Traffic Flow Forecast for Traffic with Forecastable Sporadic Events

Abstract

The prosperity of the social economy, tourism, and entertainment industry are important factors to cause traffic congestion. In addition to commuting hours and holidays, if a large-scale event, such as a concert, a sporting event or an exhibition is held, it is easy to make traffic congestion even worse. If we know in advance the time and place of the large-scale event, then we can accurately forecast the future traffic flow and plan the driving route. It helps effectively relieve traffic flow, reduce travel time and carbon emissions. In this study, we used the Vehicle Detector (VD) [12] data from the Taipei City Government Open Data Platform as a source of regular traffic data as well as the data of Forecastable Sporadic Event (FSE), such as a large-scale event, to forecast traffic flow. The information of time and place of the FSE are collected from various information websites (ticketing websites, tourist websites, etc.) by web crawlers. We proposed a Long Short-Term Memory (LSTM) deep learning model for traffic flow forecast, which was trained with both VD and FSE data. We further used Adam Optimizer to adjust the weight and bias of the model to optimize the forecast accuracy. The implementation of the LSTM model was conducted in TensorFlow, a machine learning framework developed by Google. Finally, we evaluated the forecast accuracy of the model by Mean Absolute Percentage Error (MAPE) and analyzed the effectiveness of applying FSE data to traffic forecast.