A Modular Neural Block to Enhance Power System Stability

Abstract

Conventional supplementary controllers (CSCs) can still be widely observed in power system utilities. This work aims to develop a modular neural block (MNB) to improve control performance and stability of CSCs-aided power systems. The proposed MNB is actually a one-to-one offline trained self-recurrent wavelet neural network (SRWNN) which can be modularity added to the PI/PD/PID/Lag-Lead controllers to enhance their performance by adding an adaptive property to them. Independent of the plant model, the MNB is initially trained offline using virtual training-samples. As a prefabricated one-to-one neural block, it can then be copied to required numbers and added to the lag-lead controller or any/all branches of the PI/PD/PID controller in series connection. The employed MNB(s) is then re-trained online to increase control performance by minimizing a predefined cost-function. The online training is performed by back-propagation (BP) algorithm while the closed-loop stability is guaranteed by an efficient Lyapunov-based approach. The proposed approach is thus a model-free scheme which is simple enough for implementation. Ability of the MNB to enhance the performance of the CSCs and dynamic stability of power systems is demonstrated by the simulation results of two small power plants and an IEEE 10-machine 39-bus system.