New energy day-ahead power prediction optimization method based on asymmetric sample weight
Patent Information
- Application Number
- CN202611397130.2
- Authority / Receiving Office
- CN · China
- Patent Type
- Applications(China)
- Current Assignee / Owner
- Filing Date
- 2026-09-09
- Publication Date
- 2026-10-30
AI Technical Summary
[0006]总体来看，现有研究在模型构建、误差修正及策略调整方面呈现出明显的分散性，往往仅针对单一类型极端天气（如寒潮、大风等）分别进行精细化建模，且大多沿用基于对称损失函数的常规优化目标，未能反映极端场景下高估与低估风险代价的非对称性
[0038]1、采用统计分位数方法定于各种极端天气阈值以此识别极端天气发生时刻。并定义连续强度指标刻画极端天气强度。针对极端样本稀缺问题，引入基于强度的连续权重函数。该机制使得模型在训练过程中强化极端天气误差，抑制常规天气样本主导效应从而实现极端场景预测性能的提升。针对不同天气类型独立训练模型，使模型能够学习不同气象条件下的功率变化规律。在预测阶段，根据极端天气强度进行自适应融合；
Smart Images
Figure CN123013001A_ABST
Abstract
Description
Technical Field
[0001] This invention relates to the field of power system forecasting technology, and in particular to an optimization method for predicting the day-ahead power of new energy sources based on asymmetric sample weights. Background Technology
[0002] The geographical areas where new energy power generation bases are located are often in regions prone to extreme weather. Widespread low-temperature icing caused by cold waves, prolonged periods of low wind speeds and low sunlight due to stable weather conditions, load surges and equipment derating accompanied by summer heat waves, and sudden disasters such as heavy fog and strong winds all frequently impact the operating environment of new energy power plants. These extreme weather events further amplify the inherent random fluctuations on the source side of the new power system, causing the output curve to experience sharp rises or falls in a short period, posing a severe challenge to the optimized operation of the system under high-proportion new energy integration.
[0003] This challenge manifests itself on two interconnected levels. Firstly, extreme weather is characterized by low probability, frequent occurrences, strong persistence, and multiple hazard combinations. Its specific evolution path, key turning points, and formation mechanisms are influenced by multiple factors, including atmospheric circulation anomalies, local micro-topographic forcing, and multi-physics field interactions, exhibiting high nonlinearity and non-stationarity. This makes high-precision prediction of renewable energy power extremely difficult. Traditional extrapolation methods based on historical statistical patterns often exhibit systematic biases under extreme weather scenarios, resulting in severely insufficient model generalization capabilities. Secondly, when the accuracy of renewable energy power prediction is low, power system dispatch strategies formulated based on this as boundary condition will face a dual degradation in economic efficiency and safety. Low-precision prediction results not only directly weaken the feasibility of dispatch plans but may also lead to insufficient system reserve capacity, severe power imbalances, and even frequency drops and cascading failures during periods when peak loads coincide with persistent extreme weather events, posing a substantial threat to the system's adequacy and safe, stable operation.
[0004] Conventional forecasting models typically assume that training data follows an independent and identically distributed distribution and that weather samples are relatively balanced. However, in real-world datasets, normal weather samples dominate, while extreme weather events such as cold waves, icing, and prolonged high temperatures are typical low-probability, long-tailed events with extremely sparse sample sizes and severely skewed distributions. Directly training the model under these conditions will naturally tilt its learning focus towards normal samples, resulting in a severe underfit to extreme samples and amplifying prediction errors by orders of magnitude during extreme weather periods. Conventional forecasting models generally use symmetric loss functions such as mean squared error and mean absolute error as optimization objectives, implicitly assuming that overestimation and underestimation have equal economic costs. However, in power system operation scenarios triggered by extreme weather, the directionality of prediction bias has drastically different risk implications—severe overestimation of power may lead to insufficient spinning reserves, grid frequency drops, and even load shedding risks, while underestimation mainly causes economic losses such as wind and solar power curtailment. Symmetrical loss functions cannot reflect this asymmetric risk structure, making it impossible for the model to specifically avoid prediction biases in high-cost directions during training. Consequently, the model's predictions under extreme weather conditions often favor the higher-risk error directions.
