A pattern recognition approach for modeling the air change rates in naturally ventilated buildings from limited steady-state CFD simulations

Calculating the air change rates inside naturally ventilated buildings is essential for many applications including indoor temperature calculation. The air change per hour (ACH) at a particular time step and ambient conditions (wind speed, direction, temperature, etc.) is usually calculated via sophisticated simulators, e.g., CFD. However, having a mathematical model describing the relationship between ACH (output variable) and other ambient conditions (input variables), rather than mere simulated numbers, is very important for several reasons: understanding the nature of this relationship and its dominating variables; calculating the indoor temperature at arbitrary time steps; and saving the enormous simulation time when simulating long spans. In this article, a novel approach from pattern recognition literature is introduced to model ACH. A Classification and Regression Tree (CART) was designed from 180 CFD simulated values of ACH, along with their experimentally measured ambient conditions. The RMS error between the ACH values predicted by CART and those simulated by CFD is calculated using cross validation and found to be very acceptable (0.78 and 1.48) h−1 for two different rooms. The model revealed that the ambient temperature is not predictive and hence was dropped from the final model. Designed CART was then fed to TRNSYS 17 as an equation where variables defined as algebraic functions to produce an hourly output for ACH and calculate the indoor temperature along the summer season. The absolute deviance error between the indoor temperatures simulated by TRNSYS and those measured experimentally is as low as (0.3 and 0.4) °C for the two rooms respectively.