In order to investigate the impact of refined driving behavior on the predictive performance of fuel consumption models based on individual and hybrid neural networks, the vehicular jerk was introduced as one crucial input variable in the training process. A total of eight typical neural network models were employed, including Long Short-Term Memory (LSTM), Recurrent Neural Network (RNN), Nonlinear Autoregressive model with Exogenous inputs (NARX), Generalized Regression Neural Network (GRNN), Convolutional Neural Network (CNN), Gated Recurrent Unit (GRU), Multilayer Perceptron (MLP), and Convolutional Neural Network-Long Short-Term Memory (CNN-LSTM) hybrid networks. Three input parameter combinations, namely (speed, acceleration), (speed, acceleration, and Jerk) and (engine speed), were selected, and 3 working scenarios, namely the low-speed campus scenario, the medium-speed urban scenario, and the high-speed expressway scenario, were selected, resulting in a total of 69 experiments setups. The results show that, among the seven individual neural network models, the LSTM model demonstrates the best predictive performance across all input combinations and driving conditions. The CNN-LSTM hybrid model exhibited slightly superior predictive performance compared to the LSTM model. Including vehicle jerk significantly improved the predictive performance of all neural network-based fuel consumption models across different speed conditions. Among the individual models, the Root Mean Square Error (RMSE) decreased by up to 43.2% (CNN model, highway condition), the Relative Error (RE) decreased by up to 68.2% (LSTM model, urban condition), and the coefficient of determination (R2) improved by up to 41.8% (NARX model, urban condition). In the hybrid models, RMSE and RE decreased by up to 34.9% and 61.0%, respectively (urban condition).
