A method and device for predicting extreme scenarios of renewable energy power based on dual-input features and nonparametric quantile regression

Abstract

This invention provides a method and apparatus for predicting renewable energy power in extreme scenarios based on dual-input features and nonparametric quantile regression. First, historical power data and historical meteorological data are acquired. Then, based on the historical meteorological data and the physical response mechanism of renewable energy output under extreme weather conditions, a dual-input feature vector is constructed. Finally, the historical power data and the dual-input feature vector are input into a nonparametric quantile regression model to obtain the probability prediction result for the extreme scenario. This invention constructs a dual-input feature vector based on the physical response mechanism of renewable energy output under extreme weather conditions and combines it with a nonparametric quantile regression model with bandwidth parameter optimization for nonlinear fitting. This accurately characterizes the complex coupling relationship between meteorology and power output and achieves risk quantification, thereby effectively improving the prediction accuracy and reliability of renewable energy power in extreme scenarios.

Description

Technical Field

[0001] This invention relates to the field of power forecasting technology, and in particular to a method and apparatus for predicting extreme scenarios of new energy power based on dual-input features and nonparametric quantile regression. Background Technology

[0002] Against the backdrop of escalating global climate change, the frequency and intensity of extreme weather events such as blizzards, cold waves, heat waves, and strong winds are continuously increasing, posing severe challenges to the safe and stable operation of new energy power systems such as photovoltaic and wind power. Extreme weather has a significantly destructive impact on new energy output: blizzards can drastically reduce the irradiance received by photovoltaic modules, and wet snow and ice accumulation can even lead to a complete interruption of photovoltaic power generation; extreme low temperatures caused by cold waves can trigger the protective shutdown of wind turbines, creating a "wind but no power" dilemma; high temperatures and strong winds can respectively lead to a decrease in the conversion efficiency of photovoltaic modules and an increase in the overload risk of wind turbines. These extreme scenarios directly restrict the reliable output of new energy and the balance between power grid supply and demand, and have become core risk points that must be overcome in the development of new power systems.

[0003] Current mainstream renewable energy power prediction technologies are insufficient to meet the accurate prediction needs of extreme weather scenarios, exhibiting significant technical shortcomings. Traditional prediction models are mostly built based on normal meteorological and power output data, relying on fixed parameters or linear correlation assumptions. They cannot characterize the unsteady and nonlinear coupling relationship between meteorological parameters (such as sudden drops in irradiance, extreme temperatures, and drastic changes in wind speed) and renewable energy power output (such as nonlinear attenuation and protective shutdown) under extreme weather conditions, resulting in a significant decrease in prediction accuracy under extreme scenarios. Summary of the Invention

[0004] This invention provides a method and apparatus for predicting new energy power in extreme scenarios based on dual-input features and nonparametric quantile regression, which addresses the problem that existing methods lead to a significant decrease in prediction accuracy under extreme scenarios.

[0005] A first aspect of this invention provides a method for predicting extreme scenarios of renewable energy power based on dual-input features and nonparametric quantile regression, comprising: Acquire historical power data and historical meteorological data; Based on historical meteorological data and the physical response mechanism of new energy output under extreme weather conditions, a dual-input feature vector is constructed. Historical power data and dual-input feature vectors are input into a nonparametric quantile regression model to obtain probability prediction results for extreme scenarios; Among them, the kernel function of the nonparametric quantile regression model is used to perform nonlinear fitting of the relationship between the dual input features and the renewable energy power; the nonparametric quantile regression model also includes a bandwidth parameter optimization module; the bandwidth parameter optimization module is used to optimize the nonlinear mapping between the dual input features and different fractions of renewable energy power.

[0006] In one possible implementation, a dual-input feature vector is constructed based on historical meteorological data and the physical response mechanism of renewable energy output under extreme weather conditions, including: Historical meteorological data were preprocessed in multiple dimensions and screened using two criteria to obtain core basic parameters and auxiliary correction parameters; Based on core basic parameters and auxiliary correction parameters, a dual-input feature vector with a coupling correction mechanism is constructed for different extreme weather scenarios and equipment types. The dual-input feature vector includes instantaneous feature values ​​determined based on the device physical equations and derived dimensions determined based on time-series statistical features.

[0007] In one possible implementation, based on core fundamental parameters and auxiliary correction parameters, a dual-input feature vector with a coupled correction mechanism is constructed for different extreme weather scenarios and equipment types, including: Based on the physical response mechanism under different extreme weather scenarios, the core basic parameters are physically corrected to obtain real-time characteristic values; Calculate the time-series derived dimension based on instantaneous feature values ​​and auxiliary correction parameters; The instantaneous feature values ​​and time-series derived dimensions are weighted and fused according to the coupling weights to construct a dual-input feature vector. The coupling weights are determined based on the statistical distribution or physical mechanism verification of historical sample data in extreme scenarios.

[0008] In one possible implementation, historical power data and a dual-input feature vector are fed into a nonparametric quantile regression model to obtain extreme scenario probability predictions, including: The dual-input feature vectors and historical power data are grouped according to extreme weather scenario types, divided into training and validation sets, and the dual-input feature vectors are subjected to scenario-based normalization. A Gaussian kernel function is used to perform a weighted mapping on the sample points in the training and validation sets to fit the nonlinear coupling relationship between the dual-input feature vector and the new energy power. According to the bandwidth parameter optimization module, the k-fold cross-validation algorithm is adopted, and the bandwidth parameter is iteratively optimized based on minimizing the quantile loss function on the validation set. The quantile loss function is used to quantify the deviation between the predicted value and the true value at different quantile levels. An extreme scenario penalty coefficient is introduced to adjust the fitting weights for extreme low and high quantile values; The dual-input feature vectors for the prediction period are input into the optimized nonparametric quantile regression model, which outputs a multi-level quantile prediction sequence of new energy power. Based on the comparison between the quantile prediction sequence and the preset threshold, the probability of extreme scenario risks is quantified.

[0009] In one possible implementation, based on the bandwidth parameter optimization module, a k-fold cross-validation algorithm is used to iteratively optimize the bandwidth parameter by minimizing the quantile loss function on the validation set, including: Based on the dimension of the dual-input feature vector and the statistical characteristics of historical samples of extreme weather scenarios, the initial search range of the bandwidth parameter is determined, and the training set corresponding to each extreme weather scenario is divided into k mutually exclusive subsets. Training and validation are carried out in a loop through k-fold cross-validation. Quantile loss functions are defined for different quantile levels; for extremely low and extremely high quantiles, a scene adaptation coefficient is introduced to modify the quantile loss function in order to adjust the model's fitting weights for extreme features. Within the bandwidth parameter search range, the optimal bandwidth parameter is determined with the goal of minimizing the average loss value after correction at each quantile level, and the robustness of the optimal bandwidth parameter is verified. If the verification result meets the preset threshold, the bandwidth parameter is determined as the final optimization result. If it does not meet the threshold, the search range is narrowed and iterative optimization is performed again until the preset threshold is met.

[0010] In one possible implementation, an extreme scenario penalty coefficient is introduced to adjust the fitting weights for extreme low and high quantiles, including: Based on the intensity level of extreme weather and the operational risk threshold of new energy equipment, a penalty coefficient classification system is constructed to determine the initial value of the penalty coefficient for extreme scenarios. Based on the equipment type and risk focus, the penalty coefficients for extreme low-value quantiles and extreme high-value quantiles are configured differently, and the configured extreme scenario penalty coefficients are embedded into the quantile loss function to form a weighted loss function; Based on the prediction error feedback from historical sample data of extreme scenarios, the penalty coefficient for extreme scenarios is dynamically calibrated to maintain a balance between prediction errors of extreme quantiles and non-extreme quantiles. For different extreme weather scenarios of the same equipment type, the scenario similarity is calculated, and the penalty coefficient of the calibrated extreme scenario is transferred to the similar scenario for adjustment based on the scenario similarity.

[0011] In one possible implementation, the probability of extreme scenario risk is quantified based on the comparison between the quantile prediction sequence and a preset threshold, including: Based on the safety standards for the operation of new energy equipment and the constraints of power grid dispatch, a hierarchical preset threshold system is constructed. The extreme quantiles in the quantile prediction sequence are associated and mapped with a hierarchical preset threshold system; Based on the relative positional relationship between extreme quantiles and stratified preset thresholds, as well as the fluctuation characteristics of the quantile sequence, the probability of a single risk is calculated, and combined with the risk priority of extreme weather scenarios, the overall risk probability is calculated. Based on the deviation between historical measured data and predicted data in extreme scenarios, the comprehensive risk probability is corrected, and the corresponding risk level is matched according to the corrected comprehensive risk probability to output quantitative decision results.

[0012] In one possible implementation, a Gaussian kernel function is used to perform a weighted mapping on the sample points in the training and validation sets to fit the nonlinear coupling relationship between the dual-input feature vector and the new energy power, including: Based on the dimensionality of the dual-input feature vector and the numerical distribution characteristics of sample data under extreme weather scenarios, the core parameters of the Gaussian kernel function are initialized. The Gaussian kernel function calculates the similarity between samples based on the Euclidean distance between the dual-input feature vector to be predicted and the historical feature sample vector, and adjusts it according to the dynamic optimization value output by the bandwidth parameter optimization module. The sample points are weighted by distance based on the Gaussian kernel function; Based on the differences in the physical response mechanisms of different extreme weather scenarios, the Gaussian kernel function is modified according to the specific scenarios. The modified Gaussian kernel function is embedded into the fitting layer of the nonparametric quantile regression model to construct a mapping function from the dual-input feature vector to different quantiles of new energy power, thus completing the fitting of the nonlinear coupling relationship.

