From CO ₂ Forecasting to Occupancy Inference: Performance Limits of LSTM Networks with Temporal Feature Engineering

Accurate forecasting of indoor environmental variables is essential for occupancy-aware building energy management, yet the interplay between neural network architecture, forecasting strategy, and downstream occupancy inference remains poorly understood. This study presents a comprehensive controlled evaluation framework for long short-term memory (LSTM)-based multihorizon indoor CO ₂ forecasting and downstream occupancy-proxy inference using two years of residential environmental sensor data. Established forecasting approaches are evaluated across six forecast horizons from 1 h to 48 h. A crossed strategy-by-hyperparameter experiment descriptively partitions observed mean absolute error (MAE) variation: hyperparameter configuration accounts for 80.7% at [Formula: see text] h, whereas strategy accounts for 39.1% at [Formula: see text] h, the larger of the two main-effect shares at that horizon. Temporal feature augmentation reduces MAE by 9.3% at [Formula: see text] h and 8.5% at [Formula: see text] h, while increasing MAE at [Formula: see text] h. Finally, we propose an unsupervised occupancy inference framework applying absolute threshold, rate-of-change, and adaptive threshold methods to both measured and LSTM-forecasted CO ₂ . These findings provide operational guidance for deploying neural network forecasters in intelligent building management systems.