Skip to main content

Operational decisions for multi-period industrial gas pipeline networks under uncertainty


An industrial gas company (IGC) manages a large pipeline network consisting of several gas production plants and many customers. Natural gas is an input to the production process at each plant, and a different industrial gas (e.g., hydrogen or oxygen) is the plant output. Each day the company must propose 24 h in advance a “nomination value” for the amount of natural gas they plan to use at each plant over the next day. The nomination value goes to the natural gas provider who then offers a set price for the nominated amount. If the nomination amount underestimates actual usage, the IGC must purchase the extra gas at a (typically higher) spot market price. If the nomination value overestimates usage, the IGC sells extra gas at a (typically lower) spot market price. In this study, we develop a multi-period model to determine decisions for daily nomination values of natural gas for multiple plants as well as production and distribution decisions for a steady-state behaving industrial gas at distribution network, where the direction of the gas flow is not known at each pipe. The model considers demands from multiple customers to be uncertain and varying throughout the day. Daily planning requires a multi-period formulation that includes ramping constraints (physical constraints that limit how much a plant’s production can change over each period). The model also dictates that customer demands be met with high probability and that the physical constraints of the pipeline are met. The result is a challenging multi-period, chance constrained, nonlinear integer optimization problem. We propose an optimization-based heuristic approach to find a high-quality feasible solution for industry-sized problems. Also, we use a convex relaxation of the problem to obtain valid lower bound for the problem in order to obtain a measure of the near-optimality of the problem’s feasible solution found via the introduced heuristic approach. Numerical experiments show that the proposed heuristic is able to efficiently find near-optimal solutions for the problem instances in which using an exact solution approach is prohibitively time consuming.

This is a preview of subscription content, access via your institution.

Fig. 1
Fig. 2
Fig. 3
Fig. 4
Fig. 5
Fig. 6
Fig. 7
Fig. 8
Fig. 9
Fig. 10


  • Behrooz HA (2016) Managing demand uncertainty in natural gas transmission networks. J Nat Gas Sci Eng 34:100–111

    Article  Google Scholar 

  • Behrooz HA, Boozarjomehry RB (2017) Dynamic optimization of natural gas networks under customer demand uncertainties. Energy 134:968–983

    Article  Google Scholar 

  • Benson HY, Vanderbei RJ (2003) Solving problems with semidefinite and related constraints using interior-point methods for nonlinear programming. Math Program 95(2):279–302

    MathSciNet  Article  MATH  Google Scholar 

  • Bonami P, Biegler LT, Conn AR, Cornuéjols G, Grossmann IE, Laird CD, Lee J, Lodi A, Margot F, Sawaya N et al (2008) An algorithmic framework for convex mixed integer nonlinear programs. Discrete Optim 5(2):186–204

    MathSciNet  Article  MATH  Google Scholar 

  • Borraz-Sánchez C, Bent R, Backhaus S, Hijazi H, Hentenryck PV (2016) Convex relaxations for gas expansion planning. INFORMS J Comput 28(4):645–656

    MathSciNet  Article  MATH  Google Scholar 

  • Bradley S, Hax A, Magnanti T (1977) Applied mathematical programming. Addison-Wesley Publishing Company. ISBN 9780201004649

  • Bruno SV, Moraes LA, de Oliveira W (2015) Optimization techniques for the Brazilian natural gas network planning problem. Energy Systems, pp 1–21

  • Cay P, Esmali A, Mancilla C, Storer RH, Zuluaga LF (2018) Solutions with performance guarantees on tactical decisions for industrial gas network problems. IISE Trans 50(8):654–667

    Article  Google Scholar 

  • Chebouba A, Yalaoui F, Smati A, Amodeo L, Younsi K, Tairi A (2009) Optimization of natural gas pipeline transportation using ant colony optimization. Comput Oper Res 36(6):1916–1923

    Article  MATH  Google Scholar 

  • Coelho PM, Pinho C (2007) Considerations about equations for steady state flow in natural gas pipelines. J Braz Soc Mech Sci Eng 29(3):262–273

    Article  Google Scholar 

  • CPLEX II (2009) V12. 1: Users manual for cplex. Int Bus Mach Corp 46(53):157

    Google Scholar 

  • De Wolf D, Smeers Y (2000) The gas transmission problem solved by an extension of the simplex algorithm. Manag Sci 46(11):1454–1465

