Forecasting of Solar Irradiances using Time Series and Machine Learning Models: A Case Study from India

Abstract

With the focus on renewable energy resources due to environmental reasons, reliable forecasting of renewable energy has great societal importance. This study focuses on the analysis and forecasting of GHI data at two different locations in India, namely Pokhran and Bitta. Since the GHI time series plots exhibit seasonality and randomness, the time series SARIMA model along with two machine learning models, namely MLP and LSTM, are implemented for daily, weekly and monthly forecasting. The efficacy of these competitive models is assessed using MAPE and RMSE values. We also perform residual analysis as a post processing step of the implemented models. For monthly forecasting, the SARIMA model has the best performance, as it precisely captures monthly seasonality in comparison to the machine learning models. However, for short term daily forecasting, machine learning models provide much better estimates with MLP as the most desirable one. Since the SARIMA model fails to fully capture the high amount of fluctuation (mostly, seasonal fluctuation) in the daily and weekly observations, we additionally implement an ARIMA model with sliding windows to improve modelling efficacy. The present study therefore provides a clear guideline on the selection of forecasting models based on the desired time horizon.