Dynamical stability of water distribution networks

Abstract

Water distribution networks are hydraulic infrastructures that aim to meet water demands at their various nodes. Water flows through pipes in the network create nonlinear dynamics on networks. A desirable feature of water distribution networks is high resistance to failures and other shocks to the system. Such threats would at least transiently change the flow rate in various pipes, potentially undermining the functionality of the whole water distribution system. Here we carry out a linear stability analysis for a nonlinear dynamical system representing the flow rate through pipes that are interconnected through an arbitrary pipe network with reservoirs and consumer nodes. We show that the steady state is always locally stable and develop a method to calculate the eigenvalue that corresponds to the mode that decays the most slowly towards the equilibrium, which we use as an index for resilience of the system. We show that the proposed index is positively correlated with the recovery rate of the pipe network, which was derived from a realistic and industrially popular simulator. The present analytical framework is expected to be useful for deploying tools from nonlinear dynamics and network analysis in the design, resilience management and scenario testing of water distribution networks.

1. Introduction

Water distribution systems aim to securely provide drinking water to consumers and for firefighting and hence are key infrastructures in our society. They are subject to drastic changes in demand on daily and seasonal time scales [1] and to various threats such as contamination [2] and power outage [3]. For a safe and secure water supply, it is important that a water distribution system is well designed and is operated in such a way that potential failures are minimized.

In risk management in water engineering, the risk is estimated to be equal to the product of the probability of failure and the consequence of the failure [4,5]. In contrast to risk management, resilience management is a relatively new framework that can be used to prepare for potential failures in water distribution systems. Resilience is a broad concept and its definition varies according to studies in engineering [6] and in other fields [7]. The concept of resilience is equally wide for water distribution systems; it is construed as the general capacity of a system to resist, absorb, withstand and recover from stresses [6,8]. It is a useful complement to risk management, as many failures that happen in real life are unforeseeable and are not targeted in traditional risk management because of the low estimated probability.

Water distribution systems are composed of pipes combined with other functional structures, such as reservoirs, pumps and valves. Network dysfunctions such as pipe failure are an obvious threat to the functionality of the water distribution system. Network science has accumulated knowledge on how network structure shapes the network's capacity to withstand random failures and intentional attacks [9,10]. Therefore, there has been a surge of interest in the water engineering community to deploy network analysis to assess the degree of resilience of water distribution networks. However, to the best of our knowledge, most of the existing proposals of resilience measures for water distribution networks are based on the network structure [11–18] or steady-state or quasi-steady-state flow rates [19–24]; as such, the dynamic, transient water flows in the pipe network cannot be represented, which is in fact the key to the resilience performance of a water distribution system. Furthermore, one needs to apply stresses to the system to measure resilience [18], which cannot exhaustively include all failures that might occur in a system. Therefore, it is necessary to examine the transient dynamics to reveal many of the resilience features of the system, in addition to the static network structure or flow. In fact, the transient dynamics of water flow in pipe networks are not a new issue in water engineering. The examination of transient flows is also a practical requirement because live water distribution networks are constantly undergoing changes because of, for example, routine operational adjustments, breakdowns of their elements or human error. In turn, transient water flows are known to cause various problems in water distribution systems, such as pump failure, the collapse of vapour cavities and compromised water quality. Motivated by these needs, various methods to simulate and understand transient flows have been developed [25–29]. While useful, these developments consist of mathematical modelling of the transient water dynamics and their numerical simulation methods. These methods do not themselves provide a measure of resilience.

In the present study, we propose an index of resilience via linear stability analysis of a standard set of dynamical equations modelling steady-state flow in a given water distribution network. The equations are equivalent to a type of nonlinear electric circuit system. The proposed resilience index is defined in terms of the dominant eigenvalue of the Jacobian matrix around the steady state. Therefore, the index corresponds to the speed at which the transient dynamics decay towards the steady state in response to a small perturbation in the flow rate. The index is calculated from the structure of the network, the diameter and length of the individual pipes, and other constraints such as the energy level at the reservoir nodes and water demand at individual consumer nodes. We test the proposed index against different resilience measures that we previously proposed based on realistic numerical simulations of water distribution systems. The Matlab code for calculating the so-called local stability index is available at Github (https://github.com/naokimas/LSI_water_distribution_net).

2. Model

Consider a connected and undirected network of pipes. By definition, a pipe connects two nodes, where nodes are junctions, reservoirs and households. A dead-end node that is only connected to a single pipe (hence, the node's degree is equal to 1) corresponds to a consumer node such as a household. It should be noted that nodes whose degree is larger than 1 are junction nodes but can also demand water if they are connected to households. Let us denote by N and M the number of nodes and edges, respectively. Each edge has its own diameter and length, assuming a circular shape of the pipe. These properties affect how much water flows through the pipe under a given condition, thus effectively specifying the weight (i.e. capacity) of the edge. However, we consider the network to be an unweighted network and explicitly model the effect of the pipe diameter and length.

