Load prediction method and system under complex weather based on generative adversarial learning

Abstract

The application provides a load prediction method and system under complex weather based on generative adversarial learning, and relates to the technical field of power grids.The application generates synthetic load data under extreme weather by using conditional generative adversarial networks, breaks through the technical bottleneck of the scarcity of historical samples under extreme weather, and significantly enhances the generalization ability of the model to rare meteorological scenarios.Through meteorological data coupling and load component decomposition, the training features of the refined model are refined to ensure that the load prediction model accurately identifies refined features and improves the prediction accuracy of the model.The load prediction model is embedded with physical consistency constraints to ensure that the output results conform to the operation rules of the power system and avoid physical irrationality.The application effectively solves the problem of low accuracy of power load prediction under complex weather by introducing generative adversarial learning, multi-modal feature fusion and physical constraint mechanism, and improves the accuracy of power load prediction under complex weather.

Description

Technical Field

[0001] This invention relates to the field of power grid technology, and in particular to a load forecasting method and system based on generative adversarial learning under complex weather conditions. Background Technology

[0002] Power load forecasting is a key technology for power grid dispatching and energy management, and its accuracy directly affects the safe and economical operation of the power system. Especially during complex weather events such as high temperatures, cold waves, and sudden temperature drops, load characteristics exhibit high nonlinearity and strong abrupt changes, posing a significant challenge to traditional forecasting methods. Existing technologies mainly have the following limitations:

[0003] Extreme weather events occur infrequently, and historical data is scarce. This makes it difficult for statistical models (such as ARIMA and regression analysis) and machine learning models (such as SVM and BP neural networks) to fully learn load patterns under complex weather conditions, resulting in significantly increased prediction bias and lower accuracy in power load forecasting under complex weather conditions. Summary of the Invention

[0004] This invention provides a method and system for load forecasting under complex weather conditions based on generative adversarial learning, which solves the problem of low accuracy in power load forecasting under complex weather conditions and improves the accuracy of power load forecasting under complex weather conditions.

[0005] In a first aspect, the present invention provides a load forecasting method for complex weather conditions based on generative adversarial learning. The method includes: acquiring historical extreme weather data and corresponding actual load data; expanding the sample data under extreme weather conditions in historical periods based on the extreme weather data, actual load data, and a pre-defined conditional generative adversarial network to determine an expanded sample set; refining the features of each expanded sample in the expanded sample set using meteorological data coupling and load component decomposition to determine a refined sample set; performing machine learning based on the refined sample set, a pre-defined prediction model architecture, and physical consistency constraints to generate a load forecasting model for complex weather conditions; and performing load forecasting under complex weather conditions based on the load forecasting model.

[0006] Secondly, embodiments of the present invention provide a load forecasting device for complex weather conditions based on generative adversarial learning. The load forecasting device includes: a communication module and a processing module. The communication module is used to acquire extreme weather data from historical periods and the corresponding actual load data. The processing module is used to expand sample data from historical extreme weather conditions based on the extreme weather data, actual load data, and a preset conditional generative adversarial network to determine an expanded sample set. Based on the expanded sample set, meteorological data coupling and load component decomposition are used to refine the features of each expanded sample in the expanded sample set to determine a refined sample set. Based on the refined sample set, a preset prediction model architecture, and physical consistency constraints, machine learning is performed to generate a load forecasting model for complex weather conditions. Based on the load forecasting model, load forecasting is performed under complex weather conditions.

[0007] Thirdly, embodiments of the present invention provide an electronic device including a memory and a processor. The memory stores a computer program, and the processor is configured to call and run the computer program stored in the memory to perform the steps of the method as described in the first aspect and any possible implementation thereof.

[0008] Fourthly, embodiments of the present invention provide a computer-readable storage medium storing a computer program, characterized in that, when the computer program is executed by a processor, it implements the steps of the method as described in the first aspect and any possible implementation thereof.

[0009] This invention provides a method and system for load forecasting under complex weather conditions based on generative adversarial learning. By utilizing conditional generative adversarial networks to generate synthetic load data under extreme weather conditions, this invention overcomes the technical bottleneck of scarce historical samples for extreme weather and significantly enhances the model's generalization ability to rare meteorological scenarios. Subsequently, through meteorological data coupling and load component decomposition, the training features of the model are refined, improving feature specificity and interpretability, ensuring that the load forecasting model accurately identifies refined features and improving forecast accuracy. Furthermore, physical consistency constraints are embedded in the load forecasting model to ensure that the output results conform to the operating laws of the power system, avoiding the physical irrationality of traditional pure data-driven models. This invention effectively solves the problem of low accuracy in power load forecasting under complex weather conditions by introducing generative adversarial learning, multimodal feature fusion, and physical constraint mechanisms, thereby improving the accuracy of power load forecasting under complex weather conditions. Attached Figure Description

[0010] To more clearly illustrate the technical solutions in the embodiments of the present invention, 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.

[0011] Figure 1 This is a flowchart illustrating a load forecasting method for complex weather conditions based on generative adversarial learning, provided in an embodiment of the present invention.

[0012] Figure 2 This is a schematic diagram of the structure of a load forecasting device under complex weather conditions based on generative adversarial learning, provided in an embodiment of the present invention.

[0013] Figure 3 This is a schematic diagram of the structure of an electronic device provided in an embodiment of the present invention. Detailed Implementation

[0014] In the following description, specific details such as particular system architectures and techniques are set forth for illustrative purposes and not for limitation, in order to provide a thorough understanding of the embodiments of the invention. However, those skilled in the art will understand that the invention can be implemented in other embodiments without these specific details. In other instances, detailed descriptions of well-known systems, apparatuses, circuits, and methods are omitted so as not to obscure the description of the invention with unnecessary detail.

[0015] In the description of this invention, unless otherwise stated, " / " means "or". For example, A / B can mean A or B. The term "and / or" in this document is merely a description of the relationship between related objects, indicating that three relationships can exist. For example, A and / or B can represent: A alone, A and B simultaneously, and B alone. Furthermore, "at least one" and "more than one" refer to two or more. The terms "first," "second," etc., do not limit the quantity or order of execution, and "first," "second," etc., do not necessarily imply differences.

[0016] 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 terms such as "exemplary" or "for example" is intended to present the relevant concepts in a specific manner to facilitate understanding.

[0017] Furthermore, the terms "comprising" and "having," and any variations thereof, used in the description of this application are intended to cover non-exclusive inclusion. For example, a process, method, system, product, or device that includes a series of steps or modules is not limited to the steps or modules listed, but may optionally include other steps or modules not listed, or may optionally include other steps or modules inherent to such process, method, product, or device.

[0018] To make the objectives, technical solutions, and advantages of the present invention clearer, the following description will be provided in conjunction with the accompanying drawings and specific embodiments.

