Abstract
This paper presents a framework for formulating and training artificial neural networks (ANNs) to learn a representation of primary frequency reserves in power systems from stochastic operational data. There is growing concern regarding the frequency quality of power systems. This has led some transmission system operators to implement prequalification requirements for primary frequency reserve units that describe a series of tests aimed at mapping the characteristics of the potential reserve providing unit. However, in their current form, these tests are costly and few, if any, published alternative approaches currently exist. To address this, we present a method of determining the relevant characteristics—e.g., droop—from ordinary operational data. We train an ANN to approximate the unit behavior by matching the predicted and data-estimated moments of active power, using the unscented transform. We demonstrate the method on a standard test system and investigate its robustness. This method not only provides a potentially cheaper and less invasive identification method for primary frequency reserves, but also a more expressive, stochastic description of reserve providing units, compared to the deterministic methods currently in use.