New research project at Paderborn University is using learning algorithms to control intelligent power electronics systems
Modern industry and information societies are driven by power electronics energy converters that harness electrical energy for various applications. Examples include electric drive units in vehicles, power supplies for server farms, and integrating wind farms and photovoltaic power plants into the power grid. However, power electronics is a complex field: requirements and circuit designs vary depending on specific applications. The challenge is to get to grips with this complexity, especially as regards systems control. Thus far, experts have had to design and program the control system manually and individually for each application. This is time-consuming and is becoming an issue given the increasing skills shortage. Researchers at Paderborn University are therefore working on a new approach to control power electronics systems using artificial intelligence (AI). The ‘VIP+’ project, which has been awarded around 1.8 million euros of funding by the Federal Ministry of Education and Research (BMBF), begins in October. This funding programme reviews promising research results in terms of their practicality and feasibility for subsequent application. The aim is to then translate these findings into products, processes and services.
‘Control includes a mathematical program to operate the power electronics, and the corresponding control software’s implementation on a digital platform. At present, heuristic approaches are largely used. Although these are quick and simple in design, they involve compromises in terms of performance, in particular as regards energy efficiency and dynamic behaviour. Optimum control procedures could theoretically increase performance, but despite intensive research in recent decades, these have not been implemented in a practical industrial environment due to the very high cost of implementation’, explains Dr.-Ing Oliver Wallscheid, who is heading up the project. This is something that the engineer and his team are hoping to change.
In their preliminary work, the researchers involved have already demonstrated the fundamental feasibility of a fully-automated, learning control architecture under laboratory conditions. It hugely reduces the human workload whilst also achieving top system performance. This initial success is based on a combination of data-driven reinforcement learning methods and model-predictive control procedures – and thus, ultimately, on a blend of artificial and human intelligence within an algorithm. During the validation phase that will now follow, the experts will check the potential industrial applicability of the new approach. Specifically, it is about whether the previous results can be transferred across to any power electronics system. Aspects of functional security, robustness and the computing resources required for AI-based control systems will also be examined. AI-based control procedures could prompt revolutionary changes to the development and commissioning of power electronics systems in countless industrial sectors.
