An Artificial Neural Network-based Real Time DSS to Manage the

Discharges of a Wastewater Treatment Plant and Reduce the

Flooding Risk

Loris Francesco Termite

, Emanuele Bonamente

, Alberto Garinei

3,4

, Daniele Bolpagni

Lorenzo Menculini

, Marcello Marconi

3,4

, Lorenzo Biondi

3,4

, Andrea Chini

and Massimo Crespi

K-Digitale S.r.l., Perugia, Italy

Department of Engineering, University of Perugia, Perugia, Italy

Department of Sustainability Engineering, Guglielmo Marconi University, Rome, Italy

Idea-Re S.r.l., Perugia, Italy

A2A Ciclo Idrico S.p.A., Brescia, Italy

Radarmeteo S.r.l., Due Carrare (PD), Italy

Keywords: Decision Support System, Artificial Neural Networks, Flood Management, Flood Forecasting, Smart

Infrastructures.

Abstract: An approach for sewerage systems monitoring based on Artificial Neural Networks is presented as a feasible

and reliable way of providing operators with a real-time Decision Support System that is able to predict critical

events and suggest a proper mitigation strategy. A fully-working prototype was developed and tested on a

sewerage system in the city of Brescia, Italy. The system is trained to forecast flows and water levels in critical

points of the grid based on their measured values as well as rainfall data. When relying on observed rainfall

only, key parameters can be predicted up to 60 minutes in advance, whereas including very-short-term

Quantitative Precipitation Estimates – nowcasting – the time horizon can be extended further, up to 140

minutes in the current case study. Unlike classical hydraulic modelling, the proposed approach can be

effectively used run-time as the execution is performed with a negligible computational cost, and it is suitable

to increase safety measures in a Smart City context.

1 INTRODUCTION

In many urban scenarios, the management of

wastewater is performed by combined sewer systems,

collecting both storm water and black/greywater in

order to send them to the wastewater treatment plant

(WWTP). In case of intense rainfall events, the sewer

discharge may reach its capacity limit and

occasionally produce combined sewer overflow

(CSO), leading to pollution risk. Moreover, excessive

discharges of treated water may produce floods if the

stage of the receiving bodies is already high due to

rainfall. Specific structures – spillway gates, buffer

tanks, pumps, etc – are designed to mitigate such

occurrences. Clearly, an optimal intervention strategy

during a critical event strongly relies on the accurate

knowledge of the system status (water flows and

levels, gate openings) and, possibly, the weather

conditions (i.e. observed and expected rainfalls). The

lack of a comprehensive monitoring and forecast

network requires the operators to choose the

intervention strategy by relying mostly on their

experience. In such contexts, a decision support

system (DSS) may represent a fundamental aid

(Pereira et al., 2019).

DSSs are platforms dedicated to providing

support to the human operator in deciding which

operations to perform. Several examples are available

in the literature for sewerage systems-related

applications (Park and Kim, 2013; Rao, 2015), their

focus being on the design, renovation, and upgrade of

the physical system or on the offline reanalysis of past

events in order to optimize future intervention

strategies. Instead, the monitoring of sewerage

networks using real-time data has been mainly

focused on pollution emissions and concentrations

(Rechdaoui-Guérin et al., 2018).

Termite, L., Bonamente, E., Garinei, A., Bolpagni, D., Menculini, L., Marconi, M., Biondi, L., Chini, A. and Crespi, M.

An Artiﬁcial Neural Network-based Real Time DSS to Manage the Discharges of a Wastewater Treatment Plant and Reduce the Flooding Risk.

DOI: 10.5220/0010396500150026

In Proceedings of the 10th International Conference on Smart Cities and Green ICT Systems (SMARTGREENS 2021), pages 15-26

ISBN: 978-989-758-512-8

In recent years, smart network monitoring began

to find development and applications. The advances

in technology can be exploited by municipalities,

utilities, and related organizations to implement smart

data infrastructure for wet weather control. More

specifically, advanced monitoring data are used to

support wet weather control and decision-making in

real time or near real time. The United States

Environmental Protection Agency defines smart

infrastructures as “the integration of emerging and

advanced technology to enhance the collection,

storage, and/or reanalysis of water-related data”,

making use of “hardware, communication and

management analytics to provide real and tangible

benefits to utilities”, as “maximizing existing

infrastructure and optimizing operations and

responses to be proactive, not reactive” (US EPA,

2018). Smart infrastructures are generally

implemented in connection with a supervisory control

and data acquisition (SCADA) system. Their main

objective is usually the real-time monitoring of

CSO’s flow rates and effluent level, in order to assess

potential flooding and pollution incidents and to

support real-time or quasi real-time decision making

about actions to be taken (Botturi et al., 2020).

However, to the best of the authors’ knowledge –

with particular reference to the Italian territory – the

diffusion of smart infrastructures is still quite limited.

Despite the presence of few studies aimed at reducing

CSO through Real Time Control of gates and sluices

(Carbone et al., 2014; Campisano et al., 2016), some

issues are not addressed properly, as the potential risk

of flooding due to excessive discharge of treated

water from the WWTP to the receiving body. Most of

real-world DSS implementations aimed at flood

mitigation still rely on traditional hydraulic models

and are often characterized by a high computational

demand making them unsuitable for real-time usage.

Sometimes such models proved to be outdated and

are not resilient to climate change. Moreover, there is

lack of use of very-short-term precipitation forecasts,

which are crucial in case of high-intense and short-

lasting rainfall events.