[0005] Currently, research on power prediction for renewable energy under extreme weather conditions mainly focuses on the identification of weather processes and the mechanism modeling of their impact on the power system, and often focuses on single extreme weather types. In the case of winter cold waves, existing research has established power loss prediction models for wind farms due to icing. Regarding the identification of transitional weather events, some scholars have introduced gated recurrent unit classifiers and object-oriented diagnostic testing methods to construct short-term wind power prediction models, focusing on the identification and testing of transitional wind speed sequences to improve the ability to capture uphill events. Furthermore, for power loss prediction under cold waves, researchers have proposed a binary classification model based on an attention mechanism and used temporal generative adversarial networks to expand cold wave samples to alleviate the problem of sample scarcity; based on this, they have integrated XGBoost and Transformer algorithms to establish prediction systems for power baseline values and loss values, respectively.
[0006] Overall, existing research exhibits significant fragmentation in model building, error correction, and strategy adjustment. It often focuses on detailed modeling of only a single type of extreme weather (such as cold waves and strong winds), and mostly employs conventional optimization objectives based on symmetric loss functions, failing to reflect the asymmetry between overestimating and underestimating the risks and costs in extreme scenarios. This categorized and symmetric research paradigm leads to fragmentation among methods in terms of weather type coverage, sample imbalance handling, and risk preference expression, failing to establish a unified prediction and optimization framework for multiple extreme weather scenarios. Summary of the Invention
[0007] The purpose of this invention is to overcome the shortcomings of the prior art and provide a new energy day-ahead power prediction optimization method based on asymmetric sample weights.
[0008] The objective of this invention is achieved through the following technical solution: a new energy day-ahead power prediction optimization method based on asymmetric sample weights, the method comprising the following steps:
[0009] S1. Acquire multi-source meteorological data and historical power data of new energy power stations, perform data cleaning, spatiotemporal alignment and feature fusion on the multi-source meteorological data, and obtain fused meteorological feature data.
[0010] S2. Based on the fused meteorological feature data and historical power data, construct a multi-dimensional feature vector, which includes at least physical features and historical power statistical features.
[0011] S3. Input the multidimensional feature vector into the pre-constructed gradient boosting decision tree model for training, and use the mean square error or mean absolute error as the loss function to output the basic power prediction result under normal weather conditions.
[0012] S4. Define extreme weather thresholds for each meteorological variable using the statistical quantile method, determine the time of extreme weather occurrence based on the extreme weather thresholds, and construct a continuous extreme intensity index to characterize the intensity level of extreme weather; based on the extreme intensity index, construct an asymmetric sample weight function, assign higher weights to samples in extreme weather conditions than to samples in normal weather conditions, and use a weighted loss function to replace the original loss function for model training, thereby obtaining multiple sub-models for different extreme weather types.
[0013] S5. Real-time identification of extreme weather types and corresponding extreme intensity indices at the current forecast time; adaptive fusion of the forecast results of the multiple sub-models based on the intensity of each extreme weather type; and output of the final new energy day-ahead power forecast result.
[0014] Specifically, the multi-source meteorological data includes at least numerical weather forecast data; the data cleaning uses the interquartile range method or physical threshold method to remove outliers; the spatiotemporal alignment includes time alignment using linear interpolation and spatial alignment using bilinear interpolation; and the feature fusion includes basic feature fusion and derived feature fusion.
[0015] Specifically, the physical characteristics include one or more of air density, wind shear index, solar altitude angle, and solar azimuth angle; the historical power statistical characteristics include one or more of daily historical power, intraday historical power, historical power average, and historical power standard deviation.
[0016] Specifically, the gradient boosting decision tree model uses the LightGBM algorithm and utilizes gradient-based one-sided sampling and mutually exclusive feature binding strategies for training acceleration.
[0017] The one-sided sampling strategy is as follows:
[0018] ;
[0019] In the formula, This is the final weighted gradient for sample i; Let be the original gradient of sample i, where i is the sample index, representing the current training sample. For a large gradient sample set, Let be the random sampling set of gradient samples, where a is the proportion of large gradient samples and b is the random sampling rate.
[0020] The mutual exclusion feature binding strategy is as follows:
[0021] ;
[0022] In the formula, Original features; Let m be the i-th original feature; m is the total number of original features. Features after binding; Let k be the number of bundles for the i-th feature. For the j-th feature bundle; Let n be the mutually exclusive feature of the bundle containing the j-th feature.
