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.
维普
万方数据