This paper presents a case study about the design

of a DSS dedicated to WWTP management. It allows

real-time monitoring of the system status through a

network of ground-based meters and weather radars.

It also includes a forecasting tool that provides

additional information to help operators in planning

manoeuvres, thanks to a methodology based on

Artificial Intelligence – specifically Artificial Neural

Networks (ANNs) (Maier et al., 2010) – estimating

water levels and flows inside the network. ANNs are

suitable tools for the purposes of this study, since they

allow for a reduction in the input data variety –

selecting only the most statistically significant in the

input-output relation – and the non-use of parameters

needed by traditional models, as for example the soil

properties in rainfall-runoff modelling. Moreover,

after being calibrated, their execution is immediate

and perfectly matches the real-time requirements.

They can also be easily updated, well facing possible

changes in the real environment.

The aim of the implemented DSS is to give

operators useful insights about the current status of

the receiving water body in strategic locations, the

potential CSO amount, the WWTP inflow, the current

and recent rainfall and the estimated evolution of the

strategic variables, collecting all data and showing

them in a single User Interface, to eventually guide

them to apply a proper risk mitigation strategy.

As regards rainfall information, the DSS also

exploits high resolution very-short-term quantitative

precipitation forecasting, also known as

“nowcasting” (Wilson et al., 1998). Quantitative

precipitation forecasting is performed by means of

algorithms working on rainfall measures provided by

weather radars (Bellon et al., 2010; Lee et al., 2010).

It has been proved that radar-based nowcasting can

give more reliable results than classic numerical

weather prediction within a time horizon of 3-8 hours

(Mandapaka et al., 2012), being therefore a valuable

tool for real-time decision-making. In the present

context, nowcasting is used to further extend the time

horizon of ANNs predictions for hydraulic variables,

allowing to manage the potential emergencies with a

reasonable advance.

The designed DSS provides real-time monitoring

of the hydraulic system status and – thanks to ANNs

– predicts its future evolution. Once the model

calibration has been performed, the ANNs algorithms

consist in linear algebra operations on matrices which

require negligible computational times, unlike more

traditional hydraulic models that as of today require

large computational times and computing resources

(Clark et al., 2017).

2 METHODOLOGY

2.1 Description of the System Layout

The WWTP for which the DSS was developed

collects the sewage and urban runoff in the city of

Brescia (Italy) and its surrounding area, serving a

total of 296.000 inhabitants over a 146 km

surface.

Figure 1 shows the system layout.

SMARTGREENS 2021 - 10th International Conference on Smart Cities and Green ICT Systems

Figure 1: System layout.

An urban drainage channel named Vaso Fiume

(VF) runs parallel to the final sewer collector,

upstream the WWTP. When the sewer flow raises

above a critical level, the excess is directed into the

VF trough two lateral spillways, one located just

before the WWTP and one a few km upstream. The

VF also receives the treated wastewater. A bigger

urban drainage channel, named Vaso Garzetta (VG),

collects water from the VF approximately 2.4 km

downstream and eventually flows into the Mella river

after a 3.4 km path. The whole network of receiving

water bodies is characterized by potential hydraulic

risk in case of extreme rainfall events: over the last

years urban flooding has occurred several times along

the VG channel, in a critical location corresponding

to the gz_lt01 measuring point (approximately 1.7 km

downstream the confluence with VF). In such

circumstances, the water flowing from VF into VG

should be constrained and the WWTP technicians

must pay close attention in order to minimize the risk

of flooding. To this aim, the VF channel is equipped

with a series of inline or lateral gates, which can be

used effectively for flood prevention. Two lateral

gates (g10 and g15), which are closed in standard

conditions, can be gradually opened to direct the flow

in the surrounding rural area, within environmental

law limitations. A couple of paired inline gates (g14a

and g14b), which are kept open in standard

conditions, can be progressively closed to reduce, and

in certain cases completely arrest, the VF flow into

the VG. A storage area is also present next to the

WWTP and is used to reduce the peak discharge into

the VF: an inline gate (g2) can be partially closed in

order to increase the upstream water level and

activate a lateral spillway into the storage area.

The WWTP and the sewage network are managed

by a leading multiutility society which has been

playing much effort in Smart City projects over the

last years. The wish of the WWTP managers was to

have a tool able to assist them in taking proper

decision especially when dealing with severe rainfall

events. Indeed, prior to the development of the

present DSS platform, the gate-opening strategy

relied only on operator experience – based on

information from the upstream gz_lt02 meter – to

foresee impending flood waves. Operators decided

whether to use the storage area and to open lateral

spillways in order to reduce the outflow towards the

VG channel and prevent downstream overflows.

However, the users lacked an interface allowing a

thorough monitoring of the system status. In

particular, information about forthcoming rainfall

amounts, water levels and sewer flows was

completely missing. In case of severe events, such an

approach was not able to completely avoid risks.

In addition to the VG level at gz_lt02, particular

attention is also devoted to the WWTP inlet (ml01),

as treatment cycles can be optimized thanks to the

presence of an internal buffer tank. Since during

extreme events the flow at ml01 may exceed the

WWTP processing capacity, knowing this flow in

advance can help to properly manage the internal

operation of the plant. Therefore, the VG level at

gz_lt02 and the flow at ml01 are key parameters to be

monitored. Their expected values are forecasted by

the ANNs developed for the DSS.

