In India, there are power grids connecting through some of the busiest cities and crowded streets. In many parts of the country, we regularly face power failures, voltage fluctuations and short-circuiting, often leading to a huge drop in productivity. Now, what if our authorities can detect anomalies in our power grid failure way before these failures happen?This timely detection can aid in the smooth functioning of the power grid and avoid any danger that occurs due to fluctuations. 

Scientists at the MIT-IBM Watson AI Lab have developed an AI-based solution to this problem. The computationally efficient technique developed by them can automatically identify any anomalies in the data streams in an instant. The team stated that their technique, which learns to model the interconnectedness of the power grid, is much more skilled at spotting these problems than several other conventional methods.  

This Machine Learning model developed by the team does not require data on power grid glitches. Therefore, the daily application of this technique becomes much easier, where high-quality labelled data sets are often difficult to find.  

The model is versatile in its application. It can be applied to other circumstances where there are interconnected sensors gather and report, like traffic tracking systems. Deploying this model to traffic systems would help detect traffic bottlenecks or to expose how traffic jams would flow.  

In normal cases, when voltage surges, the grid operator should be alerted. This process might be time-consuming and might require a lot of labour force. The use of the AI model will make the process easier and faster.  

Functioning of the model  

The team started the study by labelling an event that has less chance for an occurrence, such as an unexpected spike in voltage. According to the team, this power grid is a probability distribution, so if they can approximate the probability densities, then they will be able to detect the low-density values in the datasets. The data points which have the least chance for occurrence correspond to irregularities.  

Since there are numerous time series and each time series is a set of multidimensional data points captured over time, approximating those probabilities is not an easy task. Also, the sensors that record the data are linked in a manner that one sensor can occasionally influence others.  

Therefore, to understand the distribution of data, scientists made use of a deep learning model called normalizing the flow. This model is most effective at approximating the probability density of a sample.  

To learn the casual relationship structure between the sensors, researchers made use of a graph called the Bayesian network. According to Jim Chen, Senior Study Author and Research Staff Member and Manager at IBM Watson AI Lab, MIT, this graph helps the scientists to observe patterns in data that can approximate anomalies much more accurately. It breaks down the probability of the numerous time series into less intricate probabilities that are easier to parameterize, learn and assess. The technique is robust because the model can learn the graph without any supervision.  

Result of Testing  

The team verified the functioning of the framework by observing how well it could detect glitches in traffic data, power grid data and water system data. The datasets that were used for testing had anomalies detected by humans. Therefore, the scientists were able to compare both findings and understand the effectiveness.  

According to the result gathered, the team states that their model outdid all the baselines by sensing a higher percentage of anomalies in each dataset. With a large unlabeled dataset, they could tweak the model to make estimations in real-time. Once the model is installed, it would learn from the steady flow of the new sensor data. Chen and his team are looking forward to employing this method to create models that map other intricate relationships. They are also investigating how they can competently learn these models when the graph is massive with millions of interlinked nodes.  

Moving forward, they could use this method to enhance the accuracy of forecasts based on datasets or simplify other classification methods.