EGU General Assembly 2023
© Author(s) 2023. This work is distributed under
the Creative Commons Attribution 4.0 License.

Water level forecasting of Reservoir downstream by machine learning

I-Hsiu Chuang, Gwo-Fong Lin, Ming-Jui Chang, and Yuan-Fu Zeng
I-Hsiu Chuang et al.
  • Department of Civil Engineering, National Taiwan University, Taipei

Taiwan is located in the subtropical monsoon region, both typhoons and vigorous convection caused by strong southwesterly flow develop seasonal compound disasters.
In addition, the response time for early warning systems of the reservoirs and downstream riverbanks has been shortened due to higher frequency and greater intensity of short-duration rainfall events in recent years. Past studies pointed out that the current water level forecast does not consider the outflow discharge of the reservoir. Therefore, this study proposes a downstream water level forecasting model that considers the outflow discharge of the reservoir, and the model is provided to relevant hazard mitigation centers.
This research has selected the water level of the Taipei bridge as target status and collected data of typhoon and storm events from 2014-2021. These data included the precipitation in the watershed of upstream of Taipei bridge, outflow discharge of Shimen reservoir, outflow discharge of Feitsui reservoir, and tidal of Tamsui river estuary as the alternative factors. Subsequently, building several models based on multiple machine learning, such as RNN, SVM, and LSTM to interface with the constant-quantity rainfall forecast of the Central Weather Bureau, then produce the forecast in the future 12 hours with Multi-Step Forecasting (MSF) about the water level of Taipei bridge.
The result shows that SVM, RNN, LSTM forecast in the future 1 hour precisely, which statistical values of CC are more than 0.97, and root mean square errors of water level are around 0.2 m. As the forecast time is longer, the statistical values of CC decrease around 0.93 and root mean square errors of water level increase around 0.3 m.
However, LSTM is able to learn dependencies between the time series and get more precise outcomes than the SVM and RNN, which is not outstanding initially then performs the best at the last. The proposed water level forecast is proved to improve the accuracy of the forecast in the future 12 hours about the water level of Taipei bridge. Moreover, by coordinating the Quantitative Precipitation Forecast (QPF) and warning water level, the model provides early warning of the future twelve-hour water level, which is not only beneficial to evacuation and operating traversing dock-gate and evacuation gates efficiently, but also conducive to reducing the risk of losses in life and property.
Keywords: Water level, Quantitative Precipitation Forecast, Machine learning, Multi-Step Forecasting

How to cite: Chuang, I.-H., Lin, G.-F., Chang, M.-J., and Zeng, Y.-F.: Water level forecasting of Reservoir downstream by machine learning, EGU General Assembly 2023, Vienna, Austria, 24–28 Apr 2023, EGU23-17521,, 2023.