In response to the widespread and severe problem of air pollution, especially the mounting threat of ozone pollution to environmental and public health, an ensemble forecast model that combines three deep learning models LSTM, TCN, Transformer and the metaheuristic algorithm Kepler optimization algorithm is proposed. Predictive factors are selected based on the correlation analysis among ozone, other air pollutants and meteorological elements. Independent forecast is conducted by LSTM, TCN and Transformer respectively. The Kepler optimization algorithm is employed to integrate the predictive outputs of above three deep learning models to obtain final predictive results. Experimental results indicate that, on the multi-step forecast for hourly concentration of ozone in Beijing, the proposed ensemble forecast model surpasses deep learning models LSTM, TCN and Transformer as well as classical multiple linear regression and random forest models with RMSE, MAE and R2. This demonstrates that the ensemble forecast strategy integrating deep learning with metaheuristic algorithm could effectively improve the accuracy and stability of predictor, and the ensemble forecast approach based on deep learning is feasible to ozone concentration prediction.