基于ARIMA-BiGRU模型的水质分类预测

Water Quality Classification Prediction Based on ARIMA-BiGRU Model

摘要: 本文聚焦于地表水水质分类预测，选取2018年至2024年间秦皇岛市戴河口监测点的水质数据，通过结合传统时间序列模型(ARIMA)和双向门控循环单元模型(BiGRU)，系统评估了未来水质变化的预测性能。研究结果表明，两者的组合模型(ARIMA-BiGRU)进一步发挥了协同优势：先利用ARIMA提取序列的线性成分和周期性特征，再通过BiGRU对残差中的复杂非线性动态进行深度解析，从而实现对水质参数多层次时空关联的精准建模。相较于单一模型，ARIMA-BiGRU组合模型在短期和中期预测任务中均展现出更强的鲁棒性，对水质分类的预测误差显著降低。该模型不仅为水质预测提供了兼顾线性统计规律与深度特征挖掘的融合框架，其多尺度预测能力还可支持管理者制定从短期污染预警到长期生态治理的差异化决策方案，对提升水环境管理的科学性与前瞻性具有重要实践价值。

Abstract: This study focuses on the classification prediction of surface water quality, utilizing water quality data from the Daihekou monitoring site in Qinhuangdao City between 2018 and 2024. By integrating the traditional time series model (ARIMA) and the bidirectional gated recurrent unit model (BiGRU) into a combined model (ARIMA-BiGRU), the predictive performance for future water quality changes is systematically evaluated. The results demonstrate that the ARIMA model effectively captures linear trends and seasonal patterns in water quality indicators through its differencing operations and autoregressive mechanisms. Meanwhile, the BiGRU model enhances nonlinear prediction accuracy by employing bidirectional temporal modeling, which integrates both historical and potential future information. The combined ARIMA-BiGRU model further leverages synergistic advantages: it first utilizes ARIMA to extract linear components and periodic features from the sequence, then employs BiGRU to deeply analyze the complex nonlinear dynamics within the residuals, achieving precise modeling of multi-level spatiotemporal correlations in water quality parameters. Compared to individual models, the ARIMA-BiGRU combined model exhibits stronger robustness in short- and medium-term prediction tasks, significantly reducing prediction errors in water quality classification. This model not only provides a hybrid framework that balances linear statistical principles and deep feature extraction for water quality forecasting but also supports differentiated decision-making plans—from short-term pollution alerts to long-term ecological management—through its multi-scale predictive capability. This approach holds significant practical value for enhancing the scientific rigor and forward-looking perspective of water environment management.