This paper describes an approach to model the transient and steady state behavior of differential microwave oscillators at the system level. Furthermore, phase noise characteristics can be emulated by the resulting models. The mathematical structure consist of a set of algebraic and differential equations that are solved by a VHDL-AMS interpreter. The nonlinearity occurring in the oscillator is described by multi-layer perceptron artificial neural networks (ANNs). In particular, it is shown how to model accurately the two mutually dependent outputs of differential oscillators without having an input signal as (phase) reference. The presented methodology can be employed independently of the used technology and the initial complexity of the oscillators.