Energy-Efficient Resource Accounting for IoT-Based AMI Using Federated Reinforcement Learning and TinyML Optimization
Abstract
Advanced Metering Infrastructure (AMI) and Internet of Things (IoT) deployments require energy-efficient resource accounting to scale in power distribution systems. Existing approaches require centralized data collection, neglect edge compute and quantization costs, and poorly handle temporal nonstationarity. This paper proposes EERAF-Tiny, a Federated Reinforcement Learning (FRL) agent integrated with TinyML optimization for on-device inference and compressed deployment. EERAF-Tiny used Proximal Policy Optimization (PPO) in a federated aggregation loop and applied structured pruning, post-training quantization, and compiler-level kernel generation to reduce inference energy and model size. The method learned per-node accounting policies that balanced estimation RMSE, computed energy, communication, and carbon proxies. Evaluation used a public AMI dataset simulated into 1,000 heterogeneous IoT meter nodes under realistic network conditions. Experiments compared EERAF-Tiny to XGBoost and a federated DDQN hybrid baseline. EERAF-Tiny reduced accounting RMSE by 28.3% versus XGBoost and 15.1% versus Fed-DDQN while lowering per-inference energy by 42 mJ, model size by 78 KB, and communication per round by 67 kB. These results indicate that FRL combined with TinyML enables accurate, low-energy resource accounting for IoT-based AMI while preserving data locality and privacy and security.