Time-Series Predictions for People-Flow with Simulation Data

Abstract

In order to prevent accidents, it is important that the administrators of large-scale facilities or event organizers be able to analyze and predict human flow. Time series prediction is generally used for such situations. However, some cases have no historical data available such as the construction of new stadium. In such cases, the multi-agent simulator (MAS) is useful for generating sufficient simulation data to support the assessment of navigation plans, and predictions can be made more accurate by comparing simulation results to monitored data. In this paper, to predict the number of passengers at the multiple observation points, we use simulation data (generated by MAS) as a learning dataset for long short-term memory (LSTM). To compare the prediction accuracy of the proposed approach, we use the real world data collected at the music live events. In addition, for the comparison, we use the nearest neighbor approach that searches the most similar result from the pre-simulated results and predicts the human flow.