2.2 The Decision Support System

2.2.1 Data Acquisition

Prior to the DSS implementation, the acquired data

(rainfall, levels, flows, gate openings) were collected

in diverse databases and visualized in separate

interfaces, preventing the WWTP operators from

having a comprehensive view of the system status.

Rainfall data used to be provided only by six

pluviometers within the catchment; for the purposes

of the present study it was instead decided to also

exploit advanced meteorological radar

measurements. To this end, distributed rainfall

measurements performed by radars were calibrated

through the punctual pluviometer records, eventually

obtaining reliable information in terms of both value

accuracy and spatial variability. The resulting product

has 1×1 km spatial resolution as in Panziera et al.

(2011) and a 10-minutes time resolution.

The resulting integrated rainfall measurements

were spatially averaged over the whole catchment to

obtain a single value to be used as an input for the

ANNs. Preliminary attempts demonstrated that

distinguishing over rainfalls precipitated in different

zones only makes water levels and flows forecasts

An Artiﬁcial Neural Network-based Real Time DSS to Manage the Discharges of a Wastewater Treatment Plant and Reduce the Flooding

Risk

less stable and does not produce any significant

improvement over using the mean areal rainfall.

The readout of each sensor was synchronized, the

time step was fixed to 10 minutes, and a dedicated

server was realized to host the system database

containing all the records of measured data, computed

quantities and final parameters.

2.2.2 ANNs Setup

Rainfall, gz_lt02 level and ml01 discharge data were

made available for a period spanning from October

2016 to August 2018.

A set of four ANNs was set up, two predicting VG

level variation and the other two predicting sewer

flow variation.

The effect produced on target variables by rainfall

is visible with a lag time that was found to vary

between 30 minutes and 2 hours. Thus, it was decided

to predict them up to 60 minutes beyond the last

known rainfall information, whether measured or

predicted.

More specifically, the ANN named ANN

(level/short-term) makes use of measured rainfalls

only and provides gz_lt02 level forecast up to 60

minutes; ANN

(level/long-term) uses also

nowcasting up to +80 minutes, thus extending the

forecast horizon to 140 minutes. Similarly, ANN

(flow/short-term) and ANN

(flow/long-term)

forecast sewer discharge at ml01. ANN

and ANN

were trained and validated using 18 suitably trimmed

rainfall events in the analysed period, corresponding

to a total 1714 datapoints; ANN

and ANN

used 17

events, with 12057 datapoints. Due to the available

data, the two samples refer to different sets of events.

The higher number of datapoints used to train sewer-

related ANNs depends on the fact that events were

trimmed in larger chunks, because flows at ml01 take

longer to return to the unperturbed value after

rainfalls with respect to levels at gz_lt02.

Measured rainfall inputs for ANN

and ANN

cover the antecedent 2-hours interval. This extension

was considered appropriate, covering the whole range

of observed lags. A longer period (six hours) was

required for ml01 forecast, due to the longer-lasting

observed perturbation induced on the sewer collector

by precipitation.

Rainfall measurements are updated every 10

minutes, while nowcasting is updated every 20

minutes due to computing limitations. This results in

two different updating frequencies for short-term and

long-term predictions, consistent with rainfall input

updates. For convenience, the rainfall measurements

provided to the ANNs are also aggregated in 20-

minute bins, although updated every 10 minutes.

Moreover, to keep track of the initial conditions, the

current level/flow is also used as input.

For all the ANNs, the targets are the level/flow

variations induced by rainfall, with respect to the

current value, evaluated on a 10-minute basis. Thus,

at every execution, ANN

and ANN

produce six

outputs (10-minute bins from +10 to +60), while

ANN

and ANN

produce fourteen outputs (10-

minute bins from +10 to +140).

The expected levels are obtained by the algebraic

sum of current values and predicted variations. The

expected flows calculation needs an additional term

to be considered in the sum, i.e. the characteristic

daily modulation of sewer discharge. Thus, the

average flow profile at ml01 was obtained

disregarding rainy days and was found to lie in the

range 0.5-1.2 m

/s (Figure 2), then the corresponding

144 average flow variations – on a 10-minute basis

from 0:00 to 23:50 – were computed.

Figure 2: Average ml01 profile during dry days (blue line).

The Multilayer Perceptron structure was chosen

to build the ANNs, that were developed in a

MATLAB environment. After trial-and-error

attempts, the layout was chosen so that all the ANNs

share a common structure, with a 20-nodes single

hidden layer connecting the input and output layers.

All the inputs and targets were normalized between 0

and 1. A logistic activation function is used in the

hidden layer, and a linear activation function is used

in the output layer to produce the results. It is

customary in Machine Learning to use three different

sets of data, namely the calibration, validation and

testing set: the calibration and validation set are used

to train the networks, with the first used to fix weights

and biases and the latter to adjust the

hyperparameters; the testing set is used to test the

ANNs performance on unseen data. However, due to

some uncertainties in the available data, and since the

aim of this study was not to explore ANNs’ theory but

instead it was necessary to exploit as much data as

possible to build a ready-to-use product, it was

SMARTGREENS 2021 - 10th International Conference on Smart Cities and Green ICT Systems

decided to disregard the testing set. Therefore, the

available datasets were split using the last four events

(380 datapoints, i.e. 22.17% for level forecast; 2274

datapoints, i.e. 18.86% for flow forecast) for the

validation set and the previous ones for the calibration

set.

