A Spatial Model Averaging Approach for Predicting the Traffic Flow in Taiwan

基於空間模型平均法的台灣車流量預測

張志浩(Chih-Hao Chang) ； 許湘伶(Hsiang-Ling Hsu) ； 莊穎斌(Ying-Pin Zhuang) ； 楊洪鼎(Hong-Ding Yang)

《中國統計學報》 60卷4期 (2022/12) Pp. 261-301

摘要

When analyzing spatially dependent data with noise, predicting the spatial variables of interest is important. Many spatial prediction methods have been developed, each with its applicable occasion. Therefore, under the unknown mechanism of generating spatial variables of interest, there will be uncertainty in the choice of prediction by a single method. On the other hand, the experimenter will conduct data analysis under the assumption of stationary spatial correlation. However, the spatial correlation of data often does not satisfy stationary, and inferences under inappropriate assumptions may lead to unreasonable conclusions. Therefore, we propose three model average methods for spatial predicting based on the thin-plate smoothing spline and the spatially deformed Kriging predictions. For each model average prediction, the corresponding weight of each candidate prediction will be determined by the prediction variance meter of the candidate prediction methods and has the characteristics of fast computation and local adjustment of coordinates for prediction. After determining the final prediction method based on the simulation results, we analyze Taiwan's provincial highway traffic volume in 2020. We estimated the average daily number of vehicles (Amount) and the passenger car unit (PCU) at each monitoring station. The results illustrate that the uncertainty influence caused by the choice of forecasting method is avoidable, and the predicted performance is improved.

關鍵字

Spatial prediction ； Non-stationary covariogram ； Prediction uncertainty ； Prediction variation ； Model averaging

並列摘要

在分析帶有雜訊的空間相依資料時，如何對感興趣之空間變數進行預測是一重要課題。目前有許多空間預測方法已被發展且各自有適用的時機，因此在感興趣之空間變數的產生機制未知之下，由單一方法進行預測將存在預測方法選擇的不確定性。另一方面，常見的分析方法中，實驗者會在穩定態相關性假設之下進行資料分析與後續推論。然而實務上數據的空間相關性時常具有不滿足穩定態之現象，在不合適的假設下進行推論可能導致不合理的結論。因此，我們提出三種基於薄板樣條曲線預測法與空間變形克利金預測法的模型平均預測方法，各預測方法對應之權重是透過前述兩種候選預測法之預測變異數決定並具有計算快速和可隨進行預測之座標局部調整等特性。在透過模擬實驗的結果決定合適的最終預測方法後，我們將其應用於台灣2020年公路交通量分析，分別對各監測站的平均每日各車種總車輛數及標準車當量數進行估計，預期可以避免預測方法選擇所導致的不確定性，進而增進預測表現。

並列關鍵字

空間預測 ； 非穩定態共變異結構 ； 預測不確定性 ； 預測變異 ； 模型平均

參考文獻

Akaike, H. (1973). Information Theory and an Extension of the Maximum Likelihood Principle. In: B. N. Petrov, and F. Csáki eds. 2nd International Symposium on Information Theory. pages 267-281.

Bates, J. M., and Granger, C. W. J. (1969). The Combination of Forecasts. Operational Research Quarterly, 20(4), pages 451-468.

Bornn, L., Shaddick, G., and Zidek, J. V. (2012). Modeling Non-Stationary Processes through Dimension Expansion. Journal of the American Statistical Association, 107(497), pages 281-289.

Buckland, S. T., Burnham, K. P., and Augustin, N. H. (1997). Model Selection: An Integral Part of Inference. Biometrics, 53(2), pages 603-618.

Burnham, K. P., and Anderson, D. R. (2002). Model Selection and Multimodel Inference: a Practical Information-Theoretic Approach. Springer.