<p>Due to the challenge of high energy consumption and high delay in the current networks, we propose a new multi-objective multicast optimization algorithm based on deep reinforcement learning. We design a three-layer BP neural network to train Q-network by continuously interaction learning with the environment. We optimize neural network weights of the deep Q-network by using Northern Goshawk Optimization algorithm. We use Huber loss function to improve model prediction accuracy and network model robustness. Above measures not only improve the quality of learned policies, but also enable the network to better capture the feature and state information of the environment. Then we leverage the Q-network to seek path with low network energy consumption and delay for transmission demand. Simulation result shows that compared with state-of-the-art algorithms, this new algorithm can effectively reduce network energy consumption and delay. Furthermore, the advantage of its performance becomes more apparent as the network size increases.</p>