Ultra-Short-term Photovoltaic Output Power Forecasting using Deep Learning Algorithms

Abstract

Photovoltaic (PV) is becoming an attractive alternative in Norway in new zero-emission housing projects and in connection with hydropower reservoirs. However, fast-moving clouds result in abrupt changes in PV output power, which makes grid integration in such areas more challenging. One solution is to forecast the amount and variation of PV output power in advance. This paper, therefore, evaluates the performance of various DL (Deep Learning)-based forecasting models for a 20.15 kWp PV plant in Trondheim, Norway. The results show that a forecast model based on LSTM (Long Short-term Memory) network gives better performance in terms of RMSE (Root Mean Squared Error) for 15 minutes ahead forecast. This study can serve as the groundwork for future research into techniques and approaches that can result in a high-performing forecast model both in terms of accuracy and stability for the Norwegian climate.