兩步驟類神經網路車輛偵測器遺漏資料之填補及其應用

TWO-STAGE DATA IMPUTATION FOR MISSING VALUE OF VEHICLE DETECTORS AND ITS APPLICATIONS USING ARTIFICIAL NEURAL NETWORKS

本研究採用兩步驟資料填補方式，針對雪山隧道路段車輛偵測器遺漏資料之填補進行實證分析，以期找出其中最為適用之填補方法，並發展其可能之應用。於資料填補時，首先採用K-means法將資料分群，而後再以最具代表性之三種類神經網路分別進行填補。測試結果發現，將資料分為兩群，並採用回饋式類神經網路進行填補時可獲得最高之填補績效。最後，依據填補績效發展兩種可能之應用，即遺漏資料填補及偵測器佈設間距。在遺漏資料填補方面，速率填補之績效最高，無論是以上、下游任何一對偵測器資料作為輸入，準確度均高達97.5%以上。其次為流率，其準確度可達90%以上，並可以上、下游2或10對偵測器資料作為群1或群2填補之輸入。最差者為占有率，僅當準確度門檻降至80%時，群1資料方能進行填補，群2資料則無此限制。在偵測器佈設間距方面，若合併考量流率、速率與占有率三者，則佈設間距由填補績效最差之占有率決定。僅在整體準確度降至85%以下時，方可將現行之350m佈設間距擴增至3,500m。若僅考慮隨機性較低之群2資料，則在準確度高達90%以上時，即可將佈設間距增加至4,200m。

Using a two-stage data imputation method based on artificial neural networks, we carried out, in this study, an empirical analysis of the missing value of vehicle detectors in Hshehshan Tunnel to search for the optimal alternative, and developed its possible applications accordingly. By testing data imputation, we, at first, clustered all the data into groups using K-means, and then chose three typical artificial neural networks to impute the missing data. The result shows that two-group data clustering combined with a recurrent neural network can achieve the highest imputation performance. We, finally, developed two possible applications based on it, including data imputation and installation spacing of vehicle detectors. In respect to data imputation, speed performed the best with an accuracy of greater than 97.5%, and all pairs of vehicle detectors could be input for imputation. Flow performed the second best with an accuracy of over 90%, and the nearest two or ten pairs of detectors up- and downstream could be input for the imputation of data group 1 or 2, respectively. Occupancy performed the worst. Only by an accuracy threshold lowered to 80%, data points in group 1 could be imputed, and those in group 2 were not restricted, nevertheless. In respect to installation spacing, occupancy would dominate due to its relatively poor performance by considering all the three traffic attributes. Only when the overall accuracy decreased to fewer than 85% could we extend the current spacing of 350 m to 3,500 m. If only considering data group 2, we could extend it to 4,200 m with an accuracy of over 90% due to lower randomness.