A spatiotemporal joint prediction method and system for rural distributed source and load resources

Abstract

The present application relates to the technical field of power system, and more particularly to a kind of rural distributed source storage load resource space-time joint prediction method and system, by obtaining the multivariate uncertain influence factor set involving source load storage prediction, the key influence factors affecting distributed output are screened out in multivariate uncertain influence factor set, and then the output prediction scene of distributed power output is generated by using generative adversarial network to generate distributed power output, and clustering reduction is carried out, the distributed power output prediction result of distributed power output typical scene is determined;Considering the user behavior state and physical operation boundary of different types of loads, multi-type load prediction is carried out, and the prediction results of different types of loads are obtained, and then the available response power of user-side energy storage in the regulation period is determined in combination with the distributed power output prediction result.The present application realizes the high-precision space-time joint prediction of rural source load storage heterogeneous resources, and provides reliable support for in-situ consumption of distributed energy and flexible dispatching of microgrid.

Description

TECHNICAL FIELD

[0001] The present application relates to the technical field of power systems, in particular to a spatiotemporal joint prediction method and system for rural distributed source and load resources. BACKGROUND

[0002] The prediction of rural distributed resources is the key to realizing the on-site consumption of renewable resources in rural areas and improving the consumption rate of new energy. With the large-scale access of distributed power sources to the power grid in various regions across the country, problems such as voltage out-of-limit, power reverse transmission, and curtailment of wind and light have also attracted widespread attention in the field of power systems. Therefore, a reasonable prediction method plays an important role in the economic operation of the power grid and the improvement of power quality.

[0003] Currently, the prediction methods for distributed resources can be mainly divided into two categories: physical model-based prediction methods and statistical model-based prediction methods. The physical model studies the mechanism of wind power generation or photovoltaic power generation, analyzes factors affecting the output of wind turbines and photovoltaic components, such as terrain, topography, component parameters of power stations, solar radiation data, etc., and establishes physical equations such as solar radiation transfer equations for prediction. The statistical new energy output prediction method mainly uses historical numerical weather prediction (NWP) statistical data and new energy station operation data to establish a linear or nonlinear mapping relationship between the input and output of a certain method, such as autoregressive moving average model (AR, ARMA, ARIMA), Kalman filter, etc. This method is commonly used for short-term or ultra-short-term new energy output prediction.

[0004] However, with the large-scale access of rural distributed photovoltaic, wind power and energy storage, the source-side output and the load-side demand show strong coupling and strong uncertainty characteristics. The existing technology usually predicts wind and light output, load and energy storage state independently, ignoring the dynamic coupling relationship between multi-source data, resulting in inconsistency of the prediction results at the system level, which is difficult to meet the scheduling demand of high proportion of distributed energy. SUMMARY

[0005] In order to solve the above technical problems, the purpose of the present application is to provide a spatiotemporal joint prediction method and system for rural distributed source and load resources, and the technical scheme adopted is as follows: The present application provides a spatiotemporal joint prediction method for rural distributed source and load resources, comprising the following steps: Obtain a set of multiple uncertain influencing factors related to source, load and storage prediction, and select key influencing factors affecting distributed output from the set of multiple uncertain influencing factors; Based on the screened key influencing factors, a generative adversarial network is used to generate a distributed power output prediction scene, the distributed power output prediction scene is clustered and reduced, and a distributed power output prediction result of a typical scene of the distributed power output is determined; Considering the user behavior state and physical operation boundary of different types of loads, multi-type load prediction is performed to obtain a prediction result of different types of loads; Based on the prediction result of different types of loads and the distributed power output prediction result, the available response power of the user-side energy storage in the regulation period is determined.

[0006] Further, the key influencing factors affecting the distributed output are screened out from the multiple uncertain influencing factors, including: Based on the influence relationship between different influencing factors in the multiple uncertain influencing factors, an adjacency matrix for representing the binary relationship between different influencing factors is constructed; Based on the adjacency matrix, a reachable matrix is determined, and based on the reachable matrix, a reachable set and a precedence set of each influencing factor in the multiple uncertain influencing factors are determined; Based on the reachable set and the precedence set, and in combination with the reachable matrix, the key influencing factors affecting the distributed output are screened out from the multiple uncertain influencing factors.

[0007] Further, the generative adversarial network includes a generative model G and a discriminative model D; the generative adversarial network is used to generate a distributed power output prediction scene, including: A generative adversarial network is constructed, which includes a generative model G and a discriminative model D; The screened key influencing factors and noise data are jointly input into the generative model G to generate distributed power output scene data; The obtained historical real data of the distributed power output and the distributed power output scene data generated by the generative model G are respectively input into the discriminative model D, and a scoring value of each data in the real distribution is output; Based on the scoring value, a value function is established to make the generative model G and the discriminative model D perform an adversarial game, and the loss function of the generative model G and the loss function of the discriminative model D are determined according to the value function; The generative adversarial network is trained based on the loss function of the generative model G and the loss function of the discriminative model D, and the network weight parameters of the discriminative model D are pruned to be limited within a preset constant range during the training process until the network converges, and a trained generative model G is obtained; The trained generative model G is used to generate a distributed power output scene as an output prediction scene.

[0008] Further, in the process of training the generative adversarial network based on the loss function of the generative model G and the loss function of the discriminative model D, the Wasserstein distance is taken as an optimization target for measuring the distance between the historical real data of the distributed power output and the generated output prediction scenario, and the corresponding calculation formula is: In the formula: represents the historical real data of the distributed power output and the generated distributed power output scenario data Wasserstein distance between them; represents the historical real data of the distributed power output; represents the probability distribution of the historical real data of the distributed power output; represents the noise data; represents the probability distribution of the generated output prediction scenario; represents the mapping function constructed by the discriminative model D; represents the constraint condition, and represents the mapping function must satisfy the Q-Lipschitz continuity condition, represents the Lipschitz constant; represents the upper limit operation; represents the function of taking mathematical expectation.

