Abstract: In the context of the growing demand for observing a vast number of variable celestial objects detected by sky survey equipment every day, the long sequence task planning problem consisting of high-frequency dynamic Target of Opportunity (ToO) events and follow-up observations has the characteristics of observation event randomness, high timeliness of data acquisition, multiple selectable options, and complex constraints, often considered an NP-hard problem. Consequently, obtaining labeled data for supervised learning is challenging. When applying unsupervised learning through deep reinforcement learning (DRL) methods to solve the long-sequence task planning problem, satellites as agents find it difficult to quickly converge to a global optimal strategy. To address this, this paper draws on the concept of local attention (LA) to improve the pointer network (PN), proposing the Local Attention Pointer Network (LA-PN) algorithm. This algorithm introduces a time window to focus the model on the crucial sequence parts for the current decision, reducing ineffective exploration. Simulation results demonstrate the algorithm's profitability, real-time performance, and generalization ability.