Basin-scale hydrological time series processing and modeling are crucial tools for understanding and managing water resources within a watershed, supporting decision-making processes that protect lives, property, and the environment. While artificial intelligence (AI) methods demonstrate advantages over traditional models in terms of accuracy and computational efficiency in specific scenarios, there are few significant breakthroughs in cross-task versatility and adaptability. Facing these bottlenecks, we leverage a novel SAITS neural network, use large-sample datasets from several countries, and design masking-denoising strategies to pre-train a more comprehensive encoder for diverse hydrological applications. Using the USA catchments as the study area, SAITS trained under this framework is tested on 39 hydrometeorological variables for imputation, regression, and forecasting objectives. The results indicate that our model consistently achieves considerable accuracy in most cases, reflecting its high generalization capabilities and robustness. Furthermore, the pre-trained model exhibits strong post-learning capabilities, where fine-tuning with a small amount of local data leads to significant improvements, with accuracy rivaling or even surpassing that of classical neural networks. These preliminary achievements suggest that pre-training encoders with more trainable parameters is a feasible and effective way to learn the underlying relationships between hydrometeorological variables, and this methodology contributes to the advancement of more general and practical AI models in the field of hydrological modeling, extending beyond the limitations of specific regions and objectives.