Machine Learning Approaches for Error Correction of Hydraulic Simulation Models for Canal Flow Schemes

Abstract

Modernization of today’s irrigation systems attempts to improve system efficiency and management effectiveness of every component of the system (reservoirs, canals, and gates) using automation technologies, along with hydraulic simulation models. The canal flow control scheme resulting from the coupling of the system automation and the simulation models has proven to be an excellent irrigation water management instrument around the world. Nevertheless, the harsh environment of irrigation systems can induce uncertainties or errors in the components of canal flow control that can worsen over time, misleading or confusing both human and computer controllers. These errors can be attributed to parameter measurement and conceptual sources, with the complexity of locating their individual origin. In this paper, a framework is presented to minimize the collective or aggregate error within an irrigation canal flow control scheme that uses a learning machine algorithm (multilayer perceptron and relevance vector machine) embedded in a hydraulic simulation model fed by a canal automation system. This framework is evaluated using actual data from an irrigation conveyance canal located at the Lower Sevier River Basin in Utah. The results obtained prove the adequacy of the proposed framework in minimizing the aggregate error, which affects the simulation results of the automation system (up to 91% in bias and 83% in maximum absolute error) when compared with the original values obtained in the verification period. The temporal correlation of the aggregate error was also significantly reduced, thus resulting in reduced local biases and structures in the model prediction error.