Risk indicator construction method and system for new energy output prediction

Abstract

The invention discloses a risk indicator construction method and system for new energy output prediction, and belongs to the technical field of data processing and power system management, and the method comprises the steps: obtaining the preprocessing data of a plurality of new energy stations, carrying out the spatial-temporal feature analysis, and generating a spatial-temporal feature set; obtaining meteorological prediction data, and performing medium and long term output prediction through an extreme gradient boosting tree based on the space-time feature set; comparing the output prediction result with historical real output data to obtain a prediction error, and performing probability coupling modeling and quantile regression prediction in combination with a meteorological element evolution sequence to generate a dynamic quantile value; and determining a dynamic risk interval, and constructing a graded early warning index. According to the method, probabilistic coupling modeling and quantile regression prediction are adopted, prediction errors are deeply associated with dynamic evolution of meteorological elements, predicted uncertain risks can be quantized and graded, and decision support is provided for optimal scheduling and risk management of a power grid.

Description

TECHNICAL FIELD

[0001] The present application relates to the technical field of data processing and power system management, and particularly relates to a risk index construction method and system for new energy output prediction. BACKGROUND

[0002] With the transformation of global energy structure, the penetration rate of new energy represented by wind and solar energy in power systems is increasing. Due to the natural intermittency, volatility and randomness of new energy output, its large-scale grid connection brings great challenges to the planning, dispatching and safe and stable operation of power systems. Therefore, accurate prediction of new energy output and effective management of the uncertainty of prediction have become a crucial link in modern power system operation and management, and are key management means to ensure the reliability and economy of power supply.

[0003] In the prior art, new energy output prediction usually adopts a method based on numerical weather prediction and machine learning model to provide output prediction values for a future period of time to the power grid dispatching center. The power grid dispatching personnel mainly rely on these deterministic point prediction results to develop conventional unit commitment, economic dispatching and reserve capacity plan. For example, the invention patent disclosed as CN118035806A classifies the prediction scenarios based on numerical weather prediction. In terms of risk management, a relatively fixed strategy is usually adopted, such as reserving a certain proportion of installed capacity as a reserve, or uniformly configuring a margin according to the average error range of historical statistics, which lacks differentiated consideration of risk levels under different conditions.

[0004] However, the existing technical solutions have obvious deficiencies. First, the provided deterministic point prediction values cannot reflect the uncertainty degree of the prediction results themselves, and the dispatching personnel cannot know the deviation range and probability of the prediction results that may occur under specific weather conditions, resulting in blindness in risk management. Secondly, the traditional risk response strategy is usually static and one-size-fits-all, and cannot be dynamically adjusted according to the real-time changing weather conditions and prediction reliability, which may lead to waste of reserve resources in low-risk periods and insufficient reserve in high-risk periods, affecting the economy and safety of the system. In addition, the analysis of prediction errors mostly stays at the macro statistical level and fails to deeply explore the internal correlation between error distribution and specific weather elements, so as to form a specific scene-oriented and refined risk warning. SUMMARY

[0005] To address the problems existing in the prior art, this invention provides a method for constructing risk indicators for new energy power output forecasting. It employs probabilistic coupling modeling and quantile regression forecasting to deeply correlate forecasting errors with the dynamic evolution of meteorological elements. This method can quantify and classify the uncertainty risk of forecasting, providing decision support for the optimized scheduling and risk management of the power grid. This invention also provides a system for constructing risk indicators for new energy power output forecasting.

[0006] The above objectives can be achieved through the following approach: A method for constructing risk indicators for predicting new energy output, the method comprising: Acquire historical power output data and meteorological data from multiple renewable energy power stations, perform spatiotemporal coupled data cleaning and missing value repair, and generate preprocessed data; Dynamic descriptive statistics of the power output sequence of each station in the preprocessed data are used to generate dynamic probability distribution features; time-delay cross-correlation coefficients between different new energy power stations are calculated, and spatial transmission features are generated by combining geographical location information; wavelet decomposition is performed on the power output time series in the preprocessed data to extract trend, seasonal and transient event components, generate multi-scale time series features, and finally generate a spatiotemporal feature set. Acquire meteorological forecast data, and based on the spatiotemporal feature set and the meteorological forecast data, perform medium- and long-term power output prediction using the extreme gradient boosting tree prediction method to generate power output prediction results; The predicted output results are compared with historical actual output data to obtain the prediction error. Then, combined with the evolution sequence of meteorological elements in the meteorological prediction data, probabilistic coupling modeling and quantile regression prediction are performed to generate dynamic quantile values. Based on the dynamic quantile values, dynamic risk intervals are determined, thereby constructing a tiered early warning indicator.

[0007] Optionally, the step of performing spatiotemporally coupled data cleaning and missing value repair to generate preprocessed data includes: Outlier correction is performed on the original historical power output data and meteorological data using a spatiotemporal coupled median filtering method. The spatiotemporal coupled median filtering method uses a sliding window in the spatiotemporal dimension, moving point by point on the data sequence. When replacing the data at the center point of the window, the median of all data points in this spatiotemporal neighborhood is referenced. Based on the preliminary cleaned data, a weighted k-dimensional tree algorithm based on spatiotemporal similarity is used to impute missing values, resulting in a complete dataset. The Z-score detection method with dynamic adaptive threshold is applied to the complete dataset to identify and process outliers, and the outliers are replaced to generate the final preprocessed data. The adaptive threshold is calculated in real time based on the specific situation of the current data point and the power output statistical fluctuation characteristics of the station or region in the same historical period.

[0008] Optionally, the application of a weighted k-dimensional tree algorithm based on spatiotemporal similarity for missing value imputation includes: For each time point containing missing values, a spatiotemporal weighted distance metric function is used to find several neighboring data points that are most similar in spatiotemporal dimension and meteorological background. The neighboring data points represent the data of the physical scenario that has the most similar meteorological, geographical and temporal background to the current moment in history. By performing a weighted average or interpolation calculation on the missing dimension values ​​corresponding to these neighboring data points, the current missing values ​​can be estimated and filled.

[0009] Optionally, the step of generating power output prediction results by performing medium- to long-term power output prediction using the extreme gradient boosting tree prediction method includes: A forward feature selection method based on risk contribution is used to filter features from the spatiotemporal feature set and the meteorological forecast data to generate an optimized feature set; Based on the optimized feature set, the hyperparameters of the extreme gradient boosting tree prediction method are optimized using the Bayesian optimization method. After several iterations, the optimal or near-optimal parameter combination is found with a higher probability. Finally, the parameter combination that achieves the best cross-validation performance is selected as the optimized parameter. The meteorological forecast data is processed using the optimized parameters, and an output forecast operation is performed to generate an output forecast result.