[0013] In one possible implementation, the method further includes: Decision information is determined based on the probability prediction results of extreme scenarios.

[0014] A second aspect of the present invention provides a new energy power extreme scenario prediction device based on dual-input features and nonparametric quantile regression, characterized in that it comprises: The acquisition module is used to acquire historical power data and historical meteorological data; The processing module is used to construct a dual-input feature vector based on historical meteorological data and the physical response mechanism of new energy output under extreme weather conditions; The prediction module is used to input historical power data and dual-input feature vectors into a nonparametric quantile regression model to obtain the probability prediction results for extreme scenarios. Among them, the kernel function of the nonparametric quantile regression model is used to perform nonlinear fitting of the relationship between the dual input features and the renewable energy power; the nonparametric quantile regression model also includes a bandwidth parameter optimization module; the bandwidth parameter optimization module is used to optimize the nonlinear mapping between the dual input features and different fractions of renewable energy power.

[0015] Compared to traditional technologies, this invention provides a method and apparatus for predicting renewable energy power in extreme scenarios based on dual-input features and nonparametric quantile regression. First, historical power data and historical meteorological data are acquired. Then, based on the historical meteorological data and the physical response mechanism of renewable energy output under extreme weather conditions, a dual-input feature vector is constructed. Finally, the historical power data and the dual-input feature vector are input into a nonparametric quantile regression model to obtain the probability prediction result of the extreme scenario. The kernel function of the nonparametric quantile regression model is used to perform nonlinear fitting of the relationship between the dual-input features and renewable energy power. The nonparametric quantile regression model also includes a bandwidth parameter optimization module, which optimizes the nonlinear mapping between the dual-input features and different fractions of renewable energy power. This invention, by constructing a dual-input feature vector based on the physical response mechanism of renewable energy output under extreme weather conditions and combining it with a nonparametric quantile regression model with bandwidth parameter optimization for nonlinear fitting, can accurately characterize the complex coupling relationship between meteorology and power output and achieve risk quantification, thereby effectively improving the prediction accuracy and reliability of renewable energy power in extreme scenarios. Attached Figure Description

[0016] Figure 1 This is a flowchart illustrating the implementation of the new energy power extreme scenario prediction method based on dual-input features and nonparametric quantile regression provided in this embodiment of the invention. Figure 2 This is a schematic diagram of the structure of the new energy power extreme scenario prediction device based on dual-input features and nonparametric quantile regression provided in the embodiment of the present invention. Detailed Implementation

[0017] The embodiments of the present invention will now be described in detail with reference to the accompanying drawings.

[0018] Figure 1 This is a flowchart illustrating the implementation of the new energy power extreme scenario prediction method based on dual-input features and nonparametric quantile regression provided in this embodiment of the invention. Figure 1 As shown, the method includes: S110, acquires historical power data and historical meteorological data; S120, based on historical meteorological data and the physical response mechanism of new energy power output under extreme weather conditions, constructs a dual-input feature vector; S130, input historical power data and dual-input feature vectors into the nonparametric quantile regression model to obtain the probability prediction results of extreme scenarios; Among them, the kernel function of the nonparametric quantile regression model is used to perform nonlinear fitting of the relationship between the dual input features and the renewable energy power; the nonparametric quantile regression model also includes a bandwidth parameter optimization module; the bandwidth parameter optimization module is used to optimize the nonlinear mapping between the dual input features and different fractions of renewable energy power.

[0019] In this embodiment of the invention, the first step is to systematically collect historical power data and historical meteorological data. For the target renewable energy power plants, including photovoltaic power plants and wind farms, long-term continuous historical power output data needs to be collected, covering power output changes under different seasons and weather conditions. Particular emphasis should be placed on accumulating power response data during extreme weather events such as blizzards, cold waves, high temperatures, and strong winds to ensure the data reflects the power output characteristics under extreme scenarios. Simultaneously, corresponding high-precision, multi-dimensional meteorological data is collected, covering key parameters such as irradiance, ambient temperature, wind speed, humidity, air pressure, snowfall, icing thickness, and cloud cover level. The collection frequency of meteorological data needs to match that of power data to ensure temporal consistency, providing a high-quality data foundation for subsequent analysis of the correlation between meteorological factors and power output. After collection, a preliminary integrity check of the raw data is performed to remove data with obvious sensor malfunctions.

[0020] After data collection, the core feature construction stage begins, which involves building a customized dual-input feature vector based on historical meteorological data and the physical response mechanism of new energy power output under extreme weather conditions. This process is not simply parameter selection and combination, but rather based on in-depth physical mechanism analysis and data mining. First, the collected multi-dimensional historical meteorological data undergoes multi-step preprocessing, including identifying and correcting outliers using the Grubbs criterion, filling in missing data using time series interpolation, and then performing standardization transformation and dynamic time-series feature extraction to form a multi-dimensional parameter matrix. Subsequently, a dual-criteria screening system is used to select core parameters. On the one hand, statistical methods such as Pearson correlation coefficient, mutual information entropy, and grey relational analysis are used to quantify the correlation strength between each meteorological parameter and historical power data. On the other hand, physical significance verification is performed in conjunction with the operating mechanism of new energy equipment to eliminate redundant parameters that are statistically correlated but physically meaningless. Finally, irradiance, ambient temperature, and wind speed are determined as the core basic parameters, while humidity, snowfall, and air pressure are used as auxiliary correction parameters. Based on this, feature combinations are designed for different extreme weather scenarios and equipment types: photovoltaic power plants focus on the coupling effect of irradiance intensity and ambient temperature. For example, in blizzard scenarios, the synergistic effect of effective irradiance intensity after snow cover correction and cell operating temperature needs to be considered. In high-temperature scenarios, the correlation between irradiance saturation correction intensity and module heat dissipation limit temperature is considered. Wind farms focus on the combination of wind speed and ambient temperature. In strong wind scenarios, the effective wind power density after wind shear correction and the temperature associated with unit load need to be incorporated. In cold wave scenarios, the synergistic effect of low-temperature correction effective wind speed and unit critical protection temperature is analyzed. At the same time, it is also necessary to combine time series statistical features to calculate the instantaneous mutation rate, cumulative duration, peak / valley values ​​and other derived dimensions of each core parameter. Finally, the instantaneous feature values ​​and derived dimensions are weighted and fused through coupling weights to construct a dual-input feature vector that can accurately characterize the synergistic mechanism of multiple meteorological factors under extreme weather conditions.

[0021] After the feature vectors are constructed, they are input together with the preprocessed historical power data into the nonparametric quantile regression model to initiate the core calculation process for extreme scenario probability prediction. The core advantage of this model lies in its ability to flexibly adapt to complex nonlinear correlations under extreme weather conditions, with its built-in kernel function playing a crucial role in nonlinear fitting. The kernel function uses a Gaussian kernel as the basic fitting unit. By calculating the Euclidean distance between the feature vector to be predicted and the feature vectors of historical samples, a similarity-weighted mapping is applied to the sample points. That is, historical samples with closer distances and higher similarity are given greater weight in the fitting process, allowing the model to prioritize learning from historical extreme weather samples similar to the current prediction scenario, effectively avoiding interference from dissimilar samples. This weighted mapping method does not require a pre-defined fixed correlation function between meteorological features and renewable energy power. It can adaptively fit the non-steady-state and nonlinear coupling relationship between the two under extreme weather conditions based entirely on the distribution characteristics of the sample data itself. Whether it's the nonlinear output decay of photovoltaic power plants under blizzards or the critical tripping response of wind farms under cold waves, both can be accurately characterized through the flexible mapping of the kernel function.

[0022] To further enhance the model's predictive robustness and accuracy, a dedicated bandwidth parameter optimization module is included in the nonparametric quantile regression model. This module optimizes the nonlinear mapping effect between dual-input features and different quantiles of renewable energy power. First, based on the dimensionality of the dual-input feature vectors and the statistical characteristics of historical samples from extreme weather scenarios, the module determines the initial search interval for the bandwidth parameter. Then, using a k-fold cross-validation algorithm, the training set is divided into multiple mutually exclusive subsets. Through alternating training and validation, the bandwidth parameter is iteratively optimized by minimizing the quantile loss function on the validation set. During optimization, a scenario fit coefficient is introduced to correct the loss function for extremely low and high quantiles, strengthening the model's fitting weight for extreme features. This dynamic adjustment of the bandwidth parameter effectively balances the model's fitting accuracy for local features with its global generalization ability, avoiding overfitting caused by the sparseness and drastic fluctuations of historical extreme weather samples, and ensuring robust learning even with limited samples. Ultimately, the model outputs a multi-level quantile prediction sequence containing the 5% extreme low quantile, 25% quantile, 50% median, 75% quantile, and 95% extreme high quantile, enabling precise quantification of the range and probability of power fluctuations and risks of new energy sources under extreme scenarios, and providing comprehensive decision-making basis for grid dispatching and power plant operation and maintenance.

