The real-time strategy (RTS) game has become a suitable testbed for artificial intelligence (AI) technologies owing to the large state space, limited decision times, partial observation environment and adversarial behavior involved. To address these challenges, the approach named AHTN which combined hierarchical task network (HTN) with game tree search has been applied in RTS games and achieved favorable performance. However, both the opponent strategy and time constraint of generating the sub-game tree are not considered by AHTN during planning. Therefore, this paper proposes a novel method which considers the time constraint when generating the subgame tree by HTN planning and introduces the opponent strategy as the adversarial player. The empirical results are presented based on a $\mu$RTS game and show our algorithm performs better than original HTN planning methods.