[0009] Further, the output prediction scenarios are clustered and reduced to determine the distributed power output prediction result of the typical distributed power output scenario, including: Based on the DTW distance between any two output prediction scenarios, the metric distance between any two output prediction scenarios is determined; Based on the metric distance between any two output prediction scenarios, the K-medoids clustering algorithm is used to cluster the output prediction scenarios to obtain multiple clustering clusters and clustering centers, and the clustering centers of all clustering clusters are determined as the distributed power output prediction result of the typical distributed power output scenario.

[0010] Further, the output prediction scenarios are clustered and reduced to determine the distributed power output prediction result of the typical distributed power output scenario, and further including: Based on the actual output power and the output prediction value of the distributed power output, a spatio-temporal disturbance factor is determined; Based on the spatio-temporal disturbance factor, the distributed power output prediction result of the typical distributed power output scenario is corrected to obtain the final distributed power output prediction result of the typical distributed power output scenario.

[0011] Furthermore, the multi-type loads include at least industrial loads; for industrial loads, load forecasting is performed, and the calculation formula corresponding to the load forecast results is as follows: In the formula: Indicates industrial load During the adjustment period Actual power consumption; Indicates industrial load During the adjustment period The power consumption of the original production plan; This indicates that the industrial load i is during the adjustment period. The adjustable status of the production plan is represented by a 0-1 variable, where 0 indicates that it is not adjustable and 1 indicates that it is adjustable. Indicates industrial load During the adjustment period The amount of power adjustment; Indicates industrial load During the adjustment period Available response power; Indicates industrial load The baseline power consumption; Indicates the period during which the power grid command requires an increase in load; This indicates the period during which the power grid requires a load reduction.

[0012] Furthermore, the multi-type load includes at least electric vehicle load; for electric vehicle load, load forecasting is performed, and the calculation formula corresponding to the load forecast result is as follows: In the formula: and These represent the adjustment periods for electric vehicles. and The remaining battery power; and These respectively represent the adjustment period. The 0-1 variables that control the charging and discharging states; and These represent the charging efficiency and discharging efficiency of an electric vehicle, respectively. and respectively represent the rated charging power and the rated discharging power of the electric vehicle; represent the grid-connected state of the electric vehicle in the regulation period ; and respectively represent the grid-connected time and the off-grid time; represent the available response power of the electric vehicle load in the regulation period ; represent the reference charging power of the electric vehicle in the period ; represent the charging power of the electric vehicle in the regulation period ; represent the load up-regulation period required by the grid instruction; represent the load down-regulation period required by the grid instruction.

[0013] Further, based on the prediction results of different types of loads and the distributed power output prediction results, the available response power of the user-side energy storage in the regulation period is determined, including: According to the distributed power output prediction results and the prediction results of different types of loads, the source-load net power deviation amount is calculated; Based on the source-load net power deviation amount and the source-load transmission coefficient, and combined with the charging and discharging power, the charging and discharging state and the charging and discharging efficiency of the user-side energy storage, the state of charge of the user-side energy storage is dynamically updated; Based on the actual operating power and the reference operating power of the user-side energy storage, the available response power of the user-side energy storage in the regulation period is calculated.

[0014] In a second aspect, the present application further provides a rural distributed source-storage-load resource space-time joint prediction system, comprising a memory and a processor. The memory is used to store executable computer program code, and the processor is used to call and run the executable computer program code from the memory, so that the system executes the method in any one of the possible implementation manners described above.

[0015] The application has the following beneficial effects: firstly, the application effectively eliminates redundant interference variables and reduces the model training burden by screening key influence factors in a set of multiple uncertain influence factors, thereby significantly improving the generalization ability and solving efficiency of the subsequent prediction model; secondly, the generated output prediction scenarios are clustered and reduced based on the screened key influence factors, which not only accurately depict the strong randomness and fluctuation distribution characteristics of distributed power output under complex weather conditions, but also avoid the curse of dimensionality by dimensionality reduction to extract typical scenarios, thereby balancing the high accuracy and operation timeliness of output prediction; further, the physical operation boundary is considered for multi-type load prediction, which accurately restores the real adjustment potential and response boundary of elastic loads such as industry, electric vehicles and air conditioners; finally, the available response power of the energy storage is determined based on the prediction results of the load and the power, and the source-load prediction deviation is coupled as a physical disturbance to the state evolution constraint of the energy storage, thereby breaking the limitations of traditional independent prediction and realizing the accurate joint deduction of source-storage-load space-time dynamic interaction, which provides a reliable decision basis for flexible scheduling and local consumption of microgrids under high proportion of renewable energy access. BRIEF DESCRIPTION OF DRAWINGS

[0016] In order to clearly illustrate the technical features of the scheme, the embodiments of the application include the following drawings: Figure 1 A step flow chart of a space-time joint prediction method of rural distributed source storage load resources according to an embodiment of the application. DETAILED DESCRIPTION

[0017] The space-time joint prediction method and system of rural distributed source storage load resources provided by the embodiments of the application will be described in detail below with reference to the drawings.

[0018] Figure 1 A basic flow diagram of a space-time joint prediction method of rural distributed source storage load resources according to an embodiment of the application is shown, as shown in Figure 1 The method specifically includes the following steps: Step S100: acquiring a set of multiple uncertain influence factors related to source load storage prediction, and screening key influence factors affecting distributed output from the set of multiple uncertain influence factors.

