An Attention-Driven and Autoencoder-Based Bidirectional LSTM for Long Interval Gap-Filling of a Water Treatment Process Data Set

Water quality monitoring with distributed water quality probes is crucial for environmental protection and public health. But missing data in time series poses a significant challenge in environmental management. This paper introduces a novel long interval gap-filling method using a Bidirectional Long Short-Term Memory Autoencoder with multi-head Attention Mechanism (BiLSTM-AAM). The multi-head attention mechanism allows the model to focus on informative regions within the data, even with extensive gaps while the autoencoder module aids in dimensionality reduction. The efficacy of BiLSTM-AAM was evaluated by a big real-world water quality dataset and further tested in other environmental contexts, demonstrating superior performance compared to three other similar models. The results highlight the model's ability to accurately reconstruct missing values, capture complex dependencies, and restore data integrity even with high percentages of missing data. Our findings indicate significant implications for pattern analysis and machine intelligence, ultimately contributing to better monitoring for water process engineering and beyond.