Power supply and demand factor probability prediction method and system in extreme meteorological scene

Abstract

The invention provides a power supply and demand factor probability prediction method and system in an extreme meteorological scene, and the method comprises the steps: obtaining power supply and demand factor historical data and historical real meteorological data, and carrying out the standardization processing of the historical real meteorological data according to an extreme meteorological early warning standard, constructing an extreme meteorological scene library: extracting multi-dimensional features for extreme meteorological data and historical data of power supply and demand factors; constructing a multi-modal fusion prediction model; training, verifying and testing the multi-modal fusion prediction model by using the multi-dimensional features to obtain a target prediction model; and acquiring actual weather forecast data, and inputting the actual weather forecast data into the target prediction model to obtain a prediction result containing the power supply and demand factor prediction mean value, the confidence information and the probability density information. According to the invention, the links of data processing, feature engineering, model architecture, training evaluation and the like are improved, and power supply and demand factor prediction with high precision, reliability and scene adaptability in an extreme meteorological scene is realized.

Description

Technical Field

[0001] This invention relates to the field of power system forecasting technology, specifically to a method and system for probabilistic prediction of power supply and demand factors under extreme weather scenarios. Background Technology

[0002] Driven by the "dual carbon" goals, the global energy structure is accelerating its transformation towards low-carbon energy. The installed capacity of renewable energy sources such as wind and solar power continues to expand, gradually becoming one of the core forces in electricity supply. The International Energy Agency predicts that renewable energy will surpass coal to become the world's largest source of electricity as early as 2025. At the same time, intensifying global climate change has led to more frequent and increasingly severe extreme weather events such as typhoons, torrential rains, extreme heat waves, and strong cold waves, posing a serious challenge to the safe and stable operation of new power systems.

[0003] Meteorological conditions have become a key factor affecting the supply and demand balance of the power system. On the one hand, extreme weather directly impacts the stability of renewable energy output. For example, extremely low wind speeds and solar radiation can lead to "dark valleys" in wind and solar power generation, while extreme high temperatures can reduce the output power of photovoltaic modules. On the other hand, extreme weather can trigger a rigid surge in power load, forming a "double peak in winter and summer" pattern. For instance, every 1°C increase in maximum temperature will lead to a 4.5% increase in peak power load. The combination of extreme high temperatures and drought can also cause insufficient hydropower output, further exacerbating the supply and demand imbalance. The power curtailment incident in Sichuan in 2022 is a typical example. Against this backdrop, accurately predicting power supply and demand factors (including power load, wind and solar power generation, etc.) under extreme weather scenarios has become a core prerequisite for ensuring grid dispatch optimization, improving power supply reliability, and promoting the consumption of new energy sources. It is also a key technical support for building a climate-adaptive new power system.

[0004] With the deep application of artificial intelligence (AI) technology in the energy sector, power supply and demand forecasting has gradually upgraded from traditional trend extrapolation methods and ARIMA models to deep learning-based forecasting schemes, significantly improving prediction accuracy. Some AI-driven solutions can achieve prediction accuracy of over 90%. Current industry research focuses on multi-source data fusion (such as meteorological data, power data, and user behavior data) and multi-model collaborative forecasting, aiming to further enhance the adaptability of forecasting models to complex scenarios. However, existing technologies still have significant shortcomings in the special scenario of extreme weather. The strong randomness, destructiveness, and complexity of the impact mechanisms of extreme weather events cause a significant decrease in the accuracy and reliability of conventional forecasting models, making them unable to meet the high requirements of emergency dispatching in power systems. Therefore, targeted technological innovation is urgently needed to overcome this bottleneck. Summary of the Invention

[0005] To address this, the present invention provides a method and system for probabilistic prediction of power supply and demand factors under extreme weather scenarios, aiming to solve the technical problem that existing technologies for predicting power supply and demand factors under extreme weather scenarios cannot meet the core requirements of power system dispatching decisions for prediction accuracy, reliability, and scenario adaptability.

[0006] To achieve the above objectives, the present invention adopts the following technical solution:

[0007] According to a first aspect of the present invention, the present invention provides a method for probabilistic prediction of power supply and demand factors under extreme weather scenarios, the method comprising: Historical data on power supply and demand factors and historical real-time meteorological data are acquired. The historical real-time meteorological data is then standardized according to extreme weather warning standards to construct an extreme weather scenario database. For the extreme weather data in the extreme weather scenario database and the historical data of power supply and demand factors, a multi-dimensional feature set is extracted, which includes the core features of extreme weather, the historical evolution features of power supply and demand, and the cross-correlation features between extreme weather and power supply and demand. Construct a multimodal fusion prediction model that includes a feature encoding module, an attention fusion module, and a probability prediction output module; The multi-dimensional feature set is used to train, validate, and / or test the multimodal fusion prediction model to obtain a target prediction model for the probability prediction of power supply and demand factors. Obtain actual weather forecast data, input the actual weather forecast data into the target prediction model, and obtain prediction results including the predicted mean of power supply and demand factors, confidence information and / or probability density information.

[0008] Furthermore, the acquisition of historical real meteorological data, the division of the historical real meteorological data according to extreme weather prediction standards, and the generation of an extreme weather dataset consistent with the format of actual weather forecast data include: Based on the time granularity consistent with the actual weather forecast data, multi-dimensional historical real weather data are collected for the target area within the target time period; Based on the aforementioned extreme weather prediction standards, threshold values ​​for classifying various extreme weather scenarios are determined. The extreme weather sample set in the historical real meteorological data is filtered based on the aforementioned classification threshold. Referring to the data format of the actual weather forecast data, the extreme weather dataset is restructured to obtain an extreme weather dataset with field types and time granularity consistent with the actual weather forecast data; The extreme weather dataset is labeled with scene tags to form a multi-dimensional extreme weather scene library; wherein the scene tags include weather type and warning level.