[0023] Specifically, the continuous extreme intensity index softens the extreme weather threshold by using an S-shaped function or a piecewise linear function to convert the hard threshold judgment into a continuous value, so that the value of the extreme intensity index in the normal weather range approaches zero, while the value in the extreme weather range increases monotonically with the degree of deviation, and the value range is constrained to the [0,1] interval.
[0024] Specifically, the asymmetric sample weighting function is:
[0025] ;
[0026] In the formula, >1 represents the nonlinear enhancement parameter; The extreme weather type to which the i-th sample belongs, such as strong winds, low temperatures, icing, smog, etc. The normalized weather intensity is calculated based on meteorological indicators and judgment thresholds corresponding to the extreme weather type, and its value ranges from 0 to 1. =1 and =3 represents the minimum and maximum weight boundaries uniformly set by the sample weighting mechanism; The final weight coefficient obtained for the i-th sample is denoted as .
[0027] Specifically, the weighted loss function is:
[0028] ;
[0029] In the formula, L is the weighted loss function value, which measures the overall error between the model's prediction and the true value; N is the total number of training samples; and i is the sample index, representing the i-th training sample. The weight coefficient for the i-th sample is set according to the importance of the sample or the difficulty of prediction, and is used to enhance the training contribution of key samples (such as extreme weather samples). This represents the actual renewable energy power value of the i-th sample. Let be the new energy power value predicted by the model for the i-th sample.
[0030] Specifically, the adaptive fusion adopts a softmax weight fusion strategy based on extreme strength, and the fusion weight calculation formula is as follows:
[0031] ;
[0032] In the formula, This is a smoothing parameter used to adjust the degree of difference in weights between different models; This is the predicted output value for the k-th extreme weather model. This represents the day-ahead power forecast of the integrated new energy sources; Let be the extreme weather intensity evaluation index corresponding to the k-th model; e is a natural constant. is the k-th model fusion weight calculated based on the Softmax function; j is the model index, used to traverse each extreme weather prediction model participating in the fusion; Let be the extreme weather intensity evaluation index corresponding to the j-th model.
[0033] Specifically, when multiple extreme weather events exist simultaneously, the adaptive fusion in S5 further includes: firstly calculating the prediction results of the sub-models corresponding to each extreme weather type, and then, based on the intensity of each extreme weather event, using the softmax weight to allocate the contribution ratio of each sub-model, so that the sub-model corresponding to the more intense extreme weather type occupies a higher weight in the final prediction.
[0034] Specifically, the continuous extreme intensity index is constructed as follows:
[0035] ;
[0036] In the formula, is a continuous extreme intensity index; x is a meteorological variable; T is the extreme threshold of the corresponding meteorological variable; This represents the maximum value of the meteorological variable. This is the minimum value of the meteorological variable; To prevent tiny positive numbers with a denominator of zero; high-value represents high-value extreme weather variables (such as wind speed, temperature, etc., when the extreme degree increases as the variable increases when the threshold is exceeded); low-value represents low-value extreme weather variables (such as low temperature, visibility, etc., when the extreme degree increases as the variable decreases when the threshold is exceeded).
[0037] The present invention has the following advantages:
[0038] 1. Statistical quantile methods are used to define various extreme weather thresholds to identify the timing of extreme weather events. A continuous intensity index is defined to characterize the intensity of extreme weather. To address the scarcity of extreme weather samples, an intensity-based continuous weighting function is introduced. This mechanism allows the model to amplify extreme weather errors during training and suppress the dominance effect of regular weather samples, thereby improving the prediction performance of extreme scenarios. The model is trained independently for different weather types, enabling it to learn the power variation patterns under different meteorological conditions. In the prediction phase, adaptive fusion is performed based on the intensity of extreme weather events.
