Resource-efficient Machine Learning for Power Electronic Systems

The transition towards a decarbonized global energy system is fundamentally reliant on power electronics, which enable the grid integration of renewable energy sources, energy storage systems, and many variable loads. As power electronics become increasingly present in the electrical grid, their growing digitalization opens numerous possibilities for enhancing efficiency and reliability, as well as ensuring the optimal utilization of the resources they interface. In this context, machine learning methods have emerged as a powerful tool to extract value from the extensive amounts of data produced by power electronic systems. However, the computational and data requirements of many machine learning solutions introduce significant challenges to their widespread adoption in power electronics applications. This thesis identifies two such difficulties and proposes methodologies to overcome them.

The first challenge involves data collection: in problems where obtaining sufficient training data is time consuming and costly, such as in remaining useful life prediction, an alternative to extensive laboratory testing is to train models on field device data. This is often achieved by means of cloud computing platforms, which enable the logging and processing of vast amounts of data but introduce additional requirements in the form of reliable communication and service costs. In contrast, this thesis proposes harnessing the growing computational resources in edge computing devices, employing online machine learning methods to train models without relying on centralized computing platforms. The stability, reliability, and computational burden of the training process then become key concerns: this thesis demonstrates how it can fail to produce adequate results in memory-constrained environments, and proposes a method based on data selection that enhances the speed and stability of the training process. Crucially, the proposed method introduces no additional memory requirements and only limited computational overhead. The method is demonstrated by means of a case study in thermal anomaly detection for a variable-frequency motor drive, where it is shown to outperform other training techniques.

The second challenge revolves around surrogate modeling, which refers to the process of training computationally light approximators of comparatively expensive models. In power electronic systems, surrogate models are often used to enable the real-time implementation of complex control algorithms which cannot be evaluated at the required update rates on microcontrollers or similarly constrained embedded platforms. In such cases, the computational footprint of the surrogate models must also be minimal, a constraint that often results in reduced performance in the approximation process and therefore also in the controlled system. This thesis proposes a novel training methodology that incorporates additional application-specific knowledge, resulting in surrogate models that can specifically address the most relevant operating regions of the target system. This methodology is experimentally verified in the training of a surrogate controller for a dual active bridge power converter, where the resulting model is able to more closely match the performance achieved with the target control algorithm compared to conventional surrogate model training techniques.