<p>Dissolved oxygen (DO) levels are a critical indicator of water quality in aquaculture environments and are essential for the healthy development of fish. However, the dynamic nonlinear changes and complex interactions between multiple water quality parameters make accurate multivariate long-sequence DO prediction challenging. Previous long-sequence multivariate DO multi-step prediction models primarily relied on LSTM and GRU architectures. However, these models are limited by issues such as cross-dimensional dependencies, noise, and computational complexity, making accurate long-sequence DO prediction difficult. To address these issues, this paper proposes the PSG-Crossformer model, an innovative approach that integrates principal component analysis (PCA), Savitzky-Golay (SG) filter, and the Crossformer architecture. PCA reduces input dimensionality, alleviating computational complexity. SG filtering removes noise from the data, enhancing model accuracy. The Crossformer, featuring dimension-segment-wise (DSW) embedding, a two-stage attention (TSA) layer, and a hierarchical encoder-decoder (HED) structure, effectively captures cross-dimensional dependencies between variables. Experimental results show that the PSG-Crossformer model outperforms LSTM, GRU, Transformer, SVM, and other benchmark models in long-sequence DO prediction. Additionally, the PSG-Crossformer can predict DO changes up to 6&#xa0;h in advance and effectively handles seasonal variations, especially in extreme environments such as winter, where its prediction accuracy remains high, demonstrating outstanding adaptability and robustness. The proposed model provides more accurate predictions for water quality management and facilitates efficient aquaculture environment monitoring and timely interventions, thereby improving farming efficiency and reducing risks.</p>