Application of the Machine Learning LightGBM Model to the Prediction of the Water Levels of the Lower Columbia River

Min Gan, Shunqi Pan, Yongping Chen, Cheng Chen, Haidong Pan, Xian ZhuJournal of Marine Science and Engineering 2021 · DOI

TL;DR

This paper introduces the LightGBM machine learning model for predicting water levels in the complex, tide-influenced lower Columbia River estuary. The model, trained on river discharge and tidal data, achieves high accuracy, outperforming traditional physics-based methods and offering valuable insights into the interplay of subtidal constituents and river flow.

What Problem Does This Paper Address?

Estuaries are dynamic environments where river discharge and tidal forces interact in complex, non-linear ways, making water level prediction a significant challenge. Accurately forecasting these levels is crucial for managing flood risks, implementing early warning systems, and supporting various economic activities within these regions. Traditional physics-based models often struggle to fully capture these intricate interactions, especially when dealing with non-stationary tidal behavior influenced by varying river flows. This study aims to address this by applying a powerful machine learning approach, LightGBM, to improve the accuracy and understanding of estuarine water level prediction.

Methods

The core of this research is the application and optimization of the LightGBM (Light Gradient Boosting Machine) model. LightGBM is a gradient boosting framework based on decision trees, known for its speed and efficiency, particularly with large datasets. Unlike traditional level-wise tree growth, LightGBM employs a leaf-wise growth strategy, which prioritizes splitting the leaf node that yields the largest error reduction. This approach can lead to faster training and potentially more accurate models.

The study also details the hyperparameter optimization process using GridSearchCV from the Scikit-learn library. This involved systematically searching through various combinations of hyperparameters (e.g., learning_rate, num_iterations, max_depth, num_leaves) to find the optimal settings for the model. The model's performance was evaluated using metrics like Root Mean Square Error (RMSE), Mean Absolute Error (MAE), Correlation Coefficient (CC), and nondimensional Skill Score (SS).

Datasets

The study utilized historical data from the lower Columbia River. Specifically:

  • Water Level Measurements: Hourly data from five hydrometric stations (Astoria, Wauna, Longview, St.Helens, and Vancouver) were collected from the National Oceanic and Atmospheric Administration (NOAA) website for the period 2003-2020 (with some gaps for Wauna).
  • River Discharge Data: Synchronized hourly river discharge data for the Columbia River at Bonneville Dam (Qc) and the Willamette River at Portland (Qw) were obtained from the U.S. Geological Survey (USGS) for the same period (2003-2020).

The data was divided into an 80% training set and a 20% testing set for model development and evaluation.

Key Insights and Ideas

  • LightGBM's Efficacy in Estuaries: The study demonstrates that LightGBM is a highly effective tool for predicting water levels in complex estuarine environments, outperforming traditional methods.
  • Importance of Subtidal Constituents: The analysis revealed that subtidal constituents, often overlooked, play a significant role in estuarine water level dynamics, particularly in interaction with river discharge.
  • Spatial Variability of Influences: The contribution of different input parameters (river discharge, tide range, tide constituents) to the prediction model varies spatially along the estuary, highlighting the need for localized modeling approaches.
  • Impact of River Discharge on Tidal Phases: The research identified a phase lag in predicted water levels when river discharge is strong, suggesting that the model's accuracy can be influenced by the timing of upstream flow.

Key Results

  • High Prediction Accuracy: The LightGBM model achieved a Root Mean Square Error (RMSE) as low as 0.14 m across various stations.
  • Superior Performance: The model demonstrated statistically better performance than the physics-based Nonstationary Tidal Harmonic Analysis model (NS_TIDE), with lower RMSE values and higher Skill Scores (SS) ranging from 0.941–0.972 for LightGBM compared to 0.938–0.965 for NS_TIDE.
  • Strong Correlation: Correlation Coefficients (CC) for the LightGBM model were consistently high, ranging from 0.975–0.987.
  • Parameter Importance: River discharges (Qc and Qw) were consistently ranked among the top three most important parameters for prediction across all stations. The M2 tide constituent was also highly ranked, followed by several subtidal constituents.
  • Flood Season Errors: While generally accurate, the model showed slightly larger errors during flood seasons, particularly in the upstream reaches, which were attributed to factors like river discharge propagation time and tidal conditions.

Limitations and Future Work

  • Propagation Time Ignored: The study acknowledges that the model did not explicitly account for the propagation time of upstream river discharge or tides, which can lead to phase lags and larger errors during periods of strong flow.
  • Flood Season Accuracy: While comparable to physics-based models, the errors during flood seasons, especially in upstream areas, suggest room for improvement.
  • Future Directions: The authors suggest that incorporating propagation time into the model could further enhance its accuracy. The success of LightGBM also opens avenues for its application in other similar estuarine systems worldwide.

Summary updated

Abstract

Due to the strong nonlinear interaction with river discharge, tides in estuaries are characterised as nonstationary and their mechanisms are yet to be fully understood. It remains highly challenging to accurately predict estuarine water levels. Machine learning methods, which offer a unique ability to simulate the unknown relationships between variables, have been increasingly used in a large number of research areas. This study applies the LightGBM model to predicting the water levels along the lower reach of the Columbia River. The model inputs consist of the discharges from two upstream rivers (Columbia and Willamette Rivers) and the tide characteristics, including the tide range at the estuary mouth (Astoria) and tide constituents. The model is optimized with the selected parameters. The results show that the LightGBM model can achieve high prediction accuracy, with the root-mean-square-error values of water level being reduced to 0.14 m and the correlation coefficient and skill score being in the ranges of 0.975–0.987 and 0.941–0.972, respectively, which are statistically better than those obtained from physics-based models such as the nonstationary tidal harmonic analysis model (NS_TIDE). The importance of subtide constituents in interacting with the river discharge in the estuary is clearly revealed from the model results.
BibTeX
@article{gan2021application,
  author = {Min Gan and Shunqi Pan and Yongping Chen and Cheng Chen and Haidong Pan and Xian Zhu},
  title = {{Application of the Machine Learning LightGBM Model to the Prediction of the Water Levels of the Lower Columbia River}},
  journal = {Journal of Marine Science and Engineering},
  year = {2021},
  doi = {10.3390/jmse9050496},
}
Figure