An Artificial-Intelligence-Based Predictive Maintenance Strategy Using Long Short-Term Memory Networks for Optimizing HVAC System Performance in Commercial Buildings

This study addresses the persistence of avoidable failures and efficiency losses in HVAC plants by introducing a field-validated predictive maintenance (PdM) framework that estimates component-level RUL from multiyear BMS telemetry and translates forecasts into schedule-aware maintenance actions. The objective was to determine whether an LSTM ensemble with mode-aware segmentation and isotonic calibration could yield decision-quality RUL forecasts that reduce unplanned outages, downtime, and electricity use in a large Riyadh office building. Two years of 1 min BMS data from chillers, primary pumps, and AHU fans were cleaned, standardized, and segmented by operating mode; RUL labels were derived from time-stamped work orders and failure confirmations; the LSTM produced per-minute RUL estimates trained with a Huber loss, calibrated to lower quantiles, and converted to sustained triggers compared against a fixed-interval program. On the held-out test set, the model achieved a weighted MAE of 19.8 ± 2.1 h and RMSE of 29.1 ± 3.3 h, with quantile calibration error (QCE) ≤0.06 and lead-time accuracy (LTA; fraction of triggers whose calibrated lower-quantile RUL is ≥ the planning threshold) of 0.79 at a 10-day threshold. When deployed in counterfactual evaluation, triggers reduced unplanned outages by 47.6% (paired bootstrap p = 0.008) and total downtime by 41.3% (p = 0.012), and yielded a 10.6% reduction in HVAC electricity (95% CI: 7.7–13.2%) and a 9.7% decrease in total operating cost. The findings indicate that calibrated sequence models coupled to simple sustained triggers can convert routine BMS data into reliable maintenance schedules with quantifiable reliability and energy benefits. Practically, conservative calibration (q approximately 0.25) with thresholds of 10–12 days provided stable lead windows; future work should assess transferability across climates and facility types using transfer learning and integrate uncertainty-aware triggering with MPC for joint operational and maintenance optimization.