Real-Time Water Level Prediction in Open Channel Water Transfer Projects Based on Time Series Similarity

Luyan Zhou, Zhao Zhang, Weijie Zhang, Kaijun An, Xiaohui Lei, Ming HeWater 2022 · DOI

TL;DR

This paper introduces a novel "slope-similar shape" method for predicting water levels in open-channel water transfer projects, a critical task for ensuring operational safety. By leveraging historical data with similar water level change trends, the method offers a feasible, effective, and interpretable approach to real-time prediction, especially when gate opening changes are known.

What Problem Does This Paper Address?

Open-channel water transfer projects, vital for alleviating water scarcity, are complex systems with numerous gates. Fluctuations in water levels, driven by gate operations, pose significant challenges for real-time prediction. Sudden changes can lead to downstream overflow or channel damage. Existing prediction methods often lack interpretability or struggle with the dynamic nature of these projects, particularly when gate openings change frequently. This paper tackles the challenge of accurately predicting water levels in real-time, specifically when the change in gate opening for the next moment is known, aiming for a method that is both effective and understandable.

Methods

The core of this paper is the proposed "slope-similar shape" method. It operates by using a short time series of three consecutive water level data points as a "query." This query is then used to search a historical dataset for segments exhibiting similar "slope-similar shapes." These shapes are defined by the trend and direction of water level changes, characterized by line segments and their slopes.

The method identifies historical data points that not only have a similar trend (slope-similarity) but also consider the magnitude of change and the direction of the water level fluctuation. To find the "best" match, the method uses similarity metrics (maximum deviation and cumulative deviation) and crucially incorporates the known change in gate opening at the prediction moment. Once the best historical match is found, the water level change observed at that historical point is used to predict the water level change for the current query. The predicted water level is then calculated by adding this predicted change to the last known water level of the query. This approach aims to be more interpretable than black-box machine learning models.

Datasets

The study utilizes a dataset from the Middle Route of the South-to-North Water Diversion Project in China. This dataset comprises 15,570 rows of data collected at 2-hour intervals between October 1, 2017, and April 28, 2021. The data includes water levels before gates, gate opening information, and flow rates. For the case study, specific gates (gates 4, 5, and 6) were selected for analysis.

Key Insights and Ideas

  • Time Series Similarity for Water Levels: The paper pioneers the application of time series similarity techniques for predicting water levels in open-channel water transfer projects, moving beyond traditional hydrological models or black-box ML approaches.
  • "Slope-Similar Shape" as a Predictive Feature: The novel concept of "slope-similar shape" effectively captures the dynamic trends of water level changes, which are crucial for predicting future levels in these systems.
  • Gate Opening as a Key Predictor: Explicitly incorporating the known change in gate opening at the next moment significantly enhances the prediction accuracy, highlighting its direct influence on water levels.
  • Interpretability Matters: The proposed method offers a more interpretable alternative to complex machine learning models, which is valuable for understanding the underlying physical processes and building trust in the predictions.
  • Stable Operations Lead to Similar Patterns: The analysis of the South-to-North Water Diversion Project reveals that the majority of water level changes fall within a narrow range, indicating relatively stable operations and a high prevalence of "slope-similar shapes," which supports the method's applicability.

Key Results

The proposed slope-similar shape method demonstrated promising results on the Middle Route of the South-to-North Water Diversion Project:

  • For datasets with water level variations less than 0.06 m, 87.5% of predictions had an error less than 0.03 m.
  • Of those, 71.4% had an error less than 0.02 m.
  • The method is described as feasible, effective, and interpretable.

Limitations and Future Work

The authors acknowledge several limitations:

  • Data Requirements: The method relies on historical datasets collected over a long period under various conditions, which might make it less suitable for newly constructed projects.
  • Non-Stationary Conditions: The current approach is less applicable to real-time rolling predictions under highly non-stationary conditions.
  • Monitoring Frequency: The relatively low monitoring frequency might not fully capture water level changes under certain conditions, potentially leading to misdiagnoses.
  • Single Data Source: Reliance on a single data source and potential water level outliers can affect prediction accuracy.

Future work will focus on:

  • Developing real-time water level prediction for non-stationary conditions using the slope-similar shape approach.
  • Establishing hydrodynamic models to improve predictions under extreme conditions.

Summary updated

Abstract

Changes in the opening of gates in open channel water transfer projects will cause fluctuations in the water level and flow of adjacent open channels and thus bring great challenges for real-time water level prediction. In this paper, a novel slope-similar shape method is proposed for real-time water level prediction when the change of gate opening at the next moment is known. The water level data points of three consecutive moments constitute the query. The slope similarity is used to find the historical water level datasets with similar change trend to the query, and then the best slope similarity dataset is determined according to the similarity index and the gate opening change. The water level difference of the next moment of the best similar data point is the water level difference of the predicted moment, and thus the water level at the next moment can be obtained. A case study is performed with the Middle Route of the South-to-North Water Diversion Project of China. The results show that 87.5% of datasets with a water level variation of less than 0.06 m have an error less than 0.03 m, 71.4% of which have an error less than 0.02 m. In conclusion, the proposed method is feasible, effective, and interpretable, and the study provides valuable insights into the development of scheduling schemes.
BibTeX
@inproceedings{zhou2022real,
  author = {Luyan Zhou and Zhao Zhang and Weijie Zhang and Kaijun An and Xiaohui Lei and Ming He},
  title = {{Real-Time Water Level Prediction in Open Channel Water Transfer Projects Based on Time Series Similarity}},
  booktitle = {Water},
  year = {2022},
  doi = {10.3390/w14132070},
}
Figure