How do ML and AI work in power conversion? part 1

The use of machine learning (ML) and artificial intelligence (AI) in power converters represents the latest development in the field of digital power. They are being used in advanced converter control schemes, performance monitors of electrolytic capacitors to support preventative maintenance, and power management systems to improve efficiency and enhance reliability.

Part 2 will look at how ML and AI contribute to optimization and integration in green energy systems and electric vehicles.

AI and ML-based models can handle non-linearities and complex conditions better than fixed algorithms. They use neural networks, fuzzy logic, reinforcement learning (RL), and other AI/ML tools. The results can be dynamic, data-driven, and faster control algorithms.

A framework has been suggested for classifying and analyzing AI/ML tools and applications in power electronics (Figure 1). Possible AI algorithms range from generative adversarial networks (GANs) to RL. Functional layers are optimized to identify objectives, regression is used to determine correlations between input and target variables, and classifications sort the various types of system applications.

Figure 1. Framework for considering ML and AI applications in power electronics applications. (Image: Applied Energy)

BLDC motor drives

ML techniques like neural networks and RL can be used for smarter adaptive control in BLDC motors. Those advanced ML techniques, combined with Hall, EMF, and optical sensors in multi-sensor fusion systems, support precise rotor tracking and control systems that can maximize the performance of complex loads in real-time. That supports smooth, precise, and efficient motion in applications ranging from robotics to large traction motors in EVs.

These techniques can also be applied to simpler applications like pumping systems (Figure 2):

The RL agent adjusts the motor control parameters to maximize efficiency.
The MCU generates control signals for the inverter that implement the adaptive parameters from the RL agent.
The (optional) wireless signal processing module enables bidirectional control, receiving sensor inputs from the motor drive unit and sending the required control signals back. Control can also be implemented using a hardwired solution in place of the wireless module.
Figure 2. Block diagram of a pumping system using RL-based control. (Image: MDPI sustainability)

Electrolytic capacitors

AI and ML are being used to predict failures of large electrolytic capacitors like those used for output filtering or DC-link applications. Techniques like artificial neural networks (ANN), convolutional neural networks (CNN), and random forest classifiers can take inputs including average current and ripple voltage to predict remaining useful life (RUL) and detect capacitor degradation, enabling preventative maintenance.

Predicting RUL can be especially important for DC-link capacitors in MW-scale applications like Maglev choppers. In one case, a two-input ANN has been developed using inputs of average levitation current (C) and capacitor ripple voltage (Figure 3).

Figure 3. Two-input ANN for estimating C and ESR in a Maglev chopper system. (Image: MDPI energies)

The integrated voltage sensor used in a protection circuit and an existing current sensor used for levitation control are used to provide the current and ripple voltage inputs for the ANN. That simplifies the implementation.

Further simplification is achieved by using only a single hidden layer with multiple neurons. That speeds training and reduces the risk of producing local optima and inaccurate predictions. The use of AI and ML is not limited to MW-scale power systems; they can also be helpful in power management for microprocessors or systems on chip (SoC) applications.

Anticipatory power management and DVFS

Anticipatory power management (APM) in dynamic voltage and frequency scaling (DVFS) uses AI, ML, or predictive algorithms to forecast future workload demands, especially in large microprocessors or SoC solutions, rather than just reacting to past or current usage.

APM can optimize power efficiency by adjusting the voltage and frequency immediately before performing power-intensive tasks. That can also reduce latency.

Predictive modeling is one way to implement APM in DVFS. These models forecast the frequency sensitivity of near-term operation by analyzing processor wavefronts instead of waiting for a workload assessment that is inherently slower and less accurate. ML algorithms have been developed that can rapidly analyze hundreds of signals.

Summary

AI- and ML-based digital power techniques are a rapidly emerging area in power electronics and are being applied to all sizes of power systems from small SoCs to massive MW-class inverters. Their use is equally broad and spans control and switching optimization, preventative maintenance, and predictive modeling. Applications range from high-performance computing to robotics and EVs, to industrial pumps and mass transport systems.