22 June 2023 in IEEE Transactions on Intelligent Transportation Systems
Accurately exploiting the spatial correlation of passenger flow related to the metro network is important for improving metro passenger flow prediction. In this paper, a novel metro passenger flow forecasting model termed TMFO-AGGRU is proposed based on parameter estimation optimization for a graph convolutional gated recurrent neural network. First, we model a graph convolutional gated recurrent neural network, which is capable of real-time spatial and temporal information training on the passenger flow at a target station. Second, we design graph convolution operations to replace the linear operations of gated recurrent neural networks. To fully utilize the weighted distribution of output data based on different traffic characteristics, we convert the adjacent static matrices into dynamic characteristic matrices using feature learning and construct the attention mechanism module. Last, an adaptive moth-flame optimization method based on T-variance is proposed in this paper to achieve dynamic optimization of structural parameters and to avoid affecting the accuracy of prediction due to improper parameter selection. Experimental results show that the proposed TMFO-AGGRU effectively improves the prediction accuracy and convergence speed, which outperforms traditional metro passenger flow prediction methods.
3 December 2021 in IEEE Transactions on Intelligent Transportation Systems
In this paper, considering spatiotemporal correlations, we propose a novel short-term traffic flow prediction method that is based on diverse ensemble deep learning. First, a new measurement function of spatial correlation is established, quantifying the spatial correlation of traffic flow parameters. Second, the convolutional neural network model embedded long short-term memory (LSTM-CNN) is used, enhancing the learning ability of the spatiotemporal correlation of traffic flow parameters. Third, according to the network structure, ensemble rules of diverse LSTM-CNNs are constructed, combining a series of different and moderately accurate LSTM-CNN models and improving the robustness of the algorithm. Finally, a dynamic optimization method for the weight parameters of the ensemble elements is proposed to accommodate the changes in the actual road networks. The experiments show that the measuring method of spatial correlations, the ensemble combination rules of diverse LSTM-CNN and the dynamic optimization method of ensemble element weight parameters, which are proposed in this paper, have positive effects on improving the performance of the prediction algorithm. In addition, the short-term traffic flow prediction method proposed in this paper can adapt to changes in the traffic flow. It can obtain a better prediction effect even with a small sample, and the prediction performance is better than that of the other four classical short-term traffic flow prediction methods.
14 May 2020 in IEEE Access
In order to eliminate outliers in traffic flow data collection and promote the generalization performance of traffic flow prediction, this paper proposes a dynamic optimization long short-term memory (LSTM) model based on data preprocessing for short-term traffic flow prediction. A new classification algorithm named Asym-Gentle Adaboost with Cost-sensitive support vector machine (AGACS) is used for preprocessing traffic flow data. AGACS tries to employ Cost-sensitive SVM (CS-SVM) as weak component classifier in Asymmetric Gentle AdaBoost, and divide the data collection into outlier data and normal data. Only normal data is used for training LSTM to predict traffic flow and an improved chaotic Particle Swarm Optimization (CPSO) is used for dynamic optimizing hidden layer structure of LSTM to promote the generalization and robustness performance of model. The efficiency of the proposed method is proved with real traffic flow data, and the experimental results show that preprocess collecting data and dynamic optimize model structure are conducive to improve the performance of algorithm, and the proposed method achieved better performance than those of three other classical deep learning prediction methods.