Ensemble Learning Based Photovoltaic Power Nowcasting Through STL Decomposition

Based on the intermittent nature of photovoltaic (PV) power, accurate prediction is essential for an efficient energy management system due to its instability. Despite advancements in weather forecasting, PV forecasting remains an open challenge for power system stability. An Ensemble Learning (EL) based approach for forecasting is proposed in the proposed work, which is critical in the hourly nowcasting of PV power. The proposed model utilizes a Seasonal Trend decomposition using Losses (STL) of feature vectors extracted from the Savona Campus of the University of Genova, Italy, to further enhance accuracy. The EL structure adds the Transformers (TF), Long Short-Term Memory (LSTM), and the Tree-Based (TB) learner in the base learner and TB or TF in the meta learner. The proposed model is compared further through the evaluation matrix for the datasets, including the original and STL decomposed variables. Hence, the results confirm the ability of EL by having the maximum error of around 5.9918 compared to LSTM of 55.1485. On the other hand, the Mean Absolute Error (MAE) is 0.5 kW for PV power forecasting for an 81 kW PV plant. The result validates the effectiveness of the proposed work by minimizing the maximum error bound and improving its accuracy compared to other existing results.