Predicting Indoor Temperature and Humidity in a Naturally Ventilated Office Room Using Long Short-Term Memory Networks Model in a Tropical Climate

Abstract

This study evaluates the efficacy of long short-term memory (LSTM) neural networks in predicting indoor temperature and humidity dynamics for naturally ventilated office environments in Ho Chi Minh City, Vietnam (10.8231°N, 106.6297°E). The tropical climate of this location, characterized by consistently warm temperatures year-round, provides a specific context for the model's performance and applicability. A simulated office room model with diverse window opening scenarios was developed using EnergyPlus simulations, generating a synthetic dataset of hourly indoor-outdoor conditions across varying seasons. An LSTM neural network, trained on 70% of the data and tested on the remaining 30%, was employed to forecast indoor temperature and humidity at multiple time horizons. Results indicate that the LSTM achieves near-perfect short-term (1–30 min) predictions, with performance degrading at longer horizons (60–120 min) while remaining competitive with existing approaches. The model effectively captured the strong influence of window opening area and solar irradiance on indoor conditions. A comparison of different window configurations revealed their significant impact on predicted thermal dynamics. Model accuracy was assessed using the coefficient of determination, root mean square error, and mean absolute error metrics. The study demonstrates that LSTM networks can effectively learn complex non-linear building physics to forecast climate in naturally ventilated spaces. With sub-second response times, this approach shows potential for supporting real-time control and optimization of natural ventilation strategies based on probabilistic forecasts, ultimately enhancing occupant comfort and energy efficiency in buildings.