Photovoltaic power value oriented prediction method based on combined prediction

Abstract

The invention provides a photovoltaic power value oriented prediction method based on combined prediction, and relates to the technical field of photovoltaic power generation power prediction and power system optimization scheduling, and the method comprises the steps: 1, constructing a photovoltaic power combined prediction model which is formed by the weighted combination of a plurality of basic prediction models; step 2, clustering photovoltaic output and electricity price data by using a self-organizing mapping (SOM) algorithm, and dividing samples into a plurality of categories; step 3, aiming at each category, taking the maximization of the operation income of the optical storage power station as an objective function, and adopting a hunger game search (HGS) algorithm to optimize the weight coefficient of the combined prediction model; and step 4, performing combined prediction on the photovoltaic power according to the optimized weight coefficient, and outputting a value-oriented photovoltaic power prediction result. According to the method, the problem of poor economic benefit of traditional precision-oriented prediction in the dispatching of the optical storage power station is solved, and the operation benefit of the optical storage power station in the electric power market is effectively improved.

Description

Technical Field

[0001] This invention belongs to the field of photovoltaic power generation prediction and power system optimal dispatch technology, specifically a photovoltaic power value-oriented prediction method based on combined prediction. Background Technology

[0002] In recent years, the penetration rate of photovoltaic (PV) power generation in the power system has gradually increased. However, the randomness and volatility of PV system output power can have adverse effects on renewable energy consumption and the safe and stable operation of the power system. PV power forecasting can assist PV power plants in formulating power generation plans, optimizing dispatch schemes, participating in the energy market, and improving the utilization rate and economic efficiency of renewable energy.

[0003] As an important form of distributed energy, photovoltaic (PV) and energy storage (ESS) power stations need to report their next day's power generation to the power dispatching department based on the day-ahead PV forecast results. Traditional PV power forecasting is accuracy-oriented; however, due to the long time span and high uncertainty of day-ahead PV forecasts, the accuracy is relatively low, leading to discrepancies between the actual PV power generation and the reported power. This error not only reduces the operating revenue of the power station itself but also affects the stable operation of the power grid. However, the relationship between the operating revenue of PV and ESS power stations and PV forecast errors is non-linear and asymmetric. To further improve the operating revenue of PV and ESS power stations and other new energy manufacturers, it is necessary to conduct value-oriented forecasting that integrates new energy forecasting and dispatching models. This involves training the PV forecasting model with the goal of maximizing the operating revenue of new energy manufacturers, and further exploring the commercial value of PV and ESS power stations when it is difficult to significantly improve forecast accuracy. Summary of the Invention

[0004] The purpose of this invention is to provide a photovoltaic power value-oriented prediction method based on combined prediction. It uses combined prediction as the upper-level prediction model, reduces the parameter scale and optimization range of the prediction model, and effectively ensures the accuracy of the prediction model. The photovoltaic prediction model is trained with the goal of maximizing operating benefits, which significantly improves the economic benefits of photovoltaic and energy storage power stations.

[0005] To achieve the above objectives, the present invention adopts the following technical solution: A photovoltaic power value-oriented prediction method based on combined prediction includes: Step 1: Construct a photovoltaic power combined prediction model, which is composed of a weighted combination of multiple basic prediction models; Step 2: Use the self-organizing map (SOM) algorithm to cluster the photovoltaic power output and electricity price data, dividing the samples into multiple categories; Step 3: For each category, with the objective function of maximizing the operating revenue of the photovoltaic-storage power station, the weight coefficients of the combined prediction model are optimized using the Hunger Games Search (HGS) algorithm. Step 4: Perform combined prediction of photovoltaic power based on the optimized weighting coefficients, and output value-oriented photovoltaic power prediction results.

[0006] Furthermore, step 1 specifically includes: Step 1.1: Training the Support Vector Machine Regression (SVR) basic model: The principle of Support Vector Machines is to minimize the structural empirical risk of the function set. Specifically, it projects low-dimensional sample data into a high-dimensional space to find the classification hyperplane that achieves the highest classification accuracy and is furthest from the sample points. For the sample set... First, input variables Transform to a high-dimensional feature space The classification hyperplane can be represented as: Since the hyperplane needs to correctly classify the samples, it must satisfy the constraints. The classification margin between the hyperplane and the sample points is The hyperplane construction problem can be transformed into an optimization problem: By introducing an insensitive loss function, support vector machines can be extended to solve regression problems, denoted as... : In the formula: c This is a constant used to balance the loss value and the regularization term; This is the insensitive loss coefficient.

