Deep Learning Based Distributed Meta-Learning for Fast and Accurate Online Adaptive Powertrain Fuel Consumption Modeling

Connected vehicle-based distributed meta-regression (CV-DMR) algorithms from our previous work facilitate online adaptive microscopic fuel consumption modeling, which can support model-based vehicle and powertrain control (MB-VPC) for fuel-efficient vehicle operations. By taking advantage of accessible online computational resources from connected vehicles (CVs) and CV remote data center, CV-DMR can adapt the microscopic fuel consumption model to an unseen driving condition with limited training data. However, the computational complexity of obtaining the normalized feature map in CV-DMR during training is high. It can take long time to adapt the model to a new driving condition, especially when the new driving condition is considerably different from the pre-learnt driving conditions. To reduce the computational complexity of learning/training, we propose to apply deep neural networks as the model representation method and newly design efficient deep learning based CV-supported distributed meta-learning (CV-DML) algorithms for adaptive fuel consumption modeling. To further generalize the model and reduce model adaptation time, the concept of meta-suggestion is newly proposed in CV-DML. Substantial proof-of-concept experiments are conducted with steady-state and transient vehicle engine data to evaluate the model performance in terms of model prediction accuracy, model adaptation speed, and fuel saving. Compared to the physical model and non-CV-supported model adaptation, the prediction accuracy of CV-DML is improved by 34%–87% and 17%–30%, respectively. Due to the improved accuracy, up to 9.4% of fuel is saved. To achieve the same the prediction accuracy and fuel saving, the model adaptation of CV-DML is 18 to 1,300 times faster than that of CV-DMR.