Day-Ahead Photovoltaic Power Forcasting Using Convolutional-LSTM Networks

Efficient and reliable forcasting of power generation from PV plants is important for grid management optimization and power dispatch allocation. In this paper, we propose a hybrid model for short-term photovoltaic power forcasting (PVPF), Convolutional-LSTM (Conv-LSTM), and extend its attention mechanism architecture and shortcut connection architecture, called Conv-LSTM-A and Conv-LSTM-S, respectively. The input data used are hourly data of commercial PV plants for one year and contain only four features: historical active power, solar radiation, panel temperature and current hour. The effects of different historical data lengths on the forcasting results are compared. The hyperparameters of the model were adjusted using bayesian optimization to find the model with the best hyperparameter configuration. The experimental results show that the RMSE of the Conv-LSTM-A model is 11.13% and the MAPE of the Conv-LSTM-S model is 7.0030 using 7 days of historical data for day-ahead PV power forcasting.