DOI: https://doi.org/10.15407/uhmi.conference.01.10

FUZZY LOGIC BASED FLASH FLOOD FORECAST

 

Ing. Petr Janál, Ph.D., Ing. Tomáš Kozel, Ph.D. 

Czech Hydrometeorological Institute

Corresponding author: Ing. Petr Janál, Ph.D., Czech Hydrometeorological Institute, Brno Regional Office,

Kroftova 43, Brno, 61667, Czech Republic, petr.janal@chmi.cz

 

ABSTRACT

The flash flood forecasting remains one of the most difficult tasks in the operative hydrology worldwide. The torrential rainfalls bring high uncertainty included in both forecasted and measured part of the input rainfall data. The hydrological models must be capable to deal with such amount of uncertainty. The artificial intelligence methods work on the principles of adaptability and could represent a proper solution. The application of different methods, approaches, hydrological models and usage of various input data is necessary.

The tool for real-time evaluation of the flash flood occurrence was assembled on the bases of the fuzzy logic. The model covers whole area of the Czech Republic and the nearest surroundings. The domain is divided into 3245 small catchments of the average size of 30 km2. Real flood episodes were used for the calibration and future flood events can be used for recalibration (principle of adaptability). The model consists of two fuzzy inference systems (FIS). The catchment predisposition for the flash flood occurrence is evaluated by the first FIS. The geomorphological characteristics and long-term meteorological statistics serve as the inputs. The second FIS evaluates real-time data. The inputs are: The predisposition for flash flood occurrence (gained from the first FIS), the rainfall intensity, the rainfall duration and the antecedent precipitation index. The meteorological radar measurement and the precipitation nowcasting serve as the precipitation data source. Various precipitation nowcasting methods are considered. The risk of the flash flood occurrence is evaluated for each small catchment every 5 or 10 minutes (the time step depends on the precipitation nowcasting method).

The Fuzzy Flash Flood model is implemented in the Czech Hydrometeorological Institute (CHMI) – Brno Regional Office. The results are available for all forecasters at CHMI via web application for testing. The huge uncertainty inherent in the flash flood forecasting causes that fuzzy model outputs based on different nowcasting methods could vary significantly. The storms development is very dynamic and hydrological forecast could change a lot of every 5 minutes. That is why the fuzzy model estimates are intended to be used by experts only.

The Fuzzy Flash Flood model is an alternative tool for the flash flood forecasting. It can provide the first hints of danger of flash flood occurrence within the whole territory of the Czech Republic. Its main advantage is very fast calculation and possibility of variant approach using various precipitation nowcasting inputs. However, the system produces large number of false alarms, therefore the long-term testing in operation is necessary and the warning releasing rules must be set.

Keywords: Fuzzy Logic, Flash Flood, Operative Hydrology.

 

INTRODUCTION

 

The flash flood forecasting has always represented a major challenge for hydrologists. A causal torrential rainfall has substantial dynamics in both space and time and it brings high amount of uncertainty, which we will probably not be able to eliminate sufficiently in the near future. Mentioned uncertainty must be taken into account in the process of forecasting as well as when interpreting the results. The Czech hydrometeorological institute (CHMI) is the national service for meteorology, hydrology and air quality and ensures the flood forecasting service (FFS) in the Czech Republic. Standard hydrological forecast based on outputs from numerical prediction models is issued for more than one hundred forecasting profiles but in the case of flash floods it is not sufficient and a different approach is needed. A prime product of CHMI for flash flood forecasting is Flash Flood Guidance (Daňhelka et al., 2015) and output from this model is published in CHMI web site. Simultaneously, attention to a development of other tools is being paid. For example, the distributive hydrological models used for flood forecasting on bigger catchments can be applied but it requires very detailed schematization and calculation is time-consuming. Currently this approach is tested only on selected small catchments and it doesn’t cover whole area of the Czech Republic (Daňhelka et al., 2015). The artificial intelligence based methods are also tested in the CHMI, the Fuzzy Flash Flood model is introduced in following text. Its advantage consists particularly of very fast calculation, which enables us to evaluate the most up-to-date input data in more variants. Moreover, the adaptability principle becomes more and more important in the current climate change context. The fuzzy model assembly comes from the real flash flood episodes (2009-2019) and the new episodes are being added constantly. The model is so able to reflect possible changes in the rainfall runoff processes. The difficulty of flash flood forecasting requires the usage of more tools and that is why the variety in the modeling approaches will be always beneficial.

 

METHODS AND DATA  

 

In general, a process of a hydrological forecast could be divided into the three elementary steps:

  1. An input data preparation.
  2. A calculation of a hydrological model.
  3. A results evaluation and publication.

 

Let us compare a standard hydrological forecast (it means forecast for a catchment of size of hundreds of square kilometers, usually based on outputs from numerical weather prediction models) and flash flood forecast in each mentioned step.

