Evaluation of Machine-Learning Approaches in the Automation of Irrigation Canals Using a Variable-Height Weir

• Published in: Journal of irrigation and drainage engineering 2024-12, Vol.150 (6)
• Main Authors: Zamani, Shahla, Parvaresh Rizi, Atefeh, Kouchakzadeh, Salah, Sajedi, Hedieh
• Format: Article
• Language: English
• Subjects: Algorithms; Artificial neural networks; Automation; Canals; Control algorithms; Control theory; Correlation coefficient; Correlation coefficients; Feedback control; Height; Irrigation; Irrigation canals; Irrigation water; K-nearest neighbors algorithm; Machine learning; Neural networks; Statistical analysis; Statistical models; Technical Papers; Water depth; Water discharge; Water distribution; Water engineering; Water levels; Weirs
• ISSN: 0733-9437; 1943-4774
• DOI: 10.1061/JIDEDH.IRENG-10327

AbstractAutomatic check structures can be important for water distribution in irrigation networks. In this research, a control algorithm was developed for a variable height whirling (VHW) weir, as a regulating structure equipped with a control mechanism. A local feedback controller was established for adjusting the flow depth upstream of the weir within a marginal target range. The control performance of the VHW weir was investigated using two methods: (1) K nearest neighbor (KNN); and (2) artificial neural network (ANN). The required data for methods were compiled in a long trapezoidal canal using different water depth targets. The inputs consisted of the discharge at the canal entrance, the variation of the discharge in three sequential periods, the water level deviation from the target value, and the offtake discharge. The model output was the set point of the instantaneous weir angle value, which represents the crest weir height, for maintaining the water depth within the target range. Different statistical indicators were employed to investigate the control performance. The results indicated that the ANN models, which were applied to cases with and without offtake in operation, provided 0.95 and 0.93 correlation coefficients, respectively. Also, the proposed neural model performed slightly better than the KNN algorithm, which yielded marginally higher error in output predictions.