基於大廳人流時間序列數據的混合預測模型

Abstract

目前有關時間序列的預測模型在各種的時間序列的研究中提供了不同的預測準確度和分析。大多數模型在一般的時間序列在不同頻率預測 : 例如月度，季度或年度的頻率已經非常成熟。本文介紹了一種基於大廳的時間序列數據。基於大廳時間序列數據是大廳行人/訪問者的數量，本文的實驗數據是基於一個校園建築的大廳影像記錄並轉換為每分鐘經過的行人數目。這個基於大廳數據集的預測的挑戰和特性是人流數據稀疏，高方差的變化和高頻率，在不同的時間段中會有各種不同的變化差異的特點。本文在此基礎大堂時間序列數據提出了一種混合模型，並主要目的是提高對從高頻率的歷史數據預測未來一日人流數量的準確性。本文注重研究和修改三種基本時間序列預測模型，我們著重強調於數據的特點，我們提出了一個時間區段的學習和預測策略，選擇在不同時間區段中使用合適的預測模型。另外，採用了不同級別的時間序列的聚集去改進預測準確性。我們的實驗，建立了在真實世界中的數據，實際的數據集被用於評估和預測精度進行比較的三個基本模式。結果表明，我們的混合模式可以在真實世界的數據很好地應用。

The prior research in forecasting time series using different models has provided on different predictive accuracy. Most of model are well established on general time series of monthly, quarterly or annual frequency. This paper introduces a lobby-based time series data. The lobby-based time series data are the number of pedestrian/visitor, were collected by a camera at a campus lobby and converted to number of pedestrian minutely. The characteristics and the most forecasting challenges of this lobby-based dataset are sparse, high variance and high frequency. This paper proposed a hybrid model on this lobby-based time series data and the main purpose is to improve the accuracy on forecasting the future day from the high-frequency historical data. This paper investigates and modifies three basic forecasting models which are popular in time series forecasting area. We strongly emphasize the characteristics of lobby-based data, we propose a segmentation learning and forecasting strategies to select the model in different period as a hybrid model. The temporal aggregation of different levels leads to improvements in predictive accuracy. Our experiment was established in real-world data, the real data set is used to evaluate and compare the forecasting accuracy to the three basic models. The result is shown that our hybrid model can be well applied in the real-world dataset.