A Multitask Learning Neural Network for Short-Term Traffic Speed Prediction and Confidence Estimation

To improve predictive accuracy on short-term traffic speed, we proposed a multitask learning neural network (MLNN). MLNN carries out the speed prediction task for three short-terms by the combination of convolution neural network (CNN) and gated recurrent units’ network (GRU), and accomplishes the confidence estimation task on predicted speed with the confidence network. A multitask loss function with weighted sub loss terms for multitask learning is employed. In the experiment, our method was tested on the data set of Shanghai Expressway at 2014. Conventional methods such as auto-regressive integrated moving average (ARIMA) and Gaussian maximum likelihood (GML), and time series models, recurrent neural network (RNN), GRU and long short-term memory (LSTM), were also used to compare. The results show that MLNN with square loss obtained the smallest mean squared error (MSE) on most cases. For four road types, MLNN obtained the overall smallest mean absolute percentage error (MAPE) on these cases. We also proved that as compared to single-term prediction, multitask learning outperformed 12.4% in MSE and 9.91% in MAPE for 10-min and 15-min prediction. To improve the forecast on low speed, MAP-loss term is additionally used in multitask loss function. It efficiently improved the predictive accuracy on low speed. The confidence estimation network gave a 89.93% estimation accuracy on the predicted speed, efficiently avoiding the inaccurate speed prediction.