Search

CN-122000882-A - Photovoltaic power value guiding prediction method based on combined prediction

CN122000882ACN 122000882 ACN122000882 ACN 122000882ACN-122000882-A

Abstract

The invention provides a combined prediction-based photovoltaic power value guiding prediction method, which relates to the technical field of photovoltaic power generation power prediction and power system optimization scheduling, and comprises the following steps of 1, constructing a photovoltaic power combined prediction model, wherein the combined prediction model is formed by weighting and combining a plurality of basic prediction models; the method comprises the steps of step 2, clustering photovoltaic output and electricity price data by adopting a self-organizing map (SOM) algorithm, dividing a sample into a plurality of categories, step 3, optimizing a weight coefficient of a combined prediction model by adopting a Hunger Game Search (HGS) algorithm by taking the maximum operation income of an optical storage station as an objective function aiming at each category, and step 4, carrying out combined prediction on photovoltaic power according to the optimized weight coefficient, and outputting a value-oriented photovoltaic power prediction result. The method solves the problem of poor economic benefit of traditional precision oriented prediction in the dispatching of the optical storage power station, and effectively improves the operation income of the optical storage power station in the electric power market.

Inventors

  • FANG CHEN
  • LIU SHU
  • XU XIAOYUAN
  • WANG MENGYUAN
  • LIU JINSONG
  • YAN ZHENG
  • QI GUODONG

Assignees

  • 国网上海市电力公司
  • 上海交通大学

Dates

Publication Date
20260508
Application Date
20260127

Claims (11)

