A new machine-learning method could detect potential power grid problems or traffic bottlenecks in real time.

It can be difficult to identify a problem in a nation’s power grid. Many thousands of sensors connected across the U.S. collect real-time voltage and electric current data. Sometimes, multiple recordings are made per second.

Researchers from the MIT IBM Watson AI Lab have developed a computationally efficient way to detect anomalies in these data streams in real time. Their artificial intelligence method, which models the interconnectedness and power grids, could detect these glitches better than other methods.

The machine-learning model they created does not require any annotated data about power grid anomalies to be used for training. It would also be easier to use in real-world situations, where it is often difficult to find high-quality, labelled data. It is flexible enough to be used in other situations, such as traffic monitoring systems, where many interconnected sensors gather and report data. It could identify bottlenecks in traffic or show how traffic jams are accumulated.

“In the case of a power grid, people tried to collect the data using statistics and then create detection rules with the domain expertise to say that, for instance, if voltage surges by a certain per cent, the grid operator should alert. Even with statistical data analysis, such rule-based systems require expertise and a lot of labour. “We show that we can automate the process and also learn patterns using advanced machine-learning technologies,” said Jie Chen, senior author and manager of MIT-IBM Watson AI Lab.

Probabilities of probing

An anomaly is an event with a low probability of happening, such as a spike in voltage. The power grid data is treated as a probability distribution. This allows them to estimate probability densities and identify low-density values within the dataset. Anomalies are data points that are less likely to occur.

It is difficult to estimate these probabilities, particularly since each sample contains multiple time series. Each time series is a collection of multidimensional data points recorded over time. The sensors that collect all this data are conditional, meaning they can impact each other in certain ways.

Researchers used a special deep-learning model called a normalizing flow to determine the conditional probability distribution. This is especially effective for estimating the probability density.

A Bayesian graph augmented the normalizing flow model. This type of graph can learn complex causal relationships between sensors. Chen explained that this graph structure allows researchers to spot patterns in data and more accurately estimate anomalies.

“The sensors interact with one another, and they have causal relations and depend on each others,” he says. He says that it is necessary to incorporate this dependency information in the way we calculate probabilities.

The Bayesian network factors or breaks down the combined probability of multiple time series data into simpler, conditional probabilities, which are easier to parameterize and learn. Researchers can then estimate the probability of certain sensor readings being observed and identify anomalies that are a low probability of occurring.

The model can learn graphs on its own in an unsupervised way, which is why it is so powerful.

A powerful technique

The researchers tested the framework to identify traffic, power grid, and water system data anomalies. They used datasets that contained anomalies previously identified by humans to test the framework. The researchers were then able to compare the anomalies they identified with the real glitches found in each system.

Their model detected a higher percentage of true anomalies than all other baselines, allowing them to outperform them.

“A lot of baselines don’t include graph structure. This supports our hypothesis. Chen said that figuring out the dependency relationships among the graph nodes is a great help.

They are also flexible in their methodology. They can also tune their model using large unlabeled data to predict anomalies in other situations like traffic patterns.