[0007] Step 1.2: Training the Gradient Boosting Decision Tree (GBDT) base model: Gradient boosting decision tree is a decision tree algorithm based on iterative error reduction. In each iteration, a new decision tree is built in the negative gradient direction of the error function. All decision trees are weighted to obtain the final prediction model. For the sample set, GBDT iterates multiple times. M Second-rate, This represents the initial decision tree, i.e., the predicted values. The objective function of GBDT is: In the formula: L Represents the loss function; Indicate the complexity of the decision tree; T To determine the number of leaf nodes; Represents a leaf node; The splitting coefficient; The penalty coefficient is used. During each iteration, a decision tree is generated according to the greedy algorithm. The gradient descent direction of equation (4) is calculated and used as a leaf node. The generated decision tree is then added to the model. Step 1.3: Train the basic Random Forest (RF) model: Random Forest is also an ensemble learning algorithm based on decision trees. For the sample set... The RF regression algorithm first uses the bootstrap sampling method to extract samples from the sample set. k For each sample, a corresponding decision tree is constructed. Next, attributes are randomly selected from the decision tree nodes, and the nodes are split to construct individual classification and regression trees. In each regression tree, leaf nodes are visited sequentially according to threshold order, and the average of all leaf nodes is used as the predicted value. Then, the above two steps are repeated to construct a random forest, and the average of the classification and regression tree outputs is taken to obtain the final result.

[0008] Step 1.4: Training the basic Multilayer Perceptron (MLP) model: The multilayer perceptron is the most basic neural network structure, mainly consisting of an input layer, hidden layers, and an output layer. Taking the simplest three-layer MLP model as an example, the sample set is... The input layer contains m There are 10 neural nodes. The hidden layer transforms the data passed from the input layer: In the formula: These are the weighting coefficients. b 1 represents the bias, a common activation function. Including ReLU, Sigmoid, etc. The number of neurons in the output layer is the same as the dimension of the output variable; data is passed to the hidden layer and then transformed. Step 1.5: Training the K-Nearest Neighbors (KNN) basic model: The basic principle of KNN is a vector space model, which searches for the K-Nearest Neighbors model within the feature space. k Find the nearest neighbors and use the mean of the nearest neighbors as the estimate. For the sample set... Using Euclidean distance as the distance between input features: Step 1.6: Construct a combined prediction model using basic models: Combine two or more basic prediction models using weighting coefficients, and obtain the prediction results through a weighted summation method. In the formula: The prediction results are from the combined prediction. n The number of basic prediction models; and The first i The weight coefficients and prediction results of each prediction model, where... Must meet: Predictive model Transform into a combined prediction form: In the formula: Indicates the first s In the nth sample i The photovoltaic prediction values ​​of each prediction model. Substituting the combined prediction model into the prediction-scheduling two-level problem, we get: Equation (13) transforms the prediction model into a combined prediction form, with the optimization parameters being the weight coefficients of the combined prediction. This effectively reduces the parameter size and optimization range of the prediction model, transforms the complex power prediction model into a linear function form, and reduces the difficulty of solving the two-layer problem.

[0009] Furthermore, the SOM algorithm in step 2 is an unsupervised neural network algorithm, consisting of an input layer and a competition layer. It determines the neighborhood of the winning neuron and calculates the weights of neurons within that neighborhood. By updating the network, stable data classification is achieved. Step 2 specifically includes: Step 2.1, Data Preprocessing: Clean and filter feature variables, and preprocess the input features. Normalization is performed, where N Input feature dimension; Step 2.2, Network Initialization: Set the maximum number of iterations to [value missing]. K Parameter learning rate Neighborhood of neurons All vary with the number of iterations k Increase and decrease, competitive layer neurons are M their weights Assign random initial values; Step 2.3, Neuron Competition: Calculate the distance between the input feature vector and the weight vectors of all neurons in the competing layer. The neuron with the smallest distance is the winning neuron. c Its corresponding weight vector is Euclidean distance is used as the distance metric. Step 2.4, Network Weights and Input Variable Update: Update the neuron weights in the neighborhood of the winning neuron: weight function h ci The definition is as follows: Select the next training sample as the network input variable, return to step 3, and continue until the training set of samples has been traversed; Step 2.5, Convergence Check: Determine if the maximum number of iterations has been reached. K If the error between the two iterations is less than the set value, or if it is not reached, return to step 2.3.

