In an article published in the journal Nature, researchers introduced and evaluated four metaheuristic algorithms—teaching–learning-based optimization (TLBO), sine cosine algorithm (SCA), water cycle algorithm (WCA), and electromagnetic field optimization (EFO)—integrated with multi-layer perceptron neural network (MLPNN) for predicting dissolved oxygen (DO) concentration in the Klamath River. These algorithms optimized the MLPNN computational variables, enhancing the accuracy of DO prediction in water quality assessment.
Study: Enhancing Dissolved Oxygen Prediction in Rivers Using Metaheuristic Algorithms and Neural Networks. Image credit: Andrii Medvediuk/Shutterstock
Background
Water quality, crucial for ecosystem health, relies on parameters like DO. Measuring DO is challenging due to various factors, and predictive models offer solutions. Previous research extensively used machine learning models like Artificial Neural Networks (ANNs) and Support Vector Machines (SVMs) for DO prediction, demonstrating varying degrees of accuracy. However, optimization techniques, particularly metaheuristic algorithms, have gained attention for refining predictive models.
The present study employed four metaheuristic algorithms to enhance the MLPNN's predictive capacity for DO in the Klamath River. The US Geological Survey's data, particularly from the Klamath River, was a valuable resource for such studies. The dynamic nature of DO concentration, influenced by factors like water discharge, temperature, pH, and specific conductance, makes accurate prediction challenging. The proposed metaheuristic algorithms (TLBO, SCA, WCA, and EFO) addressed this challenge by optimizing the MLPNN's computational variables.
Comparative studies have shown the effectiveness of various models, such as adaptive neuro-fuzzy inference systems (ANFIS), SVM, and Long Short-Term Memory (LSTM), in DO prediction. However, gaps remain in optimizing these models for better accuracy and efficiency. This study introduced the TLBO, SCA, WCA, and EFO algorithms to fill these gaps, emphasizing their role in achieving fast, inexpensive, and reliable DO predictions. The case study on the Klamath River contributed valuable insights and practical monolithic formulas for efficient DO concentration prediction, enhancing water quality monitoring.
Methodology
The research revolved around predicting dissolved oxygen (DO) levels in the Klamath River by employing four distinct metaheuristic algorithms—TLBO, SCA, WCA, and EFO. The methodological flow included dataset creation, model development, and accuracy assessment. Notably, these metaheuristic algorithms, like TLBO with its simulated class dynamics and EFO relying on electromagnetics-based search, deviated from conventional training mechanisms.
The evaluation of accuracy involved metrics such as mean absolute error (MAE), root mean square error (RMSE), Pearson correlation coefficient (RP), and Nash–Sutcliffe efficiency (NSE) coefficient. The ultimate goal was to optimize the prediction of DO levels, contributing to effective water quality monitoring in the Klamath River. By integrating these metaheuristic algorithms with MLPNN, the research aimed to provide insights into the development of efficient and reliable predictive models for assessing water quality in dynamic and changing environmental conditions.
Results and Discussion
The WCA showed the highest quality in training, while the EFO achieved the best results in testing, with the smallest RMSE, largest RP, and highest NSE. The TLBO, with the smallest MAE in testing, was surpassed by EFO in other metrics. The SCA performed the poorest in both training and testing.
Comparative analyses, including boxplots and Taylor Diagrams, consistently highlighted the higher accuracy of the WCA-MLPNN, EFO-MLPNN, and TLBO-MLPNN compared to SCA-MLPNN.
The efficiency comparison indicated that, despite requiring more iterations, the EFO achieved optimization in a significantly shorter time compared to TLBO, SCA, and WCA. Overall, the study demonstrated considerable improvements in DO prediction compared to benchmark models and suggests avenues for future research, including updating models with recent hydrological data and exploring additional metaheuristic algorithms.
Conclusion
In conclusion, this study harnessed the capabilities of four stochastic search strategies—TLBO, SCA, WCA, and EFO—to train an artificial neural network for predicting dissolved oxygen levels in the Klamath River, Oregon, US. Precision-tuned parameters for each algorithm facilitated effective training of the MLP, enabling accurate predictions of dissolved oxygen for new environmental conditions. The comparison of hybrid models, considering accuracy, complexity, and computation time, revealed noteworthy findings.
Despite requiring 30 times more iterations, the electromagnetic field optimization (EFO) algorithm showcased superior efficiency, achieving faster optimization and delivering the most accurate results in the testing phase, as indicated by lower RMSE, higher RP, and superior NSE. Additionally, EFO demonstrated efficacy with fewer search agents, enhancing its practical appeal. The water cycle algorithm (WCA)-based MLP emerged as the second-most efficient model, further diversifying the predictive capabilities.
Consequently, two robust dissolved oxygen predictive models, fine-tuned by the WCA and EFO algorithms, were proposed. The study's outstanding models outperformed various hybrid and conventional models from prior research, marking a notable advancement in practical dissolved oxygen predictions. These findings hold promise for addressing water quality challenges in the studied region.
Machine Learning Enhances Water Quality Monitoring