Temporally Split Benders Decomposition

• The paper demonstrates that TSBD reformulates a two-stage stochastic model by decomposing recourse problems across both scenarios and time blocks, preserving essential temporal couplings.
• It introduces compact linking and budgeting variables to manage storage and policy constraints, enabling effective parallelization and reduced memory usage.
• Empirical results in energy planning show that TSBD can significantly lower computing times while balancing the computational load between master and subproblems.

Searching arXiv for papers on temporally split Benders decomposition and closely related methods. Temporally Split Benders Decomposition (TSBD) denotes a Benders framework in which the operational or recourse side is decomposed not only across the usual scenario dimension, but also across time blocks, while the master problem carries the variables required to preserve cross-block consistency. In the formulation explicitly introduced under this name, TSBD is a variant of two-stage stochastic Benders decomposition for large energy system capacity expansion models, where the second-stage operational problems are decomposed by scenarios and time blocks, and inter-temporal coupling due to storage is represented through compact linking variables for storage levels at block boundaries in the master problem (Sasanpour et al., 14 Aug 2025). Closely related temporally decomposed Benders formulations use budgeting variables to decouple operational periods while retaining annual policy constraints such as renewable portfolio standards and CO$_2$ caps, and clustering-enhanced variants study how the resulting explosion in cuts and subproblems should be managed (Jacobson et al., 2023, Law et al., 29 May 2026).

1. Problem setting and motivation

TSBD arises from a specific computational asymmetry in large-scale stochastic capacity expansion planning. In the German power-system instances used to introduce the method, the temporal horizon has hourly resolution over a full year, $|T| = 8760$, while the number of stochastic scenarios is comparatively small; deterministic equivalent formulations reach up to $87$ million constraints, $89$ million variables, and $309$ million non-zeros (Sasanpour et al., 14 Aug 2025). Under these conditions, classical Benders decomposition already exploits scenario independence, but each scenario subproblem still contains the full temporal horizon and its explicit inter-temporal couplings, so the bottleneck shifts from the master problem to the large second-stage subproblems.

The temporal bottleneck is not limited to stochastic planning. In year-long energy planning models with detailed operations, investments couple all weeks through capacity decisions, and policy constraints couple all weeks through annual sums. A related decomposition therefore separates investments from operations and decouples operational timesteps using budgeting variables in the master model. That construction permits parallelization of operational subproblems while preserving relevant time-coupling constraints such as renewable portfolio standards and CO$_2$ caps, and its runtime scales linearly with temporal resolution (Jacobson et al., 2023).

A further motivation appears once the temporal split has been performed: the number of subproblems and cuts can become so large that the master problem becomes the dominant bottleneck. In a temporally split regularized multi-cut Benders formulation for electricity-sector capacity expansion models, each operational period $s \in \mathcal{S}$ has its own subproblem and recourse approximation, and the computational burden can shift between master and subproblems depending on period length, inter-period coupling strength, and available parallelization (Law et al., 29 May 2026). TSBD is therefore best viewed not merely as a decomposition by time, but as a family of master–subproblem reorganizations designed to expose temporal parallelism without discarding the state information that makes long-horizon operation meaningful.

2. Core mathematical structure

The canonical TSBD formulation is built on a two-stage stochastic capacity expansion problem of the form

$$total = \min \; exp + \sum_{\omega \in \Omega} \pi_\omega \, op_\omega,$$

where the first stage contains continuous investment decisions in converter, storage, and transmission capacities, and the second stage contains scenario-dependent operational recourse over the hourly horizon (Sasanpour et al., 14 Aug 2025). Classical Benders introduces a master objective

$$MP = \min \; exp + \sum_{\omega \in \Omega} \pi_\omega \, \theta_\omega,$$

with one operational subproblem per scenario. TSBD refines this by partitioning the yearly horizon into time blocks $tb \in TB$, using a mapping matrix $|T| = 8760$0 from timesteps to blocks and a selector $|T| = 8760$1 for the last hour of each block. The master objective becomes

$|T| = 8760$2

so the recourse approximation is indexed by scenario–block pairs rather than by scenarios alone (Sasanpour et al., 14 Aug 2025).

