Method for predicting resilience of clean energy power system based on multiple extreme weather scenarios

Abstract

This invention provides a method for predicting the resilience of clean energy power systems based on multiple extreme weather scenarios. It organically integrates extreme weather evolution trend data with source-load scenario generation. The method inputs identified extreme weather types into a multi-task learning framework, outputting evolution trend data. Then, it uses time-series meteorological element data reflecting the current weather conditions and evolution trend data reflecting future changes as input conditions to a joint generation model for source-load uncertainty scenarios, generating a joint scenario set. This joint scenario set not only reproduces historical source-load output characteristics but also characterizes the impact of extreme weather evolution on future source-load output under the background of climate change. Finally, it obtains system power supply resilience indicators and generates prediction results through time-series production simulation, effectively improving the foresight and accuracy of power supply resilience prediction under extreme weather scenarios.

Description

Technical Field

[0001] This invention relates to the field of power system resilience assessment technology, and in particular to a method for predicting the resilience of clean energy power systems based on multiple extreme weather scenarios. Background Technology

[0002] Against the backdrop of "dual carbon" goals and the construction of a new power system, my country's installed capacity of clean energy continues to increase, and the power system's source-load sides are highly coupled with meteorological conditions. Affected by global climate change, extreme weather events such as cold waves, high temperatures, heavy precipitation, and icing / snowfall are frequent, easily causing significant fluctuations in renewable energy output and drastic changes in electricity load, leading to power supply and demand imbalances and posing a serious threat to the resilience and security of the power grid. Provincial power grids in Qinghai and other areas with a high proportion of clean energy face even more pronounced power supply risks under extreme weather conditions due to their high renewable energy share and complex regulation resource characteristics.

[0003] Currently, the industry has conducted research on extreme weather identification, new energy power prediction, source-load scenario generation, and power system resilience assessment. Extreme weather identification mostly follows general meteorological standards, source-load simulation mostly adopts traditional probabilistic models and conventional generation methods, and power supply resilience assessment is mostly oriented towards conventional operation scenarios, mainly using reliability and adequacy indicators, and combining time-series production simulation, stochastic optimization and other methods for analysis.

[0004] Existing technologies are ill-suited to the demands of high-proportion clean energy power systems: First, the definition of extreme weather does not incorporate power supply and demand balance characteristics, resulting in insufficient accuracy in meteorological evolution predictions. Second, the source-load uncertainty coupling mechanism under extreme weather conditions is not fully revealed, limiting the accuracy of joint scenario generation. Third, power supply resilience assessments do not consider the coupling effects of multi-node uncertainty, prediction errors, and operational decisions, lacking an integrated resilience prediction and assessment method for multiple extreme weather scenarios. Fourth, the flexible resource coordination and adjustment mechanism is imperfect, making it difficult to support rapid improvement in power supply resilience under extreme weather conditions. Accurate prediction, quantitative assessment, and coordinated improvement of power supply resilience in high-proportion clean energy power systems facing multiple extreme weather scenarios have become crucial issues urgently needing to be addressed by the industry. Summary of the Invention

[0005] This invention provides a method for predicting the resilience of clean energy power systems based on multiple extreme weather scenarios, in order to solve the problem of inaccurate prediction of the power supply resilience of high-proportion clean energy power systems in multiple extreme weather scenarios in the prior art.

[0006] On the one hand, this invention provides a method for predicting the resilience of clean energy power systems based on multiple extreme weather scenarios, including:

[0007] Meteorological data and power system operation data of the target area are acquired to obtain a fused dataset, and time-series data of meteorological elements affecting the output and load of new energy sources are extracted from the fused dataset;

[0008] Feature extraction and modal analysis are performed on the fused dataset to obtain the types of extreme weather processes that affect the power supply balance of the power system;

[0009] Input the extreme weather process type into a multi-task learning framework, and output the evolution trend data of the extreme weather process, including the frequency of occurrence, spatial impact range, and temporal variation trend;

[0010] The time-series data of the meteorological elements and the evolution trend data are jointly input into the source-load uncertainty scenario joint generation model to jointly simulate the probability distribution of new energy output and load output under extreme weather conditions, and output a joint scenario set;

[0011] The combined scenario set is input into the time-series production simulation model, and the power balance simulation calculation is performed by integrating the calling strategies of various types of energy storage and the optimization model of adjustable load to obtain the system power supply resilience index.

[0012] Based on the system power supply resilience index, the power supply resilience of the clean energy power system in the target area under different extreme weather scenarios is analyzed, and power supply resilience prediction results are generated.

[0013] Optionally, feature extraction and modal analysis are performed on the fused dataset to obtain the types of extreme weather processes affecting the power supply balance of the power system, including:

[0014] Extract multi-layer high-dimensional field data from the fused dataset;

[0015] The geomorphic field data were subjected to EOF orthogonal function decomposition to obtain meteorological modal characteristics;

[0016] Based on meteorological principles, threshold analysis and spatiotemporal feature extraction are performed on the meteorological modal features to construct a training sample set.

[0017] The classification model is trained using the training sample set to obtain an extreme weather classification and identification model;

[0018] The meteorological modal features to be identified are input into the extreme weather classification and identification model, and the extreme weather process type is output.

[0019] Optionally, constructing a multi-task framework includes:

[0020] Empirical mode decomposition is performed on the fusion dataset corresponding to the extreme weather process types to obtain meteorological factors;

[0021] The meteorological factors are spatial feature aggregation and convolutional feature extraction are performed using a graph convolutional neural network to obtain a meteorological feature factor dataset.

[0022] A multi-task learning framework is constructed, comprising a shared layer and multiple task layers. The shared layer is used to perform high-dimensional feature integration on the meteorological feature factor dataset. The multiple task layers correspond to the frequency prediction task, the spatial impact range prediction task, and the temporal change trend prediction task, respectively.

[0023] The multi-task learning framework is jointly trained using historical extreme weather process sample data. The network parameters of the shared layer are shared by each task layer, and the multi-task loss function is weighted and summed to optimize the model parameters. After training, the multi-task framework is obtained.

[0024] Optionally, constructing a joint generation model for source-load uncertainty scenarios includes:

[0025] Acquire historical renewable energy output data, load data, and time-series data and evolution trend data of meteorological elements during extreme weather events to construct a training sample set;

[0026] Using the time-series meteorological data and the evolution trend data as conditions, and random noise as the input to the generator, a conditional generative adversarial network is constructed. The conditional generative adversarial network includes a generator and a discriminator. The generator is used to output simulated new energy power output and load combined power output data, and the discriminator is used to distinguish between the generated combined power output data and the real combined power output data based on the training sample set.

[0027] The loss function of the discriminator is constructed based on the Vascular distance, and a gradient penalty term is introduced to constrain the gradient norm of the discriminator.

[0028] A regularization term is constructed based on the joint distribution of KL divergence to constrain the parameter update magnitude of the generator. The conditional generative adversarial network is then trained in a game-like adversarial manner until convergence, resulting in a joint generative model for a source-load uncertainty scenario.

[0029] Optionally, constructing a time-series production simulation model includes:

[0030] Historical operation data of the source-load-storage network is obtained and subjected to time-series normalization and boundary extraction to obtain time-series operation boundary data;

[0031] The time-series operational boundary data is subjected to power grid topology analysis and cross-sectional constraint extraction to obtain topology and time-series constraint data;

[0032] The system acquires flexible and adjustable resource capabilities, and constructs a time-series production simulation model based on the topology and time-series constraint data and the joint scenario set.