  1. 1. The photovoltaic power value guiding prediction method based on the combined prediction is characterized by comprising the following steps of: Step 1, constructing a photovoltaic power combination prediction model, wherein the combination prediction model is formed by combining a plurality of basic prediction models in a weighting manner; Step 2, clustering photovoltaic output and electricity price data by adopting a self-organizing map SOM algorithm, and dividing a sample into a plurality of categories; step 3, optimizing the weight coefficient of the combined prediction model by adopting a hunger game search HGS algorithm by taking the maximum operation income of the optical storage station as an objective function according to each category; And 4, carrying out combined prediction on the photovoltaic power according to the optimized weight coefficient, and outputting a value-oriented photovoltaic power prediction result.
  2. 2. The photovoltaic power value guiding prediction method based on combined prediction according to claim 1, wherein the combined prediction model in the step 1 is formed by weighted combination of five basic prediction models, namely a support vector machine regression SVR model, a gradient lifting decision tree GDBT model, a random forest RF model, a multi-layer perceptron MLP model and a K nearest neighbor regression KNN model.
  3. 3. The method for combined prediction-based photovoltaic power value-oriented prediction as set forth in claim 2, wherein the SVR model is a support vector machine regression SVR model, wherein low-dimensional sample data is projected into a high-dimensional space, a classification hyperplane with highest classification accuracy and farthest from the sample points is found, and the sample set is obtained First, input variables Transforming to high-dimensional feature space The classification hyperplane is expressed as: Constrained as The classification interval between the hyperplane and the sample point is as follows The hyperplane construction problem translates into an optimization problem: By introducing a insensitive loss function, the support vector machine can be expanded to solve the regression problem, noted as : C is a constant for balancing the loss value and the regularization term; is a insensitive loss coefficient.
  4. 4. A combined prediction-based photovoltaic power value oriented prediction method as described in claim 3, wherein the gradient-lifting decision tree GBDT model iterates M times for the sample set, the gradient-lifting decision tree GBDT model, Representing an initial decision tree, i.e., a predicted value, the objective function of the gradient-lifting decision tree GBDT model is: Wherein L represents a loss function; T is the number of decision tree leaf nodes; Representing leaf nodes; Is a splitting coefficient; Generating a decision tree according to a greedy algorithm in each iteration process for punishment coefficients, calculating the gradient descent direction of the formula (4) and using the gradient descent direction as a leaf node, and adding the generated decision tree into a model: 。
  5. 5. The method for combined prediction-based photovoltaic power value-oriented prediction as set forth in claim 4, wherein said random forest RF model is, for a sample set Firstly, extracting k samples from a sample set by using a bootstrap sampling method to construct a corresponding decision tree, secondly, randomly extracting from an attribute set of decision tree nodes, splitting the decision tree nodes to construct a single classification and regression tree, sequentially accessing leaf nodes in the single regression tree according to a threshold sequence, taking the average value of all the leaf nodes as a predicted value, and then repeating the two steps to construct a random forest, and taking the average value of the output results of the classification and regression tree to obtain a final result.
  6. 6. The method for guiding prediction of photovoltaic power value based on combined prediction as set forth in claim 5, wherein said multi-layer perceptron MLP model comprises an input layer, a hidden layer, an output layer, and a sample set of samples The input layer contains m nerve nodes, and the hidden layer transforms the data transmitted by the input layer: Wherein: As the weight coefficient, b 1 is the bias, the common activation function Including ReLU, sigmoid; The number of neurons of the output layer is the same as the dimension of the output variable, and the hidden layer transmission data is transformed again: 。
  7. 7. the method for combined prediction-based photovoltaic power value-oriented prediction as claimed in claim 6, wherein the K nearest neighbor regression KNN model is that for a sample set Distance between input feature amounts is defined as Euclidean distance: 。
  8. 8. The method for guiding prediction of photovoltaic power value based on combined prediction according to claim 1, wherein in step 1, the combined prediction means that two or more basic prediction models are combined by weight coefficients, and a prediction result is obtained by a weighted summation mode: Wherein: n is the number of basic prediction models; And Respectively the weight coefficient and the prediction result of the ith prediction model, wherein The requirements are as follows: Model to be predicted Conversion to the combined predicted form: Wherein: and expressing the photovoltaic predicted value of the ith predicted model in the s-th sample, substituting the combined predicted model into the predicted-scheduled double-layer problem to obtain: The expression (13) transforms the prediction model into a combined prediction form, and the optimizing parameter is a weight coefficient of the combined prediction.
  9. 9. The method for guiding prediction of photovoltaic power value based on combined prediction according to claim 1, wherein the self-organizing map SOM algorithm in step 2 is composed of an input layer and a competition layer, wherein the neighborhood of the winning neuron is determined by centering on the winning neuron and the weights of the neurons in the neighborhood are calculated, and the stable classification of data is realized by updating a network, and the training steps are as follows: 1) Data preprocessing, namely cleaning and screening characteristic variables, and inputting characteristics Normalizing, wherein N is the dimension of the input feature; 2) Network initialization, setting the maximum value of iteration times as K and parameter learning rate And neuron neighborhood All decrease with the increase of the iteration number k, the competing layer neuron is M, and the competing layer neuron is weighted Assigning a random initial value; 3) Neuron competition, namely calculating the distance between an input characteristic vector and all neuron weight vectors of a competition layer, wherein the smallest distance is a winning neuron c, and the corresponding weight vector is The Euclidean distance is used as a distance measurement index: 4) Updating the network weights and input variables, namely updating the neuron weights in the neighbor of the winning neuron: the weight function h ci is defined as follows: Selecting the next training sample as a network input variable, and returning to the step 3) until the sample training set is traversed; 5) And (3) judging convergence, namely judging whether the maximum iteration number K is reached or judging that the two iteration errors are smaller than a set value, and returning to the step (3) if the maximum iteration number K is not reached.
  10. 10. The method for guiding prediction of photovoltaic power value based on combined prediction according to claim 1, wherein the input features of classification of the self-organizing map SOM algorithm in step 2 are photovoltaic predicted values and electricity price predicted values, and after classification, combined prediction model parameter training is performed on each cluster respectively: Wherein: representing the category to which the sample belongs, classify () representing a classified SOM network, x being an input variable, using a photovoltaic predicted value and an electricity price predicted value as characteristic values.
  11. 11. The photovoltaic power value-oriented prediction method based on combined prediction according to claim 1, wherein the hunger game search HGS algorithm in step 3 is a group intelligence algorithm simulating hunger creatures for food search, and the hunger game search HGS algorithm simulates creature group cooperative predation: Wherein R 1 、r 2 is a random number which is uniformly distributed in a section [0,1], randn (1) is a random number which is subjected to standard normal distribution, t represents iteration times, X b represents individual optimal positions, namely optimal weight coefficients to be searched, X (t) represents individual positions in a t-th round of iteration, namely combined prediction weights adopted by the round of iteration, l is a super parameter, E represents a search position control variable, a hyperbolic secant function is used for calculating the distance between the current search weight and the optimal weight, R represents a biological search range, R (i) E [ -1,1] is gradually reduced to 0;W 1 along with the increase of iteration times, and W 2 represents starvation weights, wherein the calculation formulas are respectively as follows: Wherein F (i) represents the fitness of the ith organism, BF represents the optimal fitness, namely the objective function value of formula (13), r 3 、r 4 、r 5 and rand are random numbers uniformly distributed in interval [0,1], N represents the set population size, and hungry (i) has the following calculation formula: wherein TH and LH are superparameter to represent upper and lower limits of starvation value, UB and LB represent upper and lower limits of search space, namely weight coefficient optimizing range, and are set to be 0,1, r 6 is interval 0,1 to uniformly distribute random numbers, and BF represents the worst fitness.