Weights and biases in the ANNs nodes were

randomly initialized, and the Levenberg-Marquadt

backpropagation algorithm was used to minimize the

cost function, specifically the mean square error

between the target/output pairs; the algorithm

execution was imposed to stop if the validation error

increased for 20 consecutive iterations. Since the

training procedure outcome varies depending on the

randomly generated initial parameters and on the

chance of the training algorithm getting stuck in local

minima, each ANN was trained 2000 times using the

above-mentioned procedure, and the best-performing

network was then selected.

Rather than considering the cost function, a more

detailed multi-objective optimization was

implemented to select the best performing ANNs,

keeping in mind their final purpose, i.e. a correct and

prompt forecast of the most severe events. Thus, for

both level and flow, three threshold values were

defined, i.e. 70, 100 and 140 cm at gz_lt02 and 2.5,

3.5 and 4.5 m

/s at ml01. Four objectives were defined

to select the best performing ANNs.

1) The Nash-Sutcliffe Efficiency index (NSE)

computed on actual network targets and outputs

(normalized values). The optimization variable to

be minimized is the subtracted ratio in the NSE

definition:

𝑁𝑆𝐸 = 1 −

∑

𝑂



−𝑆











∑

𝑂



−𝑂











(1)

where 𝑂



are the observed values, 𝑂



is their mean

value and 𝑆



are the simulated values. NSE is

computed for each prediction horizon (10 minutes, 20

minutes, …etc.) and the objective is found from the

mean of the computed values.

2) Maximum number of correct predictions of

threshold values crossing. The optimization

variable, to be minimized, is the ratio of missed

predictions to observed crossings. It is computed

for each alert level and the objective is set to the

mean of the three computed values.

3) Minimum number of fake predictions (threshold

crossing prediction not corresponding to observed

crossing). The optimization variable, to be

minimized, is the ratio of fake predictions to total

predictions, either true or false. It is computed for

each alert level and the objective is set to the mean

of the three computed values.

4) Optimal prediction timing. Every time there is an

observed threshold level crossing in the forecast

time horizon after current timestamp and there is

also a predicted crossing, the delay between the

observed and predicted time of crossing is

computed. The objective to minimize is the mean

squared delay for all alert levels.

Objectives n.2 and n.3 vary between 0 and 1. In order

to give objective n.1 the same range of variability, the

upper boundary of the subtracted ratio was set to 1, as

values greater than 1 would imply a non-acceptable

performance and the related solution should be

discarded. Objective n.4 was normalized between 0

and 1 with respect to its possible minimum and

maximum values (i.e. 0 and 50 minutes for short-term

predictions and 0 and 130 minutes for long-term

predictions). In looking for the Pareto front of

undominated solutions, the second and fourth

objectives were given a weight triple than the others.

This was mainly due to sewer management reasons.

Among these points in the 4-D resulting spaces, the

best performing combinations were selected as those

with the minimum Euclidean norm and the

corresponding sets of weights and biases matrices

were used in the algorithms running in the DSS.

Since objectives n.2, n.3 and n.4 can be computed

only on threshold crossings – and therefore on a small

amount of data – the above-mentioned optimization

procedure was performed on the whole available

dataset, comprising both training and validation

datapoints.

3 RESULTS

3.1 ANNs Performance

The multi-objective optimization led to the selection

of the best performing ANNs. A first evaluation of

their performance was made according to the

objectives described in section 2.2.2. The obtained

values are shown in Table 1. A better general

performance of level prediction with respect to sewer

discharge prediction can be noticed. Moreover, better

predictions are obtained from ANNs exploiting

nowcasting information.

An Artiﬁcial Neural Network-based Real Time DSS to Manage the Discharges of a Wastewater Treatment Plant and Reduce the Flooding

Risk

Table 1: Results of the multi-objective optimization: values

for the selected ANNs.

OBJ 1 OBJ 2 OBJ 3 OBJ 4

ANN

0.241 0.150 0.106 0.061

ANN

0.128 0.137 0.039 0.037

ANN

0.531 0.175 0.117 0.100

ANN

0.412 0.201 0.069 0.083

Statistical analyses were performed on the results.

For all the selected ANNs, the NSE index was

calculated for each event and each forecasting

horizon (+10 to +60 minutes or +10 to +140 minutes).

Calibration values for ANN

– averaged on the 14

events in the set – range between 0.484 and 0.888,

with higher values corresponding to shorter

forecasting horizons. Similarly, average validation

values are comprised between 0.294 and 0.906.

ANN

gave NSE average values ranging between

0.301 and 0.601 for calibration events and between

0.103 and 0.903 for validation ones. The ranges of

average NSE values are 0.987÷0.909 and

0.904÷0.989 for ANN

calibration and validation

events, respectively. Finally, NSE average ranges for

ANN

are 0.843÷0.986 (calibration) and

0.789÷0.988 (validation). As a term of comparison,

Jeong et al. (2010) obtained NSE = 0.74 for

calibration data and NSE = 0.63 for validation data

when modelling stream flows in a small watershed

using the SWAT tool. However, this kind of analysis

does not give useful insights on the ANNs efficiency

in the DSS and may even be misleading. Indeed, the

main goal of a real-time DSS like as the one presented

Table 2: Performance of the selected ANNs in predicting threshold crossings. Values outside brackets refer to the whole

dataset; values in brackets refer to calibration and validation sets, respectively.

