Multi-stage robust and multi-role integrated allocation of emergency resources for hazard-induced disasters Authors: Xin Wu, Kai Zou, Wenjie Kang Published: 2026 ABSTRACT The allocation of emergency resources is a lifeline in hazard-induced disaster relief, directly determining survival rates, rescue efficiency, and the restoration of social order. Existing allocation models face two limitations: (1) insufficient cross-role integration of funds and resource flows, and (2) limited application of multi-stage robust optimization. To address these issues, this study proposes a tripartite multi-role integrated scheduling framework that jointly considers the objectives and constraints of decision-making (fund allocation), execution (procurement and transportation), and demand roles (affected populations) to optimize economic cost, life-saving utility, and supply–demand balance. We adopt multi-stage adaptive robust optimization to reformulate the allocation model into a feedback framework integrating pre-disaster planning with post-disaster periodic adjustments under demand, procurement, and transportation uncertainties. For model tractability and compatibility with off-the-shelf solvers, we develop a linear reformulation technique that converts the nonlinear robust model into a linear program using duality theory and semi-infinite constraint handling. A case study based on an earthquake scenario shows that multi-stage adjustment and multi-role integration improve allocation optimality, stability, and robustness under multiple uncertainty sources. Sensitivity analyses of key parameters, such as time discretization and historical dependency depth, provide practical guidance for model deployment. Compared with deterministic, static robust, and rolling-horizon methods, the proposed approach shows improvements in allocation safety and overall rescue performance. Introduction Accelerating climate change and urbanization have led to increasingly frequent hazard-induced disasters (e.g., earthquakes, floods, typhoons) and public emergencies (e.g., pandemics, safety incidents)1. The IPCC AR6 report indicates that the probability of disasters such as heatwaves, torrential rains, and floods has surged over fivefold compared to pre-industrial levels2. In Q1 2025 alone, direct economic losses from hazard-induced disasters in China reached ¥10.16 billion, with compound disasters (e.g., earthquake-triggered landslides, typhoons compounded by flooding) accounting for over 40%3. These events pose severe threats to socioeconomic stability and public safety. Importantly, disaster losses stem not only from the events themselves but also from inefficient emergency response—delayed decisions or suboptimal resource allocation significantly reduce rescue effectiveness, exacerbating human suffering and economic damage. During the post-disaster “golden 72-h window,” survival rates decline precipitously—from ~ 90% within the first 24 h to 20–30% after 72 h3. This imposes high demands on the timeliness, coordination, and robustness of emergency resource allocation. Effective allocation of emergency resources is essential for minimizing casualties, mitigating property losses, and alleviating human suffering in the aftermath of disasters4. Disaster management is commonly conceptualized through a four-phase cycle—mitigation, preparedness, response, and recovery—a framework established in operations research for structuring disaster operations5. The response phase, particularly the immediate aftermath, demands rapid and coordinated action to save lives and alleviate suffering, and emergency resource allocation is its core operational challenge6. Consequently, optimizing this allocation process has become a major focus within humanitarian logistics, leading to the development of various optimization models for facility location, relief distribution, and casualty transportation, as reviewed in the literature7. The effective and equitable distribution of limited resources directly determines survival rates and the extent of socioeconomic losses6,8,9. Researchers continue to refine these models, applying OR methodologies to enhance the speed, efficiency, and fairness of emergency responses under uncertainty10,11. Effective post-disaster response depends on the coordinated actions of the decision role (e.g., government bodies allocating funds), the execution role (e.g., logistics units handling procurement and transportation), and the demand role (the affected population)12. While substantial research has focused on optimizing individual segments of this chain—such as fund appropriation models13, supplier selection and procurement models under contracts4,14, and vehicle routing for distribution15,16—these efforts often operate in silos. There is a pronounced lack of models that explicitly coordinate the financial flow (budget al.location), the physical resource flow (procurement, inventory, and transportation), and the information flow (dynamic demand feedback) among all roles. This disconnection between financially constrained sourcing, physically constrained logistics, and dynamically evolving needs creates bottlenecks and hinders efficient information feedback and resource handover. It ultimately leads to suboptimal overall system performance, including fund misallocation, inventory imbalances, and unmet critical demand17–20. An integrated framework that coordinates these processes is therefore essential to bridge the current coordination gap. Uncertainty is inherent in emergency logistics, manifesting in fluctuating demands, unreliable supply procurement, disrupted transportation networks, and evolving disaster dynamics4,17–19. To address this challenge, a spectrum of optimization methods under uncertainty has been developed and applied. Stochastic Programming (SP) employs probability distributions to model uncertainties, optimizing expected performance. It has been used for integrated pre- and post-disaster management of supplies, coordinating inventory pre-positioning with post-disaster procurement21,22. However, SP’s reliance on accurate distribution data is a key limitation in disasters where historical data is scarce or non-representative. Traditional (Static) Robust Optimization (RO) circumvents the need for distributional knowledge by optimizing against the worst-case realization within a bounded uncertainty set, thus providing guaranteed feasibility. It has been applied to facility location, vehicle routing, and resource allocation under demand uncertainty13,23,24. While ensuring hard safety, this approach is often criticized for excessive conservatism, as the resulting solutions may perform poorly under average or likely scenarios, leading to unnecessarily high costs and inefficient resource utilization—a critical drawback when economic resources are strained24. Compared with SP and RO, Distributionally Robust Optimization (DRO) represents a middle ground, protecting against a family of distributions (an ambiguity set) rather than a single distribution or all point-wise realizations, thereby balancing robustness and conservatism. DRO models have been formulated for disaster relief logistics with option contracts under demand ambiguity14. To improve adaptability to uncertainty, Two-Stage RO (2 S-RO) and Two-Stage DRO (2 S-DRO) explicitly separate decisions into a “here-and-now” first stage (e.g., pre-positioning) and a “wait-and-see” second stage (e.g., post-disaster distribution), allowing recourse actions. They have been widely adopted for emergency facility location, inventory planning, and post-disaster scheduling13,14,17,24. Despite these advancements, two-stage models depict only a single adaptation point and are less effective in capturing the continuous feedback cycle required over multiple periods (e.g., daily adjustments over 72 h), where later decisions benefit from information revealed earlier. This limitation can lead to suboptimal or infeasible schedules across an extended time horizon. The paradigm of Multi-Stage Adaptive Robust Optimization (MSARO), with its ability to model sequential decision-dependent recourse under uncertainty, has been successfully applied and increasingly adopted across various fields facing similar dynamic challenges. In energy management, MSARO frameworks have been developed for the adaptive scheduling of industrial oxygen systems and the coordinated expansion planning of electric vehicle charging networks, balancing operational safety with economic efficiency under demand uncertainty25,26. Within supply chain and logistics management, it has been applied to integrated fleet sizing and vehicle routing problems, providing robust and adaptable policies for closed-loop systems27. In the financial sector, multi-stage robust and distributionally robust models are employed for dynamic portfolio selection to manage risk over time28. Furthermore, MSARO principles are advancing in transportation and unmanned systems, such as in robust trajectory planning for UAVs29. These diverse applications illustrate the methodological advantages of MSARO in delivering solutions that are not only robust to worst-case scenarios but also capable of dynamically adjusting to revealed information, thereby improving average-case performance and system adaptability30,31. In contrast, the adoption of MSARO in post-disaster emergency logistics remains limited, representing a salient research gap. Current optimization models for multi-period resource allocation often lack the explicit feedback loop and adjustable decision rules that are hallmarks of MSARO, which would allow funding, procurement, and transportation plans to be systematically revised based on the partial resolution of uncertainty in prior periods. Building upon the critical analysis above, two intertwined research gaps can be identified. First, there is a lack of a holistic decision-making framework that integrates financial allocation (decision role), physical resource procurement and transportation (execution role), and demand fulfillment (demand role) into a unified and synchronized process. Second, despite its demonstrated effectiveness in managing multi-period sequential decisions under uncertainty in other fields, the methodological power of MSARO remains underexplored within the domain of post-disaster emergency resource allocation. To bridge these gaps, we propose a tripartite “Decision-Execution-Demand” integration model for emergency resource allocation, integrating MSARO and linearization techniques to enhance adaptability, solvability, and applicability. The main contributions include: Integrated multi-role allocation model: We establish a centralized optimization model that incorporates the distinct objectives and constraints of three key roles: the decision-making role (minimizing cost), the execution role (maximizing life-saving utility), and the demand role (minimizing supply-demand gap). This model provides a unified framework for coordinating fund allocation, resource procurement, transportation, and demand fulfillment, rather than relying on independent local optimizations (“Multi-role integrated emergency resource allocation model”).Multi-stage robust emergency resource allocation: To address the dynamic and uncertain nature of disaster response, we cast our integrated model into an MSARO formulation, which employs affine decision rules to dynamically adjust fund allocation, procurement, and transportation quantities across multiple periods in response to the sequential realization of uncertainties in demand, supply, and logistics. Mainstream MSARO methods only focus on dynamic adjustments made when realized uncertain parameters are observed. In contrast, the proposed MSARO model distinguishes between predicted values and actual deviations of uncertain parameters, thereby combining static planning based on pre-disaster predictions with dynamic responses following the realization of post-disaster uncertainties (“Multi-stage robust resource allocation model”).Computational tractability and application potential: For model tractability and compatibility with off-the-shelf solvers, we derive a linear programming (LP) reformulation of the resulting MSARO model using linear duality theory and robust counterpart techniques. Through a case study based on a realistic earthquake scenario, we show that this LP can be solved efficiently for practically relevant problem sizes, indicating its potential as a viable tool for emergency decision support (“Linear reformulation of the robust allocation model” and “Experiment analysis and discussion”). Integrated multi-role allocation model: We establish a centralized optimization model that incorporates the distinct objectives and constraints of three key roles: the decision-making role (minimizing cost), the execution role (maximizing life-saving utility), and the demand role (minimizing supply-demand gap). This model provides a unified framework for coordinating fund allocation, resource procurement, transportation, and demand fulfillment, rather than relying on independent local optimizations (“Multi-role integrated emergency resource allocation model”). Integrated multi-role allocation model: We establish a centralized optimization model that incorporates the distinct objectives and constraints of three key roles: the decision-making role (minimizing cost), the execution role (maximizing life-saving utility), and the demand role (minimizing supply-demand gap). This model provides a unified framework for coordinating fund allocation, resource procurement, transportation, and demand fulfillment, rather than relying on independent local optimizations (“Multi-role integrated emergency resource allocation model”). Multi-stage robust emergency resource allocation: To address the dynamic and uncertain nature of disaster response, we cast our integrated model into an MSARO formulation, which employs affine decision rules to dynamically adjust fund allocation, procurement, and transportation quantities across multiple periods in response to the sequential realization of