Bayesian model averaging by combining deep learning models to improve lake water level prediction.

Area of Science:

• Hydrology and Climate Science
• Environmental Monitoring
• Water Resource Management

Background:

• Lake water level (WL) is a critical indicator of climate change impacts.
• Fluctuations in lake WL affect water supply security and ecosystem stability.
• Accurate WL prediction is essential for effective water resource management and eco-environmental protection.

Purpose of the Study:

• To evaluate and enhance the accuracy of deep learning models for lake water level prediction.
• To compare the performance of Long Short-Term Memory (LSTM), Gated Recurrent Unit (GRU), and Temporal Convolutional Network (TCN) models.
• To improve prediction accuracy and quantify uncertainty using Bayesian Model Averaging (BMA) and Monte Carlo sampling.

Main Methods:

• Applied three deep learning models: LSTM, GRU, and TCN.
• Utilized Bayesian Model Averaging (BMA) to synthesize forecasts from individual DL models.
• Employed Monte Carlo sampling to calculate 90% confidence intervals for uncertainty analysis.

Main Results:

• All three deep learning models demonstrated satisfactory prediction accuracy for lake WL.
• GRU generally outperformed TCN and LSTM across most forecast scenarios.
• BMA further improved prediction accuracy (NSE and R² metrics) in 80% of scenarios, consistently ranking highly.
• Uncertainty analysis showed high Containing Ration (CR > 84%) and reliable Relative Bandwidth (RB) for 7-day ahead predictions.

Conclusions:

• The combined BMA approach offers a robust framework for accurate lake WL forecasting, surpassing individual deep learning models.
• The proposed method simplifies model selection while enhancing predictive performance and providing reliable uncertainty estimates.
• This framework is adaptable for predicting other hydrological variables, supporting broader water resource management applications.