Near real-time management of irrigation canals using support vector machines

Support vector machines (SVMs) represent a relatively new technique in machine learning methods. SVMs can be trained to mimic or forecast the behavior of nonlinear systems without explicitly modeling the physical cause-effect relationships in those systems. This makes SVMs a potentially promising technique for use in hydraulic and hydrological applications. This study explored the use of SVMs in irrigation canal management. A model for determining required canal diversions on a near real-time basis was developed using SVMs. The approach is particularly appropriate for operation of canals that deliver water to irrigators following an "on demand" rule. A support vector machine model is presented for operation of the Sevier Valley-Piute Canal in the Sevier River Basin of southern Utah. The Sevier Valley-Piute Canal has long travel time and substantial and highly variable seepage losses which contribute to uncertainty in its operation. The principal goal of the model is to predict the optimal quantity of water that should be diverted into the canal over a 24-hour period that will satisfy the coming day's irrigation needs on the one hand, and minimize the quantity of water spilled at the end of the canal, on the other.; The results of this modeling effort show excellent performance of the SVM model in predicting the required canal diversions. The performance of the model is evaluated by standard goodness-of-fit measures. Sensitivity of SVM parameters is also examined in this study. This application of SVMs demonstrates their forecasting capability and the possibility of using them for improving canal management and flow prediction in large irrigation systems.