Abstract: Under the 'Emission Peak and Carbon Neutrality' goal and the new-type power system construction, the high penetration of new energy leads to a significant increase in the randomness and the complex and diverse operation modes of the power system. So, it is difficult for the traditional single-task deep reinforcement learning to adapt to the high randomness on both the sides of the source and the load, and for the dispatching decision to meet the needs of the new power system for the wind and solar energy consumption and the power balance. Therefore, this paper proposes a multi-task deep reinforcement learning optimal dispatching method based on grid operation scenario clustering. During the offline training, this method identifies the typical operation scenarios and their important characteristics of the massive operation data through the spatial clustering and the decision tree. Whar's more, a multi-layer perceptron classifier for discriminating the scenario categories is constructed. These scenario categories are adopted to establish the multi-task model for the multi-task division. Then, the differentiated training task models are designed from the data sources to the state and action spaces. During the online decision-making, we use the classifier to identify the scenario categories of the limited operation data. Then we intelligently choose the model to quickly solve the real-time dispatching tasks, and realize multi-task fast migration learning in the highly random scenarios, which ensures the optimality of the power system optimal dispatching decisions. The experimental results show that the multi-task deep reinforcement learning optimal dispatching algorithm based on grid operation scenario clustering is able to significantly improve the economic benefits of the dispatching decisions compared with the traditional single-task algorithm.