Dilated Causal CNNs for Energy Forecasting and Optimization in LoRaWAN Networks

Battery lifetime remains a central constraint in scaling LoRaWAN deployments across diverse IoT applications. We propose a lightweight dilated causal Convolutional Neural Network (CNN) designed to forecast per-node energy consumption with high temporal fidelity. Unlike recurrent models, our approach captures both transient spikes and long-range patterns without sequential overhead, enabling efficient edge deployment. Trained on a 12-month NS-3 simulation dataset encompassing smart lighting, environmental monitoring, waste management, and agriculture, the model achieves 96.5% forecasting accuracy with mean absolute error below 0.3, improving over SARIMA and LSTM baselines by 55% and 32% respectively. We integrate this predictor into an end-to-end energy optimization pipeline where on-device inference executes every 15 minutes with under 5 ms overhead. Forecasts drive adaptive duty-cycling, transmission slotting, and data rate control, extending device lifetime by 20%, halving collision rates, and improving fairness by 15%. Real-world validation on ten STM32F407VG microcontrollers and a commercial RAK7258 gateway confirms practical feasibility: inference completes within 0.8 ms with 4.4 mW peak power draw and 95.6% packet delivery ratio. These results demonstrate the CNN's suitability for real-time, edge-centric forecasting and its potential for enabling sustainable, intelligent LoRaWAN networks.