Abstract: Considering the random and nonlinear characteristics of traffic flow that result in low prediction accuracy, a combined prediction model based on quadratic decomposition and fusion of multiple features is proposed, where the trend and periodic features in traffic flow are extracted by using the time series decomposition method, the residual components are quadratically decomposed by the optimized variational mode decomposition, the scored amount is reconstructed, the external features of the traffic flow are selected by the correlation coefficient method, and three different models are established to predict the components after fusing the external features. Reinforcement learning is used to optimize the weights of each model, and the final prediction result is obtained by weighted sum. Using the simulation analysis of traffic flow in Changsha urban area, experimental results show that compared with the long short-term memory neural network model. Combined model of convolutional neural networks and gated cyclic units, BP after quadratic decomposition and lightweight gradient lifter after quadratic decomposition, the model established in this paper has a better prediction effect on urban road traffic flow, with an average absolute error of 2.622 and a root mean square error of 3.479. The prediction errors are better than the existing models, which verifies the effectiveness of the proposed model.

Key words: short-time traffic forecasting, time series decomposition, feature selection, Q-Learning, combined model