Research On Short-term Prediction Of Photovoltaic Power Generation Based On Deep Learning

An important measure to achieve "carbon peak and carbon neutrality" is to build a new power system with renewable energy as the main body.In recent years,solar,wind,tidal,biomass and other renewable energy generation have been vigorously deployed,especially small and large photovoltaic（PV）power stations have developed rapidly.However,due to the uncertainty and intermittensity of PV power generation,the smooth operation of the power system is a great challenge.Accurate prediction of PV power generation can better allocate power reserves and provide a reference for risk decisionmaking and energy management of microgrid.Therefore,this thesis conducts short-term PV power prediction based on deep learning,to provide better decision support for the power market.In this thesis,two PV power generation power prediction methods are designed,namely the CNN-BiGRU prediction model based on the Attention mechanism and the Attention-BiGRU prediction model based on the Whale Optimization Algorithm（WOA）.The main work of the thesis includes:First,preprocessing and correlation analysis are performed on the public dataset.Since it is impossible for PV power to be negative,outliers that are less than zero are replaced by 0;missing values are filtered out and filled with the value at the same time of the previous day;data of different unit dimensions are normalized and inverse normalized;Since the PV does not generate electricity at night or in the early morning,the PV power generation data from 5:00-19:00 in a day is selected.The feature importance was analyzed by the Pearson and Spearman correlation coefficients,the input variables of the prediction model were determined,and the purpose of dimensionality reduction for high-dimensional data was achieved.The influence of main factors such as Global Horizontal Irradiance,Global Tilt Irradiance,Extraterrestrial Irradiance,temperature and solar zenith angle on the prediction of PV power generation power is analyzed in detail.Secondly,a CNN-BiGRU neural network prediction model based on the Attention mechanism is designed.The spatial features of PV data are extracted by CNN,and these spatial features extracted by CNN are sent to the BiGRU neural network,and the temporal features of PV data are extracted by BiGRU.Then,the Attention mechanism is introduced to solve the problem of the drop in prediction accuracy caused by long sequences,highlighting the influence of strongly correlated features and reducing the influence of weakly correlated features.Experimental results show that Attention-CNNBiGRU has higher accuracy than prediction methods based on BP,GRU,BiGRU,Attention-BiLSTM,and Attention-BiGRU.Finally,an Attention-BiGRU neural network prediction model based on whale optimization algorithm is designed.Use WOA to optimize the learning rate,training times,batch size,number of hidden nodes of the first layer of BiGRU,number of hidden nodes of the second layer of BiGRU,and number of hidden nodes of the fully connected layer of the Attention-BiGRU neural network to improve the prediction accuracy of the model.The experimental results show that WOA-Attention-BiGRU has higher prediction accuracy than WOA-Attention-BiLSTM and Attention-CNN-BiGRU neural network models.And,the prediction accuracy of Attention-CNN-BiGRU is better than WOA-Attention-BiLSTM model.To sum up,this thesis uses Attention-BiGRU as the basic model to improve the prediction accuracy of its PV power generation power from two aspects: On the one hand,starting from deepening the neural network model structure,by adding a CNN layer in front of the Attention-BiGRU neural network,the spatial features of the PV sequence are extracted.On the other hand,we start with the swarm intelligence optimization algorithm.Because the WOA optimization algorithm has few parameters,simple operation and a strong ability to jump out of the local optimum,WOA is selected to solve the optimal solution of the Attention-BiGRU hyperparameters.