uncertainties in demand, supply, and logistics. Mainstream MSARO methods only focus on dynamic adjustments made when realized uncertain parameters are observed. In contrast, the proposed MSARO model distinguishes between predicted values and actual deviations of uncertain parameters, thereby combining static planning based on pre-disaster predictions with dynamic responses following the realization of post-disaster uncertainties (“Multi-stage robust resource allocation model”). Multi-stage robust emergency resource allocation: To address the dynamic and uncertain nature of disaster response, we cast our integrated model into an MSARO formulation, which employs affine decision rules to dynamically adjust fund allocation, procurement, and transportation quantities across multiple periods in response to the sequential realization of uncertainties in demand, supply, and logistics. Mainstream MSARO methods only focus on dynamic adjustments made when realized uncertain parameters are observed. In contrast, the proposed MSARO model distinguishes between predicted values and actual deviations of uncertain parameters, thereby combining static planning based on pre-disaster predictions with dynamic responses following the realization of post-disaster uncertainties (“Multi-stage robust resource allocation model”). Computational tractability and application potential: For model tractability and compatibility with off-the-shelf solvers, we derive a linear programming (LP) reformulation of the resulting MSARO model using linear duality theory and robust counterpart techniques. Through a case study based on a realistic earthquake scenario, we show that this LP can be solved efficiently for practically relevant problem sizes, indicating its potential as a viable tool for emergency decision support (“Linear reformulation of the robust allocation model” and “Experiment analysis and discussion”). Computational tractability and application potential: For model tractability and compatibility with off-the-shelf solvers, we derive a linear programming (LP) reformulation of the resulting MSARO model using linear duality theory and robust counterpart techniques. Through a case study based on a realistic earthquake scenario, we show that this LP can be solved efficiently for practically relevant problem sizes, indicating its potential as a viable tool for emergency decision support (“Linear reformulation of the robust allocation model” and “Experiment analysis and discussion”). Problem description Main roles in resource allocation According to the Emergency Response Law of the People’s Republic of China32, the national emergency management system emphasizes unified leadership, comprehensive coordination, hierarchical responsibility, and local management. The allocation of emergency resources represents a critical challenge in the management of hazard-induced disaster emergencies20. Post-disaster emergency resource allocation involves multiple roles working in coordination to facilitate disaster relief efforts33. Government agencies at all levels serve as the primary decision role, providing unified leadership for disaster response. They are responsible for formulating resource allocation plans, disbursing emergency relief funds, and ensuring the initiation of resource deployment. Their primary objective is to meet disaster relief needs while minimizing economic costs. Procurement and supply departments serve as the intermediate execution role, responsible for procuring and transporting emergency resources at designated local transfer stations (TS). Their tasks are to ensure the timely preparation and accurate delivery of resources to the demand role, maximizing life-saving utility. At the disaster areas (DA), survivors at affected sites constitute the demand role. They require prompt access to emergency resources to maximize survival chances and aim to minimize the gap between resource supply and demand. Integrated emergency resource allocation The first 72 h after a disaster are regarded as the “golden window” for emergency rescue. During this period, the decision, execution, and demand roles must collaborate closely to counteract the impacts of hazard-induced disasters, as illustrated in Fig. 1. Prior to or immediately after a hazard-induced disaster, the decision role rapidly assesses the impact and estimates the resource demand of affected populations (such as first-aid kits, instant food, drinking water, etc.). They issue allocation plans and disburse emergency funds to the execution role, dynamically adjusting funding strategies based on real-time feedback regarding resource procurement and transportation. The execution role implements the allocation plan by procuring, temporarily storing, and transporting emergency resources to the demand role, adjusting procurement and transportation strategies in response to actual demands reported by the disaster areas. The demand role—survivors at the disaster areas—cooperate with emergency response teams by reporting their actual demands and waiting orderly for rescue and relief. Fig. 1Schematic diagram of integrated emergency resources allocation after hazard-induced disasters. Schematic diagram of integrated emergency resources allocation after hazard-induced disasters. It should be noted that while real-world post-disaster response involves multiple independent roles (e.g., government departments, logistics companies, affected communities), each with potentially differing information and incentives, the model presented in this paper adopts a centralized optimization perspective. This perspective synthesizes the core objectives and operational constraints of these distinct roles into a coordinated decision-making framework. This approach allows us to focus on deriving a theoretical benchmark for resource flows under uncertainty. Multi-role integrated emergency resource allocation model Immediately before or after a hazard-induced disaster, damage is rapidly assessed and an emergency resource allocation plan is formulated for the critical 72-hour post-disaster period. The planning horizon is discretized into T equal time periods (see Fig. 2). We then develop a deterministic multi-role integrated resource allocation model incorporating both operational constraints and optimization objectives. Fig. 2Multi-period discretization of 72-hour post-disaster period. Multi-period discretization of 72-hour post-disaster period. Constraints Fund flow constraint For each transfer station in every period, allocated funds must cover procurement and transportation costs, including additional expenses caused by transportation losses (e.g., damaged roads). Unused funds may be carried over to later periods, but expenditures in each period must not exceed the cumulative allocated funds. Therefore, fund allocation from the decision role to the execution role in each period is constrained by Eq. (1). 1\documentclass[12pt]{minimal} \usepackage{amsmath} \usepackage{wasysym} \usepackage{amsfonts} \usepackage{amssymb} \usepackage{amsbsy} \usepackage{mathrsfs} \usepackage{upgreek} \setlength{\oddsidemargin}{-69pt} \begin{document}$$\:\sum\:_{\tau\:=1}^{t}({Bud}_{\tau\:,i}-\sum\:_{k=1}^{K}{Pri}_{k}\cdot\:{Pur}_{\tau\:,i,k}-\sum\:_{j=1}^{J}\sum\:_{k=1}^{K}{Cos}_{\tau\:,i,j,k}\cdot\:{Tra}_{\tau\:,i,j,k}\cdot\:(1+Rat))\ge\:0\:\forall\:i=1,\ldots\:,I,\:t=1,\ldots\:,T$$\end{document} Resource flow constraint For each transfer station and resource type, surplus inventory remaining after procurement and transportation can be carried forward to later periods. The cumulative quantity transported cannot exceed available inventory. Procurement and transportation decisions therefore satisfy Eq. (2). 2\documentclass[12pt]{minimal} \usepackage{amsmath} \usepackage{wasysym} \usepackage{amsfonts} \usepackage{amssymb} \usepackage{amsbsy} \usepackage{mathrsfs} \usepackage{upgreek} \setlength{\oddsidemargin}{-69pt} \begin{document}$$\:\sum\:_{\tau\:=1}^{t}({Pur}_{\tau\:,i,k}-\sum\:_{j=1}^{J}{Tra}_{\tau\:,i,j,k})\ge\:0\:\forall\:k=1,\ldots\:,K,i=1,\ldots\:,I,\:t=1,\ldots\:,T$$\end{document} Supply-demand gap For each disaster area and resource type, the cumulative supply-demand gap Rest,j,k between execution and demand roles over time is constrained by Eq. (3). Here, Rest,j,k is only a symbolic representation and not a decision variable. 3\documentclass[12pt]{minimal} \usepackage{amsmath} \usepackage{wasysym} \usepackage{amsfonts} \usepackage{amssymb} \usepackage{amsbsy} \usepackage{mathrsfs} \usepackage{upgreek} \setlength{\oddsidemargin}{-69pt} \begin{document}$$\:{Res}_{t,j,k}=\sum\:_{\tau\:=1}^{t}(\sum\:_{i=1}^{I}{Tra}_{\tau\:,i,j,k}-{Dem}_{\tau\:,j,k})\:\forall\:k=1,\ldots\:,K,j=1,\ldots\:,J,\:t=1,\ldots\:,T$$\end{document} Non-negativity All decision variables related to funding, procurement, and transportation must be non-negative, as expressed in Eq. (4). 4\documentclass[12pt]{minimal} \usepackage{amsmath} \usepackage{wasysym} \usepackage{amsfonts} \usepackage{amssymb} \usepackage{amsbsy} \usepackage{mathrsfs} \usepackage{upgreek} \setlength{\oddsidemargin}{-69pt} \begin{document}$$\:{Bud}_{t,i}\ge\:0,{Pur}_{t,i,k}\ge\:0,{Tra}_{t,i,j,k}\ge\:0\:\forall\:k=1,\ldots\:,K,j=1,\ldots\:,J,i=1,\ldots\:,I,\:t=1,\ldots\:,T$$\end{document} Optimization objective Emergency resource allocation involves objectives corresponding to different roles. Economic cost minimization (decision role) The total allocated budget across all periods and transfer stations is minimized, as given in Eq. (5). 5\documentclass[12pt]{minimal} \usepackage{amsmath} \usepackage{wasysym} \usepackage{amsfonts} \usepackage{amssymb} \usepackage{amsbsy} \usepackage{mathrsfs} \usepackage{upgreek} \setlength{\oddsidemargin}{-69pt} \begin{document}$$\:\mathrm{min}{\:Z}_{1}=\sum\:_{t=1}^{T}\sum\:_{i=1}^{I}{Bud}_{t,i}$$\end{document} Life-saving utility maximization (execution role) Total life-saving utility generated by resource deliveries across periods, transfer stations, and disaster areas is maximized, as defined in Eq. (6). 6\documentclass[12pt]{minimal} \usepackage{amsmath} \usepackage{wasysym} \usepackage{amsfonts} \usepackage{amssymb} \usepackage{amsbsy} \usepackage{mathrsfs} \usepackage{upgreek} \setlength{\oddsidemargin}{-69pt} \begin{document}$$\:\mathrm{max}\:{Z}_{2}=\sum\:_{t=1}^{T}\sum\:_{i=1}^{I}\sum\:_{j=1}^{J}\sum\:_{k=1}^{K}{Uti}_{t,j}\cdot\:{Tra}_{t,i,j,k}$$\end{document} Supply-demand gap minimization (demand role) To reduce shortages and avoid resource waste, the total absolute supply-demand gap across all resources, areas, and periods is minimized, as defined in Eq. (7). Because the absolute value introduces nonlinearity, an auxiliary variable \documentclass[12pt]{minimal} \usepackage{amsmath} \usepackage{wasysym} \usepackage{amsfonts} \usepackage{amssymb} \usepackage{amsbsy} \usepackage{mathrsfs} \usepackage{upgreek} \setlength{\oddsidemargin}{-69pt} \begin{document}$$\:{a}_{t,j,k}$$\end{document} is introduced to linearize the formulation, resulting in Eqs. (8) and (9). 7\documentclass[12pt]{minimal} \usepackage{amsmath} \usepackage{wasysym} \usepackage{amsfonts} \usepackage{amssymb} \usepackage{amsbsy} \usepackage{mathrsfs} \usepackage{upgreek} \setlength{\oddsidemargin}{-69pt} \begin{document}$$\:\mathrm{min}{\:Z}_{3}=\sum\:_{t=1}^{T}\sum\:_{j=1}^{J}\sum\:_{k=1}^{K}{|Res}_{t,j,k}|$$\end{document} 8\documentclass[12pt]{minimal} \usepackage{amsmath} \usepackage{wasysym} \usepackage{amsfonts} \usepackage{amssymb} \usepackage{amsbsy} \usepackage{mathrsfs} \usepackage{upgreek} \setlength{\oddsidemargin}{-69pt} \begin{document}$$\:\mathrm{min}{\:Z}_{3}=\sum\:_{t=1}^{T}\sum\:_{j=1}^{J}\sum\:_{k=1}^{K}{a}_{t,j,k}$$\end{document} 9\documentclass[12pt]{minimal} \usepackage{amsmath} \usepackage{wasysym} \usepackage{amsfonts} \usepackage{amssymb} \usepackage{amsbsy} \usepackage{mathrsfs} \usepackage{upgreek} \setlength{\oddsidemargin}{-69pt} \begin{document}$$\:{a}_{t,j,k}\ge\:{Res}_{t,j,k};{a}_{t,j,k}\ge\:-{Res}_{t,j,k}\:\forall\:k=1,\ldots\:,K,j=1,\ldots\:,J,\:t=1,\ldots\:,T$$\end{document} Comprehensive objective minimization A weighted sum of the above objectives is minimized to achieve balanced performance across roles, as shown in Eq. (10). The weighted-sum method provides a transparent way to combine multiple objectives, where weights reflect decision priorities rather than model parameters. Different weight selections correspond to different trade-offs between economic cost, life-saving utility, and supply-demand balance. The proposed model does not prescribe a single “optimal” weighting scheme; instead, it provides a decision-support tool to evaluate the consequences of alternative priority settings. In practice, weights should be determined through policy or stakeholder considerations. In this study, weights are set as \documentclass[12pt]{minimal} \usepackage{amsmath} \usepackage{wasysym} \usepackage{amsfonts} \usepackage{amssymb} \usepackage{amsbsy} \usepackage{mathrsfs} \usepackage{upgreek} \setlength{\oddsidemargin}{-69pt} \begin{document}$$\:{\omega\:}_{1}$$\end{document} = 0.0000001, \documentclass[12pt]{minimal} \usepackage{amsmath} \usepackage{wasysym} \usepackage{amsfonts} \usepackage{amssymb} \usepackage{amsbsy} \usepackage{mathrsfs} \usepackage{upgreek} \setlength{\oddsidemargin}{-69pt} \begin{document}$$\:\:{\omega\:}_{2}$$\end{document} = 0.00001, and \documentclass[12pt]{minimal} \usepackage{amsmath} \usepackage{wasysym} \usepackage{amsfonts} \usepackage{amssymb} \usepackage{amsbsy} \usepackage{mathrsfs} \usepackage{upgreek} \setlength{\oddsidemargin}{-69pt} \begin{document}$$\:{\omega\:}_{3}$$\end{document} = 1, prioritizing supply-demand balance while allowing life-saving utility and economic cost to act as secondary factors. 