Threshold 1 Threshold 2 Threshold 3

ANN

n. observed. 22 (17 / 5) 9 (6 / 3) 4 (2 / 2)

n. predicted. 22 (17 / 5) 6 (4 / 2) 3 (2 / 1)

predicted % 100.0 (100.0 / 100.0) 66.7 (66.67 / 66.67) 75.0 (100.0 / 50.0)

a.w.t. (min) 36.8 (35.2 / 42.0) 25.0 (27.5 / 20.0) 33.3 (35.0 / 30.0)

a.p.d. (min) 3.2 (3.6 / 2.0) 1.7 (0.0 / 5.0) -3.3 (0.0 / -10.0)

false alerts 10 (7 / 3) 1 (1 / 0) 0 (0 / 0)

ANN

n. observed. 22 (17 / 5) 9 (6 / 3) 4 (2 / 2)

n. predicted. 22 (17 / 5) 5 (3 / 2) 3 (2 / 1)

predicted % 100.0 (100.0 / 100.0) 55.6 (50.0 / 66.67) 75.0 (100.0 / 50.0)

a.w.t. (min) 97.7 (91.8 / 118.0) 116.0 (110.0 / 125.0) 113.3 (115.0 / 110.0)

a.p.d. (min) -7.3 (-5.9 / -12.0) -10.0 (-3.3 / -20.0) 3.3 (0.0 / 10.0)

false alerts 9 (6 / 3) 4 (2 / 2) 0 (0 / 0)

ANN

n. observed. 56 (45 / 11) 11 (7 / 4) 2 (1 / 1)

n. predicted. 34 (31 / 3) 7 (4 / 3) 2 (1 / 1)

predicted % 60.7 (68.9 / 27.3) 64.7 (57.1 / 75.0) 100.0 (100.0 / 100.0)

a.w.t. (min) 35.9 (36.4 / 30.0) 22.9 (22.5 / 23.3) 45.0 (30.0 / 60.0)

a.p.d. (min) 0.0 (0.3 / -3.3) 5.7 (15.0 / -6.7) 5.0 (20.0 / -10.0)

false alerts 16 (10 / 6) 2 (1 / 1) 0 (0 / 0)

ANN

n. observed. 56 (45 / 11) 11 (7 / 4) 2 (1 / 1)

n. predicted. 36 (32 / 4) 9 (5 / 4) 2 (1 / 1)

predicted % 64.3 (71.1 / 36.4) 81.8 (71.4 / 100.0) 100.0 (100.0 / 100.0)

a.w.t. (min) 71.4 (69.7 / 85.0) 73.3 (98.0 / 42.5) 100.0 (70.0 / 130.0)

a.p.d. (min) 8.9 (8.4 / 12.5) 5.6 (2.0 / 5.0) 15.0 (30.0 / 0.0)

false alerts 15 (5 / 10) 2 (2 / 0) 0 (0 / 0)

SMARTGREENS 2021 - 10th International Conference on Smart Cities and Green ICT Systems

in this case study is to guarantee that accurate alerts

are sent sufficiently in advance, allowing operators to

act promptly. Therefore, the performances of the

selected ANNs were also evaluated according to the

metrics defined in Table 2 (please note that values

outside brackets refer to the whole datasets, while

first and second values in brackets refer to calibration

and validation sets, respectively). For each ANN, the

total number of observed crossings of the defined

threshold values (70, 100 and 140 cm at gz_lt02; 2.5,

3.5 and 4.5 m

/s at ml01) is displayed, along with the

number of corresponding predictions and related

percentages of correctly predicted crossings. Given

the DSS updating frequency, observed crossings can

be forecasted with an anticipation varying (with a 10-

minute resolution) from 60 to 10 minutes for ANN

and ANN

and from 140 to 10 minutes for ANN

and

ANN

. A crossing is considered predicted if it is

signalled to occur at least once in the available

forecasting horizon. Two other parameters that were

evaluated are the average warning time (a.w.t. in

Table 2), i.e. the mean anticipation corresponding to

the first alert of impending crossing, and the average

prediction delay (a.p.d. in Table 2), indicating the

accuracy of the prediction timing. As an example, if

the DSS first signals an impending crossing by

warning that it will happen after 40 minutes, but the

actual crossing is observed after 30 minutes, then the

warning time is 30 minutes and the prediction delay

is 10 minutes. Finally, the number of false alerts is

shown. In contrast to correct crossing predictions, an

alert is considered false if there is no actual crossing

at any time step of the forecasting horizon.

From Table 2 it is possible to see that all the ANNs

can predict the majority of the most severe gz_lt02

level occurrences: threshold 3 crossings are predicted

three times on four occurrences, with the missed

prediction referring to a validation event. The

validation set comprises only two crossings of the

higher threshold and the prediction percentage is

50%, but more data would be necessary to better

assess the performance. All the lower threshold

crossings are predicted, while some are missed for

the intermediate one, and for both the performances

on the calibration and validation sets are comparable.

The average warning time ranges between

approximately 25 and 40 minutes for short-term

predictions, while the use of nowcasting information

allows to increase the forecast anticipation to

approximately 2 hours. The absolute value of the

average prediction delay is generally lower than 10

minutes, the only higher value being the -20 minutes

delay obtained for ANN

with respect to the two

predicted Threshold 2 crossings in the validation set

(i.e. the signalled crossing times are, on average, 20

minutes early with respect to the actual ones). As

regards sewer flow predictions, they are again slightly

outperformed by level forecasts. This may be due to

observed ml01 flows during rainfall events being less

Figure 3: ANN

envelop of level forecast for a validation

event.

