Anomaly Detection in Smart Utility Grids using Temporal Fusion Transformers on Consumption Data Streams
Abstract
Smart utility grids make use of continuous consumption data streams to make the best delivery of energy and identify abnormal use patterns. Smart grids involve the detection of abnormality to prevent energy theft, equipment malfunction, as well as enhance efficiency on operation. The old methods are being spoiled by the weaknesses in terms of temporal dependency and real-time learning capabilities, so they cannot be applied to changing circumstances. It is a work that introduces Real-Time Detection and Prediction model of Abnormal Utility Usage in residential or industrial spaces driven by Temporal Fusion Transformer (TFT). The new approach takes advantage of the capability of TFT to form long term dependencies, multiple input variables, and inclusion of attention mechanisms that allow explainability in forecasting. It allows real-time and correct detection of anomaly in the use of electric power, water, or gas. The model was tested with real utility usage data and it performed better than the traditional machine learning models both in terms of speed and accuracy in detection. Through experimental evidence, TFT-based system has demonstrated to detect malicious usage patterns at a higher degree of reliability and effectiveness in smart utility grids.