Solar Energy Forecasting Based on Complex Valued Auto-encoder and Recurrent Neural Network

Renewable energy is becoming a trusted power source. Energy forecasting is an important research field, which is used to provide information about the future power generation of renewable energy plants. Energy forecasting helps to safely manage the power grid by minimizing the operational cost of energy production. Recent advances in energy forecasting based on deep learning techniques have shown great success but the achieved results still too far from the target results. Ordinary deep learning models have been used for time series processing. In this paper, a complex-valued autoencoder was coupled with an LSTM neural network for solar energy forecasting. The complex-valued autoencoder was used to process the time series with the advantage of processing more complex data with more input arguments. The energy value was used as a real value and the weather condition was considered as the imaginary value. Taking into account the weather condition helps to better predict power generation. The proposed approach was evaluated on the Fingrid open data dataset. The mean absolute error (MAE), rootmean-square error (RMSE) and mean absolute percentage error (MAPE) was used to evaluate the performance of the proposed method. A comparison study was performed to prove the efficiency of the proposed approach. Reported results have shown the efficiency of the proposed approach. Keywords—Solar energy forecasting; artificial intelligence; complex-valued autoencoder; long-short term memory; deep