[0039] 2. The advantages of this invention are that it significantly improves the accuracy and stability of day-ahead power prediction for renewable energy under various extreme weather conditions such as cold waves, high temperatures, strong winds, and smog, while avoiding output jumps during weather state transitions, providing a more reliable risk-oriented prediction boundary for grid dispatch. The method is highly versatile and can be flexibly adapted to wind power / photovoltaic power and different base models, possessing good prospects for engineering deployment and effectively supporting the power balance and safe operation of high-proportion renewable energy power systems under extreme weather conditions. Attached Figure Description
[0040] Figure 1 This is a schematic diagram of the day-ahead power prediction optimization method of the present invention;
[0041] Figure 2 This is a schematic diagram of the data processing flow of the day-ahead power prediction optimization method of the present invention. Detailed Implementation
[0042] To make the objectives, technical solutions, and advantages of this invention clearer, the invention will be further described in detail below with reference to the accompanying drawings and embodiments. It should be understood that the specific embodiments described herein are only for explaining the invention and are not intended to limit the invention; that is, the described embodiments are merely some embodiments of the invention, and not all embodiments. The components of the embodiments of the invention described and shown in the accompanying drawings can generally be arranged and designed in various different configurations.
[0043] Therefore, the following detailed description of the embodiments of the invention provided in the accompanying drawings is not intended to limit the scope of the claimed invention, but merely to illustrate selected embodiments of the invention. All other embodiments obtained by those skilled in the art based on the embodiments of the invention without inventive effort are within the scope of protection of the invention.
[0044] It should be noted that relational terms such as "first" and "second" are used merely to distinguish one entity or operation from another, and do not necessarily require or imply any such actual relationship or order between these entities or operations. Furthermore, the terms "comprising," "including," or any other variations thereof are intended to cover non-exclusive inclusion, such that a process, method, article, or apparatus that comprises a list of elements includes not only those elements but also other elements not expressly listed, or elements inherent to such a process, method, article, or apparatus. Without further limitation, an element defined by the phrase "comprising one..." does not exclude the presence of other identical elements in the process, method, article, or apparatus that includes said element.
[0045] The present invention will be further described below with reference to the accompanying drawings, but the scope of protection of the present invention is not limited to the following description.
[0046] like Figures 1 to 2 As shown, a new energy day-ahead power prediction optimization method based on asymmetric sample weights is proposed, which includes the following steps:
[0047] S1. Acquire multi-source meteorological data and historical power data of new energy power stations, perform data cleaning, spatiotemporal alignment and feature fusion on the multi-source meteorological data, and obtain fused meteorological feature data.
[0048] In this embodiment, the multi-source heterogeneous meteorological data differ in terms of encoding methods, temporal resolution, spatial resolution, and included meteorological elements. Therefore, meteorological data fusion is required before use. The multi-source meteorological data includes at least numerical weather prediction data. The data cleaning uses the interquartile range method or physical threshold method to remove outliers. The spatiotemporal alignment includes temporal alignment using linear interpolation and spatial alignment using bilinear interpolation. The feature fusion includes basic feature fusion and derived feature fusion. Through processing, meteorological grid data is interpolated to the location of new energy power stations, and time is downscaled, that is, NWP data with a time interval of 1 hour or longer is downscaled to a 15-minute resolution. Bias correction is performed by using historical observation data to correct the systematic errors of numerical weather prediction.
[0049] S2. Based on the fused meteorological feature data and historical power data, a multidimensional feature vector is constructed. The multidimensional feature vector includes at least physical features and historical power statistical features. In this embodiment, the physical features include one or more of air density, wind shear index, solar altitude angle, and solar azimuth angle. The historical power statistical features include one or more of daily historical power, intraday historical power, historical power average, and historical power standard deviation.
[0050] The air density is calculated as follows:
[0051] ;
[0052] In the formula, air density; It refers to air pressure; Absolute temperature; ρ is the gas constant for dry air; air density directly affects wind power density, which, along with the cube of wind speed, determines the magnitude of wind power output. Air density varies with temperature, altitude, and humidity. In low-temperature, high-altitude areas, air density is high, resulting in higher wind power output at the same wind speed; conversely, air density decreases during hot summer months, leading to a drop in power output. Ignoring seasonal variations in air density will introduce systematic power prediction biases.
[0053] The wind shear index is calculated as follows:
[0054] ;
[0055] In the formula, V is the wind speed; z is the vertical height; The wind shear index describes the variation of wind speed with altitude and directly affects the actual wind speed at the turbine hub height. The wind shear index varies significantly under different atmospheric stability conditions. If a fixed wind shear coefficient is used to estimate the wind speed at hub height, it will lead to large errors in power prediction under strong wind shear weather, especially affecting the accuracy of capturing climbing events.