[0033] Optionally, the ability to acquire flexible and adjustable resources includes:

[0034] Load monitoring data under extreme weather conditions were collected, and load levels and fluctuation characteristics were calculated.

[0035] The load level and fluctuation characteristics are decomposed into temperature-sensitive load and the response characteristics are fitted to obtain the flexible load response range.

[0036] The topology and timing constraint data are input into multiple types of energy storage models to obtain the adjustable energy storage capacity and the call strategy.

[0037] By integrating the flexible load response range, the adjustable energy storage capacity, and the dispatch strategy, the flexible resource adjustability is obtained.

[0038] Optionally, the construction of the time-series production simulation model further includes:

[0039] By embedding the flexibility of resource adjustability and joint scenario set into the time-series power balance framework, a basic time-series production simulation model is obtained.

[0040] The basic time-series production simulation model is subjected to hierarchical dimensionality reduction and system partitioning aggregation to obtain a simplified time-series production simulation model.

[0041] With power supply resilience, renewable energy consumption, and operational economy as optimization objectives, a multi-objective optimization function is introduced to obtain the final time-series production simulation model.

[0042] Optionally, based on the system power supply resilience index, the power supply resilience of the clean energy power system in the target area under different extreme weather scenarios is analyzed, and power supply resilience prediction results are generated, including:

[0043] By inputting different extreme weather scenarios into the time-series production simulation model, the corresponding power and electricity time-series balance simulation data are obtained;

[0044] Based on the power supply time-series balance simulation data and the system power supply resilience index, calculate the power shortage probability, power shortage probability and sufficiency value under each scenario.

[0045] By performing weighted statistics on the power shortage probability, the power supply insufficiency probability, and the sufficiency value, a quantitative value of the power supply resilience of the target area under the corresponding extreme weather scenario is obtained.

[0046] The power supply resilience quantification value is compared with a preset threshold to determine the power supply resilience prediction result, which includes risk level, weak nodes and resilience margin.

[0047] Optionally, training the joint generation model for the source-load uncertainty scenario further includes:

[0048] Cluster analysis was performed on historical renewable energy output and load data under extreme weather conditions to obtain typical scenarios and their corresponding probability distributions under each type of extreme weather.

[0049] The typical scenario is combined with a residual neural network to adaptively extract the linkage features of new energy and load, establish a parameterized coupling function, and construct a joint representation model of source-load uncertainty under extreme weather conditions; the joint standard model of source-load uncertainty is used to assist the training of the conditional generative adversarial network.

[0050] Optionally, it also includes:

[0051] A long short-term memory network with an attention mechanism is introduced into the multi-task learning framework;

[0052] Time series modeling is performed on the correlation between the meteorological characteristic factors and climate change indicators to output a climate change correlation feature vector.

[0053] The climate change-related feature vector and the meteorological feature factor dataset are input together into the shared layer for high-dimensional feature integration.

[0054] On the other hand, the present invention also provides an electronic device, including a memory, a processor, and a computer program stored in the memory and executable on the processor, wherein the processor executes the program to implement the clean energy power system resilience prediction method based on multiple extreme weather scenarios as described above.

[0055] On the other hand, the present invention also provides a non-transitory computer-readable storage medium having a computer program stored thereon, which, when executed by a processor, implements the clean energy power system resilience prediction method based on multiple extreme weather scenarios as described above.

[0056] On the other hand, the present invention also provides a computer program product, including a computer program that, when executed by a processor, implements the clean energy power system resilience prediction method based on multiple extreme weather scenarios as described above.

[0057] This invention provides a method for predicting the resilience of clean energy power systems based on multiple extreme weather scenarios. It organically integrates extreme weather evolution trend data with source-load scenario generation. The method inputs identified extreme weather types into a multi-task learning framework, outputting evolution trend data. Then, it uses time-series meteorological element data reflecting the current weather conditions and evolution trend data reflecting future changes as input conditions to a joint generation model for source-load uncertainty scenarios, generating a joint scenario set. This joint scenario set not only reproduces historical source-load output characteristics but also characterizes the impact of extreme weather evolution on future source-load output under the background of climate change. Finally, it obtains system power supply resilience indicators and generates prediction results through time-series production simulation, effectively improving the foresight and accuracy of power supply resilience prediction under extreme weather scenarios. Attached Figure Description

[0058] To more clearly illustrate the technical solutions in this invention or the prior art, the drawings used in the description of the embodiments or the prior art will be briefly introduced below. Obviously, the drawings described below are some embodiments of this invention. For those skilled in the art, other drawings can be obtained from these drawings without creative effort.

[0059] Figure 1 This is a flowchart illustrating the method for predicting the resilience of clean energy power systems based on multiple extreme weather scenarios provided in this embodiment of the invention.

[0060] Figure 2 This is a schematic diagram of the structure of the electronic device provided in an embodiment of the present invention. Detailed Implementation

[0061] To make the objectives, technical solutions, and advantages of this invention clearer, the technical solutions of this invention will be clearly and completely described below with reference to the accompanying drawings. Obviously, the described embodiments are only some, not all, of the embodiments of this invention. All other embodiments obtained by those skilled in the art based on the embodiments of this invention without creative effort are within the scope of protection of this invention.

[0062] Figure 1 This is a flowchart illustrating the clean energy power system resilience prediction method based on multiple extreme weather scenarios provided in this embodiment of the invention.

[0063] like Figure 1 As shown in the figure, the clean energy power system resilience prediction method based on multiple extreme weather scenarios provided by this invention mainly includes the following steps:

[0064] 101. Obtain meteorological data and power system operation data of the target area to obtain a fused dataset, and extract time-series data of meteorological elements that affect the output and load of new energy sources from the fused dataset.

[0065] Specifically, the meteorological data includes ERA5 historical meteorological reanalysis data, ground meteorological station observation data, and power grid micro-meteorological monitoring data. ERA5 reanalysis data provides multi-layered geopotential field information, covering multiple pressure surfaces such as 1000hPa and 850hPa. Ground meteorological station observation data provides measured records of conventional meteorological elements such as temperature, humidity, wind speed, and precipitation. Power grid micro-meteorological monitoring data comes from micro-meteorological sensors deployed on transmission towers and at renewable energy power plants. Power system operation data includes provincial or regional wind power output data, photovoltaic power output data, and electricity load data.

[0066] These multi-source heterogeneous data were cleaned, formatted uniformly, and aligned with temporal and spatial resolutions to obtain a standardized fused dataset. The temporal resolution was standardized to 15 minutes per data point. Spatial resolution was determined by grid matching based on the geographical locations of renewable energy power plants. Time-series data of meteorological elements directly related to renewable energy output and load were extracted from the fused dataset, including wind speed, irradiance, temperature, humidity, and air pressure. Through multi-source data fusion, the correlation between meteorological elements and the operating status of the power system during extreme weather events can be comprehensively depicted.

[0067] 102. Perform feature extraction and modal analysis on the fused dataset to obtain the types of extreme weather processes that affect the power supply balance of the power system.

[0068] Specifically, feature extraction and modal analysis are performed on the fused dataset to identify the types of extreme weather events affecting the power supply balance of the power system, including:

[0069] Extract multi-layer high-altitude field data from the fused dataset.

