Traffic flow forecasting is a key problem of intelligent transport systems and represents a challenging task due to the spatial-temporal correlation features and long temporal interdependence of the considered data. Conventional methods deal with this either by spatial forecasting given observed counts at previous times or by temporal forecasting given observed traffic counts in neighbouring locations. In order to fully exploit the spatio-temporal properties observed in the data, a hybrid multimodal deep learning method for short-term traffic flow forecasting called HaMDeepT is proposed. Specifically, the HaMDeepT method can jointly and adaptively learn the spatial-temporal correlation features and long temporal interdependence of multi-modality traffic data through an attention-based auxiliary multimodal deep learning architecture. The base module of this method consists of a 1D CNN and GRU with the attention mechanism. The forecasted spatio-temporal traffic demand (counts of traffic passing through different locations at regular time intervals) is dependent on far more critical spatial factors than other point sensors such as weather stations. This HaMDeepT method, in terms of 3D CNN-GRU, which uses a stack of 3D Convolutional Neural Networks (3D CNN) and Gated Recurrent Units (GRUs) combined with the Correlation and Relative Operation layers to model both the spatial context features and the temporal dependencies of traffic count data at all locations, has a better performance compared to other network architectures. It overcomes the drawback of a fixed and handcrafted graph Laplacian matrix representation of the spatial relationships of the locations used by the ST-Graph. It uses the Correlation layer to estimate the spatial correlation features for each traffic count data point with others, focusing on the stations with major impacts on the target location, and the Relative Operation layer to model the relative distances thereafter. Using these novel methods, the traffic flow forecasting results for the miniNYC dataset are more accurate and more intuitive visualisation of the spatial structure that affects the performance of the predictions.