A virtual power plant short-term load prediction method, device and equipment

Abstract

This application discloses a method, apparatus, and equipment for short-term load forecasting of virtual power plants, relating to the field of load forecasting. First, this application constructs a multi-objective optimization function encompassing forecasting accuracy and model complexity. Then, it sets up a composite method combining a radial basis function (RBF) neural network and an improved non-dominated sorting genetic algorithm (NSGA-II). By leveraging the powerful nonlinear mapping capability of the RBF neural network and the multi-objective optimization advantages of the NSGA-II algorithm, it effectively handles the multivariate and multi-objective problems in virtual power plant load forecasting, improving the forecasting accuracy and computational efficiency of short-term load forecasting for virtual power plants.

Description

Technical Field

[0001] This application relates to the field of load forecasting, and in particular to a method, apparatus and equipment for short-term load forecasting of a virtual power plant. Background Technology

[0002] With the large-scale integration of renewable energy, Virtual Power Plants (VPPs), as a flexible dispatch system integrating various distributed energy sources, loads, and energy storage resources, have become an important component of modern power systems. VPPs not only optimize power dispatch and enhance grid reliability but also improve the economics of the electricity market. However, due to the wide distribution and uncertainty of VPP resources, accurately predicting the short-term load demand of VPPs has become a pressing problem.

[0003] Short-term load forecasting is a key technology for virtual power plant dispatching and market trading, providing a basis for load balancing in the electricity market and grid operation. Existing short-term load forecasting methods mainly include time series analysis methods, machine learning methods, and hybrid modeling methods. Among them, traditional time series analysis methods, such as the Autoregressive Moving Average Model (ARMA) and the Autoregressive Integrated Moving Average Model (ARIMA), perform poorly when handling complex nonlinear relationships and cannot fully utilize large amounts of historical data and external variables. In recent years, machine learning-based forecasting methods, such as Support Vector Machines (SVM) and Neural Networks (NN), have been widely used in load forecasting due to their strong nonlinear modeling capabilities and self-learning characteristics. However, these methods are prone to problems such as high computational complexity and insufficient optimization when training on large-scale data, especially when facing the various complex resources and dispatching strategies of virtual power plants, and still cannot meet the dual requirements of accuracy and efficiency. Summary of the Invention

[0004] The purpose of this application is to provide a method, apparatus, and equipment for short-term load forecasting of virtual power plants, which can improve the forecasting accuracy and computational efficiency of short-term load forecasting of virtual power plants.

[0005] To achieve the above objectives, this application provides the following solution:

[0006] In a first aspect, this application provides a method for short-term load forecasting of a virtual power plant, including:

[0007] Obtain historical load data sequences and multiple load influencing factor sequences of the virtual power plant;

[0008] The load influencing factors in multiple load influencing factor sequences are used as the input data of the samples, and the load data in the historical load data sequence are used as the short-term load label values ​​of the samples to construct training and test sets.

[0009] Construct a multi-objective optimization function; the multi-objective optimization function includes: a prediction accuracy objective function and a model complexity objective function;

[0010] Using the training set, the test set, and the multi-objective optimization function, the NSGA-II algorithm is used to train and test the RBF neural network model to obtain a trained RBF neural network model.

[0011] A trained RBF neural network model is used to perform short-term load forecasting for a virtual power plant.

[0012] Secondly, this application provides a virtual power plant short-term load forecasting device, which applies the aforementioned virtual power plant short-term load forecasting method, and the virtual power plant short-term load forecasting device includes:

[0013] The historical load data sequence acquisition module is used to acquire historical load data sequences and multiple load influencing factor sequences of the virtual power plant;

[0014] The training and test set construction module is used to construct the training set and test set by taking load influencing factors from multiple load influencing factor sequences as input data for the sample and load data from historical load data sequences as short-term load label values ​​for the sample.

[0015] A multi-objective optimization function construction module is used to construct multi-objective optimization functions; the multi-objective optimization functions include: a prediction accuracy objective function and a model complexity objective function;

[0016] The RBF neural network model training module is used to train and test the RBF neural network model using the training set, the test set, and the multi-objective optimization function, and employing the NSGA-II algorithm to obtain a trained RBF neural network model.

[0017] The short-term load forecasting module is used to perform short-term load forecasting on the virtual power plant using a trained RBF neural network model.

[0018] Thirdly, this application provides a computer device, including: a memory, a processor, and a computer program stored in the memory and executable on the processor, wherein the processor executes the computer program to implement the steps of the above-described virtual power plant short-term load forecasting method.

[0019] According to the specific embodiments provided in this application, the following technical effects are disclosed:

[0020] This application provides a method, apparatus, and equipment for short-term load forecasting of virtual power plants. First, a multi-objective optimization function covering prediction accuracy and model complexity is constructed. Then, a composite method combining a radial basis function (RBF) neural network and an improved non-dominated sorting genetic algorithm (NSGA-II) is set up. Utilizing the powerful nonlinear mapping capability of the RBF neural network and the multi-objective optimization advantages of the NSGA-II algorithm, the multivariate and multi-objective problems in virtual power plant load forecasting are effectively addressed, improving the prediction accuracy and computational efficiency of short-term load forecasting for virtual power plants. Attached Figure Description

[0021] To more clearly illustrate the technical solutions in the embodiments of this application or the prior art, the drawings used in the embodiments will be briefly introduced below. Obviously, the drawings described below are only some embodiments of this application. For those skilled in the art, other drawings can be obtained based on these drawings without creative effort.

