COMPARATIVE ANALYSIS OF HYBRID LSTM MODELS FOR MULTIVARIATE WATER QUALITY PREDICTION

Abstract

Accurate multivariate water-quality forecasting is essential for environmental monitoring and early-warning systems. However, the comparative performance of different Long Short-Term Memory (LSTM)-based architectures across water-quality parameters with different temporal characteristics remains insufficiently established. This study quantitatively compares five deep-learning architectures: standard LSTM, Bidirectional LSTM (Bi-LSTM), LSTM-GRU, CNN-LSTM, and Attention-based LSTM (AT-LSTM). The analysis uses 30,894 observations of seven water-quality parameters—chlorophyll, temperature, dissolved oxygen, pH, salinity, specific conductance, and turbidity—recorded at 30-minute intervals in the Brisbane River, Australia, from 4 August 2023 to 27 June 2024. We preprocessed the time-series data through missing-value imputation, Min–Max normalization, and sliding-window sequence construction. We divided the dataset into training and testing subsets using an 80:20 split, and trained the models using the Adam optimizer, mean squared error loss, and early stopping. We evaluated predictive performance using mean absolute error (MAE), mean squared error (MSE), root mean squared error (RMSE), and the coefficient of determination (R²). The results indicate that AT-LSTM achieved the best overall performance for six of the seven parameters, including temperature (R² = 0.998) and turbidity (R² = 0.636), while Bi-LSTM performed best for specific conductance (R² = 0.967). Chlorophyll and turbidity remained challenging because of their irregular fluctuations and extreme events. These findings demonstrate that model performance varies across water-quality parameters and that attention mechanisms can improve forecasting accuracy under the evaluated experimental conditions.

Keywords

• Attention Mechanism
• Deep Learning
• Long Short-Term Memory
• Multivariate Time-Series Forecasting
• Water Quality Prediction