[0019] like Figure 1 As shown, this embodiment of the invention provides a load forecasting method for complex weather conditions based on generative adversarial learning. The method includes steps S101-S104.

[0020] S101. Obtain extreme weather data for historical periods, as well as the actual load data corresponding to the extreme weather data.

[0021] In some embodiments, extreme weather data refers to quantitative information on various meteorological elements characterizing extreme weather events, recorded and stored by a meteorological monitoring system over a historical period. Extreme weather events include extreme high-temperature events, extreme low-temperature events, transitional events between high and low temperatures, heavy rainfall events, typhoons, and sandstorms.

[0022] For example, extreme weather data includes time series data such as temperature, humidity, wind speed and direction, solar radiation intensity, precipitation, and air pressure;

[0023] In some embodiments, actual load data refers to the power load data actually collected and recorded by the power grid dispatch center's energy management system (EMS) or electricity consumption information collection system during the same period of the aforementioned extreme weather events. This data is strictly aligned with the extreme weather data in time, forming a condition-sample pairing relationship, which is the basis for training the model.

[0024] For example, actual load data includes total active power, regional power supply load, substation bus load, etc.

[0025] As one possible implementation, embodiments of the present invention can synchronously acquire time-aligned extreme weather data and actual load data from meteorological monitoring systems and power dispatching systems.

[0026] S102. Based on extreme weather data, actual load data, and a pre-set conditional generative adversarial network, expand the sample data under extreme weather conditions in historical periods to determine the expanded sample set.

[0027] As one possible implementation, step S102 can be specifically implemented as steps S1021-S1027.

[0028] S1021. Using extreme weather data as input, the actual load data corresponding to the extreme weather data as the real sample, and setting a random noise vector that conforms to a Gaussian distribution, multiple batches of sampling samples are constructed.

[0029] S1022. Based on the samples sampled in the current batch during the current iteration, keep the parameters of the generator in the conditional generative adversarial network unchanged, update the parameters of the discriminator in the conditional generative adversarial network, and obtain the updated discriminator.

[0030] In some embodiments, the generator employs a deep neural network based on a spatiotemporal graph convolutional network. Its input includes a random noise vector z conforming to a Gaussian distribution and meteorological data c as conditions. The network first upscales the noise vector z to a dimension matching the meteorological conditions c through a fully connected layer. Subsequently, the processed noise and meteorological conditions are concatenated along the feature dimension. The concatenated features are fed into a sequence containing multiple one-dimensional deconvolutional layers, progressively upsampled, ultimately generating a synthetic load sequence G(z|c) with the same dimension as the actual load data. Batch normalization layers are embedded between the one-dimensional deconvolutional layers and ReLU activation functions are used. The final output layer uses the Tanh activation function to constrain the output values ​​to the [-1, 1] interval, and during training, it is further mapped back to the original load dimensions through inverse normalization.

[0031] For example, in this embodiment of the invention, first generated data can be obtained based on the random noise vector in the current batch of samples during the current iteration and the generator in the conditional generative adversarial network; based on the real samples in the current batch of samples and the first generated data, the discriminator of the conditional generative adversarial network is input to perform real-fake discrimination to obtain the loss function value of the discriminator; keeping the parameters of the generator unchanged, the parameters of the discriminator are updated based on the loss function value of the discriminator to obtain the updated discriminator.

[0032] For example, embodiments of the present invention can sample a batch of conditional meteorological data c and corresponding real load data x from a real dataset. A batch of random noise z is sampled from a prior distribution (such as a standard normal distribution). z and c are input into a generator G to obtain generated data G(z|c). The real data pair (c,x) and the generated data pair (c,G(z|c)) are input into a discriminator D to calculate the discriminant loss. The discriminant's loss function L_D uses the Wasserstein distance with a gradient penalty term. θ_D is updated using a gradient descent algorithm (such as the Adam optimizer) by calculating the gradient of L_D with respect to the discriminant parameter θ_D.

[0033] S1023. Based on the samples sampled in the current batch during the current iteration, keep the parameters of the updated discriminator unchanged, update the parameters of the generator, and obtain the updated generator.

[0034] In some embodiments, the discriminator D has the following network structure: The discriminator employs a deep neural network based on a multi-scale one-dimensional convolutional neural network. Its input is a concatenated vector of real load data x or generated data G(z|c) and meteorological conditions c. The network consists of multiple stacked one-dimensional convolutional layers, with a spectral normalization layer introduced after each convolutional layer instead of a batch normalization layer, and LeakyReLU is used as the activation function (with the negative slope parameter set to 0.2). Spectral normalization forces the discriminator to satisfy Lipschitz continuity by constraining the spectral norm (i.e., the maximum singular value) of the weight matrix of each layer, thereby greatly enhancing the stability of the training process and effectively preventing mode collapse. Finally, the network outputs a single scalar value D(x|c) through the last convolutional layer, representing the probability that the input data comes from the true distribution.

[0035] For example, in an embodiment of the present invention, second generated data can be obtained based on the random noise vector and conditional input of the current batch of samples, and the generator; the loss function value of the generator can be calculated; the parameters of the discriminator can be kept unchanged, and the parameters of the generator can be updated based on the loss function value of the generator to obtain the updated generator.

[0036] For example, embodiments of the present invention can sample a new batch of random noise z from a prior distribution. z and the corresponding condition c are input into a generator G to obtain generated data G(z|c). The generated data pair (c, G(z|c)) is input into a discriminator D to calculate the generator loss. The generator's loss function L_G aims to maximize the discriminator's discriminative output on the generated data. θ_G is updated using a gradient ascent algorithm (such as the Adam optimizer) by calculating the gradient of L_G with respect to the generator parameter θ_G.

[0037] S1024. Determine whether the iteration stopping condition is met. If it is met, stop the iteration and output the conditional generative adversarial network that has completed training.

[0038] In some embodiments, the iteration stopping condition is that the loss function value of the discriminator and the loss function value of the generator are less than a loss threshold, and the discrimination accuracy of the discriminator on the sampled samples is less than an accuracy threshold.

[0039] For example, the iteration stopping conditions are met as follows: the discriminator loss L_D and the generator loss L_G both tend to stabilize, and the change of their moving average over multiple consecutive training cycles is less than a preset threshold (e.g., 1e-5); the discriminator's accuracy in distinguishing input samples (including real samples and generated samples) decreases and stabilizes at around 50% (e.g., within the range of 45% to 55%), indicating that it can no longer effectively distinguish between real and fake data; and the preset maximum number of training cycles is reached.

[0040] S1025. If the condition is not met, replace the sampled sample and repeat the process of updating the parameters of the discriminator and the generator until the iteration stop condition is met.

