Method and device for determining hybrid energy storage configuration scheme

Abstract

The embodiment of the invention discloses a hybrid energy storage configuration scheme determination method and device, and the method comprises the steps: obtaining the historical output data of new energy, and carrying out the scene clustering of the historical output data, so as to obtain scene class labels corresponding to different weather conditions; training by adopting a preset conditional generative adversarial network model based on the scene category label to obtain a trained target model, and inputting a target scene category label corresponding to the extreme scene by utilizing the target model to generate a target output data set in the extreme scene; a double-layer optimization configuration model is established, the target output data set serves as input of the double-layer optimization configuration model, and an optimal configuration scheme of hybrid energy storage is obtained through iterative solution; wherein the double-layer optimization configuration model comprises an upper-layer planning model and a lower-layer operation model. In this way, a large number of high-fidelity new energy output scenes in extreme weather can be generated, and energy storage optimization configuration considering economical efficiency and toughness is carried out on the basis.

Description

TECHNICAL FIELD

[0001] The present application relates to the technical field of energy storage planning, and particularly relates to a method and device for determining a hybrid energy storage configuration scheme. BACKGROUND

[0002] In recent years, extreme catastrophic events such as typhoons, heavy rains, and earthquakes have shown a trend of increasing frequency and intensity on a global scale. Such events have a huge impact on power systems, especially distribution networks at the end of the power grid, frequently causing serious large-scale and long-duration power outages, and exposing the short board of current distribution networks in coping with extreme disasters. These accidents not only cause significant direct and indirect economic losses, but also seriously affect the normal operation of society and public safety. Therefore, it is urgent to explore new methods for resilience improvement that are more proactive, adaptive, and economical. In this context, energy storage systems are considered as one of the key technical means to improve the resilience of distribution networks due to their unique capabilities such as energy time shifting, rapid power support, and voltage frequency regulation. By optimizing the configuration of energy storage at key nodes of the distribution network (especially considering the complementary characteristics of different types of energy storage, such as fast-responding battery energy storage and high-energy-density hydrogen energy storage suitable for long-time backup), preventive scheduling (pre-charging / hydrogen storage) can be performed before disasters occur, emergency backup power and voltage support for critical loads can be provided during disasters, and power supply can be restored quickly after disasters. This not only significantly enhances the reliability and recovery capability of the distribution network under extreme events, but also effectively reduces user outage time and economic losses.

[0003] However, when optimizing the configuration of energy storage to improve the resilience of the power grid, the existing technology faces key challenges. One key challenge is the scarcity and difficulty of depicting extreme disaster scenario data. Historical records of extreme events are often insufficient and difficult to cover future possible "black swan" or "grey rhino" scenarios that are more destructive. Relying solely on limited historical data for resilience planning and energy storage optimization may result in insufficient robustness of the scheme, which cannot effectively cope with truly severe extreme tests, and may result in waste of investment and safety risks when the energy storage system fails to perform as expected in real severe tests. Therefore, how to generate high-fidelity extreme scenario data and conduct energy storage optimization that takes into account both economy and resilience is a technical problem that needs to be solved in this field. SUMMARY

[0004] The main purpose of the present application is to provide a method and device for determining a hybrid energy storage configuration scheme, which can solve the lack of means for generating high-fidelity extreme scenario data and means for conducting energy storage optimization that takes into account both economy and resilience on this basis in the prior art.

[0005] To achieve the above-mentioned purpose, a method for determining a hybrid energy storage configuration scheme is provided in the first aspect of the present application, which comprises: acquire historical output data of a new energy source, and perform scene clustering on the historical output data to obtain scene category labels corresponding to different weather conditions; based on the scene category labels, train a preset conditional generative adversarial network model to obtain a trained target model, and input a target scene category label corresponding to an extreme scene into the target model to generate a target output data set of the new energy source under the extreme scene; establish a double-layer optimization configuration model, input the target output data set into the double-layer optimization configuration model as an input, and obtain an optimal configuration scheme of a hybrid energy storage through iterative solution; The double-layer optimization configuration model comprises an upper-layer planning model and a lower-layer operation model. The upper-layer planning model takes a total cost of energy storage and a power distribution network resilience index as optimization objectives, solves through a multi-objective optimization algorithm, optimizes the comprehensive cost and the power distribution network resilience index, and determines a configuration scheme of the energy storage; wherein the total cost comprises a construction cost, a maintenance cost, an operation cost and a comprehensive operation cost of the energy storage, and the power distribution network resilience index is an index determined according to a loss of load rate of a system under an extreme scene. The lower-layer operation model performs multi-scene operation simulation according to the configuration scheme and the target output data set determined by the upper-layer planning model, with the operation cost and the comprehensive operation cost of the energy storage being the lowest as the objective, and feeds back the obtained operation cost and comprehensive operation cost to the upper-layer planning model.

[0006] To achieve the above object, a second aspect of the present application provides a device for determining a configuration scheme of a hybrid energy storage, which comprises: a scene clustering module, configured to acquire historical output data of a new energy source, and perform scene clustering on the historical output data to obtain scene category labels corresponding to different weather conditions; a scene generation module, configured to train a preset conditional generative adversarial network model based on the scene category labels to obtain a trained target model, and input a target scene category label corresponding to an extreme scene into the target model to generate a target output data set of the new energy source under the extreme scene; The optimization configuration module is configured to establish a double-layer optimization configuration model, take the target output data set as an input of the double-layer optimization configuration model, and obtain an optimal configuration scheme of the hybrid energy storage through iterative solution; wherein the double-layer optimization configuration model comprises an upper-layer planning model and a lower-layer operation model; the upper-layer planning model takes a total cost of the energy storage and a power distribution network resilience index as optimization objectives, solves through a multi-objective optimization algorithm, optimizes the comprehensive cost and the power distribution network resilience index, and determines the configuration scheme of the energy storage; wherein the total cost comprises a construction cost, a maintenance cost, an operation cost and a comprehensive operation cost of the energy storage, and the power distribution network resilience index is an index determined according to a loss of load rate of the system under an extreme scenario; the lower-layer operation model performs multi-scenario operation simulation with the configuration scheme determined by the upper-layer planning model and the target output data set, and takes the operation cost and the comprehensive operation cost of the energy storage as the lowest objectives, and feeds back the obtained operation cost and comprehensive operation cost to the upper-layer planning model.

[0007] To achieve the above object, the third aspect of the present application provides a computer readable storage medium, which stores a computer program, and the computer program, when executed by a processor, causes the processor to perform the steps of the first aspect and any possible implementation.

[0008] To achieve the above object, the fourth aspect of the present application provides a computer device, which comprises a memory and a processor, and the memory stores a computer program, and the computer program, when executed by the processor, causes the processor to perform the steps of the first aspect and any possible implementation.

[0009] The embodiment of the present application has the following beneficial effects: The application provides a determination method of a hybrid energy storage configuration scheme, which comprises the following steps: obtaining historical output data of new energy, and performing scene clustering on the historical output data to obtain scene category labels corresponding to different weather conditions; based on the scene category labels, a preset conditional generative adversarial network model is trained to obtain a target model, and the target model is used to input a target scene category label corresponding to an extreme scene to generate a target output data set of the new energy under the extreme scene; a double-layer optimization configuration model is established, the target output data set is taken as an input of the double-layer optimization configuration model, and an optimal configuration scheme of the hybrid energy storage is obtained through iterative solving; wherein the double-layer optimization configuration model comprises an upper-layer planning model and a lower-layer running model; the upper-layer planning model takes a total cost of the energy storage and a power distribution network resilience index as optimization targets, solves through a multi-objective optimization algorithm, and cooperatively optimizes the comprehensive cost and the power distribution network resilience index, so that the configuration scheme of the energy storage is determined; wherein the total cost comprises a construction cost, a maintenance cost, an operation cost and a comprehensive operation cost of the energy storage, and the power distribution network resilience index is an index determined according to a loss of load rate of the system under the extreme scene; the lower-layer running model performs multi-scene running simulation with the operation cost and the comprehensive operation cost of the energy storage as targets according to the configuration scheme and the target output data set determined by the upper-layer planning model, and feeds back the obtained operation cost and comprehensive operation cost to the upper-layer planning model.

