Satellite-Based DNI Forecasting Pipeline for CSP-TES Applications using Exp-GPR and Bagged Trees

This code implements a machine learning framework for forecasting Direct Normal Irradiance (DNI) to optimize dispatch strategies for Concentrated Solar Power (CSP) plants with Thermal Energy Storage (TES). It processes 17 meteorological and radiative variables from satellite data, generating lagged features from 30 to 360 minutes and utilizing a strict chronological train-validation-test split to prevent data leakage. The pipeline trains two distinct models: Ensemble Bagged Trees, which offers superior speed and accuracy for short horizons (30 mins), and Exponential Gaussian Process Regression (Exp-GPR) using a Matern kernel to capture high-frequency fluctuations and provide robust performance for longer lead times. Finally, it evaluates model performance using metrics such as RMSE, nRMSE, and R2 to ensure reliability for operational planning.