|
ИСТИНА |
Войти в систему Регистрация |
ИСТИНА ЦЭМИ РАН |
||
Development of an intelligent system for early forecasting and modelling of flood situation on the example of the Republic of Bashkortostan using a proprietary machine and deep learning libraryстатья
Статья опубликована в высокорейтинговом журнале
Информация о цитировании статьи получена из
Scopus
Статья опубликована в журнале из списка Web of Science и/или Scopus
Дата последнего поиска статьи во внешних источниках: 22 мая 2024 г.
- Авторы: Palchevsky E., Antonov V., Filimonov N., others
- Журнал: Journal of Hydrology
- Том: 633
- Год издания: 2024
- Издательство: Elsevier BV
- Местоположение издательства: Netherlands
- Номер статьи: 130978
- DOI: 10.1016/j.jhydrol.2024.130978
- Аннотация: Reliable early flood forecasting models and systems are essential for minimizing and preventing the impacts of floods worldwide. However, the predictive accuracy of such models and systems is of-ten unacceptable for long-term forecasting, creating uncertainty in flood management decision-making. To address this problem, we propose an intelligent «Flood 2.0» system based on a proprie-tary machine learning and deep learning library and a flood zone visualisation module. The devel-opment of a new version of the system is preceded by a previous study «A system based on an arti-ficial neural network of the second generation for decision support in especially significant situa-tions».The machine learning and deep learning library developed within «Flood 2.0» consists of two parts: preprocessing and prediction. Initially, the study considers the method of data preprocessing. The essence of this method is the development and application of an impulse neural network for genera-tion of information filtering rules, which makes it possible to generate a high-quality initial set of retrospective data in an automated mode, cutting off unnecessary digital noise. Next, a predictive part (method) for water level prediction based on a dataset generated by a data preprocessing meth-od is discussed. This water level prediction method consists of: a) a machine learning model repre-sented as a polynomial regression; b) a deep learning model implemented as a modified recurrent neural network (RNN), a classical neural network with long-term short-term memory (LSTM) and a pulse neural network (PNN); c) a flood zone visualisation module, which is implemented on the ba-sis of the Leaflet library and allows modelling flood zones without the use of digital elevation mod-els, i.e. only on the basis of predicted water levels.To verify the effectiveness of the developed «Flood 2.0» system in general and the machine learn-ing and deep learning library in particular, many 10-day-ahead prediction experiments have been conducted. The empirical results show that the developed recurrent neural network model predicts more accurately under the proposed intelligent system “Flood 2.0″. Moreover, the difference com-pared to other models ranges from 49.22% to 78.98%, which proves the effectiveness of the devel-opment of the Flood 2.0 intelligent system in general and mathematical/technical modifications in particular. Thus, an early and accurate forecast, as well as visualised flood zones will give special services the necessary time to carry out flood control measures to prepare for the protection of tech-nical facilities of enterprises and evacuation of the population.Additional research has also shown that the Flood 2.0 system can be used to predict water levels and model flood zones in other regions.
- Добавил в систему: Филимонов Николай Борисович