ENHANCED PERFORMANCE OF SOLAR IRRADIANCE PREDICTION USING DEEP LEARNING AND DATA MINING TECHNIQUES
Abstract
This work proposes two techniques to enhance predictive performance.
The first technique addresses implementation details regarding relevancy and redundancy measures, exploring how they may, respectively, be enhanced and mitigated. The proposed technique is a novel hybrid feature selection method built to optimize feature selection using a framework based on Least Redundant/Highest- Relevant, named Weather Recursive Feature Elimination (WRFE). The WRFE approach uses feature importance to measure reductions in variance in Random Forest Regression (RFR) in addition to data perturbation in LSTM. The training set’s optimal features demonstrate strong contributions to the prediction outcome, indicating the proposed WRFE’s generalizability for hourly GHI prediction. To lessen the seasonality effect, the second proposed technique employs a deep stack of the clustering connected layer with hybrid LSTM models. This novel Seasonal Clustering Forecasting Technique (SCFT) is then compared with other forecasting strategies, revealing its superiority.