Sequence Reconstruction for River Water Level Anomaly Correction Using a Simplified Bidirectional LSTM Autoencoder

Accurate water level observations are essential for flood forecasting, hydrological analysis, and water resource management; however, sensor malfunctions and telemetry errors frequently introduce anomalous observations that compromise data quality. This study proposes a reconstruction-oriented Bidirectional Long Short-Term Memory (BiLSTM) Autoencoder for river water level anomaly correction and compares its performance with conventional first-, second-, and third-order polynomial and exponential regression models. The proposed framework incorporates a simplified encoder–decoder architecture, a dynamic block masking strategy to emulate contiguous sensor failures in highly autocorrelated water level series, and a threshold-based peak-oriented training scheme to improve reconstruction during high-flow events. Model hyperparameters were optimized using Gaussian process-based Bayesian optimization. The methodology was evaluated using hourly observed water level data from the Han River, Republic of Korea. Results showed that the proposed BiLSTM Autoencoder achieved reconstruction accuracy comparable to conventional regression models during calibration while exhibiting superior generalization to unseen validation datasets and better preserving the temporal continuity and dynamic characteristics of downstream hydrographs. Furthermore, a model calibrated using a relatively short but hydrologically representative period successfully reconstructed a substantially longer unseen record. Synthetic outlier injection experiments further demonstrated that reconstruction accuracy gradually deteriorated with increasing training data contamination, emphasizing the importance of high-quality training data for reliable sequence reconstruction. The proposed framework demonstrates potential as an effective sequence-reconstruction approach for offline river water level quality control.