[0023] In some embodiments, a dual-input feature vector is constructed based on historical meteorological data and the physical response mechanism of new energy output under extreme weather conditions. This includes: performing multi-dimensional preprocessing and dual-criteria screening on historical meteorological data to obtain core basic parameters and auxiliary correction parameters; and constructing a dual-input feature vector with a coupling correction mechanism for different extreme weather scenarios and equipment types based on the core basic parameters and auxiliary correction parameters. The dual-input feature vector includes instantaneous feature values ​​determined based on the equipment physical equations and derived dimensions determined based on time-series statistical features.

[0024] In this embodiment of the invention, the process of constructing a dual-input feature vector begins with in-depth processing and precise screening of historical meteorological data, which is the fundamental step to ensure the effectiveness of the features. First, the collected multi-dimensional raw meteorological data undergoes multi-step preprocessing. Firstly, outlier values ​​caused by sensor malfunctions are identified and removed using the Grubbs criterion. Then, time-series interpolation is used to fill in missing data caused by equipment maintenance or signal interruption, preventing incomplete data from affecting subsequent analysis. Next, the preprocessed parameters are standardized, normalizing parameters with different dimensions, such as irradiance and wind speed, to a unified range. Simultaneously, key parameters such as temperature are dimensionless based on the equipment's critical operating threshold, eliminating interference from numerical differences. Finally, the dynamic temporal characteristics of each meteorological parameter are extracted, including hourly averages, instantaneous change rates, cumulative duration, peak and trough values, and differences between adjacent time periods, forming a multi-dimensional parameter matrix covering both static values ​​and dynamic changes. Based on this, a dual-criteria screening system was used to determine the core parameters: on the one hand, the Pearson correlation coefficient was used to measure the strength of linear correlation, mutual information entropy was used to quantify the degree of nonlinear correlation, and grey relational degree was used to adapt to the characteristics of small sample data, thus statistically screening out feature sets that are highly correlated with historical power data; on the other hand, physical significance was verified by combining the operating mechanism of new energy equipment, such as verifying whether the parameters of photovoltaic modules conform to the photoelectric conversion equation, and verifying whether the parameters of wind turbines match the aerodynamic characteristic curves and protection logic. Finally, redundant parameters that are "statistically correlated but physically meaningless" were eliminated, and irradiance, ambient temperature, and wind speed were determined as the core basic parameters, while humidity, snowfall, and air pressure were used as auxiliary correction parameters.

[0025] Once the core parameters are determined, the next step is to construct targeted feature combinations. The core idea is to customize the feature design based on the unique physical response mechanisms of different extreme weather scenarios and equipment types. For photovoltaic power plants, the focus is on the "snow layer shading - low temperature synergistic damage" mechanism in blizzard scenarios, and on the "irradiation saturation - high temperature thermal decay" effect in high temperature scenarios. For wind farms, the focus is on the "wind speed threshold - structural load synergy" relationship in strong wind scenarios, and on the "low temperature turbine switching - wind speed threshold synergistic triggering" logic in cold wave scenarios. The feature combination for each scenario is based on the core basic parameters, combined with auxiliary correction parameters for physical-level optimization and adjustment. For example, photovoltaic features in blizzard weather need to incorporate the shading correction of snowfall on irradiance intensity, and wind power features in cold wave weather need to consider the impact of air pressure on air density, ensuring that the features can accurately match the physical action laws under extreme weather conditions and avoiding correlation distortion problems caused by generalized features.

[0026] In the specific process of feature quantification and dimensional expansion, each dual-input feature vector contains an instantaneous feature value derived from the equipment's physical equations, as well as a derived dimension obtained from time-series statistical analysis. Together, they constitute a complete feature system. The calculation of instantaneous feature values ​​strictly follows the physical laws of equipment operation. For example, the effective irradiance of a photovoltaic power station needs to be corrected in multiple layers using factors such as snow layer shading coefficient, cloud layer attenuation coefficient, and terrain reflection coefficient. The cell operating temperature is calculated based on the module thermal balance equation, combined with parameters such as irradiance absorption and module heat dissipation capacity. The effective wind power density of a wind farm needs to consider the comprehensive effects of air density, turbulence intensity, and wind energy utilization coefficient. The critical protection temperature of the unit needs to be corrected in combination with the nacelle insulation attenuation and ambient temperature to ensure that the instantaneous feature values ​​can truly reflect the real-time operating status of the equipment under extreme weather conditions. The derived dimensions focus on the temporal variation characteristics of meteorological parameters. By calculating the instantaneous mutation rate, cumulative duration, and number of times the critical threshold is exceeded for core parameters, the dynamic development process and cumulative effects of extreme weather can be captured. For example, the duration of temperature continuously below the critical value, the cumulative time of wind speed exceeding the cut-out threshold, and the instantaneous rate of decrease in irradiance can be captured. These derived dimensions can supplement the information missing in the time dimension of instantaneous characteristic values ​​and more comprehensively depict the dynamic impact of extreme weather on the output of new energy sources.

[0027] Finally, a coupling and fusion mechanism is used to integrate the instantaneous feature values ​​and derived dimensions into a final dual-input feature vector, ensuring that the synergistic correlation between features is fully reflected. During the coupling process, based on the statistical distribution characteristics and physical mechanism verification results of historical sample data in extreme scenarios, corresponding coupling weights are assigned to the instantaneous feature values ​​and derived dimensions. The magnitude of the weights directly reflects the degree of influence of each dimension on the output of new energy. For example, in the photovoltaic feature under the blizzard scenario, the effective irradiance intensity after snow layer shading correction has a higher weight than the irradiance mutation rate; in the wind power feature under the cold wave scenario, the critical protection temperature of the unit has a higher weight than the temperature drop rate. At the same time, the interaction influence coefficient of the two core feature dimensions is analyzed through partial least squares regression. If the coupling degree is insufficient, auxiliary correction parameters are introduced for secondary optimization. Finally, the integrated feature vector is subjected to scenario-based normalization to ensure that the feature vector of the input model conforms to both physical correlation and numerical consistency, providing high-quality and highly targeted input basis for the accurate fitting of the subsequent nonparametric quantile regression model.

[0028] In some embodiments, based on core basic parameters and auxiliary correction parameters, a dual-input feature vector with a coupling correction mechanism is constructed for different extreme weather scenarios and device types. This includes: physically correcting the core basic parameters according to the physical response mechanism under different extreme weather scenarios to obtain instantaneous feature values; calculating the time-series derived dimension based on the instantaneous feature values ​​and auxiliary correction parameters; and weighting and fusing the instantaneous feature values ​​and the time-series derived dimension according to the coupling weight to construct a dual-input feature vector. The coupling weight is determined based on the statistical distribution of historical sample data of extreme scenarios or verification of physical mechanisms.

[0029] In this embodiment of the invention, a dual-input feature vector with a coupling correction mechanism is constructed. The primary step is to refine the core basic parameters based on the unique physical response mechanisms under different extreme weather scenarios, thereby obtaining real-time feature values ​​that accurately reflect the operating status of the equipment. For photovoltaic power plants, in blizzard scenarios, the core basic parameter, irradiance, will be significantly reduced due to snow and cloud cover. Simultaneously, low temperatures will affect the cell conversion efficiency. Therefore, it is necessary to combine auxiliary correction parameters such as snowfall and cloud cover level, and perform multi-layer correction of irradiance intensity using snow cover coefficients and cloud attenuation coefficients to obtain effective irradiance intensity. Ambient temperature needs to be converted to the actual operating temperature of the cells based on the thermal resistance characteristics of the modules to avoid feature distortion caused by differences between ambient temperature and the internal temperature of the modules. In wind farms under cold wave scenarios, wind speed is affected by changes in air density. Auxiliary parameters such as air pressure and ambient temperature need to be used to correct air density to obtain effective wind speed. Ambient temperature needs to consider the insulation effect of the nacelle and the insulation degradation due to the age of the equipment, and be corrected to the actual operating temperature of the key components of the unit to ensure that the real-time feature values ​​accurately match the physical operating status of the equipment.

[0030] After obtaining the instantaneous feature values, it is necessary to combine the changing patterns of the instantaneous feature values ​​with auxiliary correction parameters to further calculate the time-series derived dimensions in order to capture the dynamic development and cumulative effects of extreme weather. The construction of time-series derived dimensions is not a simple statistical calculation, but rather revolves around the logic of the impact of extreme weather on the output of new energy sources. Taking the high-temperature scenario of photovoltaic power plants as an example, based on the corrected irradiance saturation intensity and the module heat dissipation limit temperature, the cumulative duration of irradiance intensity exceeding the optimal operating range of the module and the proportion of periods when the temperature is continuously higher than the critical value are calculated. These derived dimensions can reflect the continuous damage to the modules caused by the high-temperature environment. In the strong wind scenario of wind farms, the instantaneous wind speed fluctuation amplitude and the deviation between the 10-minute average wind speed and the instantaneous wind speed are calculated through the effective wind speed. At the same time, combined with the turbulence intensity auxiliary parameter, the cumulative duration of wind speed exceeding the cut-out threshold is extracted to accurately characterize the impact of the volatility and cumulative impact of strong winds on the unit output, making up for the limitation that instantaneous feature values ​​can only reflect the state at a single moment.

