Investigation of LSTM-based hybrid approaches for short-term PV power forecasting

Over the recent years, short-term forecasting models have been proposed and developed in order to deal with continuous penetration of PV power into power systems and maintain their optimal control and operation. Due to the development in technology and the advance of machine learning, hybrid approaches have been considered more and more for the development of power forecasting methodologies, since they are able to optimally utilize the advantages of different methods and thus solve different problems in the forecasting process and improve the total accuracy. This paper introduces and investigates two different hybrid approaches based on the Long Short-Term Memory (LSTM) network and the Auto-Regressive Integrated Moving Average (ARIMA) method. The evaluation of the proposed hybrid approaches considers a seasonal framework and aims to analyse their behavior under the daily weather fluctuations. The proposed Hybrid model 2 is able to outperform all benchmark models considered, improving the forecasting error by 5.79% and 23.79% compared to the LSTM and ARIMA models, respectively, while the proposed Hybrid model 1, manages to surpass the ARIMA model by 18.46%. Both hybrid approaches are able to maintain a low computational burden despite the increase in the complexity of their architecture.