نوع مقاله : مقاله پژوهشی
نویسندگان
1 گروه مهندسی آب، دانشکده کشاورزی، دانشگاه زنجان، زنجان، ایران.
2 گروه علوم و مهندسی آب، دانشکده کشاورزی، دانشگاه بوعلی سینا همدان، همدان، ایران.
چکیده
کلیدواژهها
موضوعات
عنوان مقاله [English]
نویسندگان [English]
Optimal water resource management in irrigation networks has been recognized as a key strategy to address water scarcity challenges. This research developed a novel hybrid method based on inverse reinforcement learning (IRL) and hydrodynamic modeling with ICSS to optimize the operation of the E1-R1 canal in the Dez network. In this method, the reward function was automatically extracted from the experiences of expert operators (simulated using ICSS), and then, by combining it with the hydrodynamic simulation model, optimal control policies for water control structures and intakes were determined. Operational data under various scenarios, including random inflow rates (1.1 to 3.1 cubic meters per second) and different water withdrawal patterns, were analyzed. The results indicated high efficiency of the proposed system, with an average water delivery efficiency of 0.97, a supply adequacy of 0.95, and a maximum depth control error of 14.3%. The cumulative depth error also remained below 10% in all scenarios, indicating long-term system stability. These findings confirm that the IRL approach, by learning the implicit knowledge of operators and converting it into policies, can effectively reduce water losses and improve the performance of irrigation networks.
کلیدواژهها [English]
EXTENDED ABSTRACT
Many people worldwide experience water scarcity in their daily lives. Therefore, water resource management is a serious challenge for communities. A significant portion of water resource consumption in a basin is used in agriculture and irrigation and drainage networks, which act like the human body's veins, serving as the main arteries for water distribution and conveyance in an agricultural area. Improper operation of these networks leads to significant water losses, making their efficient management crucial for substantially reducing these losses.
This research develops an inverse reinforcement learning algorithm integrated with a hydrodynamic canal simulation model to address this issue.
To achieve the objective, a reward function is learned based on the different actions of the learning agent (water control structures such as weirs and intakes) in various states (obtained based on an expert's experience) in the inverse reinforcement learning algorithm. Then, using this reward function, the actions corresponding to different states are extracted. Using operational data and considering various scenarios, operational patterns in the E1-R1 canal of the Dez network were extracted.
The findings of this research showed that in most scenarios, the calculated values for the delivery efficiency index are close to one. The key parameters affecting the calculation of this index include the number of off-takes, the required (or requested) discharge, and the actual delivered discharge. According to the standard range of Molden and Gates performance evaluation indices for delivery efficiency, if the values fall within the range of 0.85 to 1, the water is delivered to consumers with higher efficiency, effectively minimizing water losses or excess delivery. In this study, based on the obtained results, the minimum and maximum delivery efficiencies were 0.97 and 1, respectively, which fall within the standard range and indicate a favorable performance. Overall, the efficiency values for Off-takes No. 1, 2, 5, and 6 were in their ideal state across all scenarios. Additionally, the average delivery efficiency for Off-takes No. 1 to 6 was 1, 1, 0.98, 0.98, 1, and 1, respectively.
The cumulative absolute error values were all below 10%. The minimum and maximum cumulative absolute errors were 0.5% and 2.4%, respectively, with an average error of 1.65% across all scenarios - all indicating good performance. The cumulative absolute error values also remained below 10%. The minimum and maximum errors were 1% and 9.4%, respectively, with an average error of 4.34% across all scenarios. Overall, these results similarly fall within the acceptable performance range.
Based on the various defined scenarios and the results of simulations and learning, all evaluation indices fell within the "good" performance class. This confirms the successful performance of the inverse reinforcement learning algorithm and validates the learned reward function.
The evaluation results for water depth under the inverse reinforcement learning method demonstrated that the model consistently maintained water level variations within the permissible depth range across all scenarios. In all simulated scenarios, the depth remained within the allowable limits, and the average depth fluctuations were centered around the target depth.
All authors contributed equally to the conceptualization of the article and writing of the original and subsequent drafts.
Data available on request from the corresponding author.
The author declares no conflict of interest.
)