This article is more than 1 year old
Machine-learning model pinpoints dying power grid components
Hello, Bayesian, our old friend
Machine learning could one day help energy providers better pinpoint failing or compromised components in power grids, or better identify traffic congestion for local authorities, according to a study.
A research project led by MIT describes a technique capable of modelling complex interconnected systems made up of numerous variables that change value over time. By mapping connections in these so-called multiple time series, a Bayesian network can learn to identify anomalies in the data.
Power grids are a perfect case study, Jie Chen, co-author of the paper [PDF] and a research staff member at the MIT-IBM Watson AI Lab, explained on Friday. "A prominent example of the source of multiple time series is the power grid, where each constituent series is the grid state over time, recorded by a sensor deployed at a certain geographic location," he said.
The power grid state can be made up of many data points, including the magnitude, frequency, and angle of voltage throughout its network as well as current. Chen said detecting anomalies depended on identifying abnormal data points that might be caused by things like a cable break or damage to insulation.
Specifically, a power grid can be modeled as a collection of nodes and edges and their associated sensor readings. A probability distribution can be calculated for these readings as they change over time; any data coming in that doesn't fit this distribution is a sign that something's wrong. And while you can set up hard rules to detect this sort of thing – frequency over or under limit, phase drifting too far, etc – the Bayesian approach is supposed to be less manual, more automatic, and less cumbersome to manage.
"In the case of a power grid, people have tried to capture the data using statistics and then define detection rules with domain knowledge to say that, for example, if the voltage surges by a certain percentage, then the grid operator should be alerted," Chen said.
"Such rule-based systems, even empowered by statistical data analysis, require a lot of labor and expertise. We show that we can automate this process and also learn patterns from the data using advanced machine learning techniques."
- Alibaba gazes into crystal ball to spy coming advances in AI and silicon photonics
- Twitter's machine learning algorithms amplify tweets from right-wing politicians over those on the left
- AI tool finds hundreds of genes related to human motor neuron disease
- AI-designed drug to treat deadly disease now tested on humans
This technique uses unsupervised learning to identify what is considered an anomalous result instead of using handcrafted rules. When the researchers tested their model on two private datasets recording measurements from two power grid interconnects in the US, they found it outperformed other machine learning methods based on neural networks.
Samples of grid sensor data measured were used as input, and the Bayesian network provides an score to forecast whether the data is anomalous or not, Enyan Dai, co-author of the paper and a PhD at Pennsylvania State University, told The Register. It's a general method to detect data changing abnormally, and could even be used to sound the alarm if power grids were hacked, he said.
"It can be used to detect a power grid failure devaluation to cybersecurity attacks. Because our method essentially aims to model the power grid in normal status, it can detect anomalies regardless of the cause." The method was applied to other multi-time series systems, too; the team used it to study datasets to identify bottlenecks in highway traffic patterns and water quality.
"Unfortunately, our model cannot point out why [systems] fail," Dai told us. "But it does can detect which part of the power grid fails. The model can be applied to monitor the status of a power grid, and could report an grid failure in one minute, which potentially means it can act in real time. But for real-world applications, I believe more tests especially the model's robustness needs to be evaluated."
The proof-of-concept paper will be presented at this year's International Conference on Learning Representations conference. ®