Artificial Neural Network Based Energy Storage System Modeling for Hybrid Electric Vehicles 2000-01-1564
The modeling of the energy storage system (ESS) of a Hybrid Electric Vehicle (HEV) poses a considerable challenge. The problem is not amenable to physical modeling without simplifying assumptions that compromise the accuracy of such models. An alternative is to build conventional empirical models. Such models, however, are time-consuming to build and are data-intensive. In this paper, we demonstrate the application of an artificial neural network (ANN) to modeling the ESS. The model maps the system's state-of-charge (SOC) and the vehicle's power requirement to the bus voltage and current. We show that ANN models can accurately capture the complex, non-linear correlations accurately. Further, we propose and deploy our new technique, Smart Select, for designing ANN training data. The underlying philosophy of Smart Select is to design the training data set such that it is uniformly distributed over the entire range of an appropriate ANN output variable, which is typically the variable that is most difficult to model. In this case, we selected training data that were uniformly distributed over the current range. We show that smart-select is economical in comparison with conventional techniques for selection of training data. Using this technique and our in-house ANN software (the CUANN“), we developed an artificial neural network model (inputs=2, hidden neurons=3, outputs=2) utilizing only 1583 of the available 32,254 points. When validated on the remaining points, its predictive accuracy, measured by R-squared error, was 0.9978. Next, we describe the integration of the ANN model of the ESS into the MATLAB-SIMULINK environment of NREL's vehicle simulation software, ADVISOR. This yields a simpler implementation of the ESS module in ADVISOR and does away with certain tenuous assumptions in the original implementation.
Lastly, we also show how a dramatic reduction in the size of the training data set may be obtained by applying the model modifier approach developed by our research group at the University of Colorado at Boulder.