Combining simple and less time complex ML models with multivariate empirical mode decomposition to obtain accurate GHI forecast

Abstract

Accurate solar irradiance forecasting requires several climatological variables whose seasonal variation affects the stability of the solar irradiance forecast. A less time complex ensemble model has been integrated with multivariate empirical mode decomposition (MEMD) to resolve the non-linearity and non-stationarity of meteorological variables into several intrinsic mode functions (IMFs). The ensemble model combines three base models (kNN-distance-based nonlinear model, DTR-tree-based nonlinear model, ridge-regularized linear model) to ensure good prediction accuracy with less time complexity. For better generalization and reliability of the model, the results have been validated for three Indian locations, viz. Roorkee, Greater Noida, and Gangtok. The MEMD-stacked model outperformed the modern and more complex ML techniques with a minimum % RMSE (% MAE) reduction of 48.51 (37.52) compared to RF, 47.43 (35.11) compared to ANN, and 46.46 (29.36) compared to LSTM. Moreover, in a seasonal analysis, the reported highest RMSE by the proposed model (53.57 W/m2) for the monsoon season was significantly less than that of LSTM (102.36 W/m2). This paper reflects the potential to combine simpler, less time-complex ML algorithms with MEMD to obtain more accurate GHI forecasts with lesser time complexity vis-à-vis the more popular complex ML models.