<p>Water environment quality monitoring systems play a pivotal role in ensuring drinking water safety and maintaining ecological balance. Nevertheless, it encounters significant challenges in managing both chronic and acute water pollution events, primarily due to the absence of specialized sensing equipment, low monitoring frequency, and high operational costs. To address these issues, this study proposes a water quality time series prediction model based on a hybrid Long Short-Term Memory (LSTM)-Transformer architecture. This hybrid model integrates LSTM’s capability to capture temporal features with Transformer’s proficiency in global information modeling, thereby enhancing the accuracy and generalization of water quality predictions. Through comprehensive data preprocessing and hyperparameter optimization, the model’s performance was systematically evaluated across short-term to long-term prediction tasks. Experimental results indicate that compared to a single LSTM model, the LSTM-Transformer model achieves improvements in key evaluation metrics such as Nash–Sutcliffe Efficiency and Root Mean Square Error. The hybrid model achieves high predictive accuracy for critical water quality indicators, namely pH value, dissolved oxygen, and water temperature. Specifically, the proportions of samples with a relative error of ≤ 5% are 99.99%, 88.40%, and 87.81%, respectively. These findings demonstrate high predictive accuracy and stability, with pH showing the highest precision. The model also excels in classifying total phosphorus and total nitrogen, though there remains room for improvement in predicting complex indicators such as ammonia nitrogen. Overall, the trained LSTM-Transformer model exhibits superior predictive performance and broad application potential, indicating its promise for dynamic water quality monitoring and pollution early warning.</p>