RNN-based Forecasting of Indoor Temperature in a Naturally Ventilated Residential Building

Abstract

Natural ventilation is an effective passive control strategy to improve building energy efficiency, indoor thermal comfort and air quality. Near real-time model-based control of window openings for natural ventilation requires forecasts completed within a short time, typically seconds. However, widely-used physics-based simulation is time-consuming and entails high computation cost. This study is aimed at developing a Recurrent Neural Network (RNN) model for forecasting indoor temperature using seasonal window opening schedules and ambient conditions. The data-driven forecasting approach utilizes simulated indoor dry-bulb temperature (DBT) and relative humidity (RH) based on ambient parameters (DBT, RH, and wind speed and direction), time (time of day, day of year), building thermal parameters and window characteristics (location, opening size and type) to train the RNN model. The results show that the proposed RNN algorithm is effective in predicting indoor environmental conditions with considerable accuracy (R2 = 0.956) and outperforms statistical methods by at least 20% in the same measure. Window opening area is highly correlated to the hourly change of indoor temperature. Prediction errors for indoor operative temperature are less than 1°C for 70% of the time and less than 2°C for 93% of the time. The speed and accuracy of a trained network illustrate the potential of the method for near-real time control of buildings and systems while maintaining occupant thermal comfort.