[0019] The selection of multiple uncertain influence factors is an important part of the microgrid source load prediction model, and screening appropriate key influence factor variables can effectively improve the accuracy of the prediction model. A set of multiple uncertain influence factors affecting rural distributed resources is acquired, which is a set of all possible factors that may affect source load power.

[0020] In a specific example, all influence factors that can affect the source load power are determined by literature review and expert interviews, and these influence factors constitute a set of multiple uncertain influence factors wherein, represents the total number of influence factors in the set of multiple uncertain influence factors, represents the i-th influence factor in the set of multiple uncertain influence factors.

[0021] In order to further analyze the influence of exogenous factors on the source load power, the input features should contain sufficient influence factors, but too many features will increase the training burden and affect the training efficiency. Therefore, the key influence factors are selected and the redundant features are removed from the set of multiple uncertain influence factors, so as to build a more concise and good generalization ability prediction model.

[0022] Further, the key influence factors are selected from the set of multiple uncertain influence factors in step S100, including: Step S101: based on the influence relationship between different influence factors in the set of multiple uncertain influence factors, an adjacency matrix is constructed to represent the binary relationship between different influence factors.

[0023] According to the influence relationship between different influence factors in the set of multiple uncertain influence factors, the binary influence relationship between any two influence factors is determined. Among them, the multiple independent expert evaluation method is used to judge whether there is an influence relationship between different influence factors. Specifically, for any i-th influence factor and j-th influence factor, multiple experts judge whether the i-th influence factor has an influence on the j-th influence factor. The evaluation result of the i-th expert is represented as: Based on the evaluation results of multiple experts, the expert consensus degree of the i-th influence factor on the j-th influence factor is calculated: In the formula: represents the expert consensus degree of the i-th influence factor on the j-th influence factor; represents the number of experts participating in the evaluation. The expert consensus degree is compared with a preset value to determine the binary influence relationship between the two influence factors:

[0024] The expert consensus degree is compared with a preset value to determine the binary influence relationship between the two influence factors: ​​​​​​​​​ In the formula: represents the binary influence relationship between two influence factors; represents a preset value, and the preset value ranges from 0 to 1, for example, the preset value is set to 0.5.

[0025] Based on the binary influence relationship between the first influence factor and the second influence factor , an adjacency matrix is constructed to describe the direct influence relationship between different influence factors. The number of rows and columns of the adjacency matrix is equal to the total number of influence factors in the multivariate uncertain influence factor set. The element at the position of the i-th row and the j-th column in the adjacency matrix is the j-th influence factor , which represents the binary influence relationship between the i-th influence factor and the j-th influence factor . . .

[0026] Step S102: Based on the adjacency matrix, a reachable matrix is determined, and based on the reachable matrix, a reachable set and a predecessor set of each influence factor in the multivariate uncertain influence factor set are determined.

[0027] The reachable matrix refers to a matrix form used to represent the degree to which each influence factor can be reached by a directed link, which can reflect the direct and indirect influence relationship of each influence factor. The reachable set of an influence factor is a set of influence factors affected by the influence factor, and the predecessor set refers to a set of influence factors that have an influence on the influence factor.

[0028] In a specific example, an initial matrix is set as , which represents an adjacency matrix, , and a unit matrix, that is, a square matrix with all elements on the main diagonal being 1 and all other elements being 0, has the same order as the adjacency matrix . Using Boolean algebra operation rules, the continuous power of is calculated, that is, , , , are calculated respectively. After each calculation, the results of the adjacent two times are compared. When there is a positive integer such that = , the iteration calculation is stopped. The corresponding calculation formula is: ​​​ In a specific example, based on the reachability matrix, determining the reachability set and antecedent set of each influencing factor in the set of multivariate uncertain influencing factors includes: based on the reachability matrix For any number of Factors Extract the reachability matrix The Middle The column elements corresponding to behavior 1 constitute the influencing factors. reachable set ,express All influencing factors, whether directly or indirectly affected; extract the reachability matrix. The Middle The row element corresponding to column 1 constitutes the influencing factor. the antecedent set This indicates that it can directly or indirectly affect All influencing factors.

[0029] Step S103: Based on the reachable set and the prior set, and in conjunction with the reachable matrix, key influencing factors are selected from the set of multivariate uncertain influencing factors.

[0030] In a specific example, the current matrix is ​​set as a reachable matrix. Extract any nth element from the current matrix Factors Find the reachable set and the antecedent set, and perform an intersection check, that is, check if they exist. ,if ,illustrate The highest-level influencing factor in the current matrix is ​​identified. This highest-level influencing factor is then assigned to the current level, and the corresponding rows and columns are removed from the current matrix, resulting in an updated submatrix that serves as the current matrix for the next round. Based on the updated current matrix, the reachable set and preceding set are re-extracted for the remaining influencing factors, and the intersection judgment step described above is repeated to sequentially divide the system into the second and third levels, until all influencing factors are divided, resulting in a multi-level hierarchical structure. The preceding set of influencing factors located in the first level (i.e., at the driving source end) is extracted as key influencing factors and used as feature input for subsequent prediction models to eliminate redundant derived features from intermediate layers.

[0031] Step S200: Based on the selected key influencing factors, a generative adversarial network is used to generate output prediction scenarios for distributed power sources. The output prediction scenarios are then clustered and reduced to determine the output prediction results for typical distributed power source output scenarios.

[0032] In this process, the key influencing factors selected in step S100 serve as input features to the generative model G, characterizing the meteorological, environmental, and operational conditions corresponding to the distributed power generation output. These key influencing factors affect the generated distributed power generation scenario through nonlinear mappings within the generative model G, rather than being directly used as input to the discriminant model D.

