基于SG-RF-MTBLSTM模型的湿地水质预测研究

Research on Wetland Water Quality Prediction Based on the SG-RF-MTBLSTM Model

摘要: 为解决湿地水质预测受潮汐气象等环境、参数耦合及极端天气干扰，本文提出SG-RF-MTBLSTM混合预测模型。对近岸湿地溶解氧(DO)与原位硝酸盐(NO₃)进行预测，两者受潮汐、气象与参数耦合共同影响：先对连续小时时序作Savitzky-Golay (SG)平滑，经随机森林(RF)得到精简协变量子集，再以共享BiLSTM骨干 + DO/NO₃双头，在24 h滑动窗口上同步估计窗末时刻的溶解氧与硝酸盐。基于USGS连续探头站2020~2025年小时数据，站点为USGS-02292900，按时间划分为2020~2022年训练，2023~2024年验证，2025年测试。结果表明，模型测试集DO的R²为0.80，NO₃的R²为0.69；以6~11月为飓风季分层后，飓风季DO的MAE由正常期0.59 mg/L升至0.94 mg/L、R²由0.74降至0.65，NO₃同步轻度下降，存在季节性分布外泛化困难。

Abstract: To address the challenges of coastal wetland water quality prediction under tidal-meteorological forcing, cross-parameter coupling, and extreme-weather disturbances, this paper proposes the SG-RF-MTBLSTM hybrid prediction model. The targets are nearshore dissolved oxygen (DO) and in-situ nitrate (NO₃), which are jointly driven by tidal, meteorological, and biogeochemical factors. Continuous hourly series are first smoothed with a Savitzky-Golay (SG) filter; a random forest (RF) then yields a compact covariate subset; and a shared bidirectional LSTM (BiLSTM) backbone with dual DO/NO₃ heads jointly estimates window-end DO and nitrate over a 24 h sliding window. Experiments use USGS continuous-probe hourly data for 2020~2025 at the primary site USGS-02292900, with a chronological split into training (2020~2022), validation (2023~2024), and testing (2025). On the test set, the model attains an R² of 0.80 for DO and 0.69 for NO₃. After a June-November hurricane-season stratification, hurricane-season DO MAE rises from 0.59 mg/L in the normal period to 0.94 mg/L and R² falls from 0.74 to 0.65, with a mild concurrent decline for NO₃, indicating seasonal out-of-distribution generalization difficulty.