Decision Support Tool for Design and Reliability Assessment of Water Distribution Networks - Pattern Recognition for Reliability Assessment of Water Distribution Networks - page 256

Figures 9.4, while the pipe lengths have been calculated from nodal coordinates. The uniform
k-value of 0.5 mm was used in all the cases. These values have been set using the network
initialisation module.

Table 9.10 Overview of GA optimised networks

Batch code Total nets Non-matching

links

Optimisation
time (min)

R55-75 (U)5000 409 3 97

NfullC2(4)1000 477 0 91

NfullC4(U)1000 252 1 64

N55-75C2(4)1000 375 0 75

N55-75C4(U)1000 304 1 77

9.5 RESULTS AND DISCUSSIONS

All the optimised networks have been further analysed in the network diagnostics module.
Figures 9.8a to 9.8c show the costs against the ADF avg (9.8a-left), NBI (9.8a-right), PBI
(9.8b-left), I n (9.8b-right), p min (9.8c-left) and p max (9.8c-right), respectively. The following
observations can be made looking at these diagrams:
- The 409 randomly generated networks show better spread of the values than the rest. This

is not a surprise, also given the larger number of generated networks in this sample
(5000).

- The non-random networks show lesser variation, specifically the samples of 477 and 375

filtered networks of low complexity. This gives impression that the differences in
configuration of these networks are relatively small i.e. many more networks should have
been generated to arrive at better spread of values.

- The low-complexity samples (477 and 375) appear to be the most expensive in general,

while the least reliable according to the values of ADF avg and NBI (Figure 9.8a). This
contradicts the results in Figure 9.8b where these networks, although more expensive,
appear also to be more reliable. This originates from the higher pressures these networks
have, confirming the head-driven nature of PBI and I n , already discussed in Chapters 5
and 7. This also raises some concern about these parameters as 'reliable' reliability
measures.

- The best 'value for money' is visible in the samples of 252 and 304 networks, originating

from the same family of non-random generated networks of complexity category 4. These
are in any case the cheaper networks and more reliable according to the values of ADF avg .

- Somewhat surprising, the values of NBI in case of the samples of 252, 304 and 409 cover

the wider range of values which gives impression that the GA-optimiser in this case has
not produced clearly branched skeleton, as obviously is the case with the samples of 477
and 375 pipes. Repetitive simulations with modified GA-settings could likely throw more
light on this situation. At least, more complex networks would require larger number of
generations. The preliminary conclusion is that the level of complexity and (even) spatial
distribution of demands can influence the branched structure of secondary mains in
optimised looped networks.

- The second peculiarity regarding the optimisation runs is that the minimum pressures in

the two categories of the networks of lower complexity (the samples of 477 and 375