Safe reinforcement learning-based prescriptive maintenance of distribution grid using Bi-Mamba+ and constrained policy optimization: A Danish grid case study

As Denmark pursues ambitious climate goals—aiming to reduce CO 2 emissions by 70 % relative to 1990 levels by 2030 and achieve climate neutrality by 2050—electrification is accelerating across sectors. This transition places increasing strain on the country’s aging distribution grids, driven by rising electricity consumption from electric vehicles and more. Traditional grid expansion, while necessary, is subject to interruptions in supply chains and does not directly reduce emissions. As a result, alternative strategies that enhance operational efficiency and grid resilience are essential. This article presents a prescriptive framework that combines predictive load forecasting and real-time control to facilitate prescriptive maintenance and proactively mitigate some of these challenges. The framework is demonstrated using real data from a Danish grid. The first stage utilizes the Bi-Mamba+ model for forecasting consumption and production at the transformer level, achieving high accuracy and demonstrating the scalability of a single, well-engineered model across the grid. The second stage applies neural network-based real-time control trained via Constrained Policy Optimization (CPO), effectively reducing grid alarms by dynamically controlling storage units, though sensitive to forecast accuracy and system complexity. Finally, the two components are combined via a transformation layer to ensure compatibility between the two components. Together, these components form a forward-looking solution that supports grid stability and flexibility, reduces reliance on infrastructure expansion, and paves the way for smarter grid management. The results underline the potential of integrating optimization, forecasting, and control in distribution system operations to take advantage of the potential that prescriptive maintenance offers. • Applying safe deep reinforcement learning with CPO to control a real Danish distribution system simulated using real data. • Proposing a framework that combines predictive analysis with real-time control. • Exploring the performance of CPO on forecasted data and comparing it to the performance on observed data. • Benchmarking the performance of CPO against TRPO on a real system.