A photovoltaic power generation power prediction method based on a neural network

Abstract

The application discloses a photovoltaic power generation power prediction method based on a neural network, first collects various data of a photovoltaic power station; carries out cleaning and conversion preprocessing on the collected data, and divides the processed data into a training set, a verification set and a test set; a distance threshold and a density threshold are introduced on the basis of an original SMOTE algorithm, the original SMOTE algorithm is improved, and then the improved SMOTE algorithm is used for sample expansion on the collected data; an extreme learning machine model is optimized through fusion of an ant colony optimization algorithm and a niche algorithm, and an extreme learning machine classifier for power generation power prediction is constructed; actual photovoltaic power generation power is predicted by using the constructed extreme learning machine classifier and dataset samples, and a classification result is a plurality of intervals of the predicted power generation power. The method can effectively improve the prediction accuracy of the power generation power by optimizing the extreme learning machine classifier through the ant colony optimization algorithm and the niche algorithm.

Description

Technical Field

[0001] This invention relates to the field of photovoltaic power generation prediction technology, and in particular to a method for predicting photovoltaic power generation based on neural networks. Background Technology

[0002] With increasing public awareness of environmental issues, the utilization of renewable energy has received unprecedented attention. Photovoltaic power generation, as a clean and renewable energy source, has been widely applied globally. However, the power output of photovoltaic (PV) power generation is affected by many factors, such as weather conditions (including solar radiation, cloud cover, and temperature), season, and sunshine duration. These factors lead to significant uncertainty in PV power output. To address this issue, accurate prediction of PV power generation is necessary for better planning and control of the power system, improving grid stability and operational efficiency. Accurate power forecasting also helps improve the economic benefits of PV power plants, as energy market participants can trade electricity based on the forecast results, avoiding economic losses caused by power output fluctuations.

[0003] However, photovoltaic (PV) power generation prediction is a highly complex task. Firstly, numerous and interconnected factors influence PV power generation, such as weather conditions, equipment status, and geographical location. These factors constitute a complex nonlinear system, making power prediction difficult. Secondly, PV power generation data often contains noise and missing values, further increasing the difficulty of prediction. Furthermore, PV power generation data is typically large-scale and high-dimensional, placing high demands on the computational power and memory requirements of the models. Existing technologies generally suffer from insufficient prediction accuracy, low data utilization efficiency, and inadequate model stability and generalization ability. Summary of the Invention

[0004] The purpose of this invention is to provide a method for predicting photovoltaic power generation based on neural networks. This method utilizes ant colony optimization algorithm and niche algorithm to optimize the Extreme Learning Machine (ELM) classifier, which can effectively improve the prediction accuracy of power generation.

[0005] The objective of this invention is achieved through the following technical solution:

[0006] A method for predicting photovoltaic power generation based on neural networks, the method comprising:

[0007] Step 1: First, collect various data from the photovoltaic power station, including weather data, voltage, current, and power data of the photovoltaic panels;

[0008] Step 2: Clean and transform the collected data, and divide the processed data into training set, validation set and test set according to a certain ratio for model training, optimization and testing;

[0009] Step 3: Based on the original SMOTE algorithm, a distance threshold and a density threshold are introduced to improve the original SMOTE algorithm. Then, the improved SMOTE algorithm is used to augment the collected data.

[0010] Step 4: Optimize the Extreme Learning Machine model by fusing the ant colony optimization algorithm and the niche algorithm to construct an Extreme Learning Machine classifier for power generation prediction;

[0011] Step 5: Use the constructed Extreme Learning Machine classifier and the dataset samples from Step 3 to predict the actual photovoltaic power generation. The classification result is several intervals of the predicted power generation.

[0012] As can be seen from the technical solution provided by the present invention, the above method utilizes the ant colony optimization algorithm and the niche algorithm to optimize the extreme learning machine (ELM) classifier, which can effectively improve the prediction accuracy of power generation. Attached Figure Description

[0013] To more clearly illustrate the technical solutions of the embodiments of the present invention, the drawings used in the following description of the embodiments will be briefly introduced. Obviously, the drawings described below are only some embodiments of the present invention. For those skilled in the art, other drawings can be obtained based on these drawings without creative effort.

