Quote of the day

A person who never made a mistake never tried anything new.

- Albert Einstein

Contact

Physics-Informed Digital Twin and Safe Reinforcement Learning for Rural Microgrid Energy Management

Open Energy

Physics-Informed Digital Twin and Safe Reinforcement Learning for Rural Microgrid Energy Management

Physics-Informed Digital Twin and Safe Reinforcement Learning for Rural Microgrid Energy Management

Description

Control a rural photovoltaic microgrid with battery storage using a constraint-aware reinforcement-learning policy trained on a physics-informed digital twin.

Details

Context and Problem Statement

Rural microgrids combine intermittent solar generation, limited storage, and uncertain loads. Rule-based control is often suboptimal, while unconstrained reinforcement learning may violate voltage or battery-aging limits.

Research Question

How can strict electrical and battery-aging constraints be guaranteed throughout training and operation of a learned energy-management policy?

Proposed Approach

Formulate a constrained MDP, develop a digital twin combining power-flow and battery-degradation models, apply safe reinforcement learning with infeasible-action projection, and validate through hardware-in-the-loop simulation.

Expected Contribution

A comparison among rule-based control, model predictive control, and safe RL in terms of operating cost, load shedding, battery degradation, and constraint violations.

Expected Prototype

An embedded or PLC-deployable controller connected to the digital twin with a supervision interface for state of charge, forecasts, and safety margins.

Datasets

PVGIS, NASA POWER, open load profiles, and laboratory microgrid measurements.

Challenges

Simulation-to-hardware mismatch, battery-aging modeling, limited local load profiles, and sensor-fault robustness.

Research Question

Can a constraint-aware learned energy-management policy outperform model predictive control in a rural microgrid without violating electrical or battery limits?

Innovation

Battery degradation is incorporated as an explicit learning constraint rather than only as a soft cost penalty.

Expected Deliverable

An open microgrid simulator, a trained safe-control policy validated through hardware-in-the-loop experiments, and a documented benchmark.

Technologies

Safe Reinforcement Learning (CMDP) Digital Twin and Power-Flow Simulation Physics-Informed Neural Networks Embedded Systems and PLCs Battery-Aging Modeling

Required Skills

  • Electrical Engineering and Power Systems
  • Reinforcement Learning
  • Numerical Simulation and Modeling
  • Embedded Programming

Datasets

  • PVGIS (European Commission)
  • NASA POWER
  • Measurements from a laboratory microgrid testbed

Morocco & Africa Relevance

Decentralized rural electrification remains important across Africa and for isolated Moroccan sites, where autonomous and safe energy management can reduce operating costs.