Multiobjective Allocation of Rotational Irrigation Groups in Large-Scale Tree-Structured Pipe Networks Using a Deep Q-Network

Abstract

To address the challenge of irrigation-group scheduling in large-scale, gravity-fed, tree-structured drip irrigation pipe networks, this study develops a dynamic-weight deep multiobjective Q-network (DMOQN). The proposed method employs action masking to enforce the minimum operating head constraint and uses a differential vector reward to jointly reduce the mean and variance of surplus head across irrigation groups within one irrigation cycle. The method was evaluated through an engineering case study of a real drip-irrigation network in Xinjiang, China. Compared with conventional multiobjective optimization algorithms, DMOQN achieved broader Pareto-front coverage. Relative to the baseline schedule, the selected optimized solution reduced the mean and variance of surplus head by 3.71% and 5.24%, respectively. After offline training, the end-to-end inference time for a full irrigation cycle was 0.561 s, approximately four orders of magnitude faster than online search using nondominated sorting genetic algorithm II (NSGA-II), multiobjective evolutionary algorithm based on decomposition (MOEA/D), and strength Pareto evolutionary Algorithm 2 (SPEA2). These findings indicate that DMOQN can support rapid, deployable optimization of irrigation grouping in large-scale tree-structured irrigation pipe networks, and field monitoring validation remains an important direction for future work.