[0014] Figure 1 A schematic diagram of the process for predicting photovoltaic power generation based on neural networks provided in an embodiment of the present invention;

[0015] Figure 2 This is an architecture diagram of the Extreme Learning Machine classifier described in an embodiment of the present invention. Detailed Implementation

[0016] The technical solutions of the embodiments of the present invention will be clearly and completely described below with reference to the accompanying drawings. Obviously, the described embodiments are only some embodiments of the present invention, and not all embodiments, and do not constitute a limitation of the present invention. Based on the embodiments of the present invention, all other embodiments obtained by those skilled in the art without creative effort are within the protection scope of the present invention.

[0017] like Figure 1 The diagram shows a flowchart of a photovoltaic power generation prediction method based on a neural network provided in an embodiment of the present invention. The method includes:

[0018] Step 1: First, collect various data from the photovoltaic power station, including weather data, voltage, current, and power data of the photovoltaic panels;

[0019] In this step, the weather data is collected through meteorological bureaus or meteorological sensors, including temperature, humidity, wind speed, and sunshine duration data.

[0020] The voltage, current, and power data of the photovoltaic panel are obtained by connecting to sensors, or by directly connecting to an electricity meter or inverter.

[0021] Other data collection: If it is necessary to consider the impact of other factors on photovoltaic power generation, such as cloud cover and air pollution, data can be collected through appropriate sensors or monitoring equipment.

[0022] Step 2: Clean and transform the collected data, and divide the processed data into training set, validation set and test set according to a certain ratio for model training, optimization and testing;

[0023] In this step, the cleaning preprocessing is to clean the data and remove invalid, duplicate, or abnormal data.

[0024] Transformation preprocessing involves standardizing and normalizing data to better facilitate its application in machine learning algorithms.

[0025] Step 3: Based on the original SMOTE (Synthetic Minority Over-sampling Technique) algorithm, a distance threshold and a density threshold are introduced to improve the original SMOTE algorithm. Then, the improved SMOTE algorithm is used to augment the collected data.

[0026] In this step, the original SMOTE algorithm proceeds as follows: Assuming there is an original dataset of size N, each data point contains d features. The goal of the original SMOTE algorithm is to generate k synthetic samples, where k is a pre-set parameter. For each original data point, the original SMOTE algorithm randomly selects one of its k nearest neighbors and then generates a synthetic sample according to the following formula: x new =x i +λ C (x zi -x i )

[0027] Where x i These are the original data points; x zi It is x i Randomly selected nearest neighbor; λ CIt is a random number between 0 and 1;

[0028] To improve the performance and reliability of the SMOTE algorithm, distance thresholds and density thresholds were introduced based on the original SMOTE algorithm. Specifically:

[0029] Density is used to define the importance of samples; the higher the density of a sample, the greater its importance. A distance threshold is introduced and set to t. d The density threshold is set to t s For each original data point x i Calculate x i Find the k nearest neighbors in d-dimensional space and calculate their density ρ. i :

[0030]

[0031] Where d ij It is data point x i and x j The Euclidean distance between them; I(·) is an indicator function, which takes the value 1 when the condition in parentheses is true, and 0 otherwise; if ρ i >t s Then the data point x i They are called core samples; otherwise, they are called marginal samples.

[0032] When generating a new sample, first randomly select data point x. i A nearest neighbor x zi Then calculate x i and x zi The distance d between zi If d zi >t d Then select a new nearest neighbor x. zi If it passes through max t If, after ties of re-selecting the nearest neighbor, the condition still cannot be met, then the generation of a new sample is abandoned; where max t ries is a pre-set hyperparameter;

[0033] To ensure that newly generated samples do not affect edge samples in the original dataset, a density threshold t is introduced. s If data point x i If the data point x is an edge sample, then only one new sample is generated; if the data point x is an edge sample, then only one new sample is generated. i If the sample is a core sample, then generate k new samples, where k is calculated by the following formula:

[0034]

[0035] Here, γc is a pre-set parameter, and 1 < γc < 100, used to control the number of new samples generated; int() is the integer function;

[0036] For each new sample, first randomly select data point x. i A nearest neighbor x zi Then, generate new samples according to the following formula:

[0037] x new =x i +λu(x zi -x i )

[0038] Where λu is a random number between 0 and 1;

[0039] If the newly generated sample x new With data point x i The distance d between zc Greater than the distance threshold t d If the maximum value has been reached, then a new nearest neighbor needs to be selected; if the maximum value has been reached... t If, after ries reselecting the nearest neighbor, a nearest neighbor that still meets the conditions is still not found, then the generation of a new sample is abandoned.