[0010] In the above manner, a large number of new energy output scenes under extreme weather with high fidelity can be generated through the conditional generative adversarial network model, the bottleneck of scarce historical data is overcome, the energy storage planning scheme based thereon can withstand more realistic and more severe scene tests, and the robustness is significantly enhanced; through the establishment of the double-layer optimization model and the iterative solving of the cost and the resilience index of the energy storage in a unified framework, a scientific balance point between the economy of the energy storage system in the whole life cycle and the improvement of the power grid safety resilience can be found, and the problems of excessive investment or insufficient resilience are avoided; finally, through the optimization configuration of the hybrid energy storage, the power distribution network has stronger active support and island operation capability when extreme disasters occur, can provide continuous power supply for critical loads, significantly reduces the power outage time and economic loss, and effectively improves the overall resilience level of the power grid. BRIEF DESCRIPTION OF DRAWINGS

[0011] In order to more clearly illustrate the technical solutions in the embodiments of the present application or the prior art, brief descriptions will be given to the drawings needed in the embodiments or prior art description. Obviously, the drawings in the following description are only some embodiments of the present application, and other drawings can be obtained by those skilled in the art without creative effort.

[0012] wherein: Figure 1 A flow chart of a method for determining a hybrid energy storage configuration scheme in an embodiment of the present application; Figure 2 A schematic diagram of a network structure of a CWGAN-GP model in an embodiment of the present application; Figure 3 A schematic diagram of a long-short-term energy storage planning model in an embodiment of the present application; Figure 4 A flow chart of an NSGA-II algorithm in an embodiment of the present application; Figure 5 A structural block diagram of a determination device for a hybrid energy storage configuration scheme in an embodiment of the present application; Figure 6 A structural block diagram of a computer device in an embodiment of the present application. DETAILED DESCRIPTION

[0013] The technical solutions in the embodiments of the present application will be described clearly and completely below with reference to the drawings in the embodiments of the present application. Obviously, the described embodiments are only a part of the embodiments of the present application, rather than all the embodiments. Based on the embodiments in the present application, all other embodiments obtained by those skilled in the art without creative work fall within the scope of protection of the present application.

[0014] The present embodiment provides a method for determining a hybrid energy storage configuration scheme, specifically a hybrid energy storage configuration method based on scenario generation and double-layer optimization. The method aims to generate high-fidelity extreme weather scenario data, and on this basis, to optimize the configuration of energy storage considering both economy and resilience, so as to improve the robustness of the distribution network in response to extreme disasters.

[0015] The present application proposes a long-short-term energy storage optimization configuration method that comprehensively considers the economy and resilience of the distribution network, to meet the resilience requirements of multi-time scale power systems. In view of the characterization problem of new energy output uncertainty, an improved generative adversarial network wind-solar joint output scenario generation method containing gradient penalty is proposed, which avoids model assumptions and prior distribution, and directly generates a large number of scenario sets with similar statistical characteristics to actual data. On this basis, the uncertainty range of extreme scenario new energy output can be obtained, providing the operating boundary of new energy output for subsequent energy storage planning. Finally, a double-layer optimization model for long-short-term energy storage configuration is established. The upper model takes the construction cost of energy storage and the system resilience index as the target to obtain the configuration scheme of energy storage. The lower model combines the operation scheme of long-short-term energy storage for multi-scenario simulation, taking the minimum comprehensive operation cost as the target to develop an optimal dispatching scheme. Through continuous optimization and iteration of the upper and lower models, an optimal energy storage configuration scheme considering economy and resilience is obtained.

[0016] Referring to Figure 1 , Figure 1 is a flow chart of a method for determining a hybrid energy storage configuration scheme in an embodiment of the present application. The method mainly includes a data acquisition and clustering step 101, an extreme scenario generation step 102, a double-layer optimization model step, and an output optimal configuration scheme step 103. The following will be described in detail around this overall process, combined with Figure 2 , Figure 3 and Figure 4 .

[0017] Referring to Figure 1 , Figure 1 is a flow chart of a method for determining a hybrid energy storage configuration scheme in an embodiment of the present application, as shown in Figure 1 The method includes the following steps: 101, obtaining historical output data of new energy, and clustering the historical output data to obtain scene category labels corresponding to different weather conditions; First, the data acquisition and clustering step 101 is performed. This step aims to analyze historical data to extract new energy output characteristics under different weather patterns and provide category labels for subsequent conditional scenario generation. Specifically, this step first performs data collection and preprocessing. From the historical database of the power system, the new energy output data of the target distribution network region in the past years (for example, 5 to 10 years) is obtained, mainly including the hourly or higher resolution (such as 15 minutes) power output sequence of wind turbine generators and photovoltaic power stations. At the same time, collect the meteorological data corresponding to the above data time points, such as wind speed, wind direction, total solar radiation, temperature, humidity, air pressure, and weather condition description (such as sunny, cloudy, overcast, rainy, snowy, typhoon, etc.). The collected raw data is preprocessed, including handling missing values (for example, the mean or interpolation method of the previous and next time can be used for filling), removing outliers (for example, data points exceeding the physical limit or equipment rated capacity), and normalizing the data to unify different dimensional data to the interval [0, 1] or [-1, 1], to eliminate the influence of dimensional differences on the subsequent clustering algorithm.

[0018] After the data preprocessing, scene clustering is performed. In this embodiment, the K-means algorithm is preferably used to cluster the historical new energy output data. As a widely used unsupervised learning algorithm, the K-means algorithm aims to divide the data set into K clusters, so that each data point belongs to the cluster represented by the nearest centroid. In specific operation, the number of clusters K needs to be determined first. The selection of K value can be based on experience, for example, it can be set according to the main weather types (sunny, cloudy, rainy, windy, snowstorm, etc.); or, as a more theoretically based way, the elbow rule or silhouette coefficient method can be used to determine the optimal K value by calculating the clustering evaluation index under different K values. Suppose K = 5 is determined through analysis. Then, the algorithm randomly initializes K centroids, and each historical output scene (for example, a 24-hour output curve) is assigned to the nearest centroid. Then, the centroid of each cluster is recalculated, that is, the mean of all data points in the cluster, and the process of reassignment and updating of the centroid is repeated until the centroid position no longer changes significantly, and the algorithm converges.

[0019] After clustering, each historical output scene is assigned to one of the K clusters. At this time, each cluster needs to be labeled with a category, and a physically meaningful scene category label y is assigned. This process can be completed in combination with meteorological data. For example, if it is found through analysis that more than 90% of the dates corresponding to all scenes belonging to cluster 1 are sunny, then cluster 1 can be labeled as "sunny scene". Similarly, "rainy scene", "windy scene" and the like can be identified. For data points corresponding to extreme weather records such as typhoons and snowstorms, they may be clustered into one or more independent clusters, which are labeled as "extreme weather scene". Through this step, the original unlabeled historical data is converted into a data set with scene category labels y, laying a foundation for the next step of conditional generation.

[0020] For example, the scene clustering of the historical output data to obtain scene category labels corresponding to different weather conditions comprises: calculating the Euclidean distance between all sample points of the historical output data and the current cluster center points; according to the nearest neighbor principle, each sample point is assigned to the cluster where the nearest cluster center is located; the mean of all sample points in each cluster is recalculated, and the mean is taken as the new cluster center point; the change between the current cluster center points and the new cluster center points is judged; if the change is less than a preset threshold, the final scene clustering result is obtained, and the scene clustering result includes a scene category label corresponding to each cluster; if the change is greater than or equal to the preset threshold, the step of calculating the Euclidean distance between all sample points of the historical output data and the current cluster center points is returned.

[0021] Among them, the scene clustering method based on K-means++ algorithm aims to transform the original wind and light output data into extreme disaster scene labels with clear physical meaning through innovative feature engineering and optimized clustering process, providing high information density condition input for subsequent generative model. The specific implementation steps are as follows: Step 1: Multi-source feature extraction and feature vector construction: Collect wind and light output historical data and its corresponding high-resolution meteorological data (wind speed, irradiance, temperature, pressure) and disaster record data.