[0070] The extraction of multi-layer geomorphic field data involves using bilinear interpolation to uniformly interpolate the data from the fused dataset onto a 0.25°×0.25° grid, forming a gridded, time-continuous multi-source fused dataset. The extraction operation is performed using the xarray and cfgrib libraries in the Python programming language. We define a function that automatically extracts six core physical quantities from the fused dataset based on the start time and affected area of ​​historically recorded extreme weather events: geopotential height, temperature, relative humidity, and U / V direction wind speed from four standard isobaric surfaces at 1000 hPa (representing near-surface), 850 hPa (representing lower altitude), 500 hPa (representing mid-altitude), and 200 hPa (representing upper altitude). Finally, the data is organized and stored in a five-dimensional tensor format, resulting in multi-layer geomorphic field data.

[0071] The meteorological modal characteristics are obtained by performing EOF orthogonal function decomposition on the geomorphic field data.

[0072] After obtaining the geopotential height field data, Empirical Orthogonal Function (EOF) decomposition is performed. The purpose is to decompose the complex, high-dimensional spatiotemporal field data into several mutually orthogonal spatial modes and their corresponding time coefficients, thus using a few principal modes to represent the main variability characteristics of the original field. Specifically, the geopotential height field data for each altitude layer is used as the processing object. Assume the data matrix is ​​X, with dimensions (T, N), where T is the number of time steps (e.g., 96 hours × 4 time steps / day = 384 time steps), and N is the number of spatial grid points (e.g., longitude × latitude, approximately 200 × 200 = 40,000 points). The PCA (Principal Component Analysis) module from the scikit-learn library in Python is used because EOF and PCA are mathematically equivalent. When calling the PCA module, the n_components parameter is set to 20, meaning the algorithm is required to retain only the first 20 principal components that explain the largest variance. The algorithm outputs two key results: first, `components_`, a (20, N) matrix where each row represents a spatial mode, intuitively a two-dimensional map reflecting the spatial distribution pattern of the pressure field; and second, `scores_`, a (T, 20) matrix where each column represents the weight coefficient of the corresponding spatial mode over time. The `explained_variance_ratio_` attribute allows you to view the variance proportion explained by each mode. For example, the variance contribution rate of the first mode might be as high as 60% or more, representing the main characteristics of the background circulation. The `components_` matrices of these first few main modes and their corresponding `scores_` time series are used as meteorological mode features.

[0073] Based on meteorological principles, threshold analysis and spatiotemporal feature extraction of meteorological modal characteristics are performed to construct a training sample set.

[0074] Input the meteorological modal features to be identified into the extreme weather classification and identification model, and output the type of extreme weather process.

[0075] Specifically, based on meteorological principles, the meteorological modal characteristics are analyzed using thresholds for trough and ridge intensity, pressure gradient, temperature gradient, wind speed, and cooling rate. Spatiotemporal features such as duration, path, impact range, and intensity center are extracted. These features are then combined with historical power system imbalance events for type labeling, forming a training sample set including cold wave, heat wave, heavy precipitation, strong wind and dust storm, and icing / snowfall types. After obtaining the training sample set, the meteorological data to be identified undergoes the same processing to obtain the meteorological modal characteristics to be identified. These are then input into the trained extreme weather classification and identification model, which outputs the corresponding extreme weather process type, confidence level, and impact range, thereby accurately identifying extreme weather processes affecting the power system's power supply balance.

[0076] By extracting meteorological modal features with clear physical meaning through multi-layer high-altitude field extraction and EOF decomposition, and constructing quantitative samples by combining meteorological principles and power operation events, we can achieve accurate classification and identification of extreme weather processes. The identification results are stable, reliable and highly reproducible, and can provide accurate meteorological boundary conditions for subsequent source-load scenario generation, time-series production simulation and power supply resilience assessment, effectively adapting to the safe operation and supply guarantee needs of high-proportion clean energy power systems.

[0077] 103. Input the extreme weather process type and use a multi-task learning framework to output the evolution trend data of extreme weather processes.

[0078] The evolution trend data includes the frequency of occurrence, the spatial range of influence, and the temporal trend.

[0079] In some embodiments, building a multi-task framework includes:

[0080] Empirical mode decomposition was performed on the fusion dataset corresponding to extreme weather process types to obtain meteorological factors;

[0081] A graph convolutional neural network was used to perform spatial feature aggregation and convolutional feature extraction on meteorological factors to obtain a meteorological feature factor dataset.

[0082] A multi-task learning framework is constructed, consisting of a shared layer and multiple task layers. The shared layer is used to integrate high-dimensional features of the meteorological feature factor dataset, and the multiple task layers correspond to the frequency prediction task, the spatial impact range prediction task, and the temporal change trend prediction task, respectively.

[0083] The multi-task learning framework is jointly trained using historical extreme weather process sample data. The network parameters of the shared layer are shared among the task layers, and the multi-task loss function is weighted and summed to optimize the model parameters. After training, the multi-task framework is obtained.

[0084] Specifically, Empirical Mode Decomposition (EMD) is performed on the fused dataset corresponding to each identified extreme weather event type (e.g., cold wave, heat wave). Each meteorological variable (e.g., temperature, air pressure) in the fused dataset is a time series that varies over time. The EMD algorithm adaptively decomposes these original sequences into several intrinsic mode function (IMF) components and a residual term. Each IMF component represents the oscillation pattern of the original signal at different time scales. For example, the first IMF might be a high-frequency diurnal variation, while the last IMF is a low-frequency interannual trend. These decomposed IMF components are collectively referred to as meteorological factors.

[0085] Graph Convolutional Neural Networks (GCNs) are used to aggregate the spatial features of the aforementioned meteorological factors. This aggregation first requires constructing a spatial topology graph. Each meteorological grid point or meteorological station within the study area is defined as a node, and edges are established based on the geographical distance between nodes. For example, a distance threshold is set; when the Euclidean distance between two nodes is less than 50 kilometers, an undirected edge is established between them. The initial feature vector of each node is the statistical value (such as mean, variance, and rate of change) of the meteorological factor sequence obtained in the previous step within a certain time window. Then, a two-layer graph convolutional network is constructed. The input of the first layer of the GCN is the node feature matrix and the adjacency matrix, and the output is the hidden features that aggregate the information of the surrounding one-hop neighbors; the second layer of the GCN further aggregates the information of the two-hop neighbors. Each layer of the GCN uses the Corrected Linear Unit (ReLU) as the activation function and incorporates a Dropout mechanism to prevent overfitting. After processing by the GCN, the features of each node are fused with the spatial information of its surrounding area; this set of node features is called the meteorological feature factor dataset.

[0086] A multi-task learning framework with a shared layer and multiple task layers is constructed. The shared layer first inputs the meteorological feature factor dataset obtained in the previous step into a flattening layer, concatenating all node feature vectors into a long vector. This long vector then passes through two fully connected layers: the first has 256 neurons, and the second has 128 neurons. Each fully connected layer is followed by a batch normalization layer and a ReLU activation function. The role of this shared layer is to integrate high-dimensional, scattered spatial features into a compact, general feature representation for use by all downstream tasks. Following this shared layer, three independent task layers are defined. The first task layer predicts the frequency of extreme weather events in the future; it is a single-neuron output layer using mean squared error (MSE) as the loss function. The second task layer predicts the spatial impact range; its output is a vector equal to the number of nodes, where each element represents the probability that the corresponding node location is affected. This is a multi-label classification problem, employing a binary cross-entropy loss function. The third task layer is used to predict the trend of time change. It consists of two stacked Long Short-Term Memory (LSTM) network layers, each with 64 hidden units. The input is the feature sequence of the shared layer output, and the output is the predicted value at multiple future time points. The mean absolute error is used as its loss function.

