Page 82 - AJWEP-22-5

Jun, et al.

higher accuracy and greater training stability in model optimized by Nadam–DE achieves an MSE of
complex, non-stationary time series forecasting tasks. 0.0193. In comparison, using the Nadam optimizer
alone yielded an MSE of 0.0369, representing an error
4.4. Model transferability verification reduction of approximately 47.8%. This highlights the
To further validate the applicability of the Nadam–DE significant advantages of the hybrid approach in terms
algorithm, this study conducted experiments using of both convergence speed and prediction accuracy.
an additional water quality dataset. This dataset was However, an in-depth analysis of the convergence
collected by Dr. J. Thad Scott in 2019 under the support process reveals the presence of oscillations during the
of the Tarrant Regional Water District, with samples local iteration phase, suggesting potential instability
sourced from Eagle Mountain Lake. The data have been in the information exchange frequency or strategy
publicly released on the Kaggle platform. The loss curve scheduling. Furthermore, although Nadam–DE has
during model training on this dataset is shown in Figure 6. shown strong performance in retrospective prediction,
After model training and evaluation, the final MSE its efficacy requires further validation in scenarios
reached 0.015074709430336952, indicating a minimal involving extended prediction lead times, extreme
error between the predicted and actual DO values, thereby pollution events, and cross-watershed generalization
demonstrating the model’s accuracy. The model not only capability.
exhibits strong predictive capability on the original dataset Overall, the outstanding performance of this
but also maintains robust generalization performance and method in water quality prediction tasks underscores
transferability under cross-source data conditions. This its significant practical application value. As a critical
verification significantly enhances the practical viability of environmental parameter for aquatic ecosystems, DO
the model for diverse aquatic environments and provides is directly related to a water body’s self-purification
a solid foundation for its deployment in real-world water capacity and ecological balance. Prediction models
quality monitoring and management tasks. optimized using Nadam–DE enable high-precision,
dynamic monitoring and forecasting of DO
5. Discussion concentrations. They also provide a scientific basis
for decision-making in pollution early warning, water
To enhance the predictive ability of DO in the LSTM body remediation, and resource allocation. This is of
model, this study introduced a hybrid optimization great importance for improving the efficiency of water
algorithm, Nadam–DE, which combines the strengths resource utilization, safeguarding ecological security,
of Nadam and DE. This method integrates the local and advancing intelligent environmental management.
optimization efficiency of Nadam with the global
exploration capability of DE, augmented by an 6. Practical applications
information-sharing mechanism to achieve balanced
and dynamic optimization. The hybrid-optimized LSTM model proposed in
Experimental results demonstrated that, under this study demonstrates significant advantages in
identical hyperparameter configurations, the LSTM enhancing the prediction accuracy and robustness of
DO levels, offering a novel technological approach for
intelligent water quality monitoring and water resource
management. This method is not only suitable for
dynamic water quality modeling in diverse environments,
such as river basins, lakes, and urban water bodies, but
also exhibits strong versatility and scalability, enabling
adaptation to environmental data characteristics across
different regions and pollution backgrounds. When
integrated with Internet of Things sensor networks, the
model shows promise for incorporation into real-time
water quality monitoring systems. It can facilitate rapid
detection and predictive early warning of key water
quality indicators, thereby providing technical support
Figure 6. Loss function changes on cross-source for emergency responses to water pollution incidents
dataset and for the optimization of treatment strategies.

Volume 22 Issue 5 (2025) 76 doi: 10.36922/AJWEP025210165