    Article  MATH  Google Scholar 

  • Demissie A, Zhu W (2015) A survey on gas pipelines operation and design optimization. In: Industrial and system engineering annual research conference and expo 2015

  • Fourer R, Gay D, Kernighan B (2002) Ampl: a modeling language for mathematical programming. Duxbury Press, North Scituate

    MATH  Google Scholar 

  • Gotzes C, Heitsch H, Henrion R, Schultz R (2016) On the quantification of nomination feasibility in stationary gas networks with random load. Math Methods Oper Res 84(2):427–457

    MathSciNet  Article  MATH  Google Scholar 

  • Guldmann J-M (1983) Supply, storage, and service reliability decisions by gas distribution utilities: a chance-constrained approach. Manag Sci 29(8):884–906

    Article  Google Scholar 

  • Jiao Y, Su H, Hou W, Liao Z (2012) Optimization of refinery hydrogen network based on chance constrained programming. Chem Eng Res Des 90(10):1553–1567

    Article  Google Scholar 

  • Kalashnikov VV, Pérez-Valdés GA, Tomasgard A, Kalashnykova NI (2010) Natural gas cash-out problem: bilevel stochastic optimization approach. Eur J Oper Res 206(1):18–33

    MathSciNet  Article  MATH  Google Scholar 

  • Kaut M, Midthun KT, Werner AS, Tomasgard A, Hellemo L, Fodstad M (2013) Multi-horizon stochastic programming. Comput Manag Sci.

  • Koch T, Hiller B, Pfetsch M, Schewe L (2015) Evaluating gas network capacities. MOS-SIAM series on optimization. SIAM. ISBN 9781611973693

  • Liang X, Kang L, Liu Y (2017) Impacts of subperiod partitioning on optimization of multiperiod hydrogen networks. Ind Eng Chem Res 56(38):10733–10742

    Article  Google Scholar 

  • Liu B (2009) Theory and practice of uncertain programming. Studies in fuzziness and soft computing. Springer. ISBN 9783540894834

  • Maggioni F, Bertocchi M, Giacometti R, Vespucci M, Innorta M, Allevi E (2009) A stochastic optimization model for gas retail with temperature scenarios and oil price parameters. IMA J Manag Math 21:149–163

    MathSciNet  Article  MATH  Google Scholar 

  • McCormick GP (1976) Computability of global solutions to factorable nonconvex programs: part I–convex underestimating problems. Math Program 10(1):147–175

    Article  MATH  Google Scholar 

  • Mesfin G, Shuhaimi M (2010) A chance constrained approach for a gas processing plant with uncertain feed conditions. Comput Chem Eng 34(8):1256–1267

    Article  Google Scholar 

  • Midthun K, Hellemo L, Kaut M, Werner A (2015) Assessing production assurance in a natural gas network by using scenario generation and optimization. Energy Procedia 64:120–129

    Article  Google Scholar 

  • Odetayo B, MacCormack J, Rosehart WD, Zareipour H (2017) A sequential planning approach for distributed generation and natural gas networks. Energy 127:428–437

    Article  Google Scholar 

  • Ríos-Mercado RZ, Borraz-Sánchez C (2015) Optimization problems in natural gas transportation systems: a state-of-the-art review. Appl Energy 147:536–555

    Article  Google Scholar 

  • Schmidt M, Aßmann D, Burlacu R, Humpola J, Joormann I, Kanelakis N, Koch T, Oucherif D, Pfetsch ME, Schewe L, Schwarz R, Sirvent M (2017) Gaslib–a library of gas network instances. Data 2(4):40

    Article  Google Scholar 

  • Tian Q, Zhao D, Li Z, Zhu Q (2017) Robust and stepwise optimization design for co2 pipeline transportation. Int J Greenhouse Gas Control 58:10–18

    Article  Google Scholar 

  • Tomasgard A, Rømo F, Fodstad M, Midthun K (2007) Optimization models for the natural gas value chain. Geometric modelling, numerical simulation, and optimization. Springer. ISBN 9783540687825

  • van den Heever SA, Grossmann IE (2003) A strategy for the integration of production planning and reactive scheduling in the optimization of a hydrogen supply network. Comput Chem Eng 27(12):1813–1839

