A Hybrid ML-NWP Approach for Predicting GHI and Power to Mitigate Penalties Under CERC’s Deviation Settlement Mechanism

Accurate GHI forecasts are critical for solar plant operators in India to effectively mitigate penalties due to the Deviation Settlement Mechanism (DSM). Despite the abundance of studies on GHI prediction, none have specifically addressed forecasting frameworks aligned with CERC’s power scheduling strategy in India. Additionally, there are no published studies that compare the use of new-generation foundation models and existing machine learning (ML) based approaches for time series forecasting of GHI. This study introduces a hybrid approach that combines Numerical Weather Prediction (NWP) with machine learning for GHI prediction and evaluates it within CERC’s power scheduling framework, alongside foundation models and univariate ML models. The results show that the proposed hybrid modeling approach, which incorporates covariates from the Global Forecast System GFS (NWP), outperforms all other methods tested across all datasets considered in this study. In this hybrid approach, the LightGBM model achieved $10 \%$ penalty reduction over the best foundation model and a 22.7 % penalty reduction over best univariate ML model.