Ensemble-Based Anomaly Detection for Urban Power Efficiency Towards Sustainable Energy Management

Cities are living organisms that breathe electricity. From the lights in a home to the power running a hospital, the flow of energy keeps modern life moving. But like any complex system, this flow is not always perfect. Sometimes, power is wasted through hidden inefficiencies, equipment failures, or irregular usage patterns that go unnoticed until a bill arrives or a system fails. Finding these hidden problems is difficult because the data generated by smart meters and sensors is vast, messy, and constantly changing. Traditional methods of looking for errors, such as setting a simple limit and waiting for a number to cross it, often miss the subtle signs of trouble or raise false alarms. To solve this, researchers are turning to a branch of computer science called machine learning, where computers learn to recognize patterns without being explicitly told what to look for. By teaching these systems to understand what "normal" energy use looks like, they can spot the strange behaviors that signal waste or danger, offering a path toward cleaner, more sustainable cities.

A team of researchers from universities in India, Iraq, Malaysia, and Bahrain set out to build a system that could do exactly this for urban power grids. Their goal was to create a digital watchdog that could sift through raw energy consumption data and identify anomalies—those irregular spikes or drops that suggest something is wrong. They did not rely on a single method to find these errors. Instead, they built a team of four different computer models, each with its own way of thinking about the data. One model, known as an Isolation Forest, works by randomly chopping up the data to see which points stand alone. Another, called a Local Outlier Factor, looks at how crowded a specific area of data is, flagging points that are too far from their neighbors. A third uses a technique called One-Class Support Vector Machine to draw a boundary around normal behavior, marking anything outside that line as suspicious. The fourth and most complex model is an Autoencoder, a type of artificial neural network that tries to rebuild the data it sees; if it struggles to reconstruct a piece of information, it assumes that piece is unusual.

The researchers began by cleaning a dataset of municipal power records, which involved filling in missing numbers, adjusting for skewed data, and scaling the values so the computer could process them fairly. They then trained these four models on this cleaned data. To test how well they worked, they simulated a scenario where ten percent of the data was known to be abnormal, a common rate for testing such systems. The results showed that while all the models had their strengths, the Autoencoder stood out. It achieved a score of 0.9740 on a metric that balances finding real problems with avoiding false alarms, outperforming the other three individual models. The Isolation Forest and Local Outlier Factor also performed well, but the Autoencoder was the most precise at distinguishing between normal fluctuations and genuine inefficiencies.

However, the researchers knew that relying on just one model could be risky. If that single model made a mistake, the whole system would fail. To make the solution more robust, they combined the four models into an ensemble system. This works like a committee: a piece of data is only flagged as an anomaly if at least two of the four models agree that it is strange. This approach dramatically reduced the number of false alarms. In their tests, this combined system achieved perfect precision, meaning that whenever it raised an alert, it was correct. While its overall accuracy was slightly lower than the best single model, the trade-off was worth it for a real-world system where false alarms can waste time and resources. The team also used a tool called SHAP to explain why the models made their decisions. This analysis revealed that specific features in the data, particularly those related to peak usage times and high consumption patterns, were the primary drivers behind the alerts.

The study concludes that while deep learning models like the Autoencoder are powerful tools for spotting energy waste, combining them with other methods creates a safer, more reliable safety net for city infrastructure. The system successfully identified irregular consumption patterns without needing pre-labeled examples of every possible failure, a crucial advantage since real-world cities rarely have perfect records of every past problem. The researchers noted that their work was based on simulated ground truth data and that future efforts will need to incorporate real-time, time-series data to handle the dynamic nature of live power grids. By making these invisible inefficiencies visible, such systems offer a practical step toward the United Nations' goal of affordable and clean energy, helping city managers protect resources and ensure that the power flowing through their streets is used wisely.