    Article  Google Scholar 

  • You F, Pinto JM, Grossmann IE, Megan L (2011) Optimal distribution-inventory planning of industrial gases. II. minlp models and algorithms for stochastic cases. Ind Eng Chem Res 50(5):2928–2945

    Article  Google Scholar 

  • Zhuang DX, Jiang JN, Gan D (2011) Optimal short-term natural gas nomination decision in generation portfolio management. Eur Trans Electr Power 21(1):1–10

    Article  Google Scholar 

  • Zuluaga L, Terlaky T (2013) Modeling and optimization: theory and applications: selected contributions from the MOPTA 2012 conference. Springer Proceedings in Mathematics & Statistics. Springer, New York. ISBN 9781461489870

Download references


The first author was supported by the Pennsylvania Infrastructure Technology Alliance (PITA) Grant PIT-13-17.

Author information

Authors and Affiliations


Corresponding author

Correspondence to Pelin Cay.

Additional information

Publisher's Note

Springer Nature remains neutral with regard to jurisdictional claims in published maps and institutional affiliations.


Appendix 1: MSP-A model formulation

This problem is basically just the original problem but with a reduced number of periods (subproblems), such that it focuses only on periods in \(S_s\) (as in Eq. 3). There are two versions of this model, depending on the starting period. If we begin in period 0, then we include constraints (2j) and (2m). If the beginning period is not period 0, then we set this period’s initial production equal to the ending production level from the previous period (which exists because we solve these problems forward in time). Thus, we do not include constraints (2j) and (2m) in this case. In this problem, we solve the original formulation of the problem i.e. Model 2 with difference on considered period length (\(S_s^i\), \(\forall i \in \{1 \dots N\}\), for simplicity we refer \(S_s\) in the formulation). This set is updated for each subproblem. The model formulation for MSP-A is as follows