[0087] The multi-task learning framework is jointly trained using sample data from historical extreme weather events. During training, the overall loss function of the entire framework is defined as the weighted sum of the loss functions of each task, i.e. Based on the importance of each task, set... , , .in, Mean square error, The error of the binary classification cross-entropy loss function, Mean absolute error.

[0088] During training, the network parameters of the shared layer are updated simultaneously using gradients fed back from the three task layers, while the parameters of each task layer are updated only by its own loss function. We employ the Adam optimizer with an initial learning rate of 0.001, a batch size of 32, and an early stopping mechanism (patience=10) to prevent overfitting. Training is complete when the combined performance across multiple tasks no longer improves. At this point, we obtain a multi-task learning framework capable of simultaneously outputting predictions for frequency, range, and trend. Compared to training three independent models separately, this method reduces the number of parameters and improves training efficiency by sharing the underlying feature extractor. Furthermore, because the shared layer is forced to learn a general representation applicable to all tasks, the model's generalization ability on tasks with smaller sample sizes is enhanced.

[0089] After training to obtain the multi-task learning framework, the extreme weather process types are input into the multi-task learning framework, and the evolution trend data of extreme weather processes are output.

[0090] 104. Input the time series data and evolution trend data of meteorological elements into the joint generation model of source-load uncertainty scenario, and jointly simulate the probability distribution of new energy output and load output under extreme weather conditions, and output a joint scenario set.

[0091] In some embodiments, constructing a joint generation model for source-load uncertainty scenarios includes:

[0092] Historical renewable energy output data, load data, and time-series data and evolution trend data of meteorological elements during extreme weather events are acquired to construct a training sample set.

[0093] Using time-series meteorological data and evolution trend data as conditions, and random noise as input to the generator, a conditional generative adversarial network is constructed.

[0094] The conditional generative adversarial network includes a generator and a discriminator. The generator is used to output simulated combined power output data of new energy sources and loads, and the discriminator is used to distinguish between the generated combined power output data and the real combined power output data based on the training sample set.

[0095] The loss function of the discriminator is constructed based on the Vascular distance, and a gradient penalty term is introduced to constrain the gradient norm of the discriminator.

[0096] Regularization terms are constructed based on the joint distribution of KL divergence to constrain the generator parameter update magnitude. The conditional generative adversarial network is then trained in a game-like adversarial manner until convergence, resulting in a joint generative model for a scenario with uncertain source and load.

[0097] Specifically, samples of extreme weather events from the Qinghai power grid's historical records are collected, such as cold waves, heat waves, and heavy precipitation events that occurred between 2020 and 2024. For each event, renewable energy output data is extracted, including measured active power values ​​from wind farms and photovoltaic power stations every 15 minutes; load data, which is the province's electricity load recorded by the power grid dispatch system every 15 minutes; time-series meteorological data, including temperature, humidity, wind speed, irradiance, and air pressure, also at 15-minute intervals; and evolution trend data, obtained by calculating the rate of change through first-order difference of the meteorological element time-series data, and then fitting a linear slope within a 6-hour window as the evolution trend feature.

[0098] Continuous data from 72 hours before the event to 48 hours after the event is used as a single sample unit. After aggregating all event samples, each sample includes a multidimensional feature tensor (time step × number of meteorological elements) and a target tensor (the joint vector of renewable energy output and load at the same time step). Finally, Z-score standardization is applied to all feature and target values, i.e., subtracting the mean and dividing by the standard deviation, to obtain the training sample set.

[0099] After obtaining the training sample set, a Conditional Generative Adversarial Network (CGAN) is used as the basic architecture. The generator's input consists of two concatenated parts: the first part is a random noise vector, which can be sampled from a standard normal distribution with a dimension of 128; the second part is conditional information, namely time-series meteorological data and evolution trend data, which are flattened into a one-dimensional vector. The generator consists of four transposed convolutional layers, each followed by a batch normalization layer and a LeakyReLU activation function. The last layer uses a Tanh activation function to limit the output value to the range [-1, 1], and the length of the output sequence is the same as the actual combined output of new energy and load data. The discriminator's input includes two cases: one is the concatenation of the actual combined output data and conditional information, and the other is the concatenation of the generator's simulated combined output data and the same conditional information. The discriminator consists of five convolutional layers, each followed by a Dropout layer and a LeakyReLU activation function. Finally, a fully connected layer outputs a score between 0 and 1, representing the probability that the input is real data.

[0100] During training, the generator aims to maximize the discriminator's misclassification probability of generated data, while the discriminator aims to correctly distinguish between real and generated data. Compared to unconditional GANs, this conditional generative adversarial network ensures that the generated source-load joint output scenario strictly conforms to the given extreme weather conditions, solving the problem that traditional methods cannot incorporate external meteorological information into the generation process.

[0101] Traditional Generative Adversarial Networks (GANs) use JS divergence as the loss function for the discriminator, which is prone to training instability and mode collapse. Therefore, this invention uses Wasserstein distance to measure the difference between the real and generated data distributions. Furthermore, to satisfy the 1-Lipschitz continuity constraint of WGAN, a gradient penalty term is introduced instead of the original weight clipping. Specifically, the total loss function of the discriminator is defined as:

[0102] .

[0103] in, Total loss of the discriminator; The input noise to the generator follows a prior distribution. ; Labels are constrained by meteorological conditions and follow a distribution. ; Based on meteorological conditions For conditions, noise The source and load obtained from the input generator are used to jointly generate the scene; Based on meteorological conditions The discriminator outputs the result based on the given conditions; The data consists of real source-load scenario data measured in historical data, and follows the actual data distribution. ; The samples are linear interpolated between the real scene and the generated scene, and follow an interpolation distribution. ; For the discriminator about the interpolated samples The gradient; Calculate the L2 norm; This is the gradient penalty coefficient, used to balance the weights of the gradient penalty term.

[0104] Minimize during training To optimize the discriminator parameters iteratively, the gradient norm of the discriminator is forced to approach 1 through a gradient penalty term, which ensures training stability, avoids mode collapse, and enables the model to accurately learn the joint probability distribution of source load output under extreme weather conditions.

[0105] During the training of the joint generation model in a scenario with uncertain source and load, the generator loss function is optimized to address the mode collapse problem caused by excessively rapid updates of generator parameters.

[0106] The generator's original adversarial loss is - The training objective is to enable the discriminator to output high scores for generated scenes. Since the joint distribution of the real source and load is unknown, the k-nearest neighbor density estimation method is used. Based on the real and generated samples within the training batch, the KL divergence estimate between the generator output distribution and the real joint distribution is approximately calculated. This KL divergence is then introduced as a regularization term into the loss function. The final generator loss function is:

[0107] .

[0108] Among them, the regularization coefficient The experimental value was 0.01; The generator takes meteorological condition label c and random noise z as input and outputs the probability distribution of the source-load joint output scenario. This represents the true probability distribution of source-load combined output scenarios under extreme weather conditions, obtained from historical data collection.