We consider a rigid water column model of pipe networks under the assumption that flows in a pipe vary slowly in time [25,30–32]. The dynamical equation representing transient flow through the mth pipe (1≤m≤M), which connects the ith and jth nodes, is given by

$$I_m\frac{\mathrm{d}{Q}_m}{\mathrm{d}t}=h_i-h_j-R_m{Q}_m|Q_m|-u(Q_m).$$ 2.1

Nodes i and j depend on edge m. However, unless we state otherwise, we use i and j here and in the following text to avoid notational abuse. In equation (2.1), Q_m is the time-dependent flow rate through the mth pipe, in the direction from the ith node to the jth node; t represents the time; h_i represents the total energy at the ith node, which depends on the time; R_m is a constant called the pipe coefficient containing information on the diameter, length and roughness of pipe m; R_mQ_m|Q_m| accounts for the nonlinear head loss across the mth pipe; u(Q_m) is the input term which comes from a valve or other structure located on the mth pipe; I_m is the pipe inertia constant. The dynamical system given by equation (2.1) is equivalent to that of a nonlinear resistor–inductor electrical circuit, where Q_m is the current, h_i is the voltage, I_m is the inductance, R_m is a nonlinear resistance and u(Q_m) is other nonlinear elements placed on the mth edge.

The pipe coefficient is given by

$$R_m=\frac{8f_m{\ell }_m}{g{\pi }^2{d}_m^5},$$ 2.2

where f_m is the dimensionless Darcy–Weisbach friction factor for the mth pipe and depends on the flow regime (see below for the equation) and other factors, ℓ_m is the length of the mth pipe, g is the gravitational acceleration and is equal to 9.81 m s⁻² and d_m is the diameter of the mth pipe [26]. More generally, if the nonlinear head loss term is given by R_mQ_m|Q_m|^γ−1, one obtains $R_m=f_m{\ell }_m/2g{d}_m{\mathcal{A}}_m^{\gamma }$, where ${\mathcal{A}}_m$ is the cross-sectional area of the mth pipe, i.e. ${\mathcal{A}}_m=\pi d_m^2/4$ [26,33]. The pipe inertia constant is given by $I_m={\ell }_m/(g{\mathcal{A}}_m)$.

There are various equations to approximate the friction factor, f_m, as a function of the pipe diameter, d_m, the roughness coefficient of the pipe (which is determined by the pipe's material, age and other factors), Reynolds number and so forth. Because the Reynolds number is a function of the flow rate, Q_m, the friction factor also depends on Q_m. Our index, which is developed in §3, involves differentiation of the right-hand side of equation (2.1) with respect to Q_m, which requires the derivative of f_m with respect to Q_m. In fact, many equations for f_m can only be applicable to particular flow regimes; the flow regime is specified by the value of the Reynolds number. Because f_m may depend on pipes, different pipes may be in different flow regimes and some pipes may be situated near the boundary of two flow regimes. Therefore, we select the equation proposed in [34] that is applicable in all flow regimes including laminar and turbulent flows and gives f_m as a continuous function of the Reynolds number (and hence continuous in Q_m). This is given by

$$f_m={\left(\frac{64}{{\mathcal{R}}_m}\right)}^a{(0.75\ln \frac{{\mathcal{R}}_m}{5.37})}^{2(a-1)b}{(0.88\ln \frac{6.82d_m}{ϵ})}^{2(a-1)(1-b)}.$$ 2.3

In equation (2.3), the Reynolds number for the mth pipe is equal to

$${\mathcal{R}}_m=\frac{\left|Q_m\right|d_m}{{\mathcal{A}}_{m}v},$$ 2.4

where v = 1.007 × 10⁻⁶ m² s⁻¹ is the kinematic viscosity of water, assuming a temperature of 20°C. Furthermore, ϵ = 2.591 × 10⁻⁴ m is the roughness coefficient of cast iron [35] (also see [26] for a similar value), and

$$a={[1+{\left(\frac{{\mathcal{R}}_m}{2712}\right)}^{8.4}]}^{-1}$$ 2.5

and

$$b={[1+{\left(\frac{ϵ{\mathcal{R}}_m}{150d_m}\right)}^{1.8}]}^{-1}.$$ 2.6

3. Local stability index

In this section, we carry out the linear stability analysis of the steady state of the nonlinear dynamical system given by equation (2.1). The system turns out to be always linearly stable, and we propose a local stability index as the magnitude of the eigenvalue corresponding to the most slowly decaying mode. This task is non-trivial because some nodes act as a reservoir of water, other nodes act as a consumer of water and the friction factor, f_m, depends on the dynamical variable, Q_m.