[0041] S1026. Based on extreme weather data and a trained conditional generative adversarial network, generate synthetic load data.

[0042] For example, after training is complete, the parameters of generator G are saved. All the historical extreme weather data c are used as conditions and input into the trained generator G, and different random noise z is fed in to generate synthetic load data that is highly consistent with the real distribution in batches.

[0043] S1027. Based on the synthetic load data and the actual load data, an expanded sample set is obtained by merging them.

[0044] For example, embodiments of the present invention can merge the generated synthetic load data with the original real load data to form the expanded sample set. This expanded sample set significantly increases the number and diversity of samples under extreme weather conditions, providing a more sufficient learning foundation for subsequent prediction models.

[0045] S103. Based on the expanded sample set, meteorological data coupling and load component decomposition are used to refine the features of each expanded sample in the expanded sample set to determine the refined sample set.

[0046] As one possible implementation, step S103 can be specifically implemented as steps S1031-S1034.

[0047] S1031. Based on the expanded sample set, decompose the actual load data in each sample into load components to determine multiple types of load data.

[0048] In some embodiments, the types of load data include trend loads that characterize basic electricity consumption habits, weather-sensitive loads driven by extreme weather, and random disturbance loads caused by random events.

[0049] For example, embodiments of the present invention can use the variational mode decomposition algorithm (VMD) to adaptively decompose the actual load sequence x(t) in each sample, separating it into three eigenmode function components with clear physical meaning:

[0050] For example, trend load components The low-frequency component extracted by VMD characterizes the slowly changing, rigid electricity demand determined by the laws of industrial production and basic social life. This component varies smoothly on a diurnal scale and is almost unaffected by instantaneous weather changes.

[0051] For example, meteorological sensitive load components The mid-to-high frequency components extracted by VMD and coupled with the frequency of meteorological element changes mainly characterize the load fluctuations caused by the start-up and shutdown of temperature control equipment such as air conditioning, heating, and ventilation. These are the core components that cause the load curve to show abrupt peak-to-valley changes under extreme weather conditions.

[0052] For example, random disturbance load components The remaining random high-frequency components after decomposition (t) characterize unpredictable load fluctuations caused by random events (such as temporary activities, grid failures, measurement noise, etc.).

[0053] S1032. Based on the expanded sample set, perform meteorological data coupling on the extreme weather data in each sample to determine the meteorological feature vector.

[0054] In some embodiments, the present invention may employ a deep neural network model based on a multi-head self-attention mechanism to perform feature fusion and coupling analysis on multi-dimensional meteorological time-series data in each sample, thereby generating a comprehensive meteorological feature vector.

[0055] For example, step S1032 can be specifically implemented as steps A1-A4.

[0056] A1. Based on the extreme weather data in each expanded sample, multidimensional raw meteorological data for each expanded sample are extracted.

[0057] In some embodiments, the multidimensional raw meteorological data includes time series data of temperature, humidity, wind speed, and solar radiation intensity.

[0058] A2. Based on the multidimensional raw meteorological data of each expanded sample and the meteorological sensitive load corresponding to the multidimensional raw meteorological data, Granger causality test is performed to screen out target meteorological elements that have a significant statistical causal relationship with the changes in meteorological sensitive load.

[0059] In some embodiments, Granger causality tests are used for quantitative screening to identify core meteorological factors that have significant predictive power for meteorologically sensitive loads.

[0060] Data preparation: For each sample, the standardized meteorological element sequences (such as the temperature sequence T) are combined with the decomposed meteorological sensitive load component sequences. Alignment.

[0061] Execution of verification: For each pair (meteorological element sequence, Granger causality tests are performed on time series data. The core of this approach is to construct two vector autoregressive (VAR) models:

[0062] Constrained model: ;

[0063] Full model: ;

[0064] Here, p is the lag order, representing the number of periods in the model that include historical data. For example, p=4 means using data from the past 4 hours to predict the value at the current moment. This value is usually determined by criteria such as AIC (Akaike Information Criterion) or BIC (Bayesian Information Criterion). The autoregressive coefficient represents the meteorological sensitive load value in the i-th period of the past. For the current value L weather The degree and direction of the influence of (t) (positive or negative correlation). These coefficients are the core parameters that the model needs to estimate; The value represents the historical observation of the meteorological sensitive load in the i-th period before time t. The random error term (in the constrained model) represents the random fluctuations or noise that the constrained model cannot explain at time t. It includes the effects of all factors not included in the model.

[0065] X(t j) represents the sequence of meteorological elements to be tested. It represents the observed value of a meteorological element (such as temperature, humidity, etc.) in the j-th period before time t. We want to test whether this variable is the cause of L. weather Granger's reasons for change. Let be the regression coefficient, representing the meteorological element in the j-th period of the past. For the current meteorological sensitive load L weather The degree and direction of influence of (t). This is a key parameter for testing whether meteorological elements have predictive capabilities. This is the random error term (full model), representing the random fluctuations or noise that the full model cannot explain at time t. Since the full model includes more explanatory variables (X), its error should theoretically be smaller than that of the restricted model. .

[0066] Hypothesis testing: Calculate the F-statistic and test whether the coefficients β1, β2, ..., βp are jointly significant and not zero.

[0067] ;

[0068] in, It is the sum of squared residuals of the restricted model, which is the sum of squared differences between the predicted and actual observed values ​​of the model that does not include the variable to be tested (such as meteorological element X). The larger the value, the worse the constrained model's ability to interpret the data, and the more unexplained the variation.

[0069] The residual sum of squares (RSSU) is the sum of squares of the differences between the predicted and observed values ​​of the model (i.e., unconstrained / full model) for the variable under test (e.g., meteorological element X). A smaller RSSU indicates a better interpretability of the full model. In the Granger causality test, we expect... < T is the sample size. This is the lag order, which is also the number of constraints. This measures the loss of explanatory power due to the imposition of constraints (i.e., excluding variable X). Dividing by p is to obtain the average loss per constraint. In the denominator: These represent the degrees of freedom of the unconstrained model. The 2p is because the unconstrained model contains p L's. weather The lagged terms are denoted by X and p lagged terms of X. The value of p here is the same as the lag order in the VAR model, and is usually determined according to information criteria such as AIC or BIC.

[0070] Significance judgment: When the p-value obtained from the test is less than the significance level (usually 0.05), the null hypothesis is rejected, and the meteorological element is considered to be a Granger cause of the meteorological sensitive load, and it is selected as the target meteorological element.

[0071] A3. Construct a composite meteorological index based on the target meteorological elements.

[0072] In some embodiments, the composite meteorological index includes a temperature and humidity synergy index for characterizing perceived comfort, and a continuous high-temperature cumulative effect coefficient for characterizing the effects of sustained thermal stress.