[0009] Furthermore, the extraction of a multi-dimensional feature set, including core features of extreme weather, historical evolution features of power supply and demand, and cross-correlation features between extreme weather and power supply and demand, includes: Based on the extreme meteorological data, core characteristics of extreme meteorological events are extracted to characterize their intensity, duration, trend, and / or spatial distribution. The intensity includes at least one of the extreme values, mean values, and peak occurrence times of various extreme meteorological parameters. The duration includes the duration of the extreme meteorological event. The trend includes the intensity level change sequence of the extreme meteorological event and / or the magnitude of meteorological parameter changes in adjacent time periods. The spatial distribution includes the standard deviation of wind speed between stations and / or the spatial gradient of precipitation. as well as, Based on the historical data of the aforementioned power supply and demand factors, historical evolution characteristics of power supply and demand are extracted to characterize their time-series evolution, volatility, and / or supply-demand balance characteristics. The time-series evolution characteristics include at least one of the historical concurrent values, historical averages, and historical extreme values ​​of power load and wind / solar power. The volatility characteristics include the volatility, trend characteristics, and abrupt change characteristics of the power supply and demand factors. The supply-demand balance characteristics include the difference between the grid's power supply capacity and the power load, the proportion of wind / solar power in the total power supply, and / or the changes in the proportion of industrial load to residential load. Based on the extreme weather data and the historical data of the power supply and demand factors, the nonlinear correlation between extreme weather events and power supply and demand factors is analyzed to obtain the cross-correlation characteristics between extreme weather and power supply and demand.

[0010] Furthermore, based on the extreme weather data and the historical data of the power supply and demand factors, the nonlinear correlation between extreme weather events and power supply and demand factors is analyzed to obtain the cross-correlation characteristics of extreme weather and power supply and demand, including: Calculate the correlation coefficients between the extreme meteorological parameters and the power supply and demand factors under different lag durations to obtain the lag correlation characteristics; and / or, Sensitivity characteristics are obtained by calculating the sensitivity coefficients between the extreme meteorological parameters and the power supply and demand factors at different intensity levels using piecewise linear regression; and / or, Develop scenario-specific cross-features for different extreme weather scenarios; and / or, The time difference between the end time of the extreme weather event and the time when the power supply and demand factors recover to normal levels, as well as the load / power change rate during the recovery process, are extracted as recovery features.

[0011] Furthermore, after extracting a multi-dimensional feature set including core features of extreme weather, historical evolution features of power supply and demand, and cross-correlation features between extreme weather and power supply and demand, the method further includes: A two-step method is used to filter the features of the multi-dimensional feature set to obtain an optimized multi-dimensional feature set, specifically including: The importance score of each feature in the multi-dimensional feature set to the prediction result of power supply and demand factors is calculated based on the tree model, and the importance scores are sorted and features below the first preset threshold are removed to obtain a preliminary feature set. Calculate the mutual information entropy between each feature in the preliminary feature set. For feature pairs whose mutual information entropy is higher than a second preset threshold, remove features with lower importance scores from the feature pairs to obtain the optimized multi-dimensional feature set.

[0012] Furthermore, the construction of the multimodal fusion prediction model, which includes a feature encoding module, an attention fusion module, and a probability prediction output module, includes: A feature encoding module integrating a convolutional neural network architecture, a long short-term memory network architecture, and a fully connected layer is constructed. The convolutional neural network architecture is used to encode the spatial distribution characteristics of the extreme weather; the long short-term memory network architecture is used to encode the temporal characteristics of power supply and demand; and the fully connected layer is used to encode the cross-correlation characteristics between extreme weather and power supply and demand. An attention fusion module employing a multi-head attention mechanism is constructed; the multi-head attention mechanism is used to adaptively weight and fuse features encoded by multimodal approaches. A probability prediction output module employing a Bayesian neural network structure is constructed; the Bayesian neural network structure is used to output prediction results containing the predicted mean of power supply and demand factors, confidence information, and / or probability density information by introducing the probability distribution of parameters.

[0013] Further, the step of training, validating, and / or testing the multimodal fusion prediction model using the multi-dimensional feature set to obtain a target prediction model for the probability prediction of electricity supply and demand factors includes: The multi-dimensional feature set is divided into a training sample set, a validation sample set, and a test sample set according to a preset ratio; and... A hybrid loss function combining mean squared error loss and negative log-likelihood loss is constructed; the mean squared error loss is used to optimize the accuracy of the prediction mean, and the negative log-likelihood loss is used to optimize the fit of the probability distribution. The model is trained based on the training sample set and the hybrid loss function, and the hyperparameters are initialized using the Adam optimizer to achieve parameter learning of the multimodal fusion prediction model. Cross-validation is performed based on the validation sample set to adjust the hyperparameters of the multimodal fusion prediction model and obtain the target prediction model. The reliability of the target prediction model is evaluated based on the test sample set. It is determined whether the prediction interval coverage and average bandwidth error meet the preset criteria. If the preset criteria are not met, the target prediction model is retrained.

[0014] Furthermore, the method also includes: A Dropout layer is added to the feature encoding module of the multimodal fusion prediction model, and an L2 regularization constraint is introduced into the hybrid loss function to avoid overfitting during model training.

[0015] Furthermore, the confidence level information includes confidence intervals corresponding to different confidence levels; the probability density information includes probability density curves.

[0016] According to a second aspect of the present invention, the present invention provides a probability prediction system for power supply and demand factors under extreme weather scenarios, the system comprising: The data acquisition module is used to acquire historical real-time meteorological data, standardize the historical real-time meteorological data according to extreme weather warning standards, and construct an extreme weather scenario library. The feature extraction module is used to extract a multi-dimensional feature set from the extreme weather data in the extreme weather scenario database, including core features of extreme weather, historical evolution features of power supply and demand, and cross-correlation features between extreme weather and power supply and demand. The model building module is used to build a multimodal fusion prediction model that includes a feature encoding module, an attention fusion module, and a probability prediction output module. The model training module is used to train, validate, and test the multimodal fusion prediction model using the multidimensional feature set to obtain a target prediction model for the probability prediction of power supply and demand factors. The probability prediction module is used to acquire actual weather forecast data, input the actual weather forecast data into the target prediction model, and obtain prediction results including the predicted mean of power supply and demand factors, confidence information and / or probability density information.

[0017] The present invention, by adopting the above technical solution, has at least the following beneficial effects: This invention proposes a method for probabilistic prediction of power supply and demand factors under extreme weather scenarios. The method includes: acquiring historical data of power supply and demand factors and historical real weather data; standardizing the historical real weather data according to extreme weather warning standards to construct an extreme weather scenario database; extracting a multi-dimensional feature set containing core features of extreme weather, historical evolution features of power supply and demand, and cross-correlation features between extreme weather and power supply and demand for the extreme weather data and the historical data of power supply and demand factors in the extreme weather scenario database; constructing a multi-modal fusion prediction model including a feature encoding module, an attention fusion module, and a probability prediction output module; training, validating, and / or testing the multi-modal fusion prediction model using the multi-dimensional feature set to obtain a target prediction model for probabilistic prediction of power supply and demand factors; acquiring actual weather forecast data and inputting the actual weather forecast data into the target prediction model to obtain prediction results containing the predicted mean, confidence information, and / or probability density information of power supply and demand factors. This invention achieves high-precision, reliable, and scenario-adaptive prediction of power supply and demand factors under extreme weather scenarios through precise improvements in four core aspects: data processing, feature engineering, model architecture, and training and evaluation.

