<p>Accurate traffic flow prediction is the key to constructing an intelligent transportation system. Long-term traffic prediction is extremely challenging due to the complex spatial-temporal dependencies of traffic systems and the ever-changing nature of many influencing factors such as weather. Existing methods still have several pitfalls: (i)While traffic flow has long term dependence (e.g., the morning peak flow may affect the evening peak), many existing methods are insufficient in modeling long-term dependencies, and usually only capture short-term temporal dependencies. (ii)Emerging methods such as Transformer-based models do not make effective use of critical external factors such as forecasted weather conditions and Point-Of-Interests. In order to better model complex spatial-temporal dependencies of traffic flow and make full use of external factors on the Transformer-based model, we propose a linear network and attention-based model to solve the long-term traffic flow forecasting problem. More specifically, we first use a simple linear network to learn the historical time series correlation of each node (bus station, road intersection, etc.). Then we design a method to represent and fuse other influencing factors as covariates (additional variables or context features associated with traffic flow). Moreover, the entire input series and covariates are used as tokens for attention computation to capture the spatial-temporal dependence between nodes. Experiments on five real-world traffic datasets demonstrate the superiority of our model over several state-of-the-art methods.</p>