(a). Removing redundancy from the dynamical equations

Assume that the total energy at N₀ nodes is fixed. The N₀ nodes are typically reservoirs whose altitude is higher than that of the typical consumer and junction nodes. Without loss of generality, we assume that h_{N−N0 + 1}, h_{N−N0 + 2}, …, h_N are fixed. Then, the set of equation (2.1) has N − N₀ + M unknowns, i.e. h_i (1≤i≤N − N₀) and Q_m (1≤m≤M), whereas equation (2.1) only provides M differential equations. The remaining constraints come from Kirchhoff's current law (i.e. conservation of water mass) at each of the N − N₀ nodes whose total energy is not fixed. In other words, we obtain

$$\sum _{j=1;j\ne i}^N{A}_{ij}{Q}_{ij}=-{\tilde{Q}}_i^{\mathrm{ext}}(i=1,2,\dots ,N-N_0),$$ 3.1

where Q_ij is the flow rate through pipe (i, j) from the ith to jth nodes; ${\tilde{Q}}_i^{\mathrm{ext}}$ is the external demand (i.e. withdrawal) of water at the ith node; and A_ij( = A_ji) is the element of the adjacency matrix. In other words, A_ij = 1 if there is a pipe between the ith and jth nodes. Otherwise, A_ij = 0.

Equations (2.1) and (3.1) imply that there are N − N₀ + M unknowns and the same number of equations for solving the steady state and its linear stability. Now we erase h_i (1≤i≤N − N₀) from equation (2.1) using equation (3.1) to derive a self-contained M-dimensional dynamical system. It should be noted that we do not have to erase h_i (N − N₀ + 1≤i≤N) from equation (2.1) because these h_i's are constant. To erase h_i (1≤i≤N − N₀), we differentiate equation (3.1) to obtain

$$\sum _{j=1;j\ne i}^N{A}_{ij}\frac{\mathrm{d}{Q}_{ij}}{\mathrm{d}t}=0(i=1,2,\dots ,N-N_0).$$ 3.2

By substituting equation (2.1) in equation (3.2), one obtains

$$\sum _{j=1;j\ne i}^N{A}_{ij}\frac{h_i-h_j-R_{ij}{Q}_{ij}|Q_{ij}|-u(Q_{ij})}{I_{ij}}=0(i=1,2,\dots ,N-N_0),$$ 3.3

where I_ij is the inertia constant of pipe (i, j). Note that equation (2.2) implies that R_ij = R_ji.

We define the M-dimensional diagonal matrix D by the diagonal elements D_mm = 1/I_m (1≤m≤M). We also denote the head vector by h = (h₁, …, h_N)^→ p, where → p denotes the transposition. Furthermore, we denote q = (q₁, …, q_M)^→ p, of which the mth element is given by

$$q_m(Q_m)\equiv R_m{Q}_m|Q_m|+u(Q_m).$$ 3.4

To rewrite equation (3.3), we introduce the incidence matrix, denoted by C, which is an N × M matrix. By definition, corresponding to each mth edge (i, j), the entries of the mth column of C are given by C_i,m = 1, C_j,m =  − 1 and C_ℓ,m = 0 (1≤ℓ≤N, ℓ≠i, j). Although the edges are undirected, we assign +1 and −1 to i and j, respectively, by arbitrarily selecting either node forming the mth edge to which +1 is assigned, for later convenience. We represent C in block form, given by

$$C=\left(\begin{array}{c}{C}^{(1)}\\ C^{(2)}\end{array}\right),$$ 3.5

where the (N − N₀) × M matrix C⁽¹⁾ is the part of the incidence matrix corresponding to the nodes whose total energy is not fixed, and the N₀ × M matrix C⁽²⁾ is the part of the incidence matrix corresponding to the nodes whose total energy is fixed.

We further denote the N × N combinatorial Laplacian matrix of the network by

$$L_{ij}=\{\begin{array}{ll}\sum _{j^{\mathrm{\prime }}=1;j^{\mathrm{\prime }}\ne i}^N{A}_{i{j}^{\mathrm{\prime }}}& (i=j),\\ -A_{ij}& (i\ne j).\end{array}$$ 3.6

Using

$$L=C{C}^{\mathrm{\top }},$$ 3.7

we obtain the following block format of the Laplacian matrix:

$$L=\left(\begin{array}{cc}{L}^{(11)}& L^{(12)}\\ L^{(21)}& L^{(22)}\end{array}\right)=\left(\begin{array}{cc}{C}^{(1)}{C}^{(1)\mathrm{\top }}& C^{(1)}{C}^{(2)\mathrm{\top }}\\ C^{(2)}{C}^{(1)\mathrm{\top }}& C^{(2)}{C}^{(2)\mathrm{\top }}\end{array}\right),$$ 3.8

where L⁽¹¹⁾ is the (N − N₀) × (N − N₀) matrix corresponding to the nodes whose total energy is not fixed, L⁽¹²⁾ is an (N − N₀) × N₀ matrix and so forth.

