Analysing the Impact of Data Normalization on Levenberg-Marquardt Backpropagation (LMBP) Neural Network Predictions for Indoor Air Quality

This study investigates the influence of data normalization (DN) on the accuracy of indoor air quality predictions within building environments, specifically using Back Propagation (BP) algorithm. Including Maximum-Minimum scaling, mean normalization, and Z-score are rigorously examined using real-world experimental datasets. The results of this research highlight that the combination of the BP algorithm with Min-Max normalization leads to the most precise prediction results, as indicated by the lowest Coefficient of Variation of the Root Mean Square Error (CVRMSE) and Normalized Mean Bias Error (NMBE). The remaining three normalization methods also offer reasonable accuracy in forecasting indoor air quality, with only minor variations in performance. For the LMBP (Levenberg-Marquardt Back-propagation) model with Z-score normalization, it demonstrates effective performance in predicting indoor air quality in residential buildings, although differences between the LMBP models are minimal. This study underscores the considerable impact of data normalization on prediction accuracy when utilizing artificial neural network models. It recommends using Min-Max and Z-score normalization for LMBP and working with raw data to achieve optimal predictive results. This study offers valuable guidance for improving the precision of indoor air quality predictions through appropriate DN techniques, making these findings applicable to various applications and practitioners in the field.