[0010] Optionally, the step of performing the power output prediction operation and generating the power output prediction result includes: A hierarchical clustering algorithm based on spatiotemporal propagation characteristics is used to divide multiple new energy power stations into multiple sub-regions based on the aforementioned spatial propagation characteristics, generating regional division results; For each sub-region of the region division result, a reference station is selected, and the dynamic transfer operator between the non-reference station and the reference station is analyzed and extracted through a gated recurrent unit network with attention mechanism to generate spatiotemporal transfer features. The spatiotemporal transfer features are nonlinearly aggregated by a multilayer perceptron with a hybrid density network as the output layer to form a nonparametric probability distribution of regional aggregated power, generate regional power output prediction results, and serve as the power output prediction results.

[0011] Optionally, the step of combining the evolution sequence of meteorological elements in the meteorological forecast data to perform probabilistic coupling modeling and quantile regression prediction to generate dynamic quantile values ​​includes: Calculate the mean, standard deviation, and extreme values ​​of the prediction error to generate error statistical characteristics; Based on the aforementioned error statistical characteristics and the evolution sequence of meteorological elements, a joint distribution model is constructed using a probabilistic coupling modeling method based on Copula functions. Based on the evolution sequence of the meteorological elements and the features extracted by the joint distribution model, dynamic quantile values ​​of the prediction error are generated by the quantile regression prediction method.

[0012] Optionally, the probabilistic coupling modeling method based on Copula functions constructs a joint distribution model, including: Construct a histogram of prediction error, observe the skewness of the prediction error probability distribution through the histogram, analyze the positive and negative errors in the prediction by combining the average value, and then select a non-parametric fitting method to fit the marginal probability distribution function of the prediction error. For each meteorological element in the meteorological element evolution sequence, a non-parametric fitting method is also used to obtain their respective marginal probability distribution functions; Given meteorological elements and prediction error data, a Copula function is selected and trained to fuse all the above marginal probability distributions to obtain a joint distribution model, thereby estimating its joint probability density.

[0013] Optionally, the construction of the hierarchical early warning indicators includes: The dynamic quantile values ​​are combined to calculate the risk intervals corresponding to different confidence levels, forming a set of risk intervals; Establish a mapping relationship between the set of risk intervals and risk levels to generate a dynamic risk level mapping; Based on the dynamic risk level mapping, different early warning levels and their triggering conditions are defined, and a hierarchical early warning index is constructed.

[0014] Based on the same inventive concept, this invention also provides a risk indicator construction system for predicting new energy power output, which is applied to the aforementioned method for constructing risk indicators for predicting new energy power output. The system includes: The data preprocessing module is used to acquire historical power output data and meteorological data from multiple new energy power stations, perform spatiotemporal coupled data cleaning and missing value repair, and generate preprocessed data. The spatiotemporal feature extraction module is used to generate dynamic probability distribution features from the dynamic descriptive statistics of the power output sequences of each station in the preprocessed data; calculate the time-delay cross-correlation coefficients between different new energy power stations and generate spatial transmission features by combining geographical location information; and perform wavelet decomposition on the power output time series in the preprocessed data to extract trend, seasonal and transient event components, generate multi-scale time series features, and finally generate a spatiotemporal feature set. The power output prediction module is used to acquire meteorological forecast data and, based on the spatiotemporal feature set and the meteorological forecast data, perform medium- and long-term power output prediction using the extreme gradient boosting tree prediction method to generate power output prediction results. The error probability modeling module is used to compare the power output prediction results with historical actual power output data to obtain the prediction error, and combine the evolution sequence of meteorological elements in the meteorological prediction data to perform probability coupling modeling and quantile regression prediction to generate dynamic quantile values. The early warning indicator construction module is used to determine the dynamic risk range based on the dynamic quantile value and construct a graded early warning indicator.

[0015] Compared with the prior art, the present invention has the following advantages: This invention improves the quality and feature dimensionality of input data through spatiotemporal coupled data cleaning and repair, as well as deep feature extraction of dynamic probability distributions, spatial propagation, and multi-scale time series. This provides a solid physical foundation for subsequent hierarchical aggregation prediction frameworks based on hybrid density networks, enabling them to generate non-parametric probabilistic prediction results that more closely resemble real physical processes. This improves the accuracy and robustness of medium- and long-term power output predictions and provides a benchmark for risk assessment.

[0016] This invention innovatively combines the Copula function with quantile regression to construct a dynamic coupling model between prediction error and the evolution sequence of meteorological elements. This model accurately captures the complex nonlinear and asymmetric dependence between the two and directly predicts the error quantiles that dynamically change with meteorological conditions. This method achieves a shift from traditional static, macroscopic error statistics to dynamic, refined risk quantification tailored to specific meteorological scenarios, making the assessment of prediction uncertainty more forward-looking and accurate, and particularly enhancing the ability to identify risks under extreme weather events.

[0017] This invention constructs a complete hierarchical early warning indicator and automated dispatch response closed-loop system. It transforms complex, continuous probabilistic risk information into discrete, hierarchical early warning signals that are easy for grid dispatchers to understand and execute through dynamic risk interval mapping. This enables the grid to adopt differentiated strategies precisely matched to the degree of risk based on the early warning level, achieving precise and economical allocation of flexible resources such as reserves. While effectively ensuring the safe and stable operation of the grid, it also improves the level of new energy consumption and the overall system operating efficiency.

[0018] Other features and advantages of the invention will be set forth in the description which follows, and will be apparent in part from the description, or may be learned by practicing the invention. The objects and other advantages of the invention may be realized and obtained by means of the structures pointed out in the description, claims and drawings. Attached Figure Description

[0019] To more clearly illustrate the technical solutions in the embodiments of the present invention or the prior art, the drawings used in the description of the embodiments or the prior art will be briefly introduced below. Obviously, the drawings described below are some embodiments of the present invention. For those skilled in the art, other drawings can be obtained based on these drawings without creative effort.

[0020] Figure 1 This is a flowchart illustrating a method for constructing risk indicators for predicting new energy output according to an embodiment of the present invention. Figure 2 This is a schematic diagram of wavelet decomposition according to an embodiment of the present invention; Figure 3 This is a flowchart of the spatiotemporal feature set generation method according to an embodiment of the present invention; Figure 4 This is a flowchart of the method for generating output prediction results according to an embodiment of the present invention; Figure 5 This is a flowchart of the method for performing output prediction operation according to an embodiment of the present invention; Figure 6 This is a graph showing the relationship between meteorological elements and error statistical characteristics according to an embodiment of the present invention; Figure 7 This is a flowchart illustrating another method for constructing risk indicators for predicting new energy output according to an embodiment of the present invention; Figure 8 This is a schematic diagram of a risk indicator construction system for predicting new energy output according to an embodiment of the present invention. Detailed Implementation

[0021] To make the objectives, technical solutions, and advantages of the embodiments of the present invention clearer, the technical solutions of the embodiments of the present invention will be clearly and completely described below with reference to the accompanying drawings. Obviously, the described embodiments are only some embodiments of the present invention, not all embodiments. Based on the embodiments of the present invention, all other embodiments obtained by those skilled in the art without creative effort are within the scope of protection of the present invention.

