Prediction of power system frequency response after generator outages using neural nets

A new methodology is presented for estimating the frequency behaviour of power systems necessary for an indication of under-frequency load shedding in steady-state security assessment. It is well known that large structural disturbances such as generator tripping or load outages can initiate cascading outages, system separation into islands, and even the complete breakup. The underfrequency load shedding takes place during the beginning phase of a dynamic change of the system frequency initiated by these distrubances. In this context the authors examine the ability of neural nets to properly interpolate among training data sets and to accurately predict the system frequency variations. The neutral-net approach provides a fairly accurate method of estimating the system average frequency response without making simplifications or neglecting non-linearities and small time constants in the equations of generating units, voltage regulators and turbines. The understanding and selection of the input features comes from the developed, simple low-order system frequency response model. Additional features are defined in terms of the centre of inertia acceleration and admittance distances. The efficiency of the new procedure is demonstrated using the New England power system model for a series of characteristic perturbations. The validity of the proposed approach is verified by comparison with the simulation of short-term dynamics including effects of control and automatic devices.