Towards a machine-learning based approach for splitting cities in freight logistics context: Benchmarks of clustering and prediction models

Urban mobility consists of three basic components that create the essential functioning for the well-being of communities namely facilities, structures, and processes. Often these components have commentary roles where processes must assess infrastructures and related operations to support policies that relate to spatial structure and transportation patterns. Zoning tools are sample processes that are designed to split areas according to given criteria and purposes in order to better implement transport facilities, investments, and plans. This article proposes a sequential approach combining several machine-learning tools of clustering and forecasting that are thought efficient according to the Key Performance Indicators (KPI). In both processes, the proposed machine learning zoning approach (MLZA) has considered the location of sites requiring logistics services and the evolution of their demand, respectively, in order to accomplish a long-term splitting of urban land. For improving the performance of the clustering process, we have used 30 KPIs including all combinations of a number of built clusters. In doing so, this step has not aimed not only to validate a clustering tool but also to identify the optimal number of established zones. Based on simulated benchmarks, results have indicated that the clustering phase of the MLZA is still appropriate using the k-means algorithm. To evaluating forecast accuracy in the forecasting phase, we have measured the standard KPIs namely the MSE (Mean Squared Error), RMSE (Root Mean Square Error), MAPE (Mean Absolute Percentage Error), and R2 (R-squared). The Support Vector Machine (SVM) algorithm has been deemed to be the most efficient forecasting algorithm regarding the average values of the obtained performance measurements.