can include wind speed and other features related to daily electricity consumption (such as lag, slip, or trend features constructed from historical electricity consumption data). The output is the predicted value of daily electricity consumption for the prediction day.

[0028] S5. When the wind speed falls into the high wind speed range, the preliminary prediction result is first obtained by the main prediction model; then the preliminary prediction result is corrected by combining the physical operation constraints of the wind turbine to obtain the second predicted power value under the high wind speed condition.

[0029] The physical operating constraints of the wind turbine are used to reflect the wind turbine's power curtailment capability or historical operating range, so that the corrected predicted value does not exceed the physically feasible range; in some optional implementations, non-negative constraints can be applied, and the upper bound obtained from historical statistics can be used to limit the preliminary prediction results.

[0030] S6. Output the first, second, or third predicted power generation value based on the wind speed segment, as the daily power generation prediction result for the wind power station on the prediction day. In some optional implementations, the branch identifier of the wind speed segment used and the values ​​before and after correction can be output simultaneously for easy operation verification.

[0031] By adopting the above three-layer prediction architecture and intelligently switching according to wind speed segments, the prediction process incorporates historical similar day experience information in the low wind speed segment to alleviate the impact of sample sparsity and fluctuations. In the normal wind speed segment, it leverages the overall fitting ability of the main prediction model. In the high wind speed segment, it uses physical constraints to suppress prediction outputs that do not conform to the unit's operating boundaries, thereby reducing systematic deviations under wind speed segment boundaries and extreme operating conditions and improving the usability of prediction results.

[0032] In another embodiment, in order to simultaneously consider the robustness of empirical estimation and the pattern recognition capability of the low-wind-speed-specific prediction model in the low-wind-speed segment, this embodiment introduces a dynamic weight allocation mechanism for the fusion prediction in the low-wind-speed segment.

[0033] Specifically: In the low wind speed range, the matching historical similar daily power generation data is first searched from historical data through multi-level similarity conditions; Then, the rule-based predicted value is calculated based on historical similar daily power generation data; Then the low-wind-speed-specific prediction model outputs the model prediction value; Subsequently, the coefficient of variation (cv) of the electricity generation data corresponding to the historical similar days is calculated. cv is used to characterize the dispersion of the historical similar day sample set. The relative difference (δ) between the rule-predicted values ​​and the model-predicted values ​​is calculated. δ characterizes the degree of disagreement between the two types of predictions. δ can be calculated as follows:

[0034] in, For the rule-predicted value, The "1" in the denominator is used to prevent the ratio from being abnormal due to an excessively small denominator; Based on cv and δ, the first weight of the rule prediction value and the second weight of the model prediction value are determined by a preset dynamic weight allocation function, and the two are made to meet the normalization constraint (e.g., the weight sum is 1). The dynamic weight allocation function follows these principles: when cv increases or δ increases, the second weight is increased and the first weight is decreased; when cv is small and δ is small, the first weight is increased. After the dynamic weights are determined, the rule-predicted value and the model-predicted value are weighted and fused to obtain the first predicted power value.

[0035] For example, CV and δ can be categorized into three levels—"low," "medium," and "high"—based on historical statistics. When either CV or δ is high, the second weight is higher than the first weight (e.g., 0.7 and 0.3); when both CV and δ are low, the first weight is higher than the second weight (e.g., 0.7 and 0.3); in other cases, a near-balanced weight is used (e.g., 0.5 and 0.5). The level division boundary can be determined by the statistical distribution of historical similar day sample sets.

[0036] This dynamic weighting mechanism reduces reliance on a single empirical estimate when historical samples fluctuate significantly or when there is a large discrepancy between the two types of predictions in the low-wind-speed segment. It also increases the contribution of empirical estimates when the samples are stable and the predictions are consistent, thereby reducing the unstable output of predictions in the low-wind-speed segment.

