LSTnet-GRU-A: Traffic Flow Forecasting Based on Attention Mechanism

With the rapid advancement of big data and multi-sensor fusion technology, it becomes easier to obtain traffic flow information. According to these related data, some methods are designed to forecast the traffic flow, which can provide decision support for scheduling and optimizing the intelligent transportation systems (ITS). Thus, the forecasting methodology of traffic flow becomes one of the most popular fields of ITS. To predict traffic flow more accurately, we propose a novel LSTnet-GRU-A model, which adds an attention module to the progressive LSTnet model, then, by regulating the weight of the hiding state, the LSTnet-GRU-A model is developed to model multi-dimensional traffic flow data. The dataset MITV is used as the sample library and the performance of different models is given to show the superiority of LSTnet-GRU-A. To optimize the model, the strategies of learning rate automatic decline and early stopping are adopted. Furthermore, the five checkpoints of our model are saved for snapshot ensembles to generate the optimal performance. Multiple experimental results prove that the LSTnet-GRU-A proposed in this paper achieves SOTA performance consistently.