A risk early warning method, device and equipment under an extreme weather scenario and a medium

Abstract

The application discloses a risk early warning method and device under an extreme weather scenario, equipment and medium, relates to the new energy and power system technical field, and the method comprises the following steps: acquiring historical meteorological data and constructing an association database; acquiring first meteorological data, and extracting first meteorological characteristics of the first meteorological data according to the association database; constructing a target network model; inputting the first meteorological characteristics into the target network model to obtain a first parameter of a prediction error probability distribution function; establishing a risk assessment index system based on the prediction error statistical characteristics in the historical meteorological data; setting a multi-level error threshold according to the first parameter and the risk assessment index system; calculating a first probability that the prediction error exceeds the multi-level error threshold; dividing a risk level according to the first probability and a preset index, and generating comprehensive early warning information. The method realizes early warning of the prediction error of wind power and photovoltaic power.

Description

Technical Field

[0001] This application relates to the field of new energy and power system technology, and in particular to a risk warning method, device, equipment and medium for extreme weather scenarios. Background Technology

[0002] With the continuous advancement of new power system construction, large-scale grid connection of new energy sources has become a prominent feature of modern power grid development. As important components of clean energy, the accuracy of power output forecasting for wind power and photovoltaic power generation directly affects power system dispatch and safe operation.

[0003] However, extreme weather events occur frequently, such as cold waves, severe convection, smog, sandstorms, and high temperatures. Under these scenarios, the accuracy of traditional numerical weather prediction (NWP) models decreases, leading to a significant increase in wind and solar power output prediction errors, posing a severe challenge to the safe and stable operation of the power grid. To ensure the safe and reliable operation of the power system under extreme weather conditions, it is urgent to develop a method that can provide early warnings of wind and solar power output prediction errors, improve prediction accuracy, reduce prediction errors, and assist in grid dispatching decisions. Summary of the Invention

[0004] This application provides a risk warning method, device, equipment, and medium for extreme weather scenarios. By constructing an extreme weather feature database, extracting meteorological features required for prediction, establishing a prediction error probability model, and classifying error risk levels, it enables early warning of prediction errors for wind power and photovoltaic power.

[0005] To achieve the above objectives, this application adopts the following technical solution:

[0006] Firstly, this application provides a risk warning method for extreme weather scenarios, the method comprising:

[0007] Acquire historical meteorological data and construct an associated database based on the historical meteorological data;

[0008] Acquire first meteorological data, and extract the first meteorological feature of the first meteorological data based on the associated database;

[0009] Construct the target network model;

[0010] The first meteorological feature is input into the target network model to obtain the first parameter of the prediction error probability distribution function;

[0011] A risk assessment index system is established based on the statistical characteristics of forecast errors in historical meteorological data.

[0012] Based on the first parameter and the risk assessment indicator system, set multi-level error thresholds;

[0013] Calculate the first probability that the prediction error exceeds the multi-level error threshold;

[0014] Risk levels are classified based on the primary probability and preset indicators, and comprehensive early warning information is generated.

[0015] In some possible implementations, constructing the target network model includes:

[0016] The historical meteorological data is used as training data for the first network model; the first network model is improved, and the target network model is obtained by combining the training data and the improved first network model.

[0017] In some possible implementations, constructing a relational database based on the historical meteorological data includes:

[0018] Extreme weather events are identified based on the historical meteorological data, meteorological parameters of the extreme weather events are extracted, a second meteorological feature is calculated based on the meteorological parameters, and an association database is constructed based on the second meteorological feature; wherein, the second meteorological feature includes gradient change feature, acceleration change feature, extreme value marking feature, and time series fluctuation feature.

[0019] In some possible implementations, acquiring the first meteorological data includes:

[0020] Acquire initial meteorological data and perform spatiotemporal matching on the initial meteorological data; perform data preprocessing on outliers and missing values ​​in the matched initial meteorological data and unify the data format to obtain the first meteorological data.

[0021] In some possible implementations, the improvement of the first network model includes:

[0022] Design the loss function;

[0023] The first network model is improved by combining the first network model and the loss function; wherein, the first network model refers to a hybrid neural network architecture, which adopts a hybrid neural network that integrates a long short-term memory network and a feedforward neural network, including multi-scale branches, an extreme weather perception attention mechanism, and a gated adaptive feature fusion module.

[0024] In some possible implementations, the establishment of a risk assessment indicator system includes:

[0025] The risk assessment indicators in the risk assessment indicator system include the power curtailment loss risk index and the equipment fatigue risk index; a dynamic threshold adjustment mechanism is established based on the power curtailment loss risk index and the equipment fatigue risk index; and a risk assessment indicator system is constructed based on the threshold adjustment mechanism.

[0026] In some possible implementations, setting multi-level error thresholds includes:

[0027] The first, second, and third thresholds are determined based on the standard deviation of historical prediction errors. The first, second, and third thresholds are dynamically adjusted by combining seasonal cycles, real-time grid load factor, new energy penetration rate, and risk assessment index system to obtain multi-level error thresholds.

[0028] Secondly, this application provides a risk warning device for extreme weather scenarios, the device comprising:

[0029] The acquisition module is used to acquire historical meteorological data, construct an associated database based on the historical meteorological data, acquire first meteorological data, and extract a first meteorological feature from the first meteorological data based on the associated database.

[0030] The training module is used to construct the target network model; the first meteorological feature is input into the target network model to obtain the first parameter of the prediction error probability distribution function;

[0031] The prediction module is used to establish a risk assessment index system based on the statistical characteristics of prediction errors in historical meteorological data; set multi-level error thresholds according to the first parameter and the risk assessment index system; and calculate the first probability that the prediction error exceeds the multi-level error thresholds.

[0032] The early warning module is used to classify risk levels based on the first probability and preset indicators, and generate comprehensive early warning information.

[0033] Thirdly, this application provides a computing device, including a memory and a processor;

[0034] The memory stores one or more computer programs, the one or more computer programs including instructions; when the instructions are executed by the processor, the computing device performs the method as described in any one of the first aspects.

[0035] Fourthly, this application provides a computer-readable storage medium for storing a computer program for performing the method as described in any one of the first aspects.

[0036] Fifthly, this application provides a computer program product comprising one or more computer instructions, wherein when the computer instructions are executed by a computer, the computer performs the method as described in any one of the first aspects.

[0037] As can be seen from the above technical solution, this application has at least the following beneficial effects:

[0038] In this application, historical meteorological data is acquired, and a related database is constructed based on the historical meteorological data; first meteorological data is acquired, and first meteorological features of the first meteorological data are extracted based on the related database; a target network model is constructed; the first meteorological features are input into the target network model to obtain the first parameter of the prediction error probability distribution function; a risk assessment index system is established based on the statistical characteristics of prediction errors in historical meteorological data; multi-level error thresholds are set according to the first parameter and the risk assessment index system; the first probability that the prediction error exceeds the multi-level error thresholds is calculated; and risk levels are divided according to the first probability and preset indicators to generate comprehensive early warning information.

[0039] In existing technologies, wind power output forecasting suffers from significant deviations due to inaccurate wind speed and direction predictions, while photovoltaic (PV) power output forecasting is distorted by forecast errors related to solar radiation, cloud cover, and particulate matter concentration. When forecast errors exceed the grid's safety threshold, they pose multiple challenges to power system operation: in the short term, they may cause power supply and demand imbalances, leading to grid frequency fluctuations, voltage overruns, and even triggering relay protection actions causing localized power outages; in the medium to long term, they increase dispatching difficulties, forcing the grid to maintain redundant spinning reserve capacity to cope with uncertainties, which not only increases operating costs but may also lead to cascading failures due to insufficient reserves. Therefore, this application achieves early warning of wind and PV forecasting errors by constructing an extreme weather feature database, extracting meteorological features required for forecasting, establishing a forecasting error probability model, and classifying error risk levels.

[0040] It should be understood that the descriptions of technical features, technical solutions, beneficial effects, or similar language in this application do not imply that all features and advantages can be achieved in any single embodiment. Rather, it is understood that the description of a feature or beneficial effect means that a specific technical feature, technical solution, or beneficial effect is included in at least one embodiment. Therefore, the descriptions of technical features, technical solutions, or beneficial effects in this specification do not necessarily refer to the same embodiment. Furthermore, the technical features, technical solutions, and beneficial effects described in this embodiment can be combined in any suitable manner. Those skilled in the art will understand that embodiments can be implemented without one or more specific technical features, technical solutions, or beneficial effects of a particular embodiment. In other embodiments, additional technical features and beneficial effects may be identified in specific embodiments that do not embody all embodiments. Attached Figure Description

[0041] Figure 1 A flowchart illustrating a risk warning method for extreme weather scenarios provided in this application embodiment;

[0042] Figure 2 A schematic diagram of a comprehensive prediction error probability distribution provided for an embodiment of this application;

[0043] Figure 3 A schematic diagram of a risk warning device for extreme weather scenarios provided in this application embodiment;

[0044] Figure 4 This is a schematic diagram of a computing device provided in an embodiment of this application. Detailed Implementation

[0045] The terms "first," "second," and "third," etc., used in this application specification and accompanying drawings are used to distinguish different objects, not to limit a specific order.

