LoRaWAN-Enabled Predictive Maintenance for Distribution Transformers: A 24-Hour Simulation Study Using Machine Learning

The operational health of distribution transformers is critical for ensuring uninterrupted power delivery across smart grid infrastructures. This paper presents a predictive maintenance framework for 11kV/433V distribution transformers using machine learning, supported by a simulated LoRaWAN communication layer. A synthetic dataset comprising 876,000 hourly samples from 100 transformers was generated in MATLAB, incorporating realistic fault conditions such as thermal anomalies and oil level deviations. Two ensemble models-Light Gradient Boosting Machine (LightGBM) and Extreme Gradient Boosting (XGBoost)-were trained on the imbalanced set, achieving 99.9% overall accuracy, 0.97 precision, and 1.00 recall on the fault class. To simulate real-time grid monitoring, a Python-based LoRaWAnmodule introduced key parameters (e.g., spreading factor, bandwidth, duty cycle), packet loss, latency, and missing sensor values across 24 hourly windows. Fault predictions were visualized on a transformer-hour heatmap and a temporal fault histogram to assess system responsiveness. The simulation showed that the models remained robust despite transmission constraints, with an estimated energy cost of approximately 3.36 mWh per day per device. This integrated framework offers a scalable and energy-efficient alternative to SCADA systems and is suitable for edge-deployable IoT gateways in rural or low-infrastructure areas. Future work will explore hardware-in-the-loop validation and cybersecurity resilience.