[0109] During model iterative training, the generator parameters are fixed, and the discriminator parameters are updated three times (i.e., the discriminator update step is 3). Then, the discriminator parameters are fixed again, and the generator parameters are updated once based on the generator loss function mentioned above. Both the generator and discriminator use the Adam optimizer, with learning rates set to 0.0001 and 0.00005, respectively. Gradient pruning is performed on the generator to limit its gradient norm to within 1.0. During training, the temporal autocorrelation and spatial correlation errors between generated samples and real samples are calculated every 500 batches. If the error does not decrease after five consecutive detections, training is terminated early. After training, the generator network weights are saved, resulting in a converged joint generation model for source-load uncertainty scenarios. Through the constraint of the KL divergence regularization term, the generator is prevented from over-optimizing on a single output mode, forcing the model to cover the entire support set of the real data distribution. This effectively improves the diversity of generated scenarios and the ability to cover extreme output scenarios, ensuring the accuracy and reliability of scenario generation.

[0110] After the joint generation model for source-load uncertainty scenarios is constructed and trained, the time series data of meteorological elements and the evolution trend data are input into the joint generation model for source-load uncertainty scenarios. This allows for the joint simulation of the probability distribution of new energy output and load output under extreme weather conditions, and the output of a joint scenario set.

[0111] 105. Input the joint scenario set into the time-series production simulation model, and combine the calling strategies of various types of energy storage and the optimization model of adjustable load to perform power balance simulation calculations and obtain the system power supply resilience index.

[0112] In some embodiments, constructing a time-series production simulation model includes:

[0113] Historical operation data of the source-load-storage network is obtained and subjected to time-series normalization and boundary extraction to obtain time-series operation boundary data.

[0114] Specifically, historical operational data from at least the past three years is exported from the power grid dispatch automation system. This includes the actual output sequence of renewable energy sources (wind power and photovoltaic) every 15 minutes, the power consumption on the load side every 15 minutes, the charging and discharging power and state of charge of energy storage stations (electrochemical energy storage and pumped storage), and the transmission power and tie-line switching power of key grid sections. These data come from different acquisition systems, and time stamps may contain biases and omissions. First, using 15 minutes as a unified time granularity, linear interpolation is used to fill in short-term missing values; for continuous missing data exceeding 2 hours, the corresponding date range is directly removed. Next, the daily operational boundaries are extracted: maximum load, minimum load, maximum renewable energy output, minimum renewable energy output, maximum charging power of energy storage, and maximum discharging power. Then, a sliding window method is used to identify and remove outliers. For example, if the load at a certain moment exceeds three times the standard deviation of the mean of the preceding and following moments, it is considered an anomaly and is smoothed out. After the above regularization is completed, the data is organized into a time-series matrix with days as the unit. Each row contains 288 time points (24 hours × 4 points / hour), and each column corresponds to a physical quantity. The final output time-series operational boundary data is a structured table that records the upper and lower limits of load, the upper and lower limits of renewable energy output, the adjustable range of energy storage, and the planned values ​​of tie lines at each time of day.

[0115] Power grid topology analysis and cross-sectional constraint extraction are performed on the time-series operational boundary data to obtain topology and time-series constraint data.

[0116] Specifically, based on the CIM / E or CDF format model file of the power grid, the physical topology of the power grid is parsed. This involves reading the connection relationships of all buses, AC lines, transformers, generators, loads, and energy storage devices to construct an undirected graph. Nodes in the graph represent buses, and edges represent line or transformer branches. Next, a depth-first search algorithm is used to identify electrical islands in the power grid, i.e., subgraphs connected by the closure of switches / disconnectors. Then, using time-series boundary data, constraints are extracted for each critical transmission section.

[0117] Critical sections are typically predefined by scheduling and operation procedures. For example, the Qingla section includes three parallel lines. The total transmission power of this section at every 15-minute interval is calculated from historical data. The 2.5% and 97.5% quantiles for each of the past three years at that time are used as the lower and upper limits of the section's transmission power, respectively. For important lines without a clearly defined section affiliation, their maximum long-term operating current-carrying capacity is also extracted. Furthermore, the constraint range for node voltage amplitude is extracted, typically between 0.95 and 1.05 pu. Finally, the topology information (node-branch correlation matrix, electrical island partitioning results) is summarized with the time-series transmission limits of all sections and lines to form topology and time-series constraint data. This data is used for power balance and network security verification in subsequent time-series production simulations to ensure that the simulation results do not violate the actual physical limitations of the power grid.

[0118] It acquires the ability to adjust flexible resources and constructs a time-series production simulation model based on topological and temporal constraint data and joint scenario sets.

[0119] Specifically, flexible resources mainly include energy storage systems (electrochemical, pumped storage, hydrogen storage), adjustable loads (such as electric vehicles, air conditioning, and industrial interruptible loads), and the regulation margin of conventional units. For energy storage systems, rated power, rated capacity, charge / discharge efficiency, and upper and lower limits of state of charge are read from the equipment parameter table, and their continuous charge / discharge time (e.g., 2 hours or 4 hours) is calculated. For adjustable loads, based on historical response records, a curve of their response capability changing with the incentive price is fitted, and the maximum upward and downward adjustment ranges are set (e.g., air conditioning load can be increased by 10% or decreased by 5%). For conventional units (thermal power, hydropower), the regulation capability is described by their technical output upper and lower limits and ramp rate.

[0120] After acquiring the adjustability of flexible resources, the adjustability of flexible resources, topology and time-series constraint data, and a joint scenario set are used as input to construct a time-series production simulation model. The time-series production simulation model is a mixed-integer linear programming (MILP) problem. The objective function is set to minimize the total system operating cost (including fuel cost, start-up and shutdown cost, wind and solar curtailment penalties, and load shedding penalties). Constraints include: node power balance constraints, line / section power flow constraints (using a DC power flow model), unit output upper and lower limits and ramping constraints, energy storage charging and discharging and energy balance constraints, adjustable load response range constraints, and spinning reserve constraints. Open-source solvers (such as CBC) or commercial solvers (such as Gurobi) are used for solving the problem. The time scale of the time-series production simulation model is 15 minutes, and the simulation duration is typically one year (35,040 time segments). To reduce computational complexity, a rolling optimization strategy is adopted: each time the scheduling plan for the next 24 hours is optimized, then only the decision at the first time point is executed, and the calculation is repeated by moving the window. Ultimately, the time-series production simulation model can quantitatively assess how to meet load demands by utilizing flexible resources under different extreme weather scenarios, thereby providing power supply resilience indicators.

[0121] Among these, the ability to acquire flexible and adjustable resources includes:

[0122] Load monitoring data under extreme weather conditions are collected, and load levels and fluctuation characteristics are calculated.

[0123] Specifically, load monitoring data for the period of extreme weather events and the three days before and after them are exported from the power grid dispatching system, with a time resolution of 15 minutes. Each sample data includes a timestamp, total active power, and meteorological parameters such as temperature and relative humidity during the same period.

[0124] When calculating load level characteristics, the average of 96 sampling points per day is taken as the daily load level, and the difference between the maximum and minimum values ​​for that day is taken as the daily load peak-to-valley difference. Next, when calculating volatility characteristics, the average of the absolute values ​​of the first-order differences of the load series is calculated. Volatility characteristics reflect the minute-level drastic changes in load. Simultaneously, the standard deviation series with a 6-hour sliding window (24 points) is calculated, and the maximum value is taken as the intraday volatility amplitude. These calculations are implemented in Python using the `diff()` and `rolling()` methods of the pandas library. Finally, the load level and volatility characteristics for each extreme weather date are output. This step compresses the raw load data into quantifiable characteristic indicators, providing input for subsequent temperature-sensitive load decomposition.

[0125] The load level and fluctuation characteristics are decomposed by temperature-sensitive load and the response characteristics are fitted to obtain the response range of flexible load.

