A stacked generalization approach for day ahead hourly photovoltaic power forecasting

• Machine learning (ML) techniques have gained prominence in the solar power forecasting field since they can overcome the drawbacks associated with traditional models. • AI-based model, and hybrid methods can be developed to enhance the overall forecasting accuracy of single models. • The socio-economic and fundamental activities taking place throughout the zonal area may be negatively impacted due to insufficient power supply. • PV power forecasting model that resembles the short-term forecasting that support operational decision-making, facilitate grid stability, and ensure reliability by providing precise and consistent performance over various conditions. Short-term PV power generation forecasting relies on some key meteorological feature behaviors. This study aims to design a robust model for day-ahead PV power generation forecasting that performs well under various sub-conditions. It proposes a stacked generalization approach with optimized ANN and XGBoost models as base learners. The outputs of these models are used as new features in validation data that function as training data for the ETR meta-model. The diversity of the base models provides complementary advantages for the meta-learner. The model is verified on real-time data from DKASC (Desert Knowledge Australia Solar Centre), a solar center in Alice Springs, Australia. A comparative analysis is conducted with various machine learning models including Random Forest, ANN, and XGBoost, both individually and in altered stacked configurations. The result shows that the proposed stacked generalization model has higher accuracy than the other models examined. Finally, the model is evaluated on different PV technologies to assess the model's compatibility with different solar technologies. The results indicate that the proposed model provides precise and consistent performance over various conditions.