[0073] For example, the temperature and humidity synergy index: used to quantify the muggy feeling under hot and humid weather, its calculation formula is as follows: ;

[0074] Where T is temperature in Celsius (°C) and RH is relative humidity (expressed as a decimal, e.g., 0.5 for 50% humidity). The higher this value, the greater the thermal stress felt by the human body, and the stronger the demand for air conditioning cooling. THI is the temperature and humidity synergy index.

[0075] For example, the cumulative effect coefficient of continuous high temperature: used to characterize the superimposed impact of sustained high temperature weather on electricity load, and its calculation method is as follows:

[0076] ;

[0077] Where k represents the number of consecutive days of high temperatures. The highest temperature on day i. ω is the high-temperature trigger threshold (e.g., 35°C), and ω is the attenuation factor (usually 0.8 < ω < 1), indicating that the more recent the high-temperature weather, the greater the impact on the current load. CHC is the cumulative effect coefficient of continuous high temperatures.

[0078] A4. Based on the target meteorological elements and composite meteorological indices of each expanded sample, generate a meteorological feature vector for each expanded sample.

[0079] S1033. Based on the load data of multiple types in each sample and the meteorological feature vector, perform coupled reconstruction to generate refined samples.

[0080] For example, step S1033 can be specifically implemented as steps B1-B5.

[0081] B1. For each expanded sample, feature-level fusion is performed based on meteorological sensitive loads and meteorological feature vectors to generate joint meteorological load features.

[0082] For example, embodiments of the present invention can employ a weighted fusion method based on an attention mechanism to deeply interact with meteorological load components and meteorological feature vectors to generate a joint meteorological-load feature vector. First, a fully connected layer projects the meteorological load sequence and meteorological feature vectors onto the same feature space dimension. Then, the dynamic attention weights of the meteorological features on each time step of the load sequence are calculated to capture the differences in meteorological impact at different times. Finally, the projected meteorological load features are weighted and summed using the attention weights, and then concatenated with the projected meteorological features to form the final joint meteorological-load feature vector.

[0083] B2. Extract the load time series data and meteorological data for the specified time period after the time of each expanded sample.

[0084] For example, to construct a predictive sample, it is necessary to extract real data within a future time window after the current sample time. The actual load value at a specific time step after each expanded sample time point is extracted as the prediction target output vector for the refined sample. Simultaneously, raw meteorological data within the same time period is extracted; this data is used to calculate the meteorological feature vector for future periods.

[0085] B3. Based on the trend load, meteorological sensitive load, random disturbance load and meteorological load joint characteristics of each expanded sample, as well as the meteorological data of the time period after the time of each expanded sample, perform time-series alignment and feature splicing to generate the input vector of refined samples.

[0086] For example, embodiments of the present invention can strictly align multiple features along the time dimension and concatenate them into a high-dimensional input feature vector. This ensures that the time windows of all features completely correspond to the current sample time. The features most effective for predicting future load are selected for concatenation, including: trend-based load components, representing a stable electricity base; weather-sensitive load components, representing load fluctuations driven by current weather; random disturbance load components, representing random noise; a joint weather-load feature vector; a future weather feature vector, providing prior information about future weather changes; and timestamp features. All features are then concatenated into a comprehensive input feature vector.

[0087] B4. Based on the load time series data of a set period after the time of each expanded sample, generate the output vector of the refined sample.

[0088] B5. Generate refined samples based on the input and output vectors of the refined samples.

[0089] For example, embodiments of the present invention can combine the concatenated input feature vector with future real load data to form a complete supervised learning sample pair. A refined sample contains a multi-dimensional fused feature input vector and a real load value output vector at a specific future time step, which serves as the target for model training.

[0090] S1034. Generate a fine sample set based on multiple fine samples.

[0091] For example, for each sample in the expanded sample set, the steps of feature fusion, data extraction, feature concatenation, and sample construction are repeated to generate a large number of refined sample pairs, ultimately forming a refined sample set for training the final prediction model. Through the above refined reconstruction process, the originally simple samples are transformed into high-quality samples with highly condensed information, explicit feature decoupling, and direct orientation towards multi-step prediction tasks, laying a solid data foundation for the subsequent prediction model to achieve high-precision predictions under complex weather conditions.

[0092] S104. Based on a refined sample set, a pre-defined prediction model architecture, and physical consistency constraints, machine learning is performed to generate a load prediction model under complex weather conditions.

[0093] As one possible implementation, step S104 can be specifically implemented as steps S1041-S1046.

[0094] S1041. For each iteration in the machine learning process, based on the input vector of the refined sample in the current iteration and the attention module of the prediction model architecture, perform attention calculation to determine the weights of various types of load data in the input vector.

[0095] For example, the attention module in the prediction model architecture receives input vectors from refined samples and analyzes the importance of various load data in the input vectors through a multi-layer perceptron mechanism. This module first takes the feature representations of trend-based loads, meteorologically sensitive loads, and random disturbance loads as input, and generates attention scores by calculating the correlation between query vectors, key vectors, and value vectors. These scores are then normalized to form attention weights, which are used to identify the importance of different load components to the prediction task. Higher weights indicate that the corresponding load component features have a stronger influence on the prediction results under current meteorological conditions.

[0096] S1042. Based on the input vector and the weights of various load data in the input vector, input the neural network module of the prediction model architecture to obtain the prediction result of the current iteration process.

[0097] For example, the neural network module receives a weighted input feature vector and performs forward propagation computation. This module contains multiple fully connected layers and activation functions, enabling it to learn the complex nonlinear relationship between load changes and meteorological characteristics. The weighted feature vector is first processed by batch normalization, and then a high-level feature representation is extracted through a nonlinear transformation layer. These feature representations are used for information transfer and aggregation at different time steps through a recurrent neural network structure or a temporal convolutional layer, ultimately outputting a load forecast sequence for a specific future period. The prediction result at each time step incorporates the combined effects of historical information, current conditions, and future meteorological conditions.

[0098] S1043. Based on the prediction results and the output vector of the refined samples in the current iteration, calculate the joint loss of the current iteration.

[0099] In some embodiments, the joint loss includes prediction error loss and physical consistency constraint loss, wherein the physical consistency constraint loss includes load non-negative penalty term, load change smoothness penalty term, and meteorological sensitivity penalty term.

[0100] For example, embodiments of the present invention can calculate two types of losses: prediction error loss, which measures the difference between the predicted result and the actual load value, using mean absolute error or root mean square error as the metric; and physical consistency constraint loss, which ensures that the prediction result conforms to the basic physical laws of power system operation, including three penalty terms: a load non-negativity penalty term, which ensures that all predicted load values ​​are greater than or equal to zero; a load change smoothness penalty term, which limits the drastic degree of load change at adjacent time points; and a meteorological sensitivity penalty term, which requires the predicted load to respond to changes in meteorological factors in a manner consistent with physical common sense. The final joint loss is a weighted sum of these individual losses.