[0037] In another embodiment, the meteorological data also includes wind direction. Based on a preset wind direction sector division rule, the wind direction can be mapped to wind direction efficiency.

[0038] The wind direction sector division rule is used to divide the wind direction angle from 0° to 360° into several sectors (the number of sectors and the angle range can be set according to the site terrain, unit layout or historical statistics).

[0039] Wind direction efficiency can be mapped in one of the following ways: First, the typical power generation level of each sector in historical data (such as the average / median daily power generation or normalized power generation under comparable wind speed conditions) is used as the efficiency index of that sector. Secondly, sectors are grouped according to the size of the efficiency index to obtain categories such as primary high efficiency, secondary high efficiency, and low efficiency, and the category is used as a discrete expression of wind direction efficiency. Third, the efficiency index is normalized to obtain a continuous efficiency coefficient, which is then used as the input for wind direction efficiency in subsequent retrieval.

[0040] Through the above mapping, wind direction information is transformed into an efficiency representation related to power generation capacity, which helps to reduce the impact of "output deviation caused by the same wind speed but different wind direction efficiency" in similar condition retrieval.

[0041] To address this, the retrieval of historical similar daily power generation data employs multi-level similarity criteria, which must include at least two of the following: wind speed, wind direction efficiency, temperature, and month. In some optional implementations, wind speed and month can be prioritized as the first-level filtering criteria, with wind direction efficiency and temperature introduced as further subdivision criteria.

[0042] Multi-level retrieval can gradually relax conditions according to a "strict to lenient" strategy to obtain sufficient samples. For example, matching can be done first based on wind speed within ±0.2; if the sample is insufficient, the range can be expanded to ±0.5. Temperature conditions can be matched first at ±5, and if the sample is insufficient, the range can be expanded to ±10. Monthly conditions are used to suppress deviations caused by seasonal differences. Wind direction efficiency is used to reflect the differences in power generation efficiency under different wind directions due to wake, terrain, or layout. It can be obtained from historical operation statistics and discretized into groups according to preset intervals.

[0043] After obtaining a historical sample set of similar daily power generation, the daily power generation of the sample set is aggregated to obtain the rule-based prediction value. The aggregation method can be either weighted average or median: when there are a few outliers in the sample set, using the median can reduce the impact of outliers; when the samples have different confidence levels, a weighted average can be used, and samples that are closer to the conditions of the prediction day can be given higher weights.

[0044] Subsequently, the wind speed and temperature prediction values ​​can be fine-tuned: when the predicted daily wind speed or temperature is near the boundary of the sample set conditions, the prediction values ​​can be slightly corrected according to the direction of deviation to reduce the system bias caused by condition discretization.

[0045] Through multi-level similarity condition retrieval and robust aggregation, the rule-based predictions can provide interpretable empirical estimates in low-wind-speed, sparse-sample scenarios and provide a stable baseline for subsequent fusion.

[0046] In another embodiment, after obtaining the weighted fusion result for the low wind speed segment, the historical quantile interval of its power generation is calculated based on the historical similar daily power generation data, and the fusion result is restricted to this interval. In a specific example, the 25th quantile and the 75th quantile can be used to construct the interval: when the fusion result is below the 25th quantile, the output is truncated to the 25th quantile; when the fusion result is above the 75th quantile, the output is truncated to the 75th quantile; when the fusion result is within the interval, it remains unchanged. The value of the quantile interval can be adjusted according to the site's preference, as long as the statistical range under historical similar conditions is used as a reasonable boundary.

[0047] When the number of historical similar day samples is small, leading to unstable quantiles, one can choose to skip pruning or use a more robust boundary (e.g., using the minimum and maximum values ​​of the samples to form the boundary) to ensure the feasibility and numerical stability of the constraint operation.

[0048] This pruning constraint makes the output of the low wind speed segment less susceptible to amplification by a small number of outlier samples or fluctuations in the fusion weights, thereby reducing the extreme deviation probability of the low wind speed segment prediction and improving the smoothness of the prediction sequence.