[0040] The improved SMOTE algorithm also introduces a new parameter max. d ensity is used to control the density threshold used when generating new samples. Specifically:

[0041] First, calculate x for all original data points. i density ρ i Sort them according to density from largest to smallest; then select the top-ranked maximum density. d The data points of ensity are used as core samples; the purpose of this is to ensure that the newly generated samples are more distributed in high-density areas, thereby improving the performance and reliability of the algorithm.

[0042] Different density thresholds are used when generating new samples. Specifically, the density threshold t of the new sample is calculated using the following formula. s :

[0043]

[0044] Here, γp is a pre-defined parameter, and 1 < γp < N, where N is the size of the original dataset. The purpose of this is to make the generated new samples more closely fit the density distribution of the dataset, thereby improving the performance and reliability of the algorithm.

[0045] With the above improvements, the improved SMOTE algorithm can more accurately simulate the density distribution of the original dataset, thereby generating more reasonable new samples. At the same time, it can also control the number of new samples generated by adjusting parameters, thus better meeting the needs of practical applications.

[0046] Step 4: Optimize the Extreme Learning Machine model by fusing the ant colony optimization algorithm and the niche algorithm to construct an Extreme Learning Machine classifier for power generation prediction;

[0047] In this step, the input training data for the extreme learning machine is first optimized based on a clustering sampling strategy. Specifically:

[0048] Let T be the dataset input to the Extreme Learning Machine, where m are the number of samples and n are the number of features. The k-means algorithm is used to cluster the dataset T, and the clustering result is expressed as:

[0049] S = k-means(T, K)

[0050] Where S = {S1, S2, ..., S} K} represents the clustering result; K is the number of clusters;

[0051] Then from each cluster S k If p samples are sampled from the sample, and the sampling proportion is α, then p = α|S k |, where |S k | is clustering S k The sample size; the sampling result is T′ k =sample(S k ,p), where T′ k From cluster S k The sample set obtained by sampling from the middle; combining the sampling results, the input training data T′ of the optimized extreme learning machine is obtained, denoted as:

[0052]

[0053] Among them, S k It is the kth cluster; sample(S) k p) is from cluster S k The sample set obtained by sampling p samples from the middle;

[0054] Use T′ instead of T for training the Extreme Learning Machine;

[0055] This invention optimizes the Extreme Learning Machine (ELM) model by integrating the Ant Colony Optimization (ACO) algorithm and the Niche Algorithm. The ACO algorithm is a metaheuristic optimization algorithm that simulates the foraging behavior of ant colonies in nature. This invention uses this algorithm to select the number of hidden layer neurons in the ELM. Specifically, the number of hidden layer neurons in the ELM is defined as the solution space of the problem. Each ant selects a path in the solution space based on pheromone intensity and heuristic information. Each path corresponds to a number of hidden layer neurons. The ant colony searches in the solution space, and the concentration of pheromones is updated based on the feedback of classification performance to guide the search.

[0056] The niche algorithm is a metaheuristic algorithm that simulates an individual organism searching for the optimal living space in its environment. This invention uses this algorithm to optimize the output layer weights of an extreme learning machine. Specifically, the output layer weights of the extreme learning machine are defined as the solution space of the problem, and each individual corresponds to a set of output layer weights. Each individual searches in the solution space based on its fitness, which is the classification performance of the extreme learning machine. The optimal living space corresponds to the optimal output layer weights.

[0057] like Figure 2 The diagram shown is an architecture diagram of the Extreme Learning Machine classifier according to an embodiment of the present invention, including an input layer, a hidden layer, and an output layer. Based on this architecture, the specific process of optimizing the Extreme Learning Machine model by fusing ant colony optimization algorithm and niche algorithm is as follows:

[0058] 1) Initialize the hidden layer weight matrix W and bias vector b of the Extreme Learning Machine.

[0059] Suppose that the feature matrix obtained after feature extraction of the text is X, where m is the number of samples and n is the number of features; and the number of neurons in the hidden layer is Ls, then the hidden layer weight matrix W and the bias vector b can be randomly initialized as follows:

[0060] W = rand(n, Ls) b = rand(Ls)

[0061] 2) Use ant colony optimization algorithm to select the number of neurons in the hidden layer.

