Parallel xLSTM-informer model for long-term dissolved oxygen prediction using STL decomposition and Bayesian optimization

Dissolved oxygen (DO) is a vital indicator of water quality in surface water ecosystems. Existing prediction models are limited in their ability to comprehensively extract features from water quality data. To address this limitation, this study proposes a parallel STL-BO-xLSTM-Informer model, in which the STL module captures seasonal and trend components, the Informer module captures long-term temporal dependencies, and the xLSTM module extracts local temporal features and fluctuation patterns. Four-hourly water quality data from six sampling sites (S1-S6) in the Chongqing section of the Yangtze River, spanning January 2021 to December 2023, were used as the dataset. The mode was validated across S1-S6 through comparisons with baseline models (iTransformer, Autoformer, Crossformer, TimesNet, and Informer) and through ablation experiments using at S6 using the xLSTM-Informer as the base model, with ablation variants including STL-BO-Base, BO-Base, and STL-Base. The results demonstrated that the STL-BO-xLSTM-Informer model effectively captured the intrinsic features and spatiotemporal characteristics of water quality data. Compared with the baseline models, MAE, RMSE, and SMAPE decreased by 27.54%, 21.55%, and 27.27%, respectively, and R 2 increased by 6.6%. The STL-BO-xLSTM-Informer model outperformed all baseline models in prediction accuracy, with performance at S6 ranked as follows: STL-BO-xLSTM-informer > Informer > iTransformer > TimesNet > Crossformer > Autoformer.