$$\begin{aligned}&\displaystyle \mathop {{\text{ minimize }}}\limits _{y,x,{\bar{p}},q,{\mathcal {H}},s,w}&\displaystyle \;\;\;\;\; \sum _{i \in {\mathcal {P}}} {\hat{p}}_{i}{\mathcal {H}}_{i} + q_{i} \end{aligned}$$
$$\begin{aligned}&\text{ subject } \text{ to }&\displaystyle \;\;\;q_{i} \ge p^+_i\left( \sum _{t \in {\mathcal {S}}_s,k \in \{1 .. {\tilde{s}}\}}(\delta _{kit}s_{kit})- {\mathcal {H}}_{i}\right) \qquad&\qquad { \forall i \in {\mathcal {P}}} \end{aligned}$$
$$\begin{aligned}&\displaystyle q_{i} \ge p^-_i\left( {\mathcal {H}}_{i} -\sum _{t \in {\mathcal {S}}_s, k \in \{1 .. {\tilde{s}} \}}(\delta _{kit}s_{kit})\right) \qquad&\qquad { \forall i \in {\mathcal {P}}} \end{aligned}$$
$$\begin{aligned}&\displaystyle y_{it} = \sum _{k \in \{1 .. {\tilde{s}}\}}s_{kit} \qquad&\qquad {\forall i \in {\mathcal {P}}, \forall t \in {\mathcal {S}}_s} \end{aligned}$$
$$\begin{aligned}&\displaystyle s_{ki}^Uw_{kit} \le s_{kit} \le s_{ki}^U \qquad&\qquad {\forall i \in {\mathcal {P}}, \forall t \in {\mathcal {S}}_s, k \in \{1\}} \end{aligned}$$
$$\begin{aligned}&\displaystyle s_{ki}^Uw_{kit} \le s_{kit} \le s_{ki}^Uw_{k-1it} \qquad&\qquad {\forall i \in {\mathcal {P}}, \forall t \in {\mathcal {S}}_s, k \in \{2 .. {\tilde{s}}-1\}} \end{aligned}$$
$$\begin{aligned}&\displaystyle 0 \le s_{kit} \le s_{kit}^Uw_{k-1it} \qquad&\qquad {\forall i \in {\mathcal {P}}, \forall t \in {\mathcal {S}}_s,k \in \{{\tilde{s}}\}} \end{aligned}$$
$$\begin{aligned}&\displaystyle \;\;\;y_{it} + \sum _{(j,i) \in {\mathcal {E}}} x_{jit} = \sum _{(i,j) \in {\mathcal {E}}} x_{ijt} \qquad&\qquad {\forall i \in {\mathcal {P}}, \forall t \in {\mathcal {S}}_s} \end{aligned}$$
$$\begin{aligned}&\displaystyle \sum _{(j,i) \in {\mathcal {E}}} x_{jit} = \sum _{(i,j) \in {\mathcal {E}}} x_{ijt} \qquad&\qquad {\forall i \in {\mathcal {K}},\forall t \in {\mathcal {S}}_s} \end{aligned}$$
$$\begin{aligned}&\displaystyle \sum _{(j,i) \in {\mathcal {E}}} x_{ji1} - \sum _{(i,j) \in {\mathcal {E}}} x_{ij1} \ge d_{i} \qquad&\qquad {\forall i \in {\mathcal {C}}} \end{aligned}$$
$$\begin{aligned}&\displaystyle \sum _{(j,i) \in {\mathcal {E}}} x_{jit} - \sum _{(i,j) \in {\mathcal {E}}} x_{ijt} \ge \mu _{it} +\Phi ^{-1}(\eta )\sigma _{it} \qquad&\qquad {\forall i \in {\mathcal {C}}, \forall t \in {\mathcal {S}}_s \setminus \{1\}} \end{aligned}$$
$$\begin{aligned}&\displaystyle \gamma _{ij}x_{ijt}\sqrt{x_{ijt}^2 + \epsilon }={\bar{p}}_{it}-{\bar{p}}_{jt} \qquad&\qquad {\forall (i,j) \in {\mathcal {E}}, \forall t \in {\mathcal {S}}_s} \end{aligned}$$
$$\begin{aligned}&\displaystyle r_i^L \le y_{i1} - {\hat{y}}_{i0} \le r_i^U \qquad&\qquad {\forall i \in {\mathcal {P}}} \end{aligned}$$
$$\begin{aligned}&\displaystyle r_i^L \le y_{it} - y_{i(t-1)} \le r_i^U \qquad&\qquad {\forall i \in {\mathcal {P}}, \forall t \in {\mathcal {S}}_s \setminus \{1\}} \end{aligned}$$
$$\begin{aligned}&\displaystyle y^L_i \le y_{it} \le y^U_i \qquad&\qquad {\forall i \in {\mathcal {P}}, \forall t \in {\mathcal {S}}_s} \end{aligned}$$
$$\begin{aligned}&\displaystyle \pi _i^- \le {\bar{p}}_{it} \le \pi _i^+ \qquad&\qquad {\forall i \in {\mathcal {V}}, \forall t \in {\mathcal {S}}_s} \end{aligned}$$
$$\begin{aligned}&\displaystyle x_{ijt} \le x^U_{ij} \qquad&\qquad {\forall (i,j) \in {\mathcal {E}}, \forall t \in {\mathcal {S}}_s} \end{aligned}$$
$$\begin{aligned}&\displaystyle - x_{ijt} \le x^U_{ij} \qquad&\qquad {\forall (i,j) \in {\mathcal {E}}, \forall t \in {\mathcal {S}}_s} \end{aligned}$$
$$\begin{aligned}&\displaystyle w_{kit} \in \{0,1\} \qquad&\qquad {\forall i \in {\mathcal {P}}, \forall t \in {\mathcal {S}}_s, k \in \{1 .. {\tilde{s}}-1\}} \end{aligned}$$

If the subproblem does not include period 1, then we drop constraint (4j) and (4m) from problem definition and solve without these constraints. At these cases, we introduce a new constraint such that we set previous subproblems’ last period production levels as starting production levels for the next subproblem and apply ramping constraint for the first period of the considered subproblem. For the rest of the problem formulations, if the time index is zero, it means that information is used either from previous subproblem’s last period information or current production information. We introduce a parameter \({\hat{y}}_{i {\hat{T}}_e}\) for all plant \(i \in {\mathcal {P}}\) which is used when the considered problem does not include period 1 in their subproblem. This parameter equals to the production levels for final period found from the previous subproblem. Then, the (4m) becomes

$$\begin{aligned} r_i^L \le y_{iT_s} - {\hat{y}}_{i{\hat{T}}_e} \le r_i^U \;\; \qquad \forall i \in {\mathcal {P}} \end{aligned}$$

This constraint mimics what (4m) does with the difference of using the solution from the previous subproblem.