[0031] Determining the coupling weights is crucial for effective feature fusion. Its core lies in scientifically allocating the influence weights of each dimension of features based on the statistical distribution characteristics and physical mechanism verification results of historical sample data from extreme scenarios. At the statistical distribution level, the correlation strength between each feature and new energy output is analyzed in a large number of historical samples from extreme scenarios. For example, the correlation between photovoltaic output and effective irradiance and cell operating temperature under blizzard scenarios is determined to establish the basic weights for different features. At the physical mechanism level, the rationality of the weights is verified by combining the equipment's operating patterns. For instance, in strong wind scenarios in wind farms, the influence of wind speed on output is greater than that of temperature; therefore, the weight of features related to effective wind speed should be higher than that of features related to temperature. For high-temperature scenarios in photovoltaic power plants, the output ceiling constraint mechanism caused by irradiance saturation is stronger than that of high-temperature thermal decay; therefore, the weight of irradiance-related instantaneous feature values ​​is slightly higher than that of temperature-related derived dimensions. In cold wave scenarios in wind farms, the physical priority of low-temperature-triggered turbine tripping is higher than that of wind speed changes; therefore, the weight of temperature-corrected instantaneous feature values ​​is higher than that of wind speed-derived dimensions.

[0032] After calculating the instantaneous feature values ​​and determining the coupling weights for the time-series derived dimensions, the process proceeds to the weighted fusion stage, organically integrating the two types of feature dimensions into a dual-input feature vector. The fusion process is not a simple weight superposition, but rather fully considers the synergistic coupling effect between features. Taking a cold wave scenario in a photovoltaic power plant as an example, the effective irradiance intensity after low-temperature correction (instantaneous feature value) and the cumulative irradiance duration (time-series derived dimension) are fused according to weights. Simultaneously, the freeze-thaw cycle-affected temperature (instantaneous feature value) and the temperature drop rate (time-series derived dimension) are weighted and combined to form two sets of core feature vectors, reflecting both the impact of individual features and highlighting their synergistic effect. In a strong wind scenario in a wind farm, the effective wind power density (instantaneous feature value) and wind speed fluctuation coefficient (time-series derived dimension) are fused, as are the unit load-related temperature (instantaneous feature value) and lubricating oil viscosity coefficient (time-series derived dimension). This allows the feature vector to comprehensively cover the dual constraints of wind speed and temperature on unit output under strong winds, avoiding the problem that a single-dimensional feature cannot characterize complex coupling effects.

[0033] The fused dual-input feature vectors still need to undergo final coupling correction and verification to ensure they fit the requirements of subsequent models. Partial least squares regression analysis is used to examine the interaction coefficients of the two core feature dimensions. If one dimension is found to have insufficient explanatory power for output or redundant correlation with another dimension, auxiliary correction parameters such as humidity and snowfall are introduced for secondary adjustments. For example, in a blizzard scenario at a photovoltaic power plant, if the coupling between effective irradiance and cell operating temperature is low, the snow thickness change rate can be introduced as an auxiliary parameter to correct the weight of effective irradiance. In a high-temperature scenario at a wind farm, if the synergy between effective wind speed and winding equivalent temperature is insufficient, the influence of humidity on heat dissipation efficiency can be used to correct the weight of time-series derived dimensions. Finally, feature vectors for all scenarios are normalized to ensure that feature vectors from different extreme scenarios and different equipment types remain consistent on the numerical scale, preserving the physical meaning and correlation logic of each feature while laying a solid foundation for the nonparametric quantile regression model to capture complex nonlinear relationships.

[0034] In some embodiments, historical power data and dual-input feature vectors are input into a nonparametric quantile regression model to obtain extreme scenario probability prediction results. This includes: grouping the dual-input feature vectors and historical power data according to extreme weather scenario types, dividing them into training and validation sets, and performing scenario-based normalization on the dual-input feature vectors; using a Gaussian kernel function to perform weighted mapping on sample points in the training and validation sets to fit the nonlinear coupling relationship between the dual-input feature vectors and new energy power; using a k-fold cross-validation algorithm based on the bandwidth parameter optimization module, iteratively optimizing the bandwidth parameter by minimizing the quantile loss function on the validation set, whereby the quantile loss function quantifies the deviation between the predicted and actual values ​​at different quantile levels; introducing an extreme scenario penalty coefficient to adjust the fitting weights for extreme low-value and extreme high-value quantiles; inputting the dual-input feature vectors for the prediction period into the optimized nonparametric quantile regression model, outputting a multi-level quantile prediction sequence for new energy power, and quantifying the extreme scenario risk probability based on the comparison between the quantile prediction sequence and a preset threshold.

[0035] In this embodiment of the invention, the primary preparation before model training is to perform refined grouping and standardization of the data, laying a solid foundation for subsequent modeling. Based on different extreme weather scenarios such as blizzards, cold waves, high temperatures, and strong winds, the constructed dual-input feature vectors and corresponding historical power data are classified and grouped to ensure that each group focuses on the characteristics of a single extreme scenario, avoiding model learning bias caused by mixing data from different scenarios. On this basis, each group of data is divided into a training set and a validation set according to a scientific ratio. The training set is used for model parameter fitting, while the validation set is specifically used to verify the model's generalization ability and prediction accuracy. Simultaneously, scenario-based normalization is performed to address the differences in the numerical distribution of feature data under different extreme scenarios. By adjusting the feature numerical scale to adapt to the distribution range of data in each scenario, the relative differences and physical meaning between features are preserved, while eliminating the interference of inconsistent numerical magnitudes under different scenarios on model fitting, ensuring uniform numerical consistency in the input data.

[0036] After data preprocessing, the model uses a Gaussian kernel function to accurately fit the complex nonlinear coupling relationship between the dual-input features and the renewable energy power. The core advantage of the Gaussian kernel function lies in its ability to flexibly weight mapping based on the similarity between samples, without requiring a pre-defined fixed correlation form. For each sample in the training and validation sets, the kernel function calculates the similarity between the feature vector to be fitted and the feature vectors of historical samples. Samples with higher similarity are assigned greater weights, allowing the model to prioritize learning from historical extreme weather experiences similar to the current scenario during the learning process. This weighting method adaptively captures the non-steady-state correlation between meteorological characteristics and power output under extreme weather conditions. Whether it's the nonlinear power decay of a photovoltaic power station during blizzards or the overload response of a wind farm in strong winds, both can be meticulously characterized through the flexible mapping of the kernel function, effectively avoiding the limitations of traditional linear or fixed-parameter models that cannot adapt to complex correlations.

[0037] To further enhance the robustness of the model's fit, the bandwidth parameter optimization module iteratively optimizes the core parameters using a k-fold cross-validation algorithm. This module first determines a reasonable search range for the bandwidth parameter based on the sparsity of extreme weather samples with dual-input features. Then, it divides the training set into multiple mutually exclusive subsets and conducts multiple rounds of training and validation by rotating different subsets as validation data. In each iteration, the bandwidth parameter is adjusted with the goal of minimizing the quantile loss function value on the validation set. This loss function accurately quantifies the deviation between predicted and true values ​​at different quantile levels, ensuring that the optimization process considers both overall prediction accuracy and the fitting effect of each quantile. Through iterative adjustments via multiple rounds of cross-validation, the bandwidth parameter reaches its optimal state, effectively balancing the model's fitting accuracy for local sample features with its global generalization ability, avoiding overfitting caused by the scarcity of extreme weather samples and drastic data fluctuations.

[0038] To address the core need for risk quantification in extreme scenarios, the model introduces an extreme scenario penalty coefficient to specifically strengthen the fitting weight of extreme quantiles. Extremely low quantiles are directly associated with the risk of "sudden power drop," while extremely high quantiles correspond to the potential for "power overload." The prediction accuracy of these two types of quantiles directly determines the reliability of the risk assessment. The penalty coefficient is not a fixed value but is dynamically adjusted based on the intensity level of extreme weather, the operational risk threshold of renewable energy equipment, and the frequency of extreme events in historical samples. When the model's prediction deviation for extreme quantiles exceeds an acceptable range, the penalty coefficient automatically amplifies the loss value corresponding to this deviation, guiding the model to focus on correcting the fitting error of extreme quantiles in subsequent training. This ensures that the model can prioritize capturing the extreme features at the tail end of the renewable energy power distribution, avoiding over-focusing on intermediate quantiles and neglecting key information about extreme scenarios.

[0039] After completing model training and parameter optimization, the actual prediction phase begins. First, meteorological data for the prediction period is acquired. Following the same logic and process as the previous dual-input feature vector construction, the dual-input feature vector for the prediction period is obtained, ensuring that the feature format and dimensions of the input model remain consistent with those in the training phase, avoiding prediction bias caused by differences in feature construction logic. Subsequently, this feature vector is input into an optimized nonparametric quantile regression model. Based on the nonlinear correlation patterns learned during training, and combined with the optimization results of bandwidth parameters and extreme scenario penalty coefficients, the model automatically outputs prediction sequences containing multiple quantile levels, covering different output ranges such as extreme low values, intermediate levels, and extreme high values, comprehensively reflecting the possible fluctuation range of renewable energy power within the prediction period.