[0022] Perform feature engineering to construct a high-dimensional comprehensive feature vector for each time series sample, which mainly includes: a. Original output features: wind and light output values at a specific time scale.

[0023] b. Time series dynamic features: Calculate the volatility index (standard deviation) of the output sequence.

[0024] c. Meteorological correlation features: Extract the spatial and temporal variation rate of key meteorological factors and calculate their correlation with the output sequence (Pears correlation coefficient).

[0025] d. Disaster evolution features: Introduce the intensity index and spatio-temporal location information of specific disasters (such as typhoon).

[0026] Standardize the constructed high-dimensional feature set and use principal component analysis (PCA) for dimension reduction to eliminate redundancy and retain the most representative feature components, forming the final clustering input feature matrix.

[0027] Step 2: Optimal clustering initialization and cluster number determination: Use K-means++ algorithm to initialize cluster center points, which can improve the stability and convergence speed of clustering results by making the initial center points as far away from each other as possible, replacing the random initialization method of traditional K-means.

[0028] Combine the elbow rule and contour coefficient analysis method to determine the optimal number of clusters k, ensuring the statistical rationality of scene division.

[0029] Step 3: Iterative clustering and center point update: Calculate the Euclidean distance of all sample points to the current cluster center points.

[0030] According to the nearest neighbor principle, assign each sample point to the cluster where the nearest cluster center is located.

[0031] Recalculate the mean of all sample points in each cluster, and take this mean as the new cluster center point.

[0032] The distance calculation, sample distribution and center point updating operations of this step are repeatedly performed until the change of the cluster center point is less than a preset threshold, iterative convergence is achieved, and finally the scene clustering result is obtained.

[0033] Step 4: Fine multi-dimensional label definition For each cluster generated by clustering, a structured multi-dimensional scene label is defined in combination with its corresponding original weather and disaster data. The label system usually includes: Weather type (such as typhoon, heavy rain, sunny).

[0034] Disaster intensity level (such as strong, medium, weak, divided according to meteorological feature threshold).

[0035] Disaster evolution stage (such as early stage, peak period, recovery period, determined by analyzing the time series characteristics of samples in the cluster).

[0036] This multi-dimensional label system converts the unsupervised clustering result into a scene label with clear physical meaning, providing accurate condition control information for the subsequent conditional generative adversarial network.

[0037] In one possible implementation, the hybrid energy storage includes one of the following combinations: a combination of battery energy storage for short-term power support and hydrogen energy storage for long-term energy reserve; or, a combination of battery energy storage and pumped storage.

[0038] 102, based on the scene category label, a preset conditional generative adversarial network model is trained to obtain a trained target model, and the target model is used to input the target scene category label corresponding to the extreme scene to generate a target output data set of new energy under the extreme scene; Then, the extreme scene generation step 102 is performed. In historical data, the number of samples of extreme weather events is often very limited, which is difficult to fully support robust optimization. The core purpose of this step is to use a generative model to artificially and massively generate new extreme scene data consistent with the distribution of historical data according to the scene category label y obtained in the previous step.

[0039] The conditional generative adversarial network model is a conditional Wasserstein generative adversarial network assisted by gradient penalty CWGAN-GP model (referred to as CWGAN-GP model), and it should be noted that, in order to solve the great uncertainty of wind and light output and overcome the difficulty of scarcity of historical data in extreme disaster scenarios, the training stability of WGAN-GP and the conditional generation ability of CGAN are combined, and an improved conditional Wasserstein generative adversarial network (CWGAN-GP) for new energy time series data generation is proposed. The model is not a simple combination of prior art, but a number of adaptive innovations for the specific application of "new energy output extreme scenario", aiming to generate massive scenario data with statistical authenticity and physical rationality.

[0040] The core improvement of the CWGAN-GP model of the application is reflected in the following aspects: 1) Hybrid network structure with time series perception: The generator (G) adopts a hybrid architecture of one-dimensional convolutional neural network (1D-CNN) and gated recurrent unit (GRU). The 1D-CNN is responsible for capturing the local fluctuations and short-term features of wind and light output, while the GRU effectively models the long-term dependencies and dynamic evolution process of the output sequence under extreme events, ensuring that the generated time series data has reasonable variation rules.

[0041] The discriminator (D) adopts a hybrid architecture of 1D-CNN and multi-head self-attention mechanism (Multi-Head Self-Attention). This design enables the discriminator to not only evaluate the local data authenticity, but also focus on the global context and key time points of the entire sequence through the self-attention mechanism, thereby more accurately judging the time sequence logic rationality of the generated sequence.

[0042] 2) Hybrid loss function with physical constraints: On the basis of the loss function of the standard CWGAN-GP, the application introduces two physical regularization terms to form a hybrid loss function: ; Wherein, The time series smoothness regularization term punishes unrealistic sharp fluctuations in generated data, ensuring the continuity of the output curve. The physical consistency regularization term forces the generated data to be physically consistent with the input meteorological condition label y through a pre-trained physical agent model. This design embeds power system domain knowledge into the data generation process, which is the key to improving the credibility of generated data.

[0043] Through these improvements, the model can stably and efficiently learn the output characteristics from wind and light historical data, and generate a large number of scenarios that are infinitely close to the real data and conform to the physical law under given refined condition labels (such as specific disaster types, intensities and stages obtained by clustering). By flexibly switching the condition labels, specific scenario sets for planning, operation, prediction and other purposes can be generated on demand, providing a solid data foundation for subsequent energy storage optimization configuration.

[0044] Further, the CWGAN-GP-based extreme scenario boundary delineation method includes the following steps P01 to P04: P01: Construct a conditional training data set. The wind and light output time series data obtained from the K-means++ algorithm with refined multi-dimensional labels is constructed as a training set for the conditional generative adversarial network. The data is preprocessed, such as normalization, and organized into a sample format suitable for input of the time series model.

[0045] P02: Train the improved CWGAN-GP model; input the training set constructed in P01 into the improved CWGAN-GP model of the present application for training, initialize the model with full-scenario data, learn the general pattern of wind and light output, and then focus on extreme scenario data, adopt a dynamically adjusted gradient penalty coefficient, and finely optimize the network. The training target is to minimize the hybrid loss function defined in the present application that combines the Wasserstein distance, the time series smoothness regular term and the physical consistency regular term, to ensure that the generator and the discriminator converge to the ideal state in the game.

[0046] P03: Directly generate extreme weather scenarios: after the model training converges, input the extreme weather label defined in the clustering algorithm into the generator as a condition, and sample different random noise vectors, thereby directly and batch generate new energy output scenario data corresponding to the specified extreme condition with high fidelity.

[0047] P04: Construct an enhanced extreme scenario set: combine the large amount of extreme weather new energy output data generated by the generator in P03 with the real existing extreme scenario data in the historical record to form an extreme scenario set that is significantly expanded in scale and enhanced in diversity. This set more completely covers possible extreme cases.

[0048] P05: Based on the enhanced extreme scenario set constructed in P04, statistical analysis is performed to depict the time-varying operating boundary of new energy output for planning purposes. The specific method is: at each time t, the maximum and minimum values of the output of all scenarios at this time are calculated, which are respectively taken as the output upper boundary Pmax(t) and the output lower boundary Pmin(t) at this time. This pair of boundary curves clearly quantifies the uncertainty range of new energy output under extreme scenarios.

[0049] Please refer to Figure 2 , Figure 2 is a schematic diagram of the network structure of a CWGAN-GP model in an embodiment of the present application; it shows the network structure schematic diagram of the conditional Wasserstein generative adversarial network supplemented with the gradient penalty model preferred in the present embodiment. The model, as an advanced variant of the generative adversarial network, is composed of a generator G and a discriminator D. In the present embodiment, the generator G and the discriminator D can adopt a recurrent neural network structure based on a long short-term memory network or a gated recurrent unit. Such a network is particularly suitable for processing time series data and can effectively capture the time sequence dependency of the new energy output curve.

[0050] The function of the generator G is to generate simulated data. It receives two inputs: one is a random noise vector z sampled from a standard normal distribution, and the other is a specified conditional label y (for example, the label of an "extreme weather scenario"). The generator G processes these two inputs and outputs a simulated new energy output curve, denoted as generated data G(z, y). The introduction of random noise z ensures that even under the same conditions, diverse curves can be generated instead of a single fixed pattern.

