Derived operating rules for a reservoir operation system: Comparison of decision trees, neural decision trees and fuzzy decision trees

This article compares the decision-tree algorithm (C5.0), neural decision-tree algorithm (NDT) and fuzzy decision-tree algorithm (FIDs) for addressing reservoir operations regarding water supply during normal periods. The conventional decision-tree algorithm, such as ID3 and C5.0, executes rapidly and can easily be translated into if-then-else rules. However, the C5.0 algorithm cannot discover dependencies among attributes and cannot treat the non-axis-parallel class boundaries of data. The basic concepts of the two algorithms presented are: (1) NDT algorithm combines the neural network technologies and conventional decision-tree algorithm capabilities, and (2) FIDs algorithm extends to apply fuzzy sets for all attributes with membership function grades and generates a fuzzy decision tree. In order to obtain higher classification rates in FIDs, the flexible trapezoid fuzzy sets are employed to define membership functions. Furthermore, an intelligent genetic algorithm is utilized to optimize the large number of variables in fuzzy decision-tree design. The applicability of the presented algorithms is demonstrated through a case study of the Shihmen Reservoir system. A network flow optimization model for analyzing long-term supply demand is employed to generate the input-output patterns. Findings show superior performance of the FIDs model in contrast with C5.0, NDT and current reservoir operating rules.