The corresponding optimality cuts are also lifted to the scenario–block level. For scenario $|T| = 8760$3, block $|T| = 8760$4, and iteration $|T| = 8760$5, TSBD uses a multi-cut of the form

$|T| = 8760$6

This exposes the central structural point: investment variables remain global, while block-boundary storage levels $|T| = 8760$7 become additional master variables that carry inter-temporal information across subproblems (Sasanpour et al., 14 Aug 2025).

A closely related temporal split appears in non-stochastic or period-indexed models with annual policy coupling. There the monolithic cross-period constraint

$|T| = 8760$8

is replaced by budgeting variables $|T| = 8760$9 satisfying

$87$0

The paper establishes the equivalence

$87$1

which turns each week $87$2 into an independent subproblem for fixed $87$3 and makes the temporal split exact for this class of coupling constraints (Jacobson et al., 2023). The same budgeting logic reappears in period-indexed electricity-sector formulations with per-period emissions budgets $87$4 and a global balance $87$5 (Law et al., 29 May 2026).

3. Mechanisms for preserving inter-temporal coupling

The principal technical challenge in any temporal split is that operational feasibility is usually defined by constraints whose semantics are inherently dynamic. In the stochastic TSBD formulation, the decisive example is storage. A naive time decomposition would sever the storage state trajectory and eliminate the possibility of optimizing multi-day or seasonal storage behavior. TSBD addresses this by introducing a master-level linking variable $87$6, interpreted as the fixed storage level at the last timestep of block $87$7. Each scenario–block subproblem then includes

$87$8

and

$87$9

with circular linkage from the first block to the last when a circular year is used (Sasanpour et al., 14 Aug 2025).

Inside each block, the storage dynamics remain explicit:

$89$0

where slack variables $89$1 are introduced to ensure subproblem feasibility even when block boundaries are tight (Sasanpour et al., 14 Aug 2025). The paper’s claim is therefore narrower and more precise than “time decomposition preserves storage”: what is preserved is long-term storage optimization through a compact reformulation of the storage level constraint into linking variables.

Budget-based temporal decompositions preserve a different kind of coupling. In year-long energy planning, renewable portfolio standards and annual CO$89$2 caps are not state equations but global resource constraints. The budgeting-variable reformulation transfers those couplings to the master, so each period solves a local problem subject to its allocated budget while the master enforces global consistency through $89$3 or $89$4 (Jacobson et al., 2023, Law et al., 29 May 2026). In this sense, TSBD-like formulations separate two categories of coupling: dynamic state propagation, represented by block-boundary variables such as $89$5, and cumulative cross-period budgets, represented by variables such as $89$6 or $89$7.

A common misconception is that any temporal partition immediately yields a valid TSBD. The literature is more restrictive. One paper states explicitly that TSBD assumes linear dynamics with storage levels as the main inter-temporal coupling and that other complex temporal couplings, including ramping, minimum up/down times, and multi-year constraints, must either be handled similarly or remain intra-block (Sasanpour et al., 14 Aug 2025). Another shows that the effectiveness of grouped or aggregated temporal cuts declines under strong inter-temporal coupling, exemplified by annual CO$89$8 emissions limits, where the benchmark multi-cut formulation performs best (Law et al., 29 May 2026). Temporal splitting is therefore a structural reformulation problem, not a generic partition heuristic.

4. Master–subproblem algorithm and parallel execution

The algorithmic realization of TSBD in stochastic planning combines a stabilized master problem, massive subproblem parallelization, and explicit cut management. Initialization begins from a deterministic capacity expansion problem that provides a starting point for $89$9. Each iteration then solves a stabilized master problem using a level bundle method with level-bundle weighting $309$0, cut inactivity horizon $309$1, cut activity threshold $309$2, and convergence tolerance $309$3. The stabilized master minimizes the squared deviation of regionally aggregated capacities from the current stability center, subject to a level-bundle constraint; an unstabilized linear master with the same cuts is also solved to recover the actual lower bound (Sasanpour et al., 14 Aug 2025).

