The Energy Paradox of AI in Power Electronics: Innovation Driver or Unsustainable Burden?

The application of artificial intelligence (AI) is rapidly growing in the power electronics (PEs) industry. This recent surge in growth in AI underscores the industry’s shift towards application-driven innovation. Unlike methods-driven research, which often falls short in real-world tasks, the integration of AI with practical challenges in PEs is fostering broader innovations in the field. This white paper explores the mainstream application of AI in PEs, expanding beyond the use of performance metrics alone, to consider the energy implications of such AI methods and the data quality requirements to achieve high performance at a reasonable economic and energy cost. This work also emphasizes the importance of explainable AI, considering both data quality and predictions, but notes that PEs as a field requires dependency on multiple independent and interdependent inputs, due to inherent cross-coupling effects. This article also advocates the use of physics-informed machine learning, which combines theoretical mechanisms with data to balance domain knowledge and data-driven predictions. This approach is particularly effective in offline settings but faces challenges in online environments without relevant high-quality data. The discussion extends to the energy demands of AI, emphasizing the need for sustainable design and implementation to avoid over-dimensioned models that consume excessive energy. In addition, this article provides a number of case studies across various applications of AI using different methodologies to illustrate the challenges in achieving high quality performance and consideration of energy as a metric. Finally, this article proposes a framework for application-driven AI innovations in PEs, addressing common deficiencies in deployment-scale assessments and concluding with the environmental sustainability of AI applications in this field.