People:Water EngD/PhD:The Stream Industrial Doctorate Centre

Sophocles Sophocleous's Project

Cohort VI (2014 - 2018)

Development of Next Generation of Water Mains Network Modelling Tools

Project sponsors

Project background

Current water mains network modelling (WMNM) is too coarse a tool for effectively supporting operational work at distribution mains level. Reasons for this include the ±1m calibration criterion for simulated pressures in the model as compared to those pressures often recorded to a transducer accuracy of 0.1m (0.1% full scale deflection) in field tests for the modelling studies. The frequent lack of flow measurements other than those into (and possibly out of) the District Metered Area being used for the model study is another contributory factor for the embedded shortfall in modelling efficacy. So too are the local demand patterns within the study area. These are often too small to generate sufficient flows in pipes for making accurate measurement. The associated hydraulic head losses are correspondingly too small to provide suitable observations to facilitate effective calibration. Other factors include the many changes that have been made to the network since water industry privatisation. They are often associated with mains rehabilitation and leakage reduction strategies. Conditions have arisen that lead to data anomalies within the geographical information systems which can then be carried over in the water mains records used for model building.

A replacement modelling practice is required that detects the state of assets and their status changes within the real network quickly, and also conveys results accurately, meaningfully and conveniently to the network operator who, in the field, can then update support systems reliably and quickly. The state of assets within the project includes node based leakage hotspots and pipe based hydraulic roughness. The status includes whether pipes or throttle control valves are open or closed.

Host University

Specific aim and objectives of project

Aim

To investigate the use of custom-and-practice Water Mains Network Modelling and Calibration methods, in order to develop a replacement modelling practice based on optimisation methods, which resolves UK Water Industry issues associated with the detection of burst-related leakage hotspots.

Objectives

Improved design of pressure and flow trending data acquired from field tests required for optimised calibration of WNWM.
Automated interrogation of field data and any avaliable contemporary real time data. This to provide informed reconcliiation of that data to reveal more targeted starting point solutions spaces for subsequent WNWM optimisations.
Further development and expansion of genetic algorithm functionality to improve computational efficiency on the group of problems identified. This wlil include optimisation of valve status and functionality to detect unknown inflows at unknown open inlet boundaries.
Identification of possible anclilary field data during optimisation process to help fine tune final optimisation outputs including improved accuracy of leakage hotspots.

Methodology

The base WMNM will be the assumed starting point. This is the model that includes valve configurations in line with existing operational understanding, leakage assigned to model nodes in proportion to domestic properties attached to those nodes and pipes with default hydraulic pipe friction values. For rural models, the leakage might alternatively be related to the nearest half lengths of pipe elements served by each node. Building on recent preliminary investigations, an optimisation-based application developed in MATLAB will then identify the flow reticulation within the WMNM by a new pipe-trees analysis, which is based on graph theory. This analysis will determine whether nodes in model are source, flow reticulation of flow converging nodes or sink nodes. Pipe trees will commence at source nodes or converging flow nodes. They will generally comprise of trunks, boughs, branches, twigs and twiglets. The pipe tree hierarchy will provide a looped representation of the WMNM. From this hierarchy the field test sites and possible unknown closed valve sites will be automatically derived. There will be options for users to modify derived selections. The new functionality will be able identify possible unknown closed valve sites from the derived pipe tree structure typically targeting elements just upstream and downstream of converging flow nodes and also those through which flow passes into significant branches.

Turning to the field testing required for model calibration, the custom-and-practice tests need modification to introduce more and larger known flows into the observed data sets provided for model calibration. Introduction into the test of controlled hydrant discharges is required. But network operators are frequently reluctant to do this because of the increased risk of reportable discolouration incidents to the water industry regulator. The risks can be mitigated by timing the hydrant discharges to take place in lower demand periods at night. Also, the application of the prediction of discolouration within the test areas based on the work that has been undertaken on this topic over last ten years will help reduce the discolouration risks.

Preliminary Results

The calibration approach

A hydraulic simulation analysis was carried out in EPANET2 by considering the true state of the network, in order to create an artificial set of field (i.e. observed) pressure and flow measurements for calibration, without accounting for noise. The artificial data were collected by means of planned hydrant discharges during Night Fire Flow Field Tests (NFFFT). Also, planned valves closures were introduced to the NFFFT while the hydrants were open. The NFFFT observations were inserted into an optimization algorithm tool for calibration that considers as decision variables valve status, pipe roughness and leakage coefficients, in order to detect the true leakage hotspot locations and their emitter coefficients. The optimization-based calibration is defined as a nonlinear optimization problem with the single objective to minimize the function ("fitness") for the weighted sum of absolute differences between the field observed and simulated values of junction pressures (hydraulic grades) and pipe flows for given boundary conditions, including valve status, pipe roughness and leakage emitter coefficients, as well as subject to implicit constraints defined by continuity (for every node) and energy loss equations (for every pipe). Two calibration problems were solved according to two case study specific cases outlined below.

