Development of a NARX State-of-Charge Predictor based on Active Power Demand

Abstract

A simple neural network state-of-charge predictor trained on one-year of energy storage system data is presented. The model uses the active power command and the state-of-charge for the current time-step, and implements a nonlinear auto-regressive network with exogenous inputs to predict the state-of-charge at the subsequent time-step. The neural network training algorithm is written in the Julia programming language, independent of any existing machine learning platforms; the resulting model is compared to one developed using Python/TensorFlow. The simulation performance was validated with data collected from the energy storage system that was dispatched to follow a standard frequency regulation duty cycle not used as part of the training data. The mean-absolute-error between the predicted state of charge and the validation data is shown to be less then 1%, despite the limited data and lack of physical information about the system.