The combinatorial Laplacian matrix for the network with the same set of edges but with the edge weight given by D_mm = 1/I_m is called the conductance matrix in electrical circuit theory [36]. We denote the conductance matrix by L_w. The conductance matrix is given by

$$L_{\mathrm{w}}=\left(\begin{array}{cc}{L}_{\mathrm{w}}^{(11)}& L_{\mathrm{w}}^{(12)}\\ L_{\mathrm{w}}^{(21)}& L_{\mathrm{w}}^{(22)}\end{array}\right)=CD{C}^{\mathrm{\top }}.$$ 3.9

Then, we can rewrite equation (3.3) as

$$(L_{\mathrm{w}}^{(11)}{L}_{\mathrm{w}}^{(12)})\mathit{h}-C^{(1)}D\mathit{q}=0.$$ 3.10

With the notation

$$\mathit{h}=\left(\begin{array}{c}{\mathit{h}}^{\mathbf{(}\mathbf{1}\mathbf{)}}\\ {\mathit{h}}^{\mathbf{(}\mathbf{2}\mathbf{)}}\end{array}\right),$$ 3.11

where h⁽¹⁾ = (h₁, …, h_N−N0)^→ p and h⁽²⁾ = (h_{N−N0 + 1}, …, h_N)^→ p, equation (3.10) is equivalent to

$$C^{(1)}D{C}^{(1)\mathrm{\top }}{\mathit{h}}^{\mathbf{(}\mathbf{1}\mathbf{)}}=-C^{(1)}D{C}^{(2)\mathrm{\top }}{\mathit{h}}^{\mathbf{(}\mathbf{2}\mathbf{)}}+C^{(1)}D\mathit{q}.$$ 3.12

Matrix C⁽¹⁾DC^{(1) → p} is the so-called loopy Laplacian matrix [36]. Because all of its eigenvalues are positive if the network is a connected network, which we have assumed, C⁽¹⁾DC^{(1) → p} has the inverse and one obtains

$${\mathit{h}}^{\mathbf{(}\mathbf{1}\mathbf{)}}={(C^{(1)}D{C}^{(1)\mathrm{\top }})}^{-1}(-C^{(1)}D{C}^{(2)\mathrm{\top }}{\mathit{h}}^{\mathbf{(}\mathbf{2}\mathbf{)}}+C^{(1)}D\mathit{q}).$$ 3.13

This procedure is essentially the same as the network Kron reduction [36].

By substituting equation (3.13) in equation (2.1), one obtains

$$\frac{\mathrm{d}\mathit{Q}}{\mathrm{d}t}=D{C}^{\mathrm{\top }}\left(\begin{array}{c}{(C^{(1)}D{C}^{(1)\mathrm{\top }})}^{-1}(-C^{(1)}D{C}^{(2)\mathrm{\top }}{\mathit{h}}^{\mathbf{(}\mathbf{2}\mathbf{)}}+C^{(1)}D\mathit{q}(\mathit{Q}))\\ {\mathit{h}}^{\mathbf{(}\mathbf{2}\mathbf{)}}\end{array}\right)-D\mathit{q}(\mathit{Q}),$$ 3.14

where Q≡(Q₁, …, Q_M)^→ p. Equation (3.14) is an M-dimensional dynamical system. However, it should be noted that the dynamics are confined on an (M − N + N₀)-dimensional hyperplane specified by equation (3.1), respecting Kirchhoff's current law.

(b). Local stability analysis

In this section, we derive the eigenequation that determines the conventional local linear stability of the steady state of the nonlinear dynamics given by equation (2.1). We use the combination of equations (2.1) and (3.1) for solving the steady state and equation (3.14) to perform the local stability analysis around the obtained steady state.

The steady state is given by setting the left-hand side of equation (2.1) to zero. By combining the resulting M constraints with the N − N₀ constraints provided by equation (3.1), one can obtain the steady state, which consists of the N − N₀ + M unknowns, i.e. h_i (1≤i≤N − N₀) and Q_m (1≤m≤M).

To set up the Newton–Raphson iteration scheme (e.g. [26]), we rewrite equation (3.1) as

$$F_i^{\mathrm{node}}(\mathit{Q})\equiv \sum _{j=1;j\ne i}^N{A}_{ij}{Q}_{ij}+{\tilde{Q}}_i^{\mathrm{ext}}=0(i=1,2,\dots ,N-N_0).$$ 3.15

We set the left-hand side of equation (2.1) to 0 to obtain

$$F_m^{\mathrm{pipe}}({\mathit{h}}^{\mathbf{(}\mathbf{1}\mathbf{)}},\mathit{Q})\equiv h_{i_m}-h_{j_m}-R_m{Q}_m|Q_m|-u(Q_m)=0,$$ 3.16

where i_m and j_m are the two nodes connected by the mth pipe; we use i_m and j_m instead of i and j to avoid possible notational conflict with equation (3.15). The Jacobian of the right-hand side of equations (3.15) and (3.16) with N + M unknowns, h₁, …, h_N−N0, Q₁, …, Q_M, is given by

$$J=\left(\begin{array}{cc}O& C^{(1)}\\ {(C^{(1)})}^{\mathrm{\top }}& -\text{diag}(q_m^{\mathrm{\prime }}(Q_m))\end{array}\right),$$ 3.17

where diag(q′_m(Q_m)) is the diagonal matrix whose diagonal elements are given by

$$q_m^{\mathrm{\prime }}(Q_m)=2R_m\mathrm{sgn}(Q_m)Q_m+u^{\mathrm{\prime }}(Q_m).$$ 3.18