The subproblem phase fixes $309$4, $309$5, $309$6, and all $309$7, then solves each $309$8 independently in parallel. TSBD is implemented with MPI and GAMS Model Instances, so each scenario–block subproblem can run on a separate distributed-memory process while reusing the generated model structure across iterations (Sasanpour et al., 14 Aug 2025). This organization is materially different from classical scenario-wise Benders: instead of $309$9 large subproblems, TSBD creates $_2$0 smaller subproblems, thereby increasing parallel workload and reducing the memory footprint per process.

The lower and upper bounds are updated in the standard Benders manner, but with scenario–block aggregation. After new cuts are added, the algorithm updates

• $_2$1,
• $_2$2,
• $_2$3, and terminates when $_2$4 (Sasanpour et al., 14 Aug 2025). Because the operational models include unmet-demand penalties, the classical stochastic subproblems are never infeasible, so only optimality cuts are required (Sasanpour et al., 14 Aug 2025). A comparable feasibility-preserving design appears in temporally split period-indexed models, where penalized slack variables are added to demand and emissions constraints so that subproblems remain feasible while still returning duals for cut generation (Law et al., 29 May 2026).

Related temporally split formulations without stochastic scenarios use the same communication pattern. In the weekly budgeting-variable decomposition, the master sends fixed investments $_2$5 and weekly budgets $_2$6 to all week-specific subproblems; each week returns its optimal value $_2$7 together with dual multipliers $_2$8 and $_2$9, and the master adds one cut per week per iteration (Jacobson et al., 2023). Algorithmically, TSBD and these budgeting-variable methods are thus aligned by a common principle: time-local optimization remains inside subproblems, while the master owns the low-dimensional variables that coordinate blocks.

5. Computational behavior, gains, and trade-offs

The empirical case for TSBD is strongest in problems where the temporal dimension dominates scenario count. In the German power-system study, temporally split Benders decomposition reduces computing times by up to $s \in \mathcal{S}$0 relative to the deterministic equivalent and reduces memory requirements; with additional enhancement strategies and distributed memory on high-performance computers, computing time improves by over $s \in \mathcal{S}$1 (Sasanpour et al., 14 Aug 2025). For a 13-region instance, adding MPI reduces computing time by $s \in \mathcal{S}$2, and adding GAMS Model Instances reduces total computing time by about $s \in \mathcal{S}$3 relative to basic Benders decomposition (Sasanpour et al., 14 Aug 2025). These gains are accompanied by a qualitative shift in bottlenecks: temporal splitting shrinks subproblem time, but the stabilized master can become dominant as the number of time blocks rises.

That trade-off is visible in the block-count experiments. For the largest 39-region case, the best performance is reported with $s \in \mathcal{S}$4 time blocks; $s \in \mathcal{S}$5 blocks already reduce subproblem times substantially, $s \in \mathcal{S}$6 blocks make overall iteration time lowest, but $s \in \mathcal{S}$7 blocks cause stabilized-master time to grow sharply and double total computing time (Sasanpour et al., 14 Aug 2025). The key observation is therefore not that “more blocks are better,” but that the optimal number of blocks is an interior compromise between subproblem shrinkage and master growth.

The related weekly budgeting-variable decomposition exhibits a different but compatible scaling law: runtime grows linearly with temporal resolution, while many monolithic problems become intractable beyond practical memory or wall-clock limits (Jacobson et al., 2023). In a 6-zone LP with $s \in \mathcal{S}$8 weeks and a CO$s \in \mathcal{S}$9 scenario, the temporally split Benders method reports runtime $$total = \min \; exp + \sum_{\omega \in \Omega} \pi_\omega \, op_\omega,$$0 s, $$total = \min \; exp + \sum_{\omega \in \Omega} \pi_\omega \, op_\omega,$$1 iterations, and average iteration time $$total = \min \; exp + \sum_{\omega \in \Omega} \pi_\omega \, op_\omega,$$2 s, while standard Benders with a full operational subproblem reports $$total = \min \; exp + \sum_{\omega \in \Omega} \pi_\omega \, op_\omega,$$3 s, $$total = \min \; exp + \sum_{\omega \in \Omega} \pi_\omega \, op_\omega,$$4 iterations, and average iteration time $$total = \min \; exp + \sum_{\omega \in \Omega} \pi_\omega \, op_\omega,$$5 s (Jacobson et al., 2023). The data support a specific computational interpretation: temporal splitting improves both iteration cost and cut informativeness when the alternative is a single full-year operational subproblem.

