Abstract <p>The maintenance tasks scheduling optimization is important and challenging for improving the oil and gas exploitation efficiency. Traditionally, this problem is addressed using exact algorithms, metaheuristic algorithms or solvers. However, due to the large-scale nature of this problem, these trials often fail in practical use. To address this, a compositional message passing neural network (CMPNN) is introduced for graph embedding, and the messages of the whole graph are obtained by aggregating the messages of neighboring nodes, which is used as the input of the subsequent framework. Based on CMPNN, a framework combining two-stage Graph Attention Networks and Q-learning (TSGAT+Q-learning) is proposed in this paper. In the first stage, the agent embedding is completed, i.e., each service technician’s messages are represented by a constructed graph; In the second phase, each maintenance task selects an agent based on probability. In this way, the task assignment scheme is obtained, and finally Q-learning is used for further optimization. In addition, a key contribution is the proposal of a novel exponential reward, designed to speed up model training using REINFORCE in reinforcement learning. To validate the effectiveness of proposed method, scenarios with different scales are provided. In most cases, TSGAT+Q-learning outperforms CPLEX, OR-Tools and other learning-based algorithms. Moreover, the trained networks can also solve the problem with varying numbers of maintenance tasks, which implies that TSGAT+Q-learning has good generalization ability. Finally, the proposed method is also proved to be effective in solving on-site maintenance tasks scheduling problem with multiple constraints.</p> Graphical abstract <p>A Two-stage Graph Attention Networks and Q-learning Framework Based Maintenance Tasks Scheduling</p>