Combining POA-VMD for multi-machine learning methods to predict ammonia nitrogen in the largest freshwater lake in China (Poyang Lake)

Abstract

Nitrogen pollution, especially ammonia nitrogen (NH₃-N), poses a serious environmental threat to the ecological health of large freshwater lakes and the sustainable use of water resources. However, accurately predicting NH₃-N levels over a period of time remains a challenging task because of the complexity and dynamics of water quality data, which are influenced by various environmental factors. Traditional time series methods and stand-alone machine learning models are often limited in making accurate predictions for complex nonlinear data. To this end, we propose a novel hybrid prediction framework that combines the Pelican Optimization Algorithm-Variable Mode Decomposition (POA-VMD) with a variety of machine learning methods, including Random Forests (RF), Transformers, Long Short-Term Memory (LSTM), Bi-directional LSTMs, and Gated Recurrent Units. Among them, POA is applied to the hyperparameter optimization of VMD, which reduces the effects of redundant patterns and noise by automatically searching for optimal hyperparameter combinations. The results show that by combining POA-VMD, the prediction performance of all the machine learning models applied in this study is improved. The comparison reveals that POA-VMD-RF has the highest prediction accuracy with an R² of 0.9153, which is 26.4 % higher than the original RF model, while the POA-VMD-LSTM, POA-VMD-BiLSTM and POA-VMD-GRU hybrid models also have good prediction performances, with an R² of more than 0.8. The results highlight the potential of the proposed hybrid model for NH₃-N prediction in large freshwater lakes, which can provide important support for water quality monitoring and management.