10\documentclass[12pt]{minimal} \usepackage{amsmath} \usepackage{wasysym} \usepackage{amsfonts} \usepackage{amssymb} \usepackage{amsbsy} \usepackage{mathrsfs} \usepackage{upgreek} \setlength{\oddsidemargin}{-69pt} \begin{document}$$\:\mathrm{min}\:f={\omega\:}_{1}\cdot\:{Z}_{1}-{\omega\:}_{2}\cdot\:{Z}_{2}+{\omega\:}_{3}\cdot\:{Z}_{3}$$\end{document} Multi-stage robust resource allocation model The deterministic allocation model described above relies on predefined parameters. However, these parameters are difficult to estimate accurately before or during a disaster, which makes the model’s solution vulnerable to uncertainty and potentially infeasible. To adapt to uncertainty, we design a multi-stage robust allocation strategy for emergency resources, integrating pre-disaster long-term planning with post-disaster periodic feedback adjustments, as illustrated in Fig. 3. During the pre-disaster or initial phase, based on parameter estimates from the execution role, the decision role formulates a resource allocation plan for the post-disaster “golden 72-hour window”. During the post-disaster adjustment phase, as each period elapses, demand uncertainties and disaster dynamics are progressively revealed. At the start of each period, the decision and execution roles adjust the current allocation plan using feedback-based decision rules to address newly revealed uncertainties. This framework is formalized using the multi-stage adaptive robust optimization (MSARO) method, involving periodic uncertainty sets and recourse decision rules. Fig. 3General idea of multi-stage robust resource allocation. General idea of multi-stage robust resource allocation. Multi-period uncertainty sets To reflect uncertainties in disaster response, we consider three sources affecting resource allocation: uncertain resource demand, procurement, and transportation. Actual demands at disaster areas are initially uncertain due to limited post-disaster information. Procurement plans may not be fully executed because of social constraints or disaster impacts at transfer stations. In addition, damaged infrastructure may disrupt planned deliveries from transfer stations to disaster areas. We define the following uncertain deviation parameters: \documentclass[12pt]{minimal} \usepackage{amsmath} \usepackage{wasysym} \usepackage{amsfonts} \usepackage{amssymb} \usepackage{amsbsy} \usepackage{mathrsfs} \usepackage{upgreek} \setlength{\oddsidemargin}{-69pt} \begin{document}$$\:{\xi\:}_{t,j,k}^{\mathrm{D}\mathrm{e}\mathrm{m}}$$\end{document} denotes the demand deviation for resource k at disaster area j in period t; \documentclass[12pt]{minimal} \usepackage{amsmath} \usepackage{wasysym} \usepackage{amsfonts} \usepackage{amssymb} \usepackage{amsbsy} \usepackage{mathrsfs} \usepackage{upgreek} \setlength{\oddsidemargin}{-69pt} \begin{document}$$\:{\xi\:}_{t,i,k}^{\mathrm{P}\mathrm{u}\mathrm{r}}$$\end{document} denotes the procurement deviation for resource k at transfer station i in period t; and \documentclass[12pt]{minimal} \usepackage{amsmath} \usepackage{wasysym} \usepackage{amsfonts} \usepackage{amssymb} \usepackage{amsbsy} \usepackage{mathrsfs} \usepackage{upgreek} \setlength{\oddsidemargin}{-69pt} \begin{document}$$\:{\xi\:}_{t,i,j,k}^{\mathrm{T}\mathrm{r}\mathrm{a}}$$\end{document} denotes the transportation deviation of resource k from transfer station i to disaster area j in period t. For compact notation, all deviations in period t are expressed in vector form as \documentclass[12pt]{minimal} \usepackage{amsmath} \usepackage{wasysym} \usepackage{amsfonts} \usepackage{amssymb} \usepackage{amsbsy} \usepackage{mathrsfs} \usepackage{upgreek} \setlength{\oddsidemargin}{-69pt} \begin{document}$$\:{\boldsymbol{\xi\:}}_{t}^{\mathrm{D}\mathrm{e}\mathrm{m}}=({\xi\:}_{t,\mathrm{1,1}}^{\mathrm{D}\mathrm{e}\mathrm{m}},\:\ldots\:,{\xi\:}_{t,J,K}^{\mathrm{D}\mathrm{e}\mathrm{m}})\in\:{\mathbb{R}}^{J\cdot\:K}$$\end{document}, \documentclass[12pt]{minimal} \usepackage{amsmath} \usepackage{wasysym} \usepackage{amsfonts} \usepackage{amssymb} \usepackage{amsbsy} \usepackage{mathrsfs} \usepackage{upgreek} \setlength{\oddsidemargin}{-69pt} \begin{document}$$\:{\boldsymbol{\xi\:}}_{t}^{\mathrm{P}\mathrm{u}\mathrm{r}}=({\xi\:}_{t,\mathrm{1,1}}^{\mathrm{P}\mathrm{u}\mathrm{r}},\:\ldots\:,{\xi\:}_{t,I,K}^{\mathrm{P}\mathrm{u}\mathrm{r}})\in\:{\mathbb{R}}^{I\cdot\:K}$$\end{document}, and \documentclass[12pt]{minimal} \usepackage{amsmath} \usepackage{wasysym} \usepackage{amsfonts} \usepackage{amssymb} \usepackage{amsbsy} \usepackage{mathrsfs} \usepackage{upgreek} \setlength{\oddsidemargin}{-69pt} \begin{document}$$\:{\boldsymbol{\xi\:}}_{t}^{\mathrm{T}\mathrm{r}\mathrm{a}}=({\xi\:}_{t,\mathrm{1,1},1}^{\mathrm{T}\mathrm{r}\mathrm{a}},\:\ldots\:,{\xi\:}_{t,I,J,K}^{\mathrm{T}\mathrm{r}\mathrm{a}})\in\:{\mathbb{R}}^{I\cdot\:J\cdot\:K}$$\end{document}. The actual values of resource demand, procurement, and transportation are given in Eqs. (11)~(13), where, \documentclass[12pt]{minimal} \usepackage{amsmath} \usepackage{wasysym} \usepackage{amsfonts} \usepackage{amssymb} \usepackage{amsbsy} \usepackage{mathrsfs} \usepackage{upgreek} \setlength{\oddsidemargin}{-69pt} \begin{document}$$\:{\widehat{Dem}}_{t,j,k}$$\end{document}, \documentclass[12pt]{minimal} \usepackage{amsmath} \usepackage{wasysym} \usepackage{amsfonts} \usepackage{amssymb} \usepackage{amsbsy} \usepackage{mathrsfs} \usepackage{upgreek} \setlength{\oddsidemargin}{-69pt} \begin{document}$$\:{\widehat{Tra}}_{t,i,j,k}$$\end{document} and \documentclass[12pt]{minimal} \usepackage{amsmath} \usepackage{wasysym} \usepackage{amsfonts} \usepackage{amssymb} \usepackage{amsbsy} \usepackage{mathrsfs} \usepackage{upgreek} \setlength{\oddsidemargin}{-69pt} \begin{document}$$\:{\widehat{Pur}}_{t,i,k}$$\end{document} denote the corresponding estimated values.11\documentclass[12pt]{minimal} \usepackage{amsmath} \usepackage{wasysym} \usepackage{amsfonts} \usepackage{amssymb} \usepackage{amsbsy} \usepackage{mathrsfs} \usepackage{upgreek} \setlength{\oddsidemargin}{-69pt} \begin{document}$$\:{Dem}_{t,j,k}={\widehat{Dem}}_{t,j,k}+{\xi\:}_{t,j,k}^{\mathrm{D}\mathrm{e}\mathrm{m}}\:\:\forall\:k=1,\ldots\:,K,j=1,\ldots\:,J,\:t=1,\ldots\:,T$$\end{document} 12\documentclass[12pt]{minimal} \usepackage{amsmath} \usepackage{wasysym} \usepackage{amsfonts} \usepackage{amssymb} \usepackage{amsbsy} \usepackage{mathrsfs} \usepackage{upgreek} \setlength{\oddsidemargin}{-69pt} \begin{document}$$\:{Tra}_{t,i,j,k}={\widehat{Tra}}_{t,i,j,k}+{\xi\:}_{t,i,j,k}^{\mathrm{T}\mathrm{r}\mathrm{a}}\:\:\forall\:k=1,\ldots\:,K,i=1,\ldots\:,I,j=1,\ldots\:,J,\:t=1,\ldots\:,T$$\end{document} 13\documentclass[12pt]{minimal} \usepackage{amsmath} \usepackage{wasysym} \usepackage{amsfonts} \usepackage{amssymb} \usepackage{amsbsy} \usepackage{mathrsfs} \usepackage{upgreek} \setlength{\oddsidemargin}{-69pt} \begin{document}$$\:{Pur}_{t,i,k}={\widehat{Pur}}_{t,i,k}+{\xi\:}_{t,i,k}^{\mathrm{P}\mathrm{u}\mathrm{r}}\:\:\forall\:k=1,\ldots\:,K,i=1,\ldots\:,I,\:t=1,\ldots\:,T$$\end{document} Periodic uncertainty sets are required in MSARO to characterize uncertain parameters in each period and support feedback-based decision adjustments. This study adopts box uncertainty sets23 and assumes that uncertainty sets are independent across periods. The bounded ranges of the uncertain parameters are defined by Eqs. (14)–(16).14\documentclass[12pt]{minimal} \usepackage{amsmath} \usepackage{wasysym} \usepackage{amsfonts} \usepackage{amssymb} \usepackage{amsbsy} \usepackage{mathrsfs} \usepackage{upgreek} \setlength{\oddsidemargin}{-69pt} \begin{document}$$\:{{\Psi\:}}_{t,\:i,j,k}^{\mathrm{L},\mathrm{T}\mathrm{r}\mathrm{a}}\le\:{\xi\:}_{t,i,j,k}^{\mathrm{T}\mathrm{r}\mathrm{a}}\le\:{{\Psi\:}}_{t,\:i,j,k}^{\mathrm{U},\mathrm{T}\mathrm{r}\mathrm{a}}\:\forall\:k=1,\ldots\:,K,i=1,\ldots\:,I,j=1,\ldots\:,J,t=1,\ldots\:,T$$\end{document} 15\documentclass[12pt]{minimal} \usepackage{amsmath} \usepackage{wasysym} \usepackage{amsfonts} \usepackage{amssymb} \usepackage{amsbsy} \usepackage{mathrsfs} \usepackage{upgreek} \setlength{\oddsidemargin}{-69pt} \begin{document}$$\:{{\Psi\:}}_{t,j,k}^{\mathrm{L},\mathrm{D}\mathrm{e}\mathrm{m}}\le\:{\xi\:}_{t,j,k}^{\mathrm{D}\mathrm{e}\mathrm{m}}\le\:{{\Psi\:}}_{t,j,k}^{\mathrm{U},\mathrm{D}\mathrm{e}\mathrm{m}}\:\forall\:k=1,\ldots\:,K,j=1,\ldots\:,J,t=1,\ldots\:,T$$\end{document} 16\documentclass[12pt]{minimal} \usepackage{amsmath} \usepackage{wasysym} \usepackage{amsfonts} \usepackage{amssymb} \usepackage{amsbsy} \usepackage{mathrsfs} \usepackage{upgreek} \setlength{\oddsidemargin}{-69pt} \begin{document}$$\:{{{\Psi\:}}_{t,\:i,k}^{\mathrm{L},\mathrm{P}\mathrm{u}\mathrm{r}}\le\:\xi\:}_{t,i,k}^{\mathrm{P}\mathrm{u}\mathrm{r}}\le\:{{\Psi\:}}_{t,\:i,k}^{\mathrm{U},\mathrm{P}\mathrm{u}\mathrm{r}}\:\forall\:k=1,\ldots\:,K,i=1,\ldots\:,I,t=1,\ldots\:,T$$\end{document} For compact representation, the periodic uncertainty set is written in vector–matrix form as Eq. (17), where \documentclass[12pt]{minimal} \usepackage{amsmath} \usepackage{wasysym} \usepackage{amsfonts} \usepackage{amssymb} \usepackage{amsbsy} \usepackage{mathrsfs} \usepackage{upgreek} \setlength{\oddsidemargin}{-69pt} \begin{document}$$\:{\boldsymbol{\xi\:}}_{t}$$\end{document} denotes the uncertainty vector in period t, \documentclass[12pt]{minimal} \usepackage{amsmath} \usepackage{wasysym} \usepackage{amsfonts} \usepackage{amssymb} \usepackage{amsbsy} \usepackage{mathrsfs} \usepackage{upgreek} \setlength{\oddsidemargin}{-69pt} \begin{document}$$\:{\boldsymbol{H}}_{t}$$\end{document} is the coefficient matrix in constraints (Eqs. (11)–(16)), \documentclass[12pt]{minimal} \usepackage{amsmath} \usepackage{wasysym} \usepackage{amsfonts} \usepackage{amssymb} \usepackage{amsbsy} \usepackage{mathrsfs} \usepackage{upgreek} \setlength{\oddsidemargin}{-69pt} \begin{document}$$\:{\boldsymbol{z}}_{t}$$\end{document} is the constant vector, \documentclass[12pt]{minimal} \usepackage{amsmath} \usepackage{wasysym} \usepackage{amsfonts} \usepackage{amssymb} \usepackage{amsbsy} \usepackage{mathrsfs} \usepackage{upgreek} \setlength{\oddsidemargin}{-69pt} \begin{document}$$\:N=J\cdot\:K+I\cdot\:K+I\cdot\:J\cdot\:K$$\end{document} is the number of uncertain parameters per period, and \documentclass[12pt]{minimal} \usepackage{amsmath} \usepackage{wasysym} \usepackage{amsfonts} \usepackage{amssymb} \usepackage{amsbsy} \usepackage{mathrsfs} \usepackage{upgreek} \setlength{\oddsidemargin}{-69pt} \begin{document}$$\:V$$\end{document} is the number of uncertainty-set constraints per period.17\documentclass[12pt]{minimal} \usepackage{amsmath} \usepackage{wasysym} \usepackage{amsfonts} \usepackage{amssymb} \usepackage{amsbsy} \usepackage{mathrsfs} \usepackage{upgreek} \setlength{\oddsidemargin}{-69pt} \begin{document}$$\:{{\Xi\:}}_{t}=\left\{{\boldsymbol{\xi\:}}_{t}=\left({\boldsymbol{\xi\:}}_{t}^{\mathrm{P}\mathrm{u}\mathrm{r}},{\boldsymbol{\xi\:}}_{t}^{\mathrm{D}\mathrm{e}\mathrm{m}},{\boldsymbol{\xi\:}}_{t}^{\mathrm{T}\mathrm{r}\mathrm{a}}\right)\in\:{\mathbb{R}}^{N}:{\boldsymbol{H}}_{t}{\boldsymbol{\xi\:}}_{t}\le\:{\boldsymbol{z}}_{t},{\boldsymbol{H}}_{t}\in\:{\mathbb{R}}^{V\times\:N},{\boldsymbol{z}}_{t}\in\:{\mathbb{R}}^{V}\right\}\:\forall\:t=1,\ldots\:,T$$\end{document} Adaptive feedback adjustment rules After period t-1 is completed, the execution role reports the realized uncertainties in resource demand, procurement, and transportation. The decision role then adjusts the allocation plan for period t based on historical uncertainty realizations, allowing the model to adapt to outcomes observed in prior periods. As illustrated in Fig. 4, recourse decision variables in each period t depend on uncertainties revealed during the preceding Δt periods, denoted by \documentclass[12pt]{minimal} \usepackage{amsmath} \usepackage{wasysym} \usepackage{amsfonts} \usepackage{amssymb} \usepackage{amsbsy} \usepackage{mathrsfs} \usepackage{upgreek} \setlength{\oddsidemargin}{-69pt} \begin{document}$$\:{\boldsymbol{\xi\:}}_{\left[t-{\Delta\:}t,t-1\right]}$$\end{document}:=(\documentclass[12pt]{minimal} \usepackage{amsmath} \usepackage{wasysym} \usepackage{amsfonts} \usepackage{amssymb} \usepackage{amsbsy} \usepackage{mathrsfs} \usepackage{upgreek} \setlength{\oddsidemargin}{-69pt} \begin{document}$$\:{\boldsymbol{\xi\:}}_{t-{\Delta\:}t}$$\end{document},…,\documentclass[12pt]{minimal} \usepackage{amsmath} \usepackage{wasysym} \usepackage{amsfonts} \usepackage{amssymb} \usepackage{amsbsy} \usepackage{mathrsfs} \usepackage{upgreek} \setlength{\oddsidemargin}{-69pt} \begin{document}$$\:{\boldsymbol{\xi\:}}_{t-1}$$\end{document}). Accordingly, decision variables for transportation, procurement, and fund allocation in period t are expressed as vector functions \documentclass[12pt]{minimal} \usepackage{amsmath} \usepackage{wasysym} \usepackage{amsfonts} \usepackage{amssymb} \usepackage{amsbsy} \usepackage{mathrsfs} \usepackage{upgreek} \setlength{\oddsidemargin}{-69pt} \begin{document}$$\:{\boldsymbol{y}}_{t}^{\mathrm{T}\mathrm{r}\mathrm{a}}\left({\boldsymbol{\xi\:}}_{\left[t-{\Delta\:}t,t-1\right]}\right)=({Tra}_{t,\mathrm{1,1},1},\ldots\:,{Tra}_{t,I,J,K})\in\:{\mathbb{R}}^{I\cdot\:J\cdot\:K}$$\end{document}, \documentclass[12pt]{minimal} \usepackage{amsmath} \usepackage{wasysym} \usepackage{amsfonts} \usepackage{amssymb} \usepackage{amsbsy} \usepackage{mathrsfs} \usepackage{upgreek} \setlength{\oddsidemargin}{-69pt} \begin{document}$$\:{\boldsymbol{y}}_{t}^{\mathrm{P}\mathrm{u}\mathrm{r}}\left({\boldsymbol{\xi\:}}_{\left[t-{\Delta\:}t,t-1\right]}\right)=({Pur}_{t,\mathrm{1,1}},\ldots\:,{Pur}_{t,I,K})\in\:{\mathbb{R}}^{I\cdot\:K}$$\end{document}, and \documentclass[12pt]{minimal} \usepackage{amsmath} \usepackage{wasysym} \usepackage{amsfonts} \usepackage{amssymb} \usepackage{amsbsy} \usepackage{mathrsfs} \usepackage{upgreek} \setlength{\oddsidemargin}{-69pt} \begin{document}$$\:{\boldsymbol{y}}_{t}^{\mathrm{B}\mathrm{u}\mathrm{d}}\left({\boldsymbol{\xi\:}}_{\left[t-{\Delta\:}t,t-1\right]}\right)=({Bud}_{t,1},\ldots\:,{Bud}_{t,I})\in\:{\mathbb{R}}^{I}$$\end{document}. This study adopts affine decision rules34 to describe how recourse decisions depend on historical uncertainty realizations, as shown in Eqs. (18)–(20). Here, vectors \documentclass[12pt]{minimal} \usepackage{amsmath} \usepackage{wasysym} \usepackage{amsfonts} \usepackage{amssymb} \usepackage{amsbsy} \usepackage{mathrsfs} \usepackage{upgreek} \setlength{\oddsidemargin}{-69pt} \begin{document}$$\:{y}_{t}^{\mathrm{T}\mathrm{r}\mathrm{a},0}$$\end{document}, \documentclass[12pt]{minimal} \usepackage{amsmath} \usepackage{wasysym} \usepackage{amsfonts} \usepackage{amssymb} \usepackage{amsbsy} \usepackage{mathrsfs} \usepackage{upgreek} \setlength{\oddsidemargin}{-69pt} \begin{document}$$\:{y}_{t}^{\mathrm{P}\mathrm{u}\mathrm{r},0}$$\end{document}, and \documentclass[12pt]{minimal} \usepackage{amsmath} \usepackage{wasysym} \usepackage{amsfonts} \usepackage{amssymb} \usepackage{amsbsy} \usepackage{mathrsfs} \usepackage{upgreek} \setlength{\oddsidemargin}{-69pt} \begin{document}$$\:{y}_{t}^{\mathrm{B}\mathrm{u}\mathrm{d},0}$$\end{document} denote intercept terms representing the initial pre-disaster allocation plan, while matrices \documentclass[12pt]{minimal} \usepackage{amsmath} \usepackage{wasysym} \usepackage{amsfonts} \usepackage{amssymb} \usepackage{amsbsy} \usepackage{mathrsfs} \usepackage{upgreek} \setlength{\oddsidemargin}{-69pt} \begin{document}$$\:{Y}_{t}^{\mathrm{T}\mathrm{r}\mathrm{a},1}$$\end{document}, \documentclass[12pt]{minimal} \usepackage{amsmath} \usepackage{wasysym} \usepackage{amsfonts} \usepackage{amssymb} \usepackage{amsbsy} \usepackage{mathrsfs} \usepackage{upgreek} \setlength{\oddsidemargin}{-69pt} \begin{document}$$\:{Y}_{t}^{\mathrm{P}\mathrm{u}\mathrm{r},1}$$\end{document}, and \documentclass[12pt]{minimal} \usepackage{amsmath} \usepackage{wasysym} \usepackage{amsfonts} \usepackage{amssymb} \usepackage{amsbsy} \usepackage{mathrsfs} \usepackage{upgreek} \setlength{\oddsidemargin}{-69pt} \begin{document}$$\:{Y}_{t}^{\mathrm{P}\mathrm{u}\mathrm{r},1}$$\end{document} denote slope coefficients measuring adjustment responses to disclosed uncertainties. Affine decision rules provide a linear approximation of the dependence between decision variables and realized uncertainties. This approach is widely adopted in MSARO to maintain tractability while preserving key adaptive features. Although nonlinear decision rules may offer greater flexibility, they typically result in intractable formulations. Affine decision rules therefore represent a practical compromise between solution quality and computational feasibility34. Fig. 4Recourse decision diagram. 18\documentclass[12pt]{minimal} \usepackage{amsmath} \usepackage{wasysym} \usepackage{amsfonts} \usepackage{amssymb} \usepackage{amsbsy} \usepackage{mathrsfs} \usepackage{upgreek} \setlength{\oddsidemargin}{-69pt} \begin{document}$$\:{\boldsymbol{y}}_{t}^{\mathrm{T}\mathrm{r}\mathrm{a}}\left({\boldsymbol{\xi\:}}_{\left[t-{\Delta\:}t,t-1\right]}\right)={\boldsymbol{y}}_{t}^{\mathrm{T}\mathrm{r}\mathrm{a},0}+{\boldsymbol{Y}}_{t}^{\mathrm{T}\mathrm{r}\mathrm{a},1}{\boldsymbol{\xi\:}}_{\left[t-{\Delta\:}t,t-1\right]}\:\forall\:t=2,\ldots\:,T$$\end{document} Recourse decision diagram. 19\documentclass[12pt]{minimal} \usepackage{amsmath} \usepackage{wasysym} \usepackage{amsfonts} \usepackage{amssymb} \usepackage{amsbsy} \usepackage{mathrsfs} \usepackage{upgreek} \setlength{\oddsidemargin}{-69pt} \begin{document}$$\:{\boldsymbol{y}}_{t}^{\mathrm{P}\mathrm{u}\mathrm{r}}\left({\boldsymbol{\xi\:}}_{\left[t-{\Delta\:}t,t-1\right]}\right)={\boldsymbol{y}}_{t}^{\mathrm{P}\mathrm{u}\mathrm{r},0}+{\boldsymbol{Y}}_{t}^{\mathrm{P}\mathrm{u}\mathrm{r},1}{\boldsymbol{\xi\:}}_{\left[t-{\Delta\:}t,t-1\right]}\:\forall\:t=2,\ldots\:,T$$\end{document} 20\documentclass[12pt]{minimal} \usepackage{amsmath} \usepackage{wasysym} \usepackage{amsfonts} \usepackage{amssymb} \usepackage{amsbsy} \usepackage{mathrsfs} \usepackage{upgreek} \setlength{\oddsidemargin}{-69pt} \begin{document}$$\:{\boldsymbol{y}}_{t}^{\mathrm{B}\mathrm{u}\mathrm{d}}\left({\boldsymbol{\xi\:}}_{\left[t-{\Delta\:}t,t-1\right]}\right)={\boldsymbol{y}}_{t}^{\mathrm{B}\mathrm{u}\mathrm{d},0}+{\boldsymbol{Y}}_{t}^{\mathrm{B}\mathrm{u}\mathrm{d},1}{\boldsymbol{\xi\:}}_{\left[t-{\Delta\:}t,t-1\right]}\:\forall\:t=2,\ldots\:,T$$\end{document} For conciseness, all recourse decisions in period t are combined into a single vector \documentclass[12pt]{minimal} \usepackage{amsmath} \usepackage{wasysym} \usepackage{amsfonts} \usepackage{amssymb} \usepackage{amsbsy} \usepackage{mathrsfs} \usepackage{upgreek} \setlength{\oddsidemargin}{-69pt} \begin{document}$$\:{\boldsymbol{y}}_{t}\left({\boldsymbol{\xi\:}}_{\left[t-{\Delta\:}t,t-1\right]}\right)\in\:{\mathbb{R}}^{P\times\:1}$$\end{document}, as shown in Eq. (21). The affine decision rule is then reformulated in Eqs. (22) and (23), where variables \documentclass[12pt]{minimal} \usepackage{amsmath} \usepackage{wasysym} \usepackage{amsfonts} \usepackage{amssymb} \usepackage{amsbsy} \usepackage{mathrsfs} \usepackage{upgreek} \setlength{\oddsidemargin}{-69pt} \begin{document}$$\:{\boldsymbol{y}}_{t}^{0}\in\:{\mathbb{R}}^{P\times\:1}$$\end{document} and \documentclass[12pt]{minimal} \usepackage{amsmath} \usepackage{wasysym} \usepackage{amsfonts} \usepackage{amssymb} \usepackage{amsbsy} \usepackage{mathrsfs} \usepackage{upgreek} \setlength{\oddsidemargin}{-69pt} \begin{document}$$\:{\boldsymbol{Y}}_{t}^{1}\in\:{\mathbb{R}}^{P\times\:(N\cdot\:{\Delta\:}t)}$$\end{document} denote compact intercept terms and slope coefficients. The term \documentclass[12pt]{minimal} \usepackage{amsmath} \usepackage{wasysym} \usepackage{amsfonts} \usepackage{amssymb} \usepackage{amsbsy} \usepackage{mathrsfs} \usepackage{upgreek} \setlength{\oddsidemargin}{-69pt} \begin{document}$$\:{\boldsymbol{y}}_{t,\tau\:}^{1}\in\:{\mathbb{R}}^{P\times\:N}$$\end{document} denotes the submatrix of \documentclass[12pt]{minimal} \usepackage{amsmath} \usepackage{wasysym} \usepackage{amsfonts} \usepackage{amssymb} \usepackage{amsbsy} \usepackage{mathrsfs} \usepackage{upgreek} \setlength{\oddsidemargin}{-69pt} \begin{document}$$\:{\boldsymbol{Y}}_{t}^{1}$$\end{document} associated with \documentclass[12pt]{minimal} \usepackage{amsmath} \usepackage{wasysym} \usepackage{amsfonts} \usepackage{amssymb} \usepackage{amsbsy} \usepackage{mathrsfs} \usepackage{upgreek} \setlength{\oddsidemargin}{-69pt} \begin{document}$$\:{\boldsymbol{\xi\:}}_{\tau\:}$$\end{document}, representing adjustment coefficients of recourse decisions in period t with respect to uncertainties in historical period \documentclass[12pt]{minimal} \usepackage{amsmath} \usepackage{wasysym} \usepackage{amsfonts} \usepackage{amssymb} \usepackage{amsbsy} \usepackage{mathrsfs} \usepackage{upgreek} \setlength{\oddsidemargin}{-69pt} \begin{document}$$\:\tau\:$$\end{document} (\documentclass[12pt]{minimal} \usepackage{amsmath} \usepackage{wasysym} \usepackage{amsfonts} \usepackage{amssymb} \usepackage{amsbsy} \usepackage{mathrsfs} \usepackage{upgreek} \setlength{\oddsidemargin}{-69pt} \begin{document}$$\:\mathrm{m}\mathrm{a}\mathrm{x}(1,t-{\Delta\:}t)\le\:\tau\:\le\:t-1$$\end{document}). The symbol \documentclass[12pt]{minimal} \usepackage{amsmath} \usepackage{wasysym} \usepackage{amsfonts} \usepackage{amssymb} \usepackage{amsbsy} \usepackage{mathrsfs} \usepackage{upgreek} \setlength{\oddsidemargin}{-69pt} \begin{document}$$\:P=I\cdot\:J\cdot\:K+I\cdot\:K+I$$\end{document}. denotes the total number of decision variables in period t.21\documentclass[12pt]{minimal} \usepackage{amsmath} \usepackage{wasysym} \usepackage{amsfonts} \usepackage{amssymb} \usepackage{amsbsy} \usepackage{mathrsfs} \usepackage{upgreek} \setlength{\oddsidemargin}{-69pt} \begin{document}$$\:{\boldsymbol{y}}_{t}\left({\boldsymbol{\xi\:}}_{\left[t-{\Delta\:}t,t-1\right]}\right)=\left({\boldsymbol{y}}_{t}^{\mathrm{T}\mathrm{r}\mathrm{a}}\left({\boldsymbol{\xi\:}}_{\left[t-{\Delta\:}t,t-1\right]}\right),{\boldsymbol{y}}_{t}^{\mathrm{P}\mathrm{u}\mathrm{r}}\left({\boldsymbol{\xi\:}}_{\left[t-{\Delta\:}t,t-1\right]}\right),{\boldsymbol{y}}_{t}^{\mathrm{B}\mathrm{u}\mathrm{d}}\left({\boldsymbol{\xi\:}}_{\left[t-{\Delta\:}t,t-1\right]}\right)\right)\forall\:t=2,\ldots\:,T$$\end{document} 22\documentclass[12pt]{minimal} \usepackage{amsmath} \usepackage{wasysym} \usepackage{amsfonts} \usepackage{amssymb} \usepackage{amsbsy} \usepackage{mathrsfs} \usepackage{upgreek} \setlength{\oddsidemargin}{-69pt} \begin{document}$$\:{\boldsymbol{y}}_{t}\left({\boldsymbol{\xi\:}}_{\left[t-{\Delta\:}t,t-1\right]}\right)={\boldsymbol{y}}_{t}^{0}+{\boldsymbol{Y}}_{t}^{1}{\boldsymbol{\xi\:}}_{\left[t-{\Delta\:}t,t-1\right]}={\boldsymbol{y}}_{t}^{0}+\sum\:_{\tau\:=\mathrm{m}\mathrm{a}\mathrm{x}(1,t-{\Delta\:}t)}^{t-1}{\boldsymbol{y}}_{t,\tau\:}^{1}{\boldsymbol{\xi\:}}_{\tau\:}\:\forall\:t=2,\ldots\:,T$$\end{document} 23\documentclass[12pt]{minimal} \usepackage{amsmath} \usepackage{wasysym} \usepackage{amsfonts} \usepackage{amssymb} \usepackage{amsbsy} \usepackage{mathrsfs} \usepackage{upgreek} \setlength{\oddsidemargin}{-69pt} \begin{document}$$\:{\boldsymbol{y}}_{t}^{0}=\left({\boldsymbol{y}}_{t}^{\mathrm{T}\mathrm{r}\mathrm{a},0},{\boldsymbol{y}}_{t}^{\mathrm{P}\mathrm{u}\mathrm{r},0},{\boldsymbol{y}}_{t}^{\mathrm{B}\mathrm{u}\mathrm{d},0}\right),{\boldsymbol{Y}}_{t}^{1}=\left({\boldsymbol{Y}}_{t}^{\mathrm{T}\mathrm{r}\mathrm{a},1},{\boldsymbol{Y}}_{t}^{\mathrm{P}\mathrm{u}\mathrm{r},1},{\boldsymbol{Y}}_{t}^{\mathrm{B}\mathrm{u}\mathrm{d},1}\right)$$\end{document} Compact multi-stage robust optimization model Building upon the deterministic resource allocation model (“Multi-role integrated emergency resource allocation model”), periodic uncertainty sets (“Multi-period uncertainty sets”), and feedback-based decision rules (“Adaptive feedback adjustment rules”), we establish an integrated multi-stage robust, multi-role resource allocation model. To facilitate linearization, the model is presented using matrix–vector notation. First, we augment the original comprehensive objective (Eq. (10)) with recourse decision variables, resulting in Eq. (24). Here, \documentclass[12pt]{minimal} \usepackage{amsmath} \usepackage{wasysym} \usepackage{amsfonts} \usepackage{amssymb} \usepackage{amsbsy} \usepackage{mathrsfs} \usepackage{upgreek} \setlength{\oddsidemargin}{-69pt} \begin{document}$$\:\boldsymbol{x}={({a}_{\mathrm{1,1},1},\ldots\:,{a}_{t,j,k},\ldots\:,{a}_{T,J,K})}^{\mathrm{T}}\in\:{\mathbb{R}}^{F\times\:1}\:(F=T\cdot\:J\cdot\:K)$$\end{document} denotes the auxiliary variable vector for supply–demand gaps and \documentclass[12pt]{minimal} \usepackage{amsmath} \usepackage{wasysym} \usepackage{amsfonts} \usepackage{amssymb} \usepackage{amsbsy} \usepackage{mathrsfs} \usepackage{upgreek} \setlength{\oddsidemargin}{-69pt} \begin{document}$$\:\boldsymbol{c}={(1,\ldots\:,1)}^{\mathrm{T}}\in\:{\mathbb{R}}^{F\times\:1}$$\end{document} is its coefficient vector. \documentclass[12pt]{minimal} \usepackage{amsmath} \usepackage{wasysym} \usepackage{amsfonts} \usepackage{amssymb} \usepackage{amsbsy} \usepackage{mathrsfs} \usepackage{upgreek} \setlength{\oddsidemargin}{-69pt} \begin{document}$$\:{\boldsymbol{q}}_{t}\in\:{\mathbb{R}}^{P\times\:1}$$\end{document} is the coefficient vector for recourse decisions \documentclass[12pt]{minimal} \usepackage{amsmath} \usepackage{wasysym} \usepackage{amsfonts} \usepackage{amssymb} \usepackage{amsbsy} \usepackage{mathrsfs} \usepackage{upgreek} \setlength{\oddsidemargin}{-69pt} \begin{document}$$\:{\boldsymbol{y}}_{t}\left({\boldsymbol{\xi\:}}_{\left[\tau\:-{\Delta\:}t,t-1\right]}\right)$$\end{document}. \documentclass[12pt]{minimal} \usepackage{amsmath} \usepackage{wasysym} \usepackage{amsfonts} \usepackage{amssymb} \usepackage{amsbsy} \usepackage{mathrsfs} \usepackage{upgreek} \setlength{\oddsidemargin}{-69pt} \begin{document}$$\:{{\Xi\:}}_{t}$$\end{document} denotes the uncertainty set for period t (Eq. (17)).24\documentclass[12pt]{minimal} \usepackage{amsmath} \usepackage{wasysym} \usepackage{amsfonts} \usepackage{amssymb} \usepackage{amsbsy} \usepackage{mathrsfs} \usepackage{upgreek} \setlength{\oddsidemargin}{-69pt} \begin{document}$$\:\begin{array}{c}min\\\:x,{\boldsymbol{y}}_{t}(1\le\:t\le\:T)\end{array}\:\left\{{\boldsymbol{c}}^{\mathrm{T}}\boldsymbol{x}+{\boldsymbol{q}}_{1}^{\mathrm{T}}{\boldsymbol{y}}_{1}+\begin{array}{c}max\\\:{\boldsymbol{\xi\:}}_{t}\in\:{{\Xi\:}}_{t}\\\:\left(1\le\:t\le\:T-1\right)\end{array}\:\sum\:_{t=2}^{T}{\boldsymbol{q}}_{t}^{\mathrm{T}}{\boldsymbol{y}}_{t}\left({\boldsymbol{\xi\:}}_{\left[t-{\Delta\:}t,t-1\right]}\right)\right\}$$\end{document} Next, incorporating uncertainty parameters \documentclass[12pt]{minimal} \usepackage{amsmath} \usepackage{wasysym} \usepackage{amsfonts} \usepackage{amssymb} \usepackage{amsbsy} \usepackage{mathrsfs} \usepackage{upgreek} \setlength{\oddsidemargin}{-69pt} \begin{document}$$\:{\boldsymbol{\xi\:}}_{t}\in\:{{\Xi\:}}_{t}$$\end{document} and recourse decisions \documentclass[12pt]{minimal} \usepackage{amsmath} \usepackage{wasysym} \usepackage{amsfonts} \usepackage{amssymb} \usepackage{amsbsy} \usepackage{mathrsfs} \usepackage{upgreek} \setlength{\oddsidemargin}{-69pt} \begin{document}$$\:{\boldsymbol{y}}_{t}\left({\boldsymbol{\xi\:}}_{\left[\tau\:-{\Delta\:}t,t-1\right]}\right)$$\end{document} transforms the original constraints (Eqs. 1–4, 9) into Eqs. (25) and (26). Equation (25) represents Period-1 constraints without historical uncertainty or adjustments, while Eq. (26) represents constraints for periods t ≥ 2 that account for historical uncertainties and corresponding adjustments. In addition, \documentclass[12pt]{minimal} \usepackage{amsmath} \usepackage{wasysym} \usepackage{amsfonts} \usepackage{amssymb} \usepackage{amsbsy} \usepackage{mathrsfs} \usepackage{upgreek} \setlength{\oddsidemargin}{-69pt} \begin{document}$$\:{\boldsymbol{\xi\:}}_{\left[t\right]}$$\end{document}:=(\documentclass[12pt]{minimal} \usepackage{amsmath} \usepackage{wasysym} \usepackage{amsfonts} \usepackage{amssymb} \usepackage{amsbsy} \usepackage{mathrsfs} \usepackage{upgreek} \setlength{\oddsidemargin}{-69pt} \begin{document}$$\:{\boldsymbol{\xi\:}}_{1},\ldots\:,{\boldsymbol{\xi\:}}_{t}$$\end{document}) collects uncertainties up to period t. \documentclass[12pt]{minimal} \usepackage{amsmath} \usepackage{wasysym} \usepackage{amsfonts} \usepackage{amssymb} \usepackage{amsbsy} \usepackage{mathrsfs} \usepackage{upgreek} \setlength{\oddsidemargin}{-69pt} \begin{document}$$\:{\boldsymbol{y}}_{t}\left({\boldsymbol{\xi\:}}_{\left[t-{\Delta\:}t,t-1\right]}\right)$$\end{document} follows