Frequency regulation is becoming increasingly important with deeper penetration of variable generation resources. Flexible loads have been proposed as a low-cost provider of frequency regulation. For example, the flexibility of loads with inherent thermal energy storage resides in their ability to vary their electricity consumption without compromising their end function. In this context, the aggregate flexibility of a collection of diverse residential air-conditioning loads has previously been shown to be well modeled as a virtual battery using first principles load models. This analytical method will not scale to more complex flexible loads such as commercial HVAC systems. This paper presents a method to identify virtual battery model parameters for these more complex flexible loads. The method extracts the parameters of the virtual battery model by stress-testing a detailed software model of the physical system. Synthetic examples reveal the effectiveness of the proposed identification technique.