Development of an LSTM broadcasting deep-learning framework for regional air pollution forecast improvement

• DOI: 10.5194/gmd-15-8439-2022
• Publication Date: 2022-11-21T09:27:35Z
• Authors: Haochen Sun; Jimmy C. H. Fung; Yiang Chen; Zhenning Li; Dehao Yuan; Wanying Chen; Xingcheng Lu
• Keywords: Interpolation

Abstract. Deep-learning frameworks can effectively forecast the air pollution data for individual stations by decoding time series data. However, most of existing time-series-based deep-learning models use offline spatial interpolation strategies and thus cannot reliably project station-based to region interest. In this study, long short-term memory (LSTM) technique was extended quality forecasting combining a novel layer, termed broadcasting which incorporates learnable weight decay parameter designed point-to-area extension. Unlike deep-learning-based methods that isolate from model training process, proposed end-to-end LSTM framework consider temporal characteristics relationships among different stations. To validate framework, PM2.5 O3 forecasts next 48 h were obtained using 3D chemical transport simulation results ground observation as inputs. The root mean square error associated with 40 % 20 lower than those Weather Research Forecasting–Community Multiscale Air Quality an combination methods, respectively. be in other regions