[0022] Figure 1 A flowchart illustrating a short-term load forecasting method for a virtual power plant, provided as an embodiment of this application;

[0023] Figure 2 A flowchart of the NSGA-II algorithm provided in one embodiment of this application;

[0024] Figure 3 This is a comparison diagram showing the effects of different methods in one embodiment of this application;

[0025] Figure 4 This is a schematic diagram of the structure of a computer device provided in an embodiment of this application. Detailed Implementation

[0026] The technical solutions of the embodiments of this application will be clearly and completely described below with reference to the accompanying drawings. Obviously, the described embodiments are only some embodiments of this application, and not all embodiments. Based on the embodiments of this application, all other embodiments obtained by those skilled in the art without creative effort are within the scope of protection of this application.

[0027] To make the above-mentioned objectives, features and advantages of this application more apparent and understandable, the application will be further described in detail below with reference to the accompanying drawings and specific embodiments.

[0028] In one exemplary embodiment, such as Figure 1 and Figure 2 As shown, a method for short-term load forecasting of a virtual power plant is provided, including the following steps 101 to 105. Wherein:

[0029] Step 101: Obtain the historical load data sequence of the virtual power plant and multiple load influencing factor sequences.

[0030] Step 102: Use the load influencing factors in multiple load influencing factor sequences as the input data of the sample, and use the load data in the historical load data sequence as the short-term load label value of the sample to construct the training set and the test set.

[0031] Step 103: Construct a multi-objective optimization function; the multi-objective optimization function includes: a prediction accuracy objective function and a model complexity objective function.

[0032] Step 104: Using the training set, the test set, and the multi-objective optimization function, the NSGA-II algorithm is used to train and test the RBF neural network model to obtain the trained RBF neural network model.

[0033] Step 105: Use the trained RBF neural network model to perform short-term load forecasting for the virtual power plant.

[0034] Implementing steps 101 to 105 above can improve the prediction accuracy and calculation efficiency of short-term load forecasting for virtual power plants.

[0035] In another exemplary embodiment, step 101 of this application may be implemented based on steps 201-203.

[0036] Step 201, Data Collection.

[0037] Historical load data and load influencing factor data of a certain park were collected, including power load data every 15 minutes, temperature, wind speed, and date type (weekday, weekend, holiday) data, to form a raw dataset.

[0038] Collect prediction-related data, as follows:

[0039] Historical load data: Collects hourly actual load data from historical periods. Let the collected historical load data sequence be L = {l1, l2, ..., l...} H}, where l h This represents the load data at time h, where h = 1, 2, ..., H, and H is the total length of the data.

[0040] Historical wind speed data: Collects hourly actual wind speed data from historical periods. The collected wind speed data sequence is V = {v1, v2, ..., v...}. H}, where v h Let h represent the wind speed data at time h, where h = 1, 2, ..., H.

[0041] Temperature data: Obtain hourly temperature data within the same time period. Let the collected temperature data sequence be T = {t1, t2, ..., t...} H}, t h This represents the temperature data at time h.

[0042] Holiday data: Information on whether it is a holiday is represented by binary variables, corresponding to each hour within the past week. Let the holiday data sequence be H' = {h1, h2, ..., h...} H If h h =1 indicates that the h-th time is a holiday, h h =1 indicates that the h-th time is not a holiday.

[0043] Step 202, data cleaning.

[0044] 1. Outlier handling.

[0045] For load data, the "3σ principle" based on normal distribution is used to detect and remove outliers. First, the mean of the load data sequence L is calculated. and standard deviation σ p The calculation formula is as follows:

[0046]

[0047] If a certain data point l h satisfy Then determine l h Outliers are identified and removed from the dataset. Alternatively, it is feasible to set reasonable threshold ranges for identifying outliers based on common sense and past experience in the electricity market. For example, if the electricity price at a certain moment is far above or far below the upper or lower limit of the normal fluctuation range for the same period in history, it can be identified as an outlier and processed.

[0048] For temperature and wind speed data, outliers can be detected based on the reasonable fluctuation range of local historical temperature and wind speed, as well as the meteorological criteria for judging extreme values.

[0049] Holiday data, being binary variables generated based on explicit calendar rules, generally does not contain outliers. However, to ensure its accuracy, it is necessary to verify the data against the officially released holiday schedule.

[0050] 2. Handling missing values.

[0051] When load data l h If a missing value is found, fill it with the load value from the same time point in the previous week:

[0052] l h =l h-7×24 ;

[0053] l h-7×24 This represents the load data at the same time point one week ago at the h-th time.

