This study presents a novel approach for predicting fuel consumption in heavy-duty vehicles using a Machine Learning-based model, which is based on feedforward neural network (FFNN). The model is designed to enhance real-time vehicle monitoring, optimize route planning, and reduce both operational costs and environmental impact, making it particularly suitable for fleet management applications. Unlike traditional physics-based approaches, the FFNN relies solely on a refined selection of input variables, including vehicle speed, acceleration, altitude, road slope, ambient temperature, and engine power. Additionally, vehicle mass is estimated using a methodology presented elsewhere (see other abstract submitted to the ession ICE601 of SAE-ICE2025) and is included as an input for a better generalization of the consumption model. This parameter significantly impacts fuel consumption and is particularly challenging to obtain for heavy-duty vehicles. Engine power is derived from both engine torque and speed (RPM), ensuring a direct relationship with fuel consumption while keeping computational complexity low.
Experimental data were collected from a fleet of heavy-duty trucks under real-world operating conditions. In particular, the study focuses on a truck powered with a turbocharged diesel engine with maximum power of 353 kW. The acquisition system is based on an On-Board Unit (OBU) featuring CANbus connectivity, GPS tracking and 4G/LTE-5G communication. The OBU enables continuous logging of vehicle and engine parameters over each mission. Additional road data, such as altitude and slope, were obtained from external mapping services.
For the purpose of FFNN development, the training dataset was compiled from 10 different routes, capturing diverse driving conditions, while validation was conducted on 10 independent routes to assess model generalization. Before training, input data were pre-processed using normalization and standardization techniques to ensure stable convergence and mitigate the impact of scale differences among input variables. A correlation analysis was performed to evaluate the relationships among available parameters and fuel consumption, guiding the selection of the most informative inputs. This step reduced redundancy in the dataset and improved network efficiency. Furthermore, hyperparameter optimization was conducted using a Randomized Grid Search algorithm, enabling the identification of an optimal network architecture (i.e. number of layers and hidden neurons) and training parameters while minimizing overfitting.
The final model demonstrated high predictive accuracy across various validation routes, confirming the effectiveness of the FFNN in estimating fuel consumption with a reduced input set. Accuracy was tested on up to 50 routes, whose data where not considered for both training and validation; it was assessed that fuel consumption percentage error per route never exceeded 2%, while pointwise error remained consistently below 1%. This approach provides a practical and computationally efficient solution for fleet operators, facilitating advanced route planning and enabling more sustainable transportation strategies through cost-effective fuel management and emissions reduction.