Day-ahead traffic flow forecasting based on a deep belief network optimized by the multi-objective particle swarm algorithm

Abstract

Traffic flow forecasting is a necessary part in the intelligent transportation systems in supporting dynamic and proactive traffic control and making traffic management plan. However, most of the previous studies attempting to build traffic flow forecasting models focus on short-term forecasting as the next step. In this paper, a deep feature leaning approach is proposed to predict short-term traffic flow in the following multiple steps using supervised learning techniques. To achieve traffic flow forecasting for the next day, an advanced multi-objective particle swarm optimization algorithm is applied to optimize some parameters in deep belief networks. The modified model can boost the accuracy of the forecasting results and enhance its multiple step prediction ability. Using real-time and historical temporal–spatial traffic data, day-ahead prediction experiment is implemented. The results of the hybrid model are compared with several commonly used benchmark models and some improved deep neural network based on evaluation criteria. Also, the proposed optimization algorithm is compared with the traditional particle swarm optimization algorithm. Furthermore, the significance in the number of hidden layers is analyzed. When the layers are increasing more than 4, the performance of the proposed model stops improving significantly. The results indicate the proposed model can extract complex features of traffic flow and therefore the forecasting accuracy and stability can be effectively improved.