Multi-Component Fusion Temporal Networks to Predict Vehicle Exhaust Based on Remote Monitoring Data

Vehicle exhaust prediction is of great importance for exhaust emissions control and public environment protection, and very challenging because it is influenced by various complex factors, including multi-component temporal dependencies (closeness, period, trend) and external factors, e.g., meteorological data and workday information. To address these issues, we propose a novel deep-learning-based framework, called multi-component fusion temporal networks (MCFT-Net), to collectively forecast emission concentration of carbon monoxide (CO). The proposed MCFT-Net consists of three parts, 1) a one-dimensional Convolution Neural Networks (1D-CNNs) module to capture temporal dependencies of vehicle emissions by merging closeness, period, and trend time sequences; 2) an artificial neural network (ANN) module to capture features of external factors, such as wind direction, weekday and weekend; and 3) a cascade fusion component to merge the temporal patterns, which are derived from temporal dependencies processed by multiple 1D-CNNs, and then assigned different weights according to an attention mechanism after fusion. The output of the attention operation is further combined with external features through sum and average operations to predict the final emission concentration. We evaluate the proposed MCFT-Net using two real-world vehicle exhaust datasets, ExhaustBMS and ExhaustHMS, which were collected from Beijing monitoring station (BMS) and Hefei monitoring station (HMS). The results demonstrate that our MCFT-Net outperforms 8 baseline methods, such as ARIMA, RNN, GRU, BiLSTM.