[0022] Reference Figure 1 One embodiment of the present invention proposes a method for constructing risk indicators for predicting new energy output. It adopts probabilistic coupling modeling and quantile regression prediction to deeply correlate prediction errors with the dynamic evolution of meteorological elements. This method can quantify and classify the uncertainty risk of prediction, providing decision support for the optimized scheduling and risk management of the power grid.

[0023] The method described in this embodiment specifically includes the following steps: S1 acquires historical power output data and meteorological data from multiple renewable energy power stations, performs spatiotemporal coupling data cleaning and missing value repair, and generates preprocessed data.

[0024] Optionally, acquiring historical power output data and meteorological data from multiple renewable energy power plants includes: Based on the installation location and power generation type of the new energy power stations, several typical areas or new energy power stations with different installation locations were selected to obtain historical power output data and meteorological data of the new energy power stations.

[0025] Specifically, the first step is to sort and classify all new energy power plants within the grid dispatch range. The classification is primarily based on two dimensions: power generation type and installation location. Power generation type clearly distinguishes power plants from different energy sources, such as wind power plants and photovoltaic power plants, because their output characteristics and dependence on meteorological factors are fundamentally different. Installation location further refines the impact of the geographical environment on output, and can be classified according to macro-geographical units and micro-topography.

[0026] For example, wind farms can be categorized into mountainous, plain, coastal, or plateau types. These different geographical environments exhibit distinct regional meteorological characteristics, such as wind resource distribution and sunlight conditions. After this classification, a "meteorological topology network" is introduced to select wind farms. For instance, when selecting wind farms, in addition to considering inland and coastal areas, a focus is placed on selecting a series of farms at specific geographical intervals along major monsoon paths and weather system movement directions. The aim is to directly observe the temporal transmission and intensity attenuation of power output changes caused by meteorological system movement between different farms, providing the most direct data support for subsequent model learning of the spatial propagation patterns of uncertainty. In this way, multiple selected renewable energy farms in typical regions or at different installation locations collectively constitute an observation matrix with a specific topological structure and physical meaning, rather than a simple sample set. Subsequently, for each selected renewable energy farm, its long-term historical power output data and synchronous meteorological data are collected. Historical power output data typically represents the actual active power transmitted from the farm to the grid, with a time resolution that meets prediction requirements, such as 15 minutes or 1 hour. Meteorological data should include the factors most directly affecting power output. For wind power plants, key meteorological data include wind speed and direction at hub height; for photovoltaic power plants, it includes total horizontal irradiance, direct radiation irradiance, diffuse radiation irradiance, and ambient temperature. This data can be obtained from on-site monitoring equipment or high-quality numerical weather prediction reanalysis data, ensuring strict timestamp alignment between power output data and meteorological data. This data acquisition strategy based on meteorological topology networks ensures that the input data source not only reflects the power output behavior and fluctuations of individual plants but also reveals the intrinsic correlation and spatiotemporal transmission characteristics of the collective response of new energy power generation to meteorological changes, thus laying a solid physical and data foundation for subsequent spatiotemporal feature analysis and extraction.

[0027] The process of generating preprocessed data includes the following steps: The spatiotemporal coupled median filtering method is applied to the historical power output data and the meteorological data to correct outliers and generate preliminary cleaned data. The preliminary cleaned data is then processed using a weighted k-dimensional tree algorithm based on spatiotemporal similarity to fill in missing values ​​and generate complete data. The Z-score detection method with dynamic adaptive threshold is applied to the complete data to identify and process outliers, generating preprocessed data.

[0028] Specifically, firstly, a spatiotemporal coupled median filtering method is applied to correct outliers in the original historical power output data and meteorological data. The core idea of ​​this method is to use a sliding window that extends to the spatiotemporal dimension. As the window moves point-by-point across the data sequence, it includes not only the immediate neighbors of the data point to be processed in the time series, but also synchronous data points from multiple geographically adjacent stations at the same time. When replacing the value of the window's center point, the median of all data points within this spatiotemporal neighborhood is referenced. This design enables the method to intelligently distinguish between isolated outliers caused by momentary sensor malfunctions or communication errors, and collective real power output fluctuations triggered by the passage of regional weather systems. Therefore, while effectively filtering out isolated impulse noise, it can retain, to the maximum extent, the spatially correlated data edge and step information driven by real physical phenomena, avoiding the ambiguity effects and misjudgments of regional extreme events that may arise from traditional linear filtering. The processed data sequence, referred to as the pre-cleaned data, exhibits significantly improved smoothness and continuity, while retaining key collective response characteristics.

[0029] Secondly, based on the preliminarily cleaned data, a weighted k-dimensional tree algorithm based on spatiotemporal similarity is applied to impute missing values. Due to equipment maintenance or communication network interruptions, data sequences often contain continuous or discrete missing segments. The k-dimensional tree algorithm is a data structure used for fast nearest neighbor search in high-dimensional space. In this application, the multidimensional data at each time point, including the station's historical output data and multiple related meteorological data, as well as synchronous data that may come from other stations, are considered as points in a high-dimensional space. First, a k-dimensional tree is constructed using all data points without missing values.

[0030] Then, for each time point containing missing values, a spatiotemporal weighted distance metric is used in this high-dimensional space to find N nearest neighbor data points that are most similar in spatiotemporal dimension and meteorological context. This weighted distance metric assigns higher weights to historical data points with closer timestamps, smaller geographical distances, and better matching of key meteorological elements. These neighbor points represent the physical scenarios in history that have the most similar meteorological, geographical, and temporal context to the current moment. Then, by performing a weighted average or other interpolation calculation on the missing dimension values ​​corresponding to these N neighbor points, the current missing values ​​are estimated and filled in. After this process, the resulting dataset is the complete dataset.

[0031] Finally, a Z-score detection method with dynamic adaptive thresholding is applied to the complete data to identify and process outliers. Outliers are numerical points that statistically deviate significantly from the overall distribution of the data set. They may be true values ​​under extreme weather events or potential errors that have not been filtered out. A dynamic adaptive thresholding mechanism is employed, where the threshold is calculated in real time based on the specific context of the current data point and the power output statistical fluctuation characteristics of the station or region during the same historical period. For example, during the afternoon windy period in the windy season, a higher threshold is automatically applied to accommodate normal high power output fluctuations; while during windless nights or low light periods in the early morning, a lower and stricter threshold is applied. This adaptive mechanism enables outlier detection to intelligently distinguish between real extreme power output events and data acquisition errors, effectively avoiding the misjudgment and rejection of valuable extreme operating condition data by traditional "one-size-fits-all" methods. For identified outliers, a replacement strategy can be adopted, such as replacing them with a preset physical upper limit value or a reasonable value calculated based on neighboring data points, thereby generating the final preprocessed data. This method is not just a simple numerical removal and filling, but an organic combination of spatiotemporal coupled filtering, weighted nearest neighbor repair and adaptive outlier detection. It can improve the integrity, accuracy and consistency of the data while preserving the real physical information and spatiotemporal correlation contained in the data.

