Deep Learning Based Imputation and Prediction Model for Vehicle Traffic Flow

Efficient traffic management system is important for traffic congestion mitigation and urban planning in all countries. Traffic control systems works on the idea of eliminating uncertainties and avoid uncertainties to reduce the traffic flow and increase the vehicle flow. Because of issues in vehicle traffic data collection and instability in short term periods, it is hard to predict the traffic congestion accurately. To manage the congestion, it is essential to predict the forthcoming traffic flow and it will be beneficial for Advanced Traffic Management Systems (ATMS), Advanced Traffic Information Systems (ATIS), and traffic analytics. Traffic data perform a major role in all transport-oriented applications. Missing values have greatly affected the process of Intelligent Transportation Systems (ITS). Hence imputation is needed to find the missing values. Spatio-temporal factors are important to estimate the missing data and also predict the vehicle traffic congestion effectively. Deep learning algorithms play a major role in both prediction and imputation. Non-linear data and uncertain entities impact the vehicle congestion at congestion hours which are not considered in traditional algorithms. Exponential Linear Unit (ELU) activation function handles nonlinearity well in deep networks. Stacked Denoise Autoencoder (SDAE) and Long Short Term Memory Network (LSTM) are deep learning algorithms, both consider the spatio-temporal factors well. This study proposes SDAE with ELU activation to impute the missing values and LSTM is proposed to predict the vehicle traffic congestion with uncertain factors. Performance Measurement Systems (PEMS) and Mesowest database are used to assess this model. Implementation results show that SDAE-LSTM multivariate imputation prediction model attained high accuracy 80.95 % than multivariate LSTM prediction model.