機器學習應用於共用單車流量分佈

Abstract

隨著共用單車在城市生活中大規模普及和應用，不同區塊之間的供需失衡問題日益嚴重，開始影響市容和居民的正常出行。因此共用單車的及時、或者是超前調度顯得尤為重要。本研究建議應用機器學習（Machine Learning）聚類分析（Cluster Analysis）於交通類工具的流量分佈，它根據在地化的數據資訊，進行模擬演算，產生即時的視覺化流量資訊，同時分辨有任意形狀或者密度特性的資訊類別。除此之外，無樁模式的共用單車具有取車和還車地點的隨機性，如何架構合理的單車存量分區方式、計算分區內的單車密度、並根據單車密度作為參數控制的條件、以提出相對應的單車疏散與調度策略成為本研究的重點。 本研究採用R語言作為處理和分析數據、進行劃分區塊操作的程式語言。以虛擬環境軟體Rstudio，作為進行聚類分析和資料處理的載體。運用機器學習聚類分析有三種方法，包括：K-平均演算分群法（K-means）、密度分群法以及DBSCAN分群法（Density-based spatial clustering of applications with noise）。 本文以上海市靜安寺周邊以及上海四環（121.75°E-121.125°E，30.95°N-31.45°N）為研究範圍，探討兩種範圍內摩拜單車在單位時段內的取得量和歸還量，借鑒紐約市計程車與城市單車在聚類分析下的分區方式，探索最佳的上海共用單車分區方式。在最佳分區方式下，本文以預測相鄰時段內單車騎行目的地區塊的趨勢分佈為基礎，提出基於閾值控制思維的逆向流動獎勵制度作為單車的疏散與調度策略。 實作發現，在僅考量不同時間段內單車的流量變化時，以上海市靜安寺及其周邊為尺度的研究範圍的情況下，應採用用密度分群法，根據取車點和還車點的分佈建立各自獨立的獎勵機制分區，分別給予獎勵，得到的單車流動趨向較為合理，會使實施後的結果趨向于平衡狀態；以上海四環為尺度的研究範圍，應採用DBSCAN分群法，根據還車點的分佈建立密度視覺化監督系統，使得運營決策者能夠及時瞭解單車在不同密度值下呈現的分佈狀態，並以此做參考，來決定下一階段新投入運營單車的數量和投放地點。 未來，隨著資料透明化的推行，精確度更高的資料可以公開使用到學術領域，大量資料的導入訓練，可增強本研究結果的分區準確度和更短週期的獎勵方案，使獎勵機制、調度策略更加具體。

With the widespread adoption and application of shared bicycles in urban life, the problem of supply and demand imbalances between different blocks has become increasingly serious, and it has begun to affect the normal appearance of city appearance and residents. Therefore, timely or advanced scheduling of shared bicycles is particularly important. This study proposes the application of Machine Learning Cluster Analysis to the traffic distribution of traffic tools. Based on the localized data information, it performs simulation calculations to generate instant visual traffic information and distinguishes any Information categories for shape or density characteristics. In addition, the shared bicycle without pile mode has the randomness of the pick-up and drop-off locations, how to structure a reasonable method of partitioning the single-vehicle inventory, calculate the density of the bicycle in the partition, and use the density of the bicycle as a parameter control condition to propose Corresponding bicycle evacuation and scheduling strategies have become the focus of this study. This study uses R language as a programming language for processing and analyzing data and dividing blocks. The virtual environment software Rstudio is used as a carrier for cluster analysis and data processing. There are three methods for cluster analysis using machine learning, including: K-means, density clustering, and DBSCAN grouping (Density-based spatial clustering of applications with noise). In this paper, around the Jing'an Temple in Shanghai and Shanghai Sihuan (121.75°E-121.125°E, 30.95°N-31.45°N) as the research scope, the acquisition and return of Mobike bicycles in two ranges within a unit time are discussed. Invest in the zoning method of the New York City Taxi under cluster analysis to explore the best Shanghai shared bike partition method. Under the optimal partition method, based on the prediction of the trend distribution of bicycle riding destination blocks in adjacent periods, a reverse flow reward system based on threshold control thinking was proposed as the evacuation and scheduling strategy of bicycles. It is found that in the case of considering only the flow of bicycles in different time periods and using the scale of the Jing'an Temple and its surroundings as the scope of the study, a density clustering method should be used, based on the point of pick-up and return of the vehicle. Distribution and establishment of independent incentive mechanism zoning, respectively rewarded, the resulting bicycle flow tends to be more reasonable, will result in the results after the implementation tends to a balanced state; the scope of the study in Shanghai Sihuan as the yardstick, the DBSCAN grouping method should be used, according to The distribution of vehicle points establishes a visual density monitoring system, which enables the operational decision makers to know the distribution status of bicycles at different density values in a timely manner, and uses this as a reference to determine the number of new bicycles to be put into operation in the next phase and the location of the new bicycles. In the future, with the implementation of transparent data, more accurate data can be publicly used in the academic field. The introduction of a large amount of data can enhance the partition accuracy of this research result and a shorter-cycle reward program, enabling the reward mechanism. The scheduling strategy is more specific.