Physics-Guided Long Short-Term Memory Network for Streamflow and Flood Simulations in the Lancang–Mekong River Basin

Binxiao Liu, Qiuhong Tang, Gang Zhao, Liang Gao, Chaopeng Shen, Baoxiang PanWater 2022 · DOI

TL;DR

This paper introduces a novel hybrid physics-data (HPD) model that combines a process-based hydrological model (VIC-CaMa-Flood) with a Long Short-Term Memory (LSTM) neural network. This approach significantly improves streamflow and flood predictions in the Lancang-Mekong River Basin compared to using either model alone, offering a more robust way to handle complex hydrological systems.

What Problem Does This Paper Address?

Accurate streamflow and flood forecasting are critical for managing water resources and mitigating disaster risks, especially in regions like the Lancang-Mekong River Basin (LMRB) which are vulnerable to climate change impacts. Traditional hydrological modeling approaches fall into two main categories: process-based models that explicitly simulate physical processes but can suffer from parametric and structural uncertainties, and data-driven models (like LSTMs) that excel at pattern recognition from data but can lack physical interpretability and struggle with limited data. The challenge lies in leveraging the strengths of both approaches to achieve more accurate and reliable predictions, particularly in data-scarce or complex environments.

Methods

The core innovation is the development of a Hybrid Physics-Data (HPD) model within a Physics-Guided Neural Network (PGNN) framework. This HPD model integrates outputs from a well-established process-based model, VIC-CaMa-Flood, with meteorological forcing data (precipitation, temperature, wind speed) as inputs to an LSTM neural network.

  • VIC-CaMa-Flood: This component simulates the physical processes of water movement and routing across the LMRB at a 0.25° spatial resolution. It generates daily runoff and streamflow based on physical principles.
  • LSTM: This recurrent neural network architecture is chosen for its ability to capture long-term dependencies in time-series data, making it suitable for streamflow prediction.
  • PGNN Framework: This overarching framework guides the integration of physical knowledge (from VIC-CaMa-Flood) into the data-driven LSTM model. This is achieved by using the simulated streamflow from VIC-CaMa-Flood as a key input feature for the LSTM.

To understand the contribution of each input, the authors employed a Gradient Boosting Regression Tree (GBRT) model to measure feature importance.

Datasets

The study utilized the following datasets for the Lancang-Mekong River Basin:

  • Precipitation: Daily gridded data (0.25° resolution) from the APHRODITE research project.
  • Temperature and Wind Speed: Daily data (0.25° resolution) from the Global Meteorological Forcing Dataset.
  • Streamflow Observations: Daily data from five hydrological stations (Chiang Saen, Luang Prabang, Vientiane, Mukdahan, Pakse) obtained from Mohammed et al. [36].
  • Soil Data: Harmonized World Soil Database (HWSD).
  • Land Cover Data: Global land cover data (1 km resolution) from the University of Maryland.
  • Digital Elevation Model (DEM): Shuttle Radar Topography Mission (SRTM) data.

The streamflow data was divided into training (1966–1992), validation (1993–1997), and testing (1998–2007) periods, with an additional "human impact period" (2008–2015) for further analysis.

Key Insights and Ideas

  • Synergy of Physics and Data: The HPD model demonstrates that combining physical model outputs with data-driven models significantly enhances prediction accuracy. This suggests that physical regularization is crucial for improving data-driven hydrological modeling.
  • Physical Model as a "Mechanistic Understanding" Feature: The VIC-CaMa-Flood streamflow output acts as a crucial input feature for the LSTM, effectively embedding mechanistic understanding into the data-driven model. This helps the LSTM learn more physically meaningful patterns.
  • Importance of Process-Based Model Output: The GBRT analysis revealed that the VIC-CaMa-Flood streamflow contributes approximately 30% to the HPD model's predictions, substantially more than any single meteorological forcing variable (<20%). This highlights the value of incorporating outputs from established hydrological models.
  • Addressing Model Deficiencies: The HPD model not only improves accuracy but also implicitly corrects biases and deviations present in the pure process-based model, showcasing the complementary nature of the two approaches.
  • Human Impact on Floods: The study identified a significant reduction in observed flood indicators during the human impact period (post-2008), particularly at the CS station, likely due to the construction and operation of large reservoirs. This underscores the need to consider human activities in flood management.

