Article Preview
TopIntroduction
With the development of social economy, the number of vehicles is increasing rapidly and the traffic congestion has seriously affected road safety and environmental pollution. The transportation departments formulate management strategy and improve service level through utilizing the existing highway facilities and transportation network resources (Chou et al., 2018). In recent years, intelligent transportation system (ITS) has become a hot research area (Wang, Chen, Cheng, Lin, and Lo, 2015; Zhuang, Luo, Pan, and Pan, 2020; Song, Pan, and Chu, 2020). In this paper, long short-term memory (LSTM) and meta-heuristics are studied here to predict the short-term traffic flow.
The prediction of traffic flow is an important part of ITS. It is a major reference in the vehicle path planning and the resources are allocated to the roads where congestion risks are most likely to occur (Chen, Lin, Chang, and Lo, 2012). The prediction is to estimate the traffic flow in the future period by the known historical data and the current traffic flow, which includes long-term and short-term predictions. The former is usually based on hours, days, months and even years to forecast the traffic flow, and it is helpful in planning the reasonable distribution of the road network. But it merely grasps the status of traffic flow from a macro perspective, it is difficult to meet people's requirement on traffic information. The short-term prediction effectively uses real-time data to forecast traffic conditions within 30 minutes in the future.
The neural network provides a simplified model that simulates the ability of biological neural assembly to solve multi-layer and nonlinear problems. Recently, it has received extensive research and attention from scholars. Lyu et al. used recurrent neural network (RNN) to obtain word representation, and then learned the importance of each word in text classification through convolutional neural network (CNN) (Lyu et al., 2021). The algorithm presented excellent performance on multiple text classification data sets. Due to the rich semantics and flexible representation of biomedical entities, the biomedical ontology matching problem remains an open challenge in the alignment quality. Xue et al. proposed an attention-based bidirectional LSTM network ontology matching technique to address this problem (Xue et al., 2021). Guo et al. suggested a method based on SIGNAN, which generated various high-fidelity images with only one input image (Guo et al., 2021). The method achieved the operations of shape change, illumination direction changes and super-resolution generation, and improved the sampling efficiency. Fantinato et al. investigated the integration of deep learning and service-oriented architecture (SOA) and discussed how SOA can be implemented by deep learning methods in different types of environments for various users (Fantinato et al., 2021).
Short-term traffic flow is a typical time series data. At present, the common prediction models of traffic flow contain autoregressive integrated moving average (ARIMA), Kalman filtering model and artificial neural network (ANN). Traditional prediction models are limited in finding the relationship between historical data and the future traffic flow, consequently, they don’t understand deep correlation and implicit information. Deep neural network (DNN) implements complex non-linear relationship by distributed hierarchical feature representation (Jiao, Wu, Bie, Umek, and Kos, 2018; Shi, Guo, Niu, and Zhan, 2020; Xia, Wang, and Guo, 2020; Chen, Song, Hwang, and Wu, 2020). Many neural networks are proposed to assist traffic prediction, such as artificial neural network, RBF neural network, RNN and long short-term memory neural network (Jiang et al., 2018; Liu, Zhang, and Chen, 2019; Ke, Shi, Guo, and Chen, 2018).