TSMixer: An Innovative Model for Advanced Bias Correction of NWP Solar Irradiance and One-Day-Ahead Power Forecasting

Accurate Photovoltaic (PV) power forecasting is essential for the efficient grid integration of solar energy systems. However, biases in Numerical Weather Prediction (NWP) models often limit prediction accuracy. This study introduces TSMixer, an innovative time-series model applied for the first time to correct biases in Weather Research and Forecasting Solar (WRF-Solar) irradiance data and improve one-day-ahead PV power forecasting. TSMixer utilizes a unique temporal mixing architecture designed to capture complex irradiance patterns and enhance correction capabilities. The model was trained using three types of inputs: measured data, NWP outputs, and satellite data from 11 solar sites across Taiwan and the Central Weather Administration (CWA). The performance of the proposed bias correction method was then benchmarked against conventional methods such as Decaying Average, XGBoost, and LightGBM. Additionally, one-day-ahead PV power forecasting results were compared with state-of-the-art deep learning models, including Informer, Transformer, XGBoost, LSTM, and GRU. Experimental results demonstrate that TSMixer significantly enhances forecast accuracy by reducing reliance on biased NWP data, proving its effectiveness for PV power forecasting and contributing to improved grid stability and solar energy utilization.