Traffic Flow Prediction Based on Optimized LSTM Model

Abstract

Intelligent transportation system (ITS) refers to comprehensive transport systems that uses advanced science and technology to ensure safety, enhance efficiency, improve environment, and save energy. Traffic flow prediction is a research focus of Intelligent Transportation Systems (ITS), which provides important basis for ITS planning and scheduling, and is crucial for ITS. As traffic flow changes are the result of multiple factors such as weather, accidents, road control, etc., fluctuations are large and irregular. Especially in the minute-level and hour-level traffic flow prediction, it is difficult to predict the trend of traffic flow accurately. To address this issue, this paper fully considers the temporal, periodic, and spatial features of traffic flow, and proposes a serial model of LSTM-LightGBM for hour-level traffic flow prediction, which uses the output of LSTM as the input of LightGBM. LightGBM can capture the spatial and periodic features of traffic flow, while LSTM is able to capture the temporal features of traffic flow. The LSTM-LightGBM model is tested on the Chicago traffic dataset. Results show that LSTM-LightGBM has the least RMSE and MAE among all tested models, and the RMSE of LSTM-LightGBM can be reduced by up to 50% compared to other base models, indicating that this model has great potential.