AGSTA-Net: Predicting Commercial Vehicle Fuel Consumption Using Adaptive Graph Spatiotemporal Attention Network

Commercial vehicles play an important role in road transport, but this is accompanied by energy consumption and environmental pollution. In this context, this paper proposes a commercial vehicle fuel consumption prediction model AGSTA-Net, which is based on vehicle operation data, firstly, the spatial dependence capture module extracts the spatial information of the feature maps by using graph neural network, which can help to capture the influence of different factors on fuel consumption; secondly, the temporal dependence capture module combines the long and short-term memory network and the improved self-attention mechanism to make full use of local and global temporal dependence in the time series of vehicle operation data. local and global time dependencies in the data time series. In addition, in order to solve the imbalance problem of fuel consumption data and improve the training efficiency, an adaptive smoothing L1 loss function is designed in this paper to learn the thresholds from the data autonomously. This helps to provide a better trade-off between L1 and L2 losses, accelerating model convergence and improving prediction accuracy. Experimental results using datasets from major Chinese commercial vehicle manufacturers show that AGSTA-Net outperforms other models in the fuel consumption prediction task with good prediction accuracy and stability.