[0051] The function of the discriminator D is to identify the authenticity of the data. It receives real data x (i.e. real output curves extracted from the labeled historical data set) or generated data G(z, y) as input, and outputs an evaluation value to judge the degree of authenticity of the input data. Unlike the traditional generative adversarial network, which outputs a probability value of [0, 1], under the framework of the Wasserstein generative adversarial network, the discriminator (also known as the critic) outputs a real number, the difference of which can measure the Wasserstein distance between the real data distribution and the generated data distribution, thereby making the training process more stable.

[0052] The training of the model is a process of an antagonistic game. In the training, the generator G and the discriminator D are alternately optimized. In the optimization of the discriminator D, the parameters of the generator G are fixed, the real data x and the generated data G(z, y) are input into the discriminator D, and the parameters of D are adjusted so that the discriminator D gives the real data x as high a score as possible and the generated data G(z, y) as low a score as possible. In the optimization of the generator G, the parameters of the discriminator D are fixed, the random noise z and the conditional label y are input into the generator G to obtain the generated data G(z, y), and then the generated data G(z, y) is input into the discriminator D. At this time, the parameters of G are adjusted, and the goal is to make the discriminator D give the generated data G(z, y) as high a score as possible to "trick" the discriminator.

[0053] In order to further improve the training stability and avoid mode collapse, the embodiment introduces a gradient penalty mechanism. This mechanism constrains the gradient norm of the discriminator D by adding a penalty term to the loss function of the discriminator D. Specifically, it calculates the norm of the discriminator gradient at a random interpolation point between the real data and the generated data, and penalizes its deviation from 1. This ensures that the discriminator satisfies the Lipschitz continuity constraint, thereby ensuring the effectiveness of the Wasserstein distance estimation.

[0054] After thousands or even more iterations of training, the generator G and the discriminator D reach a Nash equilibrium. At this time, the generator G has learned the internal distribution of the new energy output data under different conditional labels y. To generate an extreme scenario data set, repeatedly input the label y corresponding to the "extreme weather scenario" into the trained generator G, and input a completely new random noise vector z each time. In this way, the generator G generates a large number of high-fidelity and diversified new energy output curves under extreme scenarios (for example, 1000 independent 24-hour wind and light output scenarios) to form an extreme scenario new energy output data set for subsequent optimization.

[0055] 103、establishing a double-layer optimization configuration model, taking the target output data set as an input of the double-layer optimization configuration model, and obtaining an optimal configuration scheme of the hybrid energy storage through iterative solution; The double-layer optimization configuration model includes an upper-layer planning model and a lower-layer running model. The upper-layer planning model takes the total cost of the energy storage and a power distribution network resilience index as optimization objectives, solves through a multi-objective optimization algorithm, optimizes the comprehensive cost and the power distribution network resilience index, and determines the configuration scheme of the energy storage; wherein the total cost includes a construction cost, a maintenance cost, an operation cost, and a comprehensive operation cost of the energy storage, and the power distribution network resilience index is an index determined according to a loss of load rate of the system under an extreme scenario. The lower layer operation model determines the configuration scheme and the target output data set according to the upper layer planning model, performs multi-scenario operation simulation with the lowest operation cost and comprehensive operation cost of energy storage as the target, and feeds back the obtained operation cost and comprehensive operation cost to the upper layer planning model.

[0056] In a feasible implementation manner, the upper layer planning model further comprises at least one constraint: a system total carbon emission constraint, a node energy storage expansion capacity constraint, and an annual power balance constraint.

[0057] In a feasible implementation manner, in the target function of the lower layer operation model, the comprehensive operation cost comprises: a thermal power unit start-stop cost, a fuel cost and an environmental cost; the target function further comprises a carbon trading income; and the constraint condition of the lower layer operation model comprises: a thermal power unit output upper and lower limit and climbing constraint, and a state of charge operation constraint of the hybrid energy storage.

[0058] Subsequently, a two-layer optimization model step is performed. This step is the core decision-making link of the whole method, and aims to solve the optimal configuration scheme of the hybrid energy storage considering economy and resilience by using the extreme scenario data set generated in the previous step.

[0059] Please refer to Figure 3 , Figure 3 This is a schematic diagram of a long-short term energy storage planning model in an embodiment of the application, which shows a two-layer optimization model structure used in the embodiment. The model is composed of an upper layer planning model and a lower layer operation model, and the two are iteratively optimized through information interaction.

[0060] The upper layer planning model is a macro-level multi-objective optimization model. The decision variable is the configuration scheme of the hybrid energy storage, specifically including which nodes of the distribution network to install energy storage, and what type and capacity of energy storage to install. In the embodiment, the considered hybrid energy storage combination is a battery energy storage system and a hydrogen energy storage system. The battery energy storage system has the advantages of fast response speed and high conversion efficiency, and is suitable for short-term power support and frequency regulation; while the hydrogen energy storage system (including electrolytic tank, hydrogen storage tank, fuel cell) has the advantages of large capacity and long-term storage, and is suitable for long-term energy reserve. Therefore, the decision variable can be represented as: ; Among them, and represent the rated power and rated capacity of the energy storage, is the node index of the distribution network, is the set of all candidate nodes.

[0061] The upper layer planning model has two optimization objectives, reflecting the trade-off between economy and resilience: The first objective is to minimize the annualized total cost. The total cost is composed of the investment cost and the maintenance cost of the energy storage. Specifically, the annualized total cost may be expressed as ; where and are the annualized investment costs of the battery storage and the hydrogen storage at node , respectively, which are usually calculated based on the initial investment, the lifetime, and the discount rate of the equipment. and are the corresponding annual operation and maintenance costs, which are usually a fixed percentage of the investment cost. The second objective is to maximize the grid resilience index, or equivalently, to minimize the degree of insufficient resilience. In the present embodiment, the grid resilience index is determined by the loss of load probability (LOLP) of the system under extreme scenarios. Specifically, the resilience index may be defined as the inverse of the expected unserved energy of the system under all generated extreme scenarios, or the negative of it. The LOLP or the unserved energy is calculated in the lower-level operational simulation. Thus, the second objective can be expressed as ; where is the set of all generated extreme scenarios, is the probability of scenario occurring (which can be assumed to be equal here), is the total amount of load shed by the system under scenario .

[0062] Upper-level planning model: It also contains a series of constraints to ensure the reasonableness of the planning scheme. As a preferred implementation, these constraints can include: 1. System total carbon emission constraint, i.e. the total annual carbon emission of the entire system under typical annual operating conditions should not exceed the pre-set upper limit to meet the requirements of environmental protection policies; 2. Node energy storage expansion capacity constraint, i.e. the installed energy storage capacity of each node is limited by physical space or grid access conditions; 3. Annual power balance constraint, which ensures that the total power generation of the system can meet the total load demand on the scale of a typical year.

[0063] Lower-level operational model: It is a deterministic operational simulation and optimization model. Its task is to receive a specific energy storage configuration scheme (i.e. the type, location, and capacity of the energy storage) passed by the upper-level planning model, and under this configuration, conduct optimal operational simulation under multiple scenarios in combination with the generated extreme scenario new energy output data set.

[0064] The objective function of the lower-level operation model is to minimize the annual total operation cost. In the extreme disaster scenarios, this operation cost mainly reflects the cost paid by the system to maintain power supply. Specifically, the operation cost may include: the start-up and shut-down cost of thermal power units, fuel cost, and environmental cost (such as penalties for pollutant emissions). In addition, in order to encourage low-carbon operation, the model can also consider the carbon trading revenue, i.e. the revenue obtained when the actual carbon emissions are lower than the quota.

[0065] The decision variables of the lower-level operation model are the output of each generation and energy storage device in each simulation period (e.g. every hour), such as the output power of thermal power units, the charge and discharge power of battery energy storage, the working power of electrolyzer and fuel cell in hydrogen energy storage system, etc.