[0054] 3. Temperature and wind speed data imputation: K-Nearest Neighbors (KNN) imputation was used to imput the temperature data t. h Wind speed data v h Missing values ​​are imputed. The core idea of ​​the KNN algorithm is to find the k most similar neighbor samples to the missing value based on the spatial distance of the sample data, and then predict the missing value based on the values ​​of these k neighbor samples. The formula is:

[0055]

[0056] in:

[0057] x h : Indicates a missing value, which is either missing temperature or wind speed data;

[0058] K h : represents x h The number of known data points of the nearest neighbor;

[0059] x represents h The kth h Known data of the nearest neighbors.

[0060] Step 203, data normalization.

[0061] To eliminate the adverse effects of data features of different magnitudes on the subsequent neural network training process, accelerate training speed, and improve training effectiveness, it is necessary to normalize the collected and cleaned data. This application adopts the min-max normalization method to map the values ​​of each data feature to a specified interval (usually [0,1]). The calculation formula for normalization is shown below.

[0062]

[0063] in, Here, n' represents the normalized data, and n' represents the original simulation data. min Let n be the minimum value of the data. max This represents the maximum value of the data.

[0064] After this normalization operation, all input data are within the same order of magnitude, which facilitates subsequent processing by the RBF neural network.

[0065] In another exemplary embodiment, step 102 above constructs a training set using data from the previous 12 months of normalized data, and constructs a test set using data from the following month of normalized data. A sliding window method is used for sample construction. h ={v' h ,t' h ,h' h} as input, y h =l' h As output, v' h ,t' h h' h and l' h These are normalized wind speed data, temperature data, holiday data, and load data, respectively.

[0066] In another exemplary embodiment, the prediction accuracy objective function in step 103 above is the mean absolute percentage error, and the model complexity objective function is the number of hidden layer nodes in the RBF neural network model.

[0067] The formula for calculating the mean absolute percentage error is:

[0068]

[0069] Where MAPE is the mean absolute percentage error, N is the number of samples in the test set, and y n To obtain the short-term load label value of the nth sample in the test set, The short-term load forecast value for the nth sample in the test set is obtained based on the RBF neural network model.

[0070] In each iteration, the RBF neural network parameters corresponding to each individual in the population are substituted into the network. After training the network using the training set, the mean squared error (MAPE) between the predicted and actual values ​​is calculated on the validation set. The goal of algorithm optimization is to minimize this MAPE value, that is, to make the predicted value as close as possible to the actual value, thereby improving the model's prediction accuracy for short-term electricity price data.

[0071] To avoid building overly complex RBF neural network models (complex models can lead to overfitting, high computational resource consumption, and other problems), the number of hidden layer nodes, F, is used as a metric for model complexity. The goal is to minimize the value of F while maintaining prediction accuracy, thus simplifying the model and improving its generalization ability and practicality. This ensures accurate prediction of short-term electricity prices without making the model too complex to be difficult to apply and maintain.

[0072] In another exemplary embodiment, the RBF neural network model in step 104 above can be replaced by steps 301-302 as follows.

[0073] Step 301: Determine the network structure.

[0074] 1. Input layer: Select three factors as input variables: temperature, wind speed, and holidays. Set the number of input layer nodes to three.

[0075] 2. Hidden Layers: The hidden layers are a crucial part of the RBF neural network for achieving nonlinear mapping, and determining the number of nodes is a critical step. Initially, they can be roughly set based on empirical formulas or simple experiments, and will be optimized using the NSGA-II algorithm later. The neurons in the hidden layers use Gaussian radial basis functions as activation functions, the expression of which is:

[0076]

[0077] In the above formula, Let x represent the output of the f-th hidden layer neuron, and c be the input vector. f It is the center vector of the f-th Gaussian radial basis function, and its dimension is the same as the dimension of the input vector x, σ f Let f be the width parameter of the f-th radial basis function, and ||·|| denote the Euclidean norm of the vector. This function calculates the width of the input vector x and the center vector c. f The distance, and based on the width parameter σ f Adjustments are made, and the corresponding function values ​​are output to achieve nonlinear transformation of the input data.

[0078] 3. Output Layer: The number of nodes in the output layer is determined based on the dimension of the data to be predicted. Since this application aims to predict short-term load data at a single moment, the number of nodes in the output layer is set to one. The output of each neuron in the output layer is a linear combination of the outputs of the hidden layers, calculated using the following formula:

[0079]

[0080] Where y represents the output of the output layer (i.e., the predicted short-term load data), F is the number of hidden layer nodes, and wf The weight coefficient is used to connect the f-th hidden layer node and the output layer node. This coefficient is used to linearly weight and combine the nonlinear outputs of the hidden layer to obtain the final predicted output value.

[0081] Step 302: Set initial parameters.

[0082] First, regarding the structural design of the RBF neural network model, the K-means clustering algorithm is used to classify the data. The number of neurons F in the hidden layer represents the number of categories of data to be classified. The cluster centers c can be calculated using the K-means algorithm. f and the width σ of the radial basis functions f .

