Very Short-Term Solar Power Forecasting Using a Frequency Incorporated Deep Learning Model

Abstract

This paper aims to forecast solar power in very short horizons to assist in real-time distribution system operations. Popular machine learning methods for time series forecasting are studied, including recurrent neural networks with Long Short-Term Memory (LSTM). Although LSTM networks perform well in different applications by accounting for long-term dependencies, they do not consider the frequency domain patterns, especially the low frequencies in the solar power data compared to the sampling frequency. The State Frequency Memory (SFM) model in this paper extends LSTM and adds multi-frequency components into memory states to reveal a variety of frequency patterns from the data streams. To further improve the forecasting performance, the idea of Fourier Transform is integrated for optimal selection of the frequency bands by identifying the most dominant frequencies in solar power output. The results show that although the SFM model with uniform frequency selection does not significantly improve upon the LSTM model, the proper selection of frequencies yields overall better performances than the LSTM and 27% better than the persistent forecasts for forecast horizons up to one minute. Furthermore, a predictive voltage control based on solar forecasts is implemented to demonstrate the superior performance of the proposed model.