<p>With the continuous progress of the Internet of Things (IoT) technology, smart grid has become the key to improve the efficiency of the power system. At present, how to reasonably distribute the massive power computing tasks under the premise of guaranteeing the quality of service is an urgent problem of the current smart grid. Therefore, this paper focuses on the smart grid computation task offloading decision problem based on the collaboration of edge computing and cloud computing. An end-edge-cloud task scheduling model with task execution delay, energy consumption and load balance as optimization objectives is proposed. In addition, the Dyna-Actor Critic (Dyna-AC) algorithm is proposed using Deep Reinforcement Learning (DRL), which integrates the advantages of model-based and model-free learning and optimizes the performance of this algorithm through heuristic sample selection and asynchronous parameter update strategies. Simulation results show that the algorithm is very effective in reducing power business execution delay, terminal energy consumption and improving system load.</p>