[0049] In addition, in order to make the prediction results under high wind speed conditions consistent with the historical reachable boundary and power curtailment characteristics of the power station, a dynamic upper bound constraint based on similarity condition statistics is introduced into the preliminary prediction results output by the main prediction model in the high wind speed section, and a correction rule for the composite scenario of high wind speed and low temperature and a protection lower bound constraint are superimposed.

[0050] Specifically, the meteorological data obtained during forecasting includes temperature in addition to wind speed.

[0051] Based on wind direction efficiency, temperature, and wind speed, a set of statistics for similar operating conditions is retrieved from historical power generation datasets.

[0052] Similarity criteria retrieval can employ a progressively relaxed strategy to balance similarity and sample size: priority should be given to conditions such as similar wind direction efficiency (or falling into the same efficiency category), temperature within the same temperature range, and wind speed within a range similar to the predicted day. When samples are insufficient, the matching window for wind speed and temperature can be expanded (e.g., wind speed from a narrow window to a wider window, and temperature from a narrow window to a wider window), and seasonal conditions such as month can be retained to suppress the bias introduced by seasonal differences.

[0053] The retrieved set of similar conditional operation statistics includes at least the quantile, median, or maximum values ​​of the daily electricity volume corresponding to the similar sample set, which are used to characterize the historical achievable level under the meteorological combination.

[0054] After obtaining the set of statistics for similar conditions, historical quantiles or maximum values ​​are selected as dynamic upper bounds, and the preliminary prediction results of the master prediction model are restricted to these dynamic upper bounds.

[0055] The selection of the dynamic upper bound can be determined based on the wind power plant's risk appetite or the uncertainty of power curtailment. That is, when it is desirable to more conservatively suppress overestimation, a lower quantile can be selected as the upper bound; when it is desirable to reduce underestimation, the median or maximum value can be selected as the upper bound. This upper bound constraint is a boundary correction for the output layer, ensuring that the preliminary prediction results do not exceed the verifiable range under similar historical conditions, thereby reducing the systematic overestimation caused by "theoretical power generation being too high while actual power curtailment" in high wind speed sections.

[0056] After the upper bound is set, logical adjustments and protective lower bound constraints are applied to the prediction results based on the preset correction rules for the combined high wind speed and low temperature scenarios.

[0057] Composite scenario correction rules are used to handle the superimposed effects when high wind speed and low temperature occur simultaneously: for example, when the temperature falls into the low temperature range and the wind speed is in the high wind speed range, the results after the upper limit are applied with a reduction correction or further contraction, and a boundary is set for the correction magnitude to avoid over-compression.

[0058] The lower bound constraint is used to prevent the forecast result from falling into an obviously unreasonable low range after multiple corrections. It can be formed by selecting a lower quantile or other statistics from the set of statistics for similar operating conditions. When the corrected forecast result is lower than the lower bound, it is raised to the lower bound. Finally, a non-negativity constraint can be applied to avoid negative daily electricity forecast values.

[0059] In this embodiment, the wind direction difference is transformed into efficiency difference and incorporated into the similarity condition retrieval by the "wind direction sector - wind direction efficiency" mapping, so that the historical statistical boundary is closer to the actual reachability of the unit under different incoming flow directions; by the dynamic upper bound limit based on the similarity condition statistics, the output of the main prediction model in the high wind speed range is constrained to the historically verifiable range; by superimposing the high wind speed and low temperature composite rules and the protection lower bound constraint, the prediction results can suppress the risk of overestimation and avoid unreasonable underestimation, thereby improving the physical consistency and operational availability of the prediction output under high wind speed and special meteorological conditions.

[0060] Based on the above embodiments, in order to perform calculable efficiency reduction correction for high wind speed predictions in scenarios involving low temperature effects, the method for correcting the preliminary prediction results in conjunction with wind turbine physical operating constraints further includes: A temperature effect table is constructed, which includes several efficiency reduction factors corresponding to different wind speeds and temperatures.