[0010] Furthermore, in step 2, the input features for SOM classification are photovoltaic prediction values ​​and electricity price prediction values. After classification, the parameters of the combined prediction model are trained for each cluster separately. In the formula: Indicates the category to which the sample belongs; classify () indicates a classification SOM network; x The photovoltaic forecast and electricity price forecast are used as input variables and feature values.

[0011] Furthermore, HGS in step 3 is a swarm intelligence algorithm that simulates hungry organisms searching for food, possessing advantages such as simple structure, easy convergence, and high stability. HGS simulates cooperative predation by a group of organisms: In the formula: r 1. r 2 is a random number that follows a uniform distribution in the interval [0,1]. randn (1) is a random number that follows a standard normal distribution; t Indicates the number of iterations; X b This represents the optimal position of an individual, i.e., the optimal weight coefficient to be searched; X ( t ) indicates the first t The individual position in the round of iteration, i.e., the combined prediction weights used in this round of iteration; l For hyperparameters; E This represents the search position control variable, and the hyperbolic secant function is used to calculate the distance between the current search weight and the optimal weight. R Indicates the scope of the biological search. R ( i The expression ∈ [-1, 1] gradually decreases to 0 as the number of iterations increases; W 1 and W 2 represents the hunger weight, and the calculation formulas are as follows: In the formula: F ( i ) indicates the first i The fitness of an organism; BFThe optimal fitness is represented by the objective function value of equation (13); r 3. r 4. r 5. rand is a random number that follows a uniform distribution in the interval [0,1]. N Indicates the set population size; hungry ( i The calculation formula is as follows: In the formula: TH , LH These are hyperparameters representing the upper and lower limits of the starvation value; UB and LB This represents the upper and lower limits of the search space, i.e., the optimization range of the weight coefficients, and is set to [0,1]. r 6 is a random number that follows a uniform distribution in the interval [0,1]. BF This indicates the worst fitness.

[0012] Compared with the prior art, the present invention has the following beneficial effects: This invention uses combined prediction as the upper-level prediction model, which simplifies the prediction model form, reduces the range of optimization parameters, and effectively ensures the accuracy of the prediction model. The value-oriented prediction method, which is guided by operating benefits, can effectively improve the economic benefits of photovoltaic and energy storage power stations participating in the two-stage electricity market. Attached Figure Description

[0013] Figure 1 A flowchart of a photovoltaic power value-oriented prediction method based on combined prediction according to the present invention. Detailed Implementation The present invention will now be described in detail with reference to embodiments. It should be noted that, unless otherwise specified, the embodiments and features described in the embodiments of this application can be combined with each other.

[0014] The following detailed description is exemplary and intended to provide further detailed explanation of the invention. Unless otherwise specified, all technical terms used in this invention have the same meaning as commonly understood by one of ordinary skill in the art to which this application pertains. The terminology used in this invention is for the purpose of describing particular embodiments only and is not intended to limit the scope of exemplary embodiments according to the invention.

[0015] Please see Figure 1 As shown, this invention provides a photovoltaic power value-oriented prediction method based on combined prediction, comprising: Step 1: Construct a photovoltaic power combined prediction model. The combined prediction model is composed of a weighted combination of multiple basic prediction models, specifically including: Step 1.1: Training the Support Vector Machine Regression (SVR) basic model: The principle of Support Vector Machines is to minimize the structural empirical risk of the function set. Specifically, it projects low-dimensional sample data into a high-dimensional space to find the classification hyperplane that achieves the highest classification accuracy and is furthest from the sample points. For the sample set... First, input variables Transform to a high-dimensional feature space The classification hyperplane can be represented as: Since the hyperplane needs to correctly classify the samples, it must satisfy the constraints. The classification margin between the hyperplane and the sample points is The hyperplane construction problem can be transformed into an optimization problem: By introducing an insensitive loss function, support vector machines can be extended to solve regression problems, denoted as... : In the formula: c This is a constant used to balance the loss value and the regularization term; This is the insensitive loss coefficient.

