Water is a scarce resource and must be efficiently managed. One of the purposes of efficient management should be
reducing water losses and increasing supply performance in the networks through a profound knowledge of water supply
networks (WSN). Obtaining this knowledge in real networks is a complex task because distribution systems may consist of
thousands of consumption nodes interconnected by thousands of lines and the necessary elements to feed the network.
These networks are not usually the outcome of a single process of design and are the consequence of years of anarchic
responses to continually rising new demands. As a result, layouts lack a clear structure from a topological point of
view.

The division of a water supply network into 'supply clusters' enables sufficient hydraulic knowledge to be gathered to
carry out maintenance tasks and guarantee quantity and regularity to the final consumer. This approach divides large and
highly interconnected distribution networks into smaller sub-networks. These smaller networks are virtually independent
and fed by a prefixed number of sources. Independence can be physically enforced in a number of ways. For instance, by
closing valves in existing pipes, by sectioning existing pipes, or by introducing new pipes to redistribute the flow.
Each supply cluster inlet must be equipped with at least one flow-meter to accurately measure and record consumption in
short time periods. However, gauges and meters must also be placed to measure and control pressure, chlorine
concentrations, and other supply parameters.

Water network division into supply clusters should not be understood only in network configuration terms, but also as a
permanent method of management. It is essential to provide the system with a main centre that can receive and sort daily
data, as well as analyse other information (such as financial, climatological, inventory, and maintenance data). In this
way, we can establish the real performance of sectors and take appropriate decisions regarding both operational
maintenance and investment. From a classical perspective, the division of a water supply network into sectors is used
with the aim of controlling leaks, since it helps maintain a permanent pressure control system. Nevertheless, this
target has recently become more ambitious and incorporates new operational and management tasks.

This thesis proposes a suitable framework to establish efficient methods to divide the network into sectors and manage a
WSN by taking advantage of this structure. These tasks will be addressed using kernel methods and multi-agent systems.
Spectral clustering and semi-supervised learning have been shown to behave well when defining a sectorised network with
a minimum number of cut-off valves. However, their algorithms are slow (and sometimes infeasible) when tackling
divisions in a large WSN. The multi-agent system approach, firstly created as an alternative solution, is an excellent
complementary tool for clustering kernel methods that use a boosting methodology. It is therefore possible to achieve a
division of WSN into supply clusters – even in the case of large networks. This thesis also highlights other machine
learning and kernel methods, such as support vector machines, to manage a single supply cluster and facilitate the
detection, identification and monitoring of possible abnormalities in water supply. In the same sense, predictive models
are more accurate in a supply cluster than in the whole network, as they avoid biases derived from producing forecasts
in smaller areas. Finally, another variant of kernel-spectral methods, similar to Google PageRank, is adapted and
developed to assess the relative importance of the nodes of a WSN by assessing vulnerabilities and proposing a
working-line for approaching several management and operation tasks (including WSN division).