In the case of a standard hydrological forecast, the first step (input data preparation) enables checking input data both automatically and manually. Time-resolution of input data is usually 1 hour. Measured precipitation ordinarily consists of a merged product calculated as radar estimates combined with rain gauge measurements. More variants of weather forecast could be considered, according available numerical weather prediction (NWP) models. Through the consultation with meteorologists, the most probable future weather development could be determined.

In contrast, a flash flood forecasting requires the most frequent updating as possible (5-10 minutes). A manual checking or editing of input data is unrealistic. The whole process must be fully automated. Time-resolution of input data should be 5-10 minutes. Measured precipitation is being derived from radar measurement and should be significantly under/overestimated. The precipitation forecast comes from the extrapolation of radar echo (nowcasting) and includes a huge amount of uncertainty. For example, extrapolation methods do not involve the life cycle of storm cells. It is possible to consider more variants of precipitation nowcasting methods.

In the second step, a calculation of a hydrological model is carried out. When calculating a standard hydrological forecast, hydrologist’s main work lies in the adaptation of the hydrological model to the current rainfall-runoff situation. Parameters of the hydrological model could be adjusted to achieve the best possible matching of measured and simulated discharges. The estimation of the future discharge development follows.

The flash flood forecasting does not enable any real-time adjustment of the hydrological model especially because of the lack of time and high uncertainties included in the input data. Hydrologists are reliant on the automated results only. The hydrological model can be recalibrated additionally (for example according hit/miss/false analyses). It needs to be pointed out, that flash floods occur randomly and mostly hit an unobserved catchment. That means that there is the lack of relevant data for the calibration of the hydrological model.

There are differences also in the publishing options. Within the FFS provided by CHMI, the standard hydrological forecast is published on the internet and usually it is updated twice a day and during the flood situation, the updates can be done more frequently (every hour if needed). The flood reports with a verbal description of the current state and the further development are issued. The publishing of the flash flood forecast still remains a subject for discussion, in particular because of huge uncertainty of the results and the necessity of more complex interpretation. Currently, the presentation of forecast from the Flash flood guidance model is tested on CHMI website. The outputs from the Fuzzy Flash Flood model are not available for public and are used only internally.

The problem lies in the fact that we probably cannot enhance the accuracy of the input data significantly in the near future. That is the main motivation for using the artificial intelligence methods.

The theory of the fuzzy logic could be found for example in (Jang, 1993). The principles of the Fuzzy model assemblage are described in (Janál, Starý, 2012). In following text, a description of the current version of the Fuzzy model with the emphases on its operation is provided.

The whole area of the Czech Republic and certain surroundings is covered by the model. The area of interest is divided into the small catchments of the size of 30 km2 on average. There are 3245 small catchments in total and for each the input variables are considered as average values. Many different structures of the model were tested in the past. The current version consists of two fuzzy interference systems (FIS). The first FIS was developed by dr. Ježik (Ježík, 2015) and it serves to the determination of the predisposition to the occurrence of the flash flood for each mentioned small catchment. Input variables are the catchments characteristics like the area, forest cover, slope, soil type and others. The second FIS forms an operative part of the model and has 4 input variables:

 

 

The values of the first input (Potential predisposition to the occurrence of the flash flood) are calculated in advance by the first FIS for the whole set of the small catchments and they are fixed. These values could be updated by new calculation of the first FIS when the more relevant catchments characteristics are available or in terms of the analyses of the success of the model. The remaining three input variables (a precipitation characteristics) are computed operatively for the whole set of the small catchments in each time step. More variants of the precipitation forecast are considered based on different precipitation nowcasting methods. The time step (updating frequency) of the Fuzzy model is 5 or 10 minutes, according to the used precipitation nowcasting product. The time interval of 5 hours is considered in each time step, 2 hours of history and 3 hours of nowcasting. The detailed description could be found in (Janál, Starý, 2012). Currently, three precipitation nowcasting products are used as inputs for the Fuzzy model (Haiden at al., 2011), (Novák, 2007):

 

 

Additionally, the Fuzzy model is calculated in the variant when no precipitation nowcasting is taken into account and only measured precipitation is considered. In this variant, the uncertainty of input data is significantly lower, but the time for the warning is reduced.

The antecedent precipitation index is also updated in each time step (moving method).

The output variable of the Fuzzy model is the flash flood endangerment degree and it is determined for each catchment (3245 values in each time step). The modeled catchments are interlinked in the meaning that the endangerment is propagated downstream while it is reduced gradually. The exact time of the culmination is not the subject of the forecast. The interpretation is so, that flash flood could occur in the nearest future (in oncoming hours or minutes). The output variable ranges from 0 to 1, when 0 means no endangerment and 1 means the endangerment of flood with the return period of 100 years or more. This interval is divided into 5 levels represented by different colors, which are used for the operative presentation of the results through the Fuzzy Flash Flood application created by J. Brzezina, see Fig. 1.

figure1 

Fig. 1. Fuzzy Flash Flood application 

 

Through the application, the hydrological response based on different precipitation nowcasting methods can be compared in the form of the maps of endangered areas. Hydrologists can get a primary information about oncoming situation almost immediately after data from the meteorological radar are available. It means that the time for warning or some reaction is maximized. An easier interpretation could be achieved by merging the results into the bigger areas, which reflects the areas for standard warnings of CHMI (lower right map on the Fig. 1). All results are stored in relation database and are available for the retrospective analyses.  

