| Abstract | In an effort to support the International Energy Agency’s (IEA) Annex 32 goals, the National Research Council Canada has invested considerable effort towards the testing and modeling of energy storage systems (ESS). One recent development in this modeling was the creation of a battery cell surrogate model for voltage prediction using deep learning methods. The model was developed to predict a cell voltage at a given current, state of charge (SOC), and temperature operating condition. The deep learning modeling framework provided the ability to capture the intricacies of the battery cell behaviour without the need for computationally expensive analytical equations, and without the requirement for in-depth understanding of the electrochemical processes occurring within the cell. Two models were developed: the first being a standard gated recurrent unit (GRU) network implementation, and the second being a GRU implementation with transfer learning. The first model was trained exclusively using experimental data collected in the ground testing campaign of the Hybrid-Electric Aircraft Testbed (HEAT) II project. This campaign consisted of a variety of performance tests in which the battery was discharged across a large portion of its operational range. The model was observed to perform well with respect to its training and validation losses, but was visibly lacking in accuracy at low state of charge conditions due to lack of experimental data. The second model was developed using transfer learning in an attempt to reduce the first model’s deficiencies within the low state of charge regions. The model was pre-trained on a large simulated data set generated from a Modelica equivalent circuit model, and transfer learning was applied to transfer the weights to a new model which was further trained on the smaller experimental dataset. Marginal improvement was realized by the transfer learning model, however the predictions at low state of charge continued to be slightly inaccurate. The purpose of this report is to describe the development process of the model, the progress made to date, as well as recommendations for future work. |
|---|
