Green Machine Learning for Smart Grid Stability Prediction: Performance and Energy-Efficient Evaluation

Hadji, Atmane (LISI Laboratory, Computer Science Department, University Center A. Boussouf Mila, 43000 Mila, Algeria)
Boumaza, Farid (Computer Science Department, University of Mohamed El Bachir El Ibrahimi, Bordj Bou Arreridj 34030, Algeria. LAPECI Laboratory , University of Oran1, Oran 31000, Algeria)
Hadji, Fatah (Fundamental Department of Science and Technology , University of Jijel, 18000 Jijel, Algeria University of Bejaia, Faculty of Technology , Laboratory of Mechanics, Materials and Energetic, 06000 Bejaia, Algeria)

Indonesian Journal of Electrical Engineering and Informatics (IJEEI)
Vol 14, No 1: March 2026 (ACCEPTED PAPERS)

The increasing complexity of modern smart grids necessitates intelligent and sustainable predictive models that ensure system stability while minimizing computational energy consumption. This study explores the concept of Green Machine Learning (GML), which integrates high predictive accuracy with energy-efficient computation to promote sustainability in smart grid systems. Unlike conventional benchmarking studies, we propose a sustainability-oriented evaluation framework based on a dual-metric approach (GreenScore and GreenScore*), enabling the joint assessment of predictive accuracy and computational efficiency. This framework serves as a decision-support tool for selecting models under energy and operational constraints. The results demonstrate that MLP has reached the highest level predictive performance (F1 = 0.9736, AUC = 0.9957), while LightGBM offered the best compromise between accuracy and computational efficiency (F1 = 0.9685, AUC = 0.9941). Although Logistic Regression exhibited minimal energy consumption (execution time = 0.03 s), its accuracy was relatively low (0.8027). According to GreenScore and GreenScore*, LightGBM (GreenScore = 0.66, GreenScore* = 0.2646) and Extra Trees (GreenScore = 1.24, GreenScore* = 0.9449) demonstrate superior energy sustainability, while MLP (GreenScore* = 0.0161) and CatBoost (GreenScore* = 0.2171) reflect lower efficiency. Logistic Regression, despite very low computational cost, has a high GreenScore* (533.8422) due to its extremely low execution time but poor predictive performance. Overall, the study confirms that Green Machine Learning enables a multi-objective optimization between predictive performance and energy efficiency, advancing the development of sustainable smart grid management systems.