[0040] Based on the multi-level quantile prediction sequence output by the model, it is necessary to combine it with preset thresholds to complete the quantitative calculation of the risk probability of extreme scenarios. These preset thresholds are not a uniform standard, but are calibrated according to the safety constraints of power grid dispatching and the operation and maintenance management standards of new energy power plants, and are tailored to different equipment types and extreme weather scenarios. For example, the critical threshold for "sudden drop in output" in a blizzard scenario for photovoltaic power plants is lower than that in a high-temperature scenario, while the threshold for "output overload" in a strong wind scenario for wind farms is higher than that in a cold wave scenario. By comparing the extreme low-value quantiles with the critical threshold for "sudden drop in output" and the extreme high-value quantiles with the control threshold for "output overload," and combining this with the distribution characteristics of the quantile sequence, the probability of occurrence of various extreme events such as "output below the critical value" and "output exceeding the threshold" can be accurately calculated, transforming abstract prediction results into intuitive and quantifiable risk indicators.

[0041] In the process of risk probability quantification, a historical data calibration mechanism is also needed to ensure the statistical reliability of the quantification results. By retrieving historical samples of extreme weather events similar to the current prediction scenario, the degree of agreement between the predicted risk probabilities and the actual occurrence of extreme events in the historical samples is compared, and a calibration coefficient is calculated. If the predicted risk probability for a certain scenario has historically been generally higher than the actual probability of occurrence, the current predicted risk probability is lowered using the calibration coefficient; conversely, it is appropriately raised to avoid distortion in risk quantification caused by sample bias or model fitting bias. This calibration mechanism can effectively correct the deviation between model predictions and actual scenarios, making the risk probability closer to reality and providing a more reliable quantitative basis for decision-making.

[0042] Finally, the quantified extreme scenario risk probabilities are integrated with the multi-level quantile prediction sequence to form a complete prediction output. The output not only includes specific risk probability values, but also risk level assessments, such as intuitive descriptions like "extremely low risk," "high risk," and "extremely high risk," while also indicating the comparison between the predicted values ​​of key quantiles and preset thresholds.

[0043] In some embodiments, according to the bandwidth parameter optimization module, a k-fold cross-validation algorithm is used to iteratively optimize the bandwidth parameter based on minimizing the quantile loss function on the validation set. This includes: determining the initial search interval for the bandwidth parameter based on the dimension of the dual-input feature vector and the statistical characteristics of historical samples of extreme weather scenarios; dividing the training set corresponding to each extreme weather scenario into k mutually exclusive subsets; and iteratively training and validating the model through k-fold cross-validation; defining quantile loss functions for different quantile levels; wherein, for extreme low-value quantiles and extreme high-value quantiles, a scenario adaptation coefficient is introduced to correct the quantile loss function to adjust the model's fitting weights for extreme features; within the bandwidth parameter search interval, minimizing the average loss value after correction at each quantile level is the objective to determine the optimal bandwidth parameter, and performing robustness verification on the optimal bandwidth parameter; if the verification result meets a preset threshold, the bandwidth parameter is determined as the final optimization result; if it does not meet the threshold, the search interval is narrowed and iterative optimization is performed again until the preset threshold is met.

[0044] In this embodiment of the invention, the iterative optimization of the bandwidth parameter begins with a scientific initial setup and data splitting, establishing a reasonable framework for the subsequent optimization process. First, combining the dimensionality of the dual-input feature vectors with statistical information such as the quantity and distribution density of historical samples under different extreme weather scenarios, the initial search interval for the bandwidth parameter is determined. For extreme scenarios with sparse samples, such as rare extreme blizzards or super cold waves, the search interval is appropriately widened to ensure coverage of the optimal parameter range; for scenarios with relatively abundant samples, the initial interval is narrowed to improve optimization efficiency. Simultaneously, the training set corresponding to each extreme weather scenario is evenly divided into k mutually exclusive subsets. During subset division, the temporal continuity of the data is strictly maintained to avoid disrupting the development of extreme weather events, providing a structurally sound and representative data foundation for subsequent k-fold cross-validation training and validation.

[0045] The core of cross-validation lies in comprehensively testing the suitability of parameters through multiple rounds of iterative training and validation. In each round of validation, k-1 subsets are selected as the training set for model training, and the remaining subset is used as the validation set to evaluate model performance. This process is repeated, rotating the roles of the subsets, until all subsets have completed one validation. This iterative approach makes full use of limited training data, allowing the model to learn and validate repeatedly on different data subsets. This effectively reduces the random errors caused by a single data partition, ensuring that subsequent parameter optimization results are more universal, especially suitable for scenarios with limited extreme weather sample sizes, maximizing the extraction of effective information from the data.

[0046] To accurately quantify the fitting bias at different quantiles, a dedicated quantile loss function needs to be designed for each quantile level. Unlike traditional loss functions that focus on overall error, quantile loss functions can precisely focus on the degree of deviation between predicted and actual values ​​at specific quantiles, providing an independent evaluation standard for the fitting effect of each quantile. For example, for the median quantile reflecting normal output levels, the loss function focuses on characterizing the error within the normal fluctuation range; while for quantiles associated with extreme risks, the loss function needs to have higher sensitivity, laying the groundwork for subsequent weight adjustments and ensuring that the fitting quality of each quantile can be accurately measured.

[0047] To address the core need for quantifying risks in extreme scenarios, a scenario fit coefficient is introduced to specifically modify the loss function for extreme quantiles, thereby strengthening the model's weighting in fitting extreme features. The scenario fit coefficient is not fixed but dynamically adjusted based on the intensity level of extreme weather and the operational risk threshold of new energy equipment. Higher extreme weather intensity and stricter equipment risk thresholds result in a larger fit coefficient. The modified loss function amplifies the loss value corresponding to the fitting deviation of extreme low and high quantiles, allowing the model to prioritize the fitting accuracy of these key quantiles during training. This avoids overemphasizing intermediate quantiles and neglecting core information about extreme scenarios, ensuring the model can accurately capture the extreme features at the tail of the new energy power distribution.

[0048] The selection of the optimal bandwidth parameter aims to minimize the corrected average loss value, and is carried out systematically within a pre-defined search interval. By successively substituting candidate bandwidth parameter values ​​within the interval and using a k-fold cross-validation loop, the average of the corrected loss values ​​for each quantile corresponding to each candidate value is calculated. The candidate value with the smallest average loss value is initially determined as the optimal bandwidth parameter. This process balances the fitting accuracy of the median quantile with the fitting priority of the extreme quantiles through the corrected loss function, ensuring that the selected bandwidth parameter can balance the overall fitting effect with the need to capture extreme features.

[0049] To avoid the randomness of the initially selected optimal bandwidth parameters, rigorous robustness verification is necessary. During verification, the initially determined optimal bandwidth parameters are substituted into the full training set to retrain the model. Then, core evaluation metrics are calculated using an independent validation set, including prediction error at extreme quantiles and stability of fitting accuracy at different quantiles. Simultaneously, small-scale adjustments to the bandwidth parameters are simulated, and the magnitude of changes in model performance is observed. If fine-tuning the parameters leads to significant performance fluctuations, it indicates insufficient robustness of the parameters; if the performance remains stable, it proves that the parameters are well-suited, providing a reliable basis for determining the final parameters.

[0050] After robustness verification, a differentiated iterative optimization process is initiated based on the results: If the verification result meets the preset error threshold and stability requirements, i.e., the extreme quantile prediction error is controlled within an acceptable range and the performance is stable after parameter fine-tuning, then the bandwidth parameter is determined as the final optimization result; if the verification fails, it indicates that the initial search interval may not cover the true optimal parameter. In this case, the search interval is halved, centered on the preliminary optimal parameter, while increasing the step size accuracy of parameter search, and cross-validation and parameter selection are carried out again. This process of narrowing the interval and precise search will be iterated repeatedly until the optimal bandwidth parameter that meets the robustness requirements is found, ensuring that the final parameter enables the model to achieve stable and accurate nonlinear mapping under extreme weather scenarios.

[0051] In some embodiments, an extreme scenario penalty coefficient is introduced to adjust the fitting weights for extreme low-value quantiles and extreme high-value quantiles. This includes: constructing a penalty coefficient classification system based on the intensity level of extreme weather and the operational risk threshold of new energy equipment, and determining the initial value of the extreme scenario penalty coefficient; configuring the penalty coefficients for extreme low-value quantiles and extreme high-value quantiles differently based on equipment type and risk emphasis, and embedding the configured extreme scenario penalty coefficients into the quantile loss function to form a weighted loss function; dynamically calibrating the extreme scenario penalty coefficients based on the prediction error feedback of historical sample data of extreme scenarios to maintain the balance between the prediction errors of extreme quantiles and non-extreme quantiles; calculating the scenario similarity for different extreme weather scenarios of the same equipment type, and transferring the calibrated extreme scenario penalty coefficients to similar scenarios for adjustment based on the scenario similarity.

[0052] In this embodiment of the invention, the construction of the penalty coefficient for extreme scenarios begins with the establishment of a scientific classification system, the core of which is to ensure that the penalty intensity is precisely matched with the degree of extreme weather risk. Firstly, based on meteorological monitoring data, extreme weather is classified into four levels—mild, moderate, severe, and extremely severe—using indicators such as the magnitude of irradiance mutations, the duration of temperature exceeding the critical threshold, and the intensity of wind speed exceeding the cut-out threshold. Simultaneously, referring to technical manuals for new energy equipment, the operational risk thresholds for photovoltaic modules and wind turbines under different operating conditions are clarified, such as the critical temperature for low-temperature degradation of photovoltaic modules and the overload protection threshold for wind turbines. Based on the correspondence between weather intensity levels and equipment risk thresholds, a penalty coefficient classification system is constructed, setting initial penalty coefficients for different levels of scenarios. The higher the intensity of the extreme weather and the stricter the equipment risk threshold, the larger the initial penalty coefficient value, ensuring that the penalty intensity matches the actual risk level from the source.