[0046] In the embodiments of this application, the terms "exemplary" or "for example" are used to indicate that something is an example, illustration, or description. Any embodiment or design that is described as "exemplary" or "for example" in the embodiments of this application should not be construed as being more preferred or advantageous than other embodiments or design. Specifically, the use of the terms "exemplary" or "for example" is intended to present the relevant concepts in a specific manner.

[0047] To ensure clarity and conciseness in the description of the following embodiments, a brief introduction to the related technologies is given first:

[0048] The traditional hybrid neural network architecture (Feedforward Neural Network - Long Short-Term Memory, FNN-LSTM) is a network architecture that combines feedforward neural networks and long short-term memory networks. It is mainly used to handle prediction tasks that include both temporal and non-temporal features.

[0049] Among them, the Feedforward Neural Network (FNN) is a classic deep learning model, consisting of an input layer, hidden layers, and an output layer. The neurons in each layer are connected in a unidirectional transmission manner, without feedback loops. It is mainly used to extract non-temporal features and capture the static correlation between features through multi-layer nonlinear transformations. The Long Short-Term Memory (LSTM) network is an improved form of the Recurrent Neural Network (RNN). It solves the long-term dependency problem of traditional RNNs through a gating mechanism and can effectively capture the dynamic change patterns in time-series data. It is suitable for processing time-series features such as meteorological parameters and power output data that change over time.

[0050] The typical structure of a traditional FNN-LSTM network is as follows: an LSTM branch is used to process temporal features, an FNN branch is used to process non-temporal features, and finally, the outputs of the two branches are fused through simple feature concatenation for final prediction. However, this traditional architecture has limitations, such as: it does not consider the differences in feature importance at different time scales, simple concatenation is difficult to adapt to the sudden and nonlinear features under extreme weather conditions, and it is insufficient in characterizing the asymmetry of error distribution. Therefore, the prediction accuracy and adaptability in extreme weather scenarios need to be improved.

[0051] Extreme weather events occur frequently, such as cold waves, severe convection, smog, sandstorms, and high temperatures. Under these scenarios, the accuracy of traditional numerical weather prediction (NWP) models decreases, leading to a significant increase in wind and solar power output prediction errors, posing a severe challenge to the safe and stable operation of the power grid. To ensure the safe and reliable operation of the power system under extreme weather conditions, it is urgent to develop a method that can provide early warnings of wind and solar power output prediction errors to assist in grid dispatching decisions.

[0052] In view of this, embodiments of this application provide a risk warning method for extreme weather scenarios. This method involves: acquiring historical meteorological data and constructing a related database based on the historical meteorological data; acquiring first meteorological data and extracting first meteorological features from the related database; constructing a target network model; inputting the first meteorological features into the target network model to obtain the first parameter of the prediction error probability distribution function; establishing a risk assessment index system based on the statistical characteristics of prediction errors in historical meteorological data; setting multi-level error thresholds based on the first parameter and the risk assessment index system; calculating the first probability that the prediction error exceeds the multi-level error thresholds; classifying risk levels based on the first probability and preset indicators; and generating comprehensive warning information. It is evident that this application can achieve early warning of prediction errors for wind power and photovoltaic power generation by constructing an extreme weather feature database, extracting meteorological features required for prediction, establishing a prediction error probability model, and classifying error risk levels.

[0053] To make the technical solution of this application clearer and easier to understand, the following description, in conjunction with the accompanying drawings, introduces a risk warning method for extreme weather scenarios provided by an embodiment of this application. Figure 1 As shown, this figure is a flowchart of a risk warning method for extreme weather scenarios provided in an embodiment of this application. The method is applied to a processing device and includes:

[0054] S101. The processing equipment acquires historical meteorological data and constructs a related database based on the historical meteorological data.

[0055] Historical meteorological data is a collection of observation records and derived data reflecting past meteorological conditions, including historical meteorological monitoring data, historical meteorological forecast data, historical air quality monitoring data, and historical air quality forecast data. Processing equipment acquires historical meteorological data from meteorological station sensors, satellites, radars, and other instruments or from manual observation records.

[0056] The processing equipment identifies extreme weather events based on historical meteorological data, extracts meteorological parameters of these events, calculates second meteorological features based on these parameters, and constructs a correlation database based on these second meteorological features. The second meteorological features include gradient change features, acceleration change features, extreme value labeling features, and time series fluctuation features. Specifically, they include:

[0057] In this application, the processing device identifies and marks extreme weather events based on historical meteorological data, and defines specific judgment thresholds for each type of extreme weather, as shown in Table 1.

[0058] Table 1. Characteristic parameters and judgment thresholds for extreme weather events

[0059]

[0060] For each type of extreme weather event, meteorological parameters are extracted, including basic meteorological parameters, wind power-specific meteorological parameters, and photovoltaic-specific meteorological parameters.

[0061] The basic meteorological parameters (used for joint analysis of wind power and photovoltaic power) include:

[0062] Temperature (T1): Temperature at 2m altitude, in °C; Air pressure (P): Surface air pressure, in hPa; Humidity (H): Relative humidity, in %; Precipitation (R): Cumulative precipitation, in mm; Visibility (V1): Horizontal visibility, in m; PM2.5 concentration: Fine particulate matter concentration, in μg / m³; PM10 concentration: Inhalable particulate matter concentration, in μg / m³.

[0063] Specific meteorological parameters for wind power include: wind speed (WS): average wind speed at a height of 10m, in m / s; wind direction (WD): wind direction at a height of 10m, in ° (degrees); wind shear (WShear): wind speed difference at different heights, in m / s / m.

[0064] Photovoltaic-specific meteorological parameters include: Solar radiation intensity (SR): surface solar radiation intensity, unit W / m²; Direct radiation ratio (DR): direct radiation as a percentage of total radiation, unit %; Cloud cover (C): total cloud cover, unit %.

[0065] Collect actual and predicted power output data of wind farms and photovoltaic power stations for the corresponding time periods to calculate historical prediction errors as shown in the following formula.

[0066] Formula (1):

[0067]

[0068] In formula (1), Let be the historical prediction error at time t. Let t be the actual output value of the new energy source. Let t be the predicted output value of the new energy source.

[0069] To capture the dynamic changes of meteorological parameters, a second meteorological feature is calculated for each basic meteorological parameter. The second meteorological feature includes at least a first-order change feature, a second-order change feature, an extreme value marking feature, and a time series fluctuation feature. The second meteorological features mentioned below are described in detail. The extreme value marking feature is marked in the following way. The calculation methods of other second meteorological features are shown in formulas (2) to (4):

[0070] (a) The first-order variation characteristics of meteorological parameters are shown in the following formula:

[0071] Formula (2):

[0072]

[0073] In formula (2), For the first The first-order rate of change of wind speed at any given moment. For the first Wind speed data at any given time For the first Wind speed data at any given time For the first The first-order rate of change of temperature at any given time. For the first Temperature data at any given time For the first Temperature data at any given time For the first The first-order rate of change of air pressure at time t. For the first Air pressure data at any given time For the first Air pressure data at any given time For the first The first-order rate of change of humidity at any given time. For the first Humidity data at any given time For the first Humidity data at any given time For the first The first-order rate of change of solar radiation intensity at time , For the first Solar radiation intensity data at any given time. For the first Solar radiation intensity data at any given time. For the first The first-order rate of change of PM2.5 concentration at time t. For the first PM2.5 concentration data at any given time. For the first PM2.5 concentration data at any given time. For the first The first-order rate of change of PM10 concentration at time t. For the first PM10 concentration data at any given time For the first PM10 concentration data at any given time For time intervals.

[0074] (II) Second-order variation characteristics (second-order rate of change) of meteorological parameters:

[0075] Formula (3):

[0076]

[0077] In formula (3), For the first The second rate of change of wind speed at time . For the first The first-order rate of change of wind speed at any given moment. For the first The first-order rate of change of wind speed at any given moment. For the first The second rate of change of temperature at time t, For the first The first-order rate of change of temperature at any given time. For the first The first-order rate of change of temperature at any given time. For the first The second rate of change of air pressure at time t. For the first The first-order rate of change of air pressure at time t. For the first The first-order rate of change of air pressure at time t. For the first The second-order rate of change of solar radiation intensity at time t. For the first The first-order rate of change of solar radiation intensity at time , For the first The first-order rate of change of solar radiation intensity at time , For time intervals.

[0078] (III) Characteristics of extreme value marking:

[0079] For meteorological parameters used to identify extreme weather, mark their extreme value characteristics:

[0080] Extreme wind speed markings: when Extreme wind speed characteristics Mark it as 1, otherwise mark it as 0.