Eqs. (22) and (23). Matrices \documentclass[12pt]{minimal} \usepackage{amsmath} \usepackage{wasysym} \usepackage{amsfonts} \usepackage{amssymb} \usepackage{amsbsy} \usepackage{mathrsfs} \usepackage{upgreek} \setlength{\oddsidemargin}{-69pt} \begin{document}$$\:{\boldsymbol{W}}_{t,\tau\:}\in\:{\mathbb{R}}^{M\times\:P},{\boldsymbol{T}}_{t}\in\:{\mathbb{R}}^{M\times\:F},{\boldsymbol{B}}_{t}\in\:{\mathbb{R}}^{M\times\:\left(N\cdot\:t\right)}$$\end{document} denote the coefficient matrices associated with \documentclass[12pt]{minimal} \usepackage{amsmath} \usepackage{wasysym} \usepackage{amsfonts} \usepackage{amssymb} \usepackage{amsbsy} \usepackage{mathrsfs} \usepackage{upgreek} \setlength{\oddsidemargin}{-69pt} \begin{document}$$\:{\boldsymbol{y}}_{t},\boldsymbol{x},{\boldsymbol{\xi\:}}_{\left[t\right]}$$\end{document} in the original constraints, where \documentclass[12pt]{minimal} \usepackage{amsmath} \usepackage{wasysym} \usepackage{amsfonts} \usepackage{amssymb} \usepackage{amsbsy} \usepackage{mathrsfs} \usepackage{upgreek} \setlength{\oddsidemargin}{-69pt} \begin{document}$$\:\tau\:$$\end{document} indicates a historical period index prior to period t. \documentclass[12pt]{minimal} \usepackage{amsmath} \usepackage{wasysym} \usepackage{amsfonts} \usepackage{amssymb} \usepackage{amsbsy} \usepackage{mathrsfs} \usepackage{upgreek} \setlength{\oddsidemargin}{-69pt} \begin{document}$$\:{\boldsymbol{u}}_{t}\in\:{\mathbb{R}}^{M\times\:1}$$\end{document} is the corresponding constant vector, and M denotes the number of constraints per period.25\documentclass[12pt]{minimal} \usepackage{amsmath} \usepackage{wasysym} \usepackage{amsfonts} \usepackage{amssymb} \usepackage{amsbsy} \usepackage{mathrsfs} \usepackage{upgreek} \setlength{\oddsidemargin}{-69pt} \begin{document}$$\:{\boldsymbol{W}}_{\mathrm{1,1}}{\boldsymbol{y}}_{1}+{\boldsymbol{T}}_{1}\boldsymbol{x}\ge\:{\boldsymbol{B}}_{1}{\boldsymbol{\xi\:}}_{1}+{\boldsymbol{u}}_{1}\:\forall\:{{\boldsymbol{y}}_{1}\in\:{\mathbb{R}}^{P\times\:1},\boldsymbol{\xi\:}}_{1}\in\:{{\Xi\:}}_{1}$$\end{document} 26\documentclass[12pt]{minimal} \usepackage{amsmath} \usepackage{wasysym} \usepackage{amsfonts} \usepackage{amssymb} \usepackage{amsbsy} \usepackage{mathrsfs} \usepackage{upgreek} \setlength{\oddsidemargin}{-69pt} \begin{document}$$\:{\boldsymbol{W}}_{\mathrm{1,1}}{\boldsymbol{y}}_{1}+\sum\:_{\tau\:=2}^{t}{\boldsymbol{W}}_{t,\tau\:}{\boldsymbol{y}}_{\tau\:}\left({\boldsymbol{\xi\:}}_{\left[\tau\:-{\Delta\:}t,\tau\:-1\right]}\right)+{\boldsymbol{T}}_{t}\boldsymbol{x}\ge\:{\boldsymbol{B}}_{t}{\boldsymbol{\xi\:}}_{\left[t\right]}+{\boldsymbol{u}}_{t}\:\forall\:{\boldsymbol{y}}_{t}\left({\boldsymbol{\xi\:}}_{\left[t-{\Delta\:}t,t-1\right]}\right)\in\:{\mathbb{R}}^{P\times\:1},{\boldsymbol{\xi\:}}_{t}\in\:{{\Xi\:}}_{t},t=2,\ldots\:,T$$\end{document} Linear reformulation of the robust allocation model The robust optimization model for dynamic resource allocation contains uncertain parameters and semi-infinite constraints, making it computationally intractable for direct solution by off-the-shelf solvers. To overcome this difficulty, we apply linear programming duality and semi-infinite constraint techniques25 to transform the original MSARO model into a tractable linear programming formulation. The LP reformulation is exact for the ADR-based MSARO formulation and therefore does not introduce additional approximation errors beyond the affine decision rule approximation. Consequently, the robustness and optimality properties of the ADR-based model are preserved. First, substituting the affine decision rules (Eqs. 22 and 23) into the objective function (Eq. 24) yields Eq. (27). Substituting them into constraints (Eqs. 25 and 26) and rearranging terms gives Eq. (28), where \documentclass[12pt]{minimal} \usepackage{amsmath} \usepackage{wasysym} \usepackage{amsfonts} \usepackage{amssymb} \usepackage{amsbsy} \usepackage{mathrsfs} \usepackage{upgreek} \setlength{\oddsidemargin}{-69pt} \begin{document}$$\:{{\boldsymbol{B}}^{\boldsymbol{{\prime\:}}}}_{t,\tau\:}{\in\:\mathbb{R}}^{M\times\:N}$$\end{document} is the submatrix of \documentclass[12pt]{minimal} \usepackage{amsmath} \usepackage{wasysym} \usepackage{amsfonts} \usepackage{amssymb} \usepackage{amsbsy} \usepackage{mathrsfs} \usepackage{upgreek} \setlength{\oddsidemargin}{-69pt} \begin{document}$$\:{\boldsymbol{B}}_{t}\:$$\end{document}corresponding to \documentclass[12pt]{minimal} \usepackage{amsmath} \usepackage{wasysym} \usepackage{amsfonts} \usepackage{amssymb} \usepackage{amsbsy} \usepackage{mathrsfs} \usepackage{upgreek} \setlength{\oddsidemargin}{-69pt} \begin{document}$$\:{\boldsymbol{\xi\:}}_{\tau\:}$$\end{document} (\documentclass[12pt]{minimal} \usepackage{amsmath} \usepackage{wasysym} \usepackage{amsfonts} \usepackage{amssymb} \usepackage{amsbsy} \usepackage{mathrsfs} \usepackage{upgreek} \setlength{\oddsidemargin}{-69pt} \begin{document}$$\:1\le\:\tau\:\le\:t$$\end{document}). \documentclass[12pt]{minimal} \usepackage{amsmath} \usepackage{wasysym} \usepackage{amsfonts} \usepackage{amssymb} \usepackage{amsbsy} \usepackage{mathrsfs} \usepackage{upgreek} \setlength{\oddsidemargin}{-69pt} \begin{document}$$\:t,\tau\:,\:\mathrm{a}\mathrm{n}\mathrm{d}\:{\tau\:}^{{\prime\:}}$$\end{document} denote time period indices, and \documentclass[12pt]{minimal} \usepackage{amsmath} \usepackage{wasysym} \usepackage{amsfonts} \usepackage{amssymb} \usepackage{amsbsy} \usepackage{mathrsfs} \usepackage{upgreek} \setlength{\oddsidemargin}{-69pt} \begin{document}$$\:\tau\:,\:\mathrm{a}\mathrm{n}\mathrm{d}\:{\tau\:}^{{\prime\:}}$$\end{document} separately denote historical period indices preceding periods \documentclass[12pt]{minimal} \usepackage{amsmath} \usepackage{wasysym} \usepackage{amsfonts} \usepackage{amssymb} \usepackage{amsbsy} \usepackage{mathrsfs} \usepackage{upgreek} \setlength{\oddsidemargin}{-69pt} \begin{document}$$\:t\:\mathrm{a}\mathrm{n}\mathrm{d}\:\tau\:$$\end{document}.27\documentclass[12pt]{minimal} \usepackage{amsmath} \usepackage{wasysym} \usepackage{amsfonts} \usepackage{amssymb} \usepackage{amsbsy} \usepackage{mathrsfs} \usepackage{upgreek} \setlength{\oddsidemargin}{-69pt} \begin{document}$$\:\begin{array}{c}min\\\:x,{\boldsymbol{y}}_{1},{\boldsymbol{y}}_{t}^{0},{\boldsymbol{y}}_{t,\tau\:}^{1}\\\:(1\le\:t\le\:T)\end{array}\:\left\{{\boldsymbol{c}}^{\mathrm{T}}\boldsymbol{x}+{\boldsymbol{q}}_{1}^{\mathrm{T}}{\boldsymbol{y}}_{1}+\begin{array}{c}max\\\:{\boldsymbol{\xi\:}}_{t}\in\:{{\Xi\:}}_{t}\\\:(1\le\:t\le\:T-1)\end{array}\:\sum\:_{t=2}^{T}({\boldsymbol{q}}_{t}^{\mathrm{T}}{\boldsymbol{y}}_{t}^{0}+{\boldsymbol{q}}_{t}^{\mathrm{T}}\sum\:_{\tau\:=\mathrm{max}\left(1,t-{\Delta\:}t\right)}^{t-1}{\boldsymbol{y}}_{t,\tau\:}^{1}{\boldsymbol{\xi\:}}_{\tau\:})\right\}$$\end{document} 28\documentclass[12pt]{minimal} \usepackage{amsmath} \usepackage{wasysym} \usepackage{amsfonts} \usepackage{amssymb} \usepackage{amsbsy} \usepackage{mathrsfs} \usepackage{upgreek} \setlength{\oddsidemargin}{-69pt} \begin{document}$$\:\left\{\begin{array}{c}{\boldsymbol{W}}_{\mathrm{1,1}}{\boldsymbol{y}}_{1}+{\boldsymbol{T}}_{1}x\ge\:{\boldsymbol{B}}_{1}{\boldsymbol{\xi\:}}_{1}+{\boldsymbol{u}}_{1}\:\forall\:{\boldsymbol{\xi\:}}_{1}\in\:{{\Xi\:}}_{1}\:\:\left(a\right)\\\:{\boldsymbol{W}}_{\mathrm{1,1}}{\boldsymbol{y}}_{1}+\sum\:_{\tau\:=2}^{t}{\boldsymbol{W}}_{t,\tau\:}({\boldsymbol{y}}_{\tau\:}^{0}+\sum\:_{{\tau\:}^{{\prime\:}}=\mathrm{max}\left(1,\tau\:-{\Delta\:}t\right)}^{\tau\:-1}{\boldsymbol{y}}_{\tau\:,{\tau\:}^{{\prime\:}}}^{1}{\boldsymbol{\xi\:}}_{{\tau\:}^{{\prime\:}}})+{\boldsymbol{T}}_{t}x\ge\:\sum\:_{\tau\:=1}^{t}{{\boldsymbol{B}}^{\boldsymbol{{\prime\:}}}}_{t,\tau\:}{\boldsymbol{\xi\:}}_{\tau\:}+{\boldsymbol{u}}_{t}\\\:\forall\:{\boldsymbol{\xi\:}}_{t}\in\:{{\Xi\:}}_{t},2\le\:t\le\:T\:\:\:\left(b\right)\end{array}\right.$$\end{document} Through robust dual reformulation and semi-infinite constraint techniques, Eqs. (27) and (28) are transformed into the robust counterpart, namely Eqs. (42)–(44). The detailed derivation is provided in Appendix A, Eqs. (36)–(44). The nested minimization in Eq. (42) is then reformulated as a single-level minimization problem and decoupled into objective and constraints, yielding Eqs. (29) and (30). For Eqs. (43) and (44), since the embedded minimization problems are satisfied, the inequalities hold for any \documentclass[12pt]{minimal} \usepackage{amsmath} \usepackage{wasysym} \usepackage{amsfonts} \usepackage{amssymb} \usepackage{amsbsy} \usepackage{mathrsfs} \usepackage{upgreek} \setlength{\oddsidemargin}{-69pt} \begin{document}$$\:{\boldsymbol{\alpha\:}}_{1}$$\end{document} and \documentclass[12pt]{minimal} \usepackage{amsmath} \usepackage{wasysym} \usepackage{amsfonts} \usepackage{amssymb} \usepackage{amsbsy} \usepackage{mathrsfs} \usepackage{upgreek} \setlength{\oddsidemargin}{-69pt} \begin{document}$$\:{\boldsymbol{\alpha\:}}_{t,\tau\:}$$\end{document}. Therefore, Eqs. (43) and (44) simplify to Eqs. (31)~(35). The MSARO model is thus reformulated as the tractable linear program (Eqs. 29~35) with variables: \documentclass[12pt]{minimal} \usepackage{amsmath} \usepackage{wasysym} \usepackage{amsfonts} \usepackage{amssymb} \usepackage{amsbsy} \usepackage{mathrsfs} \usepackage{upgreek} \setlength{\oddsidemargin}{-69pt} \begin{document}$$\begin{aligned} & \boldsymbol{x}\in\:{\mathbb{R}}^{F\times\:1},{\boldsymbol{y}}_{1}\in\:{\mathbb{R}}^{P\times\:1},{\boldsymbol{y}}_{t}^{0}\in\:{\mathbb{R}}^{P\times\:1},{\boldsymbol{y}}_{t,\tau\:}^{1}\in\:{\mathbb{R}}^{P\times\:N}\:(\tau\:=\mathrm{m}\mathrm{a}\mathrm{x}(1,t-{\Delta\:}t),\ldots\:,t-1;t=2,\ldots\:,T), \\ & {\boldsymbol{\beta\:}}_{t}\in\:{\mathbb{R}}_{+}^{V},\:{\boldsymbol{\alpha\:}}_{1}\in\:{\mathbb{R}}_{+}^{V\times\:M},{\boldsymbol{\alpha\:}}_{t,\tau\:}\in\:{\mathbb{R}}_{+}^{V\times\:M}(\tau\:=1,\ldots\:,t;t=2,\ldots\:,T) \end{aligned}$$\end{document} 29\documentclass[12pt]{minimal} \usepackage{amsmath} \usepackage{wasysym} \usepackage{amsfonts} \usepackage{amssymb} \usepackage{amsbsy} \usepackage{mathrsfs} \usepackage{upgreek} \setlength{\oddsidemargin}{-69pt} \begin{document}$$\:\begin{array}{c}min\\\:x,{\boldsymbol{y}}_{1},{\boldsymbol{y}}_{t}^{0}{,\boldsymbol{\beta\:}}_{t}\\\:(1\le\:t\le\:T)\end{array}\:\left\{{\boldsymbol{c}}^{\mathrm{T}}\boldsymbol{x}+{\boldsymbol{q}}_{1}^{\mathrm{T}}{\boldsymbol{y}}_{1}+\sum\:_{t=2}^{T}{\boldsymbol{q}}_{t}^{\mathrm{T}}{\boldsymbol{y}}_{t}^{0}+\sum\:_{t=1}^{T-1}{\boldsymbol{z}}_{\tau\:}^{\mathrm{T}}{\boldsymbol{\beta\:}}_{t}\right\}$$\end{document} 30\documentclass[12pt]{minimal} \usepackage{amsmath} \usepackage{wasysym} \usepackage{amsfonts} \usepackage{amssymb} \usepackage{amsbsy} \usepackage{mathrsfs} \usepackage{upgreek} \setlength{\oddsidemargin}{-69pt} \begin{document}$$\:{\boldsymbol{H}}_{t}^{\mathrm{T}}{\boldsymbol{\beta\:}}_{t}=\sum\:_{\tau\:=t+1}^{\mathrm{min}\left(T,t+{\Delta\:}t\right)}{{\boldsymbol{y}}_{\tau\:,t}^{1}}^{\mathrm{T}}{\boldsymbol{q}}_{\tau\:}\:\forall\:t=1,\ldots\:,T-1$$\end{document} 31\documentclass[12pt]{minimal} \usepackage{amsmath} \usepackage{wasysym} \usepackage{amsfonts} \usepackage{amssymb} \usepackage{amsbsy} \usepackage{mathrsfs} \usepackage{upgreek} \setlength{\oddsidemargin}{-69pt} \begin{document}$$\:{\boldsymbol{W}}_{\mathrm{1,1}}{\boldsymbol{y}}_{1}+{\boldsymbol{T}}_{1}\boldsymbol{x}-{\boldsymbol{u}}_{1}\ge\:{\boldsymbol{\alpha\:}}_{1}^{T}{\boldsymbol{z}}_{1}$$\end{document} 32\documentclass[12pt]{minimal} \usepackage{amsmath} \usepackage{wasysym} \usepackage{amsfonts} \usepackage{amssymb} \usepackage{amsbsy} \usepackage{mathrsfs} \usepackage{upgreek} \setlength{\oddsidemargin}{-69pt} \begin{document}$$\:{\boldsymbol{H}}_{1}^{T}{\boldsymbol{\alpha\:}}_{1}={\boldsymbol{B}}_{1}^{T}$$\end{document} 33\documentclass[12pt]{minimal} \usepackage{amsmath} \usepackage{wasysym} \usepackage{amsfonts} \usepackage{amssymb} \usepackage{amsbsy} \usepackage{mathrsfs} \usepackage{upgreek} \setlength{\oddsidemargin}{-69pt} \begin{document}$$\:{\boldsymbol{W}}_{\mathrm{1,1}}{\boldsymbol{y}}_{1}+\sum\:_{\tau\:=2}^{t}{\boldsymbol{W}}_{t,\tau\:}{\boldsymbol{y}}_{\tau\:}^{0}+{\boldsymbol{T}}_{t}\boldsymbol{x}-{\boldsymbol{u}}_{t}\ge\:\sum\:_{\tau\:=1}^{t}{\boldsymbol{\alpha\:}}_{t,\tau\:}^{\mathrm{T}}{\boldsymbol{z}}_{\tau\:}\:\forall\:t=2,\ldots\:,T$$\end{document} 34\documentclass[12pt]{minimal} \usepackage{amsmath} \usepackage{wasysym} \usepackage{amsfonts} \usepackage{amssymb} \usepackage{amsbsy} \usepackage{mathrsfs} \usepackage{upgreek} \setlength{\oddsidemargin}{-69pt} \begin{document}$$\:{\boldsymbol{H}}_{\tau\:}^{\mathrm{T}}{\boldsymbol{\alpha\:}}_{t,\tau\:}={\boldsymbol{B}{\prime\:}}_{t,\tau\:}^{\mathrm{T}}-\sum\:_{{\tau\:}^{{\prime\:}}=\tau\:+1}^{\mathrm{min}\left(t,\tau\:+{\Delta\:}t\right)}{\left({\boldsymbol{W}}_{t,{\tau\:}^{{\prime\:}}}{\boldsymbol{y}}_{{\tau\:}^{{\prime\:}},\tau\:}^{1}\right)}^{\mathrm{T}}\:\forall\:\tau\:=1,\ldots\:,t-1;t=2,\ldots\:,T$$\end{document} 35\documentclass[12pt]{minimal} \usepackage{amsmath} \usepackage{wasysym} \usepackage{amsfonts} \usepackage{amssymb} \usepackage{amsbsy} \usepackage{mathrsfs} \usepackage{upgreek} \setlength{\oddsidemargin}{-69pt} \begin{document}$$\:{\boldsymbol{H}}_{t}^{\mathrm{T}}{\boldsymbol{\alpha\:}}_{t,t}={\boldsymbol{B}{\prime\:}}_{t,t}^{\mathrm{T}}\:\forall\:t=2,\ldots\:,T$$\end{document} Although the linear reformulation removes nonlinearities and improves solvability, it introduces additional dual variables and constraints, increasing overall model size. This trade-off is common in robust optimization, where improved tractability and numerical stability may come at the cost of higher dimensionality. Experiment analysis and discussion Case study description Earthquakes are hazard-induced disasters that threaten human lives and property. This case study considers a major earthquake event causing surface ruptures, landslides, and debris flows that damaged transportation infrastructure. After the disaster, local authorities established an emergency command center as the decision role to coordinate