 

RESULTS AND DISCUSSION

 

The fuzzy model is in the testing operation in Brno regional office of CHMI and results are shared with the other offices of CHMI for internal use. Contemporary development is focused mainly on the validation of the model. This part is very demanding because of the character of flash floods. There is often not sufficient feedback and we can hardly detect all events, which have happened. Alongside the local case studies, the more robust validation method is being developed. The validation method should be able to evaluate continuous time period and should be automatized. The essential requirement for such method is the reliable source of the impacts caused by the torrential rainfalls. Since CHMI closely cooperates with firefighters, the database of the firefighter actions was used for this purpose. This source of impact data has advantages that it is covering the whole territory of the Czech Republic and the events are localized by GPS coordinates. However, we must be aware of weaknesses of this data source. The time of the firefighter action does not always correspond with the time of flash flood. The reason of firefighter action is described by the specific code, it enables us to select only the situations that concern the flooding, for example the flooded cellars. But not all such events are caused by torrential rainfalls. The cellars could be for example flooded by the water pipe breakdown.

The algorithm of evaluation method was compiled taking into account all mentioned features of the firefighter actions database. Flash flood warnings were evaluated for the same areas as in the case of the standard warnings of CHMI (lower right map on the figure 1) and hit-miss-false statistic was processed. The goal of evaluation was to find an adequate sensitivity of the Fuzzy model. Since the evaluation method is still the subject of the development, the results published in this article cover only short time period from 26th May to 1st June 2018, nevertheless they could illustrate the potentialities of the model. Seven different levels of the model sensitivity were tested, that means that warnings were issued after exceedance of seven different thresholds, whereas the sensitivity of the model was decreasing from the first to the seventh variant.

 

 

Fig. 2. Hit-miss-false ratio

 

The hit-miss-false ratio for mentioned seven variants of the model sensitivity is depicted in Fig. 2. The high amount of false alarms will be probably always present because of the high uncertainty of input data but it could be reduced by appropriate setting of the warning threshold. It is up to as to decide which warning threshold is the most suitable for the practice. The high number of the false alarms could lead to the devaluation of the forecast in the eyes of the users. On the other hand, the issuing of the false alarm might not be connected with the same risk as in the case of the missed flood.

 

CONCLUSIONS

 

The aspiration for the flash flood forecasting comes from the possibilities that are currently available in CHMI. Fifteen years ago, the flash flood forecast seemed to be almost impossible. The major progress in the meteorological radar measurement and precipitation nowcasting brought certain ways to predict even such fast natural disasters. The successfulness of the forecast is directly dependent on the accuracy of the precipitation estimates. The error of the precipitation measurement during the convective precipitation event could be dozens of percent and the error of the precipitation nowcasting could be even higher. We cannot rely on the specific values of the precipitation inputs since they are “always wrong”. The essence of the flash flood forecast lies in the real-time evaluation of all the data we have. The artificial intelligence methods enable the very fast calculation and the ability to work with the uncertain data.

The question of the form of warning remains open. The publishing of the forecast through the websites might not by sufficient. If we were able to warn the residents of endangered municipality directly, the reaction strategy would have to be clearly specified. In the extreme cases it is not about the flooded cellars but about saving lives. Early warning may provide a few minutes for leaving the zones around watercourses.        

REFERENCES

Daňhelka. J, Janál, P., Šercl, P. a kol. Možnosti predikce přívalových povodní v podmínkách České Republiky,

Edice Sborník prací Českého hydrometeorologického ústavu , 2015, Praha, 50 s., ISBN 978-80-87577-27-1, ISSN 02320401.

Haiden et al. The Integrated Nowcasting through Comprehensive Analysis (INCA) system and its validation over the Eastern Alpine region. Weather Forecasting, 26, 2011.

Janál, P., Starý, M. Fuzzy model used for the prediction of a state of emergency for a river basin in the case of a flash flood – part 2, J. Hydrol. Hydromech. 2012. Vol. 60. No. 3. P. 162-173.

Jang, J.R. ANFIS: Adaptive-Network-Based Fuzzy Inference System, IEEE Transactions on Systems, Man, and Cybernetics. 1993. Vol. 23. No. 3.

Ježík, P. (2015): Využití vybraných metod umělé inteligence pro nalezení malých povodí nejvíce ohrožených povodněmi z přívalových dešťů. Brno, 2015, Ph.D. thesis. Brno University of Technology, Faculty of civil engineering, Institute of landscape water management.

Novák, P. The Czech Hydrometeorological Institute’s Severe Storm Nowcasting System. Atmospheric Research, 2007. Vol. 83. P. 450–457.

Back to CONTENTS