A recent study published in Engineering presents a novel framework for enhancing decision-making in energy systems through the deep integration of machine learning (ML) and mathematical programming (MP) models. The research, titled “AI-Optimized Decision-Making in Energy Systems: Toward a Decision-Aware Machine Learning Framework,” introduces an AI-native optimization (AIOpti) system that aims to bridge the gap between predictive accuracy and practical decision-making in complex energy scenarios.

The study highlights the critical role of ML and MP models in addressing the challenges posed by the integration of renewable energy sources into urban energy systems. While ML models are adept at predicting uncertain system parameters such as electricity prices and renewable power generation, MP models are essential for optimizing decision-making processes. However, existing approaches often separate the training of these models, leading to inefficiencies and suboptimal decision outcomes. The AIOpti system proposed in this study addresses this issue by enabling ML models to be aware of the impact of their predictions on downstream optimization processes.

The research demonstrates that higher prediction accuracy does not always translate into better decision-making. In fact, in some cases, improved accuracy can lead to worse decision outcomes. This counterintuitive finding underscores the need for a more integrated approach. The AIOpti system achieves this through a two-stage training framework: an initial accuracy-oriented pretraining phase followed by a problem-oriented fine-tuning phase. During pretraining, the ML model learns general patterns and features from large datasets to achieve robust predictive accuracy. In the fine-tuning phase, surrogate loss functions are employed to align the ML model’s predictions with specific optimization objectives, such as minimizing operational costs.

The study validates the AIOpti system using real-world data from the Germany–Luxembourg day-ahead electricity market. The results show significant improvements in decision quality, with reductions in decision regret ranging from 1.10% to 18.81% compared to accuracy-oriented methods. Additionally, the AIOpti system achieved notable gains in prediction accuracy (from 33.63% to 83.67%) and training efficiency (from 38.99% to 87.03%) compared to traditional problem-oriented approaches.

The proposed AIOpti system is particularly effective in addressing complex energy management tasks such as virtual power plant (VPP) operations and distributed energy system management. The study demonstrates that the AIOpti system not only enhances decision quality but also significantly reduces training time, making it a more computationally efficient solution. For instance, the training time for the SPO+ method was reduced by 47.83% when the AIOpti system was embedded.

Furthermore, the study explores the scalability of the AIOpti system by training models on larger datasets. The results indicate that while increasing the dataset size improves model performance, it also significantly increases computational demands. However, the AIOpti system mitigates this issue by achieving faster convergence and maintaining high performance with only a modest increase in training time.

The research concludes that the AIOpti system represents a significant advancement in the integration of AI and optimization models for solving complex engineering problems. By equipping AI with an awareness of how its predictions influence decision-making, the AIOpti system provides reliable decision support in fields such as demand-response programs and distributed energy systems. Future research directions include the deeper integration of multi-parameter prediction with nonlinear optimization and the incorporation of uncertainty information into the framework to further enhance decision-making robustness.

The paper “AI-Optimized Decision-Making in Energy Systems: Toward a Decision-Aware Machine Learning Framework,” is authored by Guotao Wang, Zhenjia Lin, Yuntian Chen, Haoran Ji, Dayin Chen, Haoran Zhang, Peng Li, Jinyue Yan. Full text of the open access paper: https://doi.org/10.1016/j.eng.2025.06.037. For more information about Engineering, visit the website at https://www.sciencedirect.com/journal/engineering.