[0032] S2 generates dynamic probability distribution features from the dynamic descriptive statistics of the power output sequence of each station in the preprocessed data; calculates the time-delay cross-correlation coefficient between different new energy power stations and generates spatial transmission features by combining geographical location information; and performs wavelet decomposition on the power output time series in the preprocessed data to extract trend, seasonal and transient event components, generate multi-scale time series features, and finally generate a spatiotemporal feature set.

[0033] In this embodiment, optionally, as follows: Figure 3 As shown, the generation of the spatiotemporal feature set includes the following steps: S21 calculates the dynamic descriptive statistics of the power output sequence of each station in the preprocessed data and generates dynamic probability distribution characteristics; S22 calculates the time-delay cross-correlation coefficients between different new energy power stations and, combined with geographical location information, generates spatial transmission characteristics; S23 performs wavelet decomposition on the output time series in the preprocessed data to extract trend, seasonal and transient event components, and generate multi-scale time series features; S24 integrates the dynamic probability distribution features, the spatial transmission features, and the multi-scale time series features to generate a spatiotemporal feature set.

[0034] Specifically, firstly, to generate dynamic probability distribution characteristics, dynamic statistical analysis based on a sliding window is required for the output sequence of each renewable energy power station in the preprocessed data. This method not only calculates the static statistics of the entire sequence but also calculates statistics including mean, variance, skewness, and kurtosis within each window by moving a fixed-size sliding time window across the entire time series, thus generating a series of time-varying statistical sequences. Based on this, features of these statistical sequences are further extracted, such as calculating the first difference of the sliding variance sequence to obtain the "rate of change of volatility" or calculating the trend of the sliding skewness sequence. These "second-order features" can quantify the evolution of output uncertainty itself, collectively forming a feature vector characterizing the output probability distribution profile and dynamic evolution of a single power station. Secondly, to generate spatial transmission characteristics, the spatiotemporal dynamic correlation between the outputs of different renewable energy power stations needs to be analyzed. This method does not use the static Pearson correlation coefficient but instead calculates the correlation between any two different renewable energy power stations... and The time-delay cross-correlation function between them is used to achieve this, and its calculation method is as follows: , in, and These are respectively station i at time t and station j at time t. Historical output data at any given moment; , , , These are the mean and standard deviation of the power output time series of the two stations, respectively; It's a time delay. This is achieved by calculating a series of different... under the value We can find the optimal time delay that maximizes the correlation. and the corresponding maximum correlation coefficient The key feature extracted ultimately is the "time lag-correlation pair". It directly expresses the direction and speed information of the meteorological system's movement. For example, a feature pair of (15 minutes, 0.8) physically means that the output change of one station leads that of another station by an average of 15 minutes, and the two are highly positively correlated. These features constitute spatial transmission features that can reflect the force group effect and spatiotemporal propagation laws. Furthermore, to generate multi-scale time series features, wavelet transform decomposition is needed for the output time series of each station to reveal its dynamic structure at different time scales. Wavelet transform can simultaneously provide time-domain and frequency-domain information of the signal, and is particularly good at capturing and characterizing transient events in non-stationary signals. Its model can be expressed as: , in, This is the original output sequence; These are low-frequency approximate components obtained through multi-level decomposition, representing the long-term trend of the sequence. It consists of a series of high-frequency detail components at different scales. The low-frequency detail components can characterize periodic patterns, while the high-frequency detail components can capture the violent fluctuations in output caused by gusts or momentary cloud obstruction, i.e. transient events.

[0035] like Figure 2 As shown, this decomposition process breaks down the original power output sequence into a long-term trend component and a sum of high-frequency detail components containing all periodic and transient fluctuations. The extracted trend component, the energy of each periodic scale component, and the energy, peak value, and occurrence time of high-frequency transient events together constitute multi-scale time series features. Finally, the dynamic probability distribution features, spatial transmission features, and multi-scale time series features are integrated to form a comprehensive high-dimensional feature vector. This vector set is the spatiotemporal feature set. This deep feature extraction provides subsequent prediction models with a more physically rich input containing dynamic, transmission, and transient information, enabling the model to understand the physical mechanisms and stochastic processes of new energy power output, thereby improving prediction accuracy.

[0036] S3 acquires meteorological forecast data, and based on the spatiotemporal feature set and the meteorological forecast data, performs medium- and long-term power output prediction using the extreme gradient boosting tree prediction method to generate power output prediction results.

[0037] In a preferred embodiment of this invention, such as Figure 4 As shown, the generation of the output prediction result includes the following steps: S31 uses a forward feature selection method based on risk contribution to filter features from the spatiotemporal feature set and the meteorological forecast data to generate an optimized feature set; S32 Based on the optimized feature set, the Bayesian optimization method is used to optimize the hyperparameters of the extreme gradient boosting tree prediction method to generate optimized parameters; S33 processes the meteorological forecast data using the optimized parameters, performs power output forecasting, and generates power output forecasting results.

[0038] Specifically, the spatiotemporal feature set and the acquired meteorological forecast data are merged to form a full candidate feature pool. Then, a forward feature selection method based on risk contribution is used to filter this feature pool. This method is an iterative feature optimization algorithm that starts with an empty feature set. In each iteration, it traverses all features that have not yet been selected, adding the feature that maximizes the improvement of the extreme gradient boosting tree prediction method on key risk indicators to the current feature set. This process continues until the model's risk indicators no longer show a significant decrease. Through this step, features with little or no contribution to risk prediction or those that are redundant are removed, ultimately resulting in an optimized feature set with high information density and appropriate dimensionality.

[0039] Next, based on this optimized feature set, Bayesian optimization is used to optimize the hyperparameters of the extreme gradient boosting tree prediction method. Extreme gradient boosting trees are an ensemble learning algorithm whose performance is highly dependent on the settings of a series of internal parameters. Bayesian optimization is a more efficient global optimization technique than traditional grid search. It constructs a probabilistic surrogate model between hyperparameters and model performance and uses a sampling function to intelligently select the next most promising combination of hyperparameters for evaluation. This process can find the globally optimal or near-optimal parameter combination with a higher probability within a limited number of iterations. Finally, the parameter combination that achieves the best cross-validation performance is selected as the optimized parameters.

[0040] Finally, the optimized parameters and selected optimized feature set are used to construct and execute the final power output prediction operation. First, a novel extreme gradient boosting tree prediction model is initialized using the optimized parameters. Then, the model is fully trained using all historical preprocessed data, with the optimized feature set as input and historical real power output data as output. After training, the latest meteorological forecast data is fed into this trained and optimally parameterized prediction model, which then performs prediction calculations to generate power output predictions for the future. This method, through risk-oriented feature selection and efficient Bayesian optimization, ensures that the prediction model not only has high accuracy but also a deeper understanding of uncertainty risks.

