Hierarchical machine learning and optimisation framework for intelligent energy management in smart distribution systems

Abstract

Integrated energy management frameworks that can concurrently propagate forecast uncertainty into optimization constraints, couple fault detection with automated redispatch strategies, and enable repeatable solver benchmarking on a single testbed are still lacking in current smart distribution system research. While grid-scale battery energy storage systems (BESSs), electric vehicle (EV) fleets, and distributed energy resources (DERs) have greatly increased the flexibility of distribution networks, they have also introduced significant operational complexity that traditional deterministic energy management systems (EMSs) are unable to effectively handle. Current methods limit real-time adaptability, robustness, and operational efficiency by treating forecasting, optimization, and fault management as separate jobs. In order to close this gap, this research uses hierarchically structured machine learning (ML) and mathematical optimization approaches to provide a co-design framework for intelligent real-time and day-ahead management of smart distribution systems. Four closely related layers are integrated into the architecture: (i) a stacked long short-term memory (LSTM) network for probabilistic 24-hour load forecasting; (ii) multisource heterogeneous data acquisition and preprocessing; (iii) optimization-driven EMS scheduling with uncertainty-aware constraints; and (iv) intelligent fault detection and automated redispatch mechanisms for resilient grid operation. In addition to enhancing operational dependability, forecasting precision, and decision-making effectiveness in next-generation smart distribution networks, the framework creates a consistent and repeatable platform for testing optimization solvers.