Challenges in Enhancing Power Electronic Systems with Artificial Intelligence

Artificial intelligence (AI) is one of the most prominent research areas nowadays and is growing quickly. Design, control, maintenance, and optimization are the three unique life-cycle phases of power electronic systems.

These phases are connected with one or more tasks that AI may perform, such as data structure exploration, regression, categorization, and optimization.

Three specific examples are used to demonstrate the needs of AI approaches for each stage of the life cycle.

Ⅰ. Artificial Intelligence Approaches For Design

Many variables, including weight, volume, and pattern, must be decided upon to design the heatsink of a converter system. This is essentially an optimization task.

When performing optimization via an iterative trial-and-error process, metaheuristic approaches are utilized. Even though the design job requires a lot of computing power, it is usually done offline. In this instance, there are fewer requirements for the algorithm's speed.

The majority of the time, the suboptimal heatsink design is still better and adequate, even though the optimization using the metaheuristic method does not guarantee a worldwide solution.

As a result, algorithm precision is also not important. It is not necessary to have a training dataset or for the optimization process to be interpretable.

Ⅱ. Artificial Intelligence Approaches For Control

The most important needs are algorithm speed and accuracy in the control of a converter system. It enables adaptive updating in an online mode. The real-time control errors of a converter system, such as voltage and current errors, must be fed back to the intelligent controller.

Furthermore, interpretability is important, and the stability of the controller must be guaranteed theoretically. Preparing the dataset for model training is not necessary because the intelligent controller is typically adjusted online.

Ⅲ. Artificial Intelligence Approaches For Maintenance

For maintenance, a slow algorithm speed is needed for the RUL prediction of switching devices in a converter system because devices break down slowly and a long decision-making period is manageable.

The computational effort in this application is minimal, and the degradation model for the RUL prediction can be developed in offline mode and easily modified in online mode.

The dataset requirements, such as dataset quality, dataset size, and label balance, are the most important since the dataset has a significant impact on the model's accuracy. Furthermore, it is important that the RUL prediction findings with uncertainty can be interpreted.

Table 1 presents a comparison of AI algorithms at every stage of the life-cycle of power electronic systems.

Table 1: AI algorithms at every stage of the life-cycle of power electronic systems Source: IEEE Transactions on Power Electronics

Ⅳ. What are the difficulties and features of artificial intelligence in power electronic systems?

Using AI in power electronics is different from using AI in other tech areas because of the unique features and challenges that come with power electronics systems.

The point of view is that artificial intelligence has enormous promise for power electronic systems. The following are a few of the many opportunities and challenges that need to be investigated:.

Adoption of AI

Even though there have been a lot of studies on AI for power electrical systems since the 1990s, there aren't many real-world applications yet. This is in direct opposition to what AI is said to be capable of.

Further research into tasks where AI can practically outperform traditional approaches is required. From an industrial standpoint, the advantages of AI-based solutions should be easily distinguished by contrasting them with traditional techniques.

Ⅴ. Combined AI Use During All Stages of the Lifecycle

The use of AI at all stages of the design, control, and maintenance life cycles will make flexible functional interactions possible. Procedure simplification and overall performance optimization are the benefits of combined usage of AI. It also makes it possible for the system to manage data flow between the electrical and other fields.

Merging of Multilevel Information

Power electronic systems that are vital to safety must be robust. Most of the time, several models and information sources are available for a particular power electronic system application. If these information sources and models are used together, any flaws that might exist can be lessened, making the system more stable.

Multilevel information fusion can be done at the data level, the feature level, the decision level, or any mix of these levels. This lets you use the best parts of each source of information.

For instance, differential equations for power converter systems can be combined with AI to create a hybrid answer for condition monitoring. So that both the model-driven and data-driven sides can be used together to get better accuracy and reliability,.

Computation-Light AI

One of the main distinctions between power electronic systems and other industrial domains is the absence of a powerful computing unit. However, there are strict requirements on computing performance for real-time applications, such as control.

Complex deep learning methods can improve performance, but they require a lot of computing power to power electronic systems.

Computation-light AI algorithms provide a promising solution, as they offer comparable performance to deep learning algorithms and can be implemented in a cost-effective way.

Data-light AI

The dataset is one of the obstacles preventing AI from being implemented in power electronic systems. For example, in AI-based remaining useful life prediction systems, the dataset must be sufficiently flexible to enable precise degradation behavior learning.

The dataset size is typically limited; however, the degradation trials need a lot of resources. In situations where safety is an important feature, the situation is even worse.

Therefore, one potential path is to create AI algorithms that require fewer datasets—that is, data-light AI systems that can function satisfactorily even when faced with poor datasets.

Explainable AI

One problem with most AI algorithms used in power technology is the "black boxes’’ character. The majority of AI-based systems for remaining useful life prediction are only able to offer an approximate estimate. They do not offer sensitivity analysis or quantify uncertainty.

It causes AI-based solutions to become unclear and less attractive for industry applications, particularly in circumstances where safety is at risk. Enhancing the transparency of algorithms is crucial to achieving better reliability.

Data privacy

Data privacy has received more attention lately; one example is the European Union's General Data Protection Regulation. Standard AI algorithms find it difficult to train in the presence of these crucial constraints since centralized data collection could not be possible in the future.

Therefore, creating a cooperative learning approach for AI algorithms for power electronics applications—without jointly collecting data from several places—shows promise. It is in line with the trend of data privacy laws when it comes to using AI to execute solutions.

Database on Power Electronics

Large datasets are needed for model training, particularly for maintenance applications, because of the complicated features of power electronics.

On the other hand, collecting data through experimental testing is typically costly and time-consuming. Establishing a standard data and knowledge base for power electronics is highly demanded.

The performance of benchmark algorithms and the acceleration of application development depend heavily on these publicly available datasets. It will help the academic and industrial power electronics communities worldwide.

Ⅵ. Summarizing the Key Points

● Artificial intelligence has enormous potential to enhance the performance and reliability of power electronic systems.

● Artificial intelligence can be applied across the distinct life-cycle phases of power electronic systems, including design, control, and maintenance.

● Computation-light and data-light AI algorithms offer promising solutions for power electronic systems with limited computing power and datasets.

● Explainable artificial intelligence and data privacy are crucial considerations for artificial intelligence-based solutions in power electronic systems.

Ⅶ. Reference

Zhao, Shuai, Frede Blaabjerg, and Huai Wang. “An Overview of Artificial Intelligence Applications for Power Electronics.” IEEE Transactions on Power Electronics 36, no. 4 (April 2021): 4633–58. https://doi.org/10.1109/tpel.2020.3024914.