Modeling and Evaluation of Dynamical Properties of Different Energy Storage Systems Using Machine Learning Methods

Energy storage systems recently have played a significant role in energy management system utilized in urban areas. The importance of energy storages is beside the energy buffering ability, mainly related to smoothing energy flows, and minimizing generated energy. Therefore, modeling such systems has become increasingly important. To address the dynamic behavior of storage systems, machine learning (ML) techniques require extensive datasets including a variety of features of the systems internal states. However, obtaining such datasets often requires a large number of experiments and measurements. In this paper, the focus is shifted to datasets related to external system's states. A storage system modeling based on ML methods is investigated, considering the power forecasting or the state of charge/discharge (SOC and DOD). The focus is on the following technologies: Batteries (Lithium-ion) and polymer electrolyte membrane fuel cells. Unlike the actual state of art using different storage systems related features such as current, voltage, and inner variables such as inner cell temperature, the actual work is using external features such as environment temperature, input/demand energy and fuel, and time. A real experimental dataset is used to train nonlinear autoregressive exogenous (NARX) and long-short-term memory (LSTM) models to predict the SOC and DOD of the Lithium-ion battery in charge and discharge modes, as well as the power output of the fuel cell. The results show that the NARX model performs better for the battery system, while the LSTM model performs better for the fuel cell system.