An operational real-time flood forecasting system

in Southern Italy
Enrique Ortiz (1), Gabriele Coccia (1)(2), Ezio Todini (1)(3)

ABSTRACT
A real-time flood forecasting system has been operating since year 2012 as a non-structural measure for mitigating the flood risk in Campania Region (Southern Italy), within the Sele river basin (3.240 km
2). The Sele Flood Forecasting System (SFFS) has been built within the FEWS (Flood Early Warning System) platform developed by Deltares and it assimilates the numerical weather predictions of the

COSMO LAM family: the deterministic COSMO-LAMI I2, the deterministic COSMO-LAMI I7 and the ensemble numerical weather predictions COSMO-LEPS (16 members).

Sele FFS is composed by a cascade of three main models. The first model is a fully distributed continuous hydrological model, named TOPKAPI-eXtended (Idrologia&Ambiente s.r.l., Naples, Italy), simulating
the dominant processes controlling the soil water dynamics, runoff generation and discharge with a spatial resolution of 250 m. The second module is a set of neural-networks built for forecasting the river
stages at a set of monitored cross-sections. The third component is a Model Conditional Processor (MCP) , which provides the predictive uncertainty (i.e., the probability of occurrence of a future flood event)
within the framework of a multi-temporal forecast, according to the most recent advancements on this topic (Coccia and Todini, HESS, 2011). The MCP provides information about the probability of exceedance
of a maximum river stage within the forecast lead time, by means of a discrete time function representing the variation of cumulative probability of exceeding a river stage during the forecast lead time and the
distribution of the time occurrence of the flood peak, starting from one or more model forecasts. This work shows the Sele FFS performance after two years of operation, evidencing the added-values that can
provide to a flood early warning and emergency management system.

Real-time monitoring network & System Platform COSMO Model Forecast Assimilation
The real-time monitoring network of Functional Centre, managed by the Civil Protection of Campania Region Weather forecasts assimilated by the System are the total precipitation,
(13,590 km2), has 192 stations instrumented with 577 sensors of which 178 are raingauges, 99 thermometers, 3 temperature and wind speed and direction.
snow gages and 63 hydrometric stations with ten minutes time step.The flood forecast system is based on the
Delft-FEWS platform (Deltares), which in recent years has established itself as a world leader, thanks to the easy • Forecast model COSMO i2 to 48 hours: Resolution of 2.8x2.8 km, made
integration of all types of hydrologic and hydraulic models, as well as their robustness in the acquisition, twice daily (00Z and 12Z) with hourly time interval.
management and processing all types of hydro meteorological data (Observation networks, meteorological
forecasts derived from NWP models, weather meteorological radar images, data from satellite platforms, etc.). • Forecast model COSMO i7 to 72 hours: Resolution of 7.0x7.0 km, made
twice daily (00Z and 12Z) with hourly time interval.

• Forecast model COSMO-LEPS to 132 hours (16 Members): Resolution of

7.0x7.0 km, made twice daily (00Z and 12Z) with three-hourly time interval.

Forecast Hydrological modeling Chain: TOPKAPI-X / Artificial Neural Network / Model Conditional Processor (MCP)
TOPKAPI-X (TOPographic and Kinematic APproximation and Integration EXtended) is a physically based
distributed hydrological model originated by Prof. Todini at the University of Bologna in the ‘90s. The most recent
version: TOPKAPI-X, developed and owned by Idrologia&Ambiente S.r.l., includes, with respect to previous
realizations, several additional features such as:
Additional soil response in order to reproduce different hydrological situations (two soil layers instead of one);
Introduction of an 8-directions drainage network; Addition of an infiltration module based on the Green-Ampt model,
to reproduce hortonian processes; Addition of a groundwater model based on the cellular automata with full 2-D IFD
scheme (integrated finite difference);Introduction of a new coefficient to take into account the sun height with
respect to the cell aspect for the radiation assessment into the snow accumulation and melting module based on
mass and energy balance.
TOPKAPI-X is based on the hypothesis that sub-surface flow (Two Layers), overland flow and channel flow can be
approximated using a Kinematic wave approach. The integration in space of the non-linear kinematic wave
equations representing subsurface flow, overland flow and channel flow results in four ‘structurally-similar’ non-
linear reservoir equations.
The application of artificial neural networks (ANN’s) to various aspects of hydrological modeling has yielded to
interesting and promising results during recent years. In particular, rainfall-runoff models and real time forecasting
systems based on ANN’s schemes have received special attention.
Their ability to incorporate in a systematic approach non linear relationships between variables represents an
attractive aspect in favor of ANN’s modelling. The vast majority of the networks are trained with the popular back-
propagation learning algorithm BPNN, which is also used herein, with the MSE (mean squared error function) as
the objective function. The number of hidden layers is fixed to one, which is sufficient to approximate any complex
nonlinear function with the desired accuracy. Following a typical and well tested configuration in the literature, the
activation function of the input and output layer has been set to a linear function, while the hidden layer incorporates
the non linear sigmoid function, providing the ability of the network to capture possible complex non linear
relationships between inputs and target outputs.
A robust set of ANN’s models has been implemented in six hydrometric stations of Sele FFS.
800
MLP (Multi Layer Perceptron) f ( y) 
1
1  ey
(0;1)

700 n
y j   wi , j xi  b j
i 1

600

500
Nivel (cm)

400

300

200 PSP = Post-Synaptic-

Potential
100

0
04/09/08 03/12/08 03/03/09 01/06/09 30/08/09

Niveles predichos t+12 Niveles Observados

Calib. Valid. Calib. Valid. Calib. Valid.