[0061] Wind speed can be discretized according to preset wind speed sub-segments, and temperature can be discretized according to preset temperature segments. Each combination corresponds to a degradation factor. The degradation factor can be obtained from historical power generation data statistics, such as the ratio between historical daily power generation statistics of different temperature segments within the same wind speed segment, or it can be determined from unit operating experience or manufacturer performance curves; the source of the temperature effect table is not limited, as long as it can reflect the impact of temperature on power generation capacity.

[0062] During forecasting, the temperature effect table is consulted based on the wind speed and temperature of the forecast day to obtain the corresponding performance degradation factor. The performance degradation factor is then combined with the forecast value after dynamic upper bound correction, for example, by multiplicatively scaling the forecast value to obtain a corrected value that takes into account the temperature effect. Subsequently, combined with the preset correction rules for the combined scenario of high wind speed and low temperature, logical adjustments and a protective lower bound constraint are applied to the corrected value: when both high wind speed and low temperature conditions are met simultaneously, the forecast value can be further reduced by the performance degradation factor based on the upper bound constraint, and it is ensured that the forecast value is not lower than the protective lower bound formed by the set of similar condition operation statistics.

[0063] In another embodiment, in order to obtain a master prediction model and a low-wind-speed-specific prediction model that can be used for the prediction phase, this embodiment provides a training method, which includes data construction, time decay weighting, candidate model training, and evaluation and selection.

[0064] First, obtain historical daily electricity consumption sequences and corresponding historical meteorological data to form a training sample set.

[0065] Historical meteorological data may include wind speed, wind direction, air pressure, and outside temperature. When the sampling time of meteorological data is inconsistent with the statistical caliber of daily electricity consumption, the meteorological data can be aggregated by day and aligned with the daily electricity consumption by date. In a specific example, the meteorological time can be offset by a fixed amount (e.g., reduced by 1 hour) before being aggregated into daily average meteorological data to reduce time alignment errors.

[0066] Optionally, samples from maintenance days can be removed from the training sample set, and suspected power-limited days can be marked to avoid abnormal operating conditions interfering with training.

[0067] Subsequently, the training sample set is weighted in segments based on the time difference between the sample date and the training deadline, so that the weight of recent samples is higher than that of distant samples.

[0068] In a specific example, weighting can be segmented into segments: high weight for the last 2 months, standard weight for 2–6 months, reduced weight for 6–12 months, and low weight for over one year. The boundaries of these segments and the weight values ​​can be adjusted according to the rate of change in the station's operating mode, as long as a monotonic decay relationship of "higher in the recent period and lower in the long term" is satisfied. These weights are used as sample weights in the model training loss calculation, thereby making the model more focused on the recent operating mode.

[0069] The model is trained in a candidate model pool and then selected to obtain the master prediction model. The candidate model pool can include Ridge Regression, Lasso, Random Forest, Gradient Boosting, XGBoost, LightGBM, CatBoost, etc. The evaluation metric can use the coefficient of determination (R²) as the primary indicator, supplemented by the mean absolute percentage error (MAPE). After selection, the master prediction model, along with its associated preprocessing tools and feature list, is saved to ensure consistency between the training and prediction phases.

[0070] In some alternative implementations, a subset of samples with wind speeds falling into the low wind speed range can be selected from the training sample set. Features can be constructed or selected separately for low wind speed scenarios, and a low wind speed-specific prediction model can be obtained by training and selecting the best model in the candidate model pool. When the number of low wind speed samples is insufficient, the low wind speed-specific modeling can be skipped, and only the main prediction model and the low wind speed empirical estimation path based on historical similar daily power generation data can be retained.

[0071] By using time decay weighting and multi-model pool training optimization, the prediction model can maintain its adaptability to recent data even as the operating mode changes over time, while providing reusable model components for the three-layer architecture of the prediction stage.

