Data series for paratransit representative driving cycles
Description
Contains time series data for representative driving cycle for paratransit vehicles, constructed from data gathered on vehicles travelling around Stellenbosch, South Africa. From the paper "Developing a representative driving cycle for paratransit with time series shape-based clustering" (forthcoming). The raw data are available in: Hull, Christopher; Giliomee, Johan; Collett, Katherine; mcculloch, malcolm; Booysen, MJ (2023), “1Hz GPS Tracking Data on Minibus Taxi Paratransit Vehicles in South Africa”, Mendeley Data, V2, doi: 10.17632/xt69cnwh56.2
Files
Steps to reproduce
Using the data from "1Hz GPS Tracking Data on Minibus Taxi Paratransit Vehicles in South Africa", available here: https://data.mendeley.com/drafts/xt69cnwh56?folder= This description follows the methodology from "Developing a representative driving cycle for paratransit with time series shape-based clustering", where a more detailed description of the methodology to develop the representative driving cycles and further applications can be found. 1. Read in all data 2. Smoothen data with rolling window average of 3 seconds. 3. Split the raw data into micro-trips, which in this application are trip segments that start and end when the vehicle is stopped (speed < 2km/h), and have a minimum length of least 20 seconds or 180 seconds depending on which driving cycle is being reproduced (20 seconds for the cycle in "representative_drive_cycle_tmin_20_nc_8.csv" and 180 seconds for the cycle in "representative_drive_cycle_tmin_180_nc_6.csv"). 4. Use DTW-C++ algorithm (found here: https://github.com/Battery-Intelligence-Lab/dtw-cpp), to cluster the micro-trips. Time series shaped similarity is defined by Dynamic Time Warping (DTW, and the clusterings are found using Mixed Integer Programming (MIP), ensuring a globally optimal clustering solution is found. The number of clusters used to develop the driving cycles for tmin 20 and 180 are 8 and 6 respectively. The optimal number of clusters was chosen based on the elbow and silhouette methods. 5. Calculate cluster weights as the size of each cluster proportional to the size of the smallest cluster, and select n-nearest micro-trips to each cluster center based on the relative weights of the cluster. 6. Sequence the micro-trips using Viterbi programming and a draw without replace strategy. Starting with the micro-trip that is most likely to be seen at the start of a trip, draw the next most likely micro-trip to be seen and append it to the driving cycle, until no more micro-trips remain. The processed data includes: Date (DD/MM/YYYY) Time (HH:MM:SS) Speed (km/h) Acceleration (m/s^2)