This article presents a novel approach for predicting fuel consumption in vehicles through a recurrent neural network (RNN) that uses only speed, acceleration, and road slope as input data. The model has been developed for real-time vehicle monitoring, route planning optimization, cost and emissions reduction and it is suitable for fleet-management purposes. To train and test the RNN, chosen after addressing several structures, experimental data have been measured on-board of a heavy-duty truck representative of a heavy-duty transportation company. Data have been acquired during typical daily missions, making use of an advanced connectivity platform, which features CANbus vehicle connection, GPS tracking, 4G/LTE - 5G connectivity, along with on-board data processing. The experimental data used for RNN train and test have been treated starting from on-board acquired raw data (e.g., speed, acceleration, fuel consumption, etc.) along with road slope downloaded from map providers. The improvement of the network performance has been achieved through a weight pruning procedure, to minimize instabilities and error amplification during fuel consumption prediction. RNN training has been performed using only one scheduled mission for both vehicles, but to distinct models (i.e., one for the bus and one for the truck) has been designed and tested on various routes, showing high accuracy in fuel consumption estimation. The achieved results proved RNN being capable of improving fuel consumption prediction on simulated routes, utilizing only few inputs, to support fleet operations in advanced route planning, with lower operating expenses and therefore reduced pollutant emissions.