[0083] σ f =min f' ||c f -c f' ||;

[0084] The number of categories in data classification directly impacts the performance of the RBF neural network model. The optimal F-value, or optimal number of categories, is obtained using the NSGA-II algorithm. Connection weights w f The initial settings are usually generated randomly within a reasonable range. For example, a random number generator can be used to generate initial values ​​for each w within the range [-1, 1]. f Initial values ​​are assigned so that the optimal weight configuration can be gradually found based on different initial weight states during subsequent training and optimization. The weights are updated using negative gradient descent. The error function and performance function of the RBF neural network are defined as follows:

[0085]

[0086] The weight adjustment strategy for the RBF output layer is as follows:

[0087]

[0088] Among them, W m+1 W is the weight matrix trained in the (m+1)th iteration. m Let be the weight matrix trained in the m-th iteration. J represents the gradient of the weight matrix during the m-th training iteration, where η is the learning rate, 0 < η < 1. m Let e ​​be the performance function of the RBF neural network model trained in the m-th iteration. m,i Let be the prediction error of the RBF neural network model trained in the m-th iteration for the i-th sample in the training set, and let be the short-term load label value of the i-th sample in the training set. The short-term load prediction value of the i-th sample in the training set is obtained by the RBF neural network model trained for the m-th iteration.

[0089] In another exemplary embodiment, step 104 above may be implemented using steps 401-410.

[0090] Step 401: Initialize the parameter optimization iteration number t to 0.

[0091] Step 402: Use the initial parameters of the RBF neural network model as individuals of the NSGA-II algorithm to initialize the population, and use the initialized population as the population for the t-th parameter optimization iteration.

[0092] Step 403: Use individuals in the population from the t-th parameter optimization iteration to set the initial parameters of the RBF neural network model, and obtain the RBF neural network model corresponding to each individual.

[0093] Step 404: Use the training set to train the RBF neural network model corresponding to each individual to obtain the trained RBF neural network model corresponding to each individual.

[0094] Step 405: Test the trained RBF neural network model for each individual using the test set to obtain the prediction accuracy objective function value for each individual, which is used as the fitness value for each individual.

[0095] Step 406: Based on the fitness value of each individual, perform non-dominated sorting on each individual in the population during the t-th parameter optimization iteration, and calculate the crowding degree of each individual in the population during the t-th parameter optimization iteration.

[0096] Step 407: Calculate the number of individuals to be discarded in each non-dominated layer.

[0097] Step 408: Based on the number of individuals to be discarded in each non-dominated layer and the crowding of each individual, discard the individuals in each non-dominated layer to obtain the parent population.

[0098] Step 409: Select, crossover, and mutate individuals in the parent population to generate the offspring population;

[0099] Step 410: Merge the parent and offspring populations to generate the population for the (t+1)th parameter optimization iteration. Increment the value of t by 1. Return to the step of "using individuals in the population for the tth parameter optimization iteration to set the initial parameters of the RBF neural network model and obtain the RBF neural network model corresponding to each individual". Continue until the maximum number of iterations is reached or an individual that meets the preset conditions is found. Output the trained RBF neural network model corresponding to the optimal individual as the trained RBF neural network model. The preset conditions are: the individual whose prediction accuracy objective function is greater than the preset accuracy threshold and whose model complexity objective function is less than the complexity threshold; the optimal individual is the individual with the largest prediction accuracy objective function value or the individual with the largest prediction accuracy objective function value among the individuals that meet the preset conditions.

[0100] In an exemplary embodiment, step 401 above further includes determining parameters such as the population size, maximum number of iterations, crossover probability, and mutation probability of the NSGA-II algorithm.

[0101] In an exemplary embodiment, in step 402 above, the population is initialized as follows: each individual in the population represents a set of parameters of the RBF neural network model (such as the number of hidden layer nodes, weights, etc.). Real-number encoding is used to encode the key parameters that need to be optimized in the RBF neural network model, so that these parameters can be incorporated into the population individual representation of the NSGA-II algorithm, facilitating subsequent genetic operations and parameter evolution.

[0102] Specific encoding format: The encoding of an individual can be represented as I = [c 11 ,c 12 ,c 13 ,c 14 ,…,c Fe ,σ1,σ2,…,σ F ,w1,w2,…,w F [,F], where F represents the hidden layer node index and F is the number of hidden layer nodes (the number of hidden layer nodes is also one of the parameters that needs to be optimized, and its value range and changes are reflected through encoding). This application's embodiment uses historical load, wind speed, temperature, and holidays as inputs to an RBF neural network, c fe (e = 1, 2, 3, 4, corresponding to the four dimensions of the input) represents the value of the center vector of the radial basis function of the f-th hidden layer node in each dimension, σ f w is the width parameter of the f-th radial basis function. f These are the weight coefficients connecting the f-th hidden layer node to the output layer node. Through this encoding form, f = 1, 2, ..., F, each individual represents a complete set of RBF neural network parameter configurations, facilitating subsequent operations and optimizations within the genetic algorithm framework.

[0103] In one exemplary embodiment, such as Figure 2 As shown, in step 403 above, it can be determined whether the size of the population in the t-th parameter optimization iteration is greater than the size of the parent population. If so, steps 403-408 are executed; otherwise, the parent population is directly generated. Usually, the population obtained from initialization will have a size that is not greater than the size of the parent population. The population in the iteration process will not have a size that is not greater than the size of the parent population.

