| Under the background of global warming and the goal of "carbon peaking and carbon neutrality",it has become a basic trend to seek efficient clean energy to replace traditional fossil energy.Due to the advantages of large solar energy resources,wide distribution,and near zero emissions,photovoltaic power generation has attracted attention from countries around the world.However,due to the periodicity,volatility,and low energy density of photovoltaic power generation,it will have a significant impact on the power grid,seriously affecting its stability and safety.Therefore,accurate prediction of photovoltaic power generation is helpful for the reasonable scheduling of the power grid,ensuring the balance of power supply and load sides,and is of great significance for the safe operation of the power grid.This article systematically studies the photovoltaic power prediction method based on Convolutional Neural Network using actual operating data of DKASC photovoltaic power plants in Australia.Firstly,by analyzing the relationship between photovoltaic power and meteorological factors,appropriate variables are selected as inputs to the model.Preprocess the actual photovoltaic operation data and meteorological element data using the Laida criterion detect abnormal data,fill in missing data using K-nearest neighbor algorithm,and then identify meteorological factors that have a significant impact on photovoltaic power by analyzing the Pearson correlation coefficient and Spearman correlation coefficient between photovoltaic power and meteorological factors.Then,by designing different convolutional kernel sizes,convolutional kernel numbers,activation functions,and network parameters,the parameters of the CNN informer model are optimized and improved,and the optimal parameters are selected for learning and training.Secondly,in response to the issues of easy loss of information and low prediction accuracy for long time series using a single prediction model,the CNN informer model is used to predict photovoltaic power generation.Using CNN to extract key feature information,connect the fully connected layer of CNN with the input layer of the informer model to predict future power generation.Finally,a CNN-LSTM model based on similar day clustering was established,and the improved FCM clustering algorithm was used to partition different weather types in the dataset to reduce the impact of uncertainty factors on photovoltaic power prediction.Three types of similar days,sunny,cloudy,and rainy,were predicted and compared with the predicted results of the other three models.The main conclusions are as follows:(1)Selecting total horizontal radiation,diffuse horizontal radiation,temperature,and wind speed as inputs to the model reduces the dimensionality of variables and improves modeling efficiency.(2)The results show that when the convolution kernel size is 5 and the convolution kernel size is 3×3.When the activation function is Relu,the learning rate is 0.00005,and the batch size is 128,the MAE,MAPE,and RMSE of the model are the smallest,and the model reaches the optimal state.(3)Compared with traditional BP,LSTM,and GRU models,the CNN informer model reduces MAE by 63.3%,55.1%,and 51.3%,respectively;MAPE decreased by 61.2%,43.6%,and 10.4%,respectively;The RMSE decreased by 55.2%,47.4%,and 45.6%,respectively.The photovoltaic power prediction method combining CNN and informer models has more advantages compared to traditional models.(4)The proposed CNN-LSTM model has MAE,MAPE,and RMSE values of 0.487 k W,0.051 k W,and 0.529 k W on similar sunny days,respectively;The MAE,MAPE,and RMSE on similar cloudy days are 0.506 k W,0.058 k W,and 0.552 k W,respectively;The MAE,MAPE,and RMSE on similar rainy days are 0.516 k W,0.059 k W,and 0.647 k W,respectively.Compared with traditional BP,LSTM,and GRU models,the evaluation indicators are all optimal,fully verifying that the prediction accuracy of the proposed model in this paper is improved compared to traditional models under different similar days. |
PDF Full Text Request