The updating equation for the Newton–Raphson method is given by

$$\left(\begin{array}{c}{\mathit{h}}^{(1)(n+1)}\\ {\mathit{Q}}^{(n+1)}\end{array}\right)=\left(\begin{array}{c}{\mathit{h}}^{(1)(n)}\\ {\mathit{Q}}^{(n)}\end{array}\right)-J^{-1}\left(\begin{array}{c}{\mathit{F}}_1^{\mathrm{node}}({\mathit{Q}}^{(n)})\\ ⋮\\ {\mathit{F}}_{N-N_0}^{\mathrm{node}}({\mathit{Q}}^{(n)})\\ {\mathit{F}}_1^{\mathrm{pipe}}({\mathit{h}}^{(1)(n)},{\mathit{Q}}^{(n)})\\ ⋮\\ {\mathit{F}}_M^{\mathrm{pipe}}({\mathit{h}}^{(1)(n)},{\mathit{Q}}^{(n)})\end{array}\right).$$ 3.19

Denote the obtained steady state by (h*₁, …, h*_N−N0, Q*₁, …, Q*_M). We now carry out the local stability analysis around this steady state. The Jacobian matrix of the right-hand side of equation (3.14), denoted by J^dyn = (J^dyn_mm′), is given by

$$J^{\mathrm{dyn}}=[D{C}^{(1)\mathrm{\top }}{(C^{(1)}D{C}^{(1)\mathrm{\top }})}^{-1}{C}^{(1)}D-D]\mathrm{diag}(q_m^{\mathrm{\prime }}(Q_m^{\ast })).$$ 3.20

Proposition 3.1. —

Matrix B≡DC^{(1) → p}(C⁽¹⁾DC^{(1) → p})⁻¹C⁽¹⁾D − D has (N − N₀)-fold zero eigenvalues and M − N + N₀ negative eigenvalues.

Proof. —

A direct substitution verifies

$$B{(C^{(1)})}^{\mathrm{\top }}=0.$$ 3.21

Therefore, B has (N − N₀)-fold zero eigenvalues, and the corresponding N − N₀ right eigenvectors are the columns of (C⁽¹⁾)^→ p.

Because (C⁽¹⁾DC^{(1) → p})⁻¹ is an (N − N₀) × (N − N₀) matrix, its rank is at most N − N₀. Therefore, the M × M matrix B′≡C^{(1) → p}(C⁽¹⁾DC^{(1) → p})⁻¹C⁽¹⁾D has at most rank N − N₀ and therefore has at least M − N + N₀ zero eigenvalues.

We denote the eigenvalues of a symmetric matrix X by λ₁(X)≤ · s≤λ_M(X). Using the fact that D is positive definite because all its diagonal elements are positive (i.e. =1/I_m), Weyl's theorem on eigenvalues applied to the sum of B′ and −D, which is equal to B, implies that

$${\lambda }_m(B)<{\lambda }_m(B^{\mathrm{\prime }}).$$ 3.22

Note that λ_m0(B′) = λ_m0 + 1(B′) =  · s = λ_{m0 + M − N + N0 − 1}(B′) = 0 for an m₀ (1≤m₀≤N − N₀ + 1). Therefore, equation (3.22) yields λ_m0(B), …, λ_{m0 + M − N + N0 − 1}(B) < 0. Because this relationship must be consistent with the fact that B has (N − N₀)-fold zero eigenvalues, we obtain m₀ = 1 and λ₁(B), …, λ_M−N+N0(B) < 0, λ_{M−N+N0 + 1}(B) =  · s = λ_M(B) = 0. □

Let us assume for any pipe m that q′_m(Q*_m) > 0, which means that the head loss owing to friction along the pipe increases as the steady-state flow rate, Q*_m, increases. We verified that q′_m(Q*_m) > 0 held true in all the numerical simulations carried out in the next section. Then, J^dyn has M − N + N₀ negative eigenvalues and N − N₀ zero eigenvalues, as matrix B does (appendix A).

Therefore, the dynamics given by equation (3.14) are locally neutrally stable in the subspace spanned by the column vectors of [diag(q′_m(Q*_m))]⁻¹(C⁽¹⁾)^→ p. However, perturbations along these directions are infeasible because such a perturbation is not on the hyperplane defined by equation (3.1) on which the dynamics are constrained. Note that the normal vectors of the hyperplane are given by the row vectors of C⁽¹⁾, i.e. column vectors of (C⁽¹⁾)^→ p. Therefore, in the local stability analysis, we should ignore these N − N₀ zero eigenvalues and inspect the other M − N + N₀ eigenvalues. With this consideration, we conclude that the steady state constrained by Kirchhoff's current law (i.e. equation (3.1)) is always locally stable.