[0016] Step 1.2: Training the Gradient Boosting Decision Tree (GBDT) base model: Gradient boosting decision tree is a decision tree algorithm based on iterative error reduction. In each iteration, a new decision tree is built in the negative gradient direction of the error function. All decision trees are weighted to obtain the final prediction model. For the sample set, GBDT iterates multiple times. M Second-rate, This represents the initial decision tree, i.e., the predicted values. The objective function of GBDT is: In the formula: L Represents the loss function; Indicate the complexity of the decision tree; T To determine the number of leaf nodes; Represents a leaf node; The splitting coefficient; The penalty coefficient is used. During each iteration, a decision tree is generated according to the greedy algorithm. The gradient descent direction of equation (4) is calculated and used as a leaf node. The generated decision tree is then added to the model. Step 1.3: Train the basic Random Forest (RF) model: Random Forest is also an ensemble learning algorithm based on decision trees. For the sample set... The RF regression algorithm first uses the bootstrap sampling method to extract samples from the sample set. kFor each sample, a corresponding decision tree is constructed. Next, attributes are randomly selected from the decision tree nodes, and the nodes are split to construct individual classification and regression trees. In each regression tree, leaf nodes are visited sequentially according to threshold order, and the average of all leaf nodes is used as the predicted value. Then, the above two steps are repeated to construct a random forest, and the average of the classification and regression tree outputs is taken to obtain the final result.

[0017] Step 1.4: Training the basic Multilayer Perceptron (MLP) model: The multilayer perceptron is the most basic neural network structure, mainly consisting of an input layer, hidden layers, and an output layer. Taking the simplest three-layer MLP model as an example, the sample set is... The input layer contains m There are 10 neural nodes. The hidden layer transforms the data passed from the input layer: In the formula: These are the weighting coefficients. b 1 represents the bias, a common activation function. Including ReLU, Sigmoid, etc. The number of neurons in the output layer is the same as the dimension of the output variable; data is passed to the hidden layer and then transformed. Step 1.5: Training the K-Nearest Neighbors (KNN) basic model: The basic principle of KNN is a vector space model, which searches for the K-Nearest Neighbors model within the feature space. k Find the nearest neighbors and use the mean of the nearest neighbors as the estimate. For the sample set... Using Euclidean distance as the distance between input features: Step 1.6: Construct a combined prediction model using basic models: Combine two or more basic prediction models using weighting coefficients, and obtain the prediction results through a weighted summation method. In the formula: The prediction results are from the combined prediction. n The number of basic prediction models; and The first i The weight coefficients and prediction results of each prediction model, where... Must meet: Predictive model Transform into a combined prediction form: In the formula: Indicates the first s In the nth sample iThe photovoltaic prediction values ​​of each prediction model. Substituting the combined prediction model into the prediction-scheduling two-level problem, we get: Equation (13) transforms the prediction model into a combined prediction form, with the optimization parameters being the weight coefficients of the combined prediction. This effectively reduces the parameter size and optimization range of the prediction model, transforms the complex power prediction model into a linear function form, and reduces the difficulty of solving the two-layer problem.

[0018] Step 2: Use the self-organizing map (SOM) algorithm to cluster the photovoltaic power output and electricity price data, dividing the samples into multiple categories, specifically including: Step 2.1, Data Preprocessing: Clean and filter feature variables, and preprocess the input features. Normalization is performed, where N Input feature dimension; Step 2.2, Network Initialization: Set the maximum number of iterations to [value missing]. K Parameter learning rate Neighborhood of neurons All vary with the number of iterations k Increase and decrease, competitive layer neurons are M their weights Assign random initial values; Step 2.3, Neuron Competition: Calculate the distance between the input feature vector and the weight vectors of all neurons in the competing layer. The neuron with the smallest distance is the winning neuron. c Its corresponding weight vector is Euclidean distance is used as the distance metric. Step 2.4, Network Weights and Input Variable Update: Update the neuron weights in the neighborhood of the winning neuron: weight function h ci The definition is as follows: Select the next training sample as the network input variable, return to step 3, and continue until the training set of samples has been traversed; Step 2.5, Convergence Check: Determine if the maximum number of iterations has been reached. K If the error between the two iterations is less than the set value, or if it is not reached, return to step 2.3.

