Extended cluster weighted modeling methods for transient recognition control
This dissertation considers cluster weighted modeling (CWM), a novel non-linear modeling technique for the electric load transient recognition problem. The original version of CWM is extended with a new training algorithm and a real-time CWM prediction method. In the training process, a new training algorithm is derived that is a mixture of expectation maximization (EM) - least mean squares (LMS). The algorithm addresses the singular matrix inversion problem in EM. A recursive EM-LMS algorithm is developed that allows the CWM to adapt to time varying systems. In the prediction process, a sequential version of CWM prediction based on the novel idea of tail prediction is proposed to improve the real-time performance of load transient recognition. This idea also gives rise to a real-time transient overlapping resolution method that is critical for robust online operation. Other aspects of real-time transient processing methods, such as transient scaling, detection under conditions of transient overlap, and off-training set transient indication are also developed and combined into the sequential CWM model. The sequential CWM technique is applied to an electric load transient recognition model for hybrid fuel cell system control. The model provides real-time information about the steady-state behavior of incoming load transients to control and allocate power between the fast sources and the fuel cell. An implementation of this system in a Xilinx FPGA is discussed.