[0018] It should be understood that the above general description and the following detailed description are exemplary and explanatory only, and are not intended to limit the invention. Attached Figure Description

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

[0020] Figure 1 A flowchart illustrating a method for predicting the probability of power supply and demand factors under extreme weather conditions, provided by an embodiment of the present invention, is shown. Figure 2 A schematic diagram of the process for constructing an extreme weather scenario database according to an embodiment of the present invention is shown; Figure 3 A schematic diagram of the process for extracting multi-dimensional feature sets according to an embodiment of the present invention is shown; Figure 4 A flowchart illustrating a probability prediction method for power supply and demand factors under extreme weather scenarios provided by another embodiment of the present invention is shown. Figure 5 A schematic diagram of the process for constructing a target prediction model according to an embodiment of the present invention is shown; Figure 6This diagram illustrates the structure of a probability prediction system for power supply and demand factors under extreme weather conditions, provided in an embodiment of the present invention. Figure 7 This invention provides a schematic diagram of the structure of a power supply and demand factor probability prediction system under extreme weather conditions, according to another embodiment of the present invention. Figure 8 A schematic diagram of the physical structure of a computer device provided in an embodiment of the present invention is shown. Detailed Implementation

[0021] Exemplary embodiments of the present disclosure will now be described in more detail with reference to the accompanying drawings. While exemplary embodiments of the present disclosure are shown in the drawings, it should be understood that the present disclosure may be implemented in various forms and should not be limited to the embodiments set forth herein. Rather, these embodiments are provided so that this disclosure will be thorough and complete, and will fully convey the scope of the disclosure to those skilled in the art.

[0022] It should be noted that, in this document, relational terms such as "first" and "second" are used merely to distinguish one entity or operation from another, and do not necessarily require or imply any such actual relationship or order between these entities or operations. Furthermore, the terms "comprising," "including," or any other variations thereof are intended to cover non-exclusive inclusion, such that a process, method, article, or apparatus that comprises a list of elements includes not only those elements but also other elements not expressly listed, or elements inherent to such a process, method, article, or apparatus. Without further limitations, an element defined by the phrase "comprising..." does not exclude the presence of additional identical elements in the process, method, article, or apparatus that includes said element.

[0023] Existing technologies for predicting power supply and demand factors under extreme weather scenarios face multiple challenges, resulting in prediction accuracy, reliability, and scenario adaptability failing to meet the core requirements of power system dispatching decisions. Specifically, the following core technical problems and their causes exist: First, extreme weather datasets are poorly constructed and poorly adapted to actual forecast scenarios. Existing technologies fail to standardize the classification of historical meteorological data according to national / local authoritative extreme weather warning standards, resulting in vague definitions and chaotic level classifications of extreme weather samples, making it impossible to accurately match actual weather warning scenarios. Furthermore, the format, field types, and time granularity of historical data differ from mainstream weather forecast data, creating a "data gap" that makes it difficult for trained models to directly connect with actual forecast data for prediction. In addition, the datasets lack refined scenario labels of "weather type - warning level," failing to support differentiated feature engineering and model training, and making it difficult to address the personalized prediction needs of different extreme weather scenarios.

[0024] Secondly, feature engineering lacks specificity, and the correlation between extreme weather and power supply and demand is not explored in depth. Existing technologies have significant limitations in feature extraction: on the one hand, the mining of extreme weather features is incomplete, failing to fully extract core features reflecting the intensity, duration, and spatial distribution of extreme weather, and thus unable to characterize the dynamic evolution of extreme weather; on the other hand, it ignores the deep correlation between extreme weather and power supply and demand factors, failing to construct cross-correlation features such as nonlinear correlations, lag effects, and scenario-specific features, resulting in low feature recognition and weak correlation in the input model, and an inability to accurately quantify the impact mechanism of extreme weather on power supply and demand. Furthermore, the feature selection methods are simplistic, easily retaining redundant features and leading to model overfitting.

[0025] Furthermore, the predictive model architecture lacks adaptability and cannot cope with the complexity and uncertainty of extreme scenarios. Existing technologies mostly employ single-modal models (such as simple LSTM and ARIMA), which struggle to simultaneously adapt to the spatial characteristics of extreme weather, the temporal characteristics of power supply and demand, and the cross-correlation between the two. The fusion effect of multi-source heterogeneous features is poor, making it difficult to highlight key influencing features. Moreover, traditional models mostly output deterministic prediction results without incorporating probability distribution design, failing to quantify prediction uncertainty. Under extreme weather conditions, the randomness of power supply and demand changes is strong, and deterministic results cannot provide a risk assessment basis for dispatch decisions. In addition, the models have weak generalization ability and are susceptible to overfitting due to uneven distribution of extreme samples and data noise, resulting in large fluctuations in prediction accuracy under new extreme weather scenarios.

[0026] Finally, the model training and evaluation system is inadequate, and the reliability of the prediction results is not guaranteed. Existing technologies do not optimize the proportion of extreme samples in their dataset partitioning, resulting in insufficient model learning for extreme scenarios. The loss function design is simplistic, focusing only on the accuracy of the prediction mean while neglecting the fit of the probability distribution, failing to balance prediction accuracy with the effectiveness of uncertainty quantification. Furthermore, the lack of specific reliability assessment metrics for extreme weather scenarios (such as prediction interval coverage and average bandwidth error) makes it impossible to effectively verify the credibility of the prediction results. Directly using substandard results for scheduling decisions could pose a risk to power grid safety.

[0027] This invention addresses the shortcomings of existing technologies in predicting power supply and demand factors under extreme weather conditions by making precise improvements in four core aspects: data processing, feature engineering, model architecture, and training and evaluation. This results in a complete process solution encompassing "standardized data construction - deep feature mining - multimodal fusion prediction - accurate reliability verification." Specifically, this invention provides a method for probabilistic prediction of power supply and demand factors under extreme weather conditions, such as... Figure 1 As shown, it may include at least the following steps S101~S105: Step S101: Obtain historical data on power supply and demand factors and historical real meteorological data. Standardize the historical real meteorological data according to the extreme meteorological warning standards to construct an extreme meteorological scenario database.