[0066] In order to ensure the authenticity of the simulation, the lower-level operation model needs to meet a series of strict physical and operation constraints. These constraints mainly include: 1. System power real-time balance constraint, i.e. at any time, the total power generation (including thermal power, new energy, and energy storage discharge) must be equal to the sum of total load and energy storage charging power, if it cannot be balanced, load shedding is required; 2. Thermal power unit output upper and lower limit and ramping constraint, i.e. the output of thermal power units cannot exceed its rated capacity, nor can it be lower than its minimum output for stable operation, and the change rate of output between adjacent time periods cannot exceed its ramping rate limit; 3. State of charge operation constraint of hybrid energy storage, for battery energy storage, its state of charge must be maintained within a safe range (e.g. 20% to 90%), and at the end of a dispatching period, its state of charge should return to the initial level; for hydrogen energy storage, the change of hydrogen amount in the hydrogen storage tank also needs to be considered, and its constraint is similar to the state of charge constraint of battery energy storage.

[0067] The iterative solution process of the bi-level model is as follows: the upper-level planning model generates one or a set of energy storage configuration schemes and sends them to the lower-level operation model. The lower-level operation model simulates the operation of each received configuration scheme one by one in all generated extreme scenarios. After the simulation is completed, the average operation cost and average load loss of the configuration scheme in all scenarios are calculated. These calculation results are fed back as operation parameters to the upper-level planning model, which are used to evaluate the two objective function values (i.e. total cost and resilience) of the configuration scheme. This "upper-level proposal - lower-level evaluation - result feedback" process constitutes a complete iteration.

[0068] Finally, the output optimal configuration scheme step is executed. Since the upper-level planning model is a multi-objective optimization problem, its solution is usually not a single solution, but a set called "Pareto optimal solution set". Each solution in this set represents a different trade-off between cost and resilience, and there is no solution that is superior to another solution in all objectives.

[0069] In an implementation, the optimal configuration scheme is obtained by iteration, comprising: using a preset NSGA-II algorithm (Non-dominated Sorting Genetic Algorithm II) to solve the double-layer optimization configuration model to obtain a set of Pareto optimal solutions; and using an entropy weight method-gray target decision method to select the final optimal configuration scheme from the set of Pareto optimal solutions.

[0070] Please refer to Figure 4 , Figure 4 The flowchart of the NSGA-II algorithm in the embodiment of the application shows the flowchart of the non-dominated sorting genetic algorithm II used to solve the upper-layer multi-objective optimization problem in the embodiment. The algorithm is a high-efficiency multi-objective optimization algorithm. At the beginning of the algorithm, a parent population D0 is randomly initialized, wherein each individual represents a energy storage configuration scheme. Then, a child population Q0 is generated from the parent population D0 through genetic operations such as simulated binary crossover and polynomial mutation. Next, the parent population and the child population are combined into a large population R u , and all individuals in R u are quickly non-dominated sorted and divided into different non-dominated levels. For individuals in the same level, their crowding degrees are further calculated to evaluate their distribution density in the solution space. Finally, the non-dominated levels (preferentially selecting individuals in earlier levels) and the crowding degrees (preferentially selecting individuals with larger crowding degrees to maintain diversity) are used as evaluation criteria to select N best individuals from R u to form a new parent population D u+1 . The process is continuously looped until a termination condition (such as reaching a maximum number of iterations) is met. The final output of the algorithm is an approximate Pareto optimal solution set.

[0071] After obtaining the Pareto optimal solution set, the decision maker needs to select a final implementation scheme from the set. To make the selection process more objective, the entropy weight method-gray target decision method is preferably used in the embodiment. First, the entropy weight method is used to objectively determine the weight of each objective (cost and resilience) according to the dispersion degree of the value of each objective function in the Pareto solution set. Then, the gray target decision theory is used to calculate the "gray target distance" from each Pareto solution to the ideal optimal solution, with each Pareto solution being regarded as a decision scheme. Finally, the solution with the smallest gray target distance is selected as the final optimal configuration scheme. For example, the final scheme can be to configure 1 MW / 2 MWh of battery energy storage at node A of the power distribution network and 0.5 MW / 10 MWh of hydrogen energy storage at node B. This scheme can control the average loss of load rate to be within 5% under the generated 1000 extreme scenarios, and its annualized total cost is reduced by 15% compared with the configuration method considering only the conventional scenarios.

[0072] Please continue to refer toFigure 3 and Figure 4 Finally, the Kmeans++ clustering algorithm and the CWGAN-GP improved adversarial generation network are combined to depict the boundary of the extreme scene, and the energy storage optimization configuration method for improving the resilience of the distribution network is as follows: The application provides an energy storage optimization configuration method for improving the resilience of a distribution network: an upper planning model is a long-short-term energy storage configuration planning layer, and a lower layer is a running simulation layer. The upper layer takes annual comprehensive cost and system resilience as optimization targets, realizes site selection and capacity determination of energy storage facilities, and determines a planning construction scheme; the lower layer takes power system running economy as an optimization target, obtains an optimal economic dispatching scheme under the energy storage configuration scheme determined by the upper layer, and feeds back running parameters of the lower layer to the upper layer model, so that the final energy storage optimization configuration scheme is obtained through iterative optimization between the upper and lower layer models.

[0073] 1) Upper layer model The upper layer model is an upper planning model, and the upper planning takes energy storage capacity and installation position as decision variables of the model, obtains an optimal decision scheme on the basis of meeting power balance constraints, energy storage capacity and cost constraints and carbon emission constraints, and taking the lowest system comprehensive cost and the optimal flexibility index as optimization targets.

[0074] Objective function ; In the formula: system comprehensive cost is represented by C; system flexibility index is represented by F; investment cost of energy storage is represented by Cinv; maintenance cost of energy storage is represented by Cmaint; running cost of energy storage is represented by Coper; system comprehensive running cost mainly includes running cost of a unit and carbon emission cost, and is represented by Ctotal; resilience deficiency penalty cost is represented by Cpen; a set of selected typical scenes is represented by S; probability of the selected scene is represented by P; system resilience index under the selected scene S is represented by F.

[0075] Investment and construction cost of energy storage: ; In the formula: and annual investment and construction cost of BESS and HESS is represented by Cinv and Cinv, respectively; and fixed investment and construction cost of BESS and HESS is represented by Cinv and Cinv, respectively; and configuration number of BESS and HESS is represented by N and N, respectively; and Pnand Cnrespectively represent the power and capacity of the nth BESS; and Pnand Cnrespectively represent the power and capacity of the nth HESS; a and b represent the unit power cost and battery conversion cost of the energy storage, respectively; is the annual capital recovery rate; y is the service life of the energy storage; r is the discount rate.

[0076] Maintenance cost of energy storage: ; In the formula: and Maintenance coefficient of BESS and HESS (proportion of maintenance cost to equipment investment cost), respectively.

[0077] Operating cost of energy storage: ; In the formula: and The purchase price and the selling price of the energy storage during the charging and discharging process at time t.

[0078] System comprehensive operating cost: ; In the formula: comprehensive operating cost is composed of the fixed operation and maintenance cost of various power sources , the generation and start-stop cost of thermal power units ; fuel and environmental cost ; is the annual conversion coefficient, which aims to convert the operating cost into the annual time scale; represents the fixed operation and maintenance cost of power unit i in the tth year; represents the capacity of the nth unit in the tth year; is the generation cost function of thermal power units; , and , represent the start-stop cost and start-stop state of thermal power units; is the coal cost of thermal power units; is the power generation of thermal power units in the tth year; Th represents the set of thermal power units; represents the CO2 emission intensity; and represent the environmental value coefficient and the penalty coefficient.

[0079] Resilience index: ; In the formula: is the resilience index; is the probability of scenario n; This refers to the load level of the power distribution network under extreme weather disasters. This represents the load level during normal operation of the distribution network. This refers to the load loss due to faults in the distribution network.

[0080] Constraints: Carbon emission constraints: ; In the formula: Characterizing carbon emission intensity, This is the carbon emission limit for the system in year t.

[0081] Energy storage expansion capacity constraints: ; In the formula: and These are the rated power of the HESS and BESS connected to node i, respectively; and These represent the maximum access capacity for each type of energy storage at each node.

[0082] Power balance constraint: ; In the formula: , , These represent the annual power generation from photovoltaic, thermal power, and wind power, respectively. Represents the projected electricity demand for year t; This represents the power loss with a power source.

