Distribution networks are an obvious target for AI. They generate continuous sensor data, the problems they pose are well-defined, and a one-percent efficiency gain on infrastructure of this scale is worth a lot. A 2023 review by Mohammadi & Mohammadi maps how neural networks, SVMs, and decision trees are being applied as grids shift from a few large power plants to many smaller, distributed sources, with forecasting renewable output, managing demand, and scheduling equipment as the headline use cases. The outlook is positive but fitting these systems into existing infrastructure and rules remains unresolved.
That qualification matters once you read the empirical record against it. In water networks, Elshazly et al. (2024) built a machine learning system for leak detection that reuses pressure, flow, and quality sensors already installed in the pipes. A March 2026 systematic review of 53 studies in Smart Cities by Zuñiga-Uribe et al. reports detection accuracies between 94 and 100%. The catch? Most of those numbers come from simulated networks, not real ones, because field measurements are expensive and operationally messy.
The reality gap is named directly on the electricity side. Shi et al. (2024), working with the New England grid operator, tested reinforcement learning agents for voltage control on three benchmark grids. Agents trained in simulation degraded sharply once the real network changed shape or behaved differently than the model predicted. A 2026 review in Renewable and Sustainable Energy Reviews by Ahmadi & Aly reaches the same diagnosis. These methods scale poorly, offer no formal guarantees of stable operation, and model the underlying physics weakly.
The regulatory question is the structural counterpart to these findings. Volkova et al. (2024) argue that accountability must be defined in measurable terms across the full lifecycle of an AI system in the grid, and flag the EU AI Act's narrow definition of safety-critical components as a shortcoming for infrastructure of this kind. Sharma & Pursiainen (2026) observe that AI trained on the kinds of uncertainty that can be modelled statistically struggles with cascading failures that fall outside its training assumptions; resilience therefore depends on bounded autonomy inside a hybrid human-machine governance setup, not full automation or a simple human override. The bottleneck today is about the architecture that decides where the model stops and where the operator starts.