Forecast of air quality pollutants' concentrations based on BP neural network multi-model ensemble method
ZHANG Heng-de1,2, ZHANG Ting-yu2, LI Tao2, ZHANG Tian-hang1
1. Nation Meteorological Center of China Meteorological Administration, Beijing 100081, China;
2. School of Electronic & Information Engineering, Nanjing University of Information Science and Technology, Nanjing 210044, China
Based on the forecast products of three numerical models including CUACE, BREMPS and WRF-Chem, a multi-model ensemble forecast system was established by BP neural network. Firstly, the training function, the nodes number of hidden layer and the length of training samples of BP neural network were determined to be trainbr, ten and fifty, respectively, by sensitivity experiments. Then, 3sites in Beijing, Tianjin and Shijiazhuang were selected to evaluate the performances of ensemble forecast system. Results showed that (1) Compared with single models, the normalized mean biases, the root mean square error, and the correlation coefficient between forecasted and observed pollutant concentrations in 3~72hours decreased from -100%~200% to -20%~20%, decreased by 15%, and increased from 0.1~0.8 to 0.3~0.85, respectively, indicating that the forecast results were better than single models. (2) The TS scores of AQI values in mild and medium pollution events in 2016 in Beijing of ensemble forecast system were 22% and 10% higher than those results of CUACE model, respectively. The rate of vacancy forecast and missing forecast of heavy pollution in Tianjin decreased by 31% and 25%, respectively. (3) The forecasted and observed trends of PM2.5 concentrations in December 2016 were consistent well with each other.
Zhou C H, Gong S, Zhang X Y, et al. Towards the improvements of simulating the chemical and optical properties of Chinese aerosols using an online coupled model CUACE/Aero[J]. Tellus Series B-chemical & Physical Meteorology, 2012,64(1):91-102.
Monache L D, Stull R B. An ensemble air-quality forecast over western Europe during an ozone episode[J]. Atmospheric Environment, 2003,37(25):3469-3474.
[15]
Pagowski M, Grell G A, Mckeen S A, et al. A simple method to improve ensemble-based ozone forecasts[J]. Geophysical Research Letters, 2005,32(7):351-394.
Zhi X F, Qi H X, Bai Y Q, et al. A comparison of three kinds of multi-model ensemble forecast techniques based on the TIGGE data[J]. Acta Meteorologica Sinica, 2012,26(1):41-51.
Ding W, Zhang J, Leung Y. Prediction of air pollutant concentration based on sparse response back-propagation training feedforward neural networks[J]. Environmental Science & Pollution Research, 2016,23(19):1-14.
