Parking demand forecasting based on improved complete ensemble empirical mode decomposition and GRU model

Abstract

Parking demand forecasting plays an important role in relieving traffic congestion and reducing greenhouse gas emissions. However, most previous studies model based on historical data on parking itself or numerous factors that influences parking demand, which increase the complexity of the data and the time taken to run the model, resulting in a poor fit of the model to the extreme value points and not meeting the needs of practical applications. To address this issue, a hybrid prediction model based on improved complete ensemble empirical mode decomposition (ICEEMDAN) and gate recurrent unit (GRU) model for predicting parking demand, as well as a method called homogeneous linear mean interpolation to fill in the missing data were proposed in this paper. The ICEEMDAN algorithm was used to decompose the parking time series and reduce its complexity and nonlinearity. Based on the decomposed sequences, GRU neural networks were constructed for simultaneous training and prediction. Finally, the subsequences of the predicted output were aggregated. The effectiveness of the proposed model was validated using parking data collected from a large transportation hub parking lot. The experimental results show that compared with the single GRU model, the RMSE of the ICEEMDAN-GRU model is reduced by 59.29%, the MAE is reduced by 64.08%, and the R-square is improved by 4.93%. The ICEEMDAN-GRU model is the closest to the real parking time series. Therefore, this method is more effective than other models in parking demand forecasting.