Ocean observation data forecasting is crucial for environmental monitoring and climate research. Although pre-training has shown great potential in enhancing model forecasting capabilities and alleviating data scarcity, existing pre-training methods struggle with multivariate forecasting due to the inconsistency in the number and types of variables observed at different marine monitoring stations. To address this issue, we propose a framework, Seaformer, which consists of two key stages: univariate pre-training and multivariate fine-tuning. In the pre-training stage, the model learns general temporal patterns from large-scale univariate data, providing strong prior knowledge for downstream tasks. During the fine-tuning stage, the model leverages a dynamic channel encoding layer and a sparse dependency graph structure to flexibly adapt to varying input dimensions and effectively capture key inter-variable dependencies. Experimental results demonstrate that the proposed method achieves superior forecasting accuracy and strong cross-site adaptability under both sufficient and limited data conditions.