[0104] In an exemplary embodiment, step 406 above improves the crowding calculation method by using Euclidean distance to represent the relative crowding of different individuals. Then, the absolute distance is calculated based on the relative crowding as the individual's crowding degree. The initial fitness is then used to measure the dominance of the solution, and the overall crowding degree of the individuals is calculated. Specifically, the crowding degree of the maximum and minimum fitness values ​​is set to infinity. The total crowding degree equals the sum of the crowding degrees of all individuals. After this improved crowding calculation, combining the previously solved Pareto front can more effectively identify offspring populations with high fitness and high individual richness, thus accelerating the optimization process of the algorithm. During optimization, individuals with higher crowding degrees are selected, as this results in lower similarity between individuals, which helps ensure the diversity of individuals in the population and reduces the likelihood of getting trapped in local optima compared to setting shared parameters.

[0105] In this application, Euclidean distance is used to represent the relative crowding of individuals. For individual k in a multi-objective optimization problem, let there be S objective functions in the multi-objective optimization function, and the coordinates of individual k in the objective space be (f 1k ,f 2k ,…,f Sk For any two individuals k and k', the Euclidean distance between them in the target space is calculated as follows:

[0106]

[0107] Suppose that there are K neighborhoods of individual k. k There are individuals, and the relative crowding degree between individual k and individual k' in the neighborhood of individual k is d. kk' The crowding degree of individual k is calculated using the following formula:

[0108]

[0109] Among them, D k Let d be the crowding degree of individual k. kk' K represents the relative crowding of individual k and individual k' within the neighborhood of individual k. k Let d be the number of individuals contained in the neighborhood of individual k. abLet f be the relative crowding degree between the a-th individual and the b-th individual, K be the number of individuals, and f be the number of individuals. s,k Let f be the value of the s-th objective function for individual k. s,k' Let S be the value of the s-th objective function for individual k', where S is the number of objective functions.

[0110] The above crowding calculation formula determines the degree of crowding in a local area by comparing the distance relationship between an individual and its surrounding individuals.

[0111] In another exemplary embodiment, steps 407 and 408 above employ an improved elite retention strategy. Traditional elite retention strategies first select individuals from lower non-dominated levels. Once the population size meets the requirements, individuals from higher non-dominated levels are not selected to enter the new population. In this method, crowding only affects the last selected non-dominated level, leading to a lack of population diversity. The improved elite retention strategy in this embodiment first determines the number of individuals to be deleted from each non-dominated level, ensuring that individuals from each non-dominated level have a chance to enter the new parent population. Furthermore, the lower the non-dominated level, the smaller the proportion of deleted individuals and the more retained individuals. This not only helps retain elite individuals but also increases population diversity.

[0112] The formula for the number of individuals to be discarded in each non-dominated layer is as follows:

[0113]

[0114] Among them, Q g Q represents the number of individuals that need to be discarded in the non-dominated layer g. e G represents the number of individuals in the current population that exceed the number in the parent population, and G represents the number of non-dominated layers. Q is the sign of the floor function. g' This represents the number of individuals that need to be discarded in the non-dominated layer g'.

[0115] In another exemplary embodiment, the selection in step 409 above is an important step to select some excellent individuals from the parent population to form the target population, generate the offspring population, so that the quality of the next generation solution is better than that of the current generation, and promote the evolution of the solution set toward the region where the optimal solution exists. In this embodiment, the binary tournament selection method is used.

[0116] In another exemplary embodiment, step 409 above employs an improved adaptive crossover and mutation probability strategy. A higher crossover probability increases the likelihood of changes in the existing individual structure, but if the original individual is superior, it may disrupt the feasible solution. Conversely, a lower crossover probability can easily cause the algorithm to get trapped in local optima. Therefore, the crossover probability of superior individuals can be set to be lower, while the crossover probability of inferior individuals can be set to be higher. Mutation can alter the values ​​of one or more genes in an individual, and its execution depends on the setting of the mutation probability, which is also adaptively adjusted.

[0117] In this embodiment, the formula for calculating the crossover probability is:

[0118]

[0119] Among them, P k,c Let P be the crossover probability of individual k. c ∈[P c2 ,P c1 ], P c2 and P c1 These are the minimum and maximum values ​​of the crossover probability, respectively, taking values ​​of 0.9 and 0.5; f k f is the fitness value of individual k. k ′ represents the maximum fitness of all individuals obtained by crossovering individual k with all individuals in the target population except individual k; f max and f avg These are the maximum and average fitness values ​​of all individuals in the target population, respectively.

[0120] In this embodiment, the formula for calculating the mutation probability is:

[0121]

[0122] Among them, P k,m Let P be the mutation probability of individual k. m1 and P m2 These represent the maximum and minimum crossover probabilities, respectively, with values ​​of 0.005 and 0.100. k "This is the fitness value of the individual obtained by mutating individual k."

[0123] In another exemplary embodiment, in step 410 above, to find a better solution, a local search operation is introduced. The parent and offspring populations are merged to obtain a merged population. A local search is then performed around the target individual in the merged population. The searched individuals are then compared with the target individual, and individuals with better performance are retained, thereby finding a better solution to ensure a uniform and widespread distribution of solutions. The target individual includes the individual with a non-dominated ranking level of 1 in the population from the previous parameter optimization iteration, and the top 5 most crowded individuals excluding that individual.