[0053] Considering the different risk focuses of different equipment types, the penalty coefficients for extreme quantiles need to be configured differently. The core risk of photovoltaic power plants lies in the "sudden drop in output" corresponding to extreme low quantiles, such as power outages caused by blizzards or cold waves. Therefore, the penalty coefficient for extreme low quantiles should be appropriately increased based on the corresponding weather level. The main hidden danger of wind farms is the "output overload" caused by extreme high quantiles, which can easily lead to structural damage to the units in strong winds. Therefore, the penalty coefficient for extreme high quantiles should be increased. This differentiated configuration avoids a "one-size-fits-all" coefficient setting, allowing the penalty intensity to accurately target the core risk points of each piece of equipment and improve the fitting priority for key extreme scenarios.

[0054] After configuring the coefficient differentiation, it is embedded into the quantile loss function to form a weighted loss function that can accurately guide the model's learning direction. Traditional quantile loss functions assign equal weights to each quantile, making it difficult to meet the special needs of extreme scenarios. However, by embedding penalty coefficients, the loss value corresponding to extreme quantiles is amplified. When the model's prediction of extreme quantiles deviates, the weighted loss function generates a larger loss value, thereby guiding the model to prioritize correcting the fitting error of extreme quantiles during training. This strengthens the ability to capture the tail features of new energy power distribution, allowing the model's learning focus to shift towards core risk scenarios.

[0055] To avoid fitting imbalance caused by a fixed penalty coefficient, dynamic calibration of the penalty coefficient is achieved by relying on the prediction error feedback from historical sample data of extreme scenarios. After each round of model training, the ratio of the prediction error of extreme quantiles to that of non-extreme quantiles is calculated: if the ratio is too large, it indicates that the fitting accuracy of extreme quantiles is insufficient, and the penalty coefficient needs to be appropriately increased; if the ratio is too small, it indicates that the fitting accuracy of non-extreme quantiles is excessively sacrificed, and the penalty coefficient should be decreased. Through this dynamic adjustment mechanism, the balance between the prediction errors of extreme and non-extreme quantiles is maintained, ensuring that the model accurately captures extreme features without losing the fitting accuracy of normal output scenarios.

[0056] To address the challenges of scarce samples and difficult coefficient calibration in certain extreme weather scenarios, a scenario similarity transfer mechanism is introduced to optimize coefficient configuration. By calculating the similarity between different extreme weather scenarios of the same equipment type—for example, the similarity in meteorological parameter distribution and output response patterns between a blizzard scenario and a cold wave scenario for a photovoltaic power station—the optimal penalty coefficient calibrated in scenarios with sufficient samples is transferred to scarce sample scenarios with high similarity. This transfer is not a direct reuse but rather a fine-tuning based on the scenario differences. The smaller the differences, the smaller the adjustment of the transfer coefficients; conversely, the greater the differences, the greater the adjustment, thus solving the problem of inaccurate penalty coefficient calibration in scenarios with scarce samples.

[0057] The penalty coefficients after scene similarity transfer need to be verified and adjusted again to ensure they fit the target scene. A limited number of historical samples from the target scene are selected, and the adjusted penalty coefficients are substituted into the model for testing. The test results are compared to the degree of agreement with actual extreme events. If the prediction error is within an acceptable range, the transfer coefficients are considered suitable. If the error exceeds the threshold, the penalty coefficients are fine-tuned again, taking into account the unique physical mechanisms of the target scene, such as the impact of specific terrain on wind speed in wind farms and the effect of local climate on the temperature of photovoltaic modules. This ultimately forms a precise penalty coefficient system that fits all extreme scenes, providing reliable support for the model to efficiently learn extreme features.

[0058] In some embodiments, the probability of extreme scenario risks is quantified based on the comparison between the quantile prediction sequence and preset thresholds, including: constructing a hierarchical preset threshold system according to the safety standards for the operation of new energy equipment and the requirements of power grid dispatch constraints; associating and mapping the extreme quantiles in the quantile prediction sequence with the hierarchical preset threshold system; calculating the probability of individual risks based on the relative positional relationship between the extreme quantiles and the hierarchical preset thresholds and the fluctuation characteristics of the quantile sequence, and calculating the comprehensive risk probability in combination with the risk priority of extreme weather scenarios; correcting the comprehensive risk probability based on the deviation between historical measured data and predicted data of extreme scenarios, and matching the corresponding risk level according to the corrected comprehensive risk probability to output the quantitative decision result.

[0059] In this embodiment of the invention, risk quantification is predicated on constructing a hierarchical preset threshold system that aligns with actual application needs. The core of this system is to balance the dual requirements of operational safety for new energy equipment and grid dispatch constraints, achieving refined and scenario-based threshold settings. By combining the technical manuals of different equipment such as photovoltaic modules and wind turbines, the tolerance limits of the equipment under extreme weather conditions are clearly defined, such as the minimum stable output threshold for photovoltaic modules and the overload protection threshold for wind turbines, serving as equipment-level safety thresholds. Simultaneously, referencing regional grid supply-demand balance requirements and voltage stability constraints, critical output thresholds to ensure grid safety are determined, such as the minimum power supply guarantee threshold and the maximum power capacity threshold for new energy sources, serving as grid-level constraint thresholds. Based on this, further refinement is achieved according to extreme weather scenario types. For example, the equipment safety threshold for a photovoltaic power station in a blizzard scenario is lower than that in a high-temperature scenario, while the grid constraint threshold for a wind farm in a strong wind scenario is higher than that in a cold wave scenario. Ultimately, a multi-level, scenario-specific preset threshold system is formed, providing accurate judgment criteria for risk probability calculation.

[0060] After constructing the threshold system, it is necessary to establish a mapping between quantile prediction sequences and hierarchical preset thresholds to clarify the risk indications of different quantiles. The quantile prediction sequences output by the model cover multiple levels from extreme low to extreme high values. Extreme low quantiles directly correspond to risks such as "sudden power drop" and need to be mapped to the minimum power supply threshold at the grid level and the minimum stable power output threshold at the equipment level. Extreme high quantiles correspond to risks such as "power overload" and need to be associated with the overload protection threshold at the equipment level and the maximum power acceptance threshold at the grid level. Intermediate quantiles serve as auxiliary references to characterize the overall range of power fluctuations, providing support for determining whether the relative positions of extreme quantiles and thresholds are statistically significant, ensuring that the mapping relationship accurately matches the risk type, and avoiding confusion in the judgment of different risk scenarios.

[0061] Based on the correlation mapping results, the probability of individual risks is calculated by combining the fluctuation characteristics of the quantile series. By analyzing the relative positions of extreme quantiles and their corresponding stratified thresholds (e.g., whether extreme low quantiles are below the grid safety critical threshold and by how much), and combining indicators such as the standard deviation and fluctuation coefficient of the quantile series, the impact of the severity of power fluctuations on the probability of risk occurrence is quantified. The more severe the fluctuation, the higher the uncertainty of the risk probability, requiring statistical methods to correct the probability estimate. For different individual risks such as "sudden power output drop" and "power overload," the corresponding probabilities are calculated separately, and then weights are assigned based on the risk priority of extreme weather scenarios: for example, in a cold wave scenario, the risk priority of "sudden power output drop" for wind farms is higher than that of "power overload," so the former is given a higher weight; in a strong wind scenario, the priority is tilted towards "power overload." Finally, a weighted average is obtained to obtain the comprehensive risk probability, fully reflecting the overall risk level within the scenario.

[0062] To improve the reliability of risk probabilities, a historical data calibration mechanism needs to be introduced to correct prediction biases. Historical extreme weather samples similar to the current prediction scenario are retrieved, and the predicted values ​​of the overall risk probability in these historical samples are compared with the actual occurrence of extreme events to calculate the bias coefficient. If historical predictions are generally higher than the actual probability of occurrence, it indicates that the model tends to overestimate, and the current overall risk probability needs to be adjusted downwards using the bias coefficient. If the predicted value is lower than the actual probability of occurrence, it should be adjusted upwards appropriately to ensure that the risk probability closely matches the risk level of the real scenario. For extreme scenarios with scarce historical samples, the mean of the bias coefficients of similar scenarios is used for correction, while the uncertainty range of the correction is marked to avoid distortion caused by insufficient samples.

[0063] The calibrated overall risk probability will be matched with a preset risk level system, transforming it into intuitive quantitative decision-making results. Based on the grading standards of new energy power plant operation and maintenance response capabilities and grid dispatch emergency plans, the overall risk probability is divided into five levels: extremely low, low, medium, high, and extremely high. Each level corresponds to a clear risk description and response recommendations. For example, "extremely high risk" corresponds to immediate activation of shutdown protection and emergency grid dispatch measures, while "medium risk" corresponds to enhanced equipment inspection and close monitoring of power output changes. The final output not only includes the overall risk probability value and risk level but also includes extreme quantile predictions, detailed breakdowns of individual risk probabilities, and comparison charts with thresholds. This provides a clear and actionable decision-making basis for new energy power plant operation and maintenance teams to formulate defensive measures and for grid dispatch departments to optimize power balance schemes.

