基于GCN和TPA混合模型的青草沙水库叶绿素a浓度预测方法

Prediction Method of Chlorophyll-a Concentration of Qingcaosha Reservoir Based on the Hybrid Model of GCN and TPA

青草沙水库是上海重要饮用水水源地,面临较大藻华暴发风险,实现对水库叶绿素a(Chl-a)浓度的准确预测,对于饮用水安全保障至关重要。文章以青草沙水库为对象,提出了一种图卷积时间模式注意力网络混合模型(GC-TPA),首先利用时间模式注意力(TPA)机制学习水质数据的时间依赖性,其次使用图卷积网络(GCN)学习不同水质参数之间的关系。另外,为进一步提高模型的预测精度,引入完全自适应噪声集合经验模态分解(CEEMDAN)以降低模型的滞后性,同时使用多层感知机(MLP)学习Chl-a浓度的突变。结果表明:(1)引入的GCN模块显著增强了TPA对Chl-a的预测能力,结合CEEMDAN和MLP的帮助,模型性能进一步提升,以纳什效率系数作为评价指标,混合模型的24 h预测精度较单独TPA提升了56.5%;(2)与单独TPA和长短期记忆网络(LSTM)的对比试验表明,在更长的预测周期(48 h)上,GC-TPA虽然精度下降,但仍表现最好,48 h预测平均绝对误差和均方误差比LSTM低25.5%和24.0%,比TPA低4.92%和8.40%;(3)GCN模块与MLP模块在结果预测中发挥了不同的作用,GCN模块增强了TPA的特征学习能力,提高了模型对Chl-a浓度变化趋势的预测精度,而MLP则对Chl-a的突变较为敏感。研究所提出的GC-TPA混合模型在青草沙水库Chl-a浓度短期预测中表现良好,可为水库水质管理提供支撑。

The Qingcaosha Reservoir in Shanghai is a crucial drinking water source facing the risk of eutrophication.Accurate prediction of chlorophyll-a(Chl-a)concentration is essential for ensuring water safety.This study proposed a graph convolutional temporal pattern attention network(GC-TPA)to predict Chl-a concentration in the reservoir.The model first utilized the temporal pattern attention(TPA)mechanism to capture the temporal dependence of water quality data.It then employed a graph convolutional network(GCN)to learn the relationships between different water quality parameters.In addition,to further improve the prediction accuracy,complete ensemble empirical mode decomposition with adaptive noise(CEEMDAN)was introduced to reduce the model′s lag,and a multi-layer perceptron(MLP)was used to learn the mutation of Chl-a.The results demonstrated that(1)the GCN module significantly enhanced the Chl-a prediction ability of TPA,and the combination with CEEMDAN and MLP further improved the model performance,with a 56.5%increase in 24-hour prediction accuracy compared to TPA;(2)on a longer prediction period(48 hours),GC-TPA still outperformed the TPA and long short-term memory(LSTM)models,with 25.5%and 24.0%lower average absolute and mean square errors,respectively,than LSTM,with 4.92%and 8.40%lower than TPA;and(3)the GCN module improved the feature learning ability of TPA and enhanced the prediction accuracy on the Chl-a trend,while the MLP module was more sensitive to Chl-a mutations.The proposed GC-TPA model performed well when predict of Chl-a concentration in Qingcaosha Reservoir,providing a viable approach for water quality management and ensuring the safety of drinking water.