Optimisation of electrical machines using data-driven dynamic thermal models and surrogate-based multi-objective evolutionary algorithms.

This thesis addresses two challenges that electrical engineers face when designing electrical machines: managing heat generated during operation and reducing the time required to obtain optimal designs. Heat negatively affects machine performance, shortens lifespan, and can ultimately cause failure. Traditional analytical and numerical modelling approaches require substantial domain knowledge, while data-driven methods demand large amounts of costly training data. On the other hand, electrical machine design is inherently a Multi-Objective Optimisation Problem (MOOP). Although Multi-Objective Evolutionary Algorithms (MOEAs) are widely used, they require the evaluation of thousands of candidate designs, making them impractical when fitness function evaluations are computationally expensive. To address these challenges, we develop low-complexity, data-driven Linear Regression models for estimating motor component temperatures as an alternative to traditional thermal models that depend on the physical aspects of the motor and extensive domain knowledge. A multi-objective thermal modelling framework is proposed to find the trade-off between the cost of data acquisition and the expected model accuracy. The framework incorporates domain knowledge and the Exponentially Weighted Moving Average (EWMA) to enhance training data in a way that captures historical trends. It also includes a stepwise regularisation mechanism for reducing model complexity by removing features based on the magnitude of their corresponding regression coefficients. Results show that between 30% and 50% of model features can be eliminated without significantly affecting accuracy, enabling more efficient and scalable thermal modelling for electrical machine design. We address the second challenge by developing fast and accurate surrogate models that can be integrated into standard MOEAs to accelerate early convergence by approximating the expensive fitness function. Since each optimisation problem is unique, training data for surrogate models must be generated on-the-fly during the optimisation process. To this end, we propose a strategy that constructs training data from the most recent generation of a MOEA and uses it to train surrogate models for estimating the fitness of candidate solutions in the subsequent generation. The framework is coupled with classification surrogate models (to pre-select viable solutions) or regression surrogate models (to estimate fitness objective values), which are attached to state-of-the-art MOEAs. We evaluate their performance on 31 widely used benchmark MOOPs, with results demonstrating that surrogate-assisted MOEAs achieve significantly faster convergence compared to their standard unassisted counterparts.

Grantor dc:publisher.institution

Robert Gordon University

Year dc:date.issued

2026

Author dc:creator

• Banda, Tiwonge Msulira

Advisor dc:contributor.advisor

• Zăvoianu, A.C. Petrovski, A. and Bramerdorfer, G.

• Electrical machine design
• Thermal modelling
• Data-driven models
• Multi-objective optimisation problem
• Surrogate models
• Multi-objective evolutionary algorithms

Language dc:language

en

Identifier

oai:rgu-repository.worktribe.com:3315809
https://doi.org/10.48526/rgu-wt-3315809

Author Identifier

0000-0002-2344-0397

OAI identifier oai:identifier

oai:rgu-repository.worktribe.com:3315809