An Artiﬁcial Neural Network-based Real Time DSS to Manage the Discharges of a Wastewater Treatment Plant and Reduce the Flooding

Risk

regular with respect to those of the VG level. The

highest flow threshold crossings are always predicted,

even if they occur just once in the calibration events

and once in the validation events. The average

warning time and prediction delay are comparable to

the ones obtained for level forecasts. All models

generate some false alerts, especially for the lowest

threshold, while never for the highest one.

Figure 4: Level 3 crossing prediction efficiency from 140 to 10 minutes before the first occurrence during a validation event.

SMARTGREENS 2021 - 10th International Conference on Smart Cities and Green ICT Systems

Some examples of the prediction performance on a

severe event in the validation set, during which all the

thresholds are crossed, are shown in Figures 3 and 4.

Figure 3 shows, for all the short-term prediction

horizons, the envelop of the forecasted level vs the

observed one. Figure 4 shows, for the same event, a

sort of “snap-shot” of the User Interface (UI) from 140

to 10 minutes prior to the first crossing of the L3

threshold. As in the actual UI described in next section,

the time axis spans from 120 minutes prior to the

current time to 180 minutes after. Measured and

forecasted rainfalls and levels are shown. The observed

future level is also shown for comparison purpose. As

it can be noticed, the first alert is given 110 minutes

before the actual crossing, thanks to the long-term

predicting ANN (red line), even if the crossing is

signalled to occur within 120 minutes. 80 minutes

before the crossing, the amount of expected level

becomes more reliable. The short-term prediction (blue

line) raises above L3 only 30 minutes before the

crossing. Indeed, up to 40 minutes before the crossing,

the amount of measured rainfall is still quite low, and

consequently ANN

predicts just a small level raise.

The considerable rainfall intensity increase that is

measured 30 minutes before the crossing is finally

allowing ANN

to predict a steep raise of gz_lt02 level.

This example underlines the importance of nowcasting

information, that permits to have alerts of incoming

critical events with a reasonable advance.

3.2 User Interface

The DSS was designed to provide a comprehensive

view of the sewerage network, with a particular focus

on the two key parameters defined in section 2. Thus,

a UI was created. Specifically, it was developed as a

QGIS plugin, to allow further improvements using

georeferenced data. Weights and biases were

extracted from the selected ANNs allowing to write

real-time running forecast algorithms consisting in

in linear algebra operations on matrices. These

algorithms, together with all the other necessary

scripts running behind the DSS, were developed

through the Python language. All measured and

processed data are stored in the DSS database. From

there, data are picked to be shown in the User

Interface. In particular, the UI (Figure 5) shows

directly measured quantities (e.g. current values of

levels along the VG channel, VF gate openings,

flows) together with derived quantities (as for

example the estimated CSO downstream ml01 that,

given the difficulty in placing a meter due to

morphological issues, was estimated by means of

water mass balance between the collector flow and

the internal WWTP flow during rainfall events). The

UI also can display the results of simulations of the

VF channel behaviour using different mitigation

strategies, as explained later on. The main section is

dedicated to the visualization of the VG levels at

gz_lt02 (blue lines) and ml01 discharge (orange lines)

as recorded for the past 6 hours (continuous lines) and

predicted for the next 60 minutes (dashed lines) or

140 minutes (dotted lines). Above the main plot, the

current and suggested openings for the 5 operable

gates (labelled as Nr.2, Nr.10, Nr.14A, Nr.14B, and

Nr.15 in the UI) is also shown. The gate-opening

strategy is driven by the measured VG level at

gz_lt02. The three threshold values defined for the

Figure 5: User Interface in the QGIS plugin.

An Artiﬁcial Neural Network-based Real Time DSS to Manage the Discharges of a Wastewater Treatment Plant and Reduce the Flooding

Risk

Figure 6: Real-time Grafana visualization dashboard.

VG level (L1=70 cm, L2=100 cm, L3=140 cm)

trigger four different combinations of suggested gate

openings (namely “A”, “B”, “C” and “D” in the UI)

that are characterized by an increasing quantity of

spilled water and storage area usage, resulting in a

decreasing discharge into the VG channel until it is

completely blocked in case most severe events. The

suggested strategy is highlighted in yellow and the

current openings are highlighted in green or red,

depending on whether they are in accordance or not

with the proposed ones, with a 5 cm tolerance. At this

first stage, the suggested strategies are based on the

long-time experience of the WWTP technicians.

However, the “Simulation” section of the UI, below

the main plots, embeds a physically-based HEC-RAS

hydraulic model of the VF (not discussed here for the

sake of brevity) that can be executed between two

selected timestamps using past boundary conditions.

Thus, by changing the threshold values of the VG

level or the suggested gate openings, ex-post analyses

can be performed with the final aim of assessing the

goodness of the adopted strategies or detecting more

efficient parameter combinations.

The described UI, allowing to change the DSS

parameters, is thought to be used by expert operators.

In addition, for all other operators, a visualization-

only UI was developed through the Grafana platform.

It allows data visualization on desktop and mobile

devices and is composed of three dashboards, the

main one showing the current system status (Figure

6), another showing past data in a selected time

interval and the last showing “snap-shots” of the main

dashboard plots at a selected past timestamp.

4 DISCUSSION

The results presented in Section 3 indicate that the

developed DSS is already able to give useful insights

to the WWTP operators and to help them in managing

potentially critical events.