[0019] The input features for SOM classification are photovoltaic forecasts and electricity price forecasts. After classification, the parameters of the combined prediction model are trained for each cluster. In the formula: Indicates the category to which the sample belongs; classify () indicates a classification SOM network; x The photovoltaic forecast and electricity price forecast are used as input variables and feature values.

[0020] Step 3: For each category, with maximizing the operating revenue of the photovoltaic-storage power station as the objective function, the weight coefficients of the combined prediction model are optimized using the Hunger Games Search (HGS) algorithm. The specific method is as follows: HGS simulates cooperative predation by biological groups: In the formula: r 1. r 2 is a random number that follows a uniform distribution in the interval [0,1]. randn (1) is a random number that follows a standard normal distribution; t Indicates the number of iterations; X b This represents the optimal position of an individual, i.e., the optimal weight coefficient to be searched; X ( t ) indicates the first t The individual position in the round of iteration, i.e., the combined prediction weights used in this round of iteration; l For hyperparameters; E This represents the search position control variable, and the hyperbolic secant function is used to calculate the distance between the current search weight and the optimal weight. R Indicates the scope of the biological search. R ( i The expression ∈ [-1, 1] gradually decreases to 0 as the number of iterations increases; W 1 and W 2 represents the hunger weight, and the calculation formulas are as follows: In the formula: F ( i ) indicates the first i The fitness of an organism; BF The optimal fitness is represented by the objective function value of equation (13); r 3. r 4. r 5. rand is a random number that follows a uniform distribution in the interval [0,1]. N Indicates the set population size; hungry ( i The calculation formula is as follows: In the formula: TH , LHThese are hyperparameters representing the upper and lower limits of the starvation value; UB and LB This represents the upper and lower limits of the search space, i.e., the optimization range of the weight coefficients, and is set to [0,1]. r 6 is a random number that follows a uniform distribution in the interval [0,1]. BF This indicates the worst fitness.

[0021] Step 4: Perform combined prediction of photovoltaic power based on the optimized weighting coefficients, and output value-oriented photovoltaic power prediction results.

[0022] Example Hourly photovoltaic (PV) output data from a photovoltaic (PV) power plant in Bend, Oregon, USA, from August 2, 2018 to October 31, 2023, was used, along with meteorological forecast data from the National Center for Environmental Prediction and market-clearing price data from the PJM electricity market's COMED node. Due to the intermittent nature of PV output, only power data from 6:00 AM to 7:00 PM local time was used during the training of the basic PV forecasting model. A comparative analysis was conducted using the basic forecasting model MLP, the accuracy-oriented forecasting (AOF) method (i.e., the traditional dispatch scheme), the linear regression (LR) method, and the value-oriented forecasting (VOF) method proposed in this invention. The steps of the photovoltaic power value-oriented prediction method based on combined prediction are as follows: Input photovoltaic data, meteorological data, and electricity price data; Train the SVR basic prediction model according to the formula in step 1.1; Train the GBDT basic prediction model according to the formula in step 1.2; Train the RF basic prediction model according to step 1.3; Train the MLP basic prediction model according to the formula in step 1.4; Train the KNN basic prediction model according to the formula in step 1.5; Construct a combined prediction model based on the formula in step 1.6; Based on the SOM algorithm in step 2, the photovoltaic power output and electricity price data are clustered to divide the samples into multiple categories. The weight coefficients of the combined prediction model for each category are optimized based on the HGS algorithm in step 3. Based on step 4, a combined prediction of photovoltaic power is performed, outputting value-oriented photovoltaic power prediction results. The results are as follows: Comparison of prediction performance and operating revenue of each model Comparison of AOF and VOF effects in models Finally, it should be noted that the above embodiments are only used to illustrate the technical solutions of the present invention and not to limit it. Although the present invention has been described in detail with reference to the above embodiments, those skilled in the art should understand that modifications or equivalent substitutions can still be made to the specific implementation of the present invention. Any modifications or equivalent substitutions that do not depart from the spirit and scope of the present invention should be covered within the scope of protection of the claims of the present invention.

