Abstract: As a key part of intelligent transportation systems,traffic flow forecasting faces the challenge of long-term prediction inaccuracy.The key factor is that the traffic flow has complicated spatial and temporal correlations.Recently,the emerging success of Transformer has shown promising results in time series analysis.However,there are two obstacles when applying Transformer to traffic flow forecasting:1)it's difficult for the static attention mechanisms to capture the dynamic changes of traffic flow along the space and time dimensions;2)the autoregressive decoder in transformer could cause error accumulation problem.To address the above problems,this paper proposes a time-aware Transformer(TAformer) for traffic flow forecasting.Firstly,it proposes a time-aware attention mechanism that can customize attention calculation solution according to the time features,so as to estimate the spatial and temporal dependencies more accurately.Secondly,it discards the teacher forcing mechanism during the training phase and proposes a non-autoregressive inference method to avoid the problem of error accumulation.Finally,extensive experiments on two real traffic datasets show that the proposed method can effectively capture the spatial-temporal dependence of traffic flow.Compared with the state-of-the-art baseline method,the proposed method improves the performance of long-term prediction by 2.09%～4.01%.

Key words: Traffic flow Forecasting, Spatial-Temporal modeling, Time-aware attention, Non-autoregressive, Transformer