[0028] The core objective of this step is to standardize and classify historical real meteorological data according to authoritative extreme weather warning standards, generating an extreme weather dataset that is completely consistent with the format of actual weather forecast data, thus laying the foundation for subsequent prediction models to adapt to actual forecast scenarios. For example... Figure 2 As shown, the specific implementation may include the following steps S101-1 to S101-5: Step S101-1: Based on the time granularity consistent with the actual weather forecast data, collect multi-dimensional historical real weather data for the target area within the target time period; Step S101-2: Determine the threshold values ​​for classifying various extreme weather scenarios based on the extreme weather prediction standards; Step S101-3: Filter the extreme weather sample set from historical real meteorological data based on the level classification threshold; Step S101-4: Refer to the data format of actual weather forecast data, reconstruct the format of the extreme weather dataset to obtain an extreme weather dataset with field types and time granularity consistent with the actual weather forecast data. Step S101-5: Label the extreme weather dataset with scene tags to form a multi-dimensional extreme weather scene library.

[0029] In this embodiment of the invention, the extreme weather prediction standard can be the extreme weather warning standard issued by national / local authorities. By defining the threshold for extreme weather sample levels through this standard, the problems of ambiguous extreme weather sample definition and chaotic level classification are solved. Next, historical real weather data can be standardized using mainstream weather forecast formats (such as JSON) to ensure that the data field types and time granularity are consistent with the actual forecast data, eliminating the "data gap" problem and achieving seamless integration between historical data and actual forecast scenarios. Finally, by adding scenario tags of "weather type - warning level," a refined extreme weather scenario library is constructed, providing a high-quality, highly adaptable data foundation for subsequent differentiated feature engineering and model training, thereby solving the problems of non-standard dataset construction and poor scenario adaptability.

[0030] Taking typhoon forecasting as an example, historical typhoon events from the past 10 years can be categorized according to the typhoon warning standards issued by the National Meteorological Administration (such as GB / T 27962-2011), classifying them into levels such as "Typhoon Blue Warning," "Typhoon Yellow Warning," "Typhoon Orange Warning," and "Typhoon Red Warning," and labeling each historical typhoon event with a corresponding "Typhoon-Warning Level" tag. Furthermore, all historical typhoon meteorological data (including wind speed, rainfall, air pressure, and path) can be uniformly converted into JSON format, consistent with the output of the real-time weather forecast system, ensuring complete consistency in data fields and time granularity. This results in a standardized typhoon dataset with finely labeled scenarios, resolving the mismatch between historical data and real-time forecasts, and the ambiguous definition of typhoon samples. This provides high-quality, directly usable typhoon data for subsequent feature extraction and model training.

[0031] Step S102: Extract a multi-dimensional feature set from the extreme weather data and historical data of power supply and demand factors in the extreme weather scenario database. This set includes core features of extreme weather, historical evolution features of power supply and demand, and cross-correlation features between extreme weather and power supply and demand.

[0032] Understandably, this step is crucial for improving the accuracy of electricity supply and demand factor predictions. The feature engineering portion focuses on uncovering the deep correlation between extreme weather events and electricity supply and demand factors, providing the model with highly identifiable and strongly correlated input features. For example... Figure 3 As shown, the specific implementation may include the following steps S102-1 to S102-3: Step S102-1: Based on extreme meteorological data, extract the core characteristics of extreme meteorological events that characterize the event intensity, persistence, trend and / or spatial distribution of extreme meteorological events.

[0033] Specifically, the extraction of core features of extreme weather events revolves around their intensity, duration, and trends. Intensity features include the extreme values, mean values, and peak times of various extreme weather parameters; duration features encompass the duration of extreme weather events, the sequence of intensity level changes, and the magnitude of changes in meteorological parameters over adjacent periods; if multiple meteorological monitoring stations exist in the target area, the spatial distribution characteristics of meteorological parameters at each station also need to be extracted, such as the standard deviation of wind speed between stations and the spatial gradient of precipitation, to reflect the spatial impact range of extreme weather events. Taking typhoon forecasting as an example, features such as the typhoon's maximum wind speed, duration, rate of wind speed change, typhoon eye path, and spatial distribution of rainfall are extracted to comprehensively characterize the typhoon's intensity, duration, and spatial impact.

[0034] Step S102-2: Based on historical data of electricity supply and demand factors, extract historical evolution characteristics of electricity supply and demand that characterize time-series evolution, fluctuation characteristics, and / or supply and demand balance characteristics.

[0035] Specifically, the extraction of historical evolution characteristics of power supply and demand mainly focuses on its time-series evolution patterns and fluctuation characteristics. These include time-series statistical features such as historical values, average values, and extreme values ​​of power load and wind / solar power; fluctuation characteristics involving the volatility and trend features of power supply and demand factors, as well as abrupt change features detected using the sliding window method; and supply-demand balance features, such as the difference between grid supply capacity and power load, the proportion of wind / solar power in total power supply, and changes in the proportion of industrial and residential loads, to adapt to time-series forecasting demand under extreme weather conditions. Taking typhoon forecasting as an example, features such as load curves, load volatility, load peak and valley values, and load recovery speed before and after typhoons can be extracted from historical typhoon periods.

[0036] Step S102-3: Based on extreme weather data and historical data of power supply and demand factors, analyze the nonlinear correlation between extreme weather events and power supply and demand factors to obtain the cross-correlation characteristics between extreme weather and power supply and demand.

[0037] In this embodiment of the invention, the focus of the extreme weather-power supply and demand cross-correlation feature is to explore the nonlinear correlation between extreme weather and power supply and demand factors, and to construct cross-features that can quantify the mutual influence between the two. Specifically, this includes the following aspects: the extraction of extreme weather-power supply and demand cross-correlation features focuses on exploring the nonlinear correlation between the two, including calculating the correlation coefficient between extreme weather parameters and power supply and demand factors under different lag durations to obtain lag correlation features; calculating the sensitivity coefficient under different intensity levels through piecewise linear regression to quantify the influence sensitivity features; constructing scenario-specific cross-features for different extreme weather scenarios; and extracting recovery features such as the time difference between the end of the extreme weather event and the time when the power supply and demand factors recover to normal levels, and the load / power change rate during the recovery process.

[0038] Taking typhoon forecasting as an example, we can analyze the nonlinear relationship between typhoon wind speed and load (for example, after the wind speed reaches a certain threshold, the load may decrease due to power outages), as well as the lag effect of load recovery after the typhoon makes landfall.