response efforts, including fund allocation. Two municipal transfer stations (TS#1–2) serve as the execution role. The demand role consists of two temporary shelters located in severely affected disaster areas (DA#1–2). Figure 5 shows the geographical distribution and resource flow directions. Table 1 lists the affected populations and transportation distances35. This case study is designed as an illustrative example to examine the behavior of the proposed framework under a controlled post-earthquake setting rather than as a comprehensive empirical validation across multiple disaster scenarios or scales. Fig. 5Schematic diagram of post-earthquake emergency resource allocation from transfer stations to disaster areas. Note: The map was generated using QGIS (version 3.44.8, https://qgis.org) with an XYZ raster tiles connection (CartoDB Voyager, https://cartodb-basemaps-a.global.ssl.fastly.net/rastertiles/voyager/{z}/{x}/{y}.png). Schematic diagram of post-earthquake emergency resource allocation from transfer stations to disaster areas. Note: The map was generated using QGIS (version 3.44.8, https://qgis.org) with an XYZ raster tiles connection (CartoDB Voyager, https://cartodb-basemaps-a.global.ssl.fastly.net/rastertiles/voyager/{z}/{x}/{y}.png). Table 1Affected populations and transportation distances.Transfer station/disaster areaDA#1DA#2TS#1 (km)143211TS#2 (km)177177Affected population34,58326,916 Table 1Affected populations and transportation distances.Transfer station/disaster areaDA#1DA#2TS#1 (km)143211TS#2 (km)177177Affected population34,58326,916 Affected populations and transportation distances. Parameter configuration. Parameter configuration. Parameter configuration. Standard 2-ton trucks are used for resource transportation. Referencing highway toll standards36 and accounting for road damage and repair progress, transportation costs per vehicle are set as shown in Table 2. Roads experience the most severe damage on Day 1, leading to the highest transportation costs. Costs gradually decrease as repairs progress. The transportation loss compensation ratio is set to 0.05. Table 2Post-disaster highway tolls.Post-disaster timeDay 1Day 2Day 3Toll per truck (¥/km)0.80.70.6 Table 2Post-disaster highway tolls.Post-disaster timeDay 1Day 2Day 3Toll per truck (¥/km)0.80.70.6 Post-disaster highway tolls. Earthquake relief resources include first-aid kits, instant food, and drinking water37. First-aid kits are single-use resources, with one kit satisfying one survivor’s needs during the rescue window, while food and water are daily consumables required by each survivor. Resource specifications are listed in Table 3. Table 3Emergency resource standards.Resource typeResourceCostLoading capacityAllocation standard1First-aid kits¥160/unit600 units1 unit/person2Instant food¥5/bag4000 bags6 bags/person/day3Drinking water¥3/L2000 L2 L/person/day Table 3Emergency resource standards.Resource typeResourceCostLoading capacityAllocation standard1First-aid kits¥160/unit600 units1 unit/person2Instant food¥5/bag4000 bags6 bags/person/day3Drinking water¥3/L2000 L2 L/person/day Emergency resource standards. To reduce subjective decision bias, marginal life-saving utility (MLSU) is calculated using the entropy method based on rescue effectiveness and transportation distance38. Regions farther from transfer stations yield higher marginal utility per unit resource. On the first day, when disaster information is limited, MLSU is estimated using average transportation distance. As site-specific information becomes available in subsequent days, MLSU values are updated accordingly, as shown in Table 4. Table 4Marginal life-saving utility per disaster area33.Disaster areasDay 1Day 2Day 3DA#10.41.02.82DA#20.52.281.21 Table 4Marginal life-saving utility per disaster area33.Disaster areasDay 1Day 2Day 3DA#10.41.02.82DA#20.52.281.21 Marginal life-saving utility per disaster area33. To characterize uncertainties in demand, procurement, and transportation, normal distributions with fixed standard deviations are adopted as a parsimonious benchmark, i.e., \documentclass[12pt]{minimal} \usepackage{amsmath} \usepackage{wasysym} \usepackage{amsfonts} \usepackage{amssymb} \usepackage{amsbsy} \usepackage{mathrsfs} \usepackage{upgreek} \setlength{\oddsidemargin}{-69pt} \begin{document}$$\:{\xi\:}_{t,j,k}^{\mathrm{D}\mathrm{e}\mathrm{m}}\sim\mathcal{N}$$\end{document}(0, \documentclass[12pt]{minimal} \usepackage{amsmath} \usepackage{wasysym} \usepackage{amsfonts} \usepackage{amssymb} \usepackage{amsbsy} \usepackage{mathrsfs} \usepackage{upgreek} \setlength{\oddsidemargin}{-69pt} \begin{document}$$\:{{\sigma\:}_{t,j,k}^{\mathrm{D}\mathrm{e}\mathrm{m}}}^{2}$$\end{document}), \documentclass[12pt]{minimal} \usepackage{amsmath} \usepackage{wasysym} \usepackage{amsfonts} \usepackage{amssymb} \usepackage{amsbsy} \usepackage{mathrsfs} \usepackage{upgreek} \setlength{\oddsidemargin}{-69pt} \begin{document}$$\:{\xi\:}_{t,i,k}^{\mathrm{P}\mathrm{u}\mathrm{r}}\sim\mathcal{N}$$\end{document}(0, \documentclass[12pt]{minimal} \usepackage{amsmath} \usepackage{wasysym} \usepackage{amsfonts} \usepackage{amssymb} \usepackage{amsbsy} \usepackage{mathrsfs} \usepackage{upgreek} \setlength{\oddsidemargin}{-69pt} \begin{document}$$\:{{\sigma\:}_{t,i,k}^{\mathrm{P}\mathrm{u}\mathrm{r}}}^{2}$$\end{document}) and \documentclass[12pt]{minimal} \usepackage{amsmath} \usepackage{wasysym} \usepackage{amsfonts} \usepackage{amssymb} \usepackage{amsbsy} \usepackage{mathrsfs} \usepackage{upgreek} \setlength{\oddsidemargin}{-69pt} \begin{document}$$\:{\xi\:}_{t,i,j,k}^{\mathrm{T}\mathrm{r}\mathrm{a}}\sim\mathcal{N}$$\end{document}(0, \documentclass[12pt]{minimal} \usepackage{amsmath} \usepackage{wasysym} \usepackage{amsfonts} \usepackage{amssymb} \usepackage{amsbsy} \usepackage{mathrsfs} \usepackage{upgreek} \setlength{\oddsidemargin}{-69pt} \begin{document}$$\:{{\sigma\:}_{t,i,j,k}^{\mathrm{T}\mathrm{r}\mathrm{a}}}^{2}$$\end{document}). This assumption does not aim to fully represent real stochastic processes, which may exhibit asymmetry or heavy tails, but provides a transparent and comparable description of uncertainty levels. The proposed MSARO does not rely on the exact distributional form, since decisions remain feasible under worst-case realizations within specified uncertainty sets; the distributional assumption mainly supports scenario parameterization. According to practical post-disaster assessment processes, ratios of uncertainty bounds to predicted values are set as shown in Table 5, enabling uncertainty sets to cover approximately 95% of uncertainty realizations. Referring to literature39, uncertainty is largest on the first day due to limited information, reaching about 20% of predicted values. Procurement and transportation are assumed not to exceed planned quantities; therefore, their deviations correspond to the non-positive part of the normal distribution. As damage assessment and infrastructure recovery progress, uncertainty gradually decreases to 5–10%. Meanwhile, improved survivor information leads to updated demand estimates, causing uncertainty intervals to change from [0.6, 0.8] to [0.95, 1.05]. Table 5Uncertainty bounds (ratio to estimates).Post-disaster timeDemandProcurementTransportationDay 1[0.6, 0.8][0.8, 1.0][0.8, 1.0]Day 2[0.8, 0.95][0.9, 1.0][0.9, 1.0]Day 3[0.95, 1.05][0.95, 1.0][0.95, 1.0] Table 5Uncertainty bounds (ratio to estimates).Post-disaster timeDemandProcurementTransportationDay 1[0.6, 0.8][0.8, 1.0][0.8, 1.0]Day 2[0.8, 0.95][0.9, 1.0][0.9, 1.0]Day 3[0.95, 1.05][0.95, 1.0][0.95, 1.0] Uncertainty bounds (ratio to estimates). (2)Model execution and evaluation metrics. (2)Model execution and evaluation metrics. Model execution and evaluation metrics. Parameters are set as = 1 and T = 3, dividing the 72-hour rescue window into three daily periods. Dynamic allocation strategies are updated based on prior-day uncertainty feedback. Monte Carlo simulation generates 1,000 uncertainty samples for testing. For each sample, model-derived allocation strategies are executed to evaluate stability and adaptability. The model is implemented in Python 3.8 using Gurobipy (v11.2) with a runtime limit of 5 min. All tests run on Windows 11 with a 6-core CPU (AMD Ryzen 5 5600G, 3.90 GHz) and 32 GB RAM. Several metrics are adopted to evaluate model performance: (1) Realized comprehensive objective: mean and standard deviation obtained from Monte Carlo simulations indicate solution optimality and stability; (2) Model scale: numbers of variables and constraints affect computational efficiency; (3) Solution time: reflects computational efficiency and influences response timeliness; (4) Constraint violation rate: proportion of samples violating constraints measures decision reliability. Uncertainty sets are constructed to cover approximately 95% of empirical uncertainty distributions, since real-world uncertainty boundaries are often unknown. Consequently, rare realizations outside the uncertainty sets may occur in simulations. In such cases, allocation decisions may violate certain constraints, such as fund flow, resource flow, and supply–demand balance constraints defined in “Constraints”. Because test samples are randomly generated, violation locations and types are stochastic and cannot be predetermined. While these metrics do not exhaust all evaluation dimensions, they capture key trade-offs relevant to emergency resource allocation under uncertainty. Analysis of emergency resource allocation results Based on the case setup above, this section analyzes the allocation process and optimization results of the model under uncertainty. The analysis covers resource demand responses, multi-stage management of funds and resources, and multi-role optimization objectives, and examines the model’s adaptability and robustness under uncertain conditions. Dynamics and uncertainty of emergency resource demand. Dynamics and uncertainty of emergency resource demand. Dynamics and uncertainty of emergency resource demand. Figure 6 illustrates the predicted demand values and the actual distributions under uncertainty samples for three types of emergency resources at the two disaster areas (DA#1 and DA#2) during the 72-h post-disaster period. On Day 1, the uncertainty range for resource demand is the largest (approximately ± 20%), consistent with the initial lack of disaster information (see Table 5). As disaster information becomes clearer on Days 2 and 3 through updated loss assessments and population statistics, the uncertainty range narrows significantly (within ± 5%), consistent with the parameter settings. Uncertainty in demand for single-use resources (first-aid kits) is mainly concentrated on Day 1. In contrast, uncertainty for daily-consumption resources (instant food and drinking water) persists throughout the golden rescue period and correlates with population estimation errors. This occurs because newly identified affected populations receive supplementary first-aid kits on subsequent days, while daily consumption of food and water continues to increase as more affected populations are identified. DA#1, which has a larger population, generally shows higher demand levels and larger fluctuation ranges than DA#2. This time-varying uncertainty and demand disparity provide suitable testing conditions for evaluating model performance. Fig. 6Resource demand at disaster areas (pre-disaster estimates vs. actual sample values). Resource demand at disaster areas (pre-disaster estimates vs. actual sample values). (2)Analysis of multi-role integrated and multi-stage robust allocation. (2)Analysis of multi-role integrated and multi-stage robust allocation. Analysis of multi-role integrated and multi-stage robust allocation. Figure 7, 8 and 9 show that the model formulates a global allocation plan in the pre-disaster phase via optimization and supports dynamic coordinated allocation across the “Decision–Execution–Demand” triad after the disaster through recourse decisions updated daily under uncertainty. Figure 7 shows the periodic and differentiated fund allocation strategy of the decision role. Fund allocation to the two transfer stations (TS#1–2) is highest on Day 1 without adjustment to address initial uncertainty and urgent procurement and transportation needs. Subsequent allocations are dynamically adjusted according to feedback from the previous day. For instance, if Day 1 results indicate higher-than-expected demand at DA#1 or increased transportation shortfalls from TS#1 to DA#1, the allocation to TS#1 on Day 2 exceeds the pre-planned budget, with minor adjustments to TS#2. Fund flow serves as the basis for subsequent allocation decisions through its dynamic updates. Fig. 7Fund allocation by the decision role (pre-disaster plan vs. dynamic adjustments in post-disaster samples). Fund allocation by the decision role (pre-disaster plan vs. dynamic adjustments in post-disaster samples). Fig. 8Resource procurement at transfer stations by the execution role (pre-disaster plan vs. dynamic adjustments in post-disaster samples). Resource procurement at transfer stations by the execution role (pre-disaster plan vs. dynamic adjustments in post-disaster samples). Fig. 9Resource transportation from execution to demand roles (pre-disaster plan vs. dynamic adjustments in post-disaster samples). Resource transportation from execution to demand roles (pre-disaster plan vs. dynamic adjustments in post-disaster samples). Figure 8 shows that the execution role coordinates resource procurement with fund allocation and updates procurement decisions according to demand changes. On Day 1, TS#1–2 procure resources based on initial demand estimates to meet urgent needs. Procurement plans in later periods are updated according to fund availability and actual procurement and demand information from previous periods. For example, if procurement difficulties at TS#1 occur on Day 2, TS#1 moderately increases procurement of that resource on Day 3, or TS#2 shares part of the procurement task. The procurement strategy links funds allocated by the decision role with