基於深度學習之城市全區群體流量長期多步預測

A Deep Learning Approach for Long-term Multi-Step Citywide Crowd Flow Prediction

黃千芸(Chian-Yun Huang)

指導教授 ： 曾新穆

國立交通大學/資訊學院/資訊科學與工程研究所/碩士(2019年)

摘要

近年來，隨著科技的發展，許多城市中的應用項目為人們帶來更便捷的生活，其中，群體流量預測因有助於城市之運輸管理及廣告投遞等應用，受到越來越多的關注。然而，過往的研究往往只考慮短期的預測，例如只預測輸入資料的一小時後的流量，而缺少針對長期預測的研究，但以交通資源的提前分配等應用而言，長期預測的框架會具有更高的應用價值。因此本研究的目標是利用行動裝置的GPS資訊統計之流量及天氣等相關資訊，研究群體流量長期預測問題，希望透過長期預測提供更進階應用之所需。群體流量雖然是時間序列資料的一環，卻因其不但是含有時間關係的時序性資料，也是具有空間關係的地域性資料，使得分析更為困難。因此，本研究採用深度學習之方法分別學習短週期、中週期、及長期之時空特徵，同時加入天氣及節日資料修正特殊情況，透過準確的預測模型幫助估算人流量或車流量以提前做有效的資源分配。在考慮時間及空間關係的同時，也需分析時間序列資料的週期性對長期預測之影響。此外，為增加長期預測準確度並節省計算資源，採用多步預測之模型架構，使其能透過重複利用萃取的特徵，使用單一模型同時預測多個時間點的流量。據我們所知，本研究是第一個針對群體流量長期多步預測的研究，且經實驗證明，本研究所提出之方法能有效的預測長期的群體流量，並且透過單一模型達成多步預測，在不損失長期預測表現的情況下有效節省計算資源。

關鍵字

群體流量預測 ； 時間序列 ； 時空資料 ； ⻑期預測 ； 多步預測 ； 深度學習

並列摘要

Recently, predicting the crowd flows in a city has attracted a lot of attention because it is valuable to city transportation, management, and business. However, most of the previous works mainly focused on short-term prediction while long-term prediction is even more valuable for some important real-world applications, like traffic resources pre-allocations. As a result, this work aims to predict long-term crowd flows in a city by trajectory data or trip data obtained from mobile devices or GPS devices in vehicles. As a kind of time series data, crowd flows, including the inflow and the outflow, are influenced by both temporal dependencies and spatial dependencies. Therefore, we apply deep learning methods, such as convolution and residual structure, to extract short term, long short-term, and long term patterns. Besides, we also integrate the model with context data, including weather data and calendar information, to improve the performance. In addition, to enhance the robustness for long-term prediction and save computational resources, we adopt a multi-step architecture that can output predictions on multiple time steps simultaneously. To the best of our knowledge, this is the first work on long-term multi-step crowd flow prediction. The experiment results on two real-world datasets demonstrate that our framework outperforms the state-of-the-art methods on long-term crowd flow prediction. We also show that the multi-step architecture can enhance the execution efficiency without losing the accuracy performance.

並列關鍵字

Crowd Flow Prediction ； Spatial and Temporal Data ； Long-term Prediction ； Multi-step Architecture ； Deep Learning

參考文獻

[1] Hannah Bast, Daniel Delling, Andrew Goldberg, Matthias Müller-Hannemann, Thomas Pajor, Peter Sanders, Dorothea Wagner, and Renato F. Werneck. “Route Planning in Transportation Networks”. In: Algorithm Engineering: Selected Results and Sur-

veys. Springer International Publishing, 2016, pp. 19–80.

[2] Cen Chen, Kenli Li, Sin G Teo, Guizi Chen, Xiaofeng Zou, Xulei Yang, Ramaseshan

C Vijay, Jiashi Feng, and Zeng Zeng. “Exploiting Spatio-Temporal Correlations with Multiple 3D Convolutional Neural Networks for Citywide Vehicle Flow Prediction”. In: 2018 IEEE International Conference on Data Mining (ICDM). IEEE. 2018,

pp. 893–898.