Ensemble Deep Learning Framework for the Traffic Accidents Duration Prediction

Traffic management and disaster response rely heavily on accident duration anticipation on expressways. There is no way around the fact that data pertaining to traffic accidents is multi-mode data. Nevertheless, the vast majority of research has been on single-mode data, and there has been surprisingly little quantitative examination of how multi-mode data affects model prediction abilities. In response to these shortcomings, the current study suggests a heterogeneous deep learning architecture that makes use of multi-modal characteristics to enhance the precision of motorway traffic accident length prediction models. To begin, by combining structured and textual data, six distinct data modes are produced. Second, to construct classification models with less prediction error, a hybrid deep learning method is used. In the end, we use a number of deep learning models to thoroughly examine how multi-mode data affects the accuracy of accident length predictions. The survey data obtained from an expressway observing system of China, is utilized to gauge the proposed method. The experimental outcomes demonstrate that Word2Vec-BiGRU-CNN—with an F1-score of 0.3648—is the most appropriate and superior for anticipating the duration of traffic accidents using text features. The results of this study show that deep learning algorithms’ prediction abilities are much improved by using newly-established structured features retrieved from text data.