Abstract: This thesis develops a spatio-temporal data mining method for uncertain water reservoir data. The goal of the data mining method is to learn from a history human reservoir operations in order to derive an automated controller for a reservoir system. Spatio-temporal data mining is a challenging task due to the reasons: (1) spatio-temporal datasets are usually much larger than spatial data sets, (2) many common spatial techniques are unable to deal with objects that change location, size or shape, and (3) complex and often non-linear spatio-temporal relationships cannot be separated into pure spatial and pure temporal relationships.
Support Vector Machines (SVMs) have been extensively and successfully applied in feature selection for many real-time applications. In this thesis, we use SVM feature selection to reduce redundant and non-discriminative features in order to improve the computational time of SVM-based data mining. We also propose combining Principal Component Analysis (PCA) with multi-class SVMs. We show that SVMs are invariant under PCA transformations and that PCA is a desirable dimension-reduction method for SVMs. We propose also an extension of the SVM Regression approach to be able to perform spatio-temporal data mining.
As a case study, we apply our spatio-temporal data mining method to derive an automated controller for the North Platte River Reservoir system. This reservoir system has multiple reservoirs, whose spatial location and the variables in each reservoir have been incorporated in our reservoir operations model. Further, each reservoir stage or status changes over time by the opening and closing of a dam to control the water levels. We show that by inputting the selected features from a spatio-temporal dataset by the PCA has achieved excellent results and could speed up the evaluation of data mining by SVM by an order of 10 while maintaining comparable accuracy. The North Platte River Reservoir system case study shows that the SVM Regression approach combined with PCA is an efficient tool for spatio-temporal data mining.
Committee members: Dr. Peter Revesz (Advisor), Dr. Jitender Deogun and Dr. Ashok Samal