A Hybrid Deep Learning Model for Water Quality Prediction

Water quality prediction is crucial for water environment management, playing a key role in preventing pollution and ensuring water resource safety. To improve prediction accuracy, this paper proposes a CEEMDAN-CNN-LSTM model enhanced by a self-attention mechanism. First, the Complete Ensemble Empirical Mode Decomposition with Adaptive Noise (CEEMDAN) algorithm decomposes the water quality time series into multiple Intrinsic Mode Functions (IMFs), reducing noise and addressing nonlinearities in the data. Next, Convolutional Neural Networks (CNN) extract spatial features from the decomposed IMFs, improving the model’s spatial learning capability. Then, a Long Short-Term Memory (LSTM) network models the temporal features, capturing long-term dependencies in the water quality data. Finally, a self-attention mechanism assigns varying importance to features at different time steps, improving the model’s focus on key information. Experimental results demonstrate that the proposed model outperforms traditional methods, achieving high accuracy across multiple water quality indicators. This model effectively addresses non-stationarity and noise in the data, demonstrating strong generalization and robustness, providing a novel and efficient approach to water quality prediction.