Edge-Coordinated Lightweight Quantized ANFIS for Adaptive Microclimate Prediction in Energy-Constrained IoT Systems

Abstract

This study addresses the challenge of deploying collaborative adaptive intelligence on energy-constrained IoT edge nodes, a critical requirement for real-time microclimate prediction in smart agriculture. We propose a novel Edge-Coordinated Lightweight Quantized Adaptive Neuro-Fuzzy Inference System (LQ-ANFIS). The framework combines a quantized fuzzy neural model for local inference on 8-bit microcontrollers (WeMOS D1 Mini) with a lightweight MQTT-based coordination mechanism. This mechanism enables distributed nodes to achieve synchronized adaptation by periodically exchanging only scalar parameters, namely bias and learning rate, through a broker (Raspberry Pi), thereby eliminating the need for cloud infrastructure or heavy model transfers. During a seven-day experimental deployment involving three nodes, each collecting more than 10,000 temperature and humidity samples, the system demonstrated robust prediction accuracy (RMSE ≈ 1.89%, MAE ≈ 1.13%) and high energy efficiency (average power < 100 mW per node). Compared with a conventional uncoordinated ANFIS baseline, the proposed method achieved a 21% improvement in prediction accuracy and a 39% reduction in energy consumption. The results confirmed rapid inter-node bias convergence (σ < 0.05) and low coordination latency (< 50 ms), validating its real-time adaptability at the edge. These findings support scalable, cooperative intelligence for resource-constrained agricultural IoT deployments worldwide.