Univariate model for hour ahead multi-step solar irradiance forecasting

The present work considers the task of forecasting the GHI for 1-11h ahead forecast horizon from the previous GHI time-series data. Dealing with the non-stationary and non-linear time series is a challenging and vital part of time-series forecasting. This paper aims to forecast the GHI, comparing the performance of two time-series decomposition techniques, i.e., EMD and EEMD. These techniques divide the original time series into a set of orthogonal series, which are termed as Intrinsic Mode Functions (IMFs). Further, the problem is converted into a supervised learning problem (input-output problem). A feature selection algorithm selects the highly correlated lag variables to reduce the complexity of the model. Finally, the forecasting is performed by utilizing machine learning algorithms i.e., Support Vector Machine (SVM) and ensemble learning-based Random Forest (RF). The parameters of ML algorithms are optimized for each IMF, and the final result is obtained by summing all the forecasted IMFs. This work also discusses the annual and seasonal variation of the forecasting model's performance.