Publication

Abstract : A two-layer Long Short-Term Memory (LSTM) net- work is proposed for forecasting indoor CO2 concentration. The model was trained on four weeks of time-series data containing occupancy information and CO2 measurements, structured with a 60-step sliding window and normalized via MinMax scaling. The LSTM achieved high predictive accuracy (MAE = 6.59 ppm, RMSE = 20.06 ppm, R2 = 0.9989) across varying occupancy levels, including peak conditions. Reliable occupancy inputs were found critical, as noisy or lagged signals reduced accuracy. Evaluation with a rolling-origin split confirmed robust generalization, and inference latency tests (< 20 ms) demonstrated feasibility for edge deployment in smart ventilation and building energy management applications.