[0033] Output prediction scenarios refer to the various time-series curves of power generation that distributed generation sources may exhibit under specific meteorological and environmental conditions. These scenarios are used to intuitively quantify the randomness and fluctuation characteristics of renewable energy output. Distributed generation output is subject to complex nonlinear influences from multiple exogenous factors. Generative adversarial networks (GANs) driven by selected key influencing factors can eliminate redundant interference and utilize the network's powerful data distribution fitting capabilities to generate a large number of output scenarios encompassing various potential fluctuations, thus fully characterizing the strong uncertainty of renewable energy.

[0034] However, when using generative adversarial networks to generate a large number of distributed power output prediction scenarios, many highly similar scenarios are generated. Furthermore, considering numerous distributed power output scenarios in the scheduling model significantly increases the difficulty of solving the scheduling problem. Therefore, to reduce the difficulty of solving the scheduling model and to accurately reflect the typical characteristics of distributed power output through more concise scenarios, a scenario reduction method is used to cluster and reduce the generated distributed power output scenarios, extracting typical distributed power output scenarios.

[0035] Furthermore, the generative adversarial network (GAN) includes a generative model G and a discriminative model D; the output prediction scenarios for distributed power sources using the GAN include: Step S201: Construct a generative adversarial network, which includes a generative model G and a discriminative model D.

[0036] A multi-layer neural network architecture is constructed to obtain a generative model G and a discriminative model D, and then a Wasserstein generative adversarial network composed of the generative model G and the discriminative model D is constructed. The generative model G adopts a multilayer perceptron, convolutional neural network, or other neural network structure capable of processing time-series data. The input layer of the generative model G is used to receive a fusion vector formed by concatenating random noise data and key influencing factor data selected in step S100. The fusion vector undergoes nonlinear mapping and feature expansion through multiple hidden layers within the generative model G, and generates a power output scene sequence with the same time-series length as the historical real power output data of the distributed power source at the output layer. The discriminative model D adopts a deep feature extraction and regression network structure. During training, the historical real power output data of the distributed power source and the power output scene data generated by the generative model G are respectively input into the same discriminative model D, which extracts the time-series features of the input data and outputs the corresponding real-valued scores. Unlike the binary classifiers in traditional generative adversarial networks (GANs), the output of the discriminant model D does not have a sigmoid activation function. Its output does not represent the probability that the input data is real data, but rather a real-valued score evaluating how closely the input data approximates the real data distribution. Since the process of constructing the Wasserstein GAN is existing technology, it will not be described in detail here.

[0037] Step S202: Input the selected key influencing factors and noise data into the generation model G to generate distributed power output scenario data.

[0038] The process involves acquiring sequence data of key influencing factors selected from historical time periods and normalizing them (e.g., Z-score standardization) to obtain conditional feature vectors. Simultaneously, random noise is sampled from a standard normal distribution to generate random noise vectors to increase data diversity. The conditional feature vectors and random noise vectors are concatenated along their feature dimensions to construct a joint feature vector. This joint feature vector is then input into the input layer of the generative model G, undergoing nonlinear mapping and feature expansion through multiple hidden layers within G. The output layer of the generative model G outputs a continuous numerical sequence with the same time series length as the actual power output data, serving as the generated distributed power generation scenario data.

[0039] Step S203: Input the historical real data of distributed power output and the distributed power output scenario data generated by the generation model G into the discrimination model D respectively, and output the score of each data in the real distribution.

[0040] For the discriminant model D, its inputs are real data of distributed power generation output and distributed power generation scenario data generated by the generative model. Its output is the probability that the input data is real data, and its value range is... Historical data on distributed power output can be obtained from the microgrid's energy management system (EMS) or from smart measurement terminals installed at photovoltaic and wind power grid connection points.

[0041] Step S204: Based on the scoring values, establish a value function that enables the generator model G and the discriminator model D to engage in adversarial game, and determine the loss function of the generator model G and the loss function of the discriminator model D respectively according to the value function.

[0042] The value function for enabling adversarial game between the generative model G and the discriminative model D is established, and the corresponding calculation formula is as follows: In the formula: Let represent the value function that enables the generative model G and the discriminative model D to engage in an adversarial game. During training, the network parameters of the generative model G are adjusted to minimize this value function, and the network parameters of the discriminative model D are adjusted to maximize this value function; z represents the noise data input to the generative model G. This represents the probability distribution that the noise data z follows, and is generally set to a normal distribution; A function representing the mathematical expectation; Data on the output scenarios of the generated distributed power sources; To determine the model function; This represents the actual output data of distributed power sources; This represents the probability distribution that the actual output data of distributed power sources follows. This represents the actual data of the distributed power generation output from the discriminant model D. Output score.

[0043] For the generative model G, its loss function The corresponding calculation formula is: For the discriminant model D, its loss function The corresponding calculation formula is: Step S205: Train the Wasserstein generative adversarial network based on the loss function of the generative model G and the loss function of the discriminative model D. During the training process, the network weight parameters of the discriminative model D are pruned to limit them to a preset constant range until the network converges, and the trained generative model G is obtained.

[0044] Loss function based on generative model G and the loss function of the discriminant model D The Wasserstein generative adversarial network is trained iteratively, and its weights are pruned to a preset constant range to satisfy the Lipschitz constraint after each update of the parameters of the discriminant model D. The training is repeated until the network converges, and then the trained generative model G is retained.