To define a local stability index, we calculate the largest eigenvalue of J^dyn except the (N − N₀)-fold zero eigenvalue. Note that all eigenvalues of J^dyn are real, although J^dyn is not a symmetric matrix in general (appendix A). We consider the largest non-zero eigenvalue of J^dyn (which is negative) and refer to its absolute value as the local stability index, denoted by ρ. A large ρ-value implies that dynamics perturbed by a small amount from the steady state would rapidly return to the steady state. Mathematically, the perturbations are given to the flow rate for some pipes. However, we expect that ρ serves as an index that can characterize dynamical stability of water distribution networks in wider contexts.

4. Numerical results

In this section, we calculate the local stability index for various networks and examine its relevance to stability and resilience in simulated dynamics of water flow.

We use the synthesized networks used in our previous study [18]. As supplementary materials to [18], the full network structure and properties of system components (e.g. pipes, nodal demands, elevations) were made open to the public in EPANET2 format, except for one network. For completeness, we have made the complete data for all networks and their system components used in the present study available in Matlab format at Github, alongside the code for calculating the local stability index.

Of the 85 networks used in our previous study [18], five networks included pumps, which would pin the flow rate to a constant value. Because our formalism does not directly cover the case of the pinned flow rate, we exclude these five networks. Of the remaining 80 networks, 10 networks have N = 102 nodes, N₀ = 2 reservoir nodes among the N nodes and M = 110 pipes; 11 networks have (N, N₀, M) = (102, 2, 130); 10 networks have (N, N₀, M) = (204, 4, 223); 10 networks have (N, N₀, M) = (204, 4, 263); nine networks have (N, N₀, M) = (306, 6, 335); nine networks have (N, N₀, M) = (306, 6, 395); five networks have (N, N₀, M) = (404, 4, 443); five networks have (N, N₀, M) = (404, 4, 523); five networks have (N, N₀, M) = (406, 6, 525); five networks have (N, N₀, M) = (408, 8, 447); and one network has (N, N₀, M) = (506, 6, 554). We calculated ρ for each of the 80 networks.

The diameter of a majority of the pipes was equal to 400 mm. There were also a few larger values of the diameter (e.g. 900 mm). Such wider pipes were incident (i.e. directly connected) to a reservoir. The diameter of the pipes was automatically generated by the software [37].

We set the total energy of all the N₀ reservoirs in each network to 65 m [37]. In addition to these N₀ nodes, some nodes in each network had zero water demand (i.e. ${\tilde{Q}}_i^{\mathrm{ext}}=0$), because they were pure junctions connecting different pipes. We set u(Q_m) for each pipe to zero because the data were from the 80 networks that do not have valves or pumps.

In our previous study, we numerically simulated the dynamics of water flow in these networks using the software EPANET2 [23,38] and measured six strain indices to characterize the resilience of the water distribution systems. These indices are as follows (see [18] for fuller definitions). First, the time to strain is defined as the time between the application of stress and the start of service failure, which is when the level of service at a node drops below a predefined threshold. Second, the failure duration is defined as the time needed for the system to recover to normal performance since the service failure triggered by the application of stress. Third, the failure magnitude is defined as the most severe drop at a node in terms of the system service performance as a result of the administered stress, averaged over all the nodes. Fourth, the failure rate is defined as the failure magnitude divided by the time between the start of the failure and the occurrence of the worst system performance. Fifth, the recovery rate is defined as the failure magnitude divided by the time between the worst system performance and the return to the performance threshold value used in the definition of the time to strain. Sixth, the severity is defined as the threshold minus the system performance (which is positive during the system failure) integrated with respect to time over the period of system failure.

The relationship between each of the six strain indices and the local stability index, ρ, is shown in figure 1. In the figure, each circle represents one of the 80 networks, r represents the Pearson correlation coefficient and p represents the p-value for the Pearson correlation coefficient. We found a positive correlation between ρ and the recovery rate (figure 1e). This result is intuitive because the recovery rate quantifies how the system returns to normal performance after being perturbed by external stress, which agrees with the concept of linear stability when the given external stress is infinitesimally small. The obtained correlation is only moderate (i.e. r = 0.420, p ≈ 1.04 × 10⁻⁴), probably because of various nonlinearities of the system, which our linear stability analysis does not aim to explain. However, we consider that ρ captures a strain index derived from EPANET2, which is an involved numerical simulation package, reasonably well.

Figure 1.

Relationship between each of the strain indices and the local stability index. (a) Time to strain, (b) failure duration, (c) failure magnitude, (d) failure rate, (e) recovery rate and (f ) severity. A circle represents a network and its colour represents the number of nodes. The lines represent the linear regression. The Pearson correlation coefficient and its p-value are denoted by r and p, respectively. (Online version in colour.)