[0124] In another exemplary embodiment, in step 105 above, the temperature, wind speed and holiday data for the time period to be predicted are input into the trained RBF neural network model to perform short-term load forecasting, and the forecast error is calculated by comparing the predicted value with the actual value.

[0125] In another exemplary embodiment, to illustrate the effects of the embodiments of this application, relevant data of a certain park from January 1, 2018 to January 31, 2019 were first collected, including historical load, temperature, wind speed, and date type data. This data was preprocessed, and missing values ​​were handled using methods such as K-nearest neighbor imputation and imputation based on the average value of historical data for the same period. Outliers were corrected according to the 3σ principle and related factors, and the minimum-maximum normalization method was used to map the data to the [0,1] interval. Next, feature selection and extraction were performed. Features strongly correlated with electricity prices were screened using the Pearson correlation coefficient, and principal component analysis was used to reduce data dimensionality. An RBF neural network model was constructed and trained using data from the previous 12 months. The training effect was measured by the mean absolute percentage error, and an early shutdown method was used to prevent overtraining. Then, the parameters of the RBF neural network are optimized using a multi-objective optimization algorithm based on the improved NSGA-II. Parameters such as the number of hidden layer nodes are encoded as chromosomes. After initializing the population, non-dominated sorting and improved crowding distance calculations are performed based on the mean absolute percentage error on the test set (data from the last month) and the criterion of minimizing the F-value while ensuring prediction accuracy. The optimal parameter combination is determined through selection, improved adaptive crossover, and multi-generational evolutionary operations to obtain the optimized model. Finally, the preprocessed and feature-extracted data is input into the optimized model to obtain the load prediction results. The mean absolute percentage error index is used to evaluate the prediction effect, and the model is optimized accordingly. The load data of the park on a certain day is compared with other methods using the method of this application. The comparison results are as follows: Figure 3 As shown, the effectiveness of this method is verified.

[0126] It is evident that the method of this application has significant advantages over existing technologies. First, the RBF neural network model possesses powerful nonlinear mapping capabilities, effectively capturing complex nonlinear relationships in virtual power plant load data without relying on traditional linear assumptions. Therefore, compared to time series methods such as ARMA or ARIMA, the prediction accuracy is significantly improved. Second, the NSGA-II algorithm, through multi-objective optimization, effectively resolves the contradiction between accuracy and computational efficiency in load forecasting. Compared to traditional methods with a single optimization objective (such as SVM or BP neural networks), it considers multiple objectives during the optimization process, ensuring high convergence speed and low computational complexity. The method of this application, under large-scale data training and multi-resource scheduling environments, can adapt to diverse load characteristics and complex scheduling needs in virtual power plants, possessing better generalization ability and application value. It is suitable for short-term load forecasting tasks in actual power systems with high accuracy and real-time requirements. Furthermore, this application overcomes the subjectivity problem of existing congestion calculations, improves load forecasting accuracy, ensures population diversity, enhances algorithm stability and reliability, and has good adaptability and scalability.

[0127] Based on the same inventive concept, this application also provides a virtual power plant short-term load forecasting device for implementing the virtual power plant short-term load forecasting method described above. The solution provided by this device is similar to the solution described in the above method; therefore, the specific limitations in one or more embodiments of the virtual power plant short-term load forecasting device provided below can be found in the limitations of the virtual power plant short-term load forecasting method described above, and will not be repeated here.

[0128] In one exemplary embodiment, a virtual power plant short-term load forecasting device is provided, comprising:

[0129] The historical load data sequence acquisition module is used to acquire the historical load data sequence of the virtual power plant and multiple load influencing factor sequences.

[0130] The training and test set construction module is used to construct training and test sets by taking load influencing factors from multiple load influencing factor sequences as input data for samples and load data from historical load data sequences as short-term load label values ​​for samples.

[0131] A multi-objective optimization function construction module is used to construct multi-objective optimization functions; the multi-objective optimization functions include: a prediction accuracy objective function and a model complexity objective function.

[0132] The RBF neural network model training module is used to train and test the RBF neural network model using the training set, the test set, and the multi-objective optimization function, and employing the NSGA-II algorithm to obtain a trained RBF neural network model.

[0133] The short-term load forecasting module is used to perform short-term load forecasting on the virtual power plant using a trained RBF neural network model.

[0134] In one exemplary embodiment, a computer device is provided, which may be a server or a terminal, and its internal structure diagram may be as follows. Figure 4 As shown, this computer device includes a processor, memory, input / output (I / O) interfaces, and a communication interface. The processor, memory, and I / O interfaces are connected via a system bus, and the communication interface is also connected to the system bus via the I / O interfaces. The processor provides computational and control capabilities. The memory includes non-volatile storage media and internal memory. The non-volatile storage media stores the operating system, computer programs, and a database. The internal memory provides the environment for the operating system and computer programs stored in the non-volatile storage media to run. The I / O interfaces are used for exchanging information between the processor and external devices. The communication interface is used for communicating with external terminals via a network connection. When the computer program is executed by the processor, it implements a short-term load forecasting method for a virtual power plant.