[0126] Specifically, regression analysis is used to decompose the total load into two parts: a baseline load and a temperature-sensitive load. The baseline load is selected from the typical daily load during spring or autumn when the temperature is between 15-25℃, and the average daily curve of the load during this period is used as the benchmark. Temperature-sensitive load = Total load - Baseline load.

[0127] Establish a multiple linear regression model between temperature-sensitive load and temperature and humidity: ,in Temperature (°C) For relative humidity (%), the coefficients are fitted using the least squares method based on contemporaneous meteorological data and calculated temperature-sensitive load data. , , .

[0128] When fitting the response characteristics, the change in temperature-sensitive load is statistically analyzed in historical events for every 1°C increase (or decrease) in temperature. The median of multiple events is taken as the sensitivity coefficient (kW / ℃).

[0129] The flexible load response range is defined as: at the current temperature, the upward adjustable range of the temperature-sensitive load is... The downward adjustment range is These ratios can be determined based on local load structure surveys; for example, areas with a high proportion of air conditioning load will see larger increases. Ultimately, the flexible load response range will be output. This indicates the range within which the load can be actively adjusted under extreme weather conditions. The flexible load response range quantifies the demand side's ability to participate in system regulation, providing a decision boundary for flexible resources in time-series production simulation.

[0130] By inputting topology and timing constraint data into various types of energy storage models, the adjustable capacity and call strategy of energy storage can be obtained.

[0131] Specifically, a unified model for multiple energy storage types is established. These include pumped hydro storage, electrochemical energy storage, and hydrogen energy storage. For each type of energy storage, the following parameters are defined: rated power... Rated capacity Charge and discharge efficiency and State of Charge (SOC) range Self-discharge rate (0.01% per hour is generally negligible).

[0132] Then, using the node information and section constraints from the topology and timing constraint data as input, the energy storage device is connected to a specific bus. The adjustable capacity of the energy storage is calculated as follows:

[0133] ;

[0134] .

[0135] in, This is the upper limit of the discharge capacity. This is the maximum rechargeable capacity. The current state of charge. The time interval is 15 minutes. The call strategy adopts a heuristic rule based on time-of-use pricing or system net load fluctuations. When the net load (load - renewable energy output) is higher than a threshold, energy storage discharge is prioritized; when the net load is lower than another threshold, charging is prioritized. Meanwhile, to extend lifespan, the number of charge / discharge cycles per day is set to no more than twice, and the SOC (State of Charge) does not frequently reach the boundary. The final output is the time-series adjustable capacity of each energy storage device and its corresponding call priority.

[0136] By integrating flexible load response range, adjustable energy storage capacity, and dispatch strategies, flexible resource adjustability is achieved.

[0137] Specifically, after obtaining the flexible load response range, adjustable energy storage capacity, and dispatch strategy, fusion is performed. The fusion method is as follows: at each 15-minute time segment, the adjustable capacity of all flexible resources is aggregated by node to form the comprehensive flexible adjustment range of each node. The specific calculation formula is: the upward adjustable capacity of node i at time t = the sum of the dischargeable capacity of all energy storage at that node + the upward adjustable amplitude of the flexible load at that node; the downward adjustable capacity = the sum of the charging capacity of all energy storage + the downward adjustable amplitude of the flexible load.

[0138] For the invocation strategy, a priority ranking table is established: first, sorted by response speed, and then by cost. The final output is a dataset of adjustable flexible resources, which includes: the up / down adjustable power curve for each node, the invocation cost coefficient for each resource, and the invocation priority order. These adjustable flexible resources can be directly input into a time-series production simulation model as adjustable resource constraints and cost terms in the optimization problem, thereby achieving source-grid-load-storage coordinated optimization under extreme weather conditions and improving power supply resilience.

[0139] In some embodiments, constructing a time-series production simulation model further includes:

[0140] By embedding the flexibility of resource adjustability and joint scenario set into the time-series power balance framework, a basic time-series production simulation model is obtained.

[0141] The power balance framework uses a 15-minute calculation step. Within each step, the equation must satisfy the following: total power output + energy storage discharge power + tie line input power = load demand + energy storage charging power + tie line output power. Network losses are considered on both sides of the equation.

[0142] In the power balance framework, the adjustability of flexible resources is treated as an adjustable variable, and the set of joint scenarios is used as the input uncertainty scenario. A scenario-based stochastic optimization method is adopted: for each joint scenario, a set of power balance equations is established, and all scenarios share the same set of unit start-up and shutdown decisions. The unit output, energy storage charging and discharging, and flexible load adjustment in each scenario are used as adaptive variables in the second stage. The objective function is set to minimize the expected value of system operating cost in each scenario. In addition to power balance, the constraints include: upper and lower limits of thermal power unit output and ramp-up constraints, energy storage energy balance and SOC constraints, flexible load response range constraints, cross-sectional transmission limit constraints, and spinning reserve constraints. The above problem is expressed as a mixed-integer linear programming model using an algebraic modeling language and solved using a solver.

[0143] By performing hierarchical dimensionality reduction and system partitioning aggregation on the basic time-series production simulation model, a simplified time-series production simulation model is obtained.

[0144] Specifically, since the basic time-series production simulation model includes hundreds of nodes, thousands of branches, and hundreds of joint scenarios, the computational load for direct solution is extremely large. Therefore, a hierarchical dimensionality reduction and system partitioning aggregation technique is adopted for simplification. The hierarchical dimensionality reduction strategy is as follows: First layer, the original 15-minute resolution is merged into a 1-hour resolution, and the hourly average value is taken for renewable energy output and load, and the energy storage SOC is updated hourly. Second layer, the k-means clustering algorithm is used to cluster the joint scenario set, with each cluster center representing a typical scenario, and its probability in the population is used as the weight to replace the original hundreds of scenarios. The system partitioning aggregation method is based on topological and time-series constrained data, and uses a spectral clustering algorithm to divide the power grid into several electrically close regions. The division is based on the reciprocal of the reactance between nodes as a similarity measure, and the number of clusters is usually set to 5-10, so that the voltage phase angle of the nodes in each region is similar, and the power exchange between regions is limited by cross-sectional constraints.

[0145] After partitioning, the network topology within each region is represented as an equivalent virtual node. Detailed power flow within each region is no longer modeled; only the cross-sectional constraints of interconnecting lines between regions are retained. Simultaneously, the generators, loads, and energy storage within each region are summed algebraically by power to form an aggregated equivalent model for each region. The final result is a simplified time-series production simulation model.

[0146] With power supply resilience, renewable energy consumption, and operational economy as optimization objectives, a multi-objective optimization function is introduced to obtain the final time-series production simulation model.

[0147] Specifically, based on the simplified time-series production simulation model, the original single operating cost objective is expanded into three optimization objectives. First, the power supply resilience objective: represented by a weighted sum of the expected load shedding and the probability of load shedding. Second, the renewable energy consumption objective: represented by minimizing the total amount of wind and solar power curtailment. Third, the operational economy objective: represented by minimizing the total system operating cost. Since the three objectives have different and conflicting dimensions, a linear weighting method can be used to transform the multiple objectives into a single objective. Furthermore, Pareto front analysis can be used to generate a set of non-dominated solutions by changing the weighting coefficients, allowing dispatchers to choose a compromise solution based on the severity of actual extreme weather. The final model retains all constraints of the simplified model and replaces the objective function with the aforementioned weighted sum. The resulting final time-series production simulation model can ensure computational efficiency while considering power supply security, clean energy consumption, and economic efficiency, outputting the optimal unit combination, energy storage dispatch, flexible load response plan, and corresponding resilience indicators for each extreme weather scenario.

