A data-driven accurate battery model to use in probabilistic analyses of power systems

Abstract

Analytical battery models depend on a set of complex nonlinear equations that make them impractical to use in probabilistic analyses (e.g., reliability evaluation) of power systems. Machine-learning algorithms have the potential to reduce or even avoid the computational complexities of incorporating actual battery characteristics in probabilistic analyses. In this paper, a neural network (NN)-based approach is proposed to develop a battery model that captures the non-linear interactions between charging/discharging power and battery state of charge (SoC). In the proposed approach, an NN with a rectified linear unit activation function is trained using historical data generated from an experimentally validated battery model. Another NN with linear activation function is trained to capture the relationship between charging/discharging power limits and SoC. Weights and biases of the trained networks in conjunction with mixed integer linear programming are used to develop an accurate and computationally attractive battery model. Also, a mathematical model is formulated to accommodate the proposed battery model in power system reliability evaluation. Moreover, a genetic algorithm-based approach is used to determine optimal locations for batteries considering the developed model to enhance reliability of power systems. The proposed approach is demonstrated on a modified version of the IEEE 33-bus distribution system. Monte Carlo simulation is performed to calculate reliability indices. The results show that the proposed battery model is effective to incorporate actual battery characteristics in probabilistic analyses (evaluating reliability and finding optimal battery locations) of power systems.