[0056] The solar altitude angle is calculated as follows:
[0057] ;
[0058] In the formula, The solar altitude angle; Latitude; The solar declination angle; The solar altitude angle is the angle between the sun's rays and the horizontal plane, directly determining the intensity of direct solar radiation reaching the ground. A larger solar altitude angle means a shorter path for solar radiation through the atmosphere, resulting in higher irradiance received by the Earth's surface and stronger photovoltaic output. The solar altitude angle varies with latitude, season, and time of day, and is a fundamental geometric parameter affecting irradiance calculations in photovoltaic power prediction.
[0059] The solar azimuth angle is calculated as follows:
[0060] ;
[0061] In the formula, Latitude; The solar declination angle; The hour angle; The solar azimuth angle describes the direction of the sun's projection onto the horizontal plane and, together with the solar altitude angle, determines the angle of incidence of sunlight relative to the photovoltaic array. The amount of irradiance received by the photovoltaic module is affected by the cosine effect of the incident angle. An azimuth angle deviation will lead to inaccurate calculation of the irradiance on the inclined surface, thus affecting the accuracy of photovoltaic power prediction on the inclined surface.
[0062] The historical daytime power for the three days prior to forecast is calculated as follows:
[0063] ;
[0064] In the formula, This refers to the historical power level of the day before yesterday; To predict the historical power at the same time t on the first day prior; To predict the historical power at the same time t on the second day prior; Forecast historical power at the same time t three days prior. Predicting the historical power three days prior reflects recent power output levels and weather evolution trends, providing medium- to long-term power variation patterns. It helps models capture the cumulative impact of persistent weather events on renewable energy output and is an important reference input for distinguishing between normal and transitional weather events.
[0065] The intraday historical power for the first six hours is calculated as follows:
[0066] ;
[0067] In the formula, This represents the historical power data for the past six hours; the historical power data for the past six hours reflects the power response under current meteorological conditions in real time. When the weather changes rapidly, the recent power series contains the latest trends and inertial information, making it an important reference for predicting power trends in the coming hours.
[0068] The historical average power over the previous six hours is calculated as follows:
[0069] ;
[0070] In the formula, i is the time index, representing the i-th prediction time; This is the historical power average of the previous six hours. The historical power average of the previous six hours compresses short-term fluctuation noise and extracts the overall level characteristics of recent power. It can effectively characterize the current power output baseline, help the model distinguish between instantaneous power fluctuations and trend changes, and is especially helpful in determining whether the current output is in a high or low range.
[0071] The standard deviation of historical power over the previous 24 hours is calculated as follows:
[0072] ;
[0073] In the formula, The standard deviation of historical power over the previous 24 hours; This is used to predict historical power over the previous 24 hours. The standard deviation of historical power over the previous 24 hours measures the recent fluctuation range of power, reflecting the stability and drastic changes of weather processes. A larger standard deviation usually indicates frequent power fluctuations or turning points in weather, and is of significant reference value for error prediction and confidence assessment of forecasting models.
[0074] S3. Input the multidimensional feature vector into the pre-constructed gradient boosting decision tree model for training, and use the mean square error or mean absolute error as the loss function to output the basic power prediction result under normal weather conditions.
[0075] The gradient boosting decision tree model uses the LightGBM algorithm and utilizes gradient-based one-sided sampling and mutually exclusive feature binding strategies for training acceleration. In this embodiment, LightGBM extracts the main features that affect power and reduces the impact of redundant features. LightGBM has been deeply optimized for large-scale datasets.
[0076] The one-sided sampling strategy is as follows:
[0077] ;
[0078] In the formula, The final weighted gradient for sample i, Let be the original gradient of sample i, where i is the sample index, representing the current training sample. For a large gradient sample set, Let be the random sampling set of gradient samples, where a is the proportion of large gradient samples and b is the random sampling rate.
[0079] The mutual exclusion feature binding strategy is as follows:
[0080] ;
[0081] In the formula, Original features; Let m be the i-th original feature; m is the total number of original features. Features after binding; Let k be the number of bundles for the i-th feature. For the j-th feature bundle; Let n be the mutually exclusive feature of the bundle containing the j-th feature.
