Day-ahead Multi-step Photovoltaic Output Prediction Methods Integrating Multiple Features Based On Deep Learning

Exploring the adjustment and improvement of the energy structure is one of the hot issues in the energy field.Photovoltaic power generation has bright prospects as a way to harness solar energy.However,photovoltaic power generation presents the characteristics of intermittency and randomness due to the great influence of weather and other factors,which not only brings challenges to the stable operation and dispatch of the grid system,but also increases the difficulty in the consumption of photovoltaic power generation.Accurate short-term photovoltaic output forecast plays a key role in providing decision support for the dispatch of the grid system and solving the problem of photovoltaic power consumption to a certain extent.This paper designs a novel 1D CNN structure,and combines this structure with models based on seq2 seq to implement two day-ahead multi-step photovoltaic output prediction methods integrating multiple features.Both methods can effectively improve the short-term photovoltaic output prediction accuracy.The research work can mainly be concluded as follows:We analyze the relevant data of photovoltaic output forecast and study the fluctuation characteristics of photovoltaic output value in different seasons,different weather types,different times of the day and the main factors that affect the forecast of photovoltaic output.According to the analysis results,the corresponding time sequence number of a time,weather type,the historical average photovoltaic output features at that time in different seasons,different weather types,and different weather types of different seasons are integrated with numerical weather and historical photovoltaic output features to construct input data for our models.Then a day-ahead multi-step photovoltaic output prediction model based on 1D CNN-Transformer is designed,in which 1D CNN is used to capture the correlation between meteorological features,and Transformer is used to capture the time correlation between photovoltaic sequences.In the training process of the model,a new training method is introduced,which not only provides more valuable information to assist prediction but also allows Transformer to be truly parallel.Experimental results show that 1D CNN-Transformer model has higher prediction accuracy than models based on SVR,RFR,MLP,LSTM,and Bi-LSTM.Transformer will capture the two-way dependence between sequences in the prediction model based on 1D CNN-Transformer,while the dependence between photovoltaic sequences is one-way.In order to solve this problem,a day-ahead multistep photovoltaic output prediction model based on 1D CNN-Att-LSTM is further designed.This model employs the LSTM-based seq2 seq architecture and integrates the attention mechanism to serially capture the time correlation between photovoltaic sequences.The addition of the attention mechanism can also solve the problem of performance degradation caused by long sequences.The experimental results show that 1D CNN-Att-LSTM model has better results than 1D CNN-Transformer model,although it lags behind 1D CNN-Transformer model in training time.