Machine Learning-Based Prediction of Torque, Draft Force and Power Consumption in Vertical Rotary Tillage

Abstract

Tillage machines play a critical role in agricultural production by improving soil structure through aeration, disintegration, and mixing, while typically operating under high energy demand. Among the available implements, vertical rotary tillers are widely preferred for seedbed preparation owing to their intensive soil–tool interaction and favorable energy efficiency compared with conventional tillage equipment. However, the torque, draft force, and power consumption of vertical rotary tillers are strongly influenced by operational parameters such as rotor speed and blade geometry, and accurately predicting these parameters remains a key challenge for optimizing machine design and energy use. This study investigates the effects of rotor speed (220, 270, and 310 rpm), horizontal blade angles (0°, 5°, 10°, and 15°), and vertical blade angles (0°, 10°, 20°, 30°, and 40°) on torque, draft force, and power consumption of a vertical rotary tiller under controlled laboratory soil bin conditions. Torque and draft force were measured using a custom-designed torque measurement system and a load cell, respectively, while soil penetration resistance and moisture content were maintained at 0.2–0.5 MPa and 18% ± 1%. The experimental dataset, obtained from a full factorial design, was analyzed using six trained machine learning models: Gradient Boosting (GB), Light Gradient Boosting Machine (LightGBM), Decision Tree (DT), K-Nearest Neighbors (KNN), Artificial Neural Networks (ANN), and Support Vector Machine (SVM). Model performance was evaluated using R², MAPE, MAE, and RMSE metrics, supported by 5-fold cross-validation to assess generalization capability. The results showed that Gradient Boosting achieved the highest predictive accuracy across all output parameters (R² up to 0.999), while Decision Tree and LightGBM models also produced consistently high accuracy with low error metrics, whereas ANN and SVM exhibited comparatively higher prediction errors, particularly for draft force and power consumption. These results confirm the applicability and reliability of machine learning models, particularly GB and DT, for predicting key performance parameters in vertical rotary tillage under varying operational conditions.