[0082] S4. Define extreme weather thresholds for each meteorological variable using the statistical quantile method, determine the time of extreme weather occurrence based on the extreme weather thresholds, and construct a continuous extreme intensity index to characterize the intensity level of extreme weather; based on the extreme intensity index, construct an asymmetric sample weight function, assign higher weights to samples in extreme weather conditions than to samples in normal weather conditions, and use a weighted loss function to replace the original loss function for model training, thereby obtaining multiple sub-models for different extreme weather types.
[0083] The standard GBDT (Gradient Boosting Decision Tree) model uses Mean Squared Error (MSE) or Mean Absolute Error (MAE) as its loss function by default. These loss functions are symmetric and penalize both overestimation and underestimation equally. Since extreme weather samples constitute a very small percentage of the training set, the model optimization process is dominated by a large number of regular weather samples, causing the gradient direction to be biased towards intermediate values. However, when extreme samples are scarce, their contribution to the overall gradient is diluted, resulting in insufficient model fitting ability for extreme values. The stepwise residual fitting mechanism of GBDT essentially corrects the error of the previous round, but when the samples are imbalanced, subsequent trees mainly correct the residuals of regular samples, leaving the residuals of extreme samples insufficiently corrected. Therefore, this embodiment constructs a continuous extreme intensity function during the model training phase, assigning higher weights to extreme samples, thereby significantly improving the model's fitting ability in extreme scenarios while maintaining stability under regular weather conditions. In the prediction phase, adaptive fusion is performed through a weather recognition mechanism. For extreme weather events (such as strong winds, smog, high temperatures, and low temperatures), the threshold is defined using the statistical quantile method, which aligns with the meteorological definition of extreme weather.
[0084] ;
[0085] In the formula, For corresponding meteorological variables (such as wind speed, temperature, and visibility); For the sample 90th percentile; For the sample The 10th percentile; the rules for determining extreme weather are:
[0086] ;
[0087] In the formula, "Extreme" is the extreme weather indicator. It takes a value of 1 when a sample meets the preset extreme weather criteria, and a value of 0 otherwise. It is used to distinguish extreme weather samples from regular weather samples and serves as the trigger condition for asymmetric sample weight allocation and prediction model optimization. To avoid the discontinuity caused by hard thresholds, a continuous extreme intensity index is constructed as follows:
[0088] ;
[0089] In the formula, is a continuous extreme intensity index; x is a meteorological variable; T is the extreme threshold of the corresponding meteorological variable; This represents the maximum value of the meteorological variable. This is the minimum value of the meteorological variable; To prevent tiny positive numbers with a denominator of zero; high-value refers to high-value extreme weather variables (such as wind speed, temperature, etc., where the degree of extremeness increases as the variable increases when the threshold is exceeded); low-value refers to low-value extreme weather variables (such as low temperature, visibility, etc., where the degree of extremeness increases as the variable decreases when the threshold is exceeded).
[0090] And apply interval constraints:
[0091] ;
[0092] in, Extreme weather intensity;
[0093] The continuous extreme intensity index softens the extreme weather threshold by using an S-shaped function or a piecewise linear function to convert the hard threshold judgment into a continuous value, so that the value of the extreme intensity index approaches zero in the normal weather range, and the value in the extreme weather range increases monotonically with the degree of deviation, and the value range is constrained to the [0,1] interval.
[0094] To address the problem of extreme sample scarcity, an asymmetric sample weighting function is introduced as follows:
[0095] ;
[0096] In the formula, >1 represents the nonlinear enhancement parameter; The extreme weather type to which the i-th sample belongs, such as strong winds, low temperatures, icing, smog, etc. The normalized weather intensity is calculated based on meteorological indicators and judgment thresholds corresponding to the extreme weather type, and its value ranges from 0 to 1. =1 and =3 represents the minimum and maximum weight boundaries uniformly set by the sample weighting mechanism; The final weight coefficient obtained for the i-th sample is denoted as .
