MODELING AND CONTROL USING IMPROVED NEURO-FUZZY SYSTEMS

Abstract

Soft computing techniques are used to construct intelligent systems that emulate human reasoning, learning capabilities and adaptability. Fuzzy logic and the artificial neural networks are the most widely used soft computing techniques. The combination of fuzzy logic and neural networks possess advantages of both, bringing the low-level learning and computational power of neural networks into fuzzy systems and providing the high-level human-like thinking and reasoning of fuzzy systems into neural networks. Hence the combination, known as neural fuzzy system or neuro-fuzzy system, has become a popular research focus. Knowledge representation and automated learning capability of neuro-fuzzy system make it a powerful framework for machine learning problems. Nowadays learning methods in conventional neuro-fuzzy networks are improved to achieve better results in terms of accuracy and learning time. Extreme learning machine (ELM) is an emerging learning technique where hidden layer parameters are to be randomly initialized and the output parameters are calculated analytically. Incorporation of the ELM concept into neuro-fuzzy techniques will improve generalization accuracy and reduce complexity in learning algorithm. This thesis develops and uses an advanced neuro-fuzzy technique by integrating the fuzzy system and ELM learning for efficient regression and control applications. The main focus of this thesis is to investigate the applicability of advanced neuro-fuzzy systems in the field of modeling, prediction and control through software simulation and real time implementations. The core objectives of this research work are formulated as  Preliminary study of advanced neuro-fuzzy system called extreme learning adaptive neuro-fuzzy inference system (ELANFIS) and comparison with other ELM based neuro-fuzzy systems for regression problems.  Design and performance analysis of regularization based ELANFIS for accurate regression and classification.  Design and development of an ensemble time series prediction strategy using neuro-fuzzy system with empirical mode decomposition (EMD) technique.  Performance analysis of ELANFIS based controllers for model-based control strategies.  Real time implementation of ELANFIS based adaptive control strategy for renewable energy systems. iv The ELM based neuro-fuzzy systems have been recognized as a suitable tool for modeling and prediction. After a brief introduction, the thesis examines the application of ELANFIS for modeling, time series prediction and real world regression problems. A regularized version of ELANFIS is developed based on constrained optimization problem. The regularization parameter of the proposed technique has been calculated by search method and particle swarm optimization technique. Time series prediction is an important area of forecasting in which the past observations of the same variable are collected over time and used to develop a model to represent the internal structure. An ensemble prediction strategy using ELANFIS and empirical mode decomposition (EMD) is proposed for efficient time series prediction. EMD is a technique to convert nonlinear time series data into a set of stationary series consisting of a finite number of intrinsic mode functions (IMFs) and one residue. In the proposed strategy, the original nonlinear time series is first converted into a different sub-series using EMD. Then, the decomposed data is predicted using ELANFIS algorithm. Final prediction is obtained by the summation of outputs of all ELANFIS sub-models. The performance of proposed method is verified through wind speed and landslide displacement prediction. The suitability of ELANFIS models is also examined in model based control strategies. For this purpose, three model based control strategies namely inverse control, model predictive control and internal model control are introduced using ELANFIS. A significant improvement in the controller performance has been observed in ELANFIS controllers compared to other neuro-fuzzy systems. In order to establish its applicability for renewable energy system, ELANFIS based adaptive control strategy is designed for controlling the rotor side converter (RSC) of grid connected doubly fed induction generator (DFIG) based wind energy conversion system. The feasibility of the proposed ELANFIS control model for real-time applications is demonstrated through hardware-in-loop simulation in real time digital simulator (RTDS) environment. The controller is implemented on digital signal processor (DSP) based DSPACE 1104 module, interfaced thorough RTDS in the hardware-in-loop simulation environment