Appendix 2: MSP-B model formulation

This problem includes the same constraints as Model (4), but the objective function differs because of the addition of penalty objective on the initial period production level of the considered subproblem. We introduce a new decision variable \({\bar{y}}_{iT_s}\) for all \(i \in {\mathcal {P}}\) that denotes what the initial production level should be instead of considering initial production levels, which are equal to final production levels from previous subproblem (\({\hat{y}}_{i{\hat{T}}_e}\)). Considering them in the penalty objective function deals to minimize the violation that leads infeasilbe solution of Model 4. This problem includes a penalty objective based on initial production value with penalty parameter \(\rho\) in the objective function as follows:

$$\begin{aligned}&\displaystyle \mathop {{\text{ minimize }}}\limits _{y,x,{\bar{p}},q,{\mathcal {H}},s,w,{\bar{y}}_{iT_s}}&\displaystyle \;\;\;\;\; \sum _{i \in {\mathcal {P}}} {\hat{p}}_{i}{\mathcal {H}}_{i} + q_{i}+\rho \sum _{i \in {\mathcal {P}}} | {\bar{y}}_{iT_s} - {\hat{y}}_{i{\hat{T}}_e}| \end{aligned}$$
$$\begin{aligned}&\text{ subject } \text{ to }&\displaystyle \hbox {(4b), (4c), (4d), (4e), (4f), (4g), (4h)} \end{aligned}$$
$$\begin{aligned}&\displaystyle \hbox {(4i), (4k), (4l), (4n), (4o), (4p), (4q), (4r), (4s)} \end{aligned}$$
$$\begin{aligned}&\displaystyle r_i^L \le y_{iT_s} - {\bar{y}}_{iT_s} \le r_i^U \qquad&\qquad {\forall i \in {\mathcal {P}}} \end{aligned}$$
$$\begin{aligned}&\displaystyle {\bar{y}}_{iT_s} \ge 0 \qquad&\qquad {\forall i \in {\mathcal {P}}} \end{aligned}$$

In Model (5), we can see that the constraint (5d) is different than (3m*) since the ramping at (5d) is based on new introduced decision variable instead of parameter \({\hat{y}}_{i{\hat{T}}_e}\). This constraint considers which value of previous subproblem’s production levels make this subproblem feasible by considering the ramping constraints. Lastly (5e) is for non-negative production level.

Appendix 3: MSP-C model formulation

This problem includes the same constraints as Model 4 with different objective function. As in Model (5), we introduce a new parameter \({\bar{y}}_{i{\bar{T}}_e}\) for all \(i \in {\mathcal {P}}\) that denotes what the final production level should be that is passed from Model 5 solution (\({\bar{y}}_{iT_s}\) from Model 5, initial production level for the subproblem defined for Model 5). The penalty objective with penalty parameter \(\lambda\) minimizes the violation by how much the total last production level for all plants in this subproblem can be as close as the passed value \({\hat{y}}_{iT_e}\) for all \(i \in {\mathcal {P}}\) from Model 5. This model formulation becomes as follows:

$$\begin{aligned}&\displaystyle \mathop {{\text{ minimize }}}\limits _{y,x,{\bar{p}},q,{\mathcal {H}},s,w}&\displaystyle \;\;\;\;\; \sum _{i \in {\mathcal {P}}} {\hat{p}}_{i}{\mathcal {H}}_{i} + q_{i}+\lambda \sum _{i \in {\mathcal {P}}} | y_{iT_e} - {\bar{y}}_{i{\bar{T}}_e}| \end{aligned}$$
$$\begin{aligned}&\displaystyle \hbox {(4b), (4c), (4d), (4e), (4f), (4g), (4h)} \end{aligned}$$
$$\begin{aligned}&\displaystyle \hbox {(4i), (4k), (4l), (3m*), (4n), (4o), (4p), (4q), (4r), (4s)} \end{aligned}$$
$$\begin{aligned}&\displaystyle r_i^L \le y_{it} - y_{i(t-1)} \le r_i^U \qquad&\qquad {\forall i \in {\mathcal {P}}, \forall t \in {\mathcal {S}}_s\setminus \{T_s\}} \end{aligned}$$

In Model 6, the objective function is similar as in Model 5, but the penalty is applied on the last period production level and the solution provided from Model 5 for the one subproblem ahead for the considered subproblem (this model is solved relatively one of the previous subproblems where infeasible solution is first found, you can refer examples in pages 14–15). The constrant (6d) applies ramping constraint for the periods within \(S_s\) except starting period. The starting period’s ramping constraint is represented in constraint (3m*).

