Enhancing the PI Distribution Network Performance with Hybrid ML Forecasting Techniques Applied to a Case Study of Electronic Products in Thailand

The consumption rate of different products such as electronic devices has drastically increased since the onset of COVID-19 in 2020. This shift in behavior has significantly impacted on the efficiency performance of supply chain in many countries. Using Physical Internet (PI) paradigm as an open global logistics system for increased efficiency and sustainability has been proved in different studies in the literature. Different forecasting models have been developed to predict demand fluctuations that have also been explored. However, in this work, we focus on designing different hybrid forecasting models in a dynamic way based on data analysis. Single and hybrid forecasting models, utilizing traditional and neural network techniques, have been used to design our models. The experimental data is drawn from a real case study of an electronic company in Thailand. These forecast outputs are then used to manage resources within a PI distribution network dynamically. The results indicate that hybrid forecasting models outperform other techniques in predicting the sales of three electronic Stock Keeping Units (SKUs), while only one SKU performs well with a single LSTM model. The appropriate hybrid combination model is dynamically chosen based on the forecasting performance of each SKU item. The forecast results also prove that the distribution distance, time, and cost in the PI distribution network are reduced by approximately 41%, 26%, and 49% for Bangkok and its metropolitan area, and these parameters decrease by around 14%, 11%, and 34% in the Central-Eastern regions.