[0083] 2) Lower-level model The lower-level model serves as the lower-level operation model. The lower-level operation simulation is based on the energy storage configuration determined by the upper-level model. It takes the scheduling output of the system power units and flexible resources such as energy storage as decision variables, and aims to minimize the annual comprehensive cost. While ensuring power balance constraints, energy storage operation constraints, and unit output constraints during operation, it obtains the daily scheduling plan and energy storage operation conditions under various typical scenarios.

[0084] The objective function of the lower-level model: ; In the formula: For carbon trading revenue (the cost of reducing carbon emissions from conventional thermal power units by increasing power generation from new energy sources such as wind and solar); The conversion factor for CO2 produced per unit mass of coal combustion; This refers to the amount of coal consumed per unit of electricity generated by a conventional thermal power unit. For carbon trading fees; The power generation of new energy.

[0085] Constraints: Conventional unit operation constraints ; In the formula: , , Characterize the start-stop state of different types of units at time t; Characterize the minimum output level of thermal power units; , , Represent the maximum output level of different types of units; , , Characterize the power generation of different types of units at time t.

[0086] Energy storage operation constraints ; In the formula: , , , Represent the upper and lower limits of the SOC of BESS and HESS respectively; , Represent the SOC level of BESS and HESS at time t.

[0087] Power balance constraints ; In the formula: , , , Characterize the output of the i-th thermal power, photovoltaic, wind power and energy storage at time t; , , , Characterize the set of thermal power, photovoltaic, wind power and energy storage; Represent the load demand at time t.

[0088] 3) Model solution: Since the long-short term energy storage double-layer optimization configuration model proposed in the present application considers system resilience, energy storage construction cost and typical scenario operation comprehensive cost and other multiple objectives, it is a complex optimization model with multiple constraints, multiple variables, nonlinearity and strong coupling characteristics. The NSGA-II algorithm can quickly update the position of individual solutions in the search space and find the approximate optimal solution of the problem in the iteration process. In order to obtain the compromise solution in the Pareto solution set, the entropy weight method-gray target decision is used to obtain the best compromise solution, and the specific solving process can be referred to Figure 4The solving process shown.

[0089] The application provides a hybrid energy storage optimization configuration method for improving the resilience of a power distribution network, comprising the following steps: 1) scene clustering and refined label definition based on multi-source feature fusion: adopting a K-means++ algorithm, fusing new energy output data and high-dimensional derived features including the temporal and spatial variation rate of meteorological factors, output fluctuation indexes and disaster evolution characteristics, and constructing a clustering label system capable of accurately describing the coupling relationship between disasters and output; 2) extreme scene generation and boundary description based on an improved CWGAN-GP: proposing a time sequence perception CWGAN-GP model, and adopting a hybrid network structure of 1D-CNN+GRU and 1D-CNN+multi-head self-attention for the generator and the discriminator respectively to capture the output time sequence dependence relationship; and introducing a time sequence smoothness constraint and a physical consistency regularization term into the loss function to ensure that the generated data have statistical authenticity and physical rationality, so as to accurately describe the operation boundary of new energy output under extreme scenes; 3) hydrogen-electric hybrid energy storage double-layer optimization configuration: establishing a double-layer planning-operation model with economy and system resilience as dual objectives, and determining the optimal energy storage site selection and capacity and scheduling strategy through upper and lower layer iterative optimization. Through the innovative scene description method and energy storage configuration model, the application significantly improves the prevention, resistance and recovery ability of the power distribution network under extreme disasters, and effectively reduces the power outage loss.

[0090] The application provides a method for determining a hybrid energy storage configuration scheme, which can generate a large number of high-fidelity new energy output scenes under extreme weather through a conditional generative adversarial network model, overcomes the bottleneck of scarce historical data, makes the energy storage planning scheme based thereon withstand more realistic and more severe scene tests, and significantly enhances the robustness; and through the establishment of a double-layer optimization model and the iterative solution of the cost and resilience index of the energy storage in a unified framework, a scientific balance point between the economy and the power grid safety resilience improvement of the energy storage system throughout the life cycle can be found, and the problems of excessive investment or insufficient resilience are avoided; finally, through the optimization configuration of the hybrid energy storage, the power distribution network has stronger active support and island operation ability when extreme disasters occur, can provide continuous power supply for critical loads, significantly reduces the power outage time and economic loss, and effectively improves the overall resilience level of the power grid.

[0091] Further, embodiment 2 is provided, which is mainly different from embodiment 1 described above in that a different clustering algorithm is adopted in the data acquisition and clustering step 101. The remaining steps are basically the same as those described in embodiment 1.

[0092] In this embodiment, the scene clustering adopts a density-based noise application spatial clustering algorithm to replace the K-means algorithm in Embodiment 1. This algorithm is a density-based clustering algorithm that can divide closely connected samples into a cluster and effectively identify outliers (i.e. noise points), which makes it advantageous in handling irregularly shaped clusters and identifying abnormal data.

[0093] The specific working process is as follows: after data collection and preprocessing are completed, the historical new energy output data is input into the density-based noise application spatial clustering algorithm. This algorithm needs to set two key parameters: neighborhood radius eps and minimum neighborhood sample number min_pts. The algorithm starts from an arbitrary unvisited data point, checks whether there are at least min_pts data points within the neighborhood radius eps. If yes, the point is marked as a core point and a new cluster is created; then, the algorithm recursively adds all density-reachable points (including other core points and boundary points) of the core point to this cluster. If not, the point is temporarily marked as a noise point. This process continues until all data points are visited.

[0094] Unlike the K-means algorithm which requires pre-specification of the number of clusters K, the density-based noise application spatial clustering algorithm can automatically determine the number of clusters according to the distribution characteristics of the data itself. After clustering is completed, the data set is divided into several high-density clusters and an outlier set. These high-density clusters usually represent regular weather patterns such as "sunny" and "cloudy", which can be labeled in combination with meteorological data as in Embodiment 1. It should be noted that the outlier set identified by this algorithm naturally corresponds to those scenes with low statistical frequency and distinct characteristics, which are most likely to be the extreme weather scenes of interest. Therefore, all historical data identified as noise points can be directly labeled as "extreme scene".

[0095] Using the density-based noise application spatial clustering algorithm for clustering, the expected effect is to more accurately separate the samples of extreme scenes. Since extreme events themselves exhibit sparse and abnormal characteristics in data distribution, this feature of the algorithm can accurately capture these characteristics. This provides more representative "seed" samples of extreme scenes and clearer class boundaries for subsequent training of the conditional Wasserstein generative adversarial network supplemented by the gradient penalty model, thereby possibly generating a higher-quality and more distinctive extreme scene data set, ultimately improving the relevance and effectiveness of the energy storage configuration scheme in responding to real disasters.

[0096] Further, Embodiment 3 is provided, which mainly differs from Embodiment 1 in that a different generation model is used in the extreme scene generation step 102. The remaining steps are basically the same as described in Embodiment 1.

[0097] In this embodiment, a conditional variational autoencoder model is adopted to replace the conditional Wasserstein generative adversarial network supplemented by a gradient penalty model in Embodiment 1, for generating new energy output data sets under extreme scenarios. The conditional variational autoencoder is a conditional version of the variational autoencoder, which is also a powerful generative model that generates new data by learning the latent representation of the data.

[0098] The conditional variational autoencoder model consists of an encoder and a decoder. The encoder maps the input data x (i.e. the actual output curve) and its corresponding condition label y to a low-dimensional latent space, outputting the parameters of a probability distribution (usually the mean and variance of a Gaussian distribution) in the latent space. The decoder samples a latent variable z from the posterior distribution defined by the encoder, and combines z with the condition label y to attempt to reconstruct the original input data x. The training objective of the model is to maximize the lower bound of the evidence, which is equivalent to minimizing the reconstruction error (i.e. the difference between the generated data and the original data) and the regularization term (i.e. the KL divergence between the posterior distribution output by the encoder and a standard prior distribution).