Appendix 4: MSP-D model formulation

This problem includes the same constraints as Model 5 that includes additional information as a constraint such that the solution from Model 5 which is used as parameter \({\bar{y}}_{i{\bar{T}}_e}\) in Model 6 is also considered in this formulation as a new parameter \(\upsilon\) i.e, \(\upsilon _i = {\bar{y}}_{i{\bar{T}}_e}\) for each plant i in \({\mathcal {P}}\). The model formulation is as follows:

$$\begin{aligned}&\displaystyle \mathop {{\text{ minimize }}}\limits _{y,x,{\bar{p}},q,{\mathcal {H}},s,w,{\bar{y}}_{iT_s}}&\displaystyle \;\;\;\;\; \hbox {(5a)} \end{aligned}$$
$$\begin{aligned}&\text{ subject } \text{ to }&\displaystyle \;\;\; \hbox {(5b), (5c), (5d), (5e)} \end{aligned}$$
$$\begin{aligned}&\displaystyle y_{iT_e} = \upsilon _{i} \qquad&\qquad {\forall i \in {\mathcal {P}}} \end{aligned}$$

In Model 7, the objective function and the constriants are same as Model 5 with the different of constraint 7c which enforces the solution found at Model 5 is a hard constraint to for the final production levels for each plant i in \({\mathcal {P}}\).

Appendix 5: MSP-E model formulation

This problem considers the same constraints as in Model 4 with additional constraint as follows:

$$\begin{aligned}&\displaystyle \mathop {{\text{ minimize }}}\limits _{y,x,{\bar{p}},q,{\mathcal {H}},s,w}&\displaystyle \;\;\;\;\; \hbox {(4a)} \end{aligned}$$
$$\begin{aligned}&\text{ subject } \text{ to }&\displaystyle \hbox {(4b), (4c), (4d), (4e), (4f), (4g), (4h)} \end{aligned}$$
$$\begin{aligned}&\displaystyle \hbox {(4i), (4k), (4l), (3m*), (4n), (4o), (4p), (4q), (4r), (4s)} \end{aligned}$$
$$\begin{aligned}&\displaystyle \hbox {(7c)} \end{aligned}$$

The difference is, while we are solving the subproblem with the original objective function and same subproblem as in Model 4, we consider the last production value set to \(\upsilon _i\) for all \(i \in {\mathcal {P}}\), to find a feasible solution for the next subproblem.

Appendix 6: FullLp formulation of model 1

The lower bound formulation for the Model 1 is obtained by two steps of relaxation of the pressure drop constraint (1g). The first step is changing this constraint’s type from equality to inequality as introduced in Borraz-Sánchez et al. (2016)’s study which becomes as follows:

$$\begin{aligned} \gamma _{ij}x^2_{ijt} \le ({\bar{p}}_{it} - {\bar{p}}_{jt} ) (1 -2z_{ijt}) \qquad {\forall (i,j) \in {\mathcal {E}}, \forall t \in {\mathcal {S}}} \end{aligned}$$

Then, we continue by applying the steps from Cay et al. (2018) to obtain FullLp formulation of Model 1. The right hand side of (1*) is represented explicitly as follows:

$$\begin{aligned} \gamma _{ij}x^2_{ijt} \le {\bar{p}}_{it} - 2{\bar{p}}_{it}z_{ijt} - {\bar{p}}_{jt} + 2{\bar{p}}_{jt} z_{ijt} \qquad {\forall (i,j) \in {\mathcal {E}}, \forall t \in {\mathcal {S}}} \end{aligned}$$

