Machine Learning Applied to Solid-State Transformer Control: How AI

Artificial intelligence and machine learning techniques applied to power-electronics control — a topic that ETH Zurich’s Power Electronic Systems Laboratory under Professor Johann Kolar has actively researched since the mid-2010s — are entering commercial solid-state transformer (SST) products through 2025 and 2026. The application of ML to SST control is not a marketing exaggeration. There are specific control problems in modern SSTs that ML approaches solve materially better than traditional hand-coded algorithms.

The dominant traditional control architecture in commercial SSTs is the cascaded proportional-integral-derivative (PID) controller stack: an outer voltage-regulation loop, a middle current-regulation loop, and an inner switching-modulation layer. Each PID loop is tuned individually based on small-signal stability analysis. The control gains are typically constant across operating points, which means the controller is sub-optimal at most operating points and optimal only at the design operating point.

Machine-learning approaches to SST control fall into three categories. First, reinforcement-learning agents that learn control policies through simulated interaction with a model of the SST, then transfer those policies to the physical hardware. The reinforcement-learning agents can find non-linear control policies that outperform PID stacks at off-design operating points without explicit gain scheduling. Second, deep-neural-network observers that estimate hard-to-measure internal SST state (junction temperatures, capacitor health, magnetic-core saturation) from easier-to-measure terminal quantities. Third, model-predictive controllers with neural-network plant models that anticipate disturbances rather than reacting to them.

The commercial implementation challenge for ML-based SST control is certification and verification. UL 1741 and IEC 62109 certification processes presume a control architecture that can be verified by inspection of the source code. ML-based controllers are not directly inspectable in the same way. The compromise increasingly adopted by SST manufacturers is to use ML for parameter optimization and supervisory control while retaining traditional PID stacks for the safety-critical inner loops. The hybrid approach passes existing certification processes while capturing meaningful performance benefits.

The performance improvement available through ML-augmented SST control is meaningful but not transformative. Typical reported gains include 1 to 3 percent efficiency improvement at partial-load operating points, 10 to 30 percent reduction in voltage-transient response time, and 15 to 40 percent reduction in commissioning effort through automatic parameter tuning. The cumulative commercial value of these improvements at gigawatt-scale deployments is substantial. Whether ML-based control becomes a standard feature across all commercial SST products by 2028 or whether the certification challenges keep it as a premium-tier option will be visible in 2026 and 2027 product roadmaps.

Why It Matters

For control engineers, machine learning is entering commercial SST products to solve specific problems, adaptive modulation and fault classification, that hand-coded algorithms handle worse, not as marketing. That matters because it shifts part of the SST’s value into software that improves over a fleet’s operating life, which also deepens the vendor lock-in that customers already worry about with closed control code.

Critical Perspective

Editorial correction: This post is part of MGRID’s Solid-State Transformer industry coverage. As of May 2026, that body of work systematically framed manufacturer announcements, funding rounds, and laboratory demonstrations as commercial deployments. The reality is that field-deployed commercial-class SST in revenue service globally is measured in single digits, and almost every product cited in this series is at “announced” or “funded” stage, not “operational.” Readers should treat the specific claims in this post against the standards documented in our SST Industry Reality Check (the per-claim audit table maps marketing language to verifiable deployment status). The corrective article is the canonical reference for SST industry reality; this post remains published with its original framing so the editorial drift is traceable.

Related Coverage

• How a Solid-State Transformer Gets Certified: UL 1741, IEC 62109, CE Marking, and the Path to a US Utility Procurement Order: Certification challenges for ML controllers mirror SST certification hurdles.
• Amperesand Raises $80 Million Series A Led by Walden Catalyst and Temasek for AI Data Center Solid-State Transformers: AI data center SSTs may benefit from advanced ML control.
• SolarEdge and Infineon Announce Modular 2-5 MW Solid-State Transformer Targeting >99% Efficiency for AI Data Centers: AI data center SSTs may benefit from advanced ML control.
• The MGRID Solid-State Transformer Coverage: A Comprehensive Index of Manufacturers, Deployments, Programs, Technology, and Applications: Comprehensive SST index includes technology and applications.