[0064] In some embodiments, a Gaussian kernel function is used to perform a weighted mapping on sample points in the training and validation sets to fit the nonlinear coupling relationship between the dual-input feature vector and the renewable energy power. This includes: initializing the core parameters of the Gaussian kernel function based on the dimension of the dual-input feature vector and the numerical distribution characteristics of the sample data under extreme weather scenarios; wherein, the Gaussian kernel function calculates the similarity between samples based on the Euclidean distance between the dual-input feature vector to be predicted and the historical feature sample vector, and adjusts it according to the dynamic optimization value output by the bandwidth parameter optimization module; performing distance weighting on the sample points according to the Gaussian kernel function; combining the differences in physical response mechanisms of different extreme weather scenarios to perform scenario-based correction of the Gaussian kernel function; embedding the corrected Gaussian kernel function into the fitting layer of the nonparametric quantile regression model to construct a mapping function from the dual-input feature vector to different quantiles of renewable energy power, thereby completing the fitting of the nonlinear coupling relationship.

[0065] In this embodiment of the invention, the application of the Gaussian kernel function begins with the scientific initialization of its core parameters. The key to this step is adapting the parameters to the characteristics of the dual-input features and the distribution patterns of extreme weather samples. First, the dimensional composition of the dual-input feature vectors needs to be clearly defined, such as the two-dimensional combination of "irradiance intensity - temperature" in a photovoltaic scenario and the two-dimensional combination of "wind speed - temperature" in a wind power scenario. Simultaneously, the numerical distribution density of the sample data under different extreme scenarios needs to be analyzed—for example, sparse samples in a blizzard scenario and relatively concentrated samples in a high-temperature scenario. Based on this, the basic parameter range of the kernel function is determined. During initialization, the adaptability of the parameters to the feature dimensions and sample distribution is carefully ensured to avoid deviations in subsequent weighted mapping due to initial parameter values ​​deviating from a reasonable range, thus laying the foundation for accurately fitting nonlinear relationships.

[0066] The core similarity calculation logic of the Gaussian kernel function relies on the Euclidean distance between the feature to be predicted and the features of historical samples, while simultaneously adjusting the weight allocation effect by dynamically optimized bandwidth parameters. Euclidean distance intuitively reflects the closeness of two sets of feature vectors in numerical space; the smaller the distance, the more similar the sample scenarios, and the higher the corresponding weight. This allows the model to prioritize historical extreme weather experience that matches the current prediction scenario during fitting. The bandwidth parameter is dynamically adjusted by the optimization module. By changing the bandwidth size, the influence range of the weights is adjusted: a smaller bandwidth allows the model to focus more on locally similar samples, adapting to fine features in extreme scenarios; a larger bandwidth provides a wider weight coverage, helping to improve the model's generalization ability and achieving flexible adaptation of similarity weighting.

[0067] Based on the similarity calculation results, the Gaussian kernel function performs distance-weighted processing on all sample points. This process precisely matches the nonlinear correlation between features and output under extreme weather conditions. For each historical sample in the training and validation sets, a unique weight is assigned based on its similarity to the feature to be predicted. The weight of high-similarity samples is amplified, while the weight of low-similarity samples is weakened or even ignored. This weighting method allows the model to automatically focus on extreme scenario samples that are valuable for the current prediction, effectively filtering out interference from regular weather samples or extreme samples with large differences. This ensures that the subsequent fitting process can accurately capture the unique nonlinear coupling patterns under extreme weather conditions, rather than having the learning effect diluted by massive amounts of irrelevant data.

[0068] Considering the significant differences in the physical response mechanisms of various extreme weather scenarios, the Gaussian kernel function needs to be modified according to specific scenarios to further improve the fitting accuracy. For example, in the photovoltaic blizzard scenario, the coupling between irradiance intensity and temperature on power output exhibits a strong nonlinear decay, so a correction factor is introduced to strengthen the weight of samples in low irradiance and low temperature ranges. In the wind farm strong wind scenario, the critical synergistic triggering characteristics of wind speed and temperature are obvious, so the correction factor is adjusted to highlight the impact of samples with wind speed exceeding the threshold and temperature below the critical value. This scenario-based modification breaks the limitation of the kernel function's generality, allowing the weighted mapping logic to be deeply matched with the physical mechanisms of each scenario, avoiding the problem that a single kernel function cannot adapt to the nonlinear relationships of multiple extreme scenarios.

[0069] Finally, the scenario-adjusted Gaussian kernel function is embedded into the fitting layer of the nonparametric quantile regression model to construct a dedicated mapping function from dual-input features to different quantiles of renewable energy power. This mapping function abandons the drawbacks of traditional models that pre-define fixed correlation forms, and learns autonomously entirely based on weighted sample data: by minimizing the weighted quantile loss, the model adaptively captures the nonlinear correlation between features and output under different extreme scenarios. Whether it is the irradiation saturation-thermal decay coupling in the high-temperature photovoltaic scenario or the low-temperature turbine shedding-wind speed coordination relationship in the cold wave wind power scenario, it can be accurately characterized. This process does not require forcibly setting linear or fixed functional relationships, ultimately achieving efficient and accurate fitting of the complex nonlinear coupling relationship between dual-input features and renewable energy power.

[0070] In some embodiments, the method further includes: determining decision information based on the extreme scenario probability prediction results.

[0071] In this embodiment of the invention, the extreme scenario probability prediction results generated by this method are the core basis for determining subsequent decision-making information. The entire process needs to revolve around the dual goals of safe operation and maintenance of new energy power plants and stable grid dispatch, achieving a deep transformation from data to decision. First, the extreme scenario probability prediction results are analyzed from multiple dimensions. From the output multi-level quantile prediction sequences, individual risk probabilities (such as the probability of "sudden drop in output" and "overload in output"), comprehensive risk probabilities, and risk levels, core information related to decision-making is extracted. This includes the specific time period, duration, and peak range of output fluctuations that extreme output events may occur, as well as the degree of impact of risk events on different types of new energy equipment (photovoltaic modules, wind turbines). At the same time, key constraints of grid dispatch are associated, such as the estimated supply and demand gap of the regional power grid and the redundancy of voltage stability thresholds. Based on this, combined with the experience of handling historical extreme events, the similarity between the current predicted risk and historical cases is analyzed to predict the chain reactions that the risk event may trigger. For example, whether a sudden drop in photovoltaic output will lead to an expansion of the regional power supply gap, or whether wind power overload will cause unit failures and thus affect the stability of the local power grid, providing comprehensive support for the accurate formulation of decision-making information.

[0072] The determination of decision-making information is not a single conclusion output, but rather the formation of a hierarchical and categorized system of precise action guidelines. For extremely high and high-risk scenarios, the decision-making information at the operation and maintenance end needs to clearly define emergency response procedures and key measures. For example, in a high-risk scenario of blizzards at photovoltaic power plants, it is necessary to develop operational plans for preheating modules before snowfall and rapid de-icing after snowfall, clearly defining the deployment routes of de-icing equipment, the division of labor among personnel, and safety protection requirements. In a high-risk scenario of strong winds at wind farms, it is necessary to initiate the unit load reduction operation procedure in advance, reinforce the tower and blade connection points, and clear obstacles around the site that may affect the equipment. The corresponding decision-making information at the grid dispatch end includes reserving sufficient standby capacity (such as standby power for thermal power and energy storage units), adjusting the upper limit for new energy acceptance, optimizing cross-regional power sharing plans, issuing risk warning notices, and coordinating relevant departments to prepare for emergency response. For medium- and low-risk scenarios, decision-making information focuses on routine management and enhanced early warning. On the operations and maintenance side, the frequency of real-time equipment monitoring needs to be increased, with a focus on tracking the correlation between core meteorological parameters and equipment operating status, and optimizing equipment maintenance and inspection plans. On the dispatching side, dynamic monitoring of renewable energy output fluctuations is necessary, along with fine-tuning of power balance schemes to ensure grid redundancy. Furthermore, decision-making information must include effectiveness evaluation indicators, clearly defining the implementation timeframes, expected goals, and termination conditions for each measure and emergency response.

[0073] Figure 2 This is a schematic diagram of the structure of the new energy power extreme scenario prediction device based on dual-input features and nonparametric quantile regression provided in an embodiment of the present invention. Figure 2 As shown, the new energy power extreme scenario prediction device based on dual-input features and nonparametric quantile regression is characterized by comprising: Module 210 is used to acquire historical power data and historical meteorological data; Processing module 220 is used to construct a dual-input feature vector based on historical meteorological data and the physical response mechanism of new energy output under extreme weather conditions; The prediction module 230 is used to input historical power data and dual-input feature vectors into a nonparametric quantile regression model to obtain the probability prediction results of extreme scenarios. Among them, the kernel function of the nonparametric quantile regression model is used to perform nonlinear fitting of the relationship between the dual input features and the renewable energy power; the nonparametric quantile regression model also includes a bandwidth parameter optimization module; the bandwidth parameter optimization module is used to optimize the nonlinear mapping between the dual input features and different fractions of renewable energy power.

