With increasing ozone pollution and deeper understanding of its harm to humans and climate, it is important to accurately forecast ozone. In this study, training and testing data sets were constructed with hourly numerical models forecasts and monitoring station observation for the year 2018 for Sichuan-Chongqing region, China. Three machine learning methods including Lasso, random forest and long short-term memory recurrent neural network (LSTM-RNN) coupled with CMAQ model were trained to forecast the ozone concentrations. The Lasso regression and random forest were used to realize feature optimization in four sub-regions separately. Coupled model with Lasso-random forest coupled feather selection schemes showed the best performance among different models. The main conclusions of adjusting results showed that deviations of hourly ozone prediction by CMAQ alone forecasts can be significantly reduced after machine learning coupled model adjusting, and correlation coefficients can be remarkably improved. Adjusting effects varied with different sub-regions and seasons. In three basin sub-regions, adjusting with random forest had the best performance, while in the plateau sub-region, adjusting with LSTM-RNN was most satisfactory, where root mean squared error decrease rate was 80.2% and correlation coefficient reached 91%. Machine learning methods performed better in summer and autumn for the three basin sub-regions, while in the plateau sub-region, adjusting was more significant in summer compared to other seasons.