Exploring Random Forest and Multilayer Perceptron for Container Throughput Forecasting: The Case of the Port of Douala, Cameroon

Abstract

An accurate container throughput forecast is vital for any port. Since overall improvements in port performance and competitiveness can be derailed by port bottlenecks, ports need to find leverage to identify and prioritize measures to improve weak key performance indicators (KPI) to attain growth opportunities. Prior studies had modeled container throughput from socioeconomic and growth projection factors. This study aims to provide a practical method for forecasting the optimal container throughput a port can physically handle/attract given a certain level of terminal operation efficiency through random forest (RF) and multilayer perceptron (MLP) models. The study variables are derived from the port operations dimension and include ship turnaround time, vessel draft, container dwell time, berth productivity, container storage capacity, and customs declaration time. Evaluations are made based on the R-squared, NRMSE, MAE & MAPE. Model comparison is deduced with seven competing models in container throughput forecasting. The findings indicate that the RF model is a potential candidate for forecasting the engineering optimal throughput of the Douala port. Model interpretation is provided through feature importance and partial dependence plots. The findings from this study will help reduce uncertainty and provide leverage for port management to spot bottlenecks and engage in better port planning and development projects which will strengthen their international competitive advantage.