transportation decisions executed in later stages. Resource transportation connects available resources (procurement minus inventory) with evolving demand. Figure 9 shows that the execution role prioritizes routes with higher MLSU values and adjusts transportation quantities dynamically. For instance, if transportation shortfalls increase on the TS#1→DA#1 route, shipments on this route are reduced in subsequent periods compared with the plan, and part of the capacity is shifted to the TS#1→DA#2 or TS#2→DA#1 routes. Transportation quantities are also adjusted according to actual demand reported by the demand role. For example, if actual resource consumption at DA#1 on Day 1 exceeds expectations, shipments to DA#1 increase on Day 2. This demand- and procurement-feedback-based adjustment helps reduce supply–demand imbalances and reflects coordination between execution and demand roles under dynamic allocation. In summary, the integrated allocation follows the initial plan on Day 1, and feedback-based adjustments are implemented from Day 2 onward. The direction and magnitude of adjustments depend on newly disclosed uncertainty information, illustrating how post-disaster feedback influences allocation decisions. Procurement and transportation of first-aid kits peak on Day 1, while those of food and water increase over time, as shown in Figs. 8 and 9. This occurs because first-aid kits are mainly supplied early, with later deliveries serving as supplements for newly identified populations, whereas food and water are required daily for the growing affected population. (3)Dynamic evolution of key states and robustness. (3)Dynamic evolution of key states and robustness. Dynamic evolution of key states and robustness. The trajectories of cumulative fund balance, resource inventory surplus, and supply–demand gap are key state variables used to examine the robustness of allocation decisions. According to Fig. 10, the cumulative fund balance at each transfer station remains non-negative in most test samples, indicating that the core constraint (Eq. 1) is generally satisfied. The planned daily balances for TS#1 and TS#2 approach zero, maintaining balance between fund inflow and expenditure while controlling economic costs. Although uncertainty leads to fluctuations in actual balances, adjusted fund allocation and procurement decisions maintain fund availability, preventing allocation interruptions caused by fund depletion and supporting continuous allocation operations. Fig. 10Dynamic fund balance at transfer stations (pre-disaster plan vs. dynamic adjustments in post-disaster samples). Dynamic fund balance at transfer stations (pre-disaster plan vs. dynamic adjustments in post-disaster samples). Figure 11 shows that the inventory surplus for each resource type at each transfer station remains non-negative in most samples, indicating that constraint (Eq. 2) is generally satisfied. Inventory levels exhibit time-varying fluctuations, with higher surpluses on Day 1 to hedge against initial uncertainty. As demand information becomes clearer and procurement and transportation decisions stabilize, inventory levels gradually decrease in later periods while remaining within safe ranges. This pattern reflects coordinated procurement and transportation decisions under evolving uncertainty. Fig. 11Dynamic emergency resource surplus at transfer stations (pre-disaster plan vs. dynamic adjustments in post-disaster samples). Dynamic emergency resource surplus at transfer stations (pre-disaster plan vs. dynamic adjustments in post-disaster samples). According to Fig. 12, although the original plan targets resource surplus at disaster areas, resource shortages or local surpluses occur at both DA#1 and DA#2 during the initial phase in test samples due to high information uncertainty. With information feedback and dynamic adjustments in transportation decisions, resource surplus at each disaster area converges toward equilibrium (zero) from Day 2 onward, and fluctuation ranges narrow over time. This suggests improved supply–demand matching and reduces the occurrence of shortages and surplus resources at disaster areas. Fig. 12Dynamic resource supply-demand balance at disaster areas (pre-disaster plan vs. dynamic adjustments in post-disaster samples). Dynamic resource supply-demand balance at disaster areas (pre-disaster plan vs. dynamic adjustments in post-disaster samples). (4) Evaluation of comprehensive optimization effectiveness. (4) Evaluation of comprehensive optimization effectiveness. Evaluation of comprehensive optimization effectiveness. . As shown in Fig. 13, the comprehensive objective values obtained from simulation samples are generally lower than the robustness value calculated by the allocation model. Economic costs observed in simulations are slightly higher than those computed by the model, primarily because top-level fund allocation decisions are sensitive to uncertainty in disaster areas. Meanwhile, simulated life-saving utilities and supply–demand gaps show better performance than the model’s planned results, benefiting from transportation strategies dynamically adjusted according to updated demand and transportation information. Overall, the allocation model achieves stable performance with respect to the comprehensive objective. Fig. 13Optimization objectives (model calculation vs. test samples). Note: The “model computation” refers to the worst-case robust objective result by solving the optimization problem under the uncertainty sets. In contrast, the test samples correspond to the realized results through Monte Carlo simulations, where uncertainties are sampled. Optimization objectives (model calculation vs. test samples). Note: The “model computation” refers to the worst-case robust objective result by solving the optimization problem under the uncertainty sets. In contrast, the test samples correspond to the realized results through Monte Carlo simulations, where uncertainties are sampled. (5)Discussion on practical applications. (5)Discussion on practical applications. Discussion on practical applications. From a practical perspective, the proposed framework is intended to support strategic and tactical decision-making in post-disaster emergency resource allocation rather than real-time operational dispatch. Decision makers can use the model to obtain coordinated allocation plans that account for uncertainty propagation and intertemporal dependencies among funding, procurement, and transportation decisions. The results in this section show how feedback-based adjustments contribute to improved supply–demand matching as information becomes progressively available. At the same time, practitioners should note that the performance of such approaches depends on the quality and timeliness of uncertainty information updates, as well as the appropriate selection of planning horizons and model parameters. Sensitivity analysis of key parameters This section examines the effect of two core parameters, the number of time periods (T) and the decision dependency depth (Δt), on the model’s overall performance. Impact analysis of the number of time periods (T). Impact analysis of the number of time periods (T). Impact analysis of the number of time periods (T). Based on the case setup in “Case study description” (Δt = 1), we analyze the effect of varying time periods (T = 3, 4, 6, 9, 12) on model performance metrics. As shown in Fig. 14a, the total number of variables and constraints increases with T, indicating a rise in model complexity. This occurs because each additional time period introduces more decision variables and associated constraints. Figure 14b shows that solution time increases with T, reflecting the higher model complexity. Models with T ≤ 9 are solvable within the time limit. However, for T = 12, the increased scale and computational complexity prevent convergence to the optimal solution within the preset time (denoted as “Timeout”). Finer time granularity (larger T) captures more detailed dynamics but increases computational complexity. Therefore, practical deployment requires balancing response speed with available computational resources. Figure 14c shows that the sample mean and standard deviation of the comprehensive objective f decrease significantly between T = 3 and 6, indicating that more frequent decision adjustments improve scheduling effectiveness. As T increases to 9 and 12, the improvement in mean f diminishes, and its standard deviation stabilizes. For T = 12, both slightly increase due to the non-optimal solution. Figure 14d shows that the proportion of samples violating core constraints (Eqs. 1 and 2) is highest at T = 3. As T increases, the violation rate decreases, albeit at a diminishing rate. Increasing decision adjustment frequency allows for more timely responses to uncertainty, reducing risks like resource shortages and fund depletion, and improving solution feasibility. Fig. 14Comprehensive performance metrics under different numbers of time periods (T). Comprehensive performance metrics under different numbers of time periods (T). Therefore, there is a trade-off between benefits and costs for the number of periods T. T = 6 (one decision period every 12 h) represents the “performance inflection point” for this case study, offering a good balance between optimization, stability, and solving efficiency (within 10 s) while reducing violation risk. Finer partitioning (T > 6) yields diminishing returns, offset by increased computational costs and potential solution failure, making it potentially counterproductive. The timeout (≥ 5 min) observed at T = 12 in Fig. 14b reflects the computational limits of the current MSARO formulation, not its impracticality for real-world applications. In many real-world emergency planning contexts, moderate planning horizons (e.g., T ≤ 6) are commonly adopted, within which the proposed model remains computationally tractable and yields performance improvements. (2)Impact analysis of decision dependency depth (\documentclass[12pt]{minimal} \usepackage{amsmath} \usepackage{wasysym} \usepackage{amsfonts} \usepackage{amssymb} \usepackage{amsbsy} \usepackage{mathrsfs} \usepackage{upgreek} \setlength{\oddsidemargin}{-69pt} \begin{document}$$\:{\Delta\:}t$$\end{document}). (2)Impact analysis of decision dependency depth (\documentclass[12pt]{minimal} \usepackage{amsmath} \usepackage{wasysym} \usepackage{amsfonts} \usepackage{amssymb} \usepackage{amsbsy} \usepackage{mathrsfs} \usepackage{upgreek} \setlength{\oddsidemargin}{-69pt} \begin{document}$$\:{\Delta\:}t$$\end{document}). Impact analysis of decision dependency depth (\documentclass[12pt]{minimal} \usepackage{amsmath} \usepackage{wasysym} \usepackage{amsfonts} \usepackage{amssymb} \usepackage{amsbsy} \usepackage{mathrsfs} \usepackage{upgreek} \setlength{\oddsidemargin}{-69pt} \begin{document}$$\:{\Delta\:}t$$\end{document}). The parameter Δt controls the historical uncertainty depth used in dynamic adjustment rules (Eq. 22). Based on the case setup with T = 6, we set Δt to 1, 2, 3, 4, and 5, and evaluate the resulting changes in model performance. As shown in Fig. 15a, the number of model variables increases with Δt, while the constraint count remains unchanged. Increasing Δt introduces more affine slope variables and non-zero terms in the original constraints. Figure 15b shows that solution time increases with Δt, but at a lower rate compared to increasing T. Increasing Δt increases model scale and solving time, but places less strain on computational resources than increasing T. Figure 15c shows that the sample mean and standard deviation of the comprehensive objective f decrease with increasing Δt, though with diminishing returns. This suggests that incorporating more historical information improves decision stability and optimization. Longer historical windows help the model capture uncertainty propagation patterns, such as persistent demand deviations and route impacts. However, the benefit of excessive historical dependency decreases as distant information becomes less relevant or introduces noise. Figure 15d shows that the sample violation rate decreases with increasing Δt. Leveraging more historical uncertainty information allows the model to better anticipate risks (e.g., procurement difficulties or route issues) and adjust current-period decisions, such as increasing budgets or rerouting transportation. This reduces the likelihood of violating hard constraints. Increasing the historical dependency depth can improve the robustness and safety of dynamic decisions, supporting the effective execution of resource allocation. Fig. 15Comprehensive performance metrics under different decision dependency depths (Δt). Comprehensive performance metrics under different decision dependency depths (Δt). The values of T and Δt in this section are not universally optimal, but are case-dependent, reflecting the specific scale, uncertainty structure, and planning assumptions of the studied scenario. The purpose of this sensitivity analysis is to highlight the trade-offs between decision adaptability, solution quality, and computational burden, rather than prescribe fixed parameter values for all disaster response contexts. Comparative advantages of the proposed model This section compares the proposed multi-role integrated dynamic allocation model with three widely used methods in emergency resource allocation, under the same simulation conditions. These methods reflect different decision-making paradigms—deterministic myopic planning, static worst-case robustness, and adaptive rolling-horizon decision-making—allowing a focused comparison of the value added by explicitly modeling multi-stage decision dependencies and uncertainty feedback. Specifically: (1) Deterministic Model (DM) ignores uncertainties (as described in “Multi-role integrated emergency resource allocation model”), performing a single optimization based on pre-disaster parameter estimates to generate a fixed allocation scheme. (2) Static Robust Optimization (SRO) uses single-stage robust optimization (Δt = 