The "true" system state

The EPANET2 network layout of the study area is presented in Figure 1. It involves a real-life Discrete Pressure Area (DPA) system (Sage, [2]), comprising of 363 nodes, 180 pipes, 105 throttle control valves (TCVs) and a pressure reducing valve (PRV). The network model contains two closed TCVs (T58 and T196), one open cross-connection (TNUXCON) and three leakage hotspots (J202, J244, and J341). These adjustments were made for subsequent calibration purposes aligned with the paper's objectives. A fourth emitter was also introduced in the model, representing a fabricated leak at J343. This was to test the optimiser's capability to detect leakage hotspots using a perfectly calibrated test model (i.e. with known valve status and pipe roughness). The fabricated leak was set between 00:30 and 02:00 hours with an emitter of 0.296 (or 2.01l/s). Finally, six chosen pipes from the base model, associated with three different pipe groups based on pipe material (e.g. four Cast Iron pipes (p734, p836, p1006, and p1008), one Polyethylene (PE) pipe (p1219) and one Asbestos Cement (AC) pipe (p592)) were considered for the calibration problem.

Figure 1. The true DPA system configuration illustrating the three TCVs of consideration for calibration, the six hydrants that were operated, the several pressure and flow measurement points and the four leaks to be identified .

Case I

The optimization tool initiates the search process by randomly generating values for each decision variable, i.e. the valve status and pipe roughness, assumed to be the correct system configuration for the first calibration problem. Consequently the optimization tool was run to detect the leakage hotspot characteristics, by calibrating the parameters of leakage emitters and their correct index within the network corresponding to the location.

Figure 2 illustrates the optimization outcome for the task of calibrating for the correct leakage hotspot location and emitter coefficient, without considering either unknown valve status or increased pipe ks coefficients. With a fitness of 143, representing the scope of error between observed and simulated flows and pressures, the optimization process failed to calibrate the model. This resulted in severely wrong detection of leakage hotspot locations. In addition, the optimizer overestimated the emitter coefficients at the leakage nodes, consequently overestimating the emitter's discharges compared to their true values.

Figure 2. Calibration optimization problem solution for leakage hotspot location presented on the EPANET2 network layout along with their true location. This demonstrates the incorrect detection of leakage hotspot location, assuming a known status and ks coefficient for each valve and pipe, respectively.

Case II

As in Case "I", the optimisation process starts with decision variables taking randomly generated values, with respect to valve status and pipe roughness. However, a staged process was followed for the calibration. Firstly, the optimiser was used to detect the correct valve status. This involved calibrating for the valve status of the considered group of valves. Then, following the valves' status (imaginary) verifications with the true system state (Figure 1), the second stage was to calibrate for the correct pipe ks values in each pipe group, together with the emitter coefficients and indices of the group of candidate nodes, again restricted on longer pipes.

Following the staged optimisation, that detected the correct status for the unknown valves and the correct pipe ks values, good results were obtained for the leakage hotspots including a solution fitness of 0.1. This second stage took into account the correctly identified unknown closed valves and the open cross connection leading to successful identification of all the leakage hotspots. On the other hand, the optimization outcome for pipe ks coefficients did not lead on to correct calibration (Table 1), but the suggested ks values were significantly reduced illustrating that the pipes in the model should have been smoother. The three leakage hotspots locations (J202, J244, and J341 and fabricated leak at J343) were correctly detected (Table 2).

It has been found that early proving in the field of any detected unknown closed valves from the correctly calibrated model improves detection of pipe roughness and leakage hotspots during a second stage optimisation. The staged approach to optimisation also opens up the possibility for the introduction of additional local field data collection to resolve remaining ambiguities. Examples could include detected closed valves in optimisation from first stage work that were then found to be already open in real world. The proposed project considers the field test data is to be collected and improved reconciliation of the same. This is to facilitate the optimisation and possible implementation of additional field data for second stage optimisation.

Anticipated Outcomes Of The Project

The work would provide a useful replacement for the custom-and-practice modelling that is still in widespread use, but which is failing to provide adequate support for operational activities such as leakage reduction or consistently 'right-first-time' execution of planned and unplanned work on the water mains distribution network. The work will develop new graph theoretic approaches for network analysis and combine those with network modelling (simulation) analysis that will lead to a major breakthrough in developing and utilising WMNMs.

Industrial Impact

The project will help provide the next generation of Water Mains Network Models further exploiting and building on the existing genetic algorithms methods for such work. But it will also integrate new methods for field test design and real time data collection and subsequently provide improved solutions spaces thereby promoting more accurate and reliable optimisation outputs. Better modelling and analyses will lead to reduced leakage and better understanding of the operational status of the network and its elements. The roll out of the work will lead to a paradigm shift in the water industry's approach to water mains network modelling. It will transform modelling from an activity undertaken by modelling specialists to one which is highly automated and much more readily available as part of an applications' toolkit available to water mains distribution systems operational staff.

Outputs

Sophocleous, S. et al., (2015) Advances in Water Mains Network Modelling for Improved Operations. Procedia Engineering, 13th Computer Control for the Water Industry Conference 2015, 119, 593-602.