A novel hybrid solar radiation forecasting algorithm based on discrete wavelet transform and multivariate machine learning models integrated with clearness index clusters

This study presents an innovative forecasting algorithm that combines multivariate regression (MR) and discrete wavelet transform (DWT) techniques with clearness index (CI)-based clustering methods to enhance short-term (1 h ahead) solar radiation forecasting. The proposed algorithm consists of two main steps: the first involves forecasting processes using DWT and MR methods, while the second includes clustering processes determined based on CI values. In the forecasting process, the data has been decomposed into sub-signals at different levels using DWT first. Multivariate ridge regression (MRR) and lasso regression (MLR) models for the sub-signals have been determined based on input training data sets created from three different combinations of these sub-signals. Sub-forecast signals have been obtained using models that were determined in different formats. The sub-forecast signals obtained have been recombined using the DWT reconstruction to produce the final forecasts. In the clustering process, clusters have been formed based on CI values using the Kernel k-means algorithm, which has been identified as the most effective among three different algorithms. The effectiveness of forecasts generated using DWT-MRR and DWT-MLR models for all input data set versions has been evaluated within the CI-based clusters.

The study's key findings have revealed that decomposition at the first level of DWT is sufficient to achieve optimal forecasting performance. Furthermore, the input variables yielding the best results have differed across clusters: radiation and relative humidity for the mostly cloudy cluster, radiation, temperature, and relative humidity for the cloudy cluster, and radiation and temperature for the slightly cloudy cluster. The results have demonstrated that the proposed algorithm achieves a 17% improvement in root mean square error (RMSE) compared to the best-performing model developed without CI clustering. The proposed approach significantly contributes to the literature by optimizing DWT decomposition levels, adapting data modeling to cloudiness conditions, and integrating multiple forecasting techniques to improve performance.