[0062] Assume the size of the ant colony is N. a Then each ant a i The corresponding number of neurons L in a hidden layer i L i Initialize to random values; in each iteration, each ant selects a new number L of hidden layer neurons based on pheromone concentration and heuristic information. i And calculate the classification performance P of the extreme learning machine. i The pheromone concentration τ is based on P i renew:

[0063]

[0064] Where ρ is the pheromone volatility coefficient; δ(L i ) is the Kronecker function, when L i When it is the current optimal solution, δ(L′) i ) = 1, otherwise 0;

[0065] 3) Utilize niche algorithms for multi-objective optimization to optimize the output layer weights of the extreme learning machine.

[0066] Assume the population size is N p Then each individual corresponds to a set of output layer weights β i ,β i Initialize to random values; in each iteration, each individual searches for new output layer weights β′ in the solution space based on its fitness. i And calculate the classification performance P of the extreme learning machine. i ′, individual fitness f i According to P i 'renew:

[0067] f i =P i ′

[0068] Finally, the output layer weights with the highest fitness are selected as the output layer weights of the extreme learning machine.

[0069] 4) Adaptive learning rate adjustment

[0070] Traditional extreme learning machines typically use a fixed learning rate, which can lead to slow convergence or getting stuck in local optima in some cases. Therefore, this invention proposes an adaptive learning rate strategy to dynamically adjust the learning rate according to the model's training progress. The specific steps are as follows:

[0071] First, initialize the learning rate η0 and the learning rate decay coefficient δ; at the beginning of each iteration, calculate the learning rate η for the current iteration. t :

[0072]

[0073] Using η t Instead of a fixed learning rate for model training, η is used in the weight update formula for the Extreme Learning Machine. t Replacing the original learning rate, it is expressed as:

[0074]

[0075] Where, β t It is the weight of the t-th iteration; The loss function L with respect to β t The gradient;

[0076] 5) Use the optimized Extreme Learning Machine for text classification.

[0077] Introducing an attention mechanism, we define the attention weight as W. a And w a Depend on Composition; the number of neurons in the hidden layer is Ls, a i It represents the attention weight of the i-th hidden layer neuron;

[0078] Input feature matrix X, calculate hidden layer output H:

[0079] H = σ(XW + b)

[0080] Further calculate attention score s i :

[0081]

[0082] Where b a It is a bias; h i It is the i-th column of H; tanh is the activation function;

[0083] Then calculate the attention weight a i :

[0084]

[0085] Using attention weight a i Adjust the hidden layer output H to obtain the new hidden layer output H′:

[0086] H′=H⊙a

[0087] Where ⊙ represents the Hadamarda complex;

[0088] Replace the original H with the new hidden layer output H′, and calculate the output layer output 0:

[0089] O=H′β

[0090] Where β is the output layer weight; 0 represents the classification result of the Extreme Learning Machine.

[0091] Step 5: Use the constructed Extreme Learning Machine classifier and the dataset samples from Step 3 to predict the actual photovoltaic power generation. The classification result is several intervals of the predicted power generation.

[0092] In this step, the data to be predicted is input into the constructed Extreme Learning Machine (ELM) classifier after preprocessing and feature extraction. The ELM classifier outputs various intervals of the actual photovoltaic power generation prediction, including:

[0093] Category 1: Very low power, 0-10% of the maximum power range; for example, if the maximum power is 1000W, then the range for this category is 0-100W.

[0094] Category 2: Low power, 10-30% of maximum power range;

[0095] Category 3: Medium power, 30-60% of maximum power range;

[0096] Category 4: High power, 60-90% of maximum power range;

[0097] Category 5: Very high power, 90-100% of maximum power range.

[0098] As can be seen from the above scheme, the predicted power generation results are divided into several intervals, which improves the flexibility and practicality of the prediction.

[0099] It is worth noting that the contents not described in detail in the embodiments of the present invention belong to the prior art known to those skilled in the art.

[0100] In summary, the method described in the embodiments of the present invention has the following advantages:

[0101] 1. Improve prediction accuracy: Optimizing the Extreme Learning Machine (ELM) classifier using ant colony optimization algorithm and niche algorithm can effectively improve the prediction accuracy of power generation. Ant colony optimization algorithm is used to select the number of neurons in the hidden layer, while niche algorithm is used to optimize the weights of the output layer.

