Correction methods for statistical models in tropospheric ozone forecasting

This study proposes two methods to enhance the performance of statistical models for prediction tropospheric ozone concentrations. The first method corrects the statistical model based on the average daily profile of the model errors in training set. The second method estimates the model error by making the analogy with three basic modes of feedback control: proportional, integral and derivative. These correction methods were tested with multiple linear regression (MLR) and artificial neural networks (ANN) for prediction of hourly average tropospheric ozone (O 3) concentrations. The inputs of the models were the hourly average concentrations of sulphur dioxide (SO 2), carbon monoxide (CO), nitrogen oxide (NO), nitrogen dioxide (NO 2) and O 3, and some meteorological variables (temperature – T; relative humidity – RH; and wind speed – WS) measured 24 h before. The analysed period was from May to June 2003 divided in training and test periods. ANN presented slightly better performance than MLR model for prediction of O 3 concentrations. Both models presented improvements with the proposed correction methods. The first method achieved the highest improvements with ANN model. However, the second method was the one that obtained the best predictions of hourly average O 3 concentrations with the correction of MLR model. ► The model errors were estimated by making the analogy with feedback control. ► The value of R 2 increased 115% for MLR and 105% for ANN with the correction method. ► MLR overperformed ANN in prediction of tropospheric O 3 concentrations.