[0135] Those skilled in the art will understand that Figure 4 The structure shown is merely a block diagram of a portion of the structure related to the present application and does not constitute a limitation on the computer device to which the present application is applied. Specific computer devices may include more or fewer components than those shown in the figure, or combine certain components, or have different component arrangements.

[0136] In one exemplary embodiment, a computer device is also provided, including a memory and a processor, wherein the memory stores a computer program, and the processor executes the computer program to implement the steps in the above-described method embodiments.

[0137] It should be noted that the user information (including but not limited to user device information, user personal information, etc.) and data (including but not limited to data used for analysis, data stored, data displayed, etc.) involved in this application are all information and data authorized by the user or fully authorized by all parties, and the collection, use and processing of the relevant data must comply with relevant regulations.

[0138] Those skilled in the art will understand that all or part of the processes in the above embodiments can be implemented by a computer program instructing related hardware. The computer program can be stored in a non-volatile computer-readable storage medium, and when executed, it can include the processes of the embodiments described above. Any references to memory, databases, or other media used in the embodiments provided in this application can include at least one of non-volatile and volatile memory. Non-volatile memory can include read-only memory (ROM), magnetic tape, floppy disk, flash memory, optical memory, high-density embedded non-volatile memory, resistive random access memory (ReRAM), magnetic random access memory (MRAM), ferroelectric random access memory (FRAM), phase change memory (PCM), graphene memory, etc. Volatile memory can include random access memory (RAM) or external cache memory, etc. By way of illustration and not limitation, RAM can take many forms, such as Static Random Access Memory (SRAM) or Dynamic Random Access Memory (DRAM).

[0139] The databases involved in the embodiments provided in this application may include at least one type of relational database and non-relational database. Non-relational databases may include, but are not limited to, blockchain-based distributed databases. The processors involved in the embodiments provided in this application may be general-purpose processors, central processing units, graphics processing units, digital signal processors, programmable logic devices, quantum computing-based data processing logic devices, etc., and are not limited to these.

[0140] The technical features of the above embodiments can be combined in any way. For the sake of brevity, not all possible combinations of the technical features in the above embodiments are described. However, as long as there is no contradiction in the combination of these technical features, they should be considered to be within the scope of this specification.

[0141] This document uses specific examples to illustrate the principles and implementation methods of this application. The descriptions of the above embodiments are only for the purpose of helping to understand the methods and core ideas of this application. Furthermore, those skilled in the art will recognize that, based on the ideas of this application, there will be changes in the specific implementation methods and application scope. Therefore, the content of this specification should not be construed as a limitation of this application.

Claims

1. A virtual power plant short-term load forecasting method, characterized in that, The virtual power plant short-term load forecasting method includes: Obtain historical load data sequences and multiple load influencing factor sequences of the virtual power plant; The load influencing factors in multiple load influencing factor sequences are used as the input data of the samples, and the load data in the historical load data sequence are used as the short-term load label values ​​of the samples to construct training and test sets. Construct a multi-objective optimization function; the multi-objective optimization function includes: a prediction accuracy objective function and a model complexity objective function; Using the training set, the test set, and the multi-objective optimization function, the NSGA-II algorithm is used to train and test the RBF neural network model to obtain a trained RBF neural network model. Short-term load forecasting of a virtual power plant is performed using a trained RBF neural network model. The formula for calculating congestion in the NSGA-II algorithm is: ； ； wherein, the crowding of an individual, the crowding of an individual, the relative crowding of an individual within the neighborhood of an individual, the number of individuals contained within the neighborhood of an individual, the relative crowding of the i-th individual and the j-th individual, the number of individuals, the value of the i-th objective function of an individual, the value of the i-th objective function of an individual, the value of the i-th objective function of an individual, the number of objective functions;​​​​​​​​​ The virtual power plant short-term load forecasting method also includes: For load data, the "3σ principle" based on normal distribution is used to detect and remove outliers; when load data When a missing value is found, fill it with the load value from the same time point in the previous week; For temperature and wind speed data, outliers are detected based on the reasonable fluctuation range of local historical temperature and wind speed, as well as meteorological extreme value criteria. When temperature and wind speed data are missing, the K-nearest neighbor imputation method is used to fill in the missing data. Wind speed data Fill in the missing values; The min-max normalization method was used to normalize the various types of data after cleaning.

2. The short-term load forecasting method for virtual power plants according to claim 1, characterized in that, The objective function for prediction accuracy is the mean absolute percentage error, and the objective function for model complexity is the number of hidden layer nodes in the RBF neural network model. The formula for calculating the mean absolute percentage error is: ； in, The mean absolute percentage error, This represents the number of samples in the test set. To obtain the short-term load label value of the nth sample in the test set, The short-term load forecast value for the nth sample in the test set is obtained based on the RBF neural network model.

