Predição de fretes agrícolas baseado na aplicação de técnicas de aprendizado de máquina

Abstract

The grain freight cost is one of the biggest obstacles for the Brazil's competitiveness in international scenario. The road freight still is majority adopted on agricultural products transporting, and not unusual, that is the only choice to carry them. The cost weight, felt by the agribusiness operators, which work with supply chains might be decreased after the adoption of a freight management system. A promising alternative to deal with this challenge is the use of data mining techniques, which are able to extract patterns and trends in large amounts of data, which is why they have been increasingly used to support management decision in different areas, in place of intuition and resolutions based on experience. Thus, the project major goal is to develop and implement an intelligent system for forecasting road freight prices for agricultural grains. Thus, the aim of this research is predicting the freight price of agricultural commodities using models generated from Multi Layer Perceptron (MLP) and Support Vector Regression (SVR). The massive use of freight road in the logistic and grain distribution, reinforces the importance of predicting freight prices, therefore a system that could provide that kind of information to agribusiness managers and decision makers, would be critical on their daily activities. he results indicated that both techniques were efficient in estimating agricultural grain road freight prices in terms of R-squared and root mean square error (RMSE). Comparing the techniques, the SVR technique outperformed achieving an R-squared of 0.8921 and RMSE of 8.0464 for the range of routes up to 600km and 0.8924 and 21.1167 for routes longer than 600km, adopting specialized models for each data partition. In the other hand MLP resulted an R-squared 0.8785 and 8.5404 RMSE for short routes and 0.8506 and 24.8818 for long routes also making use of specialized models. In terms of using the general model or with partitioning the dataset as a function of distance traveled on the transportation route, using two individual models specialized in partitioning the set of routes into two ranges of up to 600km or greater, performed better when compared to adopting a general model for all two ranges of distances. This confirms the hypothesis that road freight prices show different behavior for different distance ranges.