[0101] S1044. Based on the joint loss of the current iteration process, update the model parameters of the attention module and the neural network module through the gradient backpropagation algorithm to obtain the updated attention module and neural network module.

[0102] For example, embodiments of the present invention can calculate the gradient of the joint loss with respect to all model parameters using the gradient backpropagation algorithm. An adaptive moment estimation optimization algorithm is employed to update the parameters of the attention module and the neural network module. This algorithm combines first-order moment estimation and second-order moment estimation to adjust the learning rate of each parameter. The learning rate is automatically adjusted according to the magnitude of the gradient, performing targeted updates in the parameter space to ensure the stability and convergence speed of the training process. After each iteration, the updated state of the model parameters is saved for use in the next forward propagation calculation.

[0103] S1045. Determine whether the iteration stopping condition is met. If it is met, generate a load prediction model under complex weather conditions based on the attention module and neural network module after the iteration is completed.

[0104] The iteration stopping conditions include: the change magnitude of the joint loss over N consecutive iterations is less than the joint threshold and the prediction accuracy between the prediction result and the output vector is greater than the accuracy threshold, or the number of iterations is greater than the maximum number of iterations.

[0105] S1046. If the condition is not met, replace the refined sample and repeat the iteration process.

[0106] For example, embodiments of the present invention can monitor two convergence metrics: the change in the joint loss value over multiple consecutive training epochs is less than a preset threshold, while the prediction accuracy reaches the required level; or the number of training iterations reaches a preset maximum value. When either condition is met, the training process is immediately terminated, and the final model parameters and structure configuration are saved. Otherwise, new batch data is sampled from the refined sample set, and the forward propagation, loss calculation, and parameter update processes are repeated until the termination condition is met.

[0107] For example, after training, this embodiment of the invention combines the trained attention module and neural network module into a complete complex weather load prediction model. This model has the ability to process multi-dimensional input features and output probabilistic prediction results. The final model is saved in a structured file format, containing all necessary information such as model architecture definition, training parameter configuration, weight matrix, and bias vector, and can be directly deployed to a production environment for real-time load prediction.

[0108] S105. Based on the load forecasting model, perform load forecasting under complex weather conditions.

[0109] As one possible implementation, step S105 can be specifically implemented as steps S1051-S1055.

[0110] S1051. Obtain the actual load data and extreme weather data before the current time, various predicted extreme weather data after the current time, and the probability of occurrence of various predicted extreme weather data.

[0111] For example, embodiments of the present invention can collect or obtain required data in real time from multiple data sources. First, actual load data for a continuous time window (e.g., the past 72 hours) prior to the current moment is obtained from a power monitoring system (such as SCADA / EMS), including key indicators such as active power and reactive power. Simultaneously, extreme weather data measured during the same period, such as wind speed, precipitation, and temperature during typhoons, rainstorms, and extreme high temperatures, are obtained from meteorological monitoring stations or meteorological department interfaces.

[0112] This invention obtains various predicted extreme weather data for the forecast period (e.g., the next 24-72 hours) from numerical weather prediction (NWP) systems or professional meteorological services, including predicted values ​​and time-series changes of temperature, humidity, wind speed, precipitation, etc. Furthermore, it also obtains the occurrence probabilities or confidence levels corresponding to these predicted extreme weather data. These probabilities are typically generated by meteorological models using ensemble forecasting or probabilistic forecasting techniques to quantify the uncertainty of weather forecasts.

[0113] S1052. Based on the actual load data and extreme weather data before the current time, as well as various predicted meteorological data after the current time, perform meteorological data coupling and load component decomposition to determine multiple refined input vectors.

[0114] For example, embodiments of the present invention can perform deep fusion and feature extraction on the acquired data. Meteorological data coupling refers to aligning and splicing historical measured meteorological data with future predicted meteorological data in the time dimension to form a continuous meteorological feature sequence. This process may include normalizing meteorological elements and using techniques such as attention mechanisms to evaluate the weights of different meteorological elements on the load.

[0115] Subsequently, load component decomposition is performed: using a pre-trained decomposition model (such as Variational Mode Decomposition (VMD) or Empirical Mode Decomposition (EMD), the historical actual load sequence is decomposed into multiple relatively independent components, such as trend load, meteorological sensitive load, and random disturbance load. Finally, the coupled meteorological feature sequence is concatenated with each decomposed load component, as well as features extracted from timestamps (such as hour, day of the week, and holiday markers), to determine multiple refined input vectors, each corresponding to a future prediction time step.

[0116] S1053. Based on multiple refined input vectors and the load forecasting model, load forecasting is performed, and the forecasting result corresponding to each refined input vector is determined.

[0117] In some embodiments, the forecast results include a load forecast sequence after the current time.

[0118] For example, in embodiments of the present invention, multiple refined input vectors can be sequentially input into a pre-trained load forecasting model. The model performs forward computation on each input vector, outputting the load forecast value for the corresponding future time step. The forecast results for all time steps are combined in chronological order to form a load forecast sequence for the forecast period following the current moment. This process is typically parallel to improve forecasting efficiency, generating a deterministic load forecast curve under given meteorological forecast conditions.

[0119] S1054. Based on various predicted extreme weather data after the current moment, and the probability of occurrence of various predicted extreme weather data, dynamic scenarios are generated to construct multiple load scenarios.

[0120] In some embodiments, the load scenarios include a baseline scenario, a high-risk scenario, and a low-risk scenario.

[0121] For example, considering the uncertainty of future extreme weather, multiple possible scenarios need to be constructed. The dynamic scenario generation module will randomly generate a large number of future weather scenarios based on the probability of occurrence of various types of predicted extreme weather data, using specific algorithms (such as Monte Carlo simulation, scenario tree generation, etc.). These scenarios cover different possible weather forecast scenarios. Subsequently, based on the statistical characteristics of these weather scenarios (such as probability density functions), they are clustered and reduced to ultimately form several representative load scenarios.

[0122] The load curves are categorized into three scenarios: Baseline Scenario: Load curves generated under the most probable weather forecast. High-Risk Scenario: Load curves generated under the most severe weather conditions (e.g., extreme heat lasting longer than expected) that pose the greatest threat to the power system. Low-Risk Scenario: Load curves generated under weather conditions better than the baseline forecast. Each scenario includes its corresponding weather sequence and its probability of occurrence.

[0123] S1055. Based on the prediction results corresponding to each refined input vector and multiple load scenarios, load curves are plotted to determine the load prediction interface under complex weather conditions.

[0124] In some embodiments, the load forecasting interface includes multiple load forecasting curves, and the risk level and probability of occurrence for each load forecasting curve.