Claims

1. A photovoltaic power value-oriented prediction method based on combined prediction, characterized in that, include: Step 1: Construct a photovoltaic power combined prediction model, which is composed of a weighted combination of multiple basic prediction models; Step 2: Use the Self-Organizing Map (SOM) algorithm to cluster the photovoltaic power output and electricity price data, dividing the samples into multiple categories; Step 3: For each category, with the objective function of maximizing the operating revenue of the photovoltaic-storage power station, the Hunger Games Search (HGS) algorithm is used to optimize the weight coefficients of the combined prediction model. Step 4: Perform combined prediction of photovoltaic power based on the optimized weighting coefficients, and output value-oriented photovoltaic power prediction results.

2. The photovoltaic power value-oriented prediction method based on combined prediction according to claim 1, characterized in that, In step 1, the combined prediction model is a weighted combination of five basic prediction models: Support Vector Machine Regression (SVR), Gradient Boosting Decision Tree (GDBT), Random Forest (RF), Multilayer Perceptron (MLP), and K-Nearest Neighbors (KNN).

3. The photovoltaic power value-oriented prediction method based on combined prediction according to claim 2, characterized in that, The Support Vector Machine Regression (SVR) model projects low-dimensional sample data into a high-dimensional space, searching for the hyperplane with the highest classification accuracy and the greatest distance from the sample points. For the sample set... First, input variables Transform to a high-dimensional feature space The classification hyperplane is represented as: , constraint is The classification margin between the hyperplane and the sample points is The hyperplane construction problem is transformed into an optimization problem: By introducing an insensitive loss function, support vector machines can be extended to solve regression problems, denoted as... : In the formula: c This is a constant used to balance the loss value and the regularization term; This is the insensitive loss coefficient.

4. The photovoltaic power value-oriented prediction method based on combined prediction according to claim 3, characterized in that, The Gradient Boosting Decision Tree (GBDT) model: For the sample set, the Gradient Boosting Decision Tree (GBDT) model iterates multiple times. M Second-rate, The initial decision tree represents the predicted value. The objective function of the Gradient Boosting Decision Tree (GBDT) model is: In the formula: L Represents the loss function; Indicate the complexity of the decision tree; T To determine the number of leaf nodes; Represents a leaf node; The splitting coefficient; As the penalty coefficient, a decision tree is generated according to the greedy algorithm in each iteration. The gradient descent direction of equation (4) is calculated and used as the leaf node. The generated decision tree is then added to the model. 。 5. A photovoltaic power value-oriented prediction method based on combined prediction according to claim 4, characterized in that, The Random Forest (RF) model: for the sample set First, the bootstrap sampling method is used to extract samples from the sample set. k For each sample, a corresponding decision tree is constructed. Then, the attribute set of the decision tree node is randomly selected, and the decision tree node is split to construct a single classification and regression tree. In the single regression tree, the leaf nodes are visited in order of threshold, and the average value of all leaf nodes is used as the predicted value. Then repeat the above two steps to construct a random forest, and take the average of the output results of the classification and regression trees to obtain the final result.

