Anne Johannet

Rainfall-runoff modeling of flash floods in the absence of rainfall forecasts: the case of Cévenol flash floods

Abstract“Cévenol flash floods” are famous in the field of hydrology, because they are archetypical of flash floods that occur in populated areas, thereby causing heavy damages and casualties. As a consequence, their prediction has become a stimulating challenge to designers of mathematical models, whether physics based or machine learning based. Because current, state-of-the-art hydrological models have difficulty performing forecasts in the absence of rainfall previsions, new approaches are necessary. In the present paper, we show that an appropriate model selection methodology, applied to neural network models, provides reliable two-hour ahead flood forecasts.

Optimization of the generalization capability for rainfall–runoff modeling by neural networks: the case of the Lez aquifer (southern France)

Neural networks are increasingly used in the field of hydrology due to their properties of parsimony and universal approximation with regard to nonlinear systems. Nevertheless, as a result of the existence of noise and approximations in hydrological data, which are very significant in some cases, such systems are particularly sensitive to increased model complexity. This dilemma is known in machine learning as bias–variance and can be avoided by suitable regularization methods. Following a presentation of the bias–variance dilemma along with regularization methods such as cross-validation, early stopping and weight decay, an application is provided for simulating and forecasting karst aquifer outflows at the Lez site. The efficiency of this regularization process is thus demonstrated on a nonlinear, partially unknown basin. As a last step, results are presented over the most intense rainfall event found in the database, which allows assessing the capability of neural networks to generalize with rare or extreme events.

Flash Flood Forecasting by Statistical Learning in the Absence of Rainfall Forecast: A Case Study

The feasibility of flash flood forecasting without making use of rainfall predictions is investigated. After a presentation of the “cevenol flash floods“, which caused 1.2 billion Euros of economical damages and 22 fatalities in 2002, the difficulties incurred in the forecasting of such events are analyzed, with emphasis on the nature of the database and the origins of measurement noise. The high level of noise in water level measurements raises a real challenge. For this reason, two regularization methods have been investigated and compared: early stopping and weight decay. It appears that regularization by early stopping provides networks with lower complexity and more accurate predicted hydrographs than regularization by weight decay. Satisfactory results can thus be obtained up to a forecasting horizon of three hours, thereby allowing an early warning of the populations.

Atmospheric dispersion modeling using Artificial Neural Network based cellular automata

Forecasting atmospheric dispersion in complex configurations is a current challenge in fluid dynamics in terms of calculation time and accuracy. CFD models provide good accuracy but require a great computation time. Simplified or empirical models are designed to quickly evaluate the dispersion but are not adapted to complex geometry. Cellular Automata coupled with an Artificial Neural Network (CA-ANN) are developed here to calculate the atmospheric dispersion of methane (CH4) in 2D. Efforts are made in reducing computation time while keeping an acceptable accuracy. A CFD simulations database is created and the Advection-Diffusion Equation is discretized to provide variables for the ANN. Neural network design is made thanks to best sampling selection, architecture selection and optimized initialization. The coefficient of determination is over 0.7 for most cases of the test set despite small errors accumulated through time steps. CA-ANN is faster than CFD models by a factor from 1.5 to 120. Display Omitted A new atmospheric dispersion model is developed based on combination of Cellular Automata and Artificial Neural Networks.Comparisons are made with CFD RANS standard k-ź model on 2D free field dispersion of methane.CA-ANN is faster than CFD standard k-ź by a factor from 1.5 to 120 in the modeled simulations while keeping accuracy.

Feed-forward vs recurrent neural network models for non-stationarity modelling using data assimilation and adaptivity

Abstract Artificial neural networks (ANN) are nonlinear models widely investigated in hydrology due to their properties of universal approximation and parsimony. Their performance during the training phase is very good, and their ability to generalize can be improved by using regularization methods such as early stopping and cross-validation. In our research, two kinds of generic models are implemented: the feed-forward model and the recurrent model. At first glance, the feed-forward model would seem to be more effective than the recurrent one on non-stationary datasets, because measured information on the state of the system (measured discharge) is used as input, thereby implementing a kind of data assimilation. This study investigates the feasibility and effectiveness of data assimilation and adaptivity when implemented in both feed-forward and recurrent neural networks. Based on the IAHS Workshop held in Göteborg, Sweden (July 2013), the hydrological behaviour of two watersheds of different sizes and different kind of non-stationarity will be modelled: (a) the Fernow watershed (0.2 km2) in the USA, affected by significant modifications in land cover during the study period, and (b) the Durance watershed (2170 km2) in France, affected by an increase in temperature that is causing a decrease in the extent of glaciers. Two methods were applied to evaluate the ability of ANN to adapt on the test set: (i) adaptivity using observed data to adapt parameter values in real time; and (ii) data assimilation using observed data to modify inaccurate inputs in real time. The goal of the study is thus re-analysis and not forecasting. This study highlights how effective the feed-forward model is compared to the recurrent model for dealing with non-stationarity. It also shows that adaptivity and data assimilation improve the recurrent model considerably, whereas improvement is marginal for the feed-forward model in the same conditions. Finally, this study suggests that adaptivity is effective in the case of changing conditions of the watershed, whereas data assimilation is better in the case of climate change (inputs modification).

