In this study, enhancements of several functional techniques are given in order to forecast sulfur dioxide levels near a power plant. The data are considered as a time series of curves. Assuming a lag one dependence, the predictions are computed using the functional kernel (with local bandwidth) and the linear autoregressive Hilbertian model. We carry out the estimation with a so-called historical matrix, which is a subsample that emphasizes uncommon shapes. A bootstrap method is introduced to evaluate the range of the forecasts, which uses Fraiman and Muniz's order for functional data. Finally, we compare our functional techniques with neural networks and semi-parametric methods, and find that the former models are often more effective.
