Online Levenberg-Marquardt algorithm for neural network based estimation and control of power systems

Levenberg-Marquardt (LM) algorithm, a powerful off-line batch training method for neural networks, is adapted here for online estimation of power system dynamic behavior. A special form of neural network compatible with the feedback linearization framework is used to enable non-linear self-tuning control. Use of LM is shown to yield better closed-loop performance compared to conventional recursive least square (RLS) approach. For successive disturbance use of LM in conjunction with non-linear neural network structure yields faster convergence compared to RLS. A case study on a test system demonstrates the effectiveness of the online LM method for both linear and nonlinear estimation over RLS estimation (linear).