[0039] As an optional embodiment of the present invention, to avoid feature redundancy leading to model overfitting, after extracting the multi-dimensional feature set, as follows: Figure 4 As shown, a method for probabilistic prediction of power supply and demand factors under extreme weather scenarios may further include the following step S106: Step S106: Use a two-step method to filter features in the multi-dimensional feature set to obtain the optimized multi-dimensional feature set.

[0040] The "two-step method" includes a preliminary screening stage and a fine screening stage. In the preliminary screening stage, the importance score of each feature in the multi-dimensional feature set to the prediction result of the power supply and demand factor is calculated based on the tree model, and the importance scores are sorted. Features below a first preset threshold are removed to obtain the preliminary screening feature set. In the fine screening stage, the mutual information entropy between the features in the preliminary screening feature set is calculated. For feature pairs with mutual information entropy higher than a second preset threshold, features with lower importance scores in the feature pairs are removed to obtain the optimized multi-dimensional feature set.

[0041] In practical applications, the initial screening stage can utilize the random forest algorithm to calculate the importance score of each feature to the prediction results of power supply and demand factors, eliminating features with excessively low scores. Taking typhoon forecasting as an example, the importance of hundreds of features such as maximum wind speed, rainfall, typhoon path, historical load, and air pressure to load forecasting is evaluated. For instance, if the importance score of a feature (such as the air pressure at a certain location 24 hours ago) is below 0.01, it is initially eliminated. In the fine screening stage, the mutual information entropy of the initially screened features is calculated, eliminating highly redundant features with a mutual information entropy greater than 0.8 and retaining features with higher importance scores. For example, if the mutual information entropy of "maximum typhoon wind speed" and "average typhoon wind speed" is found to be as high as 0.95 (far greater than 0.8), indicating high redundancy, then the more important "maximum typhoon wind speed" is retained, while "average typhoon wind speed" is eliminated. Simultaneously, cross-validation can be used to verify the improvement effect of each feature on the model's prediction accuracy, ensuring that after each feature elimination, the model's prediction accuracy on the validation set does not decrease but rather improves. Finally, an optimal feature set with low redundancy and high correlation is determined, and subsequent model training steps are performed based on the optimal feature set.

[0042] Step S103: Construct a multimodal fusion prediction model that includes a feature encoding module, an attention fusion module, and a probability prediction output module.

[0043] In this embodiment of the invention, the feature encoding module integrates a convolutional neural network (CNN), a long short-term memory (LSTM) network, and a fully connected layer, employing differentiated encoding methods for different types of input features. Specifically, CNN encoding is used to capture spatial correlation information for the spatial distribution features of extreme weather, LSTM encoding is used to capture temporal evolution patterns for the temporal features of power supply and demand, and fully connected layer encoding is used to enhance nonlinear expression for cross-correlation features. The attention fusion module employs a multi-head attention mechanism to adaptively weight and fuse the multimodal encoded features, highlighting features that significantly affect the prediction results and improving the model's adaptability to extreme weather scenarios. The probability prediction output module uses a Bayesian neural network structure, introducing the probability distribution of parameters, and outputting the probability distribution results of power supply and demand factors, including the prediction mean, confidence intervals at different confidence levels, and probability density functions, thereby quantifying prediction uncertainty.

[0044] Taking typhoon forecasting as an example, spatial features such as typhoon paths can be input into a convolutional neural network (CNN) for encoding, while temporal features such as wind speed and load time series can be input into a long short-term memory network (LSTM) for encoding. Cross-correlation features between wind speed and load can be input into a fully connected layer for encoding. A multi-head attention mechanism is employed to fuse the features encoded by the CNN, LSTM, and fully connected layers. The model automatically assigns weights to different features based on the characteristics of the current typhoon (e.g., wind speed features have higher weights when typhoon intensity is critical, and path features have higher weights when typhoon path is critical). The fused features are then input into a Bayesian neural network, which outputs the average predicted power load during the typhoon period, along with 90% and 95% confidence intervals and the probability density distribution curve of the load. This solves the problem that a single model cannot simultaneously handle multi-source heterogeneous features (spatial, temporal, and cross-correlation), and quantifies the uncertainty of the prediction through probability output, such as "the load during the typhoon is expected to be X MW, but there is a 90% probability that it will be between Y and Z MW," providing a more comprehensive risk assessment basis for power dispatch.

[0045] Step S104: Use a multi-dimensional feature set to train, validate and / or test the multi-modal fusion prediction model to obtain the target prediction model for the probability prediction of power supply and demand factors.

[0046] like Figure 5 As shown, the specific implementation of this step may include the following steps S104-1 to S104-4: Step S104-1: Divide the multi-dimensional feature set into training sample set, validation sample set and test sample set according to a preset ratio; and construct a hybrid loss function that combines mean squared error loss and negative log-likelihood loss. Step S104-2: Model training is performed based on the training sample set and the hybrid loss function. The Adam optimizer is used to initialize the hyperparameters in order to achieve parameter learning of the multimodal fusion prediction model. Step S104-3: Perform cross-validation based on the validation sample set to adjust the hyperparameters of the multimodal fusion prediction model and obtain the target prediction model; Step S104-4: Based on the test sample set, perform a reliability assessment on the target prediction model to determine whether the prediction interval coverage and average bandwidth error meet the preset compliance conditions. If the preset compliance conditions are not met, retrain the target prediction model.

[0047] First, the optimal feature set can be divided into a training sample set, a validation sample set, and a test sample set in a 7:2:1 ratio, for model parameter learning, hyperparameter tuning, and performance evaluation, respectively. In this embodiment of the invention, the loss function can be a hybrid loss function combining mean squared error loss and negative log-likelihood loss, mathematically expressed as: L=a L MSE +(1-a) L NLL .

[0048] Where L represents the mixture loss function; L MSE L represents the mean squared error loss, used to optimize the accuracy of the predicted mean; NLL This represents the negative log-likelihood loss, used to optimize the fit of the probability distribution. By adjusting the weighting coefficient 'a' (e.g., setting it to 0.5), a balance can be struck between optimizing the accuracy of the predicted mean and the reliability of the predicted interval, ensuring that the model can accurately predict the load during typhoons while reasonably assessing the range of load fluctuations. Furthermore, during model training, the Adam optimizer can be used to update model parameters and configure hyperparameters such as the initial learning rate, number of iterations, and batch size. In addition, by combining validation set cross-validation with a grid search method, the model architecture parameters of the convolutional neural network (CNN), such as the kernel size and the number of hidden layer units, were systematically adjusted to ensure the model's prediction performance under different extreme weather scenarios and avoid fluctuations in prediction accuracy caused by architecture mismatch.

