1.College of Electrical and Information Engineering, Hunan University, Changsha 410082, China;2.Economic Research Institute of State Grid Sichuan Electric Power Company, Chengdu 610041, China;3.Global Energy Interconnection Group Co., Ltd., Beijing 100031, China

Current data-driven maximum frequency deviation prediction methods for power systems ignore the effect of uneven sample distribution on the maximum frequency deviation, and its prediction accuracy can be further improved. To this end, this paper proposes a maximum frequency deviation prediction method considering the frequency deviation distribution and penalty cost. To avoid the dimensional explosion problem, the multi-source information fusion method is used to extract key input feature subsets from power system operation information. Then, a cascaded light gradient boosting machine (CasLightGBM) is constructed, and a punishment sensitivity mechanism is embedded into its loss function, which helps CasLightGBM automatically correct the sample loss value in training according to the probability distribution of frequency deviation samples and the penalty cost of conservative prediction. Simulation results based on the IEEE 118-bus test system show that the proposed method has excellent prediction accuracy and anti-noise performance.