[0045] Furthermore, during the training of the generative adversarial network based on the loss functions of the generative model G and the discriminative model D, the Wasserstein distance is used as the optimization objective to measure the distance between the historical real data of the distributed power source output and the generated output prediction scenario. The corresponding calculation formula is as follows: In the formula: Represents historical real data of distributed power generation output. With the generated distributed power output scenario data Wasserstein distance between them; This represents the historical real data of distributed power generation output; The probability distribution representing the historical real data of distributed power generation output; Represents noise data; G represents the probability distribution that the noise data z follows; G(z) represents the distributed power output scenario generated by the generative model G based on the noise data. This represents the mapping function constructed by the discriminant model D; Represents the constraint conditions, and represents the mapping function. The Q-Lipschitz continuity condition must be satisfied. Represents the Lipschitz constant; This indicates the supremum operation, which in deep learning represents "maximization," that is, adjusting the parameters of the discriminant model to find the maximum difference. It represents a function that calculates the mathematical expectation.

[0046] Step S206: Use the trained generative model G to generate a distributed power output scenario as the power output prediction scenario.

[0047] Data on key influencing factors selected for the period to be predicted are acquired and normalized in the same way as during the training phase. Multiple independent samples are taken from a predetermined probability distribution (e.g., a standard normal distribution) to obtain multiple sets of random noise data sequences. The normalized key influencing factor data for the period to be predicted are concatenated with the multiple sets of random noise data sequences to construct multiple joint feature vectors. These joint feature vectors are then input into a pre-trained generative model G with its network parameters frozen. Forward propagation is performed through nonlinear mapping within the generative model G, outputting multiple continuous numerical sequences corresponding to the time series of the period to be predicted. These output continuous numerical sequences are then de-normalized and used as the output prediction scenarios for multiple distributed power sources.

[0048] Furthermore, the output prediction scenarios are clustered and reduced to determine the distributed power generation output prediction results for typical distributed power generation scenarios, including: Step S207: Determine the metric distance between any two output prediction scenarios based on the DTW distance between any two output prediction scenarios.

[0049] In a specific example, for any i-th power prediction scenario and j-th power prediction scenario, the formula for calculating the corresponding metric distance is: In the formula: Represents any number of The metric distance between the j-th power output prediction scenario and the j-th power output prediction scenario; Represents any number of The DTW distance between the time series corresponding to the j-th power output prediction scenario and the j-th power output prediction scenario.

[0050] Step S208: Based on the metric distance between any two power output prediction scenarios, the K-medoids clustering algorithm is used to cluster the power output prediction scenarios to obtain multiple clusters and their cluster centers, and the cluster centers of all clusters are determined as typical power output prediction scenarios.

[0051] In a specific example, the K-medoids algorithm is used to cluster and reduce the power output prediction scenarios, namely: 1) Randomly select K scenarios from the generated power output prediction scenario set as initial cluster centers. 2) Calculate the metric distance between each remaining power output prediction scenario and each cluster center, and assign each remaining power output prediction scenario to the cluster containing the cluster center with the smallest distance. 3) Update the cluster centers. The calculation method for the new cluster centers of the same cluster is as follows: In the formula: This represents the j-th cluster; This represents the candidate center points to be traversed and filtered. This indicates traversing all Then, the optimal candidate point that minimizes the total distance (the final center of the Medoid cluster). Represents clusters For any ordinary sample point, iterate through all samples to accumulate and measure the distance; This represents the metric distance between power prediction scenarios.

[0052] 4) Repeat steps 2) and 3) until the cluster centers no longer change, and output the final K cluster centers as typical output prediction scenarios.

[0053] Furthermore, clustering and reducing the output prediction scenarios to determine the distributed power output prediction results for typical distributed power output scenarios also includes: determining a spatiotemporal disturbance factor based on the actual output power and output prediction value of the distributed power source; and correcting the distributed power output prediction results for typical distributed power output scenarios based on the spatiotemporal disturbance factor to obtain the final distributed power output prediction results for typical distributed power output scenarios.

[0054] In a specific example, a spatiotemporal disturbance factor is introduced to correct for disturbances in the output of distributed power sources at different time and spatial scales. The calculation formula is as follows: ,t≥2 In the formula: Indicates the adjustment period The spatiotemporal perturbation factor; Indicates the period in the previous adjustment period The actual output power of distributed power sources; Indicates the period in the previous adjustment period Forecast output value of distributed power generation; This indicates the rated installed capacity of distributed power generation.

[0055] Based on the spatiotemporal disturbance factor, the output prediction results of distributed power sources in typical scenarios are corrected. The calculation formula is as follows: In the formula: Indicates the adjustment period Corrected output prediction results for distributed generation; Indicates the adjustment period Distributed power output prediction results for typical distributed power generation scenarios.

[0056] Step S300: Consider the user behavior status and physical operation boundary of different types of loads, perform multi-type load prediction, and obtain prediction results for different types of loads.

[0057] By considering the user behavior of different types of loads, the elastic operating space of various resources is characterized to characterize the response capability of multiple types of flexible loads, thereby obtaining the prediction results of different types of loads.

[0058] Furthermore, the multi-type loads include industrial loads. The electricity consumption behavior of industrial users is affected by the adjustability and range of production plans.

[0059] For industrial loads, load forecasting is performed, and the calculation formula corresponding to the forecasted load (i.e., the power consumption of industrial users) is as follows: In the formula: Indicates industrial load During the adjustment period Actual power consumption; Indicates industrial load During the adjustment period The power consumption of the original production plan; Indicates industrial load During the adjustment period The adjustable status of the production plan is a 0-1 variable, where 0 represents non-adjustable and 1 represents adjustable. Indicates industrial load During the adjustment period The power consumption adjustment amount is indicated by a positive value for upward adjustment and a negative value for downward adjustment. This is the adjustable state. The adjustable status is determined based on the production plan and adjustable equipment information constraints pre-declared by the industrial user. Specifically, when the corresponding production task or electrical equipment is allowed to change its operating power or adjust its operating time during the adjustment period, and the adjustment will not violate the relevant constraints, the adjustable status of that adjustment period is set to 1; otherwise, it is set to 0.