[0049] Taking typhoon prediction as an example, the initial learning rate was preset to 0.001, and the batch size was 32. To find the optimal model architecture, different CNN convolutional kernel sizes (e.g., 3x3, 5x5) and LSTM hidden layer unit numbers (e.g., 64, 128) were tried through grid search, and the performance of each combination was evaluated on the validation set. For example, it was found that when the CNN convolutional kernel size was 3x3 and the LSTM hidden layer unit number was 128, the model achieved the best prediction accuracy and reliability in typhoon scenarios, thus determining the final model architecture.

[0050] As an optional embodiment of the present invention, a Dropout layer can be added to the feature encoding module and the fully connected layer, and an L2 regularization constraint can be introduced into the hybrid loss function to avoid overfitting during model training and improve the model's generalization ability.

[0051] After the prediction results are output, this embodiment of the invention also provides a closed-loop optimization mechanism based on reliability indicators. Specifically, actual meteorological forecast data consistent with historical data formats is input into the trained and optimized target prediction model. The model outputs the probability distribution of power supply and demand factors based on the input meteorological characteristics and corresponding historical power supply and demand characteristics. Subsequently, reliability is evaluated using Prediction Interval Coverage (PICP) and Mean Bandwidth Error (MPIW). If PICP reaches a preset threshold (e.g., a 90% confidence interval requires PICP ≥ 90%) and MPIW is less than a preset maximum value, the final probability prediction result is output; otherwise, the prediction result is considered unreliable. In this case, the model training process is returned to adjust the model architecture or hyperparameters and retrain until the reliability indicators meet the requirements.

[0052] In typhoon load forecasting, if the model's 90% confidence interval prediction result, after evaluation on the test set, shows a PICP of only 85% (below the 90% threshold) or an MPIW that is too large (e.g., the prediction interval is too wide, resulting in low reference value), the current model will be deemed unreliable. In this case, the prediction result will not be output directly, but the model will be automatically retrained. This may involve adjusting the Dropout rate, L2 regularization strength, or even performing a grid search again to optimize the model architecture, until the PICP and MPIW metrics meet preset requirements, ensuring that the final output prediction result is highly reliable.

[0053] Step S105: Obtain actual weather forecast data, input the actual weather forecast data into the target prediction model, and obtain prediction results including the predicted mean of power supply and demand factors, confidence information and / or probability density information.

[0054] The final evaluation of the reliable target prediction model will be used to predict the power supply and demand factors based on actual meteorological forecast data. The final output results include the predicted mean of power supply and demand factors, 90% / 95% confidence intervals, and probability density curves, providing accurate prediction results and uncertainty information for power system dispatch decisions.

[0055] This invention provides a method for probabilistic prediction of power supply and demand factors under extreme weather scenarios, which has at least the following advantages compared with the prior art: 1) More comprehensive forecasting targets, providing higher value for decision support: This invention breaks through the limitations of existing technologies that only forecast a single power load indicator, expanding the forecasting targets to include power supply and demand factors, comprehensively covering multi-dimensional indicators such as power load (demand side), output of renewable energy sources such as wind / solar power, and grid supply capacity (supply side). These comprehensive indicators can fully reflect the supply and demand balance of the power system under extreme weather conditions, providing a more comprehensive and accurate core basis for power system dispatching decisions. Compared to single load forecasting, it is better suited to the actual needs of coordinated supply and demand dispatching in new power systems.

[0056] 2) Enhanced data adaptability and higher scenario fit: This invention innovatively introduces authoritative extreme weather warning standards (such as GB / T 27962-2011) to standardize the classification of historical real meteorological data, clarify the threshold for extreme weather levels, and label the scenarios with "meteorological type - warning level." Simultaneously, it standardizes data fields and granularity according to mainstream weather forecast formats. This completely solves the problem of poor scenario adaptability caused by the "fuzzy screening of similar weather days" in existing technologies, achieving seamless integration between historical data and actual forecast scenarios, and making model training more closely match the characteristics of real extreme weather scenarios.

[0057] 3) Deeper Feature Mining, More Robust Predictive Support: This invention constructs a feature engineering system of "three-dimensional feature extraction + two-step screening," which not only comprehensively extracts the core features of extreme weather (intensity, duration, and spatial distribution) and the evolution features of power supply and demand, but also deeply mines the nonlinear correlation, lag effects, and scene-specific correlations between the two. Then, through scientific screening, it obtains an optimal feature set with low redundancy and high correlation. Compared with existing technologies that only extract basic statistical features, the feature system of this invention can more accurately characterize the impact mechanism of extreme weather on power supply and demand, providing solid support for high-precision prediction.

[0058] 4) More Adaptable Model Architecture and More Accurate Uncertainty Quantification: This invention designs a multimodal fusion model of "feature encoding-attention fusion-probability output." It uses CNN, LSTM, and fully connected layers to differentially and accurately encode multi-source heterogeneous features, and then adaptively fuses key features through a multi-head attention mechanism. Simultaneously, a Bayesian neural network is introduced to achieve probability output, providing results such as prediction mean, confidence interval, and probability density function, accurately quantifying prediction uncertainty. This architecture overcomes the adaptability limitations of existing single GA-BP models, adapting to multimodal feature fusion requirements and providing crucial uncertainty information for emergency dispatch risk decision-making in extreme scenarios.

[0059] 4) A more comprehensive training and evaluation system ensures greater predictive reliability: This invention specifically optimizes the entire training and evaluation process. Dataset partitioning ensures a uniform distribution of extreme samples, and a hybrid loss function balances the accuracy of the prediction mean with the fit of the probability distribution. Grid search, Dropout layers, and L2 regularization effectively prevent overfitting. Furthermore, it introduces specific reliability evaluation metrics such as Prediction Interval Coverage (PICP) and Mean Bandwidth Error (MPIW), forming a closed-loop mechanism of "training-validation-optimization." Compared to the single training and evaluation logic of existing technologies, this invention ensures stable predictive performance of the model under various extreme weather scenarios, significantly improving the reliability of prediction results.

[0060] Furthermore, as Figure 1 In specific implementation, embodiments of the present invention provide a probability prediction system for power supply and demand factors under extreme weather scenarios, such as... Figure 6 As shown, the system may include: a data acquisition module 610, a feature extraction module 620, a model building module 630, a model training module 640, and a probability prediction module 650.