[0041] Optionally, such as Figure 5 As shown, the output prediction operation includes the following steps: Step 31: Based on the spatial transmission characteristics, a hierarchical clustering algorithm based on spatiotemporal propagation characteristics is used to divide multiple new energy power stations into multiple sub-regions, generating region division results; Step 32: For each sub-region of the region division result, select a reference station, and analyze and extract the dynamic transfer operator between the non-reference station and the reference station through a gated recurrent unit network with attention mechanism to generate spatiotemporal transfer features; Step 33: The spatiotemporal transfer features are nonlinearly aggregated using a multilayer perceptron with a hybrid density network as the output layer to form a nonparametric probability distribution of regional aggregated power, generating regional power output prediction results as the power output prediction results.

[0042] Specifically, the system first physically groups the power plants based on the deep spatiotemporal dynamic correlations among them. Specifically, this embodiment constructs a spatiotemporal distance matrix as input to a hierarchical clustering algorithm. This matrix integrates three dimensions of information: the maximum correlation extracted from the time-delay cross-correlation coefficient; the optimal time delay for achieving maximum correlation; and the geographical distance between the power plants. A weighted function fuses these three factors into a single distance metric. This dynamic hierarchical clustering algorithm based on spatiotemporal propagation characteristics can more accurately classify power plants that not only have highly synchronized power outputs but also exhibit stable sequential response relationships during meteorological system movements into the same sub-region, giving the regional division results stronger physical meaning. After obtaining the regional division results, a deep learning-based correlation prediction is performed for each sub-region. First, within each sub-region, a benchmark power plant is selected according to preset rules, such as the largest installed capacity, the most complete historical data, or the most central geographical location. Subsequently, a gated recurrent unit (GRU) network with an attention mechanism is used to learn and quantify the power output relationship between each non-benchmark power plant and the benchmark power plant. The GRU network itself excels at capturing time-series dependencies, while the introduced attention mechanism allows the model to dynamically and selectively focus on several of the most influential moments in the historical sequence of the reference station when predicting the current output of a non-reference station. This enables the model to accurately capture complex, non-linear dynamic relationships such as "when a gust of wind passes, the output increase of the reference station is transmitted to this station at 80% of its magnitude 10 minutes later." After training, the network's learned intrinsic states and weights constitute a dynamic transmission operator, which is extracted as a spatiotemporal transmission feature. Finally, a multilayer perceptron (MLP-MDN) with a hybrid density network as the output layer is used to achieve probabilistic aggregation and distribution estimation of the total output within the region. The core innovation of this step lies in the output layer. Traditional multilayer perceptrons (MLPs) can only output a deterministic value or a parameter of a simple parameterized distribution.

[0043] The MLP-MDN model takes the medium- to long-term power output forecasts from benchmark power stations and the spatiotemporal transfer characteristics extracted from all non-benchmark power stations in the region as inputs. Its output layer is no longer a single numerical value or distribution, but a complete parameter set of a hybrid probabilistic model. This means that the model can fit complex, non-parametric probability distributions of arbitrary shapes, thus accurately capturing the non-Gaussian characteristics of regional aggregated power under extreme weather conditions, such as "fat tails," "skewness," or "multimodality." Summarizing the regional power output forecasts from all sub-regions constitutes the final power output forecast, rich in uncertainty information. This probabilistic forecast, especially its ability to capture non-Gaussian distributions, provides a far more accurate and robust data foundation for subsequently constructing risk indicators based on Copula functions and quantile regression than a single normal distribution assumption, improving the precision and foresight of risk assessment in extreme scenarios.

[0044] S4 compares the power output prediction result with the historical actual power output data to obtain the prediction error, and combines the evolution sequence of meteorological elements in the meteorological prediction data to perform probabilistic coupling modeling and quantile regression prediction to generate dynamic quantile values.

[0045] In this embodiment, generating dynamic quantile values ​​includes the following steps: Calculate the mean, standard deviation, and extreme values ​​of the prediction error to generate error statistical characteristics; Based on the aforementioned error statistical characteristics and the evolution sequence of meteorological elements, a joint distribution model is constructed using a probabilistic coupling modeling method based on Copula functions. Based on the evolution sequence of the meteorological elements and the features extracted by the joint distribution model, dynamic quantile values ​​of the prediction error are generated by the quantile regression prediction method.

[0046] Specifically, firstly, the predicted power output is compared point-by-point with historical actual power output data for the same period to calculate the prediction error sequence for each time point. This prediction error forms the basis for all subsequent analyses. Next, to more comprehensively describe the characteristics of the prediction error, the mean, standard deviation, and extreme values ​​are calculated for the prediction error data at each time point or within a specified time window, thereby generating a temporally continuous sequence of error statistical characteristics. For example... Figure 6 As shown in the figure, this diagram illustrates the quantitative relationship between meteorological elements and the statistical characteristics of forecast errors. Secondly, a probabilistic coupling modeling method based on Copula functions is employed to reveal the nonlinear and asymmetric dependency between forecast errors and the evolution sequence of meteorological elements. This process includes the following steps: The first step is to construct a histogram of prediction errors. By analyzing this histogram, the skewness of the prediction error probability distribution can be observed. Combined with the average value, the positive and negative errors in the prediction can be analyzed, and then a method for fitting the probability density function can be selected. For example, if the error exhibits a significant "fat-tailed" characteristic, a skewed t-distribution or a non-parametric kernel density estimation method can be used to fit the marginal probability distribution function of the prediction error. .

[0047] The second step is to analyze the evolution sequence of the meteorological elements. For each meteorological element, nonparametric methods such as kernel density estimation are also used to obtain their respective marginal probability distribution functions. .

[0048] The third step involves selecting and training a suitable Copula function, given the meteorological elements and prediction error data, to glue together all the aforementioned marginal distributions, thereby estimating their joint probability density. The Copula function is a mathematical tool that can "glue" multiple one-dimensional marginal distributions into a joint distribution without making assumptions about the data distribution. Its mathematical expression is: , in, It is a Copula function. It is the marginal probability density function of each variable.

[0049] Secondly, a quantile regression prediction model is introduced to directly generate risk intervals end-to-end. This involves predicting the future evolution sequence of meteorological elements. and by The correlation structure features, calculated by the model and representing the joint probability space location of the meteorological sequence, are used as input features to construct a machine learning model with quantile loss as the objective function. Unlike traditional regression that predicts the mean, quantile regression aims to directly predict specific quantiles of the target variable. This method can handle continuous data without facing the granularity requirements of discretization methods. By assigning a series of different quantiles... Train separate models for a given future weather scenario. This allows us to directly obtain a series of quantile values ​​for the prediction error: This method, employing an end-to-end learning paradigm, not only improves computational efficiency but also extracts complex probability distribution information into dynamic quantiles for risk management, providing quantitative input for the subsequent construction of refined early warning indicators.

[0050] S5 determines the dynamic risk range based on the dynamic quantile value, thereby constructing a graded early warning indicator.

[0051] Optionally, the construction of the hierarchical early warning indicators includes: The dynamic quantile values ​​are combined to calculate the risk intervals corresponding to different confidence levels, forming a set of risk intervals; Establish a mapping relationship between the set of risk intervals and risk levels to generate a dynamic risk level mapping; Based on the dynamic risk level mapping, different early warning levels and their triggering conditions are defined, and a hierarchical early warning index is constructed.

