LSTM training set analysis and clustering model development for short-term traffic flow prediction

Abstract

Long short-term memory (LSTM) is becoming increasingly popular in the short-term flow. In order to develop high-quality prediction models, it is worth investigating the LSTM potential deeply for traffic flow prediction. This study has two objectives: first, to observe the effect of using different sized training sets in LSTM training for various and numerous databases; second, to develop a clustering model that contributes to adjusting the training set size. For this purpose, 83 datasets were divided into certain sizes and LSTM model performances were examined depending on these training set sizes. As a result, enlargement of the training set size reduced LSTM errors monotonic for certain datasets. This phenomenon was modeled with the state-of-the-art clustering algorithms, such as K-nearest neighbor, support vector machine (SVM), logistic regression and pattern recognition networks (PRNet). In these models, statistical properties of datasets were utilized as input. The best results were obtained by PRNet, and SVM model performance was closest to PRNet. This study indicates that enlarging the training set size in traffic flow prediction increases the LSTM performance monotonically for specific datasets. In addition, a high-precision clustering model is presented to assist researchers in short-term traffic forecasting to adjust the size of the training set.