[0072] In summary, this invention discloses a method for predicting daily wind power output, constructing a complete technical solution that includes data weighting training, segmented collaborative prediction, and multidimensional physical post-processing. This systematic solution uses wind speed segments as the core of decision-making, overcoming the generalization limitations of traditional single data-driven models. In the prediction execution stage, for regular wind speed segments, a robust baseline is output based on the optimized master model. For low wind speed segments, the problem of data sparsity and fluctuation is overcome by dynamically weighting and fusing historical similarity patterns with a dedicated model and quantile pruning. For high wind speed segments and extreme weather conditions, wind direction efficiency mapping and temperature effects are deeply integrated, and dynamic upper bounds and composite protective constraints are constructed using historical multidimensional similarity statistics to apply strict physical boundary corrections to the initial values ​​of the model.

[0073] This three-layer prediction architecture, which organically integrates machine learning algorithms, historical operating experience, and the physical power curtailment laws of wind turbines, effectively eliminates the systematic deviation of prediction results under extreme wind speed boundaries and complex operating conditions. It ensures that the final output of daily electricity strictly conforms to the physical feasible range of the actual site, and achieves a significant improvement in prediction accuracy, robustness, and engineering interpretability across the entire wind speed range.

[0074] In another preferred embodiment of the present invention, a wind power daily power generation prediction system is also disclosed, which predicts daily power generation based on the wind power daily power generation prediction method in the above embodiments.

[0075] This invention also discloses another wind power daily energy forecasting system, which includes one or more processors, a memory, and one or more programs, wherein the one or more programs are stored in the memory and configured to be executed by the one or more processors. The programs include instructions for performing the wind power daily energy forecasting method as described above. The processor may be a general-purpose central processing unit (CPU), a microprocessor, an application-specific integrated circuit (ASIC), or one or more integrated circuits, used to execute relevant programs to implement the functions required by the modules in the wind power daily energy forecasting system of this application embodiment, or to execute the wind power daily energy forecasting method of this application embodiment.

[0076] This invention also discloses a computer-readable storage medium comprising a computer program executable by a processor to perform the wind power daily electricity forecasting method described above. The computer-readable storage medium can be any available medium accessible to a computer or a data storage device such as a server or data center integrating one or more available media. The available medium can be read-only memory (ROM), random access memory (RAM), or magnetic media, such as floppy disks, hard disks, magnetic tapes, magnetic disks, or optical media, such as digital versatile discs (DVDs), or semiconductor media, such as solid-state disks (SSDs).

[0077] This application also discloses a computer program product or computer program, which includes computer instructions stored in a computer-readable storage medium. The processor of an electronic device reads the computer instructions from the computer-readable storage medium and executes the computer instructions, causing the electronic device to perform the aforementioned wind power daily electricity prediction method.

[0078] The above-disclosed embodiments are merely preferred embodiments of the present invention and should not be construed as limiting the scope of the present invention. Therefore, any equivalent variations made in accordance with the claims of the present invention are still within the scope of the present invention.

Claims

1. A method for predicting daily wind power generation based on multi-model fusion, characterized in that, include: Acquire meteorological data for the wind power station on the forecast day, including wind speed; When the wind speed falls into the preset low wind speed range, a fusion prediction is performed based on historical similar daily power generation data and a low wind speed-specific prediction model to obtain the first predicted power value under low wind speed conditions. When the wind speed falls into the preset high wind speed range, a preliminary prediction is made based on the main prediction model, and the preliminary prediction result is corrected in combination with the physical operation constraints of the wind turbine to obtain the second predicted power value under high wind speed conditions. When the wind speed falls within the preset normal wind speed range, the main prediction model is used to make a prediction to obtain the third predicted power value under normal wind speed conditions.