[0074] In the above embodiments, the descriptions of each embodiment have their own emphasis. Parts not detailed or described in a particular embodiment can be referred to in the relevant descriptions of other embodiments. Unless otherwise specified or in conflict with logic, the terminology and / or descriptions between different embodiments are consistent and can be referenced interchangeably. Technical features in different embodiments can be combined to form new embodiments based on their inherent logical relationships.

[0075] The above embodiments are only used to illustrate the technical solutions of the present invention, and are not intended to limit it. Although the present invention has been described in detail with reference to the foregoing embodiments, those skilled in the art should understand that modifications can still be made to the technical solutions described in the foregoing embodiments, or equivalent substitutions can be made to some of the technical features. Such modifications or substitutions do not cause the essence of the corresponding technical solutions to deviate from the spirit and scope of the technical solutions of the embodiments of the present invention, and should all be included within the protection scope of the present invention.

Claims

1. A method for predicting extreme scenarios of renewable energy power based on dual-input features and nonparametric quantile regression, characterized in that, include: Acquire historical power data and historical meteorological data; Based on historical meteorological data and the physical response mechanism of new energy output under extreme weather conditions, a dual-input feature vector is constructed. The historical power data and the dual-input feature vector are input into the nonparametric quantile regression model to obtain the extreme scenario probability prediction results; The kernel function of the nonparametric quantile regression model is used to perform nonlinear fitting of the relationship between the dual-input features and the renewable energy power; the nonparametric quantile regression model also includes a bandwidth parameter optimization module; the bandwidth parameter optimization module is used to optimize the nonlinear mapping between the dual-input features and different fractions of renewable energy power.

2. The method for predicting extreme scenarios of new energy power based on dual-input features and nonparametric quantile regression according to claim 1, characterized in that, Based on historical meteorological data and the physical response mechanism of new energy power output under extreme weather conditions, a dual-input feature vector is constructed, including: The historical meteorological data were preprocessed in multiple dimensions and screened using two criteria to obtain core basic parameters and auxiliary correction parameters. Based on the core basic parameters and auxiliary correction parameters, a dual-input feature vector with a coupling correction mechanism is constructed for different extreme weather scenarios and equipment types. The dual-input feature vector includes instantaneous feature values ​​determined based on the device physical equations and derived dimensions determined based on time-series statistical features.

3. The method for predicting extreme scenarios of new energy power based on dual-input features and nonparametric quantile regression according to claim 2, characterized in that, Based on the aforementioned core fundamental parameters and auxiliary correction parameters, a dual-input feature vector with a coupled correction mechanism is constructed for different extreme weather scenarios and equipment types, including: Based on the physical response mechanism under different extreme weather scenarios, the core basic parameters are physically modified to obtain instantaneous characteristic values; Calculate the time-series derived dimension based on the instantaneous feature value and the auxiliary correction parameter; The instantaneous feature value and the time-series derived dimension are weighted and fused according to the coupling weight to construct the dual-input feature vector. The coupling weight is determined based on the statistical distribution or physical mechanism verification of historical sample data in extreme scenarios.

4. The method for predicting extreme scenarios of new energy power based on dual-input features and nonparametric quantile regression according to claim 1, characterized in that, The historical power data and the dual-input feature vector are input into the nonparametric quantile regression model to obtain extreme scenario probability prediction results, including: The dual-input feature vectors and historical power data are grouped according to extreme weather scenario types, divided into training sets and validation sets, and the dual-input feature vectors are subjected to scenario-based normalization processing. A Gaussian kernel function is used to perform a weighted mapping on the sample points in the training set and the validation set to fit the nonlinear coupling relationship between the dual-input feature vector and the new energy power. According to the bandwidth parameter optimization module, the k-fold cross-validation algorithm is used to iteratively optimize the bandwidth parameter based on minimizing the quantile loss function on the validation set. The quantile loss function is used to quantify the deviation between the predicted value and the true value at different quantile levels. An extreme scenario penalty coefficient is introduced to adjust the fitting weights for extreme low and high quantile values; The dual-input feature vectors for the prediction period are input into the optimized nonparametric quantile regression model, which outputs a multi-level quantile prediction sequence of new energy power. Based on the comparison between the quantile prediction sequence and a preset threshold, the probability of extreme scenario risks is quantified.

5. The method for predicting extreme scenarios of new energy power based on dual-input features and nonparametric quantile regression according to claim 4, characterized in that, Based on the bandwidth parameter optimization module, a k-fold cross-validation algorithm is used to iteratively optimize the bandwidth parameter by minimizing the quantile loss function on the validation set, including: Based on the dimension of the dual-input feature vector and the statistical characteristics of historical samples of extreme weather scenarios, the initial search range of the bandwidth parameter is determined, and the training set corresponding to each extreme weather scenario is divided into k mutually exclusive subsets. Training and validation are carried out in a loop through k-fold cross-validation. Quantile loss functions are defined for different quantile levels; for extremely low and extremely high quantiles, a scene adaptation coefficient is introduced to modify the quantile loss function in order to adjust the model's fitting weights for extreme features. Within the bandwidth parameter search range, the optimal bandwidth parameter is determined with the goal of minimizing the average loss value after correction at each quantile level, and the optimal bandwidth parameter is then subjected to robustness verification. If the verification result meets the preset threshold, the bandwidth parameter is determined as the final optimization result. If it does not meet the threshold, the search range is narrowed and iterative optimization is performed again until the preset threshold is met.

6. The method for predicting extreme scenarios of new energy power based on dual-input features and nonparametric quantile regression according to claim 4, characterized in that, An extreme scenario penalty coefficient is introduced to adjust the fitting weights for extreme low and high quantiles, including: Based on the intensity level of extreme weather and the operational risk threshold of new energy equipment, a penalty coefficient classification system is constructed to determine the initial value of the penalty coefficient for extreme scenarios. Based on the equipment type and risk focus, the penalty coefficients for extreme low-value quantiles and extreme high-value quantiles are configured differently, and the configured extreme scenario penalty coefficients are embedded into the quantile loss function to form a weighted loss function; Based on the prediction error feedback from historical sample data of extreme scenarios, the penalty coefficient for extreme scenarios is dynamically calibrated to maintain a balance between prediction errors of extreme quantiles and non-extreme quantiles. For different extreme weather scenarios of the same equipment type, the scenario similarity is calculated, and the penalty coefficient of the calibrated extreme scenario is transferred to the similar scenario for adjustment based on the scenario similarity.

7. The method for predicting extreme scenarios of new energy power based on dual-input features and nonparametric quantile regression according to claim 4, characterized in that, Based on the comparison between the predicted quantile sequence and a preset threshold, the probability of risk in extreme scenarios is quantified, including: Based on the safety standards for the operation of new energy equipment and the constraints of power grid dispatch, a hierarchical preset threshold system is constructed. Associate and map the extreme quantiles in the quantile prediction sequence with the hierarchical preset threshold system; Based on the relative positional relationship between the extreme quantiles and the stratified preset thresholds, as well as the fluctuation characteristics of the quantile sequence, the probability of a single risk is calculated, and combined with the risk priority of extreme weather scenarios, the overall risk probability is calculated. Based on the deviation between historical measured data and predicted data in extreme scenarios, the comprehensive risk probability is corrected, and the corresponding risk level is matched according to the corrected comprehensive risk probability to output a quantitative decision result.

8. The method for predicting extreme scenarios of new energy power based on dual-input features and nonparametric quantile regression according to claim 4, characterized in that, A Gaussian kernel function is used to perform a weighted mapping on the sample points in the training and validation sets to fit the nonlinear coupling relationship between the dual-input feature vector and the new energy power, including: Based on the dimension of the dual-input feature vector and the numerical distribution characteristics of sample data under extreme weather scenarios, the core parameters of the Gaussian kernel function are initialized; wherein, the Gaussian kernel function calculates the similarity between samples based on the Euclidean distance between the dual-input feature vector to be predicted and the historical feature sample vector, and adjusts it according to the dynamic optimization value output by the bandwidth parameter optimization module; The sample points are weighted by distance according to the Gaussian kernel function; Based on the differences in the physical response mechanisms of different extreme weather scenarios, the Gaussian kernel function is modified according to the specific scenarios. The modified Gaussian kernel function is embedded into the fitting layer of the nonparametric quantile regression model to construct a mapping function from the dual-input feature vector to different quantiles of new energy power, thus completing the fitting of the nonlinear coupling relationship.

9. The method for predicting extreme scenarios of new energy power based on dual-input features and nonparametric quantile regression according to claim 7, characterized in that, The method further includes: Decision information is determined based on the probability prediction results of extreme scenarios.

10. A new energy power extreme scenario prediction device based on dual-input features and nonparametric quantile regression, characterized in that, include: The acquisition module is used to acquire historical power data and historical meteorological data; The processing module is used to construct a dual-input feature vector based on historical meteorological data and the physical response mechanism of new energy output under extreme weather conditions; The prediction module is used to input the historical power data and the dual-input feature vector into the nonparametric quantile regression model to obtain the extreme scenario probability prediction result; The kernel function of the nonparametric quantile regression model is used to perform nonlinear fitting of the relationship between the dual-input features and the renewable energy power; the nonparametric quantile regression model also includes a bandwidth parameter optimization module; the bandwidth parameter optimization module is used to optimize the nonlinear mapping between the dual-input features and different fractions of renewable energy power.

Images