[0052] Specifically, for a given future meteorological element evolution sequence We have obtained a series of dynamic quantile values ​​for prediction errors using a quantile regression model, and then directly combined pairs of quantiles to form risk intervals. For example, a risk interval with a confidence level of α=95% has upper and lower limits. It is directly composed of the 2.5% quantile and the 97.5% quantile of the prediction error: , By combining corresponding quantile pairs for a series of different confidence levels (e.g., 95%, 90%, ..., 50%), a set of risk intervals containing multiple nested intervals and their evolution sequence over a future period can be obtained. Next, a mapping relationship is established between this set of risk intervals and specific risk levels, generating a dynamic risk level mapping. The core logic of this mapping is that the magnitude of the uncertainty in the prediction error is positively correlated with the level of risk. Uncertainty can be controlled by the width of the confidence interval. The rate of deterioration is measured by the rate of change of the width. This is reflected in the risk level mapping. Based on this, several risk levels can be defined, such as "low risk," "medium risk," and "high risk," with quantitative evaluation criteria set for each level. For example, a rule can be defined: if the width of the 95% risk range is less than the threshold W1 and its width shows a decreasing or stable trend, the risk level is "low risk"; if the width of the 80% risk range is greater than the threshold W2 or its width shows a rapid expanding trend, the risk level is "high risk." These rules together constitute a dynamic risk level mapping.

[0053] Finally, based on this dynamic risk level mapping, different warning levels and their triggering conditions are defined, ultimately constructing a tiered warning index. The abstract risk level is concretized into operational-level warning levels, such as "blue warning," "yellow warning," and "orange warning." Each warning level directly corresponds to a risk level. Its triggering condition is the risk level assessment standard defined in the previous step. For example, when future weather forecast data, after the entire analysis process, ultimately calculates a risk level of "high risk," an "orange warning" is triggered. Thus, the complete system composed of different warning levels and their precise, quantitative triggering conditions that combine the magnitude and trend of uncertainty constitutes the final constructed tiered warning index. This method successfully transforms complex, continuous probability distribution information into a user-friendly, discrete, and tiered warning signal for power grid dispatchers. It surpasses the traditional single-threshold alarm mode, providing a more refined and dynamic risk view.

[0054] Therefore, in summary, this embodiment first integrates historical power output and meteorological data from multiple renewable energy power stations selected based on a meteorological topology network, and performs refined spatiotemporal coupling data cleaning and repair to establish a solid data foundation for the model. Secondly, it utilizes deep spatiotemporal feature analysis to comprehensively capture the inherent physical laws and group effects of renewable energy power output from three dimensions: dynamic probability distribution, spatial transmission, and multi-scale temporal evolution. These deep features provide crucial information for subsequent accurate predictions. Then, this method employs a hierarchical aggregation prediction framework: first, it physically groups the power stations using a clustering algorithm based on spatiotemporal propagation characteristics; then, it uses a gated recurrent unit (GRU) network with attention mechanism to extract dynamic transmission operators between power stations within each region; finally, it uses an aggregation model centered on a hybrid density network (MDN) to generate a non-parametric probability distribution of regional aggregated power that can capture complex characteristics such as "fat tails" and "skewness," serving as the power output prediction result.

[0055] Based on this, a two-stage method is used to quantify the uncertainty of prediction.

[0056] In the first stage, the nonlinear and asymmetric dependency structure between prediction error and meteorological element evolution sequence is probabilistically coupled and modeled using the Copula function.

[0057] In the second stage, the future meteorological evolution sequence and the dependency structure features extracted by the Copula model are used as input. A quantile regression model is then used to directly and end-to-end predict a series of key quantile values ​​for the prediction error at future times. Finally, the directly predicted dynamic quantiles are combined to construct a series of dynamic risk intervals. Based on the width and changing trends of these intervals, continuous probabilistic risk information is transformed into discretized, tiered early warning indicators that are user-friendly for grid dispatchers. This achieves a quantitative assessment and grading of the uncertainty risk of renewable energy output. This method, while ensuring the safe and stable operation of the power grid, can improve the absorption rate of renewable energy, reduce wind and solar curtailment caused by prediction errors, and enhance the overall operating efficiency and economy of the power system.

[0058] In addition to the above, such as Figure 7 As shown, the method in this embodiment further includes: S6 applies the graded early warning indicators to power grid dispatch and configures a dispatch response strategy for each early warning level; During grid operation, S7 automatically invokes the corresponding scheduling response strategy based on the early warning level triggered by the real-time candidate output prediction results, and generates an optimized scheduling scheme.

[0059] Specifically, this process first requires strategy configuration in the offline phase. That is, for each warning level defined by the constructed hierarchical warning indicators, such as "blue warning," "yellow warning," and "orange warning," a corresponding dispatch response strategy is pre-designed and configured. These strategies are tailored to the potential impact of different risk levels, aiming to proactively absorb or hedge the balance risk caused by the uncertainty of new energy output forecasts. For example, for the low-risk level corresponding to "blue warning," the dispatch response strategy might only involve minor adjustments to conventional reserve capacity; for "yellow warning," the strategy might escalate to adjusting the operating mode of hydropower units, increasing spinning reserve, and pre-starting some fast-response gas turbine units; while for the highest-risk "orange warning," the strategy might include a series of strong intervention measures such as requesting the activation of the reserve capacity market, implementing demand-side response agreements signed with large users, and even adjusting inter-regional tie-line power exchange plans. These dispatch response strategies are structured and stored in the dispatch automation system, each associated with a specific warning level. Once in the online operation phase, this method transforms into an automated real-time decision-making process. The power grid dispatching system periodically acquires the latest meteorological forecast data and automatically executes the entire process from data preprocessing to generating output forecast results.

[0060] Then, the real-time candidate output prediction result is compared with the corresponding historical actual output data, and combined with meteorological forecast data, the risk level corresponding to the current forecast period is calculated in real time through error probability modeling and early warning indicator construction modules, and the corresponding early warning level is triggered. Once an early warning level is triggered, the scheduling response strategy bound to that early warning level will be automatically invoked from the pre-configured strategy library.

[0061] Finally, the invoked scheduling response strategy is incorporated as a constraint or optimization objective into the conventional economic scheduling or safety-constrained unit combination model for further optimization calculations. For example, if a "yellow alert" is triggered, the corresponding "increase spinning reserve" strategy will be transformed into an increase in the minimum limit constraint on spinning reserve capacity in the economic scheduling model. The optimization solver re-solves under the new constraints, ultimately generating an optimized scheduling scheme that satisfies both economic objectives and fully considers the risks of predictive uncertainty.

[0062] This plan may include adjusted output plans for conventional generating units, energy storage charging and discharging plans, and necessary grid structure adjustments. This forward-looking dispatching model can dynamically and economically allocate flexible resources within the system according to the severity of risks, effectively avoiding grid security and stability issues that may result from excessive prediction errors. At the same time, the tiered response strategy avoids adopting overly conservative dispatching methods in all situations, thereby enhancing the grid's adaptability and resilience to high-proportion, highly volatile renewable energy integration while ensuring system security.