[0102] 2. Improve data utilization: In the preprocessing stage, by adopting K-means clustering and cluster-based sampling strategies, training data can be utilized more effectively, thereby improving the learning efficiency of the model.

[0103] 3. Enhance model stability and generalization ability: Introducing attention mechanisms and adaptive learning rate strategies can improve model stability and generalization ability, and enhance model performance on different datasets.

[0104] Furthermore, those skilled in the art will understand that all or part of the steps in the methods of the above embodiments can be implemented by a program instructing related hardware, and the corresponding program can be stored in a computer-readable storage medium, such as a read-only memory, a disk, or an optical disk.

[0105] The above description is merely a preferred embodiment of the present invention, but the scope of protection of the present invention is not limited thereto. Any variations or substitutions that can be easily conceived by those skilled in the art within the scope of the technology disclosed in the present invention should be included within the scope of protection of the present invention. Therefore, the scope of protection of the present invention should be determined by the scope of the claims. The information disclosed in the background section is intended only to enhance the understanding of the overall background technology of the present invention and should not be construed as an admission or implication in any way that such information constitutes prior art known to those skilled in the art.

Claims

1. A neural network-based photovoltaic power generation power prediction method, characterized by, The method comprises: Step 1, first collect various data of photovoltaic power station, the data includes weather data, photovoltaic cell panel voltage, current, power data; Step 2, the collected data is cleaned and converted for pretreatment, and the processed data is divided into training set, verification set and test set according to a certain proportion, which is used for model training, optimization and test; Step 3, on the basis of the original SMOTE algorithm, distance threshold and density threshold are introduced, and the original SMOTE algorithm is improved, and then the improved SMOTE algorithm is used for sample expansion of the collected data; In step 3, the process of the original SMOTE algorithm is as follows: assuming that there is an original data set with a size of N, each data point contains d features, and the target of the original SMOTE algorithm is to generate k synthetic samples, wherein k is a pre-set parameter, for each original data point, the original SMOTE algorithm randomly selects one point in the k nearest neighbors, and then generates a synthetic sample according to the following formula: ； wherein is the original data point; is a randomly selected nearest neighbor; is a random number between 0 and 1; On the basis of the original SMOTE algorithm, distance threshold and density threshold are introduced, and the original SMOTE algorithm is improved, and then the improved SMOTE algorithm is used for sample expansion of the collected data; Using density to define the importance of a sample, the higher the density of a sample, the more important it is, the distance threshold introduced is set to , the density threshold is set to , for each original data point , calculate k-nearest neighbors in d-dimensional space, and calculate its density : ； wherein is the Euclidean distance between data points and ; is an indicator function that takes the value 1 if the condition in the brackets is true, and 0 otherwise; if then the data point is called a core sample, otherwise it is called a margin sample; When generating a new sample, first randomly select a data point One of the nearest neighbors Then calculate the distance Between And If , then reselect a nearest neighbor If after Times of reselecting the nearest neighbor, the condition is still not met, then give up generating a new sample; wherein Is a hyperparameter set in advance; To ensure that the newly generated samples do not affect the edge samples in the original dataset, a density threshold is introduced If the data point is an edge sample, only one new sample is generated; if the data point is a core sample, k new samples are generated, where k is calculated by the following formula: ； wherein, is a pre-set parameter, and for controlling the number of new samples generated; is a rounding function; For each new sample, first randomly select a data point one of the nearest neighbors and then generate the new sample according to the following equation: ； wherein is a random number between 0 and 1; If a new sample is generated With data points Distance between Greater than the distance threshold If the nearest neighbor needs to be selected again; if it has been passed through If, after reselecting the nearest neighbor, a nearest neighbor that still meets the conditions is still not found, then the generation of a new sample is abandoned. Wherein, the improved SMOTE algorithm also introduces a new parameter , which is used to control the density threshold used when generating new samples, specifically: First, calculate all the original data points. density Sort them according to density from largest to smallest; then select the top density. The data points are used as core samples; A different density threshold is used when generating new samples, specifically the density threshold of a new sample is calculated using the following formula : ； wherein is a pre-set parameter, and N is the size of the original data set; Step 4, the extreme learning machine model is optimized by fusing ant colony optimization algorithm and niche algorithm, and an extreme learning machine classifier for power generation power prediction is constructed; Step 5, the actual photovoltaic power generation power is predicted by using the constructed extreme learning machine classifier and the data set samples in step 3, and the classification result is a plurality of intervals of the predicted power generation power.