[0081] Extreme temperature markings: when or At that time, the characteristics of extreme temperature values Mark it as 1, otherwise mark it as 0.

[0082] PM2.5 extreme labels: when At that time, the characteristics of extreme PM2.5 values Mark it as 1, otherwise mark it as 0.

[0083] PM10 extreme label: when At that time, the characteristics of extreme PM10 values Mark it as 1, otherwise mark it as 0.

[0084] Visibility extreme markers: Visibility extreme value marker Mark it as 1, otherwise mark it as 0.

[0085] (iv) Time series fluctuation characteristics:

[0086] The fluctuation coefficient of meteorological parameters is calculated as shown in formula (4):

[0087] Formula (4):

[0088]

[0089] In formula (4), The coefficient of variation for wind speed. This is the average wind speed forecast for the next 24 hours. This represents the standard deviation of the wind speed forecast for the next 24 hours. Similarly, the fluctuation coefficients of other meteorological parameters can be calculated; since the calculation formulas are of the same type, they will not be elaborated upon here.

[0090] Based on all the secondary meteorological features, complete extreme weather feature vectors for wind power and photovoltaic power are constructed, as shown in formulas (5) and (6):

[0091] Formula (5):

[0092]

[0093] Formula (6):

[0094]

[0095] in, This is a vector representing the characteristics of historical wind power extreme weather events. This is a vector representing historical photovoltaic extreme weather features. For the wind speed data of the i-th sample, For the wind direction data of the i-th sample, Let i be the wind shear data for the i-th sample. For the temperature data of the i-th sample, For the air pressure data of the i-th sample, For the humidity data of the i-th sample, Let i be the precipitation data for the i-th sample. Let be the first-order rate of change of the wind speed for the i-th sample. Let be the first-order rate of change of the temperature of the i-th sample. Let be the first-order rate of change of air pressure for the i-th sample. For the solar radiation intensity data of the i-th sample, Let be the direct radiation ratio data for the i-th sample. Let i be the cloud cover data for the i-th sample, where i is the sample number.

[0096] The extracted feature vectors are mapped to the corresponding historical prediction errors to construct a correlation database for wind power and photovoltaics, as shown in formula (7):

[0097] Formula (7):

[0098]

[0099] in, This represents the associated database for wind power. This represents the associated database for photovoltaics. This is the feature vector for extreme weather events affecting wind power. This is a feature vector for photovoltaic extreme weather. Indicates extreme weather type, This represents the sample size for wind power. The number of photovoltaic samples. Labels for historical forecast errors in wind power. This is a label for historical prediction errors in photovoltaics.

[0100] By acquiring historical meteorological data and constructing a correlated database, this approach overcomes the limitations of traditional single data sources in capturing local micro-meteorological features. Utilizing spatiotemporal adaptation technology with unified latitude and longitude coordinates and minute-level timestamp alignment, it ensures a strict correspondence between meteorological parameters and renewable energy output errors, avoiding correlation biases caused by data misalignment. In the extreme weather identification stage, a dual-dimensional judgment logic of "threshold and dynamic rules" is employed to effectively distinguish between regular meteorological fluctuations and extreme weather events. Combined with a configurable sliding time window algorithm, it flexibly adapts to the sensitivity of wind and solar power to meteorological disturbances of varying durations, improving the accuracy of extreme weather labeling. In multi-dimensional feature extraction, gradient and acceleration change features quantify the sudden change trend of parameters, capturing nonlinear patterns missed by traditional mean statistics. Extreme value marking, through 0 or 1 encoding and duration recording, directly correlates the intensity and duration of extreme weather events with the prediction error. Fluctuation features are analyzed using wavelet transform to resolve multi-scale components, matching the response characteristics of renewable energy output to disturbances at different time scales. Furthermore, by constructing a key-value pair mapping relationship from "feature vector to error distribution" and combining it with a spatiotemporal index design based on regional, seasonal, and extreme weather type classifications, rapid feature matching and second-level risk prediction of forecast data are achieved. A technical closed loop is formed from data collection, event recognition to feature application, comprehensively improving the system's adaptability to complex meteorological scenarios and early warning accuracy, and laying an efficient data association foundation for early warning of new energy output prediction errors under extreme weather conditions.

[0101] S102. The processing equipment acquires the first meteorological data and extracts the first meteorological features of the first meteorological data based on the associated database.

[0102] In this application, the first meteorological data refers to the observation records and derived data set of meteorological conditions for the next 24 to 72 hours, including the first meteorological monitoring data, the first meteorological forecast data, the first air quality monitoring data, and the first air quality forecast data. In addition to acquiring the first meteorological data, the processing device also acquires satellite remote sensing data, which includes weather information such as cloud cover and solar radiation.

[0103] The process of obtaining the first meteorological data specifically includes: acquiring initial meteorological data and performing spatiotemporal matching on the initial meteorological data; preprocessing outliers and missing values ​​in the matched initial meteorological data and unifying the data format to obtain the first meteorological data.

[0104] The spatiotemporal matching of initial meteorological data is to ensure consistency in time and space between meteorological data from different sources and in different formats. The specific steps are as follows:

[0105] Spatial matching: By using inverse distance weighted interpolation, the grid-based forecast data is mapped to the coordinates of specific sites such as wind farms and photovoltaic power stations, generating higher-resolution and more refined forecasts. This solves the problem of loss of local meteorological features caused by insufficient spatial resolution in traditional numerical weather prediction models, ensuring that meteorological data and site locations are accurately matched.

[0106] Time matching: Linear interpolation or resampling is used to unify initial meteorological data with different time steps to the same time granularity (such as keeping the time step consistent with the historical database), avoiding feature calculation deviations caused by different time intervals and ensuring the consistency and comparability of data in the time dimension.

[0107] Through the above spatiotemporal matching operation, the initial meteorological data can be accurately mapped to specific stations and time points, providing a reliable data foundation for subsequent feature extraction and model prediction.

[0108] The first meteorological feature of the first meteorological data is extracted according to the method of constructing the associated database. The calculation method of the first meteorological feature is as shown in step S101. The extreme value marking feature is marked by the marking method in step S101. The other first meteorological features are calculated as shown in formulas (2) to (4), and will not be described again here.

[0109] Next, based on all the first meteorological features, the complete meteorological feature vectors required for wind power and photovoltaic forecasting are constructed, as shown in formulas (8) and (9):

[0110] Formula (8):

[0111]

[0112] Formula (9):

[0113]

[0114] in, This is the feature vector for extreme weather events affecting wind power. This is a feature vector for photovoltaic extreme weather. This is the first wind speed data at time t1. This is the first wind direction data at time t1. This is the first wind shear data at time t1. This is the first temperature data at time t1. This is the first air pressure data at time t1. This is the first humidity data at time t1. This is the first precipitation data at time t1. This is the first visibility data at time t1. This is the first PM2.5 concentration data at time t1. This is the first PM10 concentration data at time t1. Let be the first-order rate of change of the first wind speed at time t1. Let be the first-order rate of change of the first wind direction at time t1. This is the first solar radiation intensity data at time t1. This is the first direct radiation ratio data at time t1. Let be the first-order rate of change of the first solar radiation intensity at time t1. This is the first cloud cover data at time t1. The first-order rate of change of cloud cover at time t1 is the first cloud cover change rate, where t1 is a time in the next 24 to 72 hours.

[0115] Multi-source forecast data spatiotemporal matching maps gridded forecast data to station coordinates through inverse distance weighted interpolation, generating higher-resolution, more refined forecasts and addressing the problem of lost local meteorological features due to insufficient spatial resolution in traditional NWP models. Simultaneously, linear interpolation or resampling unifies the time step, ensuring consistency with the historical feature database's temporal granularity and avoiding feature calculation biases caused by temporal misalignment. In data preprocessing, outlier filtering, combined with statistical rules and physical constraints, effectively removes erroneous data. Missing value repair utilizes an LSTM sequence imputation model to repair long-term missing data using historical similarity patterns, ensuring the continuity of feature calculation. Feature extraction adapts the feature calculation logic to historical meteorological data, ensuring compatibility between the forecast data feature dimensions and the associated database. Extreme value labeling uses the same regionalized threshold, guaranteeing feature consistency across time periods and providing a reliable foundation for subsequent model input.

[0116] S103, The processing device constructs the target network model.

[0117] The processing device acquires a first network model, which is a hybrid neural network architecture that integrates a long short-term memory network (LSTM) and a feedforward neural network (FNN). The first network model is improved according to the loss function. Then, historical meteorological data is used as the training data for the first network model. After the training data is input into the improved first network model, the target network model is obtained through training.

[0118] Specifically, the improvement of the first network model includes: adopting a hybrid neural network architecture that integrates a long short-term memory network (LSTM) and a feedforward neural network (FNN), wherein the hybrid neural network architecture includes a multi-scale LSTM branch, an extreme weather perception attention mechanism, and a gated adaptive feature fusion module; designing a loss function; and improving the first network model by combining the hybrid neural network architecture and the loss function.

