Purpose Purpose Purpose Purpose Case study Case study Case study Case study Metamodel Metamodel Metamodel Metamodel Performance Metamodel Performance
Water network Application category Reference Application Location Size: Pipes in model / [area \(\mathbf{k}\mathbf{m}^{\mathbf{2}}\)] Classification by size Type Deviations from simple MLP Inputs (Number) Outputs (Number) Computational saving Accuracy
Water distribution systems Optimisation (Sayers et al., 2019) Design TLN, GOY, MOD, BIN 8, 30, 317, 454 S, S, I, I Benchmark 2 hidden layers Diameters * Rating of the network (1) Not reported Not reported
(Dini & Tabesh, 2019) Renovation planning TLN and Ahar, Azerbaijan 8 and 192 S, M Benchmark and Real case Diameters * Nodal pressure* and chlorine concentration * Not reported Not reported
(Dini & Tabesh, 2017) Model calibration TLN and Ahar, Azerbaijan 8 and 192 S, M Benchmark and Real case Observed residual chlorine * Wall Decay coefficient (1) 58x faster (98.3%) Average error (3.85%)
(Andrade et al., 2016) Design HAN and Maricopa, Arizona 34 and 1090 S, L Benchmark and Real case Comparison of ANNs varying number of inputs and outputs Diameters and Chlorine dosing rates Chlorine concentration. (HAN): 3; (Maricopa): 9 Not reported NSE (~90%)
(Bi & Dandy, 2014) Design (I) NYT, (II) modified NYT and (III) Jilin 21, 21, and 34 S, S, S (I) Benchmark, (II) modified benchmark, and (III) synthetic network Diameters and Chlorine dosing rates (I & II: 22; III: 35) Pressures at some nodes (I & II: 4; III: 5) and residual chlorine at one node (I & II: 1; III: 7) (I & II) 91%; (III) 93%, 88%, and 77% MSE (Not reported, 0.001 as one stopping criteria)
(Broad et al., 2010) Operation Wallan, Australia 2097;(Sk: 1376) L (L) Real case Trigger levels (45) and Chlorine rates (5) Pressure Head at critical node (1), Chlorine residual (1), energy value (1), or Total chlorine dosed (1) 99% NSE (~0.6 for the full model, ~0.98 for skeletonized model)
(Behzadian et al., 2009)
Sensor placement
Anytown; Mahalat, Iran
41, and 1814;(Sk: 217)
S, L (M)
Benchmark and Real case
Available sensors
Sampling design accuracy (1)
8x and 25x faster (87% and 96%)
Pareto similarity: 93%
(Salomons et al., 2007) Operation Haifa-A, Israel 126 M Modified real case Pumping status (13), Valve settings (1), DMA demands (6), Storage levels (9) Power consumption (5), pressures (4), future storage levels (9) 25x faster (96%) RMSE (0.481%) ~5 cm averaged over all tanks
(Martínez et al., 2007) Operation Valencia, Spain 772 L Modified real case Pumping status (6), Valve settings (10), DMA demands (6), Storage levels (2) Power consumption (6), flow rates (3), pressures (4), future storage levels (2) 94x faster (99%) RMSE (1.30%)
(Broad et al., 2005a) Design NYT 21 S Benchmark Diameters and Chlorine dosing rate (22) Four pressure nodes (1) or Chlorine concentration (1) 700x faster (99.85%) RMSE (0.05 - 0.250)
Real-time (Pasha & Lansey, 2014) Warm solutions for pump scheduling Modified Anytown 37 S Modified Benchmark SVM Pump combination, demand multiplier, initial tank levels Energy and final tank levels 84.25% NSE (0.99)
(Rao & Alvarruiz, 2007; Rao & Salomons, 2007) Real-time pump scheduling Modified AnyTown 41 S Modified Benchmark Number of operating pumps (1), aggregated demand (1), and tank levels (3) Power consumption (1), pressures (3), new tank levels (3) 10-fold (90%) RMSE (1.65%)
Uncertainty analysis (Yoon et al., 2020) Seismic risk assessment A-city, South Korea 85 S Anonymous real case 15 layers - Deep neural network Components’ state (218) Network performance (1) 99% <5%
(Beh et al., 2017)
Planning under deep uncertainty
Adelaide, Australia
NA
L
Real case
Combination of 4 MLPs
Supply augmentation options (9) and Uncertain variables: Population and climate change scenarios (2)
(I) PV of cost (II) PV of Greenhouse gases (III) Reliability (IV) Vulnerability
>99%
Relative error (+-5%) NSE (~0.94, 0.95, 0.78, and 0.84)
System state estimation (Lima et al., 2018) Nodal pressure estimation at near real-time Campos do Conde II and Cambuí, Brazil 153 and 167 M, M Real case Pressure in sensors Steady State: (3) - Extended (24h): 96. Cambuí: (4) Pressure in nodes Steady State: (118) - Extended (24h): 2832. Cambuí: Steady (154 and 4) Not reported Relative error (<1%) and (<4%)
(Meirelles et al., 2017) Calibration with estimated pressures Campos do Conde II, Brazil and C-Town 153 and 429 M, I Real case and Benchmark Pressure in sensors Steady State: (3) - Extended (24h): 96. C-Town: 5 MLPs, one per DMA. Pressure in nodes Steady State: (118) - Extended (24h): 2832 Not reported Average error (0.15 m) Max. Error (13.83 m)