Edge-Driven Data Acquisition and Intelligence for Real-Time Energy Management in Smart Grids

Smart Grid infrastructure development comprises efficient monitoring, anomaly detection, and control of the grid with real-time data-driven means. An Edge architecture is presented in this paper for smart grid operation, wherein Smart Grid Devices such as sensors and smart meters are networked with local edge computing. The framework combines data collection, preprocessing, and edge anomaly detection with light-weight deep models of LSTM-lite and GRU-lite to enable real-time local control decisions like load shedding and voltage control. The data is also pushed to the cloud for in-depth analysis and long-term planning. Four anomaly detection methods were compared against inference time, accuracy, precision, and recall performance metrics: LSTM-lite, GRU-lite, Isolation Forest, and One-Class SVM. LSTM-lite was overall the best rounded strategy, highly accurate at 0.94 and inference time of just 12 ms that represents highly advantageous real-time relevance. GRU-lite was just slightly faster inference time at 10 ms but with slightly worse accuracy (0.92). Traditional machine learning models such as Isolation Forest and One-Class SVM were significantly slower to make inference (45 ms and 60 ms, respectively) and were of low recall and accuracy, indicating the efficiency of deep learning-capable edge solutions in providing instant detection. This paper highlights the viability of employing light neural networks on edge devices within the Smart Grid for the purpose of delivering timely and precise anomaly detection and enabling responsive control actions.