[0063] Based on the same inventive concept, such as Figure 8 As shown, the present invention also provides a risk indicator construction system for predicting new energy output, the system comprising: The data preprocessing module is used to acquire historical power output data and meteorological data from multiple new energy power stations, perform spatiotemporal coupled data cleaning and missing value repair, and generate preprocessed data. The spatiotemporal feature extraction module is used to generate dynamic probability distribution features from the dynamic descriptive statistics of the power output sequences of each station in the preprocessed data; calculate the time-delay cross-correlation coefficients between different new energy power stations and generate spatial transmission features by combining geographical location information; and perform wavelet decomposition on the power output time series in the preprocessed data to extract trend, seasonal and transient event components, generate multi-scale time series features, and finally generate a spatiotemporal feature set. The power output prediction module is used to acquire meteorological forecast data and, based on the spatiotemporal feature set and the meteorological forecast data, perform medium- and long-term power output prediction using the extreme gradient boosting tree prediction method to generate power output prediction results. The error probability modeling module is used to compare the power output prediction results with historical actual power output data to obtain the prediction error, and combine the evolution sequence of meteorological elements in the meteorological prediction data to perform probability coupling modeling and quantile regression prediction to generate dynamic quantile values. The early warning indicator construction module is used to determine the dynamic risk range based on the dynamic quantile value and construct a graded early warning indicator.

[0064] To verify the feasibility of this invention in practice, it was applied to a regional power grid dispatch center. This dispatch center is responsible for dispatching a new energy cluster within its jurisdiction, comprising 50 wind power plants and 30 photovoltaic power plants. Due to the intermittent and highly volatile nature of new energy output, traditional dispatch methods based on deterministic prediction are insufficient to effectively address the power grid security risks posed by significant errors in output forecasting. The dispatch center aims to utilize the method of this invention to construct risk indicators for new energy output forecasting, thereby achieving quantitative assessment of forecast uncertainty and proactive risk management.

[0065] To verify the effectiveness of this invention, this embodiment collected historical power output data and synchronous meteorological data from all renewable energy power stations in the region from January 2022 to December 2024, a total of three years, with a data time resolution of 15 minutes. Based on this data, the method of this invention was used for medium- and long-term power output forecasting, risk indicator construction, and scheduling strategy optimization.

[0066] In this embodiment, the collected raw data is first subjected to spatiotemporal coupling data cleaning and repair. For example, in the data from July 2023, a photovoltaic power station generated an anomaly in instantaneous pulse due to equipment failure. The system applied a spatiotemporal coupling median filtering method, combined with the time-series data before and after the power station and the synchronous data of neighboring power stations, to successfully correct the anomaly data and avoid misjudging the actual regional power output fluctuations. Simultaneously, a coastal wind farm experienced 48 consecutive hours of missing data in November 2023 due to annual maintenance. The system applied a weighted k-dimensional tree algorithm based on spatiotemporal similarity to find the most similar samples in terms of meteorological background, geographical location, and time in the historical database, performed weighted interpolation, and accurately filled in the missing values.

[0067] In the spatiotemporal feature extraction stage, the system performs in-depth data mining on the preprocessed data. By calculating the mean, variance, and skewness of the power output sequences for each station using a sliding window, dynamic probability distribution characteristics are generated. By calculating the time-delay cross-correlation coefficients between stations, a spatial transmission characteristic with an optimal time delay of 15 minutes and a correlation coefficient as high as 0.85 is discovered between stations A1 and A2 in the coastal wind farm cluster A, revealing the wind field's movement patterns. Simultaneously, wavelet decomposition is used to extract the trend, seasonality, and transient event components caused by gusts from the power output sequences. These features collectively constitute a spatiotemporal feature set containing physical implications.

[0068] In the power output prediction stage, the system adopts a hierarchical aggregation prediction framework. First, based on the aforementioned spatial transmission characteristics, a hierarchical clustering algorithm based on spatiotemporal propagation characteristics is used to divide 80 renewable energy power plants into 8 sub-regions with highly similar power output characteristics. For each sub-region, the system selects the power plant with the largest capacity as the benchmark power plant and uses a gated recurrent unit network (GRU) with attention mechanism to learn the dynamic transfer operator between the non-benchmark power plant and the benchmark power plant. Finally, through a multilayer perceptron with a hybrid density network (MDN) as the output layer, the predicted values ​​of the benchmark power plant and the dynamic transfer operator are nonlinearly aggregated to generate a nonparametric probability distribution of aggregated power for each sub-region that can capture complex characteristics such as "fat tails" and "skewness," which serves as the final power output prediction result.

[0069] The core of this invention lies in the construction and application of risk indicators. For example, on September 20, 2024, the weather forecast indicated that a strong cold air mass would pass through the area the following day, accompanied by drastic changes in wind speed and direction. The system first calculates the error sequence between the power output prediction result based on MDN and the historical actual power output. Then, using a probabilistic coupling modeling method based on the Copula function, a joint distribution model is established for the prediction error and the evolution sequence of key meteorological elements such as wind speed and wind direction change rate. Finally, the future meteorological evolution sequence and the dependency structure features extracted from the Copula model are input into the quantile regression model to directly predict the dynamic quantile value of the wind power output prediction error for the following day. The model calculation results show that the 95% confidence interval of the prediction error from 14:00 to 16:00 the following day dynamically expands to [-80MW, +95MW].

[0070] According to the preset risk level mapping rules, this confidence interval triggered the highest level, the "Orange Alert." Correspondingly, the system automatically invoked the dispatch response strategy bound to the "Orange Alert," adding an extra 200MW of spinning reserve in the day-ahead dispatch plan for the following day and pre-starting a fast-response gas turbine unit. However, on August 15, 2024, with stable weather conditions, the system's calculated 95% confidence interval was only [-15MW, +18MW], triggering a "Blue Alert." The dispatch system maintained only the standard reserve level, achieving economical resource allocation.

[0071] It should be noted that the electrical connections between the various units described above do not necessarily represent direct or indirect connections. Any indirect connection method can be applied to the embodiments of the present invention as long as it achieves the purpose of the present invention. The above descriptions are merely exemplary embodiments of the present invention and should not be construed as limiting the scope of the present invention.

[0072] All equivalent changes and modifications made in accordance with the teachings of this invention are still within the scope of this invention. Those skilled in the art will readily conceive of other embodiments of this invention upon considering the specification and the disclosure of practical truth. This application is intended to cover any variations, uses, or adaptations of this invention that follow the general principles of this invention and include common knowledge or conventional techniques in the art not described herein.