[0097] The weighting function is asymmetric; the weight for regular weather samples is close to 1, while the weight for extreme weather samples increases significantly, thus making the model focus more on high-risk samples. In the regression model, the weighted loss function is expressed as:
[0098] ;
[0099] In the formula, L is the weighted loss function value, which measures the overall error between the model's prediction and the true value; N is the total number of training samples; and i is the sample index, representing the i-th training sample. The weight coefficient for the i-th sample is set according to the importance of the sample or the difficulty of prediction, and is used to enhance the training contribution of key samples (such as extreme weather samples). This represents the actual renewable energy power value of the i-th sample. Let be the new energy power value predicted by the model for the i-th sample.
[0100] This mechanism enables the model to amplify extreme weather errors during training and suppress the dominance effect of regular weather samples, thereby improving the prediction performance for extreme scenarios. Training the model independently for different weather types allows it to learn the power variation patterns under different meteorological conditions.
[0101] S5. Real-time identification of extreme weather types and corresponding extreme intensity indices at the current forecast time; adaptive fusion of the forecast results of the multiple sub-models based on the intensity of each extreme weather type; and output of the final new energy day-ahead power forecast result.
[0102] In this embodiment, during the prediction phase, adaptive fusion is performed based on the intensity of extreme weather events. The adaptive fusion adopts a softmax weight fusion strategy based on extreme weather intensity, and the fusion weight calculation formula is as follows:
[0103] ;
[0104] In the formula, This is a smoothing parameter used to adjust the degree of difference in weights between different models; This is the predicted output value for the k-th extreme weather model. This represents the day-ahead power forecast of the integrated new energy sources; Let be the extreme weather intensity evaluation index corresponding to the k-th model; e is a natural constant. is the k-th model fusion weight calculated based on the Softmax function; j is the model index, used to traverse each extreme weather prediction model participating in the fusion; Let be the extreme weather intensity evaluation index corresponding to the j-th model.
[0105] When multiple extreme weather events coexist, the adaptive fusion in S5 further includes: firstly calculating the prediction results of the sub-models corresponding to each extreme weather type, and then, based on the intensity of each extreme weather event, using the softmax weight to allocate the contribution ratio of each sub-model, so that the sub-model corresponding to the more intense extreme weather type occupies a higher weight in the final prediction.
[0106] The above description is merely a preferred embodiment of the present invention and does not constitute any limitation on the present invention. Any person skilled in the art can make many possible variations and modifications to the technical solution of the present invention, or modify it into equivalent embodiments, without departing from the scope of the present invention. Therefore, any modifications, equivalent changes, and alterations made to the above embodiments based on the technology of the present invention without departing from the scope of the present invention are within the protection scope of the present invention.
Claims
1. A new energy day-ahead power prediction optimization method based on asymmetric sample weights, characterized in that: The method includes the following steps: S1. Acquire multi-source meteorological data and historical power data of new energy power stations, perform data cleaning, spatiotemporal alignment and feature fusion on the multi-source meteorological data, and obtain fused meteorological feature data. S2. Based on the fused meteorological feature data and historical power data, construct a multi-dimensional feature vector, which includes at least physical features and historical power statistical features. S3. Input the multidimensional feature vector into the pre-constructed gradient boosting decision tree model for training, and use the mean square error or mean absolute error as the loss function to output the basic power prediction result under normal weather conditions. S4. Define extreme weather thresholds for each meteorological variable using the statistical quantile method, determine the time of extreme weather occurrence based on the extreme weather thresholds, and construct a continuous extreme intensity index to characterize the intensity level of extreme weather; based on the extreme intensity index, construct an asymmetric sample weight function, assign higher weights to samples in extreme weather conditions than to samples in normal weather conditions, and use a weighted loss function to replace the original loss function for model training, thereby obtaining multiple sub-models for different extreme weather types. S5. Real-time identification of extreme weather types and corresponding extreme intensity indices at the current forecast time; adaptive fusion of the forecast results of the multiple sub-models based on the intensity of each extreme weather type; and output of the final new energy day-ahead power forecast result.
2. The method for optimizing the prediction of day-ahead power of new energy sources based on asymmetric sample weights according to claim 1, characterized in that: The multi-source meteorological data includes at least numerical weather forecast data; the data cleaning uses interquartile range or physical thresholding to remove outliers; the spatiotemporal alignment includes temporal alignment using linear interpolation and spatial alignment using bilinear interpolation; the feature fusion includes basic feature fusion and derived feature fusion.
