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

Abstract

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