[0125] For example, embodiments of the present invention can integrate and visualize the obtained deterministic prediction results with multi-scenario prediction results. The load curve plotting module will plot multiple load prediction curves for the baseline scenario, high-risk scenario, and low-risk scenario, and display them in the same coordinate system using different colors or line types for comparison. On the load prediction interface under generated complex weather conditions, in addition to displaying these curves, it will also clearly indicate: the scenario corresponding to each prediction curve (e.g., baseline, high-risk, low-risk); the risk level of each curve (e.g., a high-risk curve will be marked "extremely high risk," prompting dispatchers to pay close attention); and the probability of occurrence of each curve (displayed as a percentage, providing users with an assessment of the likelihood of the scenario occurring). This interface is typically a graphical human-computer interaction (HMI) interface, which may be integrated into the power dispatching intelligent system to provide operators with intuitive decision support.

[0126] Thus, this invention not only provides a single most probable load forecast, but also provides multi-scenario probabilistic forecast results that take into account weather uncertainties, greatly enhancing the risk resistance and scientific decision-making of power grid planning and operation under complex weather conditions.

[0127] This invention provides a load forecasting method for complex weather conditions based on generative adversarial learning. By utilizing conditional generative adversarial networks to generate synthetic load data under extreme weather conditions, it overcomes the technical bottleneck of scarce historical samples for extreme weather and significantly enhances the model's generalization ability to rare meteorological scenarios. Subsequently, through meteorological data coupling and load component decomposition, the training features of the model are refined, improving feature specificity and interpretability, ensuring that the load forecasting model accurately identifies refined features and improving forecast accuracy. Furthermore, physical consistency constraints are embedded in the load forecasting model to ensure that the output results conform to the operating laws of the power system, avoiding the physical irrationality of traditional pure data-driven models. This invention, by introducing generative adversarial learning, multimodal feature fusion, and physical constraint mechanisms, effectively solves the problem of low accuracy in power load forecasting under complex weather conditions and improves the accuracy of power load forecasting under such conditions.

[0128] Optionally, the load forecasting method based on generative adversarial learning under complex weather conditions provided in this embodiment of the invention further includes steps S201-S204 after step S105.

[0129] S201. Based on the load forecasting interface under complex weather conditions, determine the risk level and sub-period of load gap during the forecast period.

[0130] For example, embodiments of the present invention can determine the risk of load gaps based on the multi-scenario prediction results displayed on the load prediction interface in the following ways: First, the load prediction curves of each scenario are compared with the system power supply capacity threshold in real time to identify the time periods in which power supply gaps may occur; then, the risk level is calculated according to the gap size and probability of occurrence, which is usually divided into three levels: high risk, medium risk and low risk; at the same time, the start time, duration and expected gap power value of each risk period are recorded to form a complete list of risk periods.

[0131] S202. Based on the risk level and sub-period of the load gap in the forecast period, risk assessment is performed to determine the trigger time and duration of the demand-side response.

[0132] For example, embodiments of the present invention can determine key parameters of demand-side response based on the identified load gap risk characteristics: prioritize response for high-risk periods, with a certain time before the start of the risk period as the response trigger time; determine the duration of demand-side response based on the duration of the risk period, typically setting the response duration to cover the entire risk period; and set reasonable response preparation and exit times by comprehensively considering user response capabilities and system reliability requirements.

[0133] S203. Based on the risk level of load gaps during the forecast period and the load forecast curves corresponding to the sub-periods, determine the expected load reduction amount of the demand-side response.

[0134] For example, embodiments of the present invention can calculate the expected load reduction based on the load forecast curve: analyze the load forecast value during the risk period, calculate the difference between it and the system power supply capacity as the basic reduction amount; consider the uncertainty of load forecast, increase a certain safety margin; and adjust the reduction amount appropriately according to historical demand-side response effect data to ensure the safe operation of the system while avoiding over-response.

[0135] S204. Generate a demand-side response strategy based on the trigger time, duration, and expected load reduction of the demand-side response.

[0136] For example, embodiments of the present invention can integrate all parameters to generate a complete demand-side response strategy: formulate a detailed response schedule, specifying the response start time, end time, and duration; determine the user groups participating in the response and their respective load reduction task allocations; compile a response execution instruction sequence, including response start instructions, execution process monitoring instructions, and response end instructions; and generate a response effect evaluation scheme for post-event analysis and strategy optimization.

[0137] For example, embodiments of the present invention can convert the generated demand-side response strategy into an executable format: output a strategy document containing detailed execution steps and technical requirements; generate automated execution instructions that can be directly issued to the demand-side management system; prepare contingency plans, including backup plans and remedial measures in case of response failure; and establish a strategy execution tracking mechanism to ensure that the response process is controllable and measurable.

[0138] Thus, the present invention can effectively transform load forecasting results into executable demand-side response strategies, fully leverage the role of demand-side resources in addressing load gap risks under complex weather conditions, and improve the reliability and economy of power system operation.

[0139] Optionally, the load forecasting method based on generative adversarial learning under complex weather conditions provided in this embodiment of the invention further includes steps S301-S305 after step S105.

[0140] S301. Based on multiple load forecast curves in the load forecast interface under complex weather conditions, perform probabilistic power flow calculation and risk scanning to identify critical lines and sub-periods with overload risk during the forecast period, as well as the overload risk probability of each critical line.

[0141] For example, embodiments of the present invention can perform power grid safety analysis based on multi-scenario prediction curves in the load forecasting interface, using probabilistic power flow calculation and risk scanning techniques. First, the load forecast results for each scenario are input into the power grid simulation model, and the load rate distribution of each power grid line under different load scenarios is analyzed using probabilistic power flow calculation methods. By setting line safety thresholds, the system automatically scans and identifies critical lines that may have overload risks during the forecast period, recording the overload risk probability, expected overload severity, and specific risk occurrence time for these lines. The system generates a list of risky lines, including line name, risk period start time, duration, overload probability, and severity level.

[0142] S302. Based on the critical path and the overload risk probability of each critical path, as well as the topology of the power grid where the critical path is located, perform network reconfiguration to determine the load transfer path of the critical path and the reconfigured power grid topology.

[0143] For example, embodiments of the present invention can design network reconfiguration schemes based on identified risky lines and the topology of the power grid in which they are located. The system first analyzes the topological connections of the power grid to find alternative paths for load transfer. Intelligent algorithms evaluate the transmission capacity, voltage stability, and reliability indicators of each alternative path to determine the optimal load transfer line. Simultaneously, a reconfigured power grid topology scheme is generated, including necessary switch state changes and network connection adjustments to ensure the overall stability of the power grid and the reliability of power supply during load transfer. The system outputs a complete transfer scheme including the primary transfer path, alternative transfer paths, and their capacity limitations.

