This paper deals with finding an appropriate data-driven model for a prediction of vehicle fuel consumption on destination. Several models are considered for this purpose, ranging from a simple polynomial-based models which include only vehicle travelled distance and payload mass as inputs to more complex neural network (NN)-based ones. Special effort is devoted to preparation of inputs for considered NNs, since driving cycles for which the fuel consumption is being predicted can vary in length, while NNs assume an input with fixed dimensions. To this end, two input types are defined and analysed: (i) 1D vector which contains counted discrete vehicle velocity values, and (ii) 2D matrix which contains counted combinations of discrete vehicle velocity and acceleration values. The structures of considered NN architectures are optimised to find the optimal number of their hidden layers and corresponding number of neurons. The proposed models are developed and validated based on a dataset collected by using GPS/GPRS equipment on a set of 10 delivery vehicles during 24h per day over three-month period. Finally, the impact of payload mass on the fuel consumption and its contribution to improvement of fuel prediction accuracy is comprehensively analysed.