Energy Manager Based on Reinforcement Learning for a Low-Power IoT Collar Worn by Dairy Cows

This study presents an energy manager (EM) based on reinforcement learning (RL) for low-power wearable IoT devices with solar energy harvesting used in precision livestock farming (PLF). The system utilizes battery level, device orientation, weather forecasts, and time of day to dynamically adjust the device’s energy consumption by modifying the monitoring duty cycle, the number of transmissions, and GPS executions to balance the energy harvested from flexible solar panels with energy consumption. Using the twin delayed deep deterministic policy gradient (TD3) algorithm, the EM learns how to perform these adjustments. Model-based simulations over extended periods, along with real-world deployments, demonstrate the system’s ability to achieve energy neutral operation (ENO), outperforming state-of-the-art algorithm in terms of harvested energy utilization. Additionally, a statistical equivalence test (TOST) is also applied to formally validate that the system maintains energy neutrality. The lightweight neural network design ensures compatibility with resource-constrained devices. This work addresses the scarcity of RL-based EM implementations in real-world wearables and PLF applications and demonstrates their potential for use in such scenarios.