A Large-Scale IoT Telemetry Framework for Edge-Aware Forecasting and Diagnostic State Tracking in Precision Apiculture
Abstract
Precision apiculture increasingly relies on Internet of Things (IoT) telemetry, yet forecasting, short-horizon state tracking, and embedded deployment are often evaluated separately. This study presents a deployment-oriented BeeHUB framework integrating multimodal sensing, hive-weight forecasting, diagnostic state tracking, and physically validated edge inference. Two non-equivalent analysis cohorts are reported separately: a full-period cohort of 2,531,844 records from 206 device streams and a weather-enriched cohort of 1,645,074 observations from 116 GPS-resolved hives collected between January 2022 and February 2025. In the 72-h full-period experiment, Linear Regression and Ridge Regression achieved a mean absolute error of 2.111 kg and an <inline-formula> <tex-math notation="LaTeX">$R^{2}$ </tex-math></inline-formula> of 0.962, while a persistence reference remained stronger with a mean absolute error of 1.293 kg. The exact frozen linear-model artifacts were converted to allocation-free C implementations and evaluated on an ESP32-D0WD-V3 platform. Each implementation increased compiled Flash usage by 1216 bytes and required no additional static RAM. Median kernel latency was approximately <inline-formula> <tex-math notation="LaTeX">$11.2~\mu $ </tex-math></inline-formula>s, while complete feature processing and inference required approximately <inline-formula> <tex-math notation="LaTeX">$60.1~\mu $ </tex-math></inline-formula>s. Numerical equivalence with the original Python artifacts was confirmed for 1024 test vectors, with maximum absolute differences below <inline-formula> <tex-math notation="LaTeX">$9\times 10^{-14}$ </tex-math></inline-formula>. Direct board-level measurements using a Nordic Power Profiler Kit II yielded approximately <inline-formula> <tex-math notation="LaTeX">$2.46~\mu $ </tex-math></inline-formula>J per kernel inference and <inline-formula> <tex-math notation="LaTeX">$13.79~\mu $ </tex-math></inline-formula>J per complete pipeline. The principal contribution lies in system-level integration, auditable experiment provenance, and physically verified edge deployment rather than in proposing a new forecasting algorithm.