[0119] The specific construction process of the target network model is as follows:

[0120] (a) Input feature decomposition and preprocessing of training data:

[0121] The training data for the model is the wind power and photovoltaic associated database constructed in step S101. The feature vectors in the extreme weather feature database are decomposed into time-series features and non-time-series features. The decomposition results are shown in formula (10).

[0122] Formula (10):

[0123]

[0124]

[0125] in, As a time series feature, It is a non-temporal feature; It represents short-term meteorological time series characteristics; This is a mid-term trend characteristic; These are long-term pattern characteristics, including seasonal and periodic patterns; Label feature vectors for extreme values; This is the volatility characteristic vector, which includes the volatility coefficient (CV) values ​​of various meteorological parameters; Contextual features include static information such as time period, season, and geographical location.

[0126] (ii) Improvements to the first network model:

[0127] The first network model is a hybrid neural network architecture that integrates Long Short-Term Memory (LSTM) and Feedforward Neural Network (FNN), obtained through A1~A4, specifically including:

[0128] A1. Multiscale Long Short-Term Memory Network Branch

[0129] To address the issue that traditional single LSTMs cannot simultaneously capture the short-term suddenness, medium-term trend, and long-term pattern characteristics of extreme weather, a multi-scale LSTM architecture is designed:

[0130] Design three parallel LSTM branches to process information at different time scales, as shown in equations (11) to (13):

[0131] Short-term branches (1-6 hours): (11)

[0132] Mid-term (6-24 hours): (12)

[0133] Long duration (24~72 hours): (13)

[0134] in, The hidden state is the output of the short-term branch after processing by the short-term LSTM model. LSTM models for processing short-term meteorological data The hidden state output by the intermediate branch after processing by the intermediate LSTM model. LSTM models for processing medium-term trend data The hidden states output by the long-term branch after processing by the long-term LSTM model. An LSTM model for processing long-term pattern data.

[0135] The multi-scale feature fusion mechanism does not employ simple splicing. ; ; This is because it fails to reflect the differences in importance across different time scales. Therefore, this application employs an attention fusion mechanism, as shown in formula (14):

[0136] Formula (14):

[0137]

[0138] in, Let be the attention weight for the j-th time scale; Output the feature vector for branch j1, where j1∈{ , , }, This is the transpose of the first weight matrix. This is the second weight matrix. For bias vectors, This is a weighted sum vector calculated based on the attention mechanism.

[0139] The attention mechanism can automatically adjust the weight of different time scales according to the current extreme weather type. For example, cold wave events pay more attention to long-term patterns, while severe convection pays more attention to short-term suddenness.

[0140] After constructing a multi-scale feature fusion mechanism, the mathematical expression of the LSTM unit adopts a classic form to accurately characterize the dynamic dependencies of meteorological time-series data. This is because LSTM has advantages in handling long-term time-series dependencies and can adapt to the temporal evolution characteristics of meteorological parameters under extreme weather conditions.

[0141] The mathematical expression for an LSTM unit retains its classic form as follows:

[0142] Formula (15):

[0143]

[0144] in, For the Gate of Oblivion For input gate, For output gate, Let be the cell state at time t. The cell state at time t-1. Candidate cell state, Let be the hidden state at time t. Let this be the hidden state at time t-1. Let be the input features at time t. It is the sigmoid activation function. Indicates element-wise multiplication; This is the third weight matrix. This is the fourth weight matrix. This is the fifth weight matrix. This is the sixth weight matrix. This is the first bias vector. This is the second bias vector. For the third bias vector, This is the fourth bias vector.

[0145] The operations of these LSTM units are closely linked to the multi-scale feature fusion mechanism described earlier: Input It can access multi-scale meteorological features processed by an attention fusion mechanism, and output the hidden state. Furthermore, it can provide basic vector support for the attention fusion mechanism, help dynamically allocate the weights of features at different time scales, and enable multi-scale fusion and time series modeling to form a technical closed loop that complements each other, jointly serving the task of predicting the power output error of new energy sources in extreme weather scenarios.

[0146] The three branches of the LSTM have similar structures, but their weight parameters are trained independently.

[0147] A2. Improved Feedforward Neural Network Branch

[0148] Feedforward neural networks are a classic deep learning model that extracts features through multiple layers of nonlinear transformations.

[0149] Traditional FNNs lack specific design for handling static features in response to extreme weather conditions. The following feedforward network for extreme weather perception is designed:

[0150] Formula (16):

[0151]

[0152] in, It is a non-temporal feature. The input feature vector; This is the activation value for layer 1. This is the activation value for layer 2; Let be the weight matrix of the l-th layer. This is the bias vector for layer 1. This is the weight matrix for the second layer. is the bias vector for the second layer; BatchNorm is for normalization, accelerating training convergence; Dropout is for random deactivation, preventing overfitting; ReLU is the activation function. is the output feature vector of the feedforward neural network branch, which is used in gated adaptive feature fusion, and p is the random deactivation probability of the Dropout layer.

[0153] A3. Extreme Weather Perception and Attention Mechanism

[0154] In response to the sudden nature of extreme weather events, a specialized attention mechanism was designed:

[0155] Formula (17):

[0156]

[0157] Final attention output:

[0158] in: Let m be the query vector for the m-th attention head. Let m be the key vector of the attention head. Let m be the value vector of the attention head. This is an extreme weather intensity vector, calculated based on the degree of abrupt changes in meteorological parameters; This includes contextual features, such as season and time of day. This is a weighted sum vector calculated based on the attention mechanism; This is the output of the m-th attention head; The final output of attention; For attention; Let m be the query weight matrix for the m-th attention head. Let m be the key weight matrix of the attention head. The weight matrix represents the values ​​of the m-th attention head; The dimension of the key vector; The output projection matrix is ​​defined by Q, where Q is the query vector, X is the key vector, V is the value vector, and Concat is the concatenation operation.

[0159] A4. Gated Adaptive Feature Fusion

[0160] Instead of the traditional simple splicing method, a gating mechanism is used to adaptively fuse LSTM and FNN features:

[0161] Formula (18):

[0162]

[0163] in, For time-series features, For static features, the gate weights For normalized vectors, For the final attention output, This is the output feature vector of the feedforward neural network branch; This is an extreme weather state vector, containing information on the current weather type and intensity. This is the first weight matrix of the gated network. This is the second weight matrix of the gated network; This is the first bias vector of the gated network. This is the second bias vector of the gated network; This is a numerically stable term, typically 1e-8; Element-wise multiplication; The 0th element of the normalized vector. The first element of the normalized vector. It is the final output of the gating mechanism after integrating multiple features.

[0164] (III) Design probability distribution output:

[0165] It is necessary to consider improving the output form of the model. Considering that the traditional symmetric Gaussian distribution cannot accurately describe the asymmetry of the error under extreme weather conditions, the prediction output of the model, i.e. the probability distribution of the prediction error, is improved by using an improved skewed mixture Gaussian distribution.

[0166] Mapping to distributed parameters via fully connected layers:

[0167] Formula (19):

[0168]

[0169] in, For distribution parameters, To output the weight matrix, This is the output bias vector.

[0170] The final skewed mixture Gaussian distribution is:

[0171] Formula (20):

[0172]

[0173] Formula (21):

[0174]

[0175] Where e is the prediction error, Let be the probability distribution function of the prediction error. Let the number of components in the mixture be 3. The weight of the k-th Gaussian component. To improve the probability density function of the skewed Gaussian distribution, Let be the mean of the k-th Gaussian component. Let Variance be the variance of the k-th Gaussian component. Let be the standard deviation of the k-th Gaussian component. The skewness parameter is specifically designed to capture the asymmetric characteristics of extreme weather. It is the standard normal cumulative distribution function.

[0176] The trained model output layer outputs the following parameters:

[0177] Mixed weights: Use Softmax to ensure ;

[0178] Mean: ;

[0179] variance: To ensure that the variance is positive;

[0180] Skewness: To capture asymmetric properties.

[0181] in, The weight of the first Gaussian component, The weight of the second Gaussian component, The weight of the third Gaussian component. The mean of the first Gaussian component. The mean of the second Gaussian component. The mean of the third Gaussian component. The variance of the first Gaussian component. The variance of the second Gaussian component. The variance of the third Gaussian component. The skewness parameter of the first Gaussian component. This is the skewness parameter for the second Gaussian component. This is the skewness parameter of the third Gaussian component.

[0182] (iv) Design the loss function:

[0183] To reflect the differences in economic impact across different error directions, an improved loss function is designed for the first network model:

[0184] Formula (22):

[0185]

[0186] in, For the improved loss function, It is the negative log-likelihood loss function. The first regularization coefficient is . It is an asymmetric loss function. This is the second regularization coefficient. For smoothness loss function, This is the third regularization coefficient. This is the extreme value penalty loss function.