The real-time monitoring of measured quantities

gives useful advice to the operator: when the crossing

of a threshold level happens, the flood wave takes

some time (approx. 30 to 90 minutes) to reach the

overflow point, allowing for a timely intervention to

mitigate the risk of overflows. Furthermore, the

system status forecast provides operators with

additional information, guiding them in taking the

proper decision. As an example, if an alert in the UI

suggests a specific strategy related to measured

gz_lt02 threshold crossing, but the forecast is

showing that level is going to decrease soon after, the

operator may decide not to adopt any mitigation

strategy. On the contrary, forecasts of incoming

critical conditions allow to be ready and possibly

anticipate the mitigation manoeuvres. The average

warning time is in fact sufficient for operators to be

ready to adopt proper risk mitigation strategies. In

particular, the most critical events are predicted with

an average warning time of approximately 30 minutes

based on observed rainfall only, while nowcasting

SMARTGREENS 2021 - 10th International Conference on Smart Cities and Green ICT Systems

allows for an alert anticipation greater than 100

minutes. The prediction timing is fairly accurate, in

particular for VG level variation, with an average

delay almost always lower than 10 minutes. Based on

these considerations, the implemented ANNs can be

considered helpful tools in the developed DSS.

However, since this is a first stage project, several

improvements may be implemented. First, the ANNs

performance may be enhanced as long as more data

become available, being all collected in the DSS

database. Also, different ANN structures may be

implemented. Recurrent neural networks as the Long

Short-Term Memory (LSTM) ones (Hochreiter and

Schmidhuber., 1997) are particularly suited in

modelling time series and an attempt could be made

using such ANN architecture. Given the DSS

structure, such an operation would not be

complicated. New ANNs can be trained offline, then

if better performances are gained the scripts may be

easily adjusted. Moreover, at the current stage it was

decided not to use a testing set to evaluate the ANNs

performance. The future availability of more records

will allow to perform analyses on unseen data.

Another aspect that should be investigated is

related to rainfall forecasts. As already said,

nowcasting allows to generate alerts with a

reasonable advance. However, in the present study

measured rainfalls were used also as predicted ones,

in a perfect forecast hypothesis. Obviously the actual

nowcasting information could be less accurate and the

effects on level and flow predictions should be

evaluated. This is why it was chosen to keep also the

short-term ANNs in the DSS, as they rely on

measured data which are not affected by

uncertainties.

As regards the risk mitigation strategies, they are

currently suggested in the UI based on the level of the

receiving channel. The four different gate openings

combinations, associated to three VG thresholds

levels, have been discussed with the WWTP

technicians and at this stage are still based on their

long-time experience. However, the hydraulic model

integrated in the DSS allows operators to perform ex-

post simulations, assessing the effectiveness of the

adopted strategy or evaluating the effects of different

gate openings or different threshold levels. On this

basis, the predefined suggested strategies may be

easily changed by expert operators directly in the UI.

Future developments will include the integration of

multi-objective optimization functionalities in the

system, benefitting from the detailed information that

will become available in the database, in order to face

conflicting objectives as the need of sending as much

water as possible to the receiving body while

minimizing the chance of overflows, eventually

obtaining case-specific threshold levels and gate

openings combinations. For example, the premature

filling of the storage area may produce negative

effects during successive intense rainfalls, and

unnecessary lateral spills may result in exceeding the

allowed discharge in surrounding rural channels.

Such improvements could be achieved by integrating

real-time execution of the hydraulic model,

performing simulations based on forecasted variables

and different sets of threshold levels and gate

openings: optimization algorithms will eventually

determine the best combination to face the incoming

events. Finally, when a considerable amount of

available data will allow to accurately understand and

model all the hydraulic processes, the VF gates could

be provided with automated actuators in order to

implement a Real Time Control System.

5 CONCLUSIONS

The case study presented in this paper illustrates a

smart infrastructure project, specifically the design of

a DSS platform able to provide real-time monitoring,

weather nowcasting and forecasts on the status of

some key variables, in the context of managing a

medium-size city wastewater treatment plant. With

respect to the state of the art, a simplified approach

for modelling hydraulic variables based on Artificial

Neural Networks is proposed.

Data are acquired with a 10-minute frequency

from a network of sensors and are stored in a single

database that hosts heterogeneous variables, some of

which are used to perform run-time analyses using

ANNs. These are trained based on past events and can

be updated as more data become available. The

update procedure can be performed “offline”, this

being an important feature of the system design as,

once implemented and calibrated under certain

assumptions, it would be possible to improve the

performance and/or adapt the algorithms to different

conditions without any substantial revision of the

software architecture. Moreover, an automated

update procedure can be foreseen in future

improvements of the proposed system: a script

reading the database with an imposed frequency – e.g.

twice a year – could then process the new acquired

data, add them to the training dataset and perform

again the training procedure, replacing the matrices if

better results are gained.

The DSS was implemented to provide support for

a WWTP with the twofold objective of monitoring

the inlet flow to the plant and reducing the chance of

An Artiﬁcial Neural Network-based Real Time DSS to Manage the Discharges of a Wastewater Treatment Plant and Reduce the Flooding

Risk

overflow during severe rainfall events. The system

management is further aided by forecasts shown in

the plot window, which provide additional

information about the forthcoming behaviour of key

variables, as predicted by ANNs.

