A time series based new hybrid approach for solar radiation forecasting

Short-term solar radiation forecasting is critically important for improving the efficiency of solar energy systems, optimizing energy production, and enhancing grid management. In particular, short-term forecasts enable rapid adaptation to atmospheric fluctuations such as cloud cover variations, thereby contributing to the balance between energy supply and demand. In the literature, numerous solar radiation forecast studies have been conducted using different datasets and techniques. Within the scope of this thesis, initial forecasts were performed using multivariate ridge regression (MRR) and multivariate lasso regression (MLR) models based on different combinations of hourly data collected at the Afyon Regional Station (ABİ), including radiation, average air temperature, and relative humidity (BRBA approach). In the second phase, the datasets were decomposed using various levels of one-dimensional discrete wavelet transform (DWT). Separate MRR and MLR models were then developed for each resulting sub-signal to perform forecasting (WRBA approach). Finally, the performance of the DWT-MRR and DWT-MLR models—constructed for various input dataset combinations—was evaluated across clusters defined by the clearness index (CI) using the kernel k-means algorithm. Within each cluster, the models yielding the most accurate forecasts were identified. A hybrid forecasting model was then developed by integrating the best-performing models in each cluster, resulting in the most accurate predictions overall (CA-WRBA approach). According to the proposed CA-WRBA approach results, the DWT-MLR-V1 model utilizing radiation and relative humidity data provided the best forecasts in the mostly cloudy (MC) cluster. In the cloudy (C) cluster, the DWT-MRR-V3 model, which used radiation, relative humidity, and average air temperature, yielded the most accurate results. For the slightly cloudy (SC) cluster, the DWT-MLR-V2 model was the most successful based on radiation and average air temperature. By hybridizing the best-performing models from each cluster, highly accurate forecasts were achieved for both six-month and one-year testing periods.