Figure 1 also indicates that ρ is strongly correlated with the failure magnitude (figure 1c) and the failure rate (figure 1d). We do not have an active interpretation of this result. In fact, it appears contradictory that a large ρ-value, which predicts that the network is resilient, is associated with a large failure magnitude and failure rate. However, the failure magnitude and failure rate quantify the magnitude of failure response induced by the external stress, which is not related to our local stability analysis even in the limit of infinitesimally small external stress. Therefore, we conclude that the failure magnitude and the failure rate characterize different aspects of the system's resilience from what the recovery rate and ρ do. This view is consistent with our additional observation that, within our previous numerical results [18], there is positive correlation between the failure magnitude and the recovery rate (r = 0.917, p < 10⁻³²) and between the failure rate and the recovery rate (r = 0.677, p ≈ 5.60 × 10⁻¹²).

In our previous study, we measured eight structural properties of the water distribution network and examined the association between each of them and each of the six strain indices [18]. Here we measured the Pearson correlation coefficient between each of these structural properties, whose definitions are in appendix B, and the recovery rate. We found that none of the eight structural properties of the network was as strongly correlated with the recovery rate as ρ was (link density: r = 0.145, p = 0.204; algebraic connectivity: r = 0.114, p = 0.316; diameter: r =  − 0.143, p = 0.207; average path length: r =  − 0.204, p = 0.0717; central point dominance: r =  − 0.0515, p = 0.652; heterogeneity: r = 0.335, p = 0.0026; spectral gap: r =  − 0.0509, p = 0.656; clustering coefficient: r = 0.198, p = 0.0798). Therefore, the local stability index is better at capturing the recovery rate than these structural properties. This is probably because these structural properties do not directly quantify the speed at which the water distribution network responds to a perturbation in the steady state.

5. Discussion

We carried out a local stability analysis of water flow in pipe networks including reservoir nodes and consumer nodes. The steady state was shown to be always linearly stable. We used the eigenvalue corresponding to the slowest relaxation mode as the local stability index. The proposed index was moderately correlated with the recovery rate in response to external stress, which was numerically obtained from an involved simulator commonly used in the water engineering research community.

We used a particular equation for the friction factor [34]. However, the present framework accepts other equations for the friction factor. Because the differentiation of the friction factor with respect to the Reynolds number is used in calculating the local stability index and the value of the friction factor depends on the pipe in general, one should use an equation that is continuous and covers a wide range of the Reynolds number and hence different flow schemes. Some possible choices of the equation different from the one used in the present paper are found in the literature [39–41]. It is straightforward to extend our local stability index to the case of different functional forms of the friction factor.

In 1972, May proposed to examine the eigenvalue spectra of interaction networks to characterize the dynamical stability of complex systems, in particular ecological systems [42]. This is a landmark example of local stability analysis. While the original analysis and many studies that ensued were for random networks [42], the understanding has been extended to the case of interaction networks with general network structure [43,44]. This family of methods is similar to the one proposed in the present study. The main differences are that we have started from a particular set of hydraulic equations modelling water flow in pipe networks and that we have found that our system is always locally stable and hence used the relaxation mode to quantify the extent of the local stability of the system.

The proposed local stability index was significantly correlated with the recovery rate, which was obtained from numerical simulations in which the applied shock was not necessarily small. However, because the present analysis is a local stability analysis, it does not tell us the behaviour of the system when a shock injected to the steady state is not small. In water engineering applications, the magnitude of perturbation is not necessarily small [3]. A recent study in network science has developed a formalism that reduces a class of dynamical systems on interaction networks into a one-dimensional effective dynamical system and assesses the resilience of the dynamical system against a perturbation that is not necessarily small [45]. Therefore, adapting this resilience formalism to the case of transient and steady-state dynamics of water flow in pipe networks is of practical relevance. In equation (2.1), there are usually only a fraction of nodes whose energy value is fixed (i.e. reservoirs) and different nodes have different water demands. In this sense, the flow dynamics on pipe networks are heterogeneous in addition to being heterogeneous in the network structure and edge weight. Other structures such as valves would add more complexity and heterogeneity to the system. For such a heterogeneous system, it is unclear whether one can transform equation (2.1) into a form that is compatible with the resilience function formalism [45,46]. This issue is left as future work.

Appendix A. Eigenvalues of J^dyn

Here we provide an elementary proof that J^dyn has M − N + N₀ zero eigenvalues and N − N₀ negative eigenvalues when q′_m(Q*_m) > 0, 1≤m≤M.

Let us introduce a short-hand notation $\overline{D}\equiv \mathrm{diag}(q_m^{\mathrm{\prime }}(Q_m^{\ast }))$. Then, one obtains $J^{\mathrm{dyn}}=B\overline{D}$. Because J^dyn is similar to the symmetric matrix ${\overline{D}}^{1/2}B{\overline{D}}^{1/2}$, the min–max theorem for the kth smallest eigenvalue implies that

