To enhance the ability of hydro-wind-solar complementary systems to cope with the uncertainty of wind and solar power output, a peak-shaving scheduling method integrating scenario adaptive discrimination and multi-strategy deep reinforcement learning for hydro-wind-solar complementary systems is proposed, based on typical scenario division and strategy adaptation. In the offline training phase, principal component analysis and the K-means clustering algorithm are used to cluster the morphological characteristics of daily historical wind and solar power output curves at the seasonal scale, identifying and constructing typical scenario datasets with different output patterns. Based on these datasets and under unified objective functions and constraints, simulation environments are independently built for each scenario to train scenario-specific deep reinforcement learning agents, thereby constructing a strategy knowledge base. In the online scheduling phase, a BiLSTM-Transformer discriminator is designed, which takes current seasonal features and standardized wind and solar power forecast sequences for the next day as input, outputs the scenario membership probability distribution and labels, then matches and calls the corresponding agent from the strategy knowledge base to generate scheduling instructions. Feasible region projection is applied to ensure constraint satisfaction, forming a dynamic mechanism of “scenario identification-strategy matching-rolling scheduling”. Case verification results show that, under the premise of satisfying all engineering constraints, the proposed method significantly reduces the standard deviation of load fluctuations and the peak-valley difference, with decision response time at the millisecond level. The scenario discrimination accuracy reaches 97.54%, and the proposed method exhibits strong cross-scenario adaptability and robustness. Keywords: hydro-wind-solar complementary; multi-strategy optimization; deep reinforcement learning; peak-shaving operation 〖FL