This study explores the possibility to forecast formaldehyde (HCHO) concentration from past observations in an office. A monitoring campaign of HCHO was performed during 96 days, with a short time-step (1 minute). The monitored formaldehyde time series exhibited, in particular, abrupt changes and structural breaks in variance, that cannot be modelled implicitly using simple models. To overcome this problem, hybrid model for forecasting HCHO time series (characterized by nonlinear and nonstationary behaviour) was used. The periodicities of the data were modelled by Fast Fourier Transform (FFT). A Self-Exciting Threshold AutoRegressive (SETAR) model was used to model the FFT component. SETAR models are typically designed to accommodate the nonlinear features and can explain two regimes in a time series. The residuals of the FFT component subtracted from raw data was modelled using a second SETAR model. The output of the two models were summed and compared to raw test data.