hydrometric stations ALBANELLA ALBANELLA CALORE PERSANO
n° Data (hours) 55386 35064 19730 17147 16989 19730
Nash-Sutcliffe 0.940 0.890 0.945 0.926 0.910 0.892
correlation coeff. 0.967 0.955 0.972 0.963 0.941 0.947
hydrometric stations SICIGNANO AULETTA CONTURSI
n° Datos (horas) 24395 10970 19730 15635 46034 48489
Nash-Sutcliffe 0.911 0.863 0.938 0.904 0.899 0.913
correlation coeff. 0.955 0.930 0.968 0.952 0.941 0.943
Albanella – Calibrazione
TOPKAPI-X + ANN + MCP TOPKAPI-X + MCP
t0 3h 6h 9h 12h t0 3h 6h 9h 12h
BIAS 0.067 0.08 -0.151 -0.55 -1.008 0.049 0.015 -0.208 -0.555 -0.975
NASH 0.999 0.989 0.98 0.969 0.947 0.999 0.986 0.97 0.951 0.927
Assessing Predictive Uncertainty DEV ST 2.889 8.011 10.983 13.821 17.908 2.795 9.194 13.393 17.259
Albanella - Validazione
21.104

TOPKAPI-X + ANN + MCP TOPKAPI-X + MCP

t0 3h 6h 9h 12h t0 3h 6h 9h 12h
BIAS 0.428 0.654 0.45 0.229 -0.056 0.057 0.084 -0.123 -0.567 -1.047
NASH 0.998 0.99 0.981 0.973 0.958 0.999 0.989 0.977 0.961 0.938
The Sele FFS includes of a Bayesian method, the Model Conditional Processor (MCP), for assessing predictive DEV ST 2.744 6.776 9.079 10.961 13.726 2.529 7.061 10.104 13.243 16.724
uncertainty in real time from the forecasts made with TOPKAPI-X model (forced with the COSMO LAM family) andAlbanella – Calibrazione
ANN model. TOPKAPI-X + ANN + MCP TOPKAPI-X + MCP
t0 3h 6h 9h 12h t0 3h 6h 9h 12h
BIAS 0.067 0.08 -0.151 -0.55 -1.008 0.049 0.015 -0.208 -0.555 -0.975
Real time flood forecasting requires predictive uncertainty to be taken into account due to
NASH a number
0.999 0.989 of reasons.
0.98 0.969 0.947 0.999 0.986 0.97 0.951 0.927
Deterministic hydrological/hydraulic forecasts give useful information about real future
DEV STevents,
2.889 but they 10.983
8.011 can not be 17.908 2.795 9.194 13.393 17.259
13.821 21.104
taken and used as real future occurrences, as usually done in practice, since they are affected by a number ofAlbanella - Validazione
TOPKAPI-X + ANN + MCP TOPKAPI-X + MCP
errors, ranging from observation, boundary conditions and modelling errors. Model forecasts t0 3h
are inherently
6h 9h 12h t0 3h 6h 9h 12h
uncertain and they must be used to reduce the uncertainty of decision-makers onBIAS future 0.428
occurrences.
0.654 Predictive
0.45 0.229 -0.056 0.057 0.084 -0.123 -0.567 -1.047
uncertainty (PU) is in fact defined as the probability of occurrence of a future valueNASH
of a predictand
0.998 0.99(such as water
0.981 0.973 0.958 0.999 0.989 0.977 0.961 0.938
level, discharge or water volume) conditional upon prior observations and knowledge DEV ST 2.744 6.776as 9.079
as well on all10.961
the 13.726 2.529 7.061 10.104 13.243 16.724
information we can obtain on that specific future value from model forecasts. PU must be quantified in terms of a Albanella – Calibrazione
probability distribution function which will be used by the emergency managers in their decision process in order to TOPKAPI-X + ANN + MCP TOPKAPI-X + MCP
improve the quality and reliability of their decisions. The proposed MCP approach allows predictive uncertainty to t0 3h 6h 9h 12h t0 3h 6h 9h 12h
BIAS 0.067 0.08 -0.151 -0.55 -1.008 0.049 0.015 -0.208 -0.555 -0.975
be assessed, combining both the meteorological and hydrological uncertainties, by deriving, at various forecast
NASH 0.999 0.989 0.98 0.969 0.947 0.999 0.986 0.97 0.951 0.927
lead times, the joint probability density function of the observations and of one (or several) model(s) forecasts. This DEV ST 2.889 8.011 10.983 13.821 17.908 2.795 9.194 13.393 17.259 21.104
is performed by converting observations and model forecasts in a multi-Normal space using the Normal Quantile Albanella - Validazione
Transform. This work also shows the possibility of using a joint Truncated Normal Multivariate distribution, with the TOPKAPI-X + ANN + MCP TOPKAPI-X + MCP
t0 3h 6h 9h 12h t0 3h 6h 9h 12h
aim of improving adaptation to low flows and high flows. The Sele FFS also deals with the combination of model
BIAS 0.428 0.654 0.45 0.229 -0.056 0.057 0.084 -0.123 -0.567 -1.047
forecasts of a different nature, including meteorological, hydrological and hydraulic models, with the aim of taking NASH 0.998 0.99 0.981 0.973 0.958 0.999 0.989 0.977 0.961 0.938
advantage of their different capabilities. DEV ST 2.744 6.776 9.079 10.961 13.726 2.529 7.061 10.104 13.243 16.724

European Geosciences Union (1) Idrologia e Ambiente s.r.l., Via Stazio 66, 80123, Napoli, Italy
Session: HS4.3/AS1.3/NH1.3
General Assembly 2015 (2) Princeton University, Princeton, NJ (USA)
Ensemble hydro-meteorological forecasting (co-organized)
Vienna | Austria | 12 – 17 April 2015 (3) BiGeA —University of Bologna, Bologna, Italy