Deep Learning AI Targets Renewable Voltage Instability

A research team at Sweden’s KTH Royal Institute of Technology has created a machine-learning algorithm that can quickly react to voltage regulation problems caused by intermittent renewables in distribution grids. 

The algorithm targets the primary disadvantages of the world’s rapid transition to renewables. By nature, solar photovoltaics (PV) and wind turbines are variable, only supplying electricity when sunshine and wind are present. This intermittency has downstream effects on power networks, including reduced performance, continuous voltage fluctuations that degrade electrical equipment, and security issues that could cause outages or require emergency load-shedding measures. 

 

Solar panels. Image used courtesy of Pexels

 

The ongoing rise in grid-connected EV charging systems also causes instability. Charging demand varies depending on EV owners’ habits and needs, resulting in stochastic fluctuations impacting power quality and stability.

The KTH researchers envisioned a machine-learning method to control distribution systems using smart converters. The solution centers around deep reinforcement learning (DRL), where computational agents make trial-and-error decisions from unstructured input data. 

 

One of the study’s authors, KTH assistant professor Qianwen Xu, summarized the project's focus. Image used courtesy of Power Circle 

 

In tests, the team demonstrated their algorithm could self-control distributed energy resources interfaced by smart power converters. Also, the deep learning approach regulated its learning process for constraint-violating actions. 

 

Voltage Control in Renewable Distribution Grids

The DRL solution targets power converters deep in the grid. The researchers wanted to develop a decentralized control that could safely coordinate and optimize large-scale energy sources while simultaneously handling sudden fluctuations without real-time communication. 

Existing centralized voltage control methods may help with voltage regulation but bring additional problems. Since they require fast calculation and communication, they can’t achieve real-time voltage control for fluctuating renewables. 

Distributed voltage control methods are another option, but their real-time calculations still have high requirements for system communications. 

On the other hand, decentralized methods can quickly respond to fluctuations and have low communication requirements since they only operate with local information. Still, most methods are based on linear droop control and thus cannot optimally control nonlinear distribution systems. They could also cause security violations in system operations. 

 

The KTH researchers’ multi-agent DRL framework. Image used courtesy of the study’s author (Figure 2, Page 5)

 

With these technical gaps in mind, the researchers developed a multi-agent DRL algorithm to eliminate the downsides of existing methods. 

 

Decentralized Control Algorithm

The researchers tested the algorithm in a modified IEEE 33-bus distribution system and compared the results with centralized and decentralized methods. Using a novel data synchronization tool, they validated its ability to provide optimal control within the physical constraints, with 100% safety and without real-time communications. These advantages make it flexible for practical application. 

The researchers could guarantee 100% safety because the model was trained to restrict the DRL agent’s action space to stay within the distribution system’s physical constraints. 

 

The multi-agent algorithm. Image used courtesy of the study’s authors (Page 5)

 

The technology was demonstrated in a smart microgrid hardware platform. The researchers validated the decentralized management system to respond quickly and provide consistent voltage within the required limits. This contrasts with centralized control methods, which are not fast enough to handle continuous fluctuations from renewables and EV charging systems. 

The study compared the algorithm’s voltage profile and power loss curves to existing methods and found minimal power losses and zero voltage security violations. Unlike competing methods, decentralized implementation also didn’t require real-time communication. 

 

Power loss comparisons for existing decentralized DRL methods (left) and centralized control methods (right). Image used courtesy of the study’s authors (Figures 19 and 21, Pages 9-10)

 

The open-source software is available on GitHub.