Robust Energy Forecasting in Combined Cycle Power Plants: Mitigating Cyberattacks on Machine Learning Models

ABSTRACT The utilization of data for generation output forecasting in combined cycle power plants (CCPPs) makes room for attackers to exploit and degrade the performance of machine learning models. This work investigates the potential impacts of cyberattacks on energy generation forecasting models for CCPPs. Attacks are implemented on the four top‐performing forecasting models: gradient boosting, extreme gradient boosting, Random Forest, and CatBoost, identified through a comparative analysis of 12 models, including various tree‐based methods, support vector machines, deep learning models, and linear regression techniques. Scaling, denial of service, fast gradient sign method, and basic iterative method attacks are employed with diverse attack volumes and perturbations to investigate the vulnerability of these models. To counteract these vulnerabilities, a two‐layer defence scheme employing ensemble adversarial training is proposed, aimed at mitigating the adverse effects of these cyberattacks. The findings underscore the significance of the proposed robust defence strategy in ensuring the reliability of forecasting models in the presence of cyberattacks.