[0061] Data acquisition module 610 can be used to acquire historical real meteorological data, standardize the historical real meteorological data according to extreme weather warning standards, and construct an extreme weather scenario library: The feature extraction module 620 can be used to extract a multi-dimensional feature set from extreme weather data in the extreme weather scenario database, including core features of extreme weather, historical evolution features of power supply and demand, and cross-correlation features between extreme weather and power supply and demand. The model building module 630 can be used to build a multimodal fusion prediction model that includes a feature encoding module, an attention fusion module, and a probability prediction output module; The model training module 640 can be used to train, validate, and test a multimodal fusion prediction model using a multi-dimensional feature set to obtain a target prediction model for the probability prediction of power supply and demand factors. The probability prediction module 650 can be used to acquire actual weather forecast data, input the actual weather forecast data into the target prediction model, and obtain prediction results including the predicted mean of power supply and demand factors, confidence information and / or probability density information.

[0062] Optionally, such as Figure 7 As shown, another embodiment of the present invention provides a probability prediction system for power supply and demand factors under extreme weather scenarios, which may further include: a feature screening module 660.

[0063] The feature filtering module 660 can be used to perform feature filtering on a multi-dimensional feature set using a two-step method to obtain a better multi-dimensional feature set, specifically including: The importance score of each feature in the multi-dimensional feature set to the prediction result of power supply and demand factors is calculated based on the tree model, and the importance scores are sorted. Features below the first preset threshold are removed to obtain the preliminary feature set. Calculate the mutual information entropy between features in the initial feature set. For feature pairs with mutual information entropy higher than the second preset threshold, remove features with lower importance scores from the feature pairs to obtain the optimized multi-dimensional feature set.

[0064] It should be noted that other corresponding descriptions of the functional modules involved in the power supply and demand factor probability prediction system under extreme weather scenarios provided in this embodiment of the invention can be found in [reference needed]. Figure 1 The corresponding description of the method shown will not be repeated here.

[0065] Based on the above, Figure 1 Accordingly, embodiments of the present invention also provide a computer-readable storage medium storing a computer program thereon, which, when executed by a processor, implements the steps of the power supply and demand factor probability prediction method under extreme weather scenarios of any of the above embodiments.

[0066] Based on the above, Figure 1 The method shown and as Figure 7 The embodiment of the system shown in the invention also provides a physical structure diagram of a computer device, such as... Figure 8 As shown, the computer device may include a communication bus, a processor, a memory, and a communication interface. It may also include input / output interfaces and a display device. The various functional units can communicate with each other via the bus. The memory stores a computer program, and the processor executes the program stored in the memory to perform the steps of the power supply and demand factor probability prediction method under extreme weather scenarios described in the above embodiment.

[0067] Those skilled in the art will clearly understand that the specific working process of the systems, devices, modules and units described above can be referred to the corresponding process in the foregoing method embodiments. For the sake of brevity, it will not be repeated here.

[0068] Furthermore, the functional units in the various embodiments of the present invention can be physically independent of each other, or two or more functional units can be integrated together, or all functional units can be integrated into one processing unit. The integrated functional units described above can be implemented in hardware, or in software or firmware.

[0069] Those skilled in the art will understand that if the integrated functional unit is implemented in software and sold or used as an independent product, it can be stored in a computer-readable storage medium. Based on this understanding, the technical solution of the present invention, or all or part of it, can be embodied in the form of a software product. This computer software product is stored in a storage medium and includes several instructions to cause a computing device (e.g., a personal computer, server, or network device) to execute all or part of the steps of the methods described in the embodiments of the present invention when running the instructions. The aforementioned storage medium includes various media capable of storing program code, such as USB flash drives, portable hard drives, read-only memory (ROM), random access memory (RAM), magnetic disks, or optical disks.

[0070] Alternatively, all or part of the steps of the foregoing method embodiments can be implemented by hardware (such as a computing device, personal computer, server, or network device) related to program instructions. The program instructions can be stored in a computer-readable storage medium. When the program instructions are executed by the processor of the computing device, the computing device executes all or part of the steps of the methods described in the various embodiments of the present invention.

[0071] Finally, it should be noted that the above embodiments are only used to illustrate the technical solutions of the present invention, and not to limit them. Although the present invention has been described in detail with reference to the foregoing embodiments, those skilled in the art should understand that within the spirit and principles of the present invention, modifications can still be made to the technical solutions described in the foregoing embodiments, or equivalent substitutions can be made to some or all of the technical features therein; and these modifications or substitutions do not cause the corresponding technical solutions to depart from the protection scope of the present invention.

Claims

1. A method for probabilistic prediction of power supply and demand factors under extreme weather scenarios, characterized in that, The method includes: Historical data on power supply and demand factors and historical real-time meteorological data are acquired. The historical real-time meteorological data is then standardized according to extreme weather warning standards to construct an extreme weather scenario database. For the extreme weather data in the extreme weather scenario database and the historical data of power supply and demand factors, a multi-dimensional feature set is extracted, which includes the core features of extreme weather, the historical evolution features of power supply and demand, and the cross-correlation features between extreme weather and power supply and demand. Construct a multimodal fusion prediction model that includes a feature encoding module, an attention fusion module, and a probability prediction output module; The multi-dimensional feature set is used to train, validate, and / or test the multimodal fusion prediction model to obtain a target prediction model for the probability prediction of power supply and demand factors. Obtain actual weather forecast data, input the actual weather forecast data into the target prediction model, and obtain prediction results including the predicted mean of power supply and demand factors, confidence information and / or probability density information.

2. The method according to claim 1, characterized in that, The process of acquiring historical real meteorological data involves dividing the historical real meteorological data according to extreme weather prediction standards to generate an extreme weather dataset consistent with the format of actual weather forecast data, including: Based on the time granularity consistent with the actual weather forecast data, multi-dimensional historical real weather data are collected for the target area within the target time period; Based on the aforementioned extreme weather prediction standards, threshold values ​​for classifying various extreme weather scenarios are determined. The extreme weather sample set in the historical real meteorological data is filtered based on the aforementioned classification threshold. Referring to the data format of the actual weather forecast data, the extreme weather dataset is restructured to obtain an extreme weather dataset with field types and time granularity consistent with the actual weather forecast data; The extreme weather dataset is labeled with scene tags to form a multi-dimensional extreme weather scene library; wherein the scene tags include weather type and warning level.