Description

Photovoltaic power value guiding prediction method based on combined prediction Technical Field The invention belongs to the technical field of photovoltaic power generation power prediction and power system optimization scheduling, and particularly relates to a photovoltaic power value guiding prediction method based on combination prediction. Background In recent years, the permeability of the photovoltaic power generation in the power system is gradually improved, but the randomness and fluctuation of the output power of the photovoltaic system can have adverse effects on the aspects of renewable energy consumption, safe and stable operation of the power system and the like. The photovoltaic power prediction can assist the photovoltaic power station to make a power generation plan, optimize a scheduling scheme and participate in an energy market, so that the utilization rate and the economical efficiency of renewable energy sources are improved. The photovoltaic power storage station is used as an important distributed energy form, and the next day of power generation power needs to be declared to the power dispatching department according to the photovoltaic day-ahead prediction result. The traditional photovoltaic power prediction is guided by precision, and has the characteristics of long time span and high uncertainty due to the fact that the photovoltaic day-ahead prediction is low in prediction accuracy, so that errors exist between the actual generated energy of the photovoltaic and the declared electric quantity. The error not only reduces the operation income of the power station, but also affects the stable operation of the power grid. However, the operation benefits of the optical storage power station and the photovoltaic prediction errors are in a nonlinear and asymmetric relation, so that in order to further improve the operation benefits of new energy manufacturers such as the optical storage power station, the value-oriented prediction of the comprehensive new energy prediction and scheduling model is necessary, the photovoltaic prediction model is trained by taking the maximum operation benefits of the new energy manufacturers as the target, and the commercial value of the optical storage power station is further mined under the condition that the prediction accuracy is difficult to be greatly improved. Disclosure of Invention The invention aims to provide a combined prediction-based photovoltaic power value guiding prediction method, which uses combined prediction as an upper prediction model, reduces the parameter scale and optimizing range of the prediction model, effectively ensures the accuracy of the prediction model, takes the maximum operation income as guiding training photovoltaic prediction model, and remarkably improves the economic benefit of a photovoltaic power station. In order to achieve the above purpose, the present invention adopts the following technical scheme: a photovoltaic power value oriented prediction method based on combined prediction, comprising: Step 1, constructing a photovoltaic power combination prediction model, wherein the combination prediction model is formed by combining a plurality of basic prediction models in a weighting manner; step 2, clustering photovoltaic output and electricity price data by adopting a self-organizing map (self-organizing map, SOM) algorithm, and dividing a sample into a plurality of categories; Step 3, optimizing the weight coefficient of the combined prediction model by adopting a hunger game search (hunger GAMES SEARCH, HGS) algorithm according to each category and with the maximum operation income of the optical storage station as an objective function; And 4, carrying out combined prediction on the photovoltaic power according to the optimized weight coefficient, and outputting a value-oriented photovoltaic power prediction result. Further, the step 1 specifically includes: And 1.1, training a support vector machine regression (SVR) basic model, namely, training a structural experience risk minimization with the support vector machine principle as a function set, wherein the specific mechanism is to project low-dimensional sample data into a high-dimensional space, and searching a classification hyperplane with highest classification precision and farthest distance from a sample point. For sample sets First, input variablesTransforming to high-dimensional feature spaceThe classification hyperplane can be expressed as: because the hyperplane requires the correct classification of the sample, constraints are satisfied The classification interval between the hyperplane and the sample point is as followsThe hyperplane construction problem may be translated into an optimization problem: By introducing a insensitive loss function, the support vector machine can be extended to solve the regression problem, noted as : C is a constant for balancing the loss value and the regularization term; is a insensitive loss coefficient. A