A Comparative Analysis of Pruning, Quantization, and Compilation for LightGBM-Based Electricity Anomaly Detection on IoT Edge Devices

Operating machine learning models designed for electricity theft detection on resource-limited Internet of Things (IoT) edge devices involves significant trade-offs among inference speed, power utilization, and detection precision. This paper provides the first detailed comparison of pruning, quantization, and compilation strategies on Raspberry Pi 4 hardware for Light Gradient Boosting Machine (LightGBM)-based anomaly detection. We tested three optimization strategies: pruned models, Open Neural Network Exchange (ONNX) 8-bit integer (INT8) quantization, and Treelite compilation against a baseline native version for 1,000 inference cycles under practical operating conditions. ONNX INT8 quantization produced a 40.66× speedup for real-time inference (0.091 ms), but exhibited significant thermal load (80.67% CPU utilization, +7.3 °C). Treelite offered best-in-class batch processing efficiency (13.41× speedup, 25.99% CPU utilization). All approaches retained 96.98% accuracy and produced an identical confusion matrix to the baseline model. Experiments indicate that the choice of the optimization strategy hinges critically on deployment scenarios, such as real-time streaming versus occasional batch processing. Our results give practical recommendations for implementers of edge-based grid smart-monitoring systems.