[0099] The specific working process is as follows: first, use the labeled data set obtained by K-means clustering in Embodiment 1 to train the conditional variational autoencoder model. After training, the decoder has the ability to generate corresponding data distribution according to the specified condition label y and the vector z sampled from the latent space. To generate extreme scenario data, we input the "extreme weather" corresponding label y and a latent vector z randomly sampled from the standard normal distribution (i.e. the prior distribution) into the trained decoder. The decoder will output a new energy output curve that meets the extreme weather condition. By repeatedly performing this sampling and decoding process, a large and diverse set of extreme scenario data can also be generated.

[0100] Compared with the conditional Wasserstein generative adversarial network supplemented by a gradient penalty model, the training process of the conditional variational autoencoder is generally more stable and less prone to mode collapse problems. At the same time, by performing interpolation and other operations in the latent space, the conditional variational autoencoder can easily generate data samples with smooth transition characteristics. Using the conditional variational autoencoder model can also effectively complete the task of generating specific scenario data according to the label, which proves that the core idea of the technical solution of the present application - i.e. "conditional generation" - is universal, and its implementation method is not limited to a specific generative adversarial network architecture, but can also be implemented through other types of conditional generation models, thereby providing broader protection for the present application.

[0101] Example 4 is provided to illustrate that the double-layer optimization model proposed in this application has good flexibility and scalability, and can adjust the types of hybrid energy storage considered and the optimization objectives according to actual needs. The remaining steps of this example are the same as those of Example 1, and the main difference is the construction of the double-layer optimization model.

[0102] In this example, the double-layer optimization model is adjusted in two aspects: first, the technical combination of hybrid energy storage is adjusted. Considering the geographical and resource conditions in different regions, for example, in some mountainous or coastal areas, pumped storage power stations are more mature and more economical large-capacity energy storage options than hydrogen energy storage. Therefore, in this example, the combination of "battery energy storage + hydrogen energy storage" in Example 1 is replaced by the combination of "battery energy storage + pumped storage". Correspondingly, in the upper planning model, the decision variable is changed to the configuration capacity of battery energy storage and pumped storage, and the investment cost and operation cost model is also replaced by the cost parameters of the corresponding technology. In the lower running model, the operation constraints of hydrogen energy storage are replaced by the operation constraints of pumped storage power stations, which include reservoir capacity limits, pumping / generating power limits, conversion efficiency, and upstream and downstream water level constraints, etc.

[0103] Second, the resilience evaluation index is deepened. The resilience index used in Example 1 mainly focuses on the loss of load, which is more intuitive, but does not fully reflect the comprehensive performance of the system under disaster. In this example, the second optimization objective of the upper planning model uses a more comprehensive resilience index. This comprehensive resilience index not only considers the total loss of load, but also considers the recovery time of key loads and the connectivity of the power grid topology. For example, higher weights can be assigned to key loads such as hospitals, communication base stations, and fire stations. In the calculation of the resilience index, the outage time and outage amount of these key loads will result in higher "penalties". At the same time, an index can be introduced to measure whether the power grid is split into too many islands after the disaster, in order to encourage configuration schemes that can maintain the integrity and connectivity of the main grid. This new comprehensive resilience index can more comprehensively and finely evaluate the real contribution of a energy storage configuration scheme to improving the resilience of the distribution grid.

[0104] The working process of this example is similar to that of Example 1. First, the extreme scenario data set is obtained by the scenario generation method. Then, the data set is input into the adjusted double-layer optimization model. The upper planning model uses the non-dominated sorting genetic algorithm II to perform multi-objective optimization under the new cost model and comprehensive resilience index. The lower running model performs running simulation considering the new energy storage operation characteristics according to the "battery energy storage + pumped storage" configuration scheme passed by the upper layer, and feeds back the evaluation results (including running cost, loss of load, key load recovery time, network connectivity, etc.) to the upper layer.

[0105] Through the embodiment, a "battery energy storage + pumped storage" optimal configuration scheme for the resource endowment of a specific region can be finally obtained. The scheme not only performs excellently in reducing overall load loss, but also, under the driving of multi-objective optimization, preferentially ensures the rapid recovery of key loads and maintains the stability of the network topology. This fully proves that the technical framework proposed in the application has strong flexibility and scalability, can adapt to different energy storage technology options and diversified resilience requirements, and provides a customized solution for the resilience improvement planning of the distribution network in different scenarios.

[0106] In summary, the method and device provided in the embodiments of the application effectively solve the problem of insufficient resilience of energy storage planning caused by the scarcity of extreme scenario data by organically combining scenario clustering, conditional scenario generation, and double-layer optimization configuration. The method not only can generate a high-fidelity data basis, but also can scientifically balance between economy and resilience, and finally obtain a reliable, effective, and economical hybrid energy storage configuration scheme, which has important theoretical value and practical significance for improving the ability of modern distribution networks to cope with extreme disasters.

[0107] Compared with the prior art, the application has the following beneficial effects: 1. By using a conditional generative adversarial network and other models, a large number of high-fidelity new energy output scenarios under extreme weather can be generated, overcoming the bottleneck of the scarcity of historical data, so that the energy storage planning scheme based thereon can withstand more realistic and more severe scenario tests, and the robustness is significantly enhanced. 2. By establishing a "planning-operation" double-layer optimization model and iteratively solving the investment cost, operation cost, and resilience index in a unified framework, a scientific balance point between the economy of the whole life cycle of the energy storage system and the improvement of the safety and resilience of the power grid can be found, and the problems of excessive investment or insufficient resilience are avoided. 3. Through the optimal configuration of the hybrid energy storage, the distribution network has stronger active support and island operation capability when an extreme disaster occurs, can provide continuous power supply for key loads, significantly reduces the power outage time and economic loss, and effectively improves the overall resilience level of the power grid.

[0108] Please refer to Figure 5 , Figure 5 The structure block diagram of the determination device for a hybrid energy storage configuration scheme in the embodiments of the application is shown in Figure 5 The device comprises: A scenario clustering module 501 is configured to obtain historical output data of new energy, and perform scenario clustering on the historical output data to obtain scenario category labels corresponding to different weather conditions. The scene generation module 502 is configured to train a preset conditional generative adversarial network model based on the scene category label, obtain a trained target model, and input a target scene category label corresponding to an extreme scene into the target model to generate a target output data set of new energy in the extreme scene. The optimization configuration module 503 is configured to establish a double-layer optimization configuration model, take the target output data set as an input of the double-layer optimization configuration model, and obtain an optimal configuration scheme of the hybrid energy storage by iterative solution, wherein the double-layer optimization configuration model comprises an upper-layer planning model and a lower-layer operation model; the upper-layer planning model takes a total cost of the energy storage and a power distribution network resilience index as optimization objectives, solves by a multi-objective optimization algorithm, cooperatively optimizes the comprehensive cost and the power distribution network resilience index, and thus determines the configuration scheme of the energy storage; wherein the total cost comprises a construction cost, a maintenance cost, an operation cost and a comprehensive operation cost of the energy storage, and the power distribution network resilience index is an index determined according to a loss of load rate of the system in the extreme scene; the lower-layer operation model performs multi-scene operation simulation with the operation cost and the comprehensive operation cost of the energy storage as the objectives according to the configuration scheme determined by the upper-layer planning model and the target output data set, and feeds back the obtained operation cost and comprehensive operation cost to the upper-layer planning model.

[0109] It should be noted that, Figure 5 The functions of each module in the device are similar to those of the device Figure 1 The contents of each step in the method are similar to those of the method Figure 1 The contents of each step in the method are similar to those of the method

[0110] The present application provides a kind of determination device of hybrid energy storage configuration scheme, by conditional generative adversarial network model, a large number of, high fidelity new energy output scene under extreme weather can be generated, the bottleneck of the scarcity of historical data is overcome, so that the energy storage planning scheme based on this can withstand more real, more severe scene test, robustness is significantly enhanced;And by establishing double-layer optimization model, and the cost of energy storage and resilience index are placed in unified framework and iteratively solved, a scientific balance point between the economy of energy storage system full life cycle and the improvement of power grid safety resilience can be found, the problem of excessive investment or insufficient resilience is avoided;Finally, by the optimization configuration of hybrid energy storage, power distribution network has stronger active support and island operation capability when extreme disaster arrives, can provide continuous power supply for critical load, significantly reduce outage time and economic loss, effectively improve the overall resilience level of power grid.