The 2026 clustering-enhanced study shows that TSBD performance is highly regime-dependent. Adaptive grouped cuts outperform fixed grouping and provide substantial benefits under weak inter-temporal coupling, especially in larger systems with shorter subproblem horizons where the master accounts for a greater share of runtime. Their effectiveness declines under strong inter-temporal coupling, such as annual CO$$total = \min \; exp + \sum_{\omega \in \Omega} \pi_\omega \, op_\omega,$$6 emissions limits, where the benchmark regularized multi-cut performs best. A representative-subproblem method, by contrast, outperforms the benchmark when parallelization is limited and subproblem solution dominates runtime (Law et al., 29 May 2026). TSBD is therefore not a uniformly dominant algorithmic choice; its preferred implementation depends on whether the computational burden lies in the master or the subproblems, and on how strongly time blocks are coupled by policy or state constraints.

6. Relation to other Benders variants and broader extensions

The literature distinguishes TSBD from several adjacent decomposition paradigms. Relative to classical Benders decomposition, TSBD adds a second decomposition axis, so the recourse is split both scenario-wise and time-wise. Relative to pure Lagrangian relaxation, it retains a structured master problem carrying capacities and block-boundary states rather than replacing them with a purely dual coordination mechanism. Relative to nested Benders, it does not impose a strict time hierarchy, but uses time-block boundaries and parallel subproblems instead (Sasanpour et al., 14 Aug 2025). These distinctions are structural, not terminological: they specify where coupling information resides and how it is transmitted through cuts.

A separate line of work studies how cut quality itself should be improved in decomposed master–subproblem loops. “Deepest” Benders cuts are defined through norm-based distance from the incumbent master point to the candidate separating hyperplane, and the paper develops the Guided Projections Algorithm, the Directed Depth-Maximizing Algorithm, and minimal infeasible subsystem interpretations for cut selection (Hosseini et al., 2021). That work does not introduce TSBD as a standalone method, but it explicitly formulates how deepest-cut separation can be applied per period or temporal segment, and how normalization functions can be aggregated across time. This suggests that temporally split schemes can be strengthened not only by changing the decomposition, but also by changing the geometry of the cuts added to the master (Hosseini et al., 2021).

Generalized Benders decompositions for MAP inference in factor graphs point in a different direction. There, original latent variables are treated as complicating variables, factor-specific clone variables are treated as non-complicating variables, and subproblem multipliers generate feasibility and optimality cuts. The paper then outlines how a temporal factor graph could be decomposed by time slices or adjacent-slice blocks, with time-indexed latent variables and temporal logical constraints in the master and per-slice factor interactions in the subproblems (Dubey et al., 2024). This is presented as a natural extension rather than as the main algorithm of the paper, but it implies that the core TSBD idea is not intrinsically tied to energy systems: what matters is a block-separable temporal structure together with a compact representation of the variables that carry information across blocks.

In that broader sense, temporally split Benders decomposition is best understood as a design pattern for block-structured optimization. Its defining components are a temporal partition of the recourse space, master-level variables that encode the residual cross-time coupling, and cut generation from blockwise subproblem duals. The research record so far shows that this pattern is especially effective when the temporal axis is much larger than the scenario axis, when coupling can be represented compactly through boundary states or budgets, and when parallel resources are abundant; it is less effective when block coupling is strong enough that master growth or cross-block coordination dominates the computational profile (Sasanpour et al., 14 Aug 2025, Law et al., 29 May 2026).