2. The wind power daily electricity generation prediction method according to claim 1, characterized in that, The method for fusing predictions to obtain the first predicted energy value includes: The system searches for matching historical similar day power generation data from historical data using multi-level similarity conditions, and calculates rule-based predicted values ​​based on the historical similar day power generation data. Obtain the model prediction value output by the low wind speed dedicated prediction model; Calculate the coefficient of variation of the electricity values ​​corresponding to the historical similar daily power generation data, and the relative difference between the rule-predicted value and the model-predicted value; Based on the coefficient of variation and the relative difference, the first weight of the rule prediction value and the second weight of the model prediction value are dynamically allocated; Using the first weight and the second weight, the rule-predicted value and the model-predicted value are weighted and fused to obtain the first predicted power value.

3. The wind power daily electricity generation prediction method according to claim 2, characterized in that, The method for calculating the rule-predicted value includes: The meteorological data also includes wind direction; based on the preset wind direction sector division rules, the wind direction is mapped to wind direction efficiency; A historical similar daily power generation sample set is obtained by searching based on at least two of the multi-level similarity conditions, including wind speed, wind direction efficiency, temperature, and month. The power generation of the historical similar daily power generation sample set is calculated using a weighted average or median, and the predicted value is obtained after fine-tuning with wind speed and temperature.

4. The wind power daily electricity generation prediction method according to claim 1, characterized in that, When the first predicted power value is obtained, the weighted fusion result is pruned and constrained using the historical quantile interval of the historical similar daily power generation data.

5. The wind power daily electricity generation prediction method according to claim 1, characterized in that, In the process of predicting the second predicted power value, the method for correcting the preliminary prediction result in combination with the physical operating constraints of the wind turbine includes: The meteorological data also includes wind direction and temperature; based on the preset wind direction sector division rules, the wind direction is mapped to wind direction efficiency; Based on the wind direction efficiency, the temperature, and the wind speed, a set of similar condition operation statistics is retrieved from the historical power generation dataset. Select historical quantiles or maximum values ​​from the set of similar condition running statistics as dynamic upper bounds, and restrict the preliminary prediction results within the dynamic upper bounds; Based on the preset correction rules for the combined high wind speed and low temperature scenario, logical adjustments and protective lower bound constraints are applied to the preliminary prediction results.

6. The wind power daily electricity generation prediction method according to claim 5, characterized in that, The method for correcting the preliminary prediction results based on the physical operating constraints of the wind turbine also includes: A temperature effect table is constructed, which includes several efficiency reduction factors corresponding to different wind speeds and temperatures. Based on the wind speed and the temperature, the temperature effect table is queried to obtain the corresponding efficiency reduction factor; Based on the obtained efficiency reduction factor and the preset correction rules for the combined high wind speed and low temperature scenario, logical adjustments and protective lower bound constraints are applied to the preliminary prediction results.

7. The wind power daily electricity generation prediction method according to claim 1, characterized in that, The training methods for the main prediction model and the low-wind-speed-specific prediction model include: Obtain historical daily electricity consumption sequences and corresponding historical meteorological data to form a training sample set; The training sample set is weighted in segments based on the time difference between the sample date and the training deadline, so that the weight of recent samples is higher than that of distant samples. The model is trained in the candidate model pool based on the weighted full training sample set, and the best model is selected based on the evaluation metrics.

8. A wind power daily electricity generation prediction system, characterized in that, The prediction system predicts daily power generation based on the wind power daily power generation prediction method according to any one of claims 1 to 7.

9. A wind power daily electricity generation prediction system, characterized in that, include: One or more processors; Memory; And one or more programs, wherein the one or more programs are stored in the memory and configured to be executed by the one or more processors, the programs including instructions for performing the wind power daily electricity forecasting method as claimed in any one of claims 1 to 7.

10. A computer-readable storage medium, characterized in that, Includes a computer program that can be executed by a processor to perform the wind power daily power prediction method as described in any one of claims 1 to 7.

Images