Enhancing Solar Radiation Forecasting with Hybrid Ensemble Empirical Mode Decomposition and Machine Learning Techniques

The importance of short-term solar radiation forecasting for power system use and management cannot be overstated. However, Non-stationarity and unpredictability make accurate forecasting difficult. Time series approaches are suitable for forecasting stationary time series derived from a non-stationary sequence. Empirical Mode Decomposition (EMD) and Ensemble Empirical Mode Decomposition (EEMD) are time-domain decomposition methods used to separate the components of the original solar radiation time series that have distinct time
¹
-scale characteristics. As a result, the components are forecasted using the Back Propagation Neural Network (BPNN) model, which is a simple and effective machine learning method. This paper presents a hybrid EMD/EEMD-BPNN model for predicting short-term solar irradiance. To identify unique data information at different time scales, EMD/EEMD first breaks up the solar radiation time series into several stationary or fundamental sub-series. The next step is to create BPNN models with specified parameters for each subsequence to forecast new ones. Finally, each subsequence's projected value is combined to generate the final forecast result. The accuracy of solar forecasts has greatly increased, especially when utilizing hybrid methods. Furthermore, using the proposed hybrid technique for multistep forecasting resulted in even more improvement. The basic BPNN model for 15 min time step achieves predicting with a root mean square error (RMSE) of roughly 416.04 W/m2 for 15min; however, this error decreases to 65.25 W/m2 with the EMD-Hybrid Model, and 32.86 W/m2 with the EEMD-Hybrid Model. In addition, the results of the skill score have proven clearly that EMD/EEMD-BPNN Model performs better compared to typical BPNN using several meteorological inputs