Explore a deep learning multi-output neural network for regional multi-step-ahead air quality forecasts Yanlai Zhou a , Fi-John Chang a, * , Li-Chiu Chang b , I-Feng Kao a , Yi-Shin Wang a a Department of Bioenvironmental Systems Engineering, National Taiwan University, Taipei,10617, Taiwan, ROC b Department of Water Resources and Environmental Engineering, Tamkang University, New Taipei City, 25137, Taiwan, ROC article info Article history: Received 2 July 2018 Received in revised form 26 September 2018 Accepted 22 October 2018 Available online 24 October 2018 Keywords: Multi-output LSTM Deep learning Artificial intelligence (AI) Multi-step-ahead forecast Air quality Taipei city abstract Timely regional air quality forecasting in a city is crucial and beneficial for supporting environmental management decisions as well as averting serious accidents caused by air pollution. Artificial Intelligence-based models have been widely used in air quality forecasting. The Shallow Multi-output Long Short-Term Memory (SM-LSTM) model is suitable for regional multi-step-ahead air quality fore- casting, while it commonly encounters spatio-temporal instabilities and time-lag effects. To overcome these bottlenecks and overfitting issues, this study proposed a Deep Multi-output LSTM (DM-LSTM) neural network model that were incorporated with three deep learning algorithms (i.e., mini-batch gradient descent, dropout neuron and L2 regularization) to configure the model for extracting the key factors of complex spatio-temporal relations as well as reducing error accumulation and propagation in multi-step-ahead air quality forecasting. The proposed DM-LSTM model was evaluated by three time series of PM 2.5 , PM 10, and NO x simultaneously at five air quality monitoring stations in Taipei City of Taiwan. Results indicated that the loss function values (mean-square-error) of the SM-LSTM and DM- LSTM models in the testing stages at horizon tþ4 were 0.87 and 0.72, respectively. The G bench values of the DM-LSTM model in the testing stages for PM 2.5 , PM 10, and NO x reached 0.95 at horizon tþ1 and exceeded 0.81 at horizon tþ4, respectively. Results demonstrated that the proposed DM-LSTM model incorporated with three deep learning algorithms could significantly improve the spatio-temporal sta- bility and accuracy of regional multi-step-ahead air quality forecasts. © 2018 Elsevier Ltd. All rights reserved. 1. Introduction Exposure to ambient air pollution is a primary environmental risk factor in relation to adverse health impacts (Apte et al., 2015; Liu et al., 2017). Fine particulate matter (PM 2.5 and PM 10 , i.e., par- ticles smaller than 2.5 or 10 mm) and nitrogen oxide (NO x ) are the dominant components of ambient air pollution associated with booming urban development (Li et al., 2017, 2018; Lin and Zhu, 2018). To date, epidemiological investigations and studies demon- strated that some air pollution-related diseases were associated with exposure to PM 2.5 , PM 10, and NO x (Reggente et al., 2014; Wang et al., 2016; Wu et al., 2018a,b). In addition, these air pollutants were acknowledged as typical representatives of particle number concentration in urban air quality (Li et al., 2018a,b; Wu et al., 2018a,b). Real-time air quality information is of great importance to air pollution control and human health protection from air pollution (Ni et al., 2017). To support environmental management decisions and avert serious accidents caused by air pollution, air quality forecasting is becoming more and more essential not only to better govern the trend of air pollution variation but to provide timely and comprehensive environmental quality information (Pournazeri et al., 2014; Yang and Christakos, 2015; Corani and Scanagatta, 2016; Lauret et al., 2016; Wakeel et al., 2017; Van et al., 2018; Yanget al., 2018). A wide variety of methods have been used to forecast or predict regional air quality. These studies primarily branched out into two major classes: physical-based and data-driven methods. Physical- based models like dispersion and chemical transport models have still been under development as a result of uncertainties in relation to source inventories and the chemical and dynamical mechanisms of aerosols in atmosphere (Afzali et al., 2017; Vijayaraghavan et al., 2016; Jianget al., 2018; Karambelas et al., 2018; Pisoni et al., 2018). Data-driven models have leant upon the empirical or statistical relationship between air quality observations and other affecting * Corresponding author. E-mail address: changfj@ntu.edu.tw (F.-J. Chang). Contents lists available at ScienceDirect Journal of Cleaner Production journal homepage: www.elsevier.com/locate/jclepro https://doi.org/10.1016/j.jclepro.2018.10.243 0959-6526/© 2018 Elsevier Ltd. All rights reserved. Journal of Cleaner Production 209 (2019) 134e145