[0148] 106. Based on the system power supply resilience index, analyze the power supply resilience of the clean energy power system in the target area under different extreme weather scenarios and generate power supply resilience prediction results.

[0149] Specifically, based on the system power supply resilience index, the power supply resilience of the clean energy power system in the target area under different extreme weather scenarios is analyzed, and power supply resilience prediction results are generated, including:

[0150] By inputting different extreme weather scenarios into the time-series production simulation model, the corresponding power and electricity time-series balance simulation data are obtained;

[0151] Based on power supply time-series balance simulation data and system power supply resilience index, calculate the probability of power shortage, the probability of power shortage and the sufficiency value under each scenario.

[0152] By weighted statistical analysis of the probability of power shortage, the probability of power shortage and the sufficiency value, the quantitative value of power supply resilience of the target area under the corresponding extreme weather scenario is obtained.

[0153] The power supply resilience prediction results are generated by comparing the quantitative value of power supply resilience with the preset threshold, including risk level, weak nodes and resilience margin.

[0154] Specifically, the source-load uncertainty joint scenarios corresponding to different extreme weather such as cold waves, high temperatures, heavy precipitation, strong winds and dust storms, and ice and snow cover are input into the trained time-series production simulation model. The time-series production simulation model performs time-by-time power balance simulation calculations and outputs power balance simulation data such as system power output, load demand, tie line transmission power, flexible resource call status, and power shortage under each scenario.

[0155] Based on the constructed system power supply resilience index system, the probability of power shortage, the probability of power shortage, and the core values ​​of system adequacy under each extreme weather scenario are calculated based on the above simulation data. The weighted statistical method based on the influence of the index is used to perform weighted fusion calculation on the probability of power shortage, the probability of power shortage, and the adequacy values ​​to obtain the quantitative value of power supply resilience of the target area under the corresponding extreme weather scenario.

[0156] The power supply resilience quantification value is compared with the preset classification threshold. Combined with the power deficit distribution and node operation constraints in the simulation data, the weak nodes of the system are identified, the resilience margin is calculated, and finally the power supply resilience prediction result containing risk level, weak nodes and resilience margin is generated.

[0157] In some embodiments, training the joint generation model for source-load uncertainty scenarios further includes:

[0158] Cluster analysis was performed on historical renewable energy output and load data under extreme weather conditions to obtain typical scenarios and their corresponding probability distributions under each extreme weather type.

[0159] This process involves selecting labeled extreme weather events from historical data, each event containing a complete time series. During clustering, the renewable energy output and load sequences for each event are concatenated into a long vector as sample features. After clustering, each cluster represents a typical source-load linkage pattern. For each cluster, the mean and standard deviation of all samples within the cluster at each time point are calculated, forming the mean curve and fluctuation band for that typical scenario. Simultaneously, the probability of occurrence of that typical scenario is calculated by dividing the number of samples in that cluster by the total number of samples. A set of typical scenarios and their corresponding probability values ​​are provided for each extreme weather type.

[0160] By combining typical scenarios with residual neural networks to adaptively extract the linkage features between new energy sources and loads, a parameterized coupling function is established to construct a joint characterization model of source-load uncertainty under extreme weather conditions.

[0161] Specifically, the source-load uncertainty joint standard model is used to assist in the training of the conditional generative adversarial network. Typical scenarios obtained are used as input to the residual neural network. The residual network consists of multiple stacked residual blocks, each containing two convolutional layers, a batch normalization layer, and a ReLU activation function. Skip connections are introduced to directly add the input to the output to address the vanishing gradient problem in deep networks.

[0162] The input data has a time step and a feature dimension, with a time step of 96 (24 hours × 4 15-minute points) and a feature dimension of 3 (wind power, solar power, load).

[0163] The network structure is designed with a first layer of 7×7 convolutional layers, outputting 64 channels; then it passes through four residual blocks, each containing two 3×3 convolutional layers with 64, 128, 256, and 512 channels respectively; finally, it is connected to a global average pooling layer and a fully connected layer, outputting a linkage feature. The linkage feature encodes the nonlinear coupling relationship between renewable energy output and load.

[0164] A parameterized coupling function is established, which takes the linked feature vector as input and outputs a conditional probability distribution parameter. Specifically, the coupling function is implemented by a two-layer fully connected network: the first layer has 128 neurons, and the second layer outputs the distribution parameter. The entire model is trained using the mean squared error loss function, with the input being complete time-series data of a typical scenario, and the output being the distribution parameter that can reconstruct that scenario. After training, a source-load uncertainty joint representation model is obtained, which can extract linked features from any input and generate the corresponding joint probability distribution.

[0165] In some embodiments, the clean energy power system resilience prediction method based on multiple extreme weather scenarios provided by the present invention further includes:

[0166] Introducing a long short-term memory network with an attention mechanism into a multi-task learning framework;

[0167] Time series modeling is performed on the correlation between meteorological characteristic factors and climate change indicators, and climate change correlation feature vector is output;

[0168] Climate change-related feature vectors and meteorological feature factor datasets are input together into a shared layer for high-dimensional feature integration.

[0169] Specifically, a long short-term memory network with an attention mechanism is introduced into the already constructed multi-task learning framework. The long short-term memory network is used to perform time series modeling of the dynamic correlation between meteorological feature factors and climate change indicators. The evolution of meteorological feature factors with climate change is deeply explored and climate change-related feature vectors are output. Then, the climate change-related feature vectors and the meteorological feature factor dataset are input into the shared layer of the multi-task learning framework. The shared layer completes the high-dimensional feature integration and unified representation of the two types of features, providing more comprehensive feature support for subsequent multi-task prediction.

[0170] Figure 2 This is a schematic diagram of the structure of the electronic device provided in an embodiment of the present invention.

[0171] like Figure 2As shown, the electronic device may include a processor 210, a communications interface 220, a memory 230, and a communication bus 240. The processor 210, communications interface 220, and memory 230 communicate with each other via the communication bus 240. The processor 210 can call logical instructions from the memory 230 to execute a clean energy power system resilience prediction method based on multiple extreme weather scenarios.

[0172] Furthermore, the logical instructions in the aforementioned memory 230 can be implemented as software functional units and, when sold or used as independent products, can be stored in a computer-readable storage medium. Based on this understanding, the technical solution of the present invention, essentially, or the part that contributes to the prior art, or a part of the technical solution, 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 computer device (which may be a personal computer, server, or network device, etc.) to execute all or part of the steps of the methods described in the various embodiments of the present invention. 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.

[0173] On the other hand, the present invention also provides a computer program product, which includes a computer program that can be stored on a non-transitory computer-readable storage medium. When the computer program is executed by a processor, the computer is able to execute the clean energy power system resilience prediction method based on multiple extreme weather scenarios provided by the above methods.

[0174] In another aspect, the present invention also provides a non-transitory computer-readable storage medium having a computer program stored thereon, which, when executed by a processor, is implemented to perform the clean energy power system resilience prediction method based on multiple extreme weather scenarios provided by the methods described above.

[0175] The device embodiments described above are merely illustrative. The units described as separate components may or may not be physically separate. The components shown as units may or may not be physical units; that is, they may be located in one place or distributed across multiple network units. Some or all of the modules can be selected to achieve the purpose of this embodiment according to actual needs. Those skilled in the art can understand and implement this without any creative effort.