Claims

1. A method for constructing risk indicators for predicting new energy output, characterized in that, The method includes: Acquire historical power output data and meteorological data from multiple renewable energy power stations, perform spatiotemporal coupled data cleaning and missing value repair, and generate preprocessed data; Dynamic descriptive statistics of the power output sequence of each station in the preprocessed data are used to generate dynamic probability distribution features; time-delay cross-correlation coefficients between different new energy power stations are calculated, and spatial transmission features are generated by combining geographical location information; wavelet decomposition is performed on the power output time series in the preprocessed data to extract trend, seasonal and transient event components, generate multi-scale time series features, and finally generate a spatiotemporal feature set. Acquire meteorological forecast data, and based on the spatiotemporal feature set and the meteorological forecast data, perform medium- and long-term power output prediction using the extreme gradient boosting tree prediction method to generate power output prediction results; The predicted output results are compared with historical actual output data to obtain the prediction error. Then, combined with the evolution sequence of meteorological elements in the meteorological prediction data, probabilistic coupling modeling and quantile regression prediction are performed to generate dynamic quantile values. Based on the dynamic quantile values, dynamic risk intervals are determined, thereby constructing a tiered early warning indicator.

2. The method for constructing risk indicators for predicting new energy output according to claim 1, characterized in that, The process of performing spatiotemporally coupled data cleaning and missing value repair to generate preprocessed data includes: Outlier correction is performed on the original historical power output data and meteorological data using a spatiotemporal coupled median filtering method. The spatiotemporal coupled median filtering method uses a sliding window in the spatiotemporal dimension, moving point by point on the data sequence. When replacing the data at the center point of the window, the median of all data points in this spatiotemporal neighborhood is referenced. Based on the obtained preliminary cleaned data, a weighted method based on spatiotemporal similarity is adopted. k The dimensional tree algorithm is used to impute missing values ​​and obtain a complete dataset. The Z-score detection method with dynamic adaptive threshold is applied to the complete dataset to identify and process outliers, and the outliers are replaced to generate the final preprocessed data. The adaptive threshold is calculated in real time based on the specific situation of the current data point and the power output statistical fluctuation characteristics of the station or region where the current data point is located in the same period in history.

3. The method for constructing risk indicators for predicting new energy output according to claim 2, characterized in that, The application uses a weighted k-dimensional tree algorithm based on spatiotemporal similarity to impute missing values, including: For each time point containing missing values, a spatiotemporal weighted distance metric function is used to find neighboring data points that are similar in spatiotemporal dimension and meteorological background. The neighboring data points represent data of physical scenarios that have similar meteorological, geographical and temporal backgrounds to the current time in the past. By performing a weighted average or interpolation calculation on the missing dimension values ​​corresponding to these neighboring data points, the current missing values ​​can be estimated and filled.

4. The method for constructing risk indicators for predicting new energy output according to claim 1, characterized in that, The method of using extreme gradient boosting tree prediction to perform medium- and long-term power output prediction and generate power output prediction results includes: A forward feature selection method based on risk contribution is used to filter features from the spatiotemporal feature set and the meteorological forecast data to generate an optimized feature set; Based on the optimized feature set, the hyperparameters of the extreme gradient boosting tree prediction method are optimized using the Bayesian optimization method. Finally, the parameter combination that obtains the best cross-validation performance is selected as the optimization parameters. The meteorological forecast data is processed using the optimized parameters, and an output forecast operation is performed to generate an output forecast result.

5. The method for constructing risk indicators for predicting new energy output according to claim 4, characterized in that, The process of performing power output prediction and generating power output prediction results includes: A hierarchical clustering algorithm based on spatiotemporal propagation characteristics is used to divide multiple new energy power stations into multiple sub-regions based on the aforementioned spatial propagation characteristics, generating regional division results; For each sub-region of the region division result, a reference station is selected, and the dynamic transfer operator between the non-reference station and the reference station is analyzed and extracted through a gated recurrent unit network with attention mechanism to generate spatiotemporal transfer features. The spatiotemporal transfer features are nonlinearly aggregated by a multilayer perceptron with a hybrid density network as the output layer to form a nonparametric probability distribution of regional aggregated power, generate regional power output prediction results, and serve as the power output prediction results.

6. The method for constructing risk indicators for predicting new energy output according to claim 5, characterized in that, The step of combining the evolution sequence of meteorological elements in the meteorological forecast data to perform probabilistic coupling modeling and quantile regression prediction, and generating dynamic quantile values, includes: Calculate the mean, standard deviation, and extreme values ​​of the prediction error to generate error statistical characteristics; Based on the aforementioned error statistical characteristics and the evolution sequence of meteorological elements, a joint distribution model is constructed using a probabilistic coupling modeling method based on Copula functions. Based on the evolution sequence of the meteorological elements and the features extracted by the joint distribution model, dynamic quantile values ​​of the prediction error are generated by the quantile regression prediction method.

7. The method for constructing risk indicators for predicting new energy output according to claim 6, characterized in that, The probabilistic coupling modeling method based on Copula functions constructs a joint distribution model, including: Construct a histogram of prediction error, observe the skewness of the prediction error probability distribution through the histogram, analyze the positive and negative errors in the prediction by combining the average value, and then select a non-parametric fitting method to fit the marginal probability distribution function of the prediction error. For each meteorological element in the meteorological element evolution sequence, a non-parametric fitting method is also used to obtain its respective marginal probability distribution function; Given meteorological elements and prediction error data, a joint distribution model is obtained by selecting and training a Copula function and then fusing the marginal probability distribution function, thereby obtaining an estimated joint probability density.

8. The method for constructing risk indicators for predicting new energy output according to claim 1, characterized in that, The constructed hierarchical early warning indicators include: The dynamic quantile values ​​are combined to calculate the risk intervals corresponding to different confidence levels, forming a set of risk intervals; Establish a mapping relationship between the set of risk intervals and risk levels to generate a dynamic risk level mapping; Based on the dynamic risk level mapping, different early warning levels and their triggering conditions are defined, and a hierarchical early warning index is constructed.

9. A risk indicator construction system for predicting new energy power output, applied to the risk indicator construction method for predicting new energy power output as described in any one of claims 1-8, characterized in that, The system includes: The data preprocessing module is used to acquire historical power output data and meteorological data from multiple new energy power stations, perform spatiotemporal coupled data cleaning and missing value repair, and generate preprocessed data. The spatiotemporal feature extraction module is used to generate dynamic probability distribution features from the dynamic descriptive statistics of the power output sequences of each station in the preprocessed data; calculate the time-delay cross-correlation coefficients between different new energy power stations and generate spatial transmission features by combining geographical location information; and perform wavelet decomposition on the power output time series in the preprocessed data to extract trend, seasonal and transient event components, generate multi-scale time series features, and finally generate a spatiotemporal feature set. The power output prediction module is used to acquire meteorological forecast data and, based on the spatiotemporal feature set and the meteorological forecast data, perform medium- and long-term power output prediction using the extreme gradient boosting tree prediction method to generate power output prediction results. The error probability modeling module is used to compare the power output prediction results with historical actual power output data to obtain the prediction error, and combine the evolution sequence of meteorological elements in the meteorological prediction data to perform probability coupling modeling and quantile regression prediction to generate dynamic quantile values. The early warning indicator construction module is used to determine the dynamic risk range based on the dynamic quantile value and construct a graded early warning indicator.

Images