[0187] While ensuring the model fit, the distribution characteristics of the error are explicitly controlled, and the error is the difference between the true value and the predicted value.

[0188] The loss function is decomposed as follows:

[0189] B1. Negative log-likelihood loss function: This ensures that the model parameters can accurately capture the core distribution characteristics of the error.

[0190] B2. Asymmetric loss function: Differential penalties are imposed on positive and negative errors.

[0191] B3. Smoothness Loss Function: The penalty is the absolute value of the error curvature, which forces the model to output a smooth error sequence and avoids drastic fluctuations in prediction error.

[0192] B4. Extreme value penalty loss function: This suppresses extreme errors exceeding the threshold and improves the model's robustness to outliers.

[0193] In the formula obtained from the loss function decomposition It is the negative log-likelihood loss function. For integers greater than or equal to 1 For model parameters, It is an asymmetric loss function. For smoothness loss function, The extreme value penalty loss function is defined, where W+ represents the positive error weight and W- represents the negative error weight, determined based on the cost of power curtailment. For extreme value thresholds, Let be the error of the k-th sample.

[0194] The first network model is improved based on the loss function. Then, the training data is input into the improved first network model, and the target network model is obtained after training. The training method is based on (V):

[0195] (v) Model training settings:

[0196] The model was trained using the Adam optimizer with adaptive moment estimation, and the learning rate employed a cosine annealing strategy.

[0197] Formula (23):

[0198] (twenty three)

[0199] In the formula, Let s be the learning rate at step s during training. To achieve the maximum learning rate, To minimize the learning rate, 's' represents the period length, and 's' represents the number of training steps. An early stopping strategy is employed during training; training stops if the validation set loss does not decrease for five consecutive epochs.

[0200] This leads to the target network model.

[0201] After completing a series of operations such as spatiotemporal matching, preprocessing, and feature extraction and adaptation of multi-source forecast data, S103 focuses on constructing a probability distribution model of prediction errors. This step is crucial for quantifying the risk of prediction errors in new energy output. This step involves building a hybrid neural network architecture and using a skewed Gaussian mixture distribution to meticulously characterize the error distribution, providing a solid data foundation for subsequent risk warnings.

[0202] When constructing the hybrid neural network, the multi-timescale characteristics of new energy output being affected by meteorological factors are fully considered. The LSTM branch, with its powerful ability to capture short-term dependencies in time series data, can keenly perceive short-term sudden features such as abrupt changes in wind speed and solar radiation within 1-6 hours, accurately locating the immediate impact of rapid fluctuations in meteorological conditions on new energy output. The improved feedforward neural network branch has a unique advantage in mining long-term pattern features, capable of identifying the changing patterns of meteorological data over long timescales such as monsoon cycles and seasonal changes, and how these patterns gradually affect the new energy production process. To further optimize feature fusion, a gated adaptive feature fusion module is introduced. This module uses a Softmax weight dynamic allocation mechanism to intelligently adjust the importance of time-series and non-time-series features in model computation based on the characteristics of different meteorological scenarios. For example, under extreme high-temperature weather, the temperature gradient is crucial to the efficiency degradation of photovoltaic panels. In this case, the module automatically increases the weights of features related to the temperature gradient and the photovoltaic panel efficiency degradation coefficient, making the model more closely resemble the actual scenario and significantly improving prediction accuracy.

[0203] Based on the features of the hybrid neural network output, a skewed Gaussian mixture distribution is used to model the prediction error. The skewed Gaussian mixture distribution overcomes the limitations of the traditional normal distribution, effectively characterizing the "long-tail" characteristics of the error distribution under extreme weather conditions. Through the skewness parameter, the degree to which the error distribution deviates from a symmetrical state can be accurately quantified. For example, under the influence of severe convective weather, the prediction error of new energy output often exhibits a right-skewed distribution. The skewness parameter can accurately reflect this asymmetry, avoiding the problem of the traditional normal distribution underestimating tail risks due to the assumption of symmetry. Simultaneously, Softmax normalization ensures that the sum of the weights of each skewed Gaussian component is always 1, maintaining the stability and rationality of the distribution model and preventing distortion of the error distribution due to weight imbalance. For example, the three skewed components in the Gaussian mixture distribution can be associated with different meteorological scenarios such as "normal weather," "weak extreme weather," and "strong extreme weather," and the probability weights of each component can be dynamically adjusted according to real-time meteorological data to achieve an accurate description of the prediction error under different scenarios.

[0204] In summary, the advantages of the improved model are:

[0205] Improved Prediction Accuracy: Compared to traditional single LSTM, the multi-scale LSTM architecture can simultaneously capture the short-term suddenness (1-6 hours), medium-term trend (6-24 hours), and long-term pattern (24-72 hours) of extreme weather, enhancing its nonlinear mapping capabilities. Adaptive Feature Fusion: Compared to simple feature concatenation, attention mechanisms and gating fusion automatically adjust the importance weights of different features according to the type of extreme weather, reducing false alarm rates. Asymmetric Modeling: Compared to traditional symmetric Gaussian distributions, skewed mixture Gaussian distributions accurately characterize the skewness of errors under extreme weather conditions. Enhanced Robustness: Improved smoothness loss and extreme value penalty terms in the loss function make the model output more stable, avoiding drastic fluctuations in prediction errors.

[0206] S104. The processing device inputs the first meteorological feature into the target network model to obtain the first parameter of the prediction error probability distribution function.

[0207] Through the data processing and model training steps described above, the processing equipment inputs the first meteorological feature into the target network model to obtain the first parameter of the prediction error probability distribution function, laying the foundation for subsequent steps.

[0208] S105. Based on the statistical characteristics of prediction errors in historical meteorological data, establish a risk assessment index system.

[0209] The risk assessment index system includes two risk assessment indices: a power curtailment loss risk index and an equipment fatigue risk index. A dynamic threshold adjustment mechanism is established based on these two indices. The risk assessment index system is then constructed based on this threshold adjustment mechanism. Specifically, this includes:

[0210] To address the issue of power curtailment losses caused by forecasting errors, a power curtailment loss risk index is established to assess the economic risks of wind / solar curtailment due to inaccurate forecasting. In particular, when peak-shaving capacity is insufficient, priority should be given to early warning, as shown in formula (24):

[0211] Formula (24):

[0212]

[0213] in, Let t be the risk index of power curtailment loss. Let be the prediction error power at time t. Let be the installed capacity at time t. The real-time peak shaving margin at time t. Let t represent the peak-shaving demand at time t. The economic loss coefficient is determined based on local electricity prices. The first weighting coefficient is determined based on historical power curtailment data, and the model sensitivity is calibrated using historical power curtailment data.

[0214] Establish an equipment fatigue risk index based on the frequency and amplitude of output fluctuations:

[0215] Formula (25):

[0216]

[0217] in, (t) represents the equipment fatigue risk index at time t. This represents the nth power change from time t-1 to time t. The time interval between time t-1 and time t is the nth power change. The cyclic fatigue factor is determined by the type of equipment. The fatigue coefficient for equipment differs between wind power and solar power. The fatigue index, i.e., the nonlinear amplification factor, is used when wind power fluctuations are frequent. =1.5, photovoltaic fluctuation stability is taken =1.2; The second weighting coefficient is determined based on the number of historical equipment failures.

[0218] Traditional fixed thresholds cannot adapt to changes in power grid conditions and preset risk assessment indicators for new energy sources in different time periods and seasons. Based on this, a dynamic threshold adjustment mechanism is established as follows:

[0219] Formula (26):

[0220]

[0221] in, Let r be the threshold of risk r at time t, where r∈{low,medium,high}. The basic threshold is determined based on the historical error standard deviation. For seasonal cycles (e.g., 365 days); This is a function of the real-time load factor of the power grid. This is a function of the penetration rate of new energy sources; The first adaptive coefficient, This is the second adaptive coefficient. The third adaptive coefficient, This is the fourth adaptive coefficient. The fifth adaptive coefficient was determined through regression analysis of historical data.

[0222] A dynamic threshold adjustment mechanism based on risk assessment indicators is adopted as the risk assessment indicator system. This mechanism solves the problem that traditional fixed thresholds cannot adapt to the time-varying characteristics of the power grid by introducing seasonal cycles, load fluctuations, renewable energy penetration rates, PLRI, and EFRI to dynamically adjust the risk threshold. For example, when the PLRI value remains high (high risk of power curtailment), we hope to adjust the threshold dynamically by... This reduces the risk threshold, making the system more sensitive and providing early warnings; when the EFRI value remains high (high risk of equipment fatigue), we hope to... Adjust the threshold to avoid frequent fluctuations that could exacerbate equipment fatigue.