Then, as a part of the first step, the integer reformulation of constraint (2*) is obtained by applying two times McCormick (1976) relaxations on bilinear terms. For this integer reformulation, extra variables are introduced \({\hat{w}}_{ijt}\) and \({\hat{s}}_{ijt}\) where \((i,j) \in {\mathcal {E}}, t \in {\mathcal {S}}\), where

$$\begin{aligned} {\hat{w}}_{ijt}&= {\bar{p}}_{it}z_{ijt} \qquad&{\forall (i,j) \in {\mathcal {E}}, \forall t \in {\mathcal {S}}} \end{aligned}$$
$$\begin{aligned} {\hat{s}}_{ijt}&= {\bar{p}}_{jt} z_{ijt} \qquad&{\forall (i,j) \in {\mathcal {E}}, \forall t \in {\mathcal {S}}} \end{aligned}$$

We can rewrite relation at (a) as follows:

$$\begin{aligned} {\hat{w}}_{ijt}&\le \pi _{it}^+ z_{ijt}&{\forall (i,j) \in {\mathcal {E}}, \forall t \in {\mathcal {S}}} \end{aligned}$$
$$\begin{aligned} {\hat{w}}_{ijt}&\ge \pi _{it}^- z_{ijt}&{\forall (i,j) \in {\mathcal {E}}, \forall t \in {\mathcal {S}}} \end{aligned}$$
$$\begin{aligned} {\hat{w}}_{ijt}&\ge {\bar{p}}_{it} - \pi _{it}^+(1- z_{ijt})&{\forall (i,j) \in {\mathcal {E}}, \forall t \in {\mathcal {S}}} \end{aligned}$$
$$\begin{aligned} {\hat{w}}_{ijt}&\le {\bar{p}}_{it} - \pi _{it}^-( 1- z_{ijt}) \qquad&{\forall (i,j) \in {\mathcal {E}}, \forall t \in {\mathcal {S}}} \end{aligned}$$

The same relation for (b) is as follows:

$$\begin{aligned} {\hat{s}}_{ijt}&\le \pi _{jt}^+ z_{ijt}&{\forall (i,j) \in {\mathcal {E}}, \forall t \in {\mathcal {S}}} \end{aligned}$$
$$\begin{aligned} {\hat{s}}_{ijt}&\ge \pi _{jt}^- z_{ijt}&{\forall (i,j) \in {\mathcal {E}}, \forall t \in {\mathcal {S}}} \end{aligned}$$
$$\begin{aligned} {\hat{s}}_{ijt}&\ge {\bar{p}}_{jt} - \pi _{jt}^+(1- z_{ijt})&{\forall (i,j) \in {\mathcal {E}}, \forall t \in {\mathcal {S}}} \end{aligned}$$
$$\begin{aligned} {\hat{s}}_{ijt}&\le {\bar{p}}_{jt} - \pi _{jt}^-( 1- z_{ijt}) \qquad&{\forall (i,j) \in {\mathcal {E}}, \forall t \in {\mathcal {S}}} \end{aligned}$$

Then, constraint (2*) is equivalent to

$$\begin{aligned} \gamma _{ij}x^2_{ijt} \le {\bar{p}}_{it} - 2{\hat{w}}_{ijt} - {\bar{p}}_{jt} + 2{\hat{s}}_{ijt} \qquad {\forall (i,j) \in {\mathcal {E}}, \forall t \in {\mathcal {S}}} \end{aligned}$$

including the constraints (a1)–(a4) and (b1)–(b4). Then at the end of the first step, we obtain the mixed-integer second order cone constraint optimization (MISOCO) relaxation with integer reformulation of Model 1.

At the second step, we focus on relaxing the squared flow component of the constraint (3*). Since the objective of this formulation is to obtain a lower bound, outer linearization is applied on the squared flow term. To apply the outer linearization on \(x_{ijt}^2\) for all \((i,j) \in {\mathcal {E}}\) and \(t \in {\mathcal {S}}\), B number of uniformly distributed points \(a_{ij}^k\) for all \((i,j) \in {\mathcal {E}}\) and \(k \in \{1 ... B\}\) within the bounds of \(x_{ijt}\) are used and a new decision variable \({\hat{q}}_{ijt}\) where \((i,j) \in {\mathcal {E}}\) for each \(x_{ijt}^2\) is defined. Then, the relation between \({\hat{q}}_{ijt}\) and \(x_{ijt}^2\) is represented as follows:

$$\begin{aligned} {\hat{q}}_{ijt} \ge (a_{ij}^k)^2 + 2a_{ij}^k(x_{ijt} - a_{ij}^k) \qquad \qquad {\forall (i,j) \in {\mathcal {E}}, \forall t \in {\mathcal {S}}, k \in 1 .. B} \end{aligned}$$