[0144] S303. Based on the sub-periods with overload risk, determine the load transfer time and duration of the critical path.

[0145] For example, embodiments of the present invention can formulate precise load transfer timing plans based on the characteristics of risk periods. The system analyzes the specific time periods during which overload risks occur, determines the optimal start time for load transfer, and typically sets it within a sufficient time window before the risk occurs to ensure operational safety. The duration for which the load transfer operation needs to be maintained is determined based on the duration of the risk, while also considering the execution and recovery times of the transfer operation. The system generates a detailed operation schedule, including the transfer start time, estimated completion time, maintenance duration, and the time point for restoring the original operating mode, ensuring coordination with the power grid dispatch plan.

[0146] S304. Based on the load transfer lines, load transfer times and durations of the critical lines, and the restructured power grid topology, generate overload handling schemes for the critical lines.

[0147] For example, embodiments of the present invention can integrate all analysis results to generate a complete overload handling plan. This plan includes a specific sequence of operation instructions, an operation schedule, an expected effect assessment, and safety measures. The system will detail the load transfer path, the amount of load to be transferred, the operation time requirements, and precautions for each risky line. Simultaneously, an emergency plan will be generated, including backup plans and emergency response measures in case of transfer failure. The final plan is output in a standardized format, including an execution procedure description, an operation instruction table, and monitoring requirements, and can be directly integrated into the power grid dispatching system for execution.

[0148] Thus, the embodiments of the present invention can effectively transform load forecasting results into power grid safety control measures, realize proactive defense and precise handling of power grid overload risks under complex weather conditions, and significantly improve the reliability and safety of power grid operation.

[0149] It should be understood that the sequence number of each step in the above embodiments does not imply the order of execution. The execution order of each process should be determined by its function and internal logic, and should not constitute any limitation on the implementation process of the embodiments of the present invention.

[0150] The following are device embodiments of the present invention. For details not described in detail, please refer to the corresponding method embodiments described above.

[0151] Figure 2 A schematic diagram of a load forecasting device for complex weather conditions based on generative adversarial learning, provided by an embodiment of the present invention, is shown. The load forecasting device 400 includes a communication module 401 and a processing module 402.

[0152] The communication module 401 is used to acquire extreme weather data from historical periods, as well as the actual load data corresponding to the extreme weather data.

[0153] The processing module 402 is used to expand the sample data under extreme weather conditions in historical periods based on extreme weather data, actual load data, and a preset conditional generative adversarial network to determine an expanded sample set; based on the expanded sample set, it uses meteorological data coupling and load component decomposition to refine the features of each expanded sample in the expanded sample set to determine a refined sample set; based on the refined sample set, a preset prediction model architecture, and physical consistency constraints, it performs machine learning to generate a load prediction model under complex weather conditions; and based on the load prediction model, it performs load prediction under complex weather conditions.

[0154] Figure 3This is a schematic diagram of the structure of an electronic device provided in an embodiment of the present invention. The electronic device 500 includes: a processor 501, a memory 502, and a computer program 503 stored in the memory 502 and executable on the processor 501. When the processor 501 executes the computer program 503, it implements the steps in the above-described method embodiments. Alternatively, when the processor 501 executes the computer program 503, it implements the functions of each module / unit in the above-described device embodiments.

[0155] For example, the computer program 503 may be divided into one or more modules / units, which are stored in the memory 502 and executed by the processor 501 to complete the present invention. The one or more modules / units may be a series of computer program instruction segments capable of performing specific functions, which describe the execution process of the computer program 503 in the electronic device 500.

[0156] The processor 501 may be a Central Processing Unit (CPU), or other general-purpose processors, digital signal processors (DSPs), application-specific integrated circuits (ASICs), field-programmable gate arrays (FPGAs), or other programmable logic devices, discrete gate or transistor logic devices, discrete hardware components, etc. A general-purpose processor may be a microprocessor or any conventional processor.

[0157] The memory 502 can be an internal storage unit of the electronic device 500, such as a hard disk or memory of the electronic device 500. The memory 502 can also be an external storage device of the electronic device 500, such as a plug-in hard disk, smart media card (SMC), secure digital card (SD), flash card, etc., equipped on the electronic device 500. Furthermore, the memory 502 can include both internal and external storage units of the electronic device 500. The memory 502 is used to store the computer program and other programs and data required by the terminal. The memory 502 can also be used to temporarily store data that has been output or will be output.

[0158] The above-described embodiments are only used to illustrate the technical solutions of the present invention, and are not intended to limit it. Although the present invention has been described in detail with reference to the foregoing embodiments, those skilled in the art should understand that modifications can still be made to the technical solutions described in the foregoing embodiments, or equivalent substitutions can be made to some of the technical features. Such modifications or substitutions do not cause the essence of the corresponding technical solutions to deviate from the spirit and scope of the technical solutions of the embodiments of the present invention, and should all be included within the protection scope of the present invention.

Claims

1. A load forecasting method under complex weather conditions based on generative adversarial learning, characterized in that, include: Acquire historical extreme weather data and the corresponding actual load data for extreme weather events; Based on the extreme weather data, the actual load data, and the preset conditional generative adversarial network, the sample data under extreme weather conditions in historical periods are expanded to determine the expanded sample set. Based on the expanded sample set, meteorological data coupling and load component decomposition are used to refine the features of each expanded sample in the expanded sample set to determine a refined sample set. This includes: decomposing the actual load data in each expanded sample into load components based on the expanded sample set, identifying multiple types of load data, including trend loads characterizing basic electricity consumption habits, meteorologically sensitive loads driven by extreme weather, and random disturbance loads caused by random events; coupling extreme weather data in each expanded sample with meteorological data based on the expanded sample set to determine meteorological feature vectors; performing coupling reconstruction based on the multiple types of load data in each expanded sample and the meteorological feature vectors to generate refined samples; and generating the refined sample set based on multiple refined samples. Based on the refined sample set, the preset prediction model architecture, and the physical consistency constraints, machine learning is performed to generate a load prediction model under complex weather conditions. Based on the load forecasting model, load forecasting is performed under complex weather conditions; The process of coupling extreme weather data in each expanded sample with meteorological data based on the expanded sample set to determine meteorological feature vectors includes: extracting multidimensional raw meteorological data for each expanded sample based on the extreme weather data in each expanded sample, wherein the multidimensional raw meteorological data includes time series data of temperature, humidity, wind speed, and solar radiation intensity; performing Granger causality tests on the multidimensional raw meteorological data of each expanded sample and the meteorological sensitive load corresponding to the multidimensional raw meteorological data to screen out target meteorological elements that have a significant statistical causal relationship with changes in meteorological sensitive loads; constructing a composite meteorological index based on the target meteorological elements, wherein the composite meteorological index includes a temperature and humidity synergy index for characterizing perceived comfort and a continuous high temperature cumulative effect coefficient for characterizing the impact of sustained thermal stress; and generating a meteorological feature vector for each expanded sample based on the target meteorological elements and the composite meteorological index.