3. The short-term load forecasting method for virtual power plants according to claim 1, characterized in that, Using the training set, the test set, and the multi-objective optimization function, the NSGA-II algorithm is employed to train and test the RBF neural network model, resulting in a well-trained RBF neural network model. Specifically, this includes: The initialization parameter optimization iteration number t is set to 0; The initial parameters of the RBF neural network model are used as individuals in the NSGA-II algorithm to initialize the population, and the population obtained from the initialization is used as the population for the t-th parameter optimization iteration. The initial parameters of the RBF neural network model are set using individuals in the population during the t-th parameter optimization iteration, thus obtaining the RBF neural network model corresponding to each individual. The RBF neural network model corresponding to each individual is trained using the training set to obtain the trained RBF neural network model for each individual. The trained RBF neural network model for each individual is tested using a test set to obtain the prediction accuracy objective function value for each individual, which is used as the fitness value for each individual. Based on the fitness value of each individual, perform non-dominated sorting on each individual in the population during the t-th parameter optimization iteration, and calculate the crowding degree of each individual in the population during the t-th parameter optimization iteration. Calculate the number of individuals to be discarded in each non-dominated layer; Based on the number of individuals to be discarded in each non-dominated layer, individuals in each non-dominated layer are discarded to obtain the parent population. Individuals in the parent population are selected, crossovered, and mutated to generate the offspring population; Merge the parent and offspring populations to generate the population for the (t+1)th parameter optimization iteration. Increment the value of t by 1 and return to the step of "setting the initial parameters of the RBF neural network model using individuals from the population for the tth parameter optimization iteration to obtain the RBF neural network model corresponding to each individual". Continue until the maximum number of iterations is reached or an individual that meets the preset conditions is found. Output the trained RBF neural network model corresponding to the optimal individual as the trained RBF neural network model. The preset conditions are: the individual whose prediction accuracy objective function is greater than the preset accuracy threshold and whose model complexity objective function is less than the complexity threshold; the optimal individual is the individual with the largest prediction accuracy objective function value or the individual with the largest prediction accuracy objective function value among the individuals that meet the preset conditions.

4. The short-term load forecasting method for virtual power plants according to claim 3, characterized in that, The training process of the RBF neural network model for each individual is conducted using the training set. The weights are updated using negative gradient descent, with the specific formula as follows: ； ； ； in, This is the weight matrix trained in the (m+1)th iteration. Let be the weight matrix trained in the m-th iteration. This represents the gradient of the weight matrix during the m-th training iteration. For learning rate, Let be the performance function of the RBF neural network model trained in the m-th iteration. Let be the prediction error of the RBF neural network model trained in the m-th iteration for the i-th sample in the training set, and let be the short-term load label value of the i-th sample in the training set. The short-term load prediction value of the i-th sample in the training set is obtained by the RBF neural network model trained for the m-th iteration.

5. The short-term load forecasting method for virtual power plants according to claim 3, characterized in that, The formula for calculating the number of individuals to be discarded in each non-dominated layer is as follows: ； in, Non-dominated layer The number of individuals that need to be discarded. The number of individuals in the current population that exceeds the number of individuals in the parent population. The number of non-dominated layers, This is the symbol for the floor function. Non-dominated layer The number of individuals that need to be discarded.

6. The short-term load forecasting method for virtual power plants according to claim 3, characterized in that, The offspring population is generated by selecting, crossovering, and mutating individuals in the parent population, specifically including: A predetermined number of individuals are selected from the parent population using a binary tournament selection method to form the target population. Calculate the crossover probability and mutation probability of each individual in the target population; Cross over each individual in the target population based on the crossover probability of each individual in the target population; Each individual in the target population is mutated based on its mutation probability.

7. The short-term load forecasting method for virtual power plants according to claim 6, characterized in that, The formula for calculating the crossover probability is: ； in, For individuals The crossover probability, , and These are the minimum and maximum values ​​of the crossover probability, respectively. Let k be the fitness value of individual k. The maximum fitness of all individuals obtained by crossover of individual k with all individuals in the target population except individual k. and These are the maximum and average fitness values ​​of all individuals in the target population, respectively. The formula for calculating the mutation probability is: ； in, For individuals The probability of mutation, and These are the maximum and minimum values ​​of the crossover probability, respectively. The fitness value of an individual obtained by mutating individual k.

8. A virtual power plant short-term load forecasting device, characterized in that, The virtual power plant short-term load forecasting device applies the virtual power plant short-term load forecasting method according to any one of claims 1-7, and the virtual power plant short-term load forecasting device comprises: The historical load data sequence acquisition module is used to acquire historical load data sequences and multiple load influencing factor sequences of the virtual power plant; The training and test set construction module is used to construct the training set and test set by taking load influencing factors from multiple load influencing factor sequences as input data for the sample and load data from historical load data sequences as short-term load label values ​​for the sample. A multi-objective optimization function construction module is used to construct multi-objective optimization functions; the multi-objective optimization functions include: a prediction accuracy objective function and a model complexity objective function; The RBF neural network model training module is used to train and test the RBF neural network model using the training set, the test set, and the multi-objective optimization function, and employing the NSGA-II algorithm to obtain a trained RBF neural network model. The short-term load forecasting module is used to perform short-term load forecasting on the virtual power plant using a trained RBF neural network model.

9. A computer device, comprising: A memory, a processor, and a computer program stored in the memory and executable on the processor, characterized in that the processor executes the computer program to implement the steps of the virtual power plant short-term load forecasting method according to any one of claims 1-7.

Images