Short-Term Traffic Speed Prediction Using Hybrid LSTM-SVR Model.

Traffic speed prediction uses historical data to model traffic patterns and generates forecasts for future steps. As the number of vehicles surges significantly, traffic congestion could negatively affect the quality of life, human health, and the environment. Thus, finding a method providing accurate and robust forecasts for road users and traffic management is a crucial task. In traffic prediction, capturing both long and short-term patterns is necessary for forecasting precisely. However, traditional models only perform well in short-term modeling and vice versa. This paper aims to create a hybrid model, namely LSTM-SVR, to overcome the mentioned difficulty. The LSTM-SVR combines long short-term memory (LSTM) for modeling long-term dependencies with support vector regression (SVR) for capturing short-term features. The experiments corroborate that the model outperforms popular selected baselines. The results also show the ability of the proposed model to capture peak hours and short-term patterns. This research provides a reliable forecasting tool for traffic engineers and new insights into hybrid model design in traffic speed prediction. The source code for the model from this paper is publicly available and can be found at https://github.com/adasken/lstm-svr .