Key Results

  • Superior Performance: The HPD model significantly outperformed both the pure VIC-CaMa-Flood model and a standalone LSTM model in streamflow and flood simulations.
  • Streamflow Accuracy: In the less impacted period (1966–2007), the HPD model achieved Nash-Sutcliffe Efficiency (NSE) values greater than 0.90 and relative errors (RE) less than 10% for streamflow.
  • Flood Accuracy: For flood indicators (MAF, Q95, Q90) in the less impacted period, the HPD model achieved NSE values greater than 0.86 and RE values less than 20%.
  • Human Impact Period Performance: While performance slightly decreased in the human impact period (2008–2015), especially at stations affected by reservoir construction (CS, VT), the model still provided satisfactory results, with NSE values for streamflow around 0.6-0.9 and RE values generally below 20%.
  • Feature Importance: VIC-CaMa-Flood streamflow was the most important input feature for the HPD model, contributing around 30% to the predictions across stations.

Limitations and Future Work

  • Scale Applicability: The authors acknowledge that the current HPD model, utilizing a large-scale hydrological model, is primarily applicable to river basins and global scales, and may not be directly transferable to small watersheds or local scales without further adaptation.
  • Topographic Data Integration: Future work could explore integrating topographic data (DEM, land use, slope) more directly into machine learning models for improved streamflow simulation, particularly for smaller scales.
  • Reservoir Impact Modeling: While the study observed the impact of reservoirs on floods, further research could focus on explicitly incorporating reservoir operation rules into hybrid models for more precise flood forecasting in regulated river systems.
  • Data Scarcity: The paper suggests that this hybrid approach offers a novel way to leverage incomplete physical understanding and insufficient data, implying its potential utility in data-scarce regions.

Summary updated

Abstract

A warming climate will intensify the water cycle, resulting in an exacerbation of water resources crises and flooding risks in the Lancang–Mekong River Basin (LMRB). The mitigation of these risks requires accurate streamflow and flood simulations. Process-based and data-driven hydrological models are the two major approaches for streamflow simulations, while a hybrid of these two methods promises advantageous prediction accuracy. In this study, we developed a hybrid physics-data (HPD) methodology for streamflow and flood prediction under the physics-guided neural network modeling framework. The HPD methodology leveraged simulation information from a process-based model (i.e., VIC-CaMa-Flood) along with the meteorological forcing information (precipitation, maximum temperature, minimum temperature, and wind speed) to simulate the daily streamflow series and flood events, using a long short-term memory (LSTM) neural network. This HPD methodology outperformed the pure process-based VIC-CaMa-Flood model or the pure observational data driven LSTM model by a large margin, suggesting the usefulness of introducing physical regularization in data-driven modeling, and the necessity of observation-informed bias correction for process-based models. We further developed a gradient boosting tree method to measure the information contribution from the process-based model simulation and the meteorological forcing data in our HPD methodology. The results show that the process-based model simulation contributes about 30% to the HPD outcome, outweighing the information contribution from each of the meteorological forcing variables (&lt;20%). Our HPD methodology inherited the physical mechanisms of the process-based model, and the high predictability capability of the LSTM model, offering a novel way for making use of incomplete physical understanding, and insufficient data, to enhance streamflow and flood predictions.
BibTeX
@inproceedings{liu2022physics,
  author = {Binxiao Liu and Qiuhong Tang and Gang Zhao and Liang Gao and Chaopeng Shen and Baoxiang Pan},
  title = {{Physics-Guided Long Short-Term Memory Network for Streamflow and Flood Simulations in the Lancang–Mekong River Basin}},
  booktitle = {Water},
  year = {2022},
  doi = {10.3390/w14091429},
}
Figure