Towards Energy-Efficient Buildings

Timothy Mulumba

Keywords: Time Series, machine learning, fault detection and diagnosis

Abstract:

Heating, Ventilation, and Air Conditioning (HVAC) systems account for a large share of building energy use. We study a deployment-oriented ARX to SVM pipeline for chiller Fault Detection and Diagnosis (FDD) that builds on prior ARX–hybrid work benchmarked on the ASHRAE RP-1043 dataset and related online extensions.
Our contributions are: (i) a practitioner-focused procedure for tuning the exponential forgetting factor on F1 to align estimator memory with detection objectives; (ii) a clean ablation contrasting dynamic ARX features with a static regression baseline at matched feature budgets; and (iii) a latency/memory audit showing sub-millisecond per-timestep updates on commodity hardware.
On RP-1043, the pipeline attains competitive macro-F1 against stronger baselines while preserving interpretability and runtime efficiency. We qualify claims with time-series-aware uncertainty estimates and paired significance tests, and discuss maintenance under concept drift.

Submission Number: 21