Time series prediction of CO2, TVOC and HCHO based on machine learning at different sampling points

Abstract

This paper presents novel findings about the prediction of TVOC and HCHO using the machine learning approach. Continuous measurements of CO2, TVOC and HCHO were conducted in five rooms of SDE, NUS. The collection data was trained and tested by 4 machine learning algorithms including support vector machine (SVM), Gaussian processes (GP), M5P and backpropagation neural network (BPNN). Overall, SVM scored the highest in performance evaluation because it has the highest average prediction accuracy and fewer overfitting in the test data. High predictability due to large autocorrelation was observed in the pattern analysis of CO2 and TVOC. Accurate results were achieved by SVM for CO2 and TVOC, with mean MAPE of being 1.87% and 2.30%, respectively. In contrast, low autocorrelation indicated the hidden mode of HCHO data was more difficult to capture than CO2 and TVOC. The small R2 between predicted and actual values of HCHO demonstrated low predictability, ranging from 0.0008 to 0.0215.