Using tide for rainfall runoff simulation with feature projection and reversible instance normalization.

Area of Science:

• Hydrology
• Machine Learning
• Time Series Analysis

Background:

• Traditional runoff forecasting models often use complex architectures like Long Short-Term Memory (LSTM) and Transformers.
• These complex structures may not always be necessary for effective hydrological modeling.

Purpose of the Study:

• Introduce RR-TiDE, a novel, simpler model for runoff forecasting based on the Time Series Dense Encoder.
• Evaluate RR-TiDE's performance in multi-basin runoff simulation and prediction in data-sparse basins.
• Analyze the contributions of individual components, such as the feature projection layer and Reversible Instance Normalization (RevIN).

Main Methods:

• Developed RR-TiDE using a fully Multilayer Perceptron architecture, incorporating Reversible Instance Normalization (RevIN) to handle hydrological data non-stationarity.
• Trained and evaluated the model on the Catchment Attributes and Meteorology for Large-Sample Studies (CAMELS) dataset.
• Compared RR-TiDE against Transformer and LSTM-based models for 7-day runoff predictions and assessed its generalization capability in data-sparse watersheds.

Main Results:

• RR-TiDE outperformed Transformer and LSTM models in multi-basin runoff simulation across all metrics for 7-day predictions.
• Achieved a median Nash-Sutcliffe Efficiency (NSE) of 0.82 for 1-day runoff forecasting in 51 data-sparse watersheds, indicating robust spatial extrapolation capabilities.
• The feature projection layer significantly enhanced RR-TiDE's performance, while RevIN contributed to stabilizing training loss.

Conclusions:

• RR-TiDE offers a highly suitable and simpler alternative for rainfall-runoff simulation compared to more complex deep learning architectures.
• The model demonstrates strong generalization capabilities, enabling effective predictions even in watersheds with limited data.
• The findings suggest that simpler, well-designed models can achieve state-of-the-art performance in hydrological forecasting.