Real-Time LNG Buses Emissions Prediction Based on a Temporal Fusion Trans-Formers Model

Abstract

Emissions from transportation are one of the key factors preventing the achievement of carbon peaking and carbon neutrality by 2050, with particular attention to emissions from buses. Specifically, few research has been conducted on the exhaust emissions characteristics of liquified natural gas (LNG) buses under different driving scenarios. This study proposed a framework for predicting exhaust emissions of LNG buses based on the portable emission measurement system and GPS collaborative perception data. Firstly, the emission distribution characteristics of CO₂, CO, HC, and NO_x from LNG buses in real-world driving were analyzed by visualization methods. Then, the real-time exhaust emissions of LNG buses were predicted based on the temporal fusion transformers model for both urban and suburban sections of Zhenjiang City, and the model validity was verified. The current and past 10 s driving states were used for predicting the emission rate of LNG buses. The results showed that the proposed model outperforms other advanced algorithms in real-time exhaust emissions prediction of LNG buses, with an average R² value higher than 0.94 and an average MAPE reduction of 14.19%. The error assessment revealed that the emission values and average emission rates are higher when driving in the urban section compared to the suburban section. Among the influencing factors, traffic conditions have the most significant impacts on the exhaust emissions of LNG buses, followed by road conditions and driving states, with relative feature importance of 48.9, 34.8, and 16.3%, respectively. Additionally, the current and past 10 s driving states also significantly influenced real-time predictions. This study provides an essential theoretical reference for reducing exhaust emissions for city buses.

Keywords: decarbonization, emission prediction, PEMS, GPS, collaborative perception, traffic condition