Abstract:

To handle inaccurate observations and errors in data analysis and assimilation, a single prediction is only one possible solution. To solve this problem, this study proposes an O₃ prediction model based on an artificial neural network ensemble prediction. To construct the artificial neural network forecast model, 8 types of meteorological factors and 2 types of pollutant factors are considered. Furthermore, the random disturbance method is used to create 15 sets of mutually independent random disturbance weather fields using data from May to September of each year, starting from 2013 and ending with 2019 as the training set and those from May to September 2020 as the test set. Results show that compared with a single artificial intelligence network prediction model, the proposed ensemble model clearly shows higher accuracy. The O₃ pollution hit rate is improved obviously, the nonresponse rate is remarkably reduced, and the empty rate is slightly higher than that of the single model. The O₃ pollution is predicted to happen more often using the proposed ensemble model, whereas that using the single model tends to be less. Considering a heavy O₃ pollution condition that occurred on July, 3 to 9, 2020 as an example, the proposed ensemble model can reflect the rapid cumulative increase and continuous process of O₃ pollution better than the single model. The proposed ensemble model can facilitate the probability of various occurrence, uncertainties and other more forecast information by providing quantitative probabilistic forecasts, which have certain practical application values.

Key words: :artificial neural network, ensemble prediction, random disturbance, air quality, the maximum 8-hour moving average of O₃, O_x