关键词: 交通运输工程;交通大数据；交通预测；深度学习；交通流；深度信任网络；花朵授粉算法

Abstract: To address the issues of excessive reliance on time series training data, insufficient consideration of spatial correlation, and overly fixed selection forms of model structural parameters in short-term traffic flow prediction, a structure-optimized short-term traffic flow prediction method for deep belief networks was proposed. The proposed model could simultaneously train three kinds of traffic data related to the traffic volume of the predicted nodes, enhance the spatial and temporal correlation of the prediction, and overcome the defect that the training data relied too much on time series. Meanwhile, the proposed model optimized the structure of the deep belief network short-term traffic flow prediction model and proposed an improved flower pollination algorithm to optimize the hidden layer structure parameters of the prediction model, avoiding the problem that the prediction results of the model were trapped in local optimal solutions and the practicality was reduced due to the overly fixed selection forms of model structural parameters. The feasibility of the proposed prediction model was evaluated by collecting the relevant traffic volume data of two intersections in Fuzhou. At the same time, the MFPA-DBN model was compared with the deep belief network model with different hidden layer structures, GA-LSTM, CNN-SVR and TGWO-BP models. The experimental results show that under the same training data conditions, the short-term traffic flow prediction method based on MFPA-DBN is feasible and effective, its prediction accuracy is better than that of other depth learning prediction models, and the real-time performance can also meet the actual requirements.

Key words: traffic and transportation engineering; traffic big data; traffic forecast; deep learning; traffic flow; deep belief network; flower pollination algorithm