2. The load forecasting method for complex weather conditions based on generative adversarial learning according to claim 1, characterized in that, The method involves expanding historical extreme weather sample data based on the extreme weather data, the actual load data, and a preset conditional generative adversarial network to determine the expanded sample set, including: Using the extreme weather data as input, the actual load data corresponding to the extreme weather data as the real sample, and setting a random noise vector that conforms to a Gaussian distribution, multiple batches of sampling samples are constructed. Based on the samples sampled in the current batch during the current iteration, while keeping the parameters of the generator in the conditional generative adversarial network unchanged, the parameters of the discriminator in the conditional generative adversarial network are updated to obtain the updated discriminator; Based on the samples sampled in the current batch during the current iteration, keep the parameters of the discriminator unchanged and update the parameters of the generator to obtain the updated generator; Determine whether the iteration stopping condition is met. If it is met, stop the iteration and output the conditional generative adversarial network that has completed training. The iteration stopping condition is that the loss function value of the discriminator and the loss function value of the generator are less than the loss threshold, and the discrimination accuracy of the discriminator on the sampled samples is less than the accuracy threshold. If the condition is not met, the sampled sample is replaced, and the process of updating the parameters of the discriminator and the generator is repeated until the iteration stopping condition is met. Based on the extreme weather data and the trained conditional generative adversarial network, synthetic load data is generated. The expanded sample set is obtained by merging the synthetic load data and the actual load data.

3. The load forecasting method for complex weather conditions based on generative adversarial learning according to claim 1, characterized in that, The process of coupling and reconstructing multiple types of load data in each expanded sample, along with the meteorological feature vector, to generate refined samples includes: For each expanded sample, feature-level fusion is performed based on the meteorological sensitive load and the meteorological feature vector to generate joint meteorological load features; Extract the load time series data and meteorological data for a set period following the time of each expanded sample; Based on the trend load, meteorological sensitive load, random disturbance load and joint features of the meteorological load of each expanded sample, as well as the meteorological data of a set period after the time of each expanded sample, time-series alignment and feature splicing are performed to generate the input vector of the refined sample. Based on the load time series data of a set period after the time of each expanded sample, an output vector of refined samples is generated; Refined samples are generated based on the input and output vectors of the refined samples.

4. The load forecasting method for complex weather conditions based on generative adversarial learning according to claim 1, characterized in that, The process of generating a load forecasting model under complex weather conditions based on the refined sample set, the preset prediction model architecture, and physical consistency constraints includes: For each iteration in the machine learning process, based on the input vector of the refined sample in the current iteration and the attention module of the prediction model architecture, attention calculation is performed to determine the weights of various types of load data in the input vector; Based on the input vector and the weights of various load data in the input vector, the neural network module of the prediction model architecture is input to obtain the prediction result of the current iteration process; Based on the prediction results and the output vector of the refined sample in the current iteration, the joint loss of the current iteration is calculated. The joint loss includes prediction error loss and physical consistency constraint loss. The physical consistency constraint loss includes load non-negative penalty term, load change smoothness penalty term, and meteorological sensitivity penalty term. Based on the joint loss of the current iteration process, the model parameters of the attention module and the neural network module are updated through the gradient backpropagation algorithm to obtain the updated attention module and neural network module. Determine whether the iteration stopping condition is met. If it is met, then generate a load prediction model under complex weather conditions based on the attention module and neural network module after the iteration is completed. The iteration stopping condition includes: the change amplitude of the joint loss over N consecutive iteration cycles is less than the joint threshold and the prediction accuracy between the prediction result and the output vector is greater than the accuracy threshold, or the number of iterations is greater than the maximum number of iterations. If the conditions are not met, a more refined sample is used, and the iterative process is repeated.

5. The load forecasting method based on generative adversarial learning under complex weather conditions according to claim 1, characterized in that, The load forecasting based on the load forecasting model under complex weather conditions includes: Acquire actual load data and extreme weather data up to the current moment, various predicted extreme weather data after the current moment, and the probability of occurrence of various predicted extreme weather data; Based on the actual load data and extreme weather data before the current moment, and various predicted meteorological data after the current moment, meteorological data coupling and load component decomposition are performed to determine multiple refined input vectors; Based on the multiple refined input vectors and the load forecasting model, load forecasting is performed to determine the forecast result corresponding to each refined input vector. The forecast result includes the load forecast sequence for the forecast period after the current time. Based on various predicted extreme weather data after the current moment, and the probability of occurrence of various predicted extreme weather data, dynamic scenarios are generated to construct multiple load scenarios, including a baseline scenario, a high-risk scenario, and a low-risk scenario. Based on the prediction results corresponding to each refined input vector and the multiple load scenarios, load curves are plotted to determine the load prediction interface under complex weather conditions; the load prediction interface includes multiple load prediction curves, as well as the risk level and probability of occurrence of each load prediction curve.

6. The method for load forecasting under complex weather conditions based on generative adversarial learning according to any one of claims 1 to 5, characterized in that, The method further includes: Based on the load forecasting interface under complex weather conditions, determine the risk level and sub-periods of load gaps during the forecast period; Based on the risk level and sub-periods of load gaps during the forecast period, risk assessment is conducted to determine the trigger time and duration of demand-side response. Based on the risk level of load gaps during the forecast period and the load forecast curves corresponding to the sub-periods, the expected load reduction amount of the demand-side response is determined. Based on the trigger time, duration, and expected load reduction of the demand-side response, a demand-side response strategy is generated.

7. The load forecasting method for complex weather conditions based on generative adversarial learning according to claim 1, characterized in that, The method further includes: Based on multiple load forecast curves in the load forecast interface under complex weather conditions, probabilistic power flow calculation and risk scanning are performed to identify critical lines and sub-periods with overload risk during the forecast period, as well as the overload risk probability of each critical line. Based on the critical lines and the overload risk probability of each critical line, as well as the topology of the power grid where the critical lines are located, network reconfiguration is performed to determine the load transfer lines of the critical lines and the reconfigured power grid topology. Based on the sub-periods where there is an overload risk, determine the load transfer time and duration of the critical path; Based on the load transfer lines, load transfer times and durations of the critical lines, as well as the restructured power grid topology, overload handling schemes for the critical lines are generated.

8. A load forecasting system for complex weather conditions based on generative adversarial learning, characterized in that, The load forecasting system includes an electronic device, which includes a memory and a processor. The memory stores a computer program, and the processor is used to call and run the computer program stored in the memory to perform the method as described in any one of claims 1 to 7.

Images