Accurate water quality prediction is paramount for the sustainable management of surface water resources. Current deep learning models face challenges in reliably forecasting water quality due to the non-stationarity of environmental conditions and the intricate interactions among various environmental factors. This study introduces a novel, multi-level coupled machine learning framework that integrates data denoising, feature selection, and Long Short-Term Memory (LSTM) networks to enhance predictive accuracy. The findings demonstrate that the LSTM model incorporates data denoising pre-processing, capturing non-stationary water quality patterns more effectively than the baseline model, enhancing prediction performance (R2 increased by 1.01%). The most adept model with wavelet transform exhibited superior adaptability and predictability, achieving a modest but statistically significant increase in R2 values of 0.81% and 0.51% relative to incorporate moving average and complete ensemble empirical mode decomposition with adaptive noise techniques, respectively. The integrated models varied in their suitability for time series characterized by different patterns of variability (stability vs. instability, periodicity vs. non-periodicity). We conducted multi-step ahead predictions (t+1 and t+3 days) and employed two training configurations (80-20% and 70-30% splits) for dissolved oxygen and the permanganate index across four monitoring stations within the world’s largest long-distance inter-basin water diversion project, to assess the reliability and robustness of the proposed water quality prediction models under varying conditions. The integration of data denoising techniques with LSTM networks substantially improves the prediction of dynamic water quality indices in complex environmental settings. Future research should explore the scalability of this framework across different geographical and climatic conditions to further validate its effectiveness and utility in global water resource management.
Water resources are a pillar of human societal functions and the sustainability of ecological systems, yet the quality of surface waters has been persistently declining under the dual influence of anthropogenic pressures and natural occurrences (Paller et al., 2019; Wu et al., 2021). Timely monitoring and precise water quality predictions are paramount for advancing sustainable water resource management. Despite this, the current automatic water quality monitoring station network lacks the predictive capabilities necessary to address this critical challenge (Lu and Ma, 2020). The development and implementation of robust and reliable models for time-series prediction in water quality are thus essential, as they can significantly aid relevant departments and managers in devising strategies to mitigate water quality degradation, while also optimizing monitoring regimens and enhancing the efficiency of resource utilization (Li et al., 2018).
Machine learning (ML) and deep learning (DL) models, renowned for the significant increase in available data due to improved monitoring techniques, have emerged as the most favored data-driven modeling algorithms in recent years, representing a cutting-edge approach to water quality dynamic prediction (Lau et al., 2019; Reichstein et al., 2019). Process-driven models necessitate a comprehensive understanding of the physical and chemical relationships between water quality parameters, and the collection of detailed geographic and environmental data, these models are characterized by their data-intensive and time-consuming nature (Banerjee et al., 2019). Data-driven models, based on statistical methods and machine learning theory, offer not only faster computational speeds but also the ability to effectively establish mapping relationships between target variables and multiple factors, thereby capturing the dynamic patterns of the predicted objects (Ghorbani et al., 2021; Hunter et al., 2018). Artificial neural networks, recurrent neural networks, and Long Short-Term Memory (LSTM) networks are frequently utilized for nonlinear time-series prediction (Farah et al., 2022). However, both artificial neural networks and recurrent neural networks, which are deep learning models consisting of three fully connected neural network layers, are limited by two main issues (Ghorbani et al., 2021; Jeong et al., 2021; Memarzadeh and Keynia, 2020). The first is their lack of interpretability, which precludes the provision of the importance of input features within the models. Secondly, fully connected neural networks are ill-equipped to capture the characteristics of time-series data (LeCun et al., 2015). Consequently, when employing these two neural network types to predict water quality, the inherent variability and nonlinearity of time-series data pose significant challenges (Li et al., 2020). Particularly, as a specialized variant of recurrent neural networks, LSTM networks are adept at handling time-series data due to their unique gate structure, effectively resolving the gradient vanishing problem (Gers et al., 2003; Hochreiter and Schmidhuber, 1997; Zhang et al., 2022b). Prior studies have empirically demonstrated LSTM’s exceptional performance in time series prediction, the LSTM achieved remarkable prediction accuracy (R2 = 0.9976), outperforming conventional ARIMA models (R2 = 0.0017) (Elsheikh et al., 2021). This significant performance gap substantiates LSTM’s superior capability in capturing complex temporal dependencies through its gated memory units.
Conclusion
In this study, we introduce a multi-tiered machine learning framework that integrates Long Short-Term Memory (LSTM) neural networks with time-honored data denoising and feature selection techniques to analyze multivariate water quality datasets exhibiting both periodic return and jitter-prone characteristics. The framework is applied to the world’s longest inter-basin water diversion project, the Middle Route of the South-to-North Water Diversion Project in China, using dissolved oxygen.
