Behaviour-Based Temperature-Sensor Role Inference and Lagged Weight-Gain Event Prediction in Multi-Year Smart-Hive Telemetry

Long-term smart-hive monitoring is complicated by uncertain temperature-sensor placement and the composite nature of hive-mass changes. We analysed BeeHUB telemetry from 2022–2025 to infer temperature-channel roles and predict daytime mass-increase events (≥0.3 kg). Role inference used fixed heuristics based on thermal offset, variability and ambient correlation. The biological analysis included 35,004 stream-days from 110 telemetry streams; prediction used 20,034 target-complete, weather-matched stream-days. During calendar-defined May–June, warm–stable days had a median inferred internal temperature of 34.8 °C and thermal buffer of 17.0 °C. Among warm–stable days with observed mass change, median daytime increase was 0.35 kg and event frequency was 52.5%. Whole-stream bootstrap intervals supported higher warm–stable mass change and event frequency than four comparison states, but not the sparse warm–unstable state. Across 50 stream-grouped holdouts, a strictly lagged unweighted histogram-gradient-boosting model achieved areas under the receiver operating characteristic curve of 0.809±0.018 in May–June and 0.797±0.026 in summer (mean ± standard deviation), compared with 0.751±0.018 and 0.727±0.030 for previous-day persistence. These findings support conditional thermal-state interpretation and autoregressive prediction, not verified nectar intake or causal thermal effects. Sensor roles lacked placement ground truth, and generalisation to unseen biological colonies has not been established.