Forecasting Delivery Time of Goods in Supply Chains Using Machine Learning Methods
Автор: Rezvanov V.K., Romakina O.M., Zaytseva E.V.
Журнал: Вестник Донского государственного технического университета @vestnik-donstu
Рубрика: Информатика, вычислительная техника и управление
Статья в выпуске: 2 т.25, 2025 года.
Бесплатный доступ
Introduction. Trade development requires the implementation of artificial intelligence and machine learning technologies to improve the accuracy of delivery forecasts. The scientific research published to date in this area appears insufficient for two reasons. First, it focuses primarily on global supply chains, although the issue is relevant for local businesses as well. Second, forecasting typically requires large amounts of data for machine learning and significant computing resources that are not available to the majority of companies. The presented study aims to fill these gaps and demonstrate the efficiency of using open, accessible data and known algorithms. The research objective is to describe a pattern of appropriate selection of the least resource-intensive delivery forecasting model based on the analysis of machine learning algorithms. Materials and Methods. The open data set DataCo Smart supply chain for big data analysis on deliveries in online trade was used. To process and analyze the information, methods of data cleaning, eliminating multicollinearity, normalization and coding of categorical features were applied. The following algorithms were used with the cleaned data: Decision tree, Random Forest, k-nearest neighbors, Naïve Bayes, Linear discriminant analysis, XGBoost, CatBoost, LightGBM, AdaBoost, and Perceptron. Results. The basic algorithm for the delivery forecasting model was the Decision Tree algorithm. This choice was due to its high accuracy, ease of use, and low risk of overfitting. The model evaluation showed a high determination coefficient close to one (0.986). Low values of the mean square error (0.0367) and mean absolute error (0.0324) were recorded. The model showed satisfactory results in terms of time spent on training (3.3087 s) and forecasting (0.0051 s). Actual and predicted values almost perfectly matched. Deviations from actual values were minimal. Discussion and Conclusion. The proposed model is efficient and has a high predictive ability. High-quality forecasting of delivery time is possible without the use of extensive databases and powerful computing resources. The study opens up the prospect of high-quality organization of logistics operations for small and medium enterprises. In further research, it is advisable to integrate weather data, traffic conditions and other indicators into the model. Using such information in real time will increase the adaptability and accuracy of forecasting.
Delivery time forecasting model, delivery forecast for small and medium enterprises, error in delivery forecasting, decision tree for logistics challenges
Короткий адрес: https://sciup.org/142244845
IDR: 142244845 | DOI: 10.23947/2687-1653-2025-25-2-120-128