关键词: 深度学习, 自编码, 人工神经网络, 伤亡预测模型, 稀疏理论

Abstract: Accurate forecast of highway accident casualties is of great significance for the assurance of China’s future traffic safety situation and the realization of transportation system optimization. This paper innovatively takes Deep Auto-Encoder（DAE） based on the theory of sparse into the prediction of casualty. Then it constructs accident forecast model using casualty historical data of 2000—2015, and achieves that the average error rates are 1.69%, 1.53% of death toll and injury separately. This paper also constructs forecast model using the same period data of car ownership, GDP, the total mileage of the highway, road area per capita as the influence factors of casualty, and average error rates are 1.76%, 2.13% separately. By comparison, high precision model can be acquired by applying DAE into casualties prediction, and the accuracy of model based on casualty time series data is higher. So this paper uses it to forecast 2016—2020 data, the death toll will remain stable, while the injury will decline significantly.

Key words: deep learning, auto-encoder, artificial neural network, casualty forecast model, sparse theory