[0223] In summary, this risk assessment index system breaks through the limitations of a single dimension. The curtailment loss risk index integrates predicted error power, renewable energy installed capacity, and grid peak-shaving margin to accurately quantify the economic cost of wind / solar curtailment. For example, by using the correlation of peak-shaving margin, it avoids "over-warning when the error meets the standard but peak-shaving is sufficient." The equipment fatigue risk index correlates the output fluctuation amplitude and frequency with equipment lifespan models, such as the fatigue curve of wind turbine gearboxes and the aging coefficient of photovoltaic inverters, to deeply characterize asset loss risk and solve the problem of traditional assessments ignoring equipment characteristics. The dynamic threshold adjustment mechanism is based on dynamic adaptation of dual indices, breaking the rigidity of fixed threshold scenarios—actively lowering the threshold when the curtailment index rises to intercept economic losses in advance; tightening constraints when the equipment fatigue index is high to avoid irreversible asset damage.

[0224] S106, Processing equipment first parameter and risk assessment index system, setting multi-level error thresholds.

[0225] The processing equipment adjusts the risk assessment indicator system based on the first parameter and the risk assessment indicator system. Then, it determines the first, second, and third thresholds based on the standard deviation of historical prediction errors. By dynamically adjusting the first, second, and third thresholds in conjunction with seasonal cycles, real-time grid load factor, renewable energy penetration rate, and the adjusted risk assessment indicator system, a multi-level error threshold system is obtained. Specifically, this includes:

[0226] Specifically, the standard deviation is first calculated by analyzing the prediction error in historical data. Three basic thresholds are set based on the statistical characteristics of historical error data:

[0227] The low-risk threshold, also known as the first threshold, is set as the standard deviation of historical prediction errors. .

[0228] The medium-risk threshold, also known as the second threshold, is set at twice the standard deviation of the prediction error. .

[0229] The high-risk threshold, also known as the third threshold, is set at three times the standard deviation of the prediction error. .

[0230] The multi-level error threshold is first anchored to the benchmark based on the standard deviation of historical prediction errors, and a scientific framework is constructed using statistical laws. Then, it is linked to seasonal cycles. In summer, due to sudden changes in cloud cover, the threshold is more sensitive to radiation gradients, while in winter, due to low temperature errors, the threshold range is flexibly widened. It is also linked to the real-time load factor of the power grid. When the load is high, the peak-shaving margin is small, and the threshold is lowered to warn of cascading risks in advance. It is linked to the penetration rate of new energy sources. In high-penetration scenarios, the error impact is amplified, and the threshold is tightened to strengthen constraints. In addition, combined with the dynamic feedback of risk assessment indicators, the threshold is corrected in real time when there are abnormalities in curtailment or fatigue index, so as to achieve intelligent optimization of "static statistical benchmark + dynamic scenario adaptation" and avoid the contradiction of "delayed and missed judgment in extreme scenarios" or "excessive alarm in normal scenarios".

[0231] S107. The processing device calculates the first probability that the prediction error exceeds the multi-level error threshold.

[0232] Based on the above steps, determine the dynamic threshold adjustment mechanism. The prediction error probability distribution function obtained from the prediction The probability of the prediction error exceeding each threshold is calculated, which is the first probability. The first probability includes the probability of exceeding the low-risk threshold, the probability of exceeding the medium-risk threshold, and the probability of exceeding the high-risk threshold, as shown below:

[0233] Formula (27):

[0234]

[0235] in, Let be the probability of exceeding the low-risk threshold at time t. Let t be the probability of exceeding the medium-risk threshold. Let t be the probability of exceeding the high-risk threshold at time t. Let be the probability distribution function of the prediction error. Let be the low-risk threshold at time t. Let be the threshold for medium risk at time t. The threshold for high risk at time t.

[0236] The probability of errors exceeding the threshold is calculated, upgrading the traditional fixed threshold judgment to a multi-level error threshold: relying on models such as skewed mixture Gaussian distribution, the tail characteristics of error distribution under extreme weather conditions are accurately captured, replacing the bias of the normal distribution assumption; probability quantification enables risk assessment to be graded, providing more detailed basis for subsequent decision-making.

[0237] S108. The processing equipment classifies risk levels based on the first probability and preset indicators, and generates comprehensive early warning information.

[0238] In this application, the processing device combines a first probability and preset indicators to establish a dynamic risk level classification criterion:

[0239] Low risk: <0.3 and <0.1 and <0.05 and PLRI<0.2 and EFRI<0.3

[0240] Medium risk: 0.3≤ <0.6 or 0.1≤ <0.3 or 0.05≤ <0.15 or 0.2≤PLRI<0.5 or 0.3≤EFRI<0.6

[0241] High risk: ≥0.6 or ≥0.3 or ≥0.15 or PLRI≥0.5 or EFRI≥0.6

[0242] Based on the above calculation method, comprehensive early warning information is generated, which includes the following:

[0243] Risk level: low risk, medium risk, high risk; Power curtailment loss risk index PLRI; Equipment fatigue risk index EFRI; Extreme weather type: extreme weather type classified based on meteorological characteristics; Error direction: predicted value is too high or too low.

[0244] Using the prediction error as the confidence interval for the first probability, assuming the calculated first probability is 95%, the 95% confidence interval is calculated based on the cumulative distribution function (CDF) of the Gaussian mixture distribution. The specific calculation steps are as follows:

[0245] First, construct the cumulative distribution function of the Gaussian mixture distribution:

[0246] Formula (28):

[0247]

[0248] in, The cumulative distribution function is... Let be the cumulative distribution function of the u-th skewed Gaussian component, where u and U are integers greater than or equal to 1, and u is less than or equal to U. denoted as the weight of the u-th skewed Gaussian component.

[0249] Then, the equation F( ) = 0.025 and F( )=0.975, thus obtaining the 95% confidence interval for the prediction error. , ];

[0250] Finally, this range is used as an important indicator for quantifying the prediction error range and incorporated into the comprehensive early warning information, providing accurate decision-making basis for power grid dispatch. This comprehensive early warning information provides power grid dispatching departments with intuitive and clear decision-making basis, helping them to formulate targeted response strategies.

[0251] In summary, by mapping probability to risk levels and pre-setting response strategies based on grid capacity: low risk triggers energy storage capacity, medium risk activates backup power, and high risk triggers demand response. This transformation translates abstract probabilities into executable dispatch commands, addressing the pain point of "knowing the risk but having no countermeasures." Simultaneously, the risk level classification can dynamically adapt to the real-time grid status (e.g., relaxing risk level standards when backup capacity is sufficient), achieving optimal matching of risk and resources, ultimately significantly improving the grid's risk response efficiency and resilience under extreme weather conditions.

[0252] Based on the above, this application acquires historical meteorological data and constructs a related database based on this data; acquires first meteorological data and extracts the first meteorological features from the related database; constructs a target network model; inputs the first meteorological features into the target network model to obtain the first parameter of the prediction error probability distribution function; establishes a risk assessment index system based on the statistical characteristics of prediction errors in historical meteorological data; sets multi-level error thresholds based on the first parameter and the risk assessment index system; calculates the first probability that the prediction error exceeds the multi-level error thresholds; and classifies the risk level based on the first probability and preset indicators to generate comprehensive early warning information. This application achieves early warning of prediction errors for wind power and photovoltaic power generation by constructing an extreme weather feature database, extracting the meteorological features required for prediction, establishing a prediction error probability model, and classifying error risk levels.

[0253] To maintain the system's high-precision early warning capability, this application implements an online model update mechanism and adaptive threshold adjustment, enabling the system to adapt to long-term changes in meteorological patterns and the evolution of power system equipment status. The model parameters are periodically updated using newly collected data to adapt to changes in meteorological conditions and equipment status. An incremental learning rate is adopted to avoid "catastrophic forgetting" of the model. The update formula is shown below:

[0254] Formula (29):

[0255]

[0256] In formula (29), The incremental learning rate is set to 0.2. For the new parameters, For old parameters, These are the parameters obtained by training with new data. This smooth update strategy can both absorb useful information from new data and retain the accumulated experience from historical data.

[0257] Simultaneously, the online updates of feature distribution statistics are maintained, as shown below:

[0258] Formula (30):

[0259]

[0260] in, The updated mean of the feature distribution. For updating the variance of the feature distribution; The mean of the old characteristic distribution. The variance of the old characteristic distribution; The mean of the characteristic distribution of the new data. The variance of the characteristic distribution of the new data, The smoothing coefficient is set to 0.1.

[0261] The basic threshold for risk level is also dynamically adjusted as statistical information is updated:

[0262] Formula (31):

[0263]

[0264] in, The updated low-risk threshold. The updated medium-risk threshold. The updated high-risk threshold, This is the updated standard deviation of the error.

[0265] This application provides an embodiment of a risk warning method for extreme weather scenarios, comprising the following steps:

[0266] Step C1: Collect historical meteorological data of wind farms, identify extreme weather events, extract extreme weather characteristics, and build a related database.

[0267] This embodiment uses a wind farm as an example. This wind farm is located in mountainous terrain and is susceptible to cold waves and severe convective weather. Historical meteorological data for this wind farm is collected, including:

[0268] Historical predicted power output data and corresponding historical actual power output data are sampled at 15-minute intervals.