6. The photovoltaic power value-oriented prediction method based on combined prediction according to claim 5, characterized in that, The Multilayer Perceptron (MLP) model includes an input layer, hidden layers, and an output layer, with the following sample set: The input layer contains m Each neural node transforms the data passed from the input layer to the hidden layer. In the formula: These are the weighting coefficients. b 1 represents the bias, a common activation function. Including ReLU and Sigmoid; The number of neurons in the output layer is the same as the dimension of the output variable. Data is then passed to the hidden layer and transformed. 。 7. A photovoltaic power value-oriented prediction method based on combined prediction according to claim 6, characterized in that, The K-Nearest Neighbors (KNN) model: for the sample set Using Euclidean distance as the distance between input features: 。 8. The photovoltaic power value-oriented prediction method based on combined prediction according to claim 1, characterized in that, In step 1, combined prediction refers to combining two or more basic prediction models using weighting coefficients and obtaining the prediction result through a weighted summation: In the formula: The prediction results are from the combined prediction. n The number of basic prediction models; and The first i The weight coefficients and prediction results of each prediction model, where... Must meet: Predictive model Transform into a combined prediction form: In the formula: Indicates the first s In the nth sample i The photovoltaic prediction values ​​from each prediction model are used to substitute the combined prediction model into the prediction-scheduling two-level problem, resulting in: Equation (13) transforms the prediction model into a combined prediction form, with the optimization parameter being the weight coefficient of the combined prediction.

9. A photovoltaic power value-oriented prediction method based on combined prediction according to claim 1, characterized in that, The Self-Organizing Map (SOM) algorithm in step 2 consists of an input layer and a competition layer. It determines the neighborhood of the winning neuron and calculates the weights of neurons within that neighborhood. By updating the network, it achieves stable data classification. The training steps are as follows: 1) Data preprocessing: Cleaning and filtering feature variables, and processing the input features. Normalization is performed, where N Input feature dimension; 2) Network initialization: Set the maximum number of iterations to [value]. K Parameter learning rate Neighborhood of neurons All vary with the number of iterations k Increase and decrease, competitive layer neurons are M their weights Assign random initial values; 3) Neuron Competition: Calculate the distance between the input feature vector and the weight vectors of all neurons in the competing layer, and the neuron with the smallest distance is the winning neuron. c Its corresponding weight vector is Euclidean distance is used as the distance metric. 4) Network weights and input variable updates: Update the weights of neurons in the neighborhood of the winning neuron: weight function h ci The definition is as follows: Select the next training sample as the network input variable and return to step 3), until the training set of samples has been traversed; 5) Convergence check: Determine if the maximum number of iterations has been reached. K Or the error between the two iterations is less than the set value; if not, return to step 3).

10. A photovoltaic power value-oriented prediction method based on combined prediction according to claim 1, characterized in that, In step 2, the input features for the self-organizing map (SOM) algorithm classification are the photovoltaic prediction value and the electricity price prediction value. After classification, the parameters of the combined prediction model are trained for each cluster. In the formula: Indicates the category to which the sample belongs; classify () indicates a classification SOM network; x The photovoltaic forecast and electricity price forecast are used as input variables and feature values.

11. A photovoltaic power value-oriented prediction method based on combined prediction according to claim 1, characterized in that, The Hunger Games Search (HGS) algorithm in step 3 is a swarm intelligence algorithm that simulates hungry organisms searching for food. The Hunger Games Search (HGS) algorithm simulates cooperative predation by a group of organisms: In the formula: r 1. r 2 is a random number that follows a uniform distribution in the interval [0,1]. randn (1) is a random number that follows a standard normal distribution; t Indicates the number of iterations; X b This represents the optimal position of an individual, i.e., the optimal weight coefficient to be searched; X ( t ) indicates the first t The individual position in the round of iteration, i.e., the combined prediction weights used in this round of iteration; l For hyperparameters; E This represents the search position control variable, and the hyperbolic secant function is used to calculate the distance between the current search weight and the optimal weight. R Indicates the scope of the biological search. R ( i The expression ∈ [-1, 1] gradually decreases to 0 as the number of iterations increases; W 1 and W 2 represents the hunger weight, and the calculation formulas are as follows: In the formula: F ( i ) indicates the first i The fitness of an organism; BF The optimal fitness is represented by the objective function value of equation (13); r 3. r 4. r 5. rand is a random number that follows a uniform distribution in the interval [0,1]. N Indicates the set population size; hungry ( i The calculation formula is as follows: In the formula: TH , LH These are hyperparameters representing the upper and lower limits of the starvation value; UB and LB This represents the upper and lower limits of the search space, i.e., the optimization range of the weight coefficients, and is set to [0,1]. r 6 is a random number that follows a uniform distribution in the interval [0,1]. BF This indicates the worst fitness.

Images