Traffic flow prediction is very important in traffic-related fields, and
increasing prediction accuracy is the primary goal of traffic prediction
research. This study proposes a new traffic flow prediction method, which uses
the CNN–BiLSTM model to extract features from traffic data, further models these
features through GBRT, and uses Optuna to tune important hyperparameters of the
overall model. The main contribution of this study is to propose a new
combination model with better performance. The model integrates two deep
learning models that are widely used in this field and creatively uses GBRT to
process the output features of the front-end model. On this basis, the optimal
hyperparameters and the robustness of the model are deeply explored, providing
an effective and feasible solution to the difficult problems in traffic flow
prediction. This model is experimentally studied using three different data
transformation methods (original data, wavelet transform, Fourier transform).
Compared with other models using similar data, the evaluation index performance
of this model under wavelet transform is better, RMSE, MAE, MAPE, and
R
2 values are 53.879, 40.641, 11.13%, and 0.97, respectively, which
are better than other comparison models. The results show that the proposed
Optuna–CNN–BiLSTM–GBRT model can significantly improve the accuracy of traffic
flow prediction and provide an effective means to solve problems in the field of
traffic prediction.
