Day-Ahead Solar Power Forecasting Using LightGBM and Self-Attention Based Encoder-Decoder Networks

The burgeoning trend of integrating renewable energy harvesters into the grid introduces critical issues for its reliability and stability. These issues arise from the stochastic and intermittent nature of renewable energy sources. Data-driven forecasting tools are indispensable in mitigating these challenges with their rugged performance. However, tools relying solely on data-driven methods often underperform when an adequate amount of recorded data is unattainable. To bridge this gap, this paper presents a novel day-ahead hybrid forecasting framework for photovoltaic applications. This framework integrates a physics-based model with Machine Learning (ML) techniques, enhancing prediction reliability in environments with scarce data. Additionally, an innovative ML pipeline is introduced for data-abundant environments. The proposed ML tool comprises two branches: a set of regressors, each tailored for specific weather conditions, and a self-attention-based encoder-decoder network. By fusing the outputs from these branches through a meta-learner, the tool achieves predictions of higher quality, as evidenced by its superior performance over benchmark models in an investigated test dataset.