Spatio-temporal traffic data prediction based on stacking multiple regression models

This paper uses the Stacking algorithm to solve the problems of low forecasting accuracy and long testing time of the spatial-temporal traffic data prediction model. Two kinds of Stacking multi-regression prediction models are put forward: Stacking fusion model 1 (layer 1: RandomForest and XGBoost, layer 2: LightGBM) and Stacking fusion model 2 (layer 1: RandomForest, GradientBoosting and XGBoost, layer 2: LightGBM). Firstly, the first-layer learner is trained and constructed by means of 5-fold cross-validation. Secondly, the results obtained by merging the predictions of the layer 1 model are used as the training and testing sets of the layer 2 meta-learner LightGBM, which can correct the prediction errors of the layer 1 model and improve its prediction accuracy and generalization ability through learning. The Chicago Traffic dataset for the period April through September 1991 is used as an example for comparison experiments with seven classical machine learning prediction methods. The results show that both of the two Stacking multi-regression prediction models have good performance, with high accuracy, short prediction time, and high stability. Although the accuracy of the first model is lower than that of the second, the training time is 79.83% less than that of the latter.