混沌車流短期交通量變化之預測－相空間局部近似法（PSLA）之應用

Abstract

智慧型運輸系統（intelligent transportation system, ITS）應能自動蒐集即時路況資料，並對數分鐘後的交通量做出預測，此為動態交通控制的必備條件。以先進交通管理及資訊系統（Advanced Traffic Management and Information System, ATMIS）為例，唯有依賴能自動蒐集交通狀況資料，據以描述系統狀態的監視技術，並具良好的預測能力，才能預先處理系統的改變，進一步修正控制策略，以達到運輸系統能更有效率的運作。可見交通預測能力的良窳是ITS成功與否之重要關鍵，而短期交通量預測則是ATMIS一個重要的課題，但對此的研究仍相當缺乏。由於交通量變化型態屬於不規則非線性系統，具混沌特性，因此可以混沌理論加以探討及預測。 混沌理論近年來已被廣泛的運用於各領域，用來描述及預測複雜的非線性系統，頗具發展潛力。本研究乃嘗試採用混沌理論為基礎來預測短期交通量之非線性變化，參採Farmer and Sidorowich之相空間局部近似法（phase space local approximation method）構建交通量預測模式，依不同時間間隔（time interval）如五、十及十五分鐘，進行短期交通量預測。 預測結果顯示，於各不同條件下，觀察值與模式值間之相關係數皆大於0.6，預測之相對誤差等指標也在可接受之範圍內，表示短期交通量的變化型態可由確定性混沌模式描繪之。對於預測結果之比較，若區分並分別針對相同時期之交通量變化進行預測，更能得到較佳之預測績效，平均可改善的程度約20%-30%。而與ARIMA模式相較，在預測能力最佳的情況下，相空間法預測之相對誤差約在0.3-0.7間，而ARIMA約在0.9左右，顯示本法之預測能力遠較ARIMA模式佳，特別是可適用於短期交通量變化之預測課題上。

Traffic prediction is a prerequisite in advanced traffic control. If real-time traffic dynamics can be automatically collected and future traffic patterns can be accurately predicted, then appropriate traffic control strategies can be adopted accordingly. However, little literature has been devoted to traffic prediction, particularly in the short-term traffic dynamics. Short-term traffic flow characteristics can be viewed as chaos because it varies over time and over space with quasi-periodic dynamics and self-similarity; thus chaotic theory may well depict such nonlinear traffic dynamics. Chaotic theory has been well developed and applied in various areas in explaining and forecasting a complex nonlinear system. Therefore, this thesis attempts to use chaotic theory to predict short-term traffic dynamics. The prediction models are developed using phase space local approximation (PSLA), originally proposed by Farmer and Sidorowich, to capture the nonlinear dynamics of traffic flow patterns in different time intervals such as five, ten, and fifteen minutes. In order to examine the model's prediction accuracy, real traffic data from urban streets are collected and validated. The results of prediction show that correlation coefficients between observations and model outputs are greater than 0.6. The relative errors of prediction are also within the practically acceptable range. It suggests that short-term traffic dynamics can be satisfactorily explained by the deterministic chaotic models. If we consolidate traffic flow patterns of similar times of days and predict it individually, we can even obtain better prediction performance by 20%-30%. Compared with conventional ARIMA, chaotic PSLA model’s relative prediction errors are within 0.3-0.7, while ARIMA roughly estimates to 0.9. It is concluded that the prediction ability of chaotic PSLA is much better than ARIMA, particularly for the shorter-term traffic dynamics.