0), generating a conservative scheme for the worst-case scenario within the uncertainty set (Eq. 17) for the entire rescue period. (3) Rolling Horizon Optimization (RHO) is a widely used dynamic optimization approach. At the start of each period t, it solves an optimization problem based on the latest information and a limited look-ahead horizon, making decisions for the current period only. Other methods, such as stochastic programming, scenario-based robust optimization, and machine learning approaches, may outperform the baselines under certain conditions. However, they typically involve additional assumptions and computational challenges in multi-stage settings, and are beyond the scope of this comparison. Using the case setup (T = 6) from “Case study description”, we simulated the comparative methods to calculate key metrics, as shown in Fig. 16. In terms of model scale and computational efficiency (Fig. 16a,b), the proposed model and SRO have a larger scale than DM and RHO, resulting in longer solution times. This is due to the introduction of auxiliary variables (\documentclass[12pt]{minimal} \usepackage{amsmath} \usepackage{wasysym} \usepackage{amsfonts} \usepackage{amssymb} \usepackage{amsbsy} \usepackage{mathrsfs} \usepackage{upgreek} \setlength{\oddsidemargin}{-69pt} \begin{document}$$\:{\boldsymbol{\beta\:}}_{t},{\boldsymbol{\alpha\:}}_{1},{\boldsymbol{\alpha\:}}_{t,\tau\:}$$\end{document}) and constraints for robust dual transformation. MSARO also introduces slope variables (\documentclass[12pt]{minimal} \usepackage{amsmath} \usepackage{wasysym} \usepackage{amsfonts} \usepackage{amssymb} \usepackage{amsbsy} \usepackage{mathrsfs} \usepackage{upgreek} \setlength{\oddsidemargin}{-69pt} \begin{document}$$\:{\boldsymbol{Y}}_{t}^{1}$$\end{document}) for affine decision rules, which is necessary for multi-period dynamic adaptability and robust feasibility. DM, a simple linear program, is solved the fastest. The computation time reported for the RHO approach in Fig. 16b represents the average solution time for a single decision period. Since the RHO framework resolves the optimization problem at each stage, the total computational burden over the entire planning horizon accumulates with the number of decision periods. Each individual RHO solved within a rolling window has a scale comparable to DM, but the cumulative computational burden across multiple optimization runs is significantly higher than DM over the entire rescue horizon. Fig. 16Performance metrics comparison under different scheduling models. Performance metrics comparison under different scheduling models. Regarding optimization performance (Fig. 16c), DM theoretically yields optimal allocation under its deterministic assumptions but suffers severe deviations in supply-demand gaps due to ignoring perturbations. This results in the highest (worst) comprehensive objective value f, which is not ideal for affected populations. SRO sacrifices optimality for worst-case feasibility, resulting in conservative solutions. RHO adapts to uncertainty through periodic re-optimization but suffers from myopic decision-making, often converging to local optima and limiting optimization potential. In contrast, the proposed model dynamically adjusts decisions based on historical uncertainty information, enabling both reactive adaptation and proactive mitigation of future impacts. It coordinates fund, procurement, and transportation flows across the entire rescue period, considering interdependencies between allocation strategies and the propagation of uncertainty during optimization. This ensures the generated scheme is near-optimal from a multi-period perspective, improving optimality (~ 50%, ~ 44%, and ~ 26% relative to DM, SRO, and RHO) and stability of the comprehensive objective. Table 6 further reports component-wise optimization performance comparisons. The proposed method achieves the smallest supply-demand gap and economic cost while maintaining the highest life-saving utility among all compared methods. This indicates the improvement is mainly due to the reduction in resource mismatch, while utility improvement remains stable compared to DM and better than SRO and RHO. In terms of constraint reliability (Fig. 16d), DM, ignoring uncertainty, exhibits the highest violation rate (~ 99%), suggesting it frequently leads to fund chain breaks or resource depletion, making the solution impractical. SRO, designed for feasibility within its uncertainty set, achieves a low violation rate (~ 3%) but sacrifices optimization performance. RHO’s violation rate (~ 35%) is higher than SRO’s, and its key weakness is insufficient cross-period coordination: decisions based on current information may lead to violations in subsequent periods due to unanticipated uncertainties. Conversely, the proposed model shows high decision reliability with the lowest violation rate (0.8%). It is constructed based on the predefined uncertainty set and models the dependence of decision variables on historical uncertainties. This ensures current-period decisions satisfy constraints while providing flexibility to handle future uncertainties, avoiding intertemporal violation risks in RHO. In summary, the SRO approach performs relatively well only in terms of constraint reliability. In comparison, the proposed model achieves comparable performance in terms of model scale and solution time, while yielding substantially lower values in both the constraint violation rate and the comprehensive optimization objective (mean ± standard deviation). These improvements indicate clear advantages in decision safety, economic cost, and solution stability. Compared with DM and RHO, the proposed approach does not exhibit superior computational efficiency. However, under a fully acceptable solving time (on the order of several seconds), it consistently achieves markedly lower constraint violation rates and better comprehensive objective values, demonstrating a favorable trade-off between computational efficiency and decision quality. Table 6Component-wise comparison of different methods in the comprehensive optimization objective. Note: Bold text indicates the performance improvement of this study compared to other methods.MethodsEconomic cost (104 ¥)Life-saving utilityResource supply-demand gapComprehensive objective (f)DM71161,917,38775,01274,993 (↓49.61%)SRO24361,035,24368,32968,319 (↓44.69%)RHO22611,451,52350,80250,787 (↓25.60%)Proposed19391,934,85737,80637,787 Table 6Component-wise comparison of different methods in the comprehensive optimization objective. Note: Bold text indicates the performance improvement of this study compared to other methods.MethodsEconomic cost (104 ¥)Life-saving utilityResource supply-demand gapComprehensive objective (f)DM71161,917,38775,01274,993 (↓49.61%)SRO24361,035,24368,32968,319 (↓44.69%)RHO22611,451,52350,80250,787 (↓25.60%)Proposed19391,934,85737,80637,787 Component-wise comparison of different methods in the comprehensive optimization objective. Note: Bold text indicates the performance improvement of this study compared to other methods. Conclusions Efficient and integrated emergency resource allocation during the post-disaster “golden 72-hour window” is crucial for mitigating disaster losses. Addressing two key gaps in existing research—lack of multi-role coordination mechanisms and limited application of MSARO—this paper proposes an allocation method integrating a “Decision-Execution-Demand” linkage mechanism. This method combines pre-disaster planning with post-disaster feedback adjustments to optimize economic costs, rescue utility, and supply-demand gaps, offering a strategy for complex disaster resource allocation. In the earthquake scenario with identical modeling assumptions, the results demonstrate: Effectiveness of the allocation strategy: In this case, the integrated model enables feedback among fund allocation (decision role), resource procurement (execution role), and transportation (execution role → demand role). Under multiple uncertainty disturbances, the model helps maintain the security of the fund chain and resource inventory buffers through multi-period feedback adjustments. It improves rescue efficacy and resource supply-demand balance while controlling economic costs.Impact of key parameters on model performance: The number of time periods (T) and the decision dependency depth (Δt) influence model performance. A benefit-cost equilibrium point is reached at T = 6: the optimization objective decreases by ~ 5%, the violation rate drops to 1%, and solution time is under 10 s. Finer time granularity (T > 9) increases computational burden with diminishing optimization gains. Extending Δt can also improve optimization and decision safety at low computational cost. These results provide insights into parameter selection for this case, rather than universally optimal settings.Comprehensive performance improvement: The proposed model improved the optimization objective (mean ± standard deviation under the same 1,000 simulative test samples) by ~ 50% (vs. DM) with solving times of only a few seconds. This performance improves economic cost, rescue utility, and supply-demand balance. Its violation rate (0.8%) is lower than DM and RHO by 98% and 34%, respectively, indicating improved decision reliability. The model overcomes the vulnerabilities of DM, conservatism of SRO, and local optimality issues of RHO. Effectiveness of the allocation strategy: In this case, the integrated model enables feedback among fund allocation (decision role), resource procurement (execution role), and transportation (execution role → demand role). Under multiple uncertainty disturbances, the model helps maintain the security of the fund chain and resource inventory buffers through multi-period feedback adjustments. It improves rescue efficacy and resource supply-demand balance while controlling economic costs. Effectiveness of the allocation strategy: In this case, the integrated model enables feedback among fund allocation (decision role), resource procurement (execution role), and transportation (execution role → demand role). Under multiple uncertainty disturbances, the model helps maintain the security of the fund chain and resource inventory buffers through multi-period feedback adjustments. It improves rescue efficacy and resource supply-demand balance while controlling economic costs. Impact of key parameters on model performance: The number of time periods (T) and the decision dependency depth (Δt) influence model performance. A benefit-cost equilibrium point is reached at T = 6: the optimization objective decreases by ~ 5%, the violation rate drops to 1%, and solution time is under 10 s. Finer time granularity (T > 9) increases computational burden with diminishing optimization gains. Extending Δt can also improve optimization and decision safety at low computational cost. These results provide insights into parameter selection for this case, rather than universally optimal settings. Impact of key parameters on model performance: The number of time periods (T) and the decision dependency depth (Δt) influence model performance. A benefit-cost equilibrium point is reached at T = 6: the optimization objective decreases by ~ 5%, the violation rate drops to 1%, and solution time is under 10 s. Finer time granularity (T > 9) increases computational burden with diminishing optimization gains. Extending Δt can also improve optimization and decision safety at low computational cost. These results provide insights into parameter selection for this case, rather than universally optimal settings. Comprehensive performance improvement: The proposed model improved the optimization objective (mean ± standard deviation under the same 1,000 simulative test samples) by ~ 50% (vs. DM) with solving times of only a few seconds. This performance improves economic cost, rescue utility, and supply-demand balance. Its violation rate (0.8%) is lower than DM and RHO by 98% and 34%, respectively, indicating improved decision reliability. The model overcomes the vulnerabilities of DM, conservatism of SRO, and local optimality issues of RHO. Comprehensive performance improvement: The proposed model improved the optimization objective (mean ± standard deviation under the same 1,000 simulative test samples) by ~ 50% (vs. DM) with solving times of only a few seconds. This performance improves economic cost, rescue utility, and supply-demand balance. Its violation rate (0.8%) is lower than DM and RHO by 98% and 34%, respectively, indicating improved decision reliability. The model overcomes the vulnerabilities of DM, conservatism of SRO, and local optimality issues of RHO. This study presents a theoretical model for integrated and adaptive emergency resource allocation. However, the model relies on a centralized optimization framework with idealized assumptions, such as perfect information sharing and fully aligned stakeholder objectives. Real-world challenges—including political constraints, information asymmetry, coordination frictions, and conflicting incentives—are not captured. The use of affine decision rules and box uncertainty sets simplifies the model for tractability, but does not capture the full complexity of real-world uncertainty. The empirical analysis is based on a simplified earthquake case study, does not verify generalizability to other hazard types, and abstracts from several operational factors, including network congestion, capacity limits, and multi-modal transportation structures. Future work should: (1) develop decentralized models to account for information asymmetry and individual incentives; (2) incorporate institutional and behavioral constraints; (3) adopt data-driven uncertainty representations and more flexible decision rules; and (4) extend the framework to multi-hazard and recovery-phase scenarios. While this study focuses on an earthquake scenario, the model can be adapted to other sudden-onset crises, such as public health emergencies, with appropriate modifications and consideration of additional operational factors.