In response to the limited accuracy and high training costs of existing traffic flow forecasting models in long-term forecasting, a long-term traffic flow forecasting model based on cross period sparse technique was constructed. Considering the complex periodicity of traffic flow data and the increased error accumulation and model complexity brought by the extended time span in long-term forecasting, cross period sparse technique was used for downsampling processing of the time series. The long-term sequence was decomposed into periods and then spliced and transposed to form a time series matrix. A mixed graph convolutional network was embedded into gated recurrent units to process the downsampled time series matrix, and finally, a time multi-head attention module with a time-dependent matrix was used to fuse relevant spatiotemporal features. After processing three public traffic flow datasets, PeMS-bay, PeMS04, and PeMS08, for effectiveness validation, the results show that the model's prediction accuracy indicators on all three datasets are superior to the baseline model, with average prediction accuracy improved by 5.59%, 8.56%, and 7.11% respectively on the three datasets, and the model parameters is between 14.81% and 27.17% of the baseline model.