APPLICATION OF MULTIREGRESSIVE LINEAR MODELS, DYNAMIC KRIGING MODELS AND NEURAL NETWORK MODELS TO PREDICTIVE MAINTENANCE OF HYDROELECTRIC POWER SYSTEMS

Abstract Since the operational conditions of a hydroelectric unit can vary within a wide range, the monitoring system must be able to distinguish between the variations of the monitored variable caused by variations of the operation conditions and those due to arising and progressing of failures and misoperations. The paper aims to identify the best technique to be adopted for the monitoring system. Three different methods have been implemented and compared. Two of them use statistical techniques: the first, the linear multiple regression, expresses the monitored variable as a linear function of the process parameters (independent variables), while the second, the dynamic kriging technique, is a modified technique of multiple linear regression representing the monitored variable as a linear combination of the process variables in such a way as to minimize the variance of the estimate error. The third is based on neural networks. Tests have shown that the monitoring system based on the kriging technique is not affected by some problems common to the other two models e.g. the requirement of a large amount of data for their tuning, both for training the neural network and defining the optimum plane for the multiple regression, not only in the system starting phase but also after a trivial operation of maintenance involving the substitution of machinery components having a direct impact on the observed variable. Or, in addition, the necessity of different models to describe in a satisfactory way the different ranges of operation of the plant. The monitoring system based on the kriging statistical technique overrides the previous difficulties: it does not require a large amount of data to be tuned and is immediately operational: given two points, the third can be immediately estimated; in addition the model follows the system without adapting itself to it. The results of the experimentation performed seem to indicate that a model based on a neural network or on a linear multiple regression is not optimal, and that a different approach is necessary to reduce the amount of work during the learning phase using, when available, all the information stored during the initial phase of the plant to build the reference baseline, elaborating, if it is the case, the raw information available. A mixed approach using the kriging statistical technique and neural network techniques could optimise the result.

[1]  S. Masri,et al.  Identification of Nonlinear Dynamic Systems Using Neural Networks , 1993 .

[2]  A. Narayanan Probability and statistics in engineering and management science , 1972 .

[3]  O. Dubrule Comparing splines and kriging , 1984 .

[4]  Rob W. Parrott,et al.  Using kriging for 3D medical imaging. , 1993, Computerized medical imaging and graphics : the official journal of the Computerized Medical Imaging Society.

[5]  R. Reyment,et al.  Statistics and Data Analysis in Geology. , 1988 .