Using wavelet transform and dynamic time warping to identify the limitations of the CNN model as an air quality forecasting system

Abstract

11As the deep learning algorithm has become a popular data analytic technique, atmospheric 12scientists should have a balanced perception of its strengths and limitations so that they can provide 13 a powerful analysis of complex data with well-established procedures. Despite the enormous 14 success of the algorithm in numerous applications, certain issues related to its applications in air 15 quality forecasting (AQF) require further analysis and discussion. This study addresses significant 16limitations of an advanced deep learning algorithm, the convolutional neural network (CNN), in 17 two common applications: (i) a real-time AQF model, and (ii) a post-processing tool in a dynamical 18 AQF model, the Community Multi-scale Air Quality Model (CMAQ). In both cases, the CNN 19 model shows promising accuracy for ozone prediction 24 hours in advance in both the United 20States and South Korea (with an overall index of agreement exceeding 0.8). For the first case, we 21 use the wavelet transform to determine the reasons behind the poor performance of CNN during 22 the nighttime, cold months, and high ozone episodes. We find that when fine wavelet modes 23(hourly and daily) are relatively weak or when coarse wavelet modes (weekly) are strong, the CNN 24 model produces less accurate forecasts. For the second case, we use the dynamic time warping 25(DTW) distance analysis to compare post-processed results with their CMAQ counterparts (as a 26 base model). For CMAQ results that show a consistent DTW distance from the observation, the 27 post-processing approach properly addresses the modeling bias with predicted IOAs exceeding 28 0.85. When the DTW distance of CMAQ-vs-observation is irregular, the post-processing approach 29is unlikely to perform satisfactorily. Awareness of the limitations in CNN models will enable 30 scientists to develop more accurate regional or local air quality forecasting systems by identifying 31 the affecting factors in high concentration episodes. 32 33