This paper presents an empirical study on the prediction of the instantaneous fuel consumption of public transport buses using LSTM type recurrent neural networks. The analyses were conducted on selected repetitive Sort 2 driving cycles. This allowed for stable test conditions and control of data variability. For the analyses, valid measurements including vehicle speed, accelerator pedal position (APP) and instantaneous fuel consumption (l/h) were used. Five LSTM modelling strategies were developed and compared: a baseline model, an in-depth model with dropout, an advanced model with callbacks, a model with a special weighted loss function for idling periods and a FuelNet model for fuel consumption prediction . The results indicate high prediction performance (MAE, RMSE, R²) and the potential for practical implementation of the model in fleet management systems.