Abstract: Using a single model to handle each regulation task in isolation can only one-sided capture the local characteristics of resources, and the efficiency of the new power system distributed resource aggregation regulation is low. Therefore, a new research on distributed resource aggregation and regulation of power systems based on multi task learning is proposed. Decompose the task of aggregating and regulating distributed resources in the new power system into four sub tasks: state prediction, feasible region calculation, multi-objective optimization regulation, and market mechanism interaction, and clarify their interrelationships. Using a hard parameter sharing architecture combined with LSTM and task specific networks, a distributed resource dynamic model based on multi task learning is constructed through a joint training strategy. Based on multi-objective optimization problems and non dominated sorting genetic algorithm to solve Pareto front, combined with model predictive control to achieve 15 min rolling optimization control instructions, a new type of distributed resource aggregation control for power systems is realized. The experiment shows that the research method can effectively achieve the aggregation and regulation of distributed resources in the new power system, significantly improving the regulation speed under different resource scales and enhancing the efficiency of the aggregation and regulation of distributed resources in the new power system.