When dealing with Smart City projects, several

challenges arise. It is important to lower the design

and operational costs to increase the probability of

implementation. The handling of heterogeneous data

from multiple sources, the analysis of Big Data and

security-related issues are also to be considered (Silva

et al., 2018). The proposed approach, fulfilling these

requirements, is a valuable step in guaranteeing safety

in a Smart City context and can be in principle

replicated and applied in all those settings where

measurements from different sensors over large areas,

meteorological data, and in general any quantitative

information needs to be processed to provide

synthetic outputs for the final user.

ACKNOWLEDGEMENTS

The study presented in this paper is part of the

INNOVA EFD3 research project financed by A2A

Ciclo Idrico S.p.A.

REFERENCES

Bellon, A., Zawadzki, I., Kilambi, A., Lee, H.C., Lee, Y.H.

and Lee, G., 2010. “McGill algorithm for precipitation

nocasting by Lagrangian extrapolation (MAPLE)

applied to the South Korean radar network. Part I:

Sensitiity studies of the Variational Echo Tracking

(VET) technique.” Asia-Pacific Journal of Atmospheric

Scences 46(3): 369-381.

Botturi, A., Gozde Ozbayram, E., Tondera, K., Gilbert,

N.I., Rouault, P., Caradot, N., Gutierrez, O., Daneshgar,

S., Frison, N., Akyol, C., Foglia, A., Eusebi, A.L. and

Fatone, F., 2020. “Combined sewer overflows: A

critical review on best practice and innovative solutions

to mitigate impacts on environment and human health”.

Critical Reviews in Environmental Science and

Technology, 1-34.

Campisano, A., Creaco, E., and Modica, C., 2016.

“Application of real-time control techniques to reduce

water volume discharges from quality-oriented CSO

devices”. Journal of Environmental Engineering,

142(1), 04015049 1-8.

Carbone, M., Garofalo, G., and Piro, P., 2014.

“Decentralized real time control in combined sewer

system by using smart objects”. Procedia Engineering,

89, 473–478.

Clark, M.P., Bierkens, M.F.P., Samaniego, L., Woods,

R.A., Uijlenhoet, R., Bennett, K.E., Pauwels, V.R.N.,

Cai, X., Wood, A. and Peters-Lidard, C.D., 2017. “The

evolution of process-based hydrologic models:

historical challenges and the collective quest for

physical realism.” Hydrology and Earth System

Sciences 21: 3427-3440.

Hochreiter, S. and Schmidhuber, J., 1997. “Long short-term

memory”. Neural computation, 9(8), 1735-1780.

Jeong, J., Kannan, N., Arnold, J., Glick, R., Gosselink, L.

and Srinivasan, R., 2010. “Development and

integration of sub-hourly rainfall–runoff modeling

capability within a watershed model”. Water Resources

Management, 24(15): 4505-4527.

Lee, H.C., Lee, Y.H., Ha, J.C., Chang, D.E., Bellon, A.,

Zawadzki, I. and Lee, G., 2010. “McGill Algorithm for

Precipitation Nowcasting by Lagrangian Extrapolation

(MAPLE) applied to the South Korean radar network.

Part II: Real-time verification for the summer season.”

Asia-Pacific Journal of Atmospheric Sciences 46(3):

383-391.

Maier, H. R., Jain, A., Dandy G.C. and. Sudheer, K.P.,

2010. “Methods used for the development of neural

networks for the prediction of water resource variables

in river systems: Current status and future directions.”

Environmental modelling & software 25(8): 891-909.

Mandapaka, P.V., Germann, U., Panziera, L. and Hering,

A., 2012. “Can Lagrangian extrapolation of radar fields

be used for precipitation nowcasting over complex

Alpine orography?” Weather and forecasting 27(1): 28-

49.

Panziera, L., Germann, U. Gabella., M. and Mandapaka,

P.V., 2011. “NORA –Nowcasting of Orographic

Rainfall by means of Analogues.” Quarterly Journal of

the Royal Meteorological Society 137(661): 2106-

2123.

Park, T., and Kim, H.A., 2013. “A data warehouse-based

decision support system for sewer infrastructure

management.” Automation in Construction 30: 37-49.

Pereira, A., Pinho, J.L.S., Vieira, J.M.P., Faria, R., and

Costa, C., 2019. “Improving operational management

of wastewater systems. A case study.” Water Science

and Technology 80(1): 173-183.

Rao, M. 2015. “A performance measurement application

for a wastewater treatment plant.” International Journal

of Services and Standards 10(3): 134-147.

Rechdaoui-Guérin, S., Bersinger, T., Bareille, G., Pigop, T.,

Le Hécho, I., Azimi, S. and Rocher, V., 2018.

“Monitoring the quality of effluents in a unitary

sanitation network” Technique-Sciences-Methodes

113: 77-90.

Silva, B. N., Khan, M., and Han, K., 2018. “Towards

sustainable smart cities: A review of trends,

architectures, components, and open challenges in

smart cities”. Sustainable Cities and Society, 38: 697-

713.

US EPA. (2018). Smart data infrastructure for wet weather

control and decision support. EPA 830-B-17-004.

Wilson, J.W., Crook, N.A., Mueller, C.K., Sun, J. and

Dixon, M., 1998. “Nowcasting thunderstorms: A status

report.” Bulletin of the American Meteorological

Society 79(10): 2079-2100.

SMARTGREENS 2021 - 10th International Conference on Smart Cities and Green ICT Systems