WaterSubject

As the world emerges into the 21st century, many of the major cities may face water crises. The World Meteorological Organisation predicts that explosive growth of urban cities will rapidly deplete our water resources. By the turn of the next decade, world water demand is likely to claim almost half of the total global water runoff that is available every year. Between 1900 and 1995, water use has increased by a factor of six, which is more than double the population increase during the same period. The world population is projected to increase from the current (near 6 billion) to 8.3 billion in 2025 and to about 10 billion in 2050. Water crises are already evident in the competition for water for agriculture, domestic, and industrial use in many parts of the world. High population densities with associated high traffic intensities, intensive agriculture and use of fertilizers, inevitably imply pollution of underground and surface waters. The lack of care in use of water in general will lead to a limitation of the availability of water resources on a long-term scale. At the same time, very few countries have policies concerning the economising of water. It seems as if neither producers nor consumers are aware that water is a resource that should be saved and used with care, being treated essentially as a non-renewable resource.

The network of pipes in a city and all components associated with this network (valves, pumps, reservoirs, etc.) constitute a water supply asset. As with any other asset, it is important to invest in its maintenance it in order to fulfill its task. However, the modelling of water supply assets is a complex issue. Such an asset is constituted of many components, it is subjected to large a number of different stresses, it is mainly situated underground and therefore not visible. Nevertheless, the maintenance and rehabilitation has to be carried out regardless of such incomplete and inaccurate knowledge about the condition of the asset. It is not financially feasible to monitor the evolution of the state of an asset on a regular basis. Information is fully available only when the pipe is laid. The knowledge about the state of the pipe is only updated when a burst occurs and the pipe is partially unearthed. Moreover the management of a pipe network must deal with administrative, environmental and social constraints upon the actions of the water company. A poor condition of water supply assets typically implies high pipe burst rates, which in turn result in high water leakage rates. Reports of leakages typically amounting between 35% and 65% total supplied volume of water are not at all uncommon.

However, in the light of the already outlined ‘water crisis’ and its most likely consequences for a future economic pricing of water (whereby it can sometimes be extrapolated that the price of water supplied through an asset may soon reach the price of bottled water!), high leakage rates cannot and should not be tolerated. Despite the enormous complexity of water supply assets and the associated intellectual challenges with respect to modelling of them, it is primarily our ethical responsibility to devise approaches for the optimal management of water.

The complexity of the asset modelling problem calls for advanced methods. This paper outlines two advanced data mining techniques for the analysis of the risk of pipe burst in water supply networks. The paper also discusses the introduction of risk assessment routines within integrated water supply modelling frameworks.The economic and social costs associated with pipe bursts and associated leakage problems in modern water supply systems are rapidly rising to unacceptably high levels.Pipe burst risks depend on a number of factors which are extremely difficult to characterise. A part of the problem is that water supply assets are mainly situated underground, and therefore not visible and under influence of various highly unpredictable forces. This paper proposes the use of advanced data mining methods in order to determine the risks of pipe bursts. For example, analysis of the database of already occurred bursts events can be used to establish a risk model as a function of associated characteristics of bursting pipe (its age, diameter, material of which it is built, etc.), soil type in which a pipe is laid, climatological factors (such as temperature), traffic loading, etc.

In addition to the immediate aid with the the choice of pipes to be replaced, the outlined approach opens completely new avenues in asset management: the one of asset modeling. The condition of an asset such as a water supply network deteriorates with age. With reliable risk models, addressing the evolution of risk with aging asset, it is now possible to plan optimal rehabilitation strategies in advance, before the burst actually occurs.

6.1. Data- and knowledge-driven modelling

In this study, the two approaches were confronted: a pure data-driven technique and one in which data and knowledge about the problem were combined. The pipe network is an underground asset, and information on its condition and evolution is not readily available. It is therefore difficult to build a model deterministically, while data-driven techniques can show relationships between the available source of information that are not pinpointed by physical laws.

6.2. Building the scoring model: the sources of information reviewed

The scoring model was built exclusively from data. Provided that data of sufficient quality and quantity exist, it is reasonable to expect that the scoring model could capture the underlying processes and provide good burst risk characterisation.

The present case study is based on an exceptionally long history of asset condition which has been collected with a great care. Therefore both quality and quantity of available data were unusually high. Bearing this in mind, it is no surprise that the resulting scoring models outperform Bayesian networks.

6.3. Building the Bayesian network: the sources of information reviewed

The sources of information for building a Bayesian network can be divided into three categories:

• data, most of the time in the form of statistical descriptions;

• literature, scientific knowledge in the form of equations or already-built models;

• human experts, that possess the empirical field knowledge and the past experience.

With a comprehensive data collection, it is possible to build the network of causal relationships and the probability tables automatically. This process allows also for the update of the network in case of new information becoming available.

Unfortunately, it is quite seldom that comprehensive data collections are available, especially in underground asset modelling. Moreover, the knowledge collected through the experience of experts is generally biased, making the learning process difficult and potentially inaccurate.

For Bayesian networks to be reliable in the field of underground asset modelling, one has to be able to assess the quality of the data and the knowledge encoded in the model. The quality of Bayesian networks to encode knowledge also represents its weakness: Bayesian networks can gather knowledge that is not relevant to the model (thus increasing its complexity with no benefit to the overall model) or even worse, it can gather ‘knowledge’ that is wrong.

Another issue to be raised is the necessity to reduce the complexity of the network: the computational time required to build the probability table increases exponentially with the number of causal relationships in the network. Adding knowledge adds relationships in a network that is already complex.

To summarize, Bayesian networks have a great potential but must be build with considerable care in order to be reliable.

6.4. A final word: a need for more comprehensive modelling

To the best knowledge of the authors, this is one of the first attempts to model an underground asset using data-mining methods, and accommodating to the objective of dealing with the uncertainty of the outcome instead of trying to reduce it.

From the perspective of uncertainty handling, Bayesian networks and scoring models are two technologies that should be further investigated in order to provide a reliable model that can be included in the framework of asset modelling.

This study represents only an introduction of asset modelling: it has opened doors to a vast domain for research and experimentation to understand the behaviour of underground assets. The demands to obtain a reliable model are two fold: a better knowledge of the pipe and its failure behaviours and better techniques to model the uncertainty.