In this study, the hybrid of combination-mixed data sampling regression model and back propagation neural network (combination-MIDAS-BP) is proposed to perform real-time forecasting of weekly carbon prices in China's Shenzhen carbon market. In addition to daily energy, economy and weather conditions, environmental factor is introduced into predictive indicators. The empirical results show that the carbon price is more sensitive to coal, temperature and AQI (air quality index) than to other factors. It is also shown that the forecast accuracy of the proposed model is approximately 30% and 40% higher than that of combination-MIDAS models and benchmark models, respectively. Given these forecast results, China's government and enterprises can effectively manage nonlinear, nonstationary, and irregular carbon prices, providing a better investing and managing tool from behavioural economics. (C) 2019 Elsevier Ltd. All rights reserved.