Comparison between inverse modelling and data assimilation to estimate rainfall from runoff using the multilayer perceptron

The ability of the multilayer perceptron to model the inverse relation of a fictitious watershed is investigated. Comparison is done between a new formulation of data assimilation and the standard multilayer perceptron applied to three kinds of models: static, feedforward and recurrent. It appears that both techniques are equivalent and allow a very good estimation of the inverse relation. This study aims at proposing methods to supplement or adapt historical databases to modern instrumentation. Datasets will thus be used over a longer time-series to better apprehend the consequences of global warming.

Flash flood forecasting using Support Vector Regression: An event clustering based approach

We present a new machine learning approach to flash flood forecasting in the absence of rainfall forecasts, based on the agglomerative hierarchical clustering of flood events. Each cluster contains events whose models have similar behaviors. Specific Support Vector Regression models are then trained from each cluster. The test results show that a specific model may be more accurate than a general model trained from all floods present in the training database.

Karst Flash Flood Forecasting Using Recurrent and Non-recurrent Artificial Neural Network Models: The Case of the Lez Basin (Southern France)

Flash floods pose significant hazards in urbanized zones and have important implications that should probably increase due to global changes. Early warning is thus a priority that could be done by using forecast models. When such events occur on karst basins, well known for their intrinsic complexity, anisotropy and heterogeneity, the lack of knowledge regarding the various hydrodynamic behaviours involved in karst systems prevents to use physical models. A generic black box method seems thus to be adequate; specifically, artificial neural network modelling seems to be a relevant method. To model hydrosystem behaviour efficiently, neural networks need to dispose of relevant data sets constituting input and output variables, and rigorous application of regularization methods. In this study, we propose to apply two kinds of models: feedforward and recurrent neural networks to flash flood forecasting. These models are designed using a specific methodology to diminish their complexity. They are applied to the Lez karst aquifer, located in southern France, and their performances are compared. Recurrent model can be used at longer lead time for operational flash flood forecasting. Nevertheless, for short horizon of prevision, performances of feedforward model are higher than those showed by recurrent one. The comparison of both models is then necessary to guide the improvement of operational flash flood forecasting.

Operational Turbidity Forecast Using Both Recurrent and Feed-Forward Based Multilayer Perceptrons

Approximately 25% of the world population drinking water depends on karst aquifers. Nevertheless, due to their poor filtration properties, karst aquifers are very sensitive to pollutant transport and specifically to turbidity. As physical processes involved in solid transport (advection, diffusion, deposit…) are complicated and badly known in underground conditions, a black-box modelling approach using neural networks is promising. Despite the well-known ability of universal approximation of multilayer perceptron, it appears difficult to efficiently take into account hydrological conditions of the basin. Indeed these conditions depend both on the initial state of the basin (schematically wet or dry), and on the intensity of rainfalls. To this end, an original architecture has been proposed in previous works to take into account phenomenon at large temporal scale (moisture state), coupled with small temporal scale variations (rainfall). This architecture, called hereafter as “two-branches” multilayer perceptron is compared with the classical two layers perceptron for both kinds of modelling: recurrent and non-recurrent. Applied in this way to the Yport pumping well (Normandie, France) with 12 h lag time, it appears that both models proved crucial information: amplitude and synchronization are better with “two-branches” feed forward model when thresholds surpassing prediction is better using classical feed forward perceptron.

Neural Networks for Karst Spring Management. Case of the Lez Spring (Southern France)

Karst hydrosystems constitute important water resource but their recharge and emptying process are poorly known and quantified. Water resource management is thus difficult. Nevertheless, it is a major issue when rainfall is not uniformly distributed during the year, as in Mediterranean climate. This study proposes a method based on neural networks permitting to simulate karst emptying as a function of the pumping volume during the dry period. Applied to the Lez karst system, the model provides excellent simulations of the water level at the main outlet of the system by using mean pumping discharge and zero rainfall hypothesis during dry period. An arbitrary extreme scenario is also provided by introducing a mean pumping volume.