[0111] Figure 6 An internal structure diagram of a computer device in an embodiment is shown.The computer device can be a terminal or a server specifically.As shown in Figure 6As shown, the computer device includes a processor, a memory and a network interface connected through a system bus. Among them, the memory includes a non-volatile storage medium and an internal memory. The non-volatile storage medium of the computer device stores an operating system, and can also store a computer program, which, when executed by the processor, can enable the processor to implement the above method. The internal memory can also store a computer program, which, when executed by the processor, can enable the processor to execute the above method. Those skilled in the art can understand that Figure 6 The structure shown in the figure is only a block diagram of part of the structure related to the scheme of the present application, and does not constitute a limitation on the computer device to which the scheme of the present application is applied. The specific computer device can include more or fewer components than those shown in the figure, or combine certain components, or have a different component arrangement.

[0112] In one embodiment, a computer device is provided, including a memory and a processor, the memory storing a computer program, the computer program being executed by the processor to enable the processor to execute the steps of the method as Figure 1 The method.

[0113] In one embodiment, a computer readable storage medium is provided, storing a computer program, the computer program being executed by a processor to enable the processor to execute the steps of the method as Figure 1 The method.

[0114] Those skilled in the art can understand that all or part of the processes in the above-mentioned embodiments can be completed by a computer program instructing related hardware. The program can be stored in a non-volatile computer readable storage medium, and when executed, can include the processes of the above-mentioned embodiments. Any reference to memory, storage, database or other medium used in the embodiments provided by the present application can include non-volatile and / or volatile memory. Non-volatile memory can include read-only memory (ROM), programmable ROM (PROM), electrically programmable ROM (EPROM), electrically erasable programmable ROM (EEPROM) or flash memory. Volatile memory can include random access memory (RAM) or external cache memory. As an illustration but not limitation, RAM is available in various forms, such as static RAM (SRAM), dynamic RAM (DRAM), synchronous DRAM (SDRAM), double data rate SDRAM (DDR SDRAM), enhanced SDRAM (ESDRAM), synchronous link (Synchlink) DRAM (SLDRAM), memory bus (Rambus) direct RAM (RDRAM), direct memory bus dynamic RAM (DRDRAM), and memory bus dynamic RAM (RDRAM), etc.

[0115] The technical features of the above embodiments can be combined in any manner. For the sake of brevity, not all possible combinations of the technical features in the above embodiments are described, however, as long as the combinations of the technical features do not contradict each other, they should be considered to be within the scope of the present disclosure.

[0116] The above embodiments only express several implementation manners of the present application, and the description is relatively specific and detailed, but it should not be understood as a limitation on the patent scope of the present application. It should be pointed out that, for ordinary skilled persons in the art, several modifications and improvements can be made without departing from the concept of the present application, and these all belong to the protection scope of the present application. Therefore, the patent protection scope of the present application should be subject to the appended claims.

Claims

1. A method for determining a hybrid energy storage configuration scheme, characterized in that, The method comprises: acquiring historical output data of a new energy source, and performing scene clustering on the historical output data to obtain scene category labels corresponding to different weather conditions; based on the scene category labels, training a preset conditional generative adversarial network model to obtain a trained target model, and inputting a target scene category label corresponding to an extreme scene into the target model to generate a target output data set of the new energy source under the extreme scene; establishing a double-layer optimization configuration model, taking the target output data set as an input of the double-layer optimization configuration model, and obtaining an optimal configuration scheme of a hybrid energy storage through iterative solution; wherein the double-layer optimization configuration model comprises an upper-layer planning model and a lower-layer operation model; the upper-layer planning model takes a total cost of energy storage and a power grid resilience index as optimization objectives, solves through a multi-objective optimization algorithm, and cooperatively optimizes the comprehensive cost and the power grid resilience index to determine the configuration scheme of the energy storage; wherein the total cost comprises a construction cost, a maintenance cost, an operation cost and a comprehensive operation cost of the energy storage, and the power grid resilience index is an index determined according to a loss of load probability of the system under the extreme scene; the lower-layer operation model performs multi-scene operation simulation with the operation cost and the comprehensive operation cost of the energy storage being the lowest according to the configuration scheme determined by the upper-layer planning model and the target output data set, and feeds back the obtained operation cost and comprehensive operation cost to the upper-layer planning model.

2. The method of claim 1, wherein, The scene clustering on the historical output data to obtain the scene category labels corresponding to different weather conditions comprises: calculating Euclidean distances of all sample points of the historical output data to current cluster center points; according to a nearest neighbor principle, assigning each sample point to a cluster in which a cluster center closest to the sample point is located; recalculating mean values of all sample points in each cluster, and taking the mean values as new cluster center points; judging a change amount between the current cluster center points and the new cluster center points; if the change amount is less than a preset threshold, obtaining a final scene clustering result, the scene clustering result comprising a scene category label corresponding to each cluster; if the change amount is greater than or equal to the preset threshold, returning to execute the step of calculating the Euclidean distances of all sample points of the historical output data to the current cluster center points.

3. The method of claim 1, wherein, The conditional generative adversarial network model is a conditional Wasserstein generative adversarial network supplemented with gradient penalty CWGAN-GP model.

4. The method of claim 1, wherein, The method further comprises: replacing the conditional generative adversarial network model with a conditional variational autoencoder CVAE model to generate the target output data set under the extreme scene.

5. The method of claim 1, wherein, The hybrid energy storage comprises one of the following combinations: a combination of battery energy storage for short-term power support and hydrogen energy storage for long-term energy reserve; or, a combination of battery energy storage and pumped storage.

6. The method of claim 1, wherein, The upper-layer planning model further comprises at least one constraint of a system total carbon emission constraint, a node energy storage expansion capacity constraint, and an annual power balance constraint.

7. The method of claim 1, wherein, The objective function of the lower-level operation model includes the comprehensive operating cost, which includes the start-up and shutdown cost of thermal power units, fuel cost, and environmental cost. The objective function also includes carbon trading revenue. The constraints of the lower-level operation model include the upper and lower limits of thermal power unit output and ramping constraints, as well as the state-of-charge operation constraints of hybrid energy storage.

8. The method of claim 1, wherein, The process of obtaining the optimal configuration scheme through iterative solution includes: The two-layer optimization configuration model is solved using the preset NSGA-II algorithm to obtain a set of Pareto optimal solutions; The final optimal configuration scheme is selected from the Pareto optimal solution set using the entropy weight method-grey target decision method.

9. A device for determining a hybrid energy storage configuration scheme, characterized in that, The device includes: The scene clustering module is used to acquire historical power output data of new energy sources and perform scene clustering on the historical power output data to obtain scene category labels corresponding to different weather conditions. The scene generation module is used to train a pre-set conditional generative adversarial network model based on the scene category labels to obtain a trained target model, and to use the target model to input the target scene category labels corresponding to the extreme scene to generate a target output dataset of new energy under the extreme scene. An optimization configuration module is used to establish a two-layer optimization configuration model. The target output dataset is used as input to the two-layer optimization configuration model, and the optimal configuration scheme for hybrid energy storage is obtained through iterative solution. The two-layer optimization configuration model includes an upper-layer planning model and a lower-layer operation model. The upper-layer planning model uses the total cost of energy storage and the distribution network resilience index as optimization objectives, and solves the problem using a multi-objective optimization algorithm to collaboratively optimize the overall cost and the distribution network resilience index, thereby determining the energy storage configuration scheme. The total cost includes the energy storage's construction cost, maintenance cost, operating cost, and comprehensive operating cost. The distribution network resilience index is an index determined based on the system's load failure rate under extreme scenarios. The lower-layer operation model, based on the configuration scheme determined by the upper-layer planning model and the target output dataset, performs multi-scenario operation simulations with the goal of minimizing the energy storage's operating cost and comprehensive operating cost, and feeds the obtained operating cost and comprehensive operating cost back to the upper-layer planning model. 10.A computer device, comprising a memory and a processor, and characterized in that, The memory stores a computer program that, when executed by the processor, causes the processor to perform the steps of the method as described in any one of claims 1 to 8.

Images