$$\begin{array}{rl}{\lambda }_k(J^{\mathrm{dyn}})={\lambda }_k({\overline{D}}^{1/2}B{\overline{D}}^{1/2})& =\underset{\begin{array}{c}S\subset {\mathbb{R}}^M\\ \dim (S)=k\end{array}}{min}\left[\underset{x\in S\mathrm{\setminus }\{0\}}{max}\frac{({\overline{D}}^{1/2}B{\overline{D}}^{1/2}x,x)}{(x,x)}\right]\\ & =\underset{\begin{array}{c}S\subset {\mathbb{R}}^M\\ \dim (S)=k\end{array}}{min}[\underset{x\in S\mathrm{\setminus }\{0\}}{max}\frac{({\overline{D}}^{1/2}B{\overline{D}}^{1/2}x,x)}{({\overline{D}}^{1/2}x,{\overline{D}}^{1/2}x)}\cdot \frac{(\overline{D}x,x)}{(x,x)}]\\ & \le \underset{\begin{array}{c}S\subset {\mathbb{R}}^M\\ \dim (S)=k\end{array}}{min}[\underset{x\in S\mathrm{\setminus }\{0\}}{max}\frac{({\overline{D}}^{1/2}B{\overline{D}}^{1/2}x,x)}{({\overline{D}}^{1/2}x,{\overline{D}}^{1/2}x)}\cdot \underset{x^{\mathrm{\prime }}\in {\mathbb{R}}^M\mathrm{\setminus }\{0\}}{max}\frac{(\overline{D}{x}^{\mathrm{\prime }},x^{\mathrm{\prime }})}{(x^{\mathrm{\prime }},x^{\mathrm{\prime }})}]\\ & ={\lambda }_k(B){\lambda }_M(\overline{D}).\end{array}$$ A 1

Similarly, one obtains

$$\begin{array}{rl}{\lambda }_k(J^{\mathrm{dyn}})& =\underset{\begin{array}{c}S\subset {\mathbb{R}}^M\\ \dim (S)=k\end{array}}{min}[\underset{x\in S\mathrm{\setminus }\{0\}}{max}\frac{({\overline{D}}^{1/2}B{\overline{D}}^{1/2}x,x)}{({\overline{D}}^{1/2}x,{\overline{D}}^{1/2}x)}\cdot \frac{(\overline{D}x,x)}{(x,x)}]\\ & \ge \underset{\begin{array}{c}S\subset {\mathbb{R}}^M\\ \dim (S)=k\end{array}}{min}[\underset{x\in S\mathrm{\setminus }\{0\}}{max}\frac{({\overline{D}}^{1/2}B{\overline{D}}^{1/2}x,x)}{({\overline{D}}^{1/2}x,{\overline{D}}^{1/2}x)}\cdot \underset{x^{\mathrm{\prime }}\in {\mathbb{R}}^M\mathrm{\setminus }\{0\}}{min}\frac{(\overline{D}{x}^{\mathrm{\prime }},x^{\mathrm{\prime }})}{(x^{\mathrm{\prime }},x^{\mathrm{\prime }})}]\\ & ={\lambda }_k(B){\lambda }_1(\overline{D}).\end{array}$$ A 2

By combining ${\lambda }_1(\overline{D})>0$, ${\lambda }_M(\overline{D})>0$ and λ₁(B), …, λ_M−N+N0 < 0, λ_{M−N+N0 + 1} =  · s = λ_M = 0, one obtains λ₁(J^dyn), …, λ_M−N+N0(J^dyn) < 0, λ_{M−N+N0 + 1}(J^dyn) =  · s = λ_M(J^dyn) = 0.

Appendix B. Structural properties of networks

The eight structural properties of the network that we measured in our previous study [18] and were used in the present article for comparison purposes are as follows.

The link density is given by 2M/[N(N − 1)]. The algebraic connectivity is the smallest positive eigenvalue of the Laplacian matrix, L, of the undirected and unweighted network. The version of the clustering coefficient used in [18] is given by 3 × (number of triangles in the network)/(number of connected triples of nodes in the network). The average path length is equal to the shortest path length between two nodes, which is averaged over all the possible N(N − 1)/2 node pairs. The central point dominance is given by $\sum _{i=1}^N(b_{max}-b_i)/(N-1)$, where b_i is the betweenness centrality of the ith node and $b_{max}=\underset{j=1,\dots ,N}{max}{b}_j$. The heterogeneity is that in the degree and is given by $\sum _{i=1}^N{(k_i-⟨k⟩)}^2/⟨k⟩$, where k_i is the degree of the ith node and $⟨k⟩=\sum _{i=1}^N{k}_i/N$ is the average degree of the node. The version of the spectral gap used in [18] is given by the difference between the two largest eigenvalues of L. The modularity is a quantity representing the quality of community structure detected in the network and approximate modularity maximization was carried out using Newman's algorithm [47].

Data accessibility

The data on the network structure and system components are available on Github.

Competing interests

We declare we have no competing interests.

Author's contributions

N.M. and F.M. conceived of and designed the study, discussed the results and drafted the manuscript. F.M. provided the data on the network structure and system components. N.M. carried out the theoretical analysis and performed the data analysis. Both authors read and approved the manuscript.

Funding

N.M. and F.M. acknowledge the support provided through the BRIM (Building Resilience into Risk Management) project, EPSRC (grant no. EP/N010329/1).