[0269] Historical weather forecast data and corresponding historical actual weather data, including parameters such as wind speed, wind direction, temperature, air pressure, and humidity.

[0270] Historical air quality data, including parameters such as PM2.5 and PM10.

[0271] Analysis of historical meteorological data for this wind farm identified 15 cold wave events and 23 severe convective weather events. The identification results for some of the cold wave events are shown in Table 2.

[0272] Table 2 Cold Wave Events

[0273]

[0274] Key meteorological parameters and forecast error data segments during the CS03 cold wave event were extracted, as shown in Table 3:

[0275] Table 3. Data snippets during the CS03 cold wave event.

[0276]

[0277] Analysis of the data in the table above reveals that as the cold wave progresses, temperatures gradually decrease, wind speeds increase, and prediction errors show a significant increasing trend. Particularly when temperatures drop below -5.5℃, the prediction error increases dramatically, reaching over 18.6MW, accounting for more than 18.6% of the installed capacity, far exceeding the prediction error level under normal circumstances. This indicates that the impact of cold wave weather on wind power output prediction is very significant and requires special attention.

[0278] In this historical data segment, we observed that the actual power output exceeded the predicted output (positive error), indicating that weather forecasting models often underestimate wind power generation in the early stages of a cold wave. This is because the cold air intrusion brought by a cold wave is accompanied by an increased pressure gradient, leading to increased wind speed, a feature that traditional forecasting models fail to fully capture. In this situation, the actual output power of the wind turbines exceeds the predicted value, resulting in a positive error.

[0279] Using this historical meteorological data as a training set, data from all extreme weather events are processed to construct a complete correlation database of wind power extreme weather. All features are normalized to unify the feature scales of different dimensions. Based on the feature database, an error prediction model is trained to establish a mapping relationship between meteorological feature vectors and historical prediction errors.

[0280] Step C2: Extract the features of the meteorological data required for prediction.

[0281] A specific cold wave event within a certain time period was selected as test data for analysis. Assuming the current time is the first time, the following data were acquired: numerical weather prediction data, air quality forecast data, and satellite remote sensing data. Taking the 09:00 forecast as an example, features were extracted using a method completely consistent with the feature library construction phase:

[0282] 1) Basic meteorological parameters: Wind speed (WS): 14.5 m / s; Wind direction (WD): 290°; Wind shear (WShear): 0.038 m / s / m; Temperature (T1): -7.2℃; Air pressure (P): 1018.3 hPa; Humidity (H): 45%; Precipitation (R): 0 mm; Visibility (V1): 5500 m; PM2.5 concentration: 42 μg / m³; PM10 concentration: 78 μg / m³;

[0283] 2) Meteorological parameter gradient characteristics: First-order rate of change of wind speed: 0.4 m / s / h; First-order rate of change of wind direction: -5° / h; First-order rate of change of air temperature: -0.83℃ / h; First-order rate of change of air pressure: 0.27 hPa / h; First-order rate of change of humidity: -1% / h; First-order rate of change of precipitation: 0 mm / h; First-order rate of change of visibility: 66.67 m / h; First-order rate of change of PM2.5: 0.67 μg / m³ / h; First-order rate of change of PM2.5: 0.91 μg / m³ / h;

[0284] 3) Meteorological parameter acceleration characteristics: Second-order rate of change of wind speed: 0.1 m / s / Second-order rate of change of temperature: -0.3℃ / Second-order rate of change of air pressure: 0.05 hPa / ;

[0285] 4) Extreme value markings: Wind speed extreme marking: 0; Temperature extreme marking: 1; PM2.5 extreme marking: 0; PM10 extreme marking: 0; Visibility extreme marking: 0;

[0286] 5) Time series volatility characteristics: wind speed fluctuation coefficient: 0.21; air temperature fluctuation coefficient: 0.32; air pressure fluctuation coefficient: 0.002; humidity fluctuation coefficient: 0.08.

[0287] In summary, the meteorological feature vector of the current forecast input is constructed. .

[0288] Step C3: Using the trained hybrid neural network model, input the feature vector of the predicted input. The first parameter of the probability distribution function for the prediction error at 9:00 is given. The parameters of the predicted Gaussian mixture distribution are as follows:

[0289] First Gaussian component: .

[0290] The second Gaussian component: .

[0291] The third Gaussian component: .

[0292] The probability distribution of the comprehensive prediction error is as follows Figure 2 As shown in the figure, this is a schematic diagram of a comprehensive prediction error probability distribution provided in an embodiment of this application, where k=1, 2, 3, and the probability distribution function is as follows:

[0293]

[0294] Step C4: Conduct a risk assessment based on the preset indicators and dynamic thresholds for new energy sources, calculate the first probability, and classify the error levels.

[0295] Based on historical data, the standard deviation of the prediction error for this wind farm under normal weather conditions is 5.5MW. Therefore, several risk thresholds are derived: low risk threshold... medium risk threshold High risk threshold .

[0296] Taking into account seasonal factors and power grid conditions, the dynamically adjusted threshold is: =6.2MW, =12.4MW, =18.6MW.

[0297] Based on this, the first probability is calculated as follows:

[0298]

[0299] Calculate the preset indicators for new energy sources: PLRI = 0.65, EFRI = 0.45

[0300] Based on the comprehensive assessment criteria, the risk level is determined as follows:

[0301] because ,and ,and Furthermore, if PLRI ≥ 0.5, the current wind power output prediction error is determined to be at a high-risk level.

[0302] Step C5: Generate comprehensive early warning information.

[0303] Figure 2 The study visually demonstrates the skewed distribution characteristics of prediction errors under extreme weather (cold wave) scenarios: all skewness parameters γ are positive, indicating that the distribution is right-skewed and the actual output is likely to be higher than the predicted value; the superposition of the three Gaussian components forms a complex multi-peak distribution, describing the complexity of the error under extreme weather; it is asymmetric and clearly biased towards the positive error range, which is consistent with the physical mechanism of wind speed enhancement under cold wave weather.

[0304] Figure 2 It is clear that the errors are mainly distributed in the positive range and exhibit significant asymmetry, indicating that the actual output is likely higher than the predicted value. The distribution curves show that most probability masses exceed the various risk thresholds, which aligns with the conclusion in the example that it was classified as "high-risk."

[0305] At this point, a high PLRI index indicates that positive errors dominate, and the power grid may not be able to fully absorb the power, requiring mandatory restrictions on wind power output, resulting in a waste of clean energy.

[0306] The cumulative distribution function (CDF) of the Gaussian mixture distribution is:

[0307]

[0308] Solve the equation using numerical methods:

[0309]

[0310]

[0311] Through numerical solution, we obtain: .

[0312] In summary, the comprehensive early warning information generated for 09:00 on January 15, 2024 is as follows:

[0313] Risk level: High risk; Estimated error range: 1.8MW to 32.5MW (95% confidence interval); Curtailment loss risk index: 0.65; Equipment fatigue risk index: 0.45; Extreme weather type: Cold wave; Error direction: Actual output is greater than predicted output.

[0314] The above text combined Figure 1 and Figure 2 The risk warning method for extreme weather scenarios provided in the embodiments of this application has been described in detail. The apparatus and equipment provided in the embodiments of this application will be described below with reference to the accompanying drawings.

[0315] This application also provides a risk warning device for extreme weather scenarios, such as... Figure 3 As shown in the figure, this figure is a schematic diagram of a risk warning device for extreme weather scenarios provided in an embodiment of this application. The device includes: an acquisition module 301, a training module 302, a prediction module 303, and a warning module 304.

[0316] The acquisition module 301 is used to acquire historical meteorological data, construct an association database based on the historical meteorological data, acquire first meteorological data, and extract first meteorological features of the first meteorological data based on the association database.

[0317] Training module 302 is used to construct the target network model; the first meteorological feature is input into the target network model to obtain the first parameter of the prediction error probability distribution function;

[0318] Prediction module 303 is used to establish a risk assessment index system based on the statistical characteristics of prediction errors in historical meteorological data; set multi-level error thresholds according to the first parameter and the risk assessment index system; and calculate the first probability that the prediction error exceeds the multi-level error thresholds.

[0319] The early warning module 304 is used to classify risk levels based on the first probability and preset indicators, and generate comprehensive early warning information.

[0320] In some possible implementations, the training module 302 is specifically used to use the historical meteorological data as training data for the first network model; to improve the first network model; and to obtain the target network model by combining the training data and the improved first network model.

[0321] In some possible implementations, the acquisition module 301 is specifically used to identify extreme weather events based on the historical meteorological data, extract meteorological parameters of the extreme weather events, calculate a second meteorological feature based on the meteorological parameters, and construct an association database based on the second meteorological feature; wherein, the second meteorological feature includes gradient change feature, acceleration change feature, extreme value marking feature, and time series fluctuation feature.