3. The method according to claim 1, characterized in that, The extracted feature set comprises a multi-dimensional feature set including core features of extreme weather, historical evolution features of power supply and demand, and cross-correlation features between extreme weather and power supply and demand, including: Based on the extreme meteorological data, core characteristics of extreme meteorological events are extracted to characterize their intensity, duration, trend, and / or spatial distribution. The intensity includes at least one of the extreme values, mean values, and peak occurrence times of various extreme meteorological parameters. The duration includes the duration of the extreme meteorological event. The trend includes the intensity level change sequence of the extreme meteorological event and / or the magnitude of meteorological parameter changes in adjacent time periods. The spatial distribution includes the standard deviation of wind speed between stations and / or the spatial gradient of precipitation. as well as, Based on the historical data of the aforementioned power supply and demand factors, historical evolution characteristics of power supply and demand are extracted to characterize their time-series evolution, volatility, and / or supply-demand balance characteristics. The time-series evolution characteristics include at least one of the historical contemporaneous values, historical averages, and historical extreme values ​​of power load and wind / solar power. The volatility characteristics include the volatility, trend characteristics, and abrupt change characteristics of the power supply and demand factors. The supply-demand balance characteristics include the difference between the grid's power supply capacity and the power load, the proportion of wind / solar power in the total power supply, and / or the changes in the proportion of industrial load to residential load. Based on the extreme weather data and the historical data of the power supply and demand factors, the nonlinear correlation between extreme weather events and power supply and demand factors is analyzed to obtain the cross-correlation characteristics between extreme weather and power supply and demand.

4. The method according to claim 3, characterized in that, The process involves analyzing the nonlinear correlation between extreme weather events and power supply and demand factors based on the extreme weather data and historical data of power supply and demand factors, thereby obtaining the cross-correlation characteristics between extreme weather and power supply and demand, including: Calculate the correlation coefficients between the extreme meteorological parameters and the power supply and demand factors under different lag durations to obtain the lag correlation characteristics; and / or, Sensitivity characteristics are obtained by calculating the sensitivity coefficients between the extreme meteorological parameters and the power supply and demand factors at different intensity levels using piecewise linear regression; and / or, Develop scenario-specific cross-features for different extreme weather scenarios; and / or, The time difference between the end time of the extreme weather event and the time when the power supply and demand factors recover to normal levels, as well as the load / power change rate during the recovery process, are extracted as recovery features.

5. The method according to claim 1, characterized in that, After extracting a multi-dimensional feature set including core features of extreme weather, historical evolution features of power supply and demand, and cross-correlation features between extreme weather and power supply and demand, the method further includes: A two-step method is used to filter the features of the multi-dimensional feature set to obtain an optimized multi-dimensional feature set, specifically including: The importance score of each feature in the multi-dimensional feature set to the prediction result of power supply and demand factors is calculated based on the tree model, and the importance scores are sorted and features below the first preset threshold are removed to obtain a preliminary feature set. Calculate the mutual information entropy between each feature in the preliminary feature set. For feature pairs whose mutual information entropy is higher than a second preset threshold, remove features with lower importance scores from the feature pairs to obtain the optimized multi-dimensional feature set.

6. The method according to claim 1, characterized in that, The construction of the multimodal fusion prediction model, which includes a feature encoding module, an attention fusion module, and a probability prediction output module, includes: A feature encoding module integrating a convolutional neural network architecture, a long short-term memory network architecture, and a fully connected layer is constructed. The convolutional neural network architecture is used to encode the spatial distribution characteristics of the extreme weather; the long short-term memory network architecture is used to encode the temporal characteristics of power supply and demand; and the fully connected layer is used to encode the cross-correlation characteristics between extreme weather and power supply and demand. An attention fusion module employing a multi-head attention mechanism is constructed; the multi-head attention mechanism is used to adaptively weight and fuse features encoded by multimodal approaches. A probability prediction output module employing a Bayesian neural network structure is constructed; the Bayesian neural network structure is used to output prediction results containing the predicted mean of power supply and demand factors, confidence information, and / or probability density information by introducing the probability distribution of parameters.

7. The method according to claim 1, characterized in that, The step of training, validating, and / or testing the multimodal fusion prediction model using the multi-dimensional feature set to obtain a target prediction model for probabilistic prediction of electricity supply and demand factors includes: The multi-dimensional feature set is divided into a training sample set, a validation sample set, and a test sample set according to a preset ratio; and... A hybrid loss function combining mean squared error loss and negative log-likelihood loss is constructed; the mean squared error loss is used to optimize the accuracy of the prediction mean, and the negative log-likelihood loss is used to optimize the fit of the probability distribution. The model is trained based on the training sample set and the hybrid loss function, and the hyperparameters are initialized using the Adam optimizer to achieve parameter learning of the multimodal fusion prediction model. Cross-validation is performed based on the validation sample set to adjust the hyperparameters of the multimodal fusion prediction model and obtain the target prediction model. The reliability of the target prediction model is evaluated based on the test sample set. It is determined whether the prediction interval coverage and average bandwidth error meet the preset criteria. If the preset criteria are not met, the target prediction model is retrained.

8. The method according to claim 7, characterized in that, The method further includes: A Dropout layer is added to the feature encoding module of the multimodal fusion prediction model, and an L2 regularization constraint is introduced into the hybrid loss function to avoid overfitting during model training.

9. The method according to any one of claims 1 to 8, characterized in that, The confidence level information includes confidence intervals corresponding to different confidence levels; the probability density information includes probability density curves.

10. A probability prediction system for power supply and demand factors under extreme weather scenarios, characterized in that, The system includes: The data acquisition module is used to acquire historical real-time meteorological data, standardize the historical real-time meteorological data according to extreme weather warning standards, and construct an extreme weather scenario library. The feature extraction module is used to extract a multi-dimensional feature set from the extreme weather data in the extreme weather scenario database, including core features of extreme weather, historical evolution features of power supply and demand, and cross-correlation features between extreme weather and power supply and demand. The model building module is used to build a multimodal fusion prediction model that includes a feature encoding module, an attention fusion module, and a probability prediction output module. The model training module is used to train, validate, and test the multimodal fusion prediction model using the multidimensional feature set to obtain a target prediction model for the probability prediction of power supply and demand factors. The probability prediction module is used to acquire actual weather forecast data, input the actual weather forecast data into the target prediction model, and obtain prediction results including the predicted mean of power supply and demand factors, confidence information and / or probability density information.

Images