As applied in Cay et al. (2018), we also include an upper bound on \({\hat{q}}_{ijt}\) for all \((i,j) \in {\mathcal {E}}\) and \(t \in {\mathcal {S}}\) by adding overestimator of that term derived from McCormick (1976) relaxation which is

$$\begin{aligned} {\hat{q}}_{ijt} \le (x_{ij}^L + x_{ij}^U)x_{ijt}-x_{ij}^Lx_{ij}^U \qquad \qquad {\forall (i,j) \in {\mathcal {E}}, \forall t \in {\mathcal {S}}} \end{aligned}$$

where \(x_{ij}^L\) is lower bound on amount of gas that can be in a pipe \((i,j) \in {\mathcal {E}}\) which is zero for this study. Then, FullLp model which is the relaxation of Model 1 is formulated as follows:

$$\begin{aligned}&\displaystyle \mathop {{\text{ minimize }}}\limits _{y,x,{\bar{p}},q,z,{\mathcal {H}},s,w,{\hat{q}},{\hat{w}},{\hat{s}}}&\displaystyle \;\;\;\;\; \hbox {(2a)} \end{aligned}$$
$$\begin{aligned}&\text{ subject } \text{ to }&\displaystyle \hbox {(2b)}-\hbox {(2k)}\end{aligned}$$
$$\begin{aligned}&\displaystyle \gamma _{ij}{\hat{q}}_{ijt} \le {\bar{p}}_{it} - 2{\hat{w}}_{ijt} - {\bar{p}}_{jt} + 2{\hat{s}}_{ijt} \qquad&\qquad {\forall (i,j) \in {\mathcal {E}}, \forall t \in {\mathcal {S}}} \end{aligned}$$
$$\begin{aligned}&\displaystyle {\hat{q}}_{ijt} \ge (a_{ij}^k)^2 + 2a_{ij}^k(x_{ijt} - a_{ij}^k) \qquad&\qquad {\forall (i,j) \in {\mathcal {E}}, k \in 1 .. N, \forall t \in {\mathcal {S}}} \end{aligned}$$
$$\begin{aligned}&\displaystyle {\hat{q}}_{ijt} \le (x_{ij}^L + x_{ij}^U)x_{ijt}-x_{ij}^Lx_{ij}^U \qquad&\qquad {\forall (i,j) \in {\mathcal {E}}, \forall t \in {\mathcal {S}}} \end{aligned}$$
$$\begin{aligned}&\displaystyle \hbox {(a1)} -\hbox {(a4)}\end{aligned}$$
$$\begin{aligned}&\displaystyle \hbox {(b1)} -\hbox {(b4)}\end{aligned}$$
$$\begin{aligned}&\displaystyle \hbox {(2m)} -\hbox {(2s)} \end{aligned}$$

Model 9 is a MILP reformulation, which is a relaxation of Model 1, with additional variables and constraints. Constraint (9d) represents the outer linearization for each flow \(x_{ijt}\) in pipe \((i,j) \in {\mathcal {E}}\) at period \(t \in {\mathcal {S}}\) and constraint (9e) is an upper bound on \({\hat{q}}_{ijt}\) by adding McCormick (1976) relaxation for all \((i,j) \in {\mathcal {E}}\) and \(t \in {\mathcal {S}}\).

Rights and permissions

Reprints and Permissions

About this article

Verify currency and authenticity via CrossMark

Cite this article

Cay, P., Mancilla, C., Storer, R.H. et al. Operational decisions for multi-period industrial gas pipeline networks under uncertainty. Optim Eng 20, 647–682 (2019).

Download citation

  • Received:

  • Revised:

  • Accepted:

  • Published:

  • Issue Date:

  • DOI:


  • Gas network optimization
  • Uncertainty
  • Multi-period
  • Heuristic
  • Decision making