[0060] Industrial users will rationally allocate their actual production power consumption based on the baseline power consumption released by the dispatch center, taking into account factors such as electricity price, incentive price, and the adjustability and range of production plans. Under these conditions, the available response power of the industrial load during a certain adjustment period can be calculated using the following formula: In the formula: Indicates industrial load During the adjustment period Available response power; Indicates industrial load During the adjustment period The actual power adjustment amount used relative to the original production plan's power consumption is used to calculate the industrial load power consumption after the production plan is adjusted. Indicates industrial load The baseline power consumption; Indicates the period during which the power grid command requires an increase in load; This indicates the period during which the power grid requires a load reduction.

[0061] Furthermore, the multi-type load also includes electric vehicle loads. The electricity consumption behavior of electric vehicles is influenced by factors such as electricity prices, incentive prices, and travel demand. Their flexible operating range is typically determined by the electric vehicle's grid-connected status, remaining electricity capacity when connected to the grid, and electricity demand when disconnected from the grid. In the grid-connected state, the electric vehicle's status is generally divided into three categories: charging, discharging, and waiting. The formulas for calculating its remaining electricity capacity and charging / discharging power are as follows: In the formula: and They represent electric vehicles. During the adjustment period t+1 and The remaining battery power; and These respectively represent the adjustment period. The 0-1 variables that control the charging and discharging states; and They represent electric vehicles. The charging efficiency and discharging efficiency; and They represent electric vehicles. Rated charging power and rated discharging power; Indicates electric vehicles During the adjustment period Grid connection status; and These represent the grid connection time and the grid disconnection time, respectively. Indicates electric vehicles During the adjustment period Available response power; Indicates electric vehicles during the adjustment period The reference charging power; Indicates electric vehicles During the adjustment period The charging power; Indicates the period during which the power grid command requires an increase in load; This indicates the period during which the load is reduced as required by the power grid. Specifically, when the electric vehicle is connected to the grid and needs to replenish its power or execute a charging command, and the battery's maximum allowable capacity will not be exceeded after charging, the charging status is set to 1 and the discharging status is set to 0. When the electric vehicle is connected to the grid, the user allows the electric vehicle to discharge to the grid, and the current remaining power meets the discharge lower limit and the power requirement for disconnection, and a discharging command is executed, the charging status is set to 0 and the discharging status is set to 1. In all other cases, both the charging and discharging statuses are set to 0, indicating that the electric vehicle is in a waiting or disconnected state.

[0062] The above formula only represents the change in the electric vehicle's power consumption when it is connected to the grid. In order to control the electric vehicle's grid-connected and off-grid states, it is also necessary to introduce 0-1 variables. The constraint calculation formula that it needs to satisfy is: In the formula: Indicates electric vehicles During the adjustment period The grid connection status is a 0-1 variable. A value of 1 indicates that the electric vehicle is connected to the grid, and it can be in any of the three states: charging, discharging, or waiting. A value of 0 indicates that the electric vehicle is disconnected from the grid and cannot be charged or discharged. The various states are represented as follows: In addition, considering that electric vehicles only make one trip per day, the grid connection time and grid disconnection time are respectively... and ,and ,but The constraint calculation formula must also be satisfied: Calculate the electric vehicle during a certain adjustment period The available response power is calculated using the following formula: In the formula: Indicates electric vehicles During the adjustment period Available response power; This indicates that the dispatching agency has released information about electric vehicles. The reference charging power on the execution date; Indicates electric vehicles During the adjustment period The charging power.

[0063] Furthermore, the multi-type load also includes air conditioning load. Air conditioning load is a temperature-controlled load, and its flexible operating range is typically determined by the equipment's start / stop status, the user-set temperature, and the outdoor temperature. The available response power during a certain adjustment period is calculated using the following formula: in: Indicates air conditioner During the adjustment period Available response power; Indicates the air conditioning issued by the dispatching agency The baseline power consumption on the execution date; Indicates air conditioner During the adjustment period The power consumption.

[0064] Step S400: Based on the forecast results of different types of loads and the forecast results of distributed power generation output, determine the available response power of user-side energy storage during the adjustment period.

[0065] Available response power refers to the flexible adjustable power space that energy storage devices can provide (or absorb) to the grid, provided they meet their own physical constraints (such as upper and lower limits of SOC, charge and discharge limits) and the user's baseline operating plan. In microgrid systems, the net power deviation between the predicted output of distributed generation sources and the predicted demand of various types of loads is directly transmitted as a physical disturbance and affects the real-time state of charge (SOC) of energy storage. By coupling the source and load prediction results for calculation, the state transitions that energy storage undergoes to smooth system fluctuations can be extrapolated forward. Then, based on the difference between the corrected physical operating boundary of energy storage and its original baseline plan, the actual upward or downward response capacity of energy storage during a specific adjustment period can be accurately defined.

[0066] Furthermore, based on the forecast results of different types of loads and the output forecast results of distributed power sources, the available response power of user-side energy storage during the adjustment period is determined, including: Step S401: Calculate the net power deviation between the source and load based on the output prediction results of the distributed power source and the prediction results of different types of loads.