[0322] In some possible implementations, the device further includes:

[0323] The preprocessing module is used to acquire initial meteorological data, perform spatiotemporal matching on the initial meteorological data, preprocess outliers and missing values ​​in the matched initial meteorological data, and unify the data format to obtain the first meteorological data.

[0324] In some possible implementations, the training module 302 is specifically used to design a loss function; and to improve the first network model by combining the first network model and the loss function; wherein the first network model refers to a hybrid neural network architecture, which adopts a hybrid neural network architecture that integrates a long short-term memory network LSTM and a feedforward neural network FNN, including a multi-scale LSTM branch, an extreme weather perception attention mechanism, and a gated adaptive feature fusion module.

[0325] In some possible implementations, the prediction module 303 is specifically used for the risk assessment indicators in the risk assessment indicator system, including the power curtailment loss risk index and the equipment fatigue risk index; to establish a dynamic threshold adjustment mechanism based on the power curtailment loss risk index and the equipment fatigue risk index; and to construct the risk assessment indicator system based on the threshold adjustment mechanism.

[0326] In some possible implementations, the prediction module 303 is specifically used to determine a first threshold, a second threshold, and a third threshold based on the standard deviation of historical prediction errors; and to dynamically adjust the first threshold, the second threshold, and the third threshold in conjunction with seasonal cycles, real-time grid load factor, new energy penetration rate, and the index of the risk assessment indicator system to obtain multi-level error thresholds.

[0327] The risk warning device for extreme weather scenarios according to the embodiments of this application can correspond to the execution of the method described in the embodiments of this application, and the other operations and / or functions of each module / unit of the risk warning device for extreme weather scenarios are respectively for implementing Figure 1For the sake of brevity, the corresponding processes of each method in the illustrated embodiments will not be described in detail here.

[0328] This application also provides a computing device. For example... Figure 4 As shown in the figure, this is a schematic diagram of a computing device provided in an embodiment of this application. The computing device 400 includes a bus 401, a processor 402, a communication interface 403, and a memory 404. The processor 402, the memory 404, and the communication interface 403 communicate with each other via the bus 401.

[0329] Bus 401 can be a Peripheral Component Interconnect (PCI) bus or an Extended Industry Standard Architecture (EISA) bus, etc. Buses can be categorized as address buses, data buses, control buses, etc. For ease of representation, Figure 4 The bus is represented by a single thick line, but this does not mean that there is only one bus or one type of bus.

[0330] Processor 402 can be any one or more of the following processors: central processing unit (CPU), graphics processing unit (GPU), microprocessor (MP), or digital signal processor (DSP).

[0331] The communication interface 403 is used for communication with external devices. For example, if the computing device is a first switch, the communication interface 403 can be used for communication between the first switch and a first user terminal, or for communication between the first switch and a second switch.

[0332] Memory 404 may include volatile memory, such as random access memory (RAM). Memory 404 may also include non-volatile memory, such as read-only memory (ROM), flash memory, hard disk drive (HDD), or solid state drive (SSD).

[0333] The memory 404 stores executable code, and the processor 402 executes the executable code to perform the aforementioned risk warning method for extreme weather scenarios.

[0334] Specifically, in achieving Figure 3 In the case of the illustrated embodiment, and Figure 3 When the modules or units of the risk warning device for extreme weather scenarios described in the embodiments are implemented through software, the following steps are performed: Figure 3 The software or program code required for the functions of each module / unit can be partially or entirely stored in memory 404. Processor 402 executes the program code corresponding to each unit stored in memory 404 to execute the aforementioned risk warning method under extreme weather scenarios.

[0335] This application also provides a computer-readable storage medium. The computer-readable storage medium can be any available medium that a computing device can store, or a data storage device such as a data center containing one or more available media. The available medium can be a magnetic medium (e.g., floppy disk, hard disk, magnetic tape), an optical medium (e.g., DVD), or a semiconductor medium (e.g., solid-state drive). The computer-readable storage medium includes instructions that instruct the computing device to execute the aforementioned risk warning method for extreme weather scenarios.

[0336] This application also provides a computer program product comprising one or more computer instructions. When the computer instructions are loaded and executed on a computing device, all or part of the processes or functions described in this application are generated.

[0337] The computer instructions may be stored in a computer-readable storage medium or transmitted from one computer-readable storage medium to another. For example, the computer instructions may be transmitted from one website, computer, or data center to another website, computer, or data center via wired (e.g., coaxial cable, fiber optic) or wireless (e.g., infrared, wireless, microwave, etc.) means.

[0338] When the computer program product is executed by a computer, the computer executes any of the aforementioned risk warning methods for extreme weather scenarios. The computer program product can be a software installation package; when any of the aforementioned risk warning methods for extreme weather scenarios is required, the computer program product can be downloaded and executed on the computer.

[0339] The descriptions of the processes or structures corresponding to the above figures each have their own emphasis. For parts of a process or structure that are not described in detail, please refer to the relevant descriptions of other processes or structures.

[0340] The above description is merely a specific embodiment of this application, but the scope of protection of this application is not limited thereto. Any changes or substitutions within the technical scope disclosed in this application should be covered within the scope of protection of this application.

Claims

1. A risk warning method for extreme weather scenarios, characterized in that, The method includes: Acquire historical meteorological data and construct an associated database based on the historical meteorological data; Acquire first meteorological data, and extract the first meteorological feature of the first meteorological data based on the associated database; Construct the target network model; The first meteorological feature is input into the target network model to obtain the first parameter of the prediction error probability distribution function; A risk assessment index system is established based on the statistical characteristics of forecast errors in historical meteorological data. Based on the first parameter and the risk assessment indicator system, set multi-level error thresholds; Calculate the first probability that the prediction error exceeds the multi-level error threshold; Risk levels are classified based on the first probability and preset indicators, and comprehensive early warning information is generated.

2. The method according to claim 1, characterized in that, The construction of the target network model includes: The historical meteorological data is used as training data for the first network model; The first network model is improved, and the target network model is obtained by combining the training data and the improved first network model.

3. The method according to claim 1, characterized in that, The step of constructing a relational database based on the historical meteorological data includes: Extreme weather events are identified based on the historical meteorological data, meteorological parameters of the extreme weather events are extracted, a second meteorological feature is calculated based on the meteorological parameters, and an association database is constructed based on the second meteorological feature; wherein, the second meteorological feature includes gradient change feature, acceleration change feature, extreme value marking feature, and time series fluctuation feature.

4. The method according to claim 1, characterized in that, The acquisition of the first meteorological data includes: Acquire initial meteorological data and perform spatiotemporal matching on the initial meteorological data; The outliers and missing values ​​in the matched initial meteorological data are preprocessed and the data format is standardized to obtain the first meteorological data.

5. The method according to claim 2, characterized in that, The improvement to the first network model includes: Design the loss function; The first network model is improved by combining the first network model and the loss function; wherein, the first network model refers to a hybrid neural network architecture, which adopts a hybrid neural network that integrates a long short-term memory network and a feedforward neural network, including multi-scale branches, an extreme weather perception attention mechanism, and a gated adaptive feature fusion module.

6. The method according to claim 1, characterized in that, The establishment of the risk assessment indicator system includes: The risk assessment indicators in the risk assessment indicator system include the power curtailment loss risk index and the equipment fatigue risk index. A dynamic threshold adjustment mechanism is established based on the aforementioned power curtailment loss risk index and equipment fatigue risk index; A risk assessment indicator system is constructed based on a threshold adjustment mechanism.

7. The method according to claim 1, characterized in that, The setting of multi-level error thresholds includes: The first, second, and third thresholds are determined based on the standard deviation of historical prediction errors. By combining seasonal cycles, real-time grid load factor, renewable energy penetration rate, and risk assessment index system, the first, second, and third thresholds are dynamically adjusted to obtain multi-level error thresholds.

8. A risk warning device for extreme weather scenarios, characterized in that, The device includes: The acquisition module is used to acquire historical meteorological data, construct an associated database based on the historical meteorological data, acquire first meteorological data, and extract a first meteorological feature from the first meteorological data based on the associated database. The training module is used to construct the target network model; the first meteorological feature is input into the target network model to obtain the first parameter of the prediction error probability distribution function; The prediction module is used to establish a risk assessment index system based on the statistical characteristics of prediction errors in historical meteorological data; set multi-level error thresholds according to the first parameter and the risk assessment index system; and calculate the first probability that the prediction error exceeds the multi-level error thresholds. The early warning module is used to classify risk levels based on the first probability and preset indicators, and generate comprehensive early warning information.

9. A computing device, characterized in that, Including memory and processor; The memory stores one or more computer programs, the one or more computer programs including instructions; when the instructions are executed by the processor, the computing device performs the method as described in any one of claims 1 to 7.

10. A computer-readable storage medium, characterized in that, The computer-readable storage medium is used to store a computer program for performing the method as described in any one of claims 1 to 7.

Images