Temporally Split Benders Decomposition (TSBD)

• TSBD is a decomposition method that partitions long-horizon operational recourse into temporal subproblems while retaining investment decisions in a master problem.
• It ensures high temporal fidelity and exact enforcement of intertemporal policy constraints like CO₂ caps and storage dynamics.
• By leveraging budgeting and linking variables along with grouped cuts, TSBD improves computational tractability in large-scale energy planning.

Temporally Split Benders Decomposition (TSBD) is a class of Benders-based decomposition methods for large-scale energy system capacity expansion planning in which long-horizon operational decisions are split into multiple temporally indexed subproblems while long-term investment decisions remain in a master problem. In the literature represented by "A computationally efficient Benders decomposition for energy systems planning problems with detailed operations and time-coupling constraints" (Jacobson et al., 2023), "Accelerating Stochastic Energy System Optimization Models: Temporally Split Benders Decomposition" (Sasanpour et al., 14 Aug 2025), and "Clustering-enhanced adaptive Benders decomposition for energy systems planning optimization" (Law et al., 29 May 2026), TSBD is used to preserve chronological operational detail while improving tractability through temporal decomposition, parallel subproblem solution, and master-level handling of complicating intertemporal couplings such as policy budgets or storage boundary states.

1. Historical emergence and problem setting

TSBD emerged from the computational difficulty of solving high-resolution energy system capacity expansion models with detailed chronology, storage, transmission, and policy constraints. These models are typically formulated as LPs or MILPs, with first-stage investment decisions and second-stage operational decisions over hourly or sub-hourly time grids. The underlying challenge is that monolithic formulations can become extremely large: the stochastic German power system instances in (Sasanpour et al., 14 Aug 2025) reach up to 87 million constraints and variables, while the macro-energy planning models in (Jacobson et al., 2023) are reported at up to $9.7\times 10^9$ variables and $3.04\times 10^{10}$ constraints.

Classical Benders decomposition separates investments from operations, but it does not necessarily address the fact that the operational recourse itself may be chronologically large. In stochastic settings, standard L-shaped decomposition splits by scenario only, which is insufficient when the temporal cardinality $|T|$ is much larger than the scenario cardinality $|\Omega|$ (Sasanpour et al., 14 Aug 2025). In deterministic macro-energy planning, a full-year operational subproblem may remain too large even when investments are isolated in the master (Jacobson et al., 2023).

The defining feature of TSBD is therefore temporal subdivision of the recourse. In (Jacobson et al., 2023), the year is decomposed into weekly subperiods, and global policy couplings are replaced by master-level budgeting variables. In (Sasanpour et al., 14 Aug 2025), the recourse is partitioned into time blocks within each stochastic scenario, and storage continuity is preserved by additional linking variables for end-of-block storage states. In (Law et al., 29 May 2026), the same temporally split structure is treated as a baseline and then extended with clustering-based cut aggregation and representative-subproblem scheduling.

Paper	TSBD role	Key coupling mechanism
(Jacobson et al., 2023)	Core temporally split Benders idea	Budgeting variables $\bm{q}_w$
(Sasanpour et al., 14 Aug 2025)	Explicit TSBD formulation for stochastic CEP	Linking variables $\ell_{b,r,s,\omega}$
(Law et al., 29 May 2026)	Clustering-enhanced TSBD variants	Budgets $q_s$, grouped cuts, representative SPs

A plausible implication is that TSBD is best understood not as a single fixed formulation, but as a decomposition pattern: temporal disaggregation of operational recourse combined with master-level variables that restore exactness for otherwise time-coupled structures.

2. Core decomposition architecture

Across the cited works, TSBD retains the standard Benders separation between a master problem (MP) and operational subproblems (SPs), but refines the recourse partition so that subproblems are indexed by time segments rather than only by scenario or by the entire year.

In the electricity-sector capacity expansion model of (Law et al., 29 May 2026), the full-space MILP is written as

$$\begin{aligned} \min_{y, \{q_s\}, \{x_{s,t}\} } \quad & c_I^\top y + \sum_{s \in \mathcal{S}} \sum_{t \in \mathcal{T}_s} c_{s,t}^\top x_{s,t} \ \text{s.t.} \quad & A_s x_s + B_s y \le d_s \quad && \forall s \in \mathcal{S} \ & \sum_{t \in \mathcal{T}_s} e_{s,t}^\top x_{s,t} \le q_s \quad && \forall s \in \mathcal{S} \ & \sum_{s \in \mathcal{S}} q_s = \bar{Q} \quad && \text{(annual CO}_2\text{ cap, if present)} \ & R y \le r \ & x_{s,t} \ge 0,\quad q_s \ge 0,\quad y \ge 0,\quad y_j \in \mathbb{Z}\;\; \forall j \in \mathcal{I}. \end{aligned}$$

The MP chooses investments $y$ and, when a global annual CO$_2$ cap is present, budgeting variables $3.04\times 10^{10}$0. Each subproblem $3.04\times 10^{10}$1 then optimizes operations for a time segment $3.04\times 10^{10}$2 conditional on the current MP decisions (Law et al., 29 May 2026).

The operational subproblem at iteration $3.04\times 10^{10}$3 is

$3.04\times 10^{10}$4

The corresponding classical multi-cut master approximation is

$3.04\times 10^{10}$5

This is the benchmark regularized multi-cut formulation in (Law et al., 29 May 2026).

The same architectural pattern appears in (Jacobson et al., 2023), where the subproblems are weekly, and in (Sasanpour et al., 14 Aug 2025), where they are indexed jointly by scenario and time block $3.04\times 10^{10}$6. The principal difference lies in how temporal couplings are represented in the MP: budgeting variables in one line of work, storage linking variables in the other.

3. Budgeting-variable TSBD for deterministic planning

The deterministic formulation in (Jacobson et al., 2023) addresses planning problems with time-coupling policy constraints such as Renewable Portfolio Standard (RPS) requirements and annual CO$3.04\times 10^{10}$7 caps. The monolithic compact MILP is

$3.04\times 10^{10}$8

The central device is the replacement of the global policy coupling

$3.04\times 10