Abstract: Xili reservoir is one of the most important drinking water sources in Shenzhen. The water quality of the reservoir affects the water supply safety of the whole city. We aim to get timely and accurate water quality prediction results for formulating a scientific and reasonable water supply plan for the reservoir and water plant. Based on data decomposition using the Complete Ensemble Empirical Mode Decomposition with Adaptive Noise (CEEMDAN), we established a long- and short-term memory network model of water quality prediction for Xili Reservoir. Through extensive simulation and calculation, the model demonstrates excellent performance. The prediction results of total nitrogen, ammonia nitrogen and total phosphorus in the water quality prediction model are in good agreement with measured results. For total nitrogen and ammonia nitrogen, the relative prediction error of the model can be controlled below 10%. This highlights the model’s ability to effectively simulate the changing water quality in the reservoir and underscores the model’s rationality. The research findings serve as vital model and technical support for water quality prediction and the development of water supply plans for the Xili Reservoir.

Key words: water quality prediction, neural network, long-short term memory(LSTM), CEEMDAN decomposition, Xili Reservoir