Actor–Critic Deep Reinforcement Learning for Multi-Objective Intelligent Irrigation Scheduling: Algorithm and Edge-Cloud Management System

Abstract

Against the backdrop of increasingly prominent climate fluctuations and water scarcity, the demand for precision and intelligence in agricultural irrigation continues to rise. This article focuses on the research of "agricultural irrigation intelligent scheduling algorithm and management system based on deep reinforcement learning", aiming to construct a technical solution that combines decision-making adaptability and resource utilization efficiency. At the algorithmic level, a deep reinforcement learning model is constructed using an improved DQN combined with policy gradient fusion, ensuring consistency between algorithm description and system implementation to map multimodal data such as soil moisture, evapotranspiration, and meteorological predictions collected by field sensing networks into state representations in the irrigation strategy space. The strategy function is optimized using the Time Difference (TD) method to enable the system to continuously update decisions in a dynamic environment. In order to avoid the limitations of single objective optimization, a multi-objective reward function was designed, which integrates crop yield, water resource utilization rate, and energy consumption into the evaluation indicators, and achieves adaptive balance through normalization and weight adjustment. At the system implementation level, a management platform integrating data collection, edge computing, cloud decision-making and mobile visualization is built to support the automatic generation, real-time adjustment and historical data backtracking analysis of irrigation plans. Field trials on a 35-ha wheat–corn site (12 plots, 4 months) evaluated a DQN–Policy Gradient hybrid, trained for 5000 episodes (200 steps each) with lr=0.0005, batch size=64, and buffer=10,000. Rewards weighted efficiency (0.5), yield (0.3), and energy (0.2).The system achieved 88.1% ± 1.7% water use (n=30, p<0.01), representing a 12.7% improvement in water resource utilization, and 8.3% ± 1.2% yield gain (n=30, p<0.05), outperforming thresholds.The research results provide a scalable technical path for intelligent management of agricultural water conservancy, and provide practical verification for the application of deep reinforcement learning in complex resource scheduling scenarios.