SIFA: A two-stage adaptive ensemble framework for solar irradiance forecasting using a wrapper-based feature selection and chaotic manta ray optimization

Engineering notes

Engineering notes will be added by the Power for Solar editorial team.

Chinese explanation / 中文解读

中文解读待补充：本站会优先为光伏效率、钙钛矿太阳能电池、储能技术、太阳能热利用、BIPV、并网技术等高价值论文补充中文说明。

Original abstract

Accurate solar irradiance forecasting is increasingly crucial for managing solar energy systems effectively, as their power output is directly dependent on solar irradiance (SI). Several models in the literature have been presented for SI forecasting; however, they still face at least one of these limitations: difficulties in modelling nonlinear data, demanding high computational resources, and often struggling to identify the best feature subsets for higher accuracy. To address these challenges, this study proposes a new multi-stage forecasting approach, termed SIFA, for accurate SI prediction, aiming to enhance the stability and efficiency of PV power plants. This approach comprises two main stages. The first stage employs a hybrid feature selection strategy combining random forest (RF) and sequential forward selection (SFS) to identify the most informative features. Specifically, SFS explores candidate feature subsets, while RF evaluates various subsets to select the most effective one. To further improve the RF performance, the number of estimators is tuned using an enhanced manta ray foraging optimizer, called IWMRFO, which employs chaotic maps instead of random generators to better balance exploration and exploitation, thereby avoiding local optima and accelerating convergence. The second stage combines three effective ML models-Huber Regressor (HR), Extra Trees (ET), and Extreme Gradient Boosting (XGB)-using a weight vector that is used to control the contribution of each base model in the hybrid ensemble approach (SIFA). This vector is optimized by the proposed IWMRFO, resulting in an adaptive ensemble that enhances predictive accuracy while maintaining stable generalization capability. This approach is tested on three popular datasets: the San Diego dataset, the Islamabad dataset, and the NASA SI dataset. Its performance is compared with several other models using several performance metrics, including RMSE, MAE, MAPE, MSE, and R². The numerical results demonstrate that SIFA achieved lower average forecasting errors than the competing models across the three evaluated datasets under repeated experiments, indicating that it is a strong alternative for predicting SI with higher accuracy.

Links and sources

• PDF from original source

Comments