Assignment 02: Constraint-Aware Edge Control for Energy Management Systems using TinyML
Summary
Design and deploy a real-time control system on the EnergyWatch edge device that uses local forecasts to optimally manage energy flows under user and device constraints.
Problem Definition
Forecasting alone does not create value unless it is translated into control actions. In EnergyWatch, control decisions must be taken locally on the edge device to ensure low latency, privacy, and resilience. However, classical optimisation-based control approaches such as Model Predictive Control are often too computationally demanding for embedded hardware, while rule-based strategies fail to adapt to changing conditions. The challenge is to develop a control framework that operates within strict computational limits while respecting physical device constraints, user preferences, and tariff structures. This controller must reliably manage flexible assets such as batteries, electric vehicles, and heating systems, even under uncertainty in forecasts and measurements.
Method
The student will formulate a home energy management problem in which control actions are derived from forecasts of load, renewable generation, and prices. The formulation will include operational constraints such as device limits, comfort requirements, and tariff structures. The student will develop and compare lightweight control approaches suitable for edge deployment, including simplified or approximate Model Predictive Control, reinforcement learning methods adapted for constrained environments, and hybrid approaches combining physics-based models with learning components.
Particular attention will be given to constraint handling, ensuring that all control actions remain safe and feasible in real-world operation. The controller will be implemented in a resource-efficient manner and deployed on the EnergyWatch edge device using an appropriate runtime environment. Robustness mechanisms such as fallback strategies and safe default policies will be included to guarantee reliable operation under communication failures or model inaccuracies. The performance of the controller will be evaluated using simulation scenarios and, where possible, hardware-in-the-loop experiments, considering metrics such as cost savings, self-consumption, peak reduction, computational efficiency, and real-time feasibility.
Research Objective
The objective is to demonstrate that advanced control strategies can be executed directly on edge devices without relying on cloud-based optimisation. The student should show that the proposed controller improves energy efficiency and cost performance while respecting constraints and operating within strict hardware limits. Success will be evaluated based on control performance, robustness, computational efficiency, and successful deployment on the EnergyWatch edge device as an operational control module.
Courses and Supervision
A strong background in control systems, optimisation, machine learning, and power/energy systems is required. Recommended courses include optimisation and control, machine learning, and embedded systems. Familiarity with systems modelling is expected.
For any questions please contact Eli Shirazi, e.shirazi@utwente.nl