Abstract: Due to the heavy workload and high repeatability of power adjustment for key transmission cross-sections of complex bulk power grids, the computational speed is difficult to meet the requirements of online assistant decision-making. Based on the deep learning theory, a method is proposed to perform feature self-learning on the cross-section power adjustment data to realize the fast automatic adjustment of the cross-section power of bulk power grids. First, the reverse equivalent matching method is used to simulate the manual adjustment operation, and the massive data set required by deep learning is constructed. Then, under the constraints of unit sensitivity and adjustment amount, the effective set of units participating in the power adjustment in the complex bulk power grid is selected. On this basis, an optimal regression model is built with the determination coefficient as the index to accurately predict the output value of the adjusting unit, thereby realizing the fast automatic adjustment of the cross-section power.Finally, the proposed method is validated by taking the inter-provincial cross-section power adjustment in a practical regional bulk power grid in China an example. Simulation results show that the determination coefficient of the optimal model and the success rate of the cross-section power adjustment are both relatively ideal, which greatly shortens the cross-section power adjustment time, and the adjustment efficiency is not affected by the system operation mode and the difference between the actual cross-section power and the target power.