Abstract: Emergency observation mission scheduling is a complex problem of combinatorial optimization with strong timeliness as the scheduling algorithm must complete the computation within the required time limit. Using machine learning methods to provide high-quality initial solutions for scheduling algorithms can effectively simplify the calculation process. For this reason, this paper proposes a multi-satellite scheduling approach for emergency scenarios based on hierarchical forecasting with Transformer network, which decomposes the scheduling into three steps. Firstly, using the Transformer-based task schedulability prediction model to predict whether the observation task will be executed or not, so as to obtain a set of tasks to be executed. After that, using the Transformer-based task allocation model to allocate the satellite to the task set, so as to obtain the initial scheduling scheme. Finally, a constraint modification algorithm based on random hill climbing is used to optimize the initial scheme, so as to obtain feasible scheduling schemes. To verify the effectiveness of the proposed method, simulation experiments are conducted, and the results are compared with those by CPLEX Optimization, standard genetic algorithm, Long Short-Term Memory and other methods. The simulation results show that the method proposed consumes a short calculation time and has high benefits, and is thus suitable for multi-satellite scheduling of emergency observation missions.

Key words: emergency observation mission, hierarchical forecasting, task schedulability, task allocation, Transformer