River water level prediction plays a crucial role in preventing flood disasters, safeguarding lives and property, and supporting water resource management and ecological environment protection. This study focuses on innovatively designing a multi-module collaborative water level prediction model to enhance the accuracy and adaptability of river water level prediction.Based on the LSTM-Seq2Seq model, this paper attempts to optimize and improve the model by introducing certain mechanisms and modules, ultimately obtaining the MultiMod-Seq2Seq model. Firstly, the attention mechanism is introduced to enhance the efficiency of extracting key features. Then, the Auto-Correlation module is integrated to analyze the self-correlation characteristics of the sequence. Subsequently, the ATFNet module is utilized to integrate time and frequency domain feature information and fully explore the characteristics of water level data at different frequencies. The Decomposition module is also employed to decompose the data into periodic and trend components, aiming to enhance the model’s adaptability and prediction accuracy in response to complex water level changes. To systematically evaluate the model’s performance, this study selects three datasets with diverse features for comparative experiments, all in the form of time series, including strongly periodic, non-periodic, and mixed datasets. The results show that the model performs well in strongly periodic datasets. In the experiment on the river flow of a certain basin in the Yangtze River using GRDC global runoff data, the MAE and RMSE are reduced by 33.9% and 21.8% respectively compared to the LSTM-Seq2Seq model, and the PeriodScore is increased by 30.0%. The model also performs well in mixed datasets, with better performance than traditional and emerging models. The MAE and RMSE are reduced by 21.6% and 21.1% respectively compared to the LSTM-Seq2Seq model, and the PeriodScore is increased by 30.9%. In the ablation experiment, the water level data of a certain node in the Yangtze River Basin’s hydrological monitoring station is used, with missing values filled by mean imputation. This helps to explore the contribution of each module to the performance improvement of the LSTM-Seq2Seq model.