Page 72 - AJWEP-22-5

Jun, et al.

decomposition, and nutrient cycling. Hypoxic or anoxic and introducing an IEM to enhance optimization
conditions can cause aquatic organism mortality and stability and prediction robustness; and (iii) validating
even ecosystem collapse. Consequently, developing a the method’s superior performance on real-world
high-precision, real-time DO prediction method is of DO prediction tasks. The structure of the article is as
paramount importance for pollution monitoring, water follows: Section 1 presents the research background and
resource management, and early warning systems. 4 motivation (Introduction); Section 2 reviews relevant
In recent years, deep learning methods have gained literature and the current state of research (Literature
significant attention in environmental modeling, Review); Section 3 describes the materials and methods
demonstrating exceptional capabilities in capturing employed in the study (Materials and Methods); Section
nonlinear relationships, particularly within hydrological 4 details the experimental design and results (Results);
time series analysis. Among these, long short-term Section 5 discusses the findings (Discussion); Section 6
5,6
memory (LSTM) networks excel at capturing temporal explores practical application prospects of this research
dependencies and have been successfully applied in (Practical Applications); and Section 7 summarizes
various environmental forecasting scenarios. However, the paper and highlights key contributions and future
7
their training process is highly sensitive to optimization research directions (Conclusions).
algorithms. Traditional gradient-based optimizers
(e.g., stochastic gradient descent, Adam) often converge 2. Literature review
prematurely or become trapped in local optima when
applied to high-dimensional, noisy environmental In recent years, DO prediction has become a vital
datasets, thereby limiting model generalization and component of water quality management and ecosystem
prediction accuracy. 8 health monitoring. Researchers have extensively
To address these limitations, this study proposes a explored data-driven approaches, including support
novel hybrid optimization strategy combining Nesterov- vector machines (SVMs), artificial neural networks, and
accelerated Adaptive Moment Estimation (Nadam) fuzzy linear regression, to address the nonlinearity and
and differential evolution (DE) to enhance the training data scarcity inherent in aquatic systems. For example,
stability and global search capability of LSTM models for Ji et al. accurately predicted DO concentration in a
13
DO prediction tasks. Nadam enhances local convergence low-oxygen river system using an SVM model in the
by incorporating a momentum-based mechanism, Wenruitang River in China. Khan and Valeo developed
9
14
while DE, a population-based stochastic algorithm, a fuzzy regression framework to handle environmental
enhances global exploration through evolutionary uncertainties, outperforming conventional regression
and mutation mechanisms. A key innovation of this techniques by robustly managing sparse datasets and
10
study is the introduction of an information exchange systemic complexity. Wu et al. integrated sliding
15
mechanism (IEM), which enables the Nadam and DE window techniques with particle swarm optimization
sub-populations to share elite solutions during training, and backpropagation neural networks to enhance DO
thereby achieving a dynamic balance between local prediction accuracy. Malek et al. applied an SVM
16
exploitation and global exploration. model to analyze data from two large lakes in Malaysia,
Beyond improving DO prediction accuracy, using the forward selection method to optimize input
the proposed method offers broad applicability for parameters and achieving a prediction accuracy of 74%.
environmental modeling. By providing early signals Although traditional forecasting methods, such as
of pollution changes and ecological stress, accurate SVM, show high accuracy with nonlinear problems,
DO trend modeling can inform the optimization of they have limitations when handling long-term and
water treatment strategies, particularly in identifying complex time series data. Specifically, they require a
pathways influenced by metal, inorganic, and organic large number of feature selections and cannot adequately
pollutants. 11,12 Therefore, this study not only extends capture dynamic changes in time series. To overcome
the application boundaries of deep learning in complex these challenges, LSTM has been introduced for DO
time series but also provides a novel decision-support prediction in recent years. Its unique memory unit
tool for intelligent pollution control and sustainable structure effectively captures long-term dependencies
water resource management. in DO time series data, improving prediction accuracy.
In summary, the main contributions of this study For example, Tan et al. combined principal component
17
are: (i) proposing a hybrid Nadam –DE optimization analysis with LSTM, achieving a 2.71% reduction in
framework for training LSTM models; (ii) designing mean squared error (MSE) and a 9.03% improvement in

Volume 22 Issue 5 (2025) 66 doi: 10.36922/AJWEP025210165