This paper presents multiple data analytics and machine learning approaches to optimize energy storage for grid services. These approaches are critical components towards an information integration and decision-making framework as a Digital Twin for energy storage systems, especially those with long-duration storage capability. The Digital Twin framework integrates three complementary components: (i) unsupervised learning to capture joint patterns of energy storage system dispatch, system states, and environmental conditions; (ii) ensemble supervised learning for mechanistic understanding and classification of dispatch; and (iii) deep learning for mid-term electricity price forecasting to support multi-day decision-making. In contrast to recent forecasting- or optimization-centric studies, the proposed Digital Twin explicitly models nonlinear cross-dependencies, embeds uncertainty information into dispatch decisions, and unifies behavioral learning with predictive forecasting within a single modular architecture. The framework integrates effective feature selection, uncertainty reduction, and improved scheduling for inter-day energy shifting. We demonstrated the effectiveness of the framework and its three key components using multi-year datasets from independent system operators and a 10 h battery energy storage system. Results show that the Digital Twin predictions yield total revenues exceeding, by more than 10%, those achieved through conventional linear programming approaches, as demonstrated in the New York Independent System Operator studies. • Modular digital twin integrating unsupervised, ensemble, and deep learning. • Joint behavioral learning captures nonlinear cross-dependencies in energy storage. • Automatically optimized energy storage dispatch for long-duration storage systems. • Greatly reduced uncertainty and optimized multi-day energy storage scheduling. • Proposed framework yields >10% revenue gain over conventional benchmarks.
HOU et al. (Wed,) studied this question.