[0176] Through the above description of the embodiments, those skilled in the art can clearly understand that each embodiment can be implemented by means of software plus necessary general-purpose hardware platforms, and of course, it can also be implemented by hardware. Based on this understanding, the above technical solutions, in essence or the part that contributes to the prior art, can be embodied in the form of a software product. This computer software product can be stored in a computer-readable storage medium, such as ROM / RAM, magnetic disk, optical disk, etc., and includes several instructions to cause a computer device (which may be a personal computer, server, or network device, etc.) to execute the methods described in the various embodiments or some parts of the embodiments.

[0177] 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 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; and these 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.

Claims

1. A method for predicting the resilience of a clean energy power system based on multi-extreme weather scenarios, the method comprising: include: ​ Meteorological data and power system operation data of the target area are acquired to obtain a fused dataset, and time-series data of meteorological elements affecting the output and load of new energy sources are extracted from the fused dataset; Feature extraction and modal analysis are performed on the fused dataset to obtain the types of extreme weather processes that affect the power supply balance of the power system; Input the extreme weather process type into a multi-task learning framework, and output the evolution trend data of the extreme weather process, including the frequency of occurrence, spatial impact range, and temporal variation trend; The time-series data of the meteorological elements and the evolution trend data are jointly input into the source-load uncertainty scenario joint generation model to jointly simulate the probability distribution of new energy output and load output under extreme weather conditions, and output a joint scenario set; The combined scenario set is input into the time-series production simulation model, and the power balance simulation calculation is performed by integrating the calling strategies of various types of energy storage and the optimization model of adjustable load to obtain the system power supply resilience index. Based on the system power supply resilience index, the power supply resilience of the clean energy power system in the target area under different extreme weather scenarios is analyzed, and power supply resilience prediction results are generated.

2. The method of claim 1, wherein the method further comprises: Feature extraction and modal analysis were performed on the fused dataset to identify the types of extreme weather events affecting the power supply balance of the power system, including: Extract multi-layer high-dimensional field data from the fused dataset; The geomorphic field data were subjected to EOF orthogonal function decomposition to obtain meteorological modal characteristics; Based on meteorological principles, threshold analysis and spatiotemporal feature extraction are performed on the meteorological modal features to construct a training sample set. The classification model is trained using the training sample set to obtain an extreme weather classification and identification model; The meteorological modal features to be identified are input into the extreme weather classification and identification model, and the extreme weather process type is output.

3. The method of claim 1, wherein the method further comprises: Building a multi-task framework includes: Empirical mode decomposition is performed on the fusion dataset corresponding to the extreme weather process types to obtain meteorological factors; The meteorological factors are spatial feature aggregation and convolutional feature extraction are performed using a graph convolutional neural network to obtain a meteorological feature factor dataset. A multi-task learning framework is constructed, comprising a shared layer and multiple task layers. The shared layer is used to perform high-dimensional feature integration on the meteorological feature factor dataset. The multiple task layers correspond to the frequency prediction task, the spatial impact range prediction task, and the temporal change trend prediction task, respectively. The multi-task learning framework is jointly trained using historical extreme weather process sample data. The network parameters of the shared layer are shared by each task layer, and the multi-task loss function is weighted and summed to optimize the model parameters. After training, the multi-task framework is obtained.

4. The method of claim 1, wherein the method further comprises: The construction of a joint generation model for source-load uncertainty scenarios includes: Acquire historical renewable energy output data, load data, and time-series data and evolution trend data of meteorological elements during extreme weather events to construct a training sample set; Using the time-series meteorological data and the evolution trend data as conditions, and random noise as the input to the generator, a conditional generative adversarial network is constructed. The conditional generative adversarial network includes a generator and a discriminator. The generator is used to output simulated new energy power output and load combined power output data, and the discriminator is used to distinguish between the generated combined power output data and the real combined power output data based on the training sample set. The loss function of the discriminator is constructed based on the Vascular distance, and a gradient penalty term is introduced to constrain the gradient norm of the discriminator. A regularization term is constructed based on the joint distribution of KL divergence to constrain the parameter update magnitude of the generator. The conditional generative adversarial network is then trained in a game-like adversarial manner until convergence, resulting in a joint generative model for a source-load uncertainty scenario.

5. The method of claim 1, wherein, The construction of a time-series production simulation model includes: Historical operation data of the source-load-storage network is obtained and subjected to time-series normalization and boundary extraction to obtain time-series operation boundary data; The time-series operational boundary data is subjected to power grid topology analysis and cross-sectional constraint extraction to obtain topology and time-series constraint data; The system acquires flexible and adjustable resource capabilities, and constructs a time-series production simulation model based on the topology and time-series constraint data and the joint scenario set.

6. The method of claim 5, wherein the method further comprises: Access to flexible and adjustable resources includes: Load monitoring data under extreme weather conditions were collected, and load levels and fluctuation characteristics were calculated. The load level and fluctuation characteristics are decomposed into temperature-sensitive load and the response characteristics are fitted to obtain the flexible load response range. The topology and timing constraint data are input into multiple types of energy storage models to obtain the adjustable energy storage capacity and the call strategy. By integrating the flexible load response range, the adjustable energy storage capacity, and the dispatch strategy, the flexible resource adjustability is obtained.

7. The method of claim 5, wherein the method further comprises: The construction of the time-series production simulation model also includes: By embedding the flexibility of resource adjustability and joint scenario set into the time-series power balance framework, a basic time-series production simulation model is obtained. The basic time-series production simulation model is subjected to hierarchical dimensionality reduction and system partitioning aggregation to obtain a simplified time-series production simulation model. With power supply resilience, renewable energy consumption, and operational economy as optimization objectives, a multi-objective optimization function is introduced to obtain the final time-series production simulation model.

8. The method of claim 1, wherein, Based on the aforementioned system power supply resilience index, the power supply resilience of the clean energy power system in the target area under different extreme weather scenarios is analyzed, and power supply resilience prediction results are generated, including: By inputting different extreme weather scenarios into the time-series production simulation model, the corresponding power and electricity time-series balance simulation data are obtained; Based on the power supply time-series balance simulation data and the system power supply resilience index, calculate the power shortage probability, power shortage probability and sufficiency value under each scenario. By performing weighted statistics on the power shortage probability, the power supply insufficiency probability, and the sufficiency value, a quantitative value of the power supply resilience of the target area under the corresponding extreme weather scenario is obtained. The power supply resilience quantification value is compared with a preset threshold to determine the power supply resilience prediction result, which includes risk level, weak nodes and resilience margin.

9. The method of claim 4, wherein the method further comprises: The training of the joint generation model for source-load uncertainty scenarios also includes: Cluster analysis was performed on historical renewable energy output and load data under extreme weather conditions to obtain typical scenarios and their corresponding probability distributions under each type of extreme weather. The typical scenarios are combined with residual neural networks to adaptively extract the linkage features of new energy sources and loads, establish parameterized coupling functions, and construct a joint representation model of source-load uncertainty under extreme weather conditions; the joint standard model of source-load uncertainty is used to assist the training of the conditional generative adversarial network.

10. The method of claim 3, wherein the method further comprises: Also includes: A long short-term memory network with an attention mechanism is introduced into the multi-task learning framework; Time series modeling is performed on the correlation between the meteorological characteristic factors and climate change indicators to output a climate change correlation feature vector. The climate change-related feature vector and the meteorological feature factor dataset are input together into the shared layer for high-dimensional feature integration.

Images