3. The method for optimizing the prediction of day-ahead power of new energy sources based on asymmetric sample weights according to claim 1, characterized in that: The physical characteristics include one or more of air density, wind shear index, solar altitude angle, and solar azimuth angle; the historical power statistical characteristics include one or more of daily historical power, intraday historical power, historical power average, and historical power standard deviation.
4. The new energy day-ahead power prediction optimization method based on asymmetric sample weights according to claim 1, characterized in that: The gradient boosting decision tree model uses the LightGBM algorithm and utilizes a gradient-based one-sided sampling strategy and a mutually exclusive feature binding strategy to accelerate training. The one-sided sampling strategy is as follows: ； In the formula, This is the final weighted gradient for sample i; Let be the original gradient of sample i, where i is the sample index, representing the current training sample; For a large gradient sample set, Let be the random sampling set of gradient samples, where a is the proportion of samples with large gradients and b is the random sampling rate; The mutual exclusion feature binding strategy is as follows: ； In the formula, Original features; Let m be the i-th original feature; m is the total number of original features. Features after binding; Let k be the number of bundles for the i-th feature. For the j-th feature bundle; It is the nth mutually exclusive feature in the bundle of the jth feature.
5. The method for optimizing the prediction of day-ahead power of new energy sources based on asymmetric sample weights according to claim 1, characterized in that: The continuous extreme intensity index softens the extreme weather threshold by using an S-shaped function or a piecewise linear function to convert the hard threshold judgment into a continuous value, so that the value of the extreme intensity index approaches zero in the normal weather range, and the value in the extreme weather range increases monotonically with the degree of deviation, and the value range is constrained to the [0,1] interval.
6. The method for optimizing the prediction of day-ahead power of new energy sources based on asymmetric sample weights according to claim 1, characterized in that: The asymmetric sample weighting function is: ； In the formula, >1 represents the nonlinear enhancement parameter; The extreme weather type to which the i-th sample belongs; The normalized weather intensity is calculated based on meteorological indicators and judgment thresholds corresponding to the extreme weather type, and its value ranges from 0 to 1. and These are the minimum and maximum weight boundaries uniformly set for the sample weighting mechanism; The final weight coefficient obtained for the i-th sample is denoted as .
7. The method for optimizing the prediction of day-ahead power of new energy sources based on asymmetric sample weights according to claim 6, characterized in that: The weighted loss function is: ； In the formula, L is the weighted loss function value; N is the total number of training samples; and i is the sample index, representing the i-th training sample. Let be the weight coefficient of the i-th sample; This represents the actual renewable energy power value for the i-th sample. Let be the new energy power value predicted by the model for the i-th sample.
8. The new energy day-ahead power prediction optimization method based on asymmetric sample weights according to claim 1, characterized in that: The adaptive fusion adopts a softmax weight fusion strategy based on extreme strength, and the fusion weight calculation formula is as follows: ； In the formula, This is a smoothing parameter used to adjust the degree of difference in weights between different models; This is the predicted output value for the k-th extreme weather model; This represents the day-ahead power forecast of the integrated new energy sources; Let be the extreme weather intensity evaluation index corresponding to the k-th model; e is a natural constant. is the k-th model fusion weight calculated based on the Softmax function; j is the model index, used to traverse each extreme weather prediction model participating in the fusion; Let be the extreme weather intensity evaluation index corresponding to the j-th model.
9. The method for optimizing the prediction of day-ahead power of new energy sources based on asymmetric sample weights according to claim 8, characterized in that: When multiple extreme weather events coexist, the adaptive fusion in S5 further includes: firstly calculating the prediction results of the sub-models corresponding to each extreme weather type, and then, based on the intensity of each extreme weather event, using the softmax weight to allocate the contribution ratio of each sub-model, so that the sub-model corresponding to the more intense extreme weather type occupies a higher weight in the final prediction.
10. The method for optimizing the prediction of day-ahead power of new energy sources based on asymmetric sample weights according to claim 1, characterized in that: The continuous extreme intensity index is constructed as follows: ； In the formula, is a continuous extreme intensity index; x is a meteorological variable; T is the extreme threshold of the corresponding meteorological variable; This represents the maximum value of the meteorological variable. This is the minimum value of the meteorological variable; To prevent tiny positive numbers with a denominator of zero; high-value refers to high-value extreme weather variables; low-value refers to low-value extreme weather variables.