IHRR Associate's research advances risk-aware control of multi-energy systems

The research was led by Muhammed Cavus, an Associate of the Institute of Hazard, Risk and Resilience (IHRR) at Durham University, and contributes to the development of more resilient energy infrastructure capable of operating safely and efficiently under uncertain and changing conditions.

The study introduces a new control framework designed to make multi-energy microgrids (MEMGs) safer, more transparent and more resilient. These systems manage multiple energy carriers simultaneously, including electricity, heating, cooling and hydrogen.

As microgrids become more complex and increasingly rely on components such as batteries, fuel cells and electrolysers, maintaining reliable operation becomes more challenging. These technologies degrade over time, reducing performance and increasing uncertainty. Conventional control methods, including standard Model Predictive Control (MPC) and machine learning-enhanced approaches, often treat forecasting, uncertainty and equipment degradation as separate problems. This can reduce operator confidence and lead to less efficient or less reliable operation, particularly under stress conditions or as assets age.

To address these challenges, the authors developed a Learning-Enhanced Stochastic Model Predictive Control (LE-SMPC) framework that combines three key elements:

• Trustworthy machine learning forecasts of battery and fuel cell performance, supported by feature attribution techniques that show which factors drive each prediction.
• Degradation-aware modelling of battery capacity fade and fuel cell and electrolyser voltage decay.
• Stochastic MPC with chance constraints, overseen by a resilience index that automatically activates more conservative control actions when system health approaches critical levels.

These components operate together and continuously exchange feedback. For example, when control constraint violations are detected, the forecasting model is automatically updated and retrained. This allows forecasting, control and degradation monitoring to evolve together rather than functioning independently.

The framework was evaluated using a year-long simulation of a multi-energy microgrid incorporating solar PV, wind generation, battery storage, an electrolyser, hydrogen storage, a fuel cell, heat pump, boiler, and both absorption and electric cooling systems.

Results showed around 40% lower forecasting error compared with LSTM, GRU and standard RNN models. The framework also maintained reliable operation under representative degradation scenarios, including up to 18% fuel cell voltage loss, 30% battery capacity fade, and seasonal heating and cooling demands ranging from 160 to 210 kW.

Compared with deterministic and standard stochastic MPC approaches, the LE-SMPC framework reduced operational costs by 8-12% and cut control constraint violations from more than 12% to less than 1%. It also improved the average resilience index from approximately 0.75 to between 0.86 and 0.90.

A distinctive aspect of the approach is that feature attribution is integrated directly into the machine learning training process rather than being applied only as a post hoc explanatory tool. This ensures that each retained input has a stable, measurable and physically meaningful influence on model predictions. As a result, operators gain clearer insight into why particular control decisions are made, helping to build trust in increasingly autonomous energy management systems.

The research aligns closely with IHRR's interests in risk, resilience and the management of complex infrastructure systems, demonstrating how explainable artificial intelligence and uncertainty-aware control can support more robust and adaptive energy networks.

This research was supported by UK Research and Innovation (UKRI) through the SAT-Guard project (MR/Z50578X/1) and by the Engineering and Physical Sciences Research Council (EPSRC) through the VPP-WARD project (EP/Y005376/1).

Authors: Muhammed Cavus Visiting Fellow Department of Engineering, Durham University, Jing Jiang School of Engineering Physics and Mathematics, University of Northumbria, Adib Allahham School of Engineering Physics and Mathematics, University of Northumbria and Hongjian Sun Director of Research in the Department of Engineering, Durham University.

Read the paper: https://doi.org/10.1016/j.apenergy.2026.128321

Code and data: https://github.com/cavusmuhammed68/Trustworthy_ML