One of the central issues in transportation science involves accurately predicting traffic flow, drawing attention from both academics and practitioners. It serves an essential function in enhancing traffic management efficiency, alleviating congestion and reduce accidents. However, current research still faces several limitations: (i) The constraint on input sequence length hinders the ability to capture long-range and long-distance dependencies; (ii) Complex spatial-temporal dependencies are difficult to model, especially in a changing traffic environment, and existing methods have limited effect on capturing spatial-temporal interaction patterns; (iii) Most models rely only on historical data, and the periodic characteristics of traffic flow are not adequately modeled, making it difficult to fully utilize periodic patterns for prediction. To this end, this study proposes a traffic flow forecasting model (STMFNet) built upon spatial-temporal masking and feature fusion. Specifically, the model first uses spatial-temporal masking mechanisms (T-Mamba and S-Mamba) to randomly mask the data and reconstruct the missing information; then it combines the graph convolutional network (GCN) to model short-term traffic; finally, an adaptive periodic modeling mechanism is designed to combine the periodic characteristics of holiday and weekday data, and generate spatial-temporal feature representations through the hidden layer fusion masking mechanism of the model, thereby achieving accurate prediction of traffic flow. Experimental results based on four real-world datasets (PEMS03, PEMS04, PEMS07, PEMS08) show that STMFNet has significant advantages in prediction performance compared with existing baseline models.