[0067] To reflect the dynamic impact of distributed power generation output and load demand on the operating status of energy storage, a dynamic correction model for energy storage status is constructed by introducing the net power deviation between the source and load based on the power generation forecast results and the load forecast results, thereby realizing information transmission and collaborative forecasting among the source, load, and storage.

[0068] Based on the obtained distributed generation output forecast results and load forecast results, the net power deviation between source and load is constructed, and the calculation formula is as follows: In the formula: Indicates the adjustment period The net power deviation of the source load; Indicates the adjustment period Total predicted load power; Indicates the adjustment period Corrected power supply predicted output.

[0069] Step S402: Based on the net power deviation of the source and load and the source and load conduction coefficient, and in combination with the charging and discharging power, charging and discharging state and charging and discharging efficiency of the user-side energy storage, dynamically update the state of charge of the user-side energy storage.

[0070] Energy storage operates in a relatively flexible manner, and its charging and discharging power, as well as the specific formulas for the state of charge considering the net power deviation between the source and load, are as follows: in: Indicates user-side energy storage During the adjustment period The operating power is positive for charging and negative for discharging; and These represent the control of user-side energy storage. During the adjustment period 0-1 variables for charging and discharging states; and These represent user-side energy storage. Rated charging power and rated discharging power; and These represent energy storage during the adjustment period. and The state of charge; Indicates the source charge conduction coefficient; and These represent user-side energy storage. The charging efficiency and discharging efficiency; Indicates the scheduling time interval; This indicates the rated capacity of the energy storage.

[0071] Step S403: Based on the actual operating power and the baseline operating power of the user-side energy storage, calculate the available response power of the user-side energy storage during the adjustment period.

[0072] In actual operation, users will rationally arrange their actual charging and discharging plans based on their baseline power consumption, taking into account factors such as electricity prices, incentive prices, and original usage needs. Under these conditions, the available response power of user-side energy storage during a certain adjustment period can be calculated using the following specific formula: In the formula: Indicates user-side energy storage During the adjustment period Available response power; This indicates the user-side energy storage released by the dispatching agency. The baseline operating power; Indicates user-side energy storage During the adjustment period The operating power is positive for charging and negative for discharging; Indicates the period during which the power grid command requires an increase in load; This indicates the period during which the power grid requires a load reduction.

[0073] Based on the above technical solutions, this paper first proposes a multivariate uncertainty influencing factors related to source-load-storage prediction. By using an interpretive structural model to screen these factors, the problem of heavy training burden due to redundancy influencing factors is solved. Then, a distributed power generation scenario generation and clustering analysis method based on an improved generative adversarial network and K-medoids clustering algorithm is adopted to predict the output of distributed power sources. Next, the user behavior states of different types of loads are considered to characterize the response capabilities of various types of flexible loads. Finally, based on the power and load output prediction, the response power of user-side energy storage during a certain adjustment period is calculated, making the output of various types of distributed energy sources clear and helping to facilitate the smooth grid connection of large-scale distributed resources.

[0074] Based on the same inventive concept, this invention also provides a spatiotemporal joint prediction system for rural distributed source-storage-load resources. The system includes: a memory, a processor, and computer program code stored in the memory and running on the processor. When the processor executes the computer program code, the system can execute any of the spatiotemporal joint prediction methods for rural distributed source-storage-load resources described above.

[0075] The aforementioned system is a software system. In this embodiment of the invention, the system can be divided into functional modules according to the method example described above. For example, each module can correspond to a separate function, or two or more functions can be integrated into a single processing module. The integrated module can be implemented in hardware. It should be noted that the module division in this embodiment is illustrative and represents only one logical functional division; in actual implementation, other division methods may be used.

[0076] It should be noted that the above embodiments are only used to illustrate the technical concept and technical solution of the present invention, and are not intended to limit the scope of protection of the present invention. Any modifications or partial improvements made to the technical solution of the present invention by those skilled in the art without departing from the inventive concept of the present invention, or equivalent substitutions or transformations of some technical features, shall fall within the scope of protection defined by the present invention.

Claims

1. A spatiotemporal joint prediction method for rural distributed source-storage-load resources, characterized in that, Includes the following steps: Obtain a set of multiple uncertain influencing factors related to source-load-storage forecasting, and screen out the key influencing factors affecting distributed power output from the set of multiple uncertain influencing factors; Based on the selected key influencing factors, a generative adversarial network is used to generate output prediction scenarios for distributed power sources. These output prediction scenarios are then clustered and reduced to determine the output prediction results for typical distributed power source output scenarios. Considering the user behavior status and physical operating boundaries of different types of loads, multi-type load forecasting is performed to obtain forecast results for different types of loads; Based on the forecast results of different types of loads and the output forecast results of distributed power sources, the available response power of user-side energy storage during the adjustment period is determined.

2. The spatiotemporal joint prediction method for rural distributed source-storage-load resources according to claim 1, characterized in that, Key influencing factors affecting distributed power output were identified from the set of multiple uncertain influencing factors, including: Based on the influence relationships between different influencing factors in the set of multivariate uncertain influencing factors, an adjacency matrix is ​​constructed to characterize the binary relationships between different influencing factors; Based on the adjacency matrix, the reachability matrix is ​​determined, and based on the reachability matrix, the reachability set and prior set of each influencing factor in the multivariate uncertain influencing factor set are determined; Based on the reachable set and the prior set, and in conjunction with the reachability matrix, key influencing factors affecting distributed power output are screened from the set of multivariate uncertain influencing factors.

