Deep Learning and statistical algorythms in searching leaks, water demands and optimal pressure sensors topology in meshed water distributions networks
PREMISE
This research work aims to identify methodologies and algorithms suitable for identifying leaks in water distribution networks characterized by a mesh conformation.
For this purpose it is supposed to install pressure sensors in the water distribution networks and to numerically process the data detected by them, in the presence and absence of leaks, as the operating conditions of the networks vary, i.e. as the supply pressures vary and as withdrawals are made from them.
;
STUDY OF TRANSIENTS STATESThis research work aims to identify methodologies and algorithms suitable for identifying leaks in water distribution networks characterized by a mesh conformation.
For this purpose it is supposed to install pressure sensors in the water distribution networks and to numerically process the data detected by them, in the presence and absence of leaks, as the operating conditions of the networks vary, i.e. as the supply pressures vary and as withdrawals are made from them.
;
Methods that analyze the pressure trend in transient conditions have first been explored.
For this purpose short-duration test pressure signals (each a few milliseconds long) have been applied in the aforementioned networks.
;
The pressure data generated in the networks by the combination of the reflections of the pressure waves due to the application of the test signals have been analyzed
;
More specifically, the differences between the data detected in the absence of leaks and the data detected in presence of leaks will be explored.
;
Due to lack of real data, the distribution networks have been simulated using a software simulator (in particular using ALTAIR ACTIVATE simulator) and the pressure and flow rate data in such a way generated have been processed by applying in Matlab framework classic data processing methods consisting in digital filters, FFT, integrations, differentiations , crosscorrelations, autocorrelations, etc.
;
STUDY OF STATIONARY STATES
Methods have then been explored that analyze the pressure trend in stationary conditions, i.e. after the effects of the aforementioned transients have completely disappeared.
In this case, again, due to absence of real data, the pressures data have been simulated using the EPANET software or have been retrieved from synthetically constructed databases.
For this purpose, a database (LeakDB) is available on internet which is normally used in literature to compare the methods and to evaluate the results obtained.
This database contains simulated data relating to a portion of the distribution network of the city of Hanoi and more specifically contains data relating to 1000 scenarios, each lasting one year, containing samples pressure signals occurring every 30 minutes.
The scenarios are characterized, from time to time, by the absence of leaks or by the presence of abrupt or incipient leaks of various durations and amounts.
The portion of the Hanoi network in question is made up of approximately 31 junctions and of a reservoir and the pressure data have a sort of weekly periodicity on which a seasonal pattern and a certain amount of instantaneous noise are superimposed.
The data from the "Hanoi" network have been analyzed with machine learning/deep learning methods and in particular with MLP-type neural networks, or with RNN-type networks (adopting, from time to time, with regard to this specific category of neural networks, standard RNN autoencoders, LSTM autoencoders, GRU autoencoders, bidirectional LSTM autoencoders, each of them with or without "code/bottleneck" layers).
In fact, it should be noted in this case, that the study actually materializes in the determination of the anomalies present in time series (each of which coincides with a scenario).
CNN-type neural networks have then also been applied to these time series and in particular CNN1D one-dimensional autoencoders and CNN2D two-dimensional autoencoders, with or without residual blocks.
STUDY OF THE INVERSION ALGORITHMS OF THE EPANET SOFTWARE
The EPANET software allows, once the withdrawals from the nodes are known, once the characteristics of the reservoirs are known, once the characteristic curves of the pumps are known, once the characteristics and topology of the pipelines are known, to calculate the pressures each node of the water distribution networks.
Therefore the pressures and flow rates in the pipes of the water distribution networks constitute the output data of the EPANET software.
The EPANET software is available in the form of a graphical application and also in the form of freely usable software libraries suitable to create own applications written in C++ language, in Python language or within Matlab framework.
However, I believe that, since the pressures can be measured in field with relatively cheap sensors, it may instead be of considerable interest to invert the EPANET model and, in such a way, to allow the water demands from the nodes of the water distribution networks to be determined starting from the knowledge of the pressures of the nodes and the from knowledge of the other characteristics of the distribution network.
Therefore, for the sake of inversion of EPANET model statistical and Deep Learning methods will be developed and, more particularly, with regard to the latter method type, MLP neural networks have been used.
The pressure data necessary for the study of the inversion of the EPANET model have been generated, using the aforementioned EPANET libraries, by writing software simulators in Matlab framework, or by writing software in Python language.
OPTIMISATION OF NUMBER AND POSITION OF INSTALLED PRESSURE SENSORS IN THE WATER DISTRIBUTION NETWORK
It is unthinkable or in however not economically sustainable to install a pressure sensor in each node of the water distribution network.
It is therefore convenient to study and apply algorithms that allow, while maintaining acceptable pressure measurement performance, to reduce the number of pressure sensors installed in field.
For this purpose, given the EPANET model of a water distribution network, a polynomial model was developed, through multiple polynomial regression on pressure samples generate using EPANET, in order to estimate the pressure in each node of the network via the pressures measured in the other nodes.
In this way it is possible to estimate, by analyzing the coefficients of the polynomial equation corresponding to each node of the network, which are the other nodes of the network that contribute most to determining the pressure of the node under examination and so it is possible to chose which are the nodes in which it may be acceptable to omit the installation of the pressure sensor.
Back to Current Students