2. The neural network-based photovoltaic power generation power prediction method according to claim 1, characterized in that, In step 1, the weather data is collected by meteorological bureau or meteorological sensor, including temperature, humidity, wind speed and sunshine duration data; The voltage, current and power data of the photovoltaic cell panel are obtained by connecting the sensor, or collected by directly connecting the electric meter or the inverter. 3.The neural network-based photovoltaic power generation power prediction method according to claim 1, characterized in that, In step 2, the cleaning pretreatment is to clean the data, remove invalid data, duplicate data or abnormal data; The conversion pretreatment is to standardize and normalize the data, so as to better apply to machine learning algorithm.

4. The neural network-based photovoltaic power generation power prediction method according to claim 1, characterized in that, In step 4, first, the input training data of extreme learning machine is optimized based on clustering sampling strategy, specifically: Define the data set input into the extreme learning machine as T, wherein the sample number is m, the feature number is n, the k-means algorithm is used for clustering the data set T, and the clustering result is represented as: ； wherein is the clustering result; K is the number of clusters; Then sample p samples from each cluster Assume the sampling ratio is Then Where is the number of samples in cluster The sampling result is Where is the sample set sampled from cluster ​ The combination sampling result obtains the input training data of the optimized extreme learning machine is expressed as: ； wherein, is the kth cluster; is a sample set of p samples drawn from the cluster Ck. Using Instead of T, the training of the extreme learning machine is performed; And the extreme learning machine model is optimized by fusing ant colony optimization algorithm and niche algorithm, specifically: 1) initialize the hidden layer weight matrix W and bias vector b of extreme learning machine Assuming that the feature matrix obtained after feature extraction of the text is X, wherein m is the sample number, n is the feature number; the number of hidden layer neurons is Ls, then the hidden layer weight matrix W and bias vector b can be randomly initialized, represented as: ； 2) select the number of hidden layer neurons using ant colony optimization algorithm Assume the size of the ant colony is , then each ant corresponds to a hidden layer neuron number , initialized to a random value; each iteration, each ant selects a new hidden layer neuron number according to the pheromone concentration and heuristic information , and calculates the classification performance of the extreme learning machine , the pheromone concentration is updated: ； wherein is the volatility factor of the pheromone; is the Kronecker function, when is the current best solution, 0 otherwise. 3) use niche algorithm for multi-objective optimization to optimize the output layer weight of extreme learning machine Assume the size of the population is , then each individual corresponds to a set of output layer weights , initialized to random values; at each iteration, each individual searches for new output layer weights in the solution space according to its fitness , and calculates the classification performance of the extreme learning machine , the fitness of the individual is updated according to : ； Finally, select the output layer weight with the highest fitness as the output layer weight of extreme learning machine; 4) adaptive learning rate adjustment The adaptive learning rate adjustment is to dynamically adjust the learning rate according to the training progress of the model, and the specific steps are as follows: First initialize the learning rate and the learning rate decay coefficient ; at the beginning of each iteration, compute the learning rate for the current iteration : ； Using Instead of fixed learning rate for model training, for extreme learning machine weight update formula, replace the original learning rate with , expressed as: ； wherein, is the weight of the tth iteration; is the gradient of the loss function L with respect to is the gradient of the loss function L with respect to 5) Text classification using optimized extreme learning machine Introducing an attention mechanism, we define attention weights as follows: ,and Depend on Composition; the number of neurons in the hidden layer is Ls, It represents the attention weight of the i-th hidden layer neuron; Input feature matrix X, compute hidden layer output H: ； Further calculating the attention score : ； wherein is a bias; is the i-th column of H; is an activation function; Recalculating attention weights : ； with attention weights adjusting the hidden layer output H to obtain a new hidden layer output : ； wherein is a Hadamard product; With the new hidden layer output Instead of the original H, the output layer output O is calculated: ； wherein is the output layer weight; O is the classification result of the extreme learning machine.

Images