3. The spatiotemporal joint prediction method for rural distributed source-storage-load resources according to claim 1, characterized in that, The generative adversarial network includes a generative model G and a discriminative model D; the generation of a distributed power source output prediction scenario using the generative adversarial network includes: Construct a generative adversarial network, which includes a generative model G and a discriminative model D; The selected key influencing factors and noise data are jointly input into the generation model G to generate distributed power output scenario data. The historical real data of distributed power output and the distributed power output scenario data generated by the generation model G are respectively input into the discrimination model D, and the score of each data in the real distribution is output. Based on the scoring values, a value function is established to enable the generative model G and the discriminative model D to engage in adversarial game, and the loss function of the generative model G and the loss function of the discriminative model D are determined according to the value function. Generative adversarial network is trained based on the loss function of generative model G and the loss function of discriminative model D. During the training process, the network weight parameters of discriminative model D are pruned to limit them to a preset constant range until the network converges, and the trained generative model G is obtained. The trained generative model G is used to generate distributed power output scenarios as power output prediction scenarios.

4. The spatiotemporal joint prediction method for rural distributed source-storage-load resources according to claim 3, characterized in that, In training the generative adversarial network based on the loss functions of the generative model G and the discriminative model D, the Wasserstein distance is used as the optimization objective to measure the distance between the historical real data of distributed power output and the generated power output prediction scenario. The corresponding calculation formula is as follows: In the formula: Represents historical real data of distributed power generation output. With the generated distributed power output scenario data Wasserstein distance between them; This represents the historical real data of distributed power generation output; The probability distribution representing the historical real data of distributed power generation output; Represents noise data; This represents the probability distribution of the generated output prediction scenario; This represents the mapping function constructed by the discriminant model D; Represents the constraint conditions, and represents the mapping function. The Q-Lipschitz continuity condition must be satisfied. Represents the Lipschitz constant; This indicates the upper bound operation; It represents a function that calculates the expected value of a mathematical expression.

5. The spatiotemporal joint prediction method for rural distributed source-storage-load resources according to claim 1, characterized in that, Clustering and reducing the power output prediction scenarios to determine the distributed power output prediction results for typical distributed power output scenarios includes: Based on the DTW distance between any two output prediction scenarios, determine the metric distance between any two output prediction scenarios; Based on the metric distance between any two output prediction scenarios, the K-medoids clustering algorithm is used to cluster the output prediction scenarios, resulting in multiple clusters and their cluster centers. The cluster centers of all clusters are then determined as the distributed power output prediction results for typical distributed power output scenarios.

6. The spatiotemporal joint prediction method for rural distributed source-storage-load resources according to claim 5, characterized in that, Clustering and reducing the output prediction scenarios to determine the distributed power generation output prediction results for typical distributed power generation scenarios also includes: Based on the actual output power and predicted output value of distributed power sources, determine the spatiotemporal disturbance factor; Based on the aforementioned spatiotemporal disturbance factor, the distributed power output prediction results for typical distributed power output scenarios are corrected to obtain the final distributed power output prediction results for typical distributed power output scenarios.

7. The spatiotemporal joint prediction method for rural distributed source-storage-load resources according to claim 1, characterized in that, The multiple load types include at least industrial loads; for industrial loads, load forecasting is performed, and the calculation formula corresponding to the load forecast results is as follows: In the formula: Indicates industrial load During the adjustment period Actual power consumption; Indicates industrial load During the adjustment period The power consumption of the original production plan; This indicates that the industrial load i is during the adjustment period. The adjustable status of the production plan is represented by a 0-1 variable, where 0 indicates that it is not adjustable and 1 indicates that it is adjustable. Indicates industrial load During the adjustment period The amount of power adjustment; Indicates industrial load During the adjustment period Available response power; Indicates industrial load The baseline power consumption; Indicates the period during which the power grid command requires an increase in load; This indicates the period during which the power grid requires a load reduction.

8. The spatiotemporal joint prediction method for rural distributed source-storage-load resources according to claim 1, characterized in that, The multi-type loads include at least electric vehicle loads; for electric vehicle loads, load forecasting is performed, and the calculation formula corresponding to the load forecast results is as follows: In the formula: and These represent the adjustment periods for electric vehicles. and The remaining battery power; and These respectively represent the adjustment period. The 0-1 variables that control the charging and discharging states; and These represent the charging efficiency and discharging efficiency of an electric vehicle, respectively. and These represent the rated charging power and rated discharging power of the electric vehicle, respectively. Indicates electric vehicles during the adjustment period Grid connection status; and These represent the grid connection time and the grid disconnection time, respectively. This indicates the load of electric vehicles during the adjustment period. Available response power; Indicates electric vehicles during the time period The reference charging power; Indicates electric vehicles during the adjustment period The charging power; Indicates the period during which the power grid command requires an increase in load; This indicates the period during which the power grid requires a load reduction.

9. The spatiotemporal joint prediction method for rural distributed source-storage-load resources according to claim 1, characterized in that, Based on the forecast results of different types of loads and the output forecast results of distributed generation, the available response power of user-side energy storage during the adjustment period is determined, including: Based on the output prediction results of distributed power sources and the prediction results of different types of loads, calculate the net power deviation between the source and load. Based on the net power deviation of the source and load and the source and load conduction coefficient, and combined with the charging and discharging power, charging and discharging state and charging and discharging efficiency of the user-side energy storage, the state of charge of the user-side energy storage is dynamically updated. Based on the actual operating power and the baseline operating power of user-side energy storage, the available response power of user-side energy storage during the adjustment period is calculated.

10. A spatiotemporal joint prediction system for rural distributed source-storage-load resources, characterized in that, It includes a memory, a processor, and executable computer program code stored in the memory and executable on the processor. When the processor executes the computer program code, it performs a spatiotemporal joint prediction method for rural distributed source-storage-load resources as described in any one of claims 1 to 9.

Images