Forecasting the spring flood of rivers with machine learning methods

Abstract

The subject of the research. The paper provides an overview of a flood forecasting problem in the Nenetsky region, Russia. The solution involves the use of the open source data on water level during the spring floods. Specifically, its collection, analysis and forecasting via machine learning models. Method. The authors describe a new forecasting approach that involves the use of the Holt-Winters model for a training sample, which is further implemented in order to train the following statistical models: XGBoost, Random Forest and Bagging. The solution is based on a sample of gauging stations’ historical indicators that provide a detailed description of weather conditions in the nearest settlements over several years. A separate sample was created for each location considered in the problem with the aim to build forecasts given a one-month or a one-year time period. Main Results. The forecast was obtained based on the results provided by individually trained models. In the future, the findings could be used when taking preventive measures during flood control. Practical relevance. Low maintenance costs of the information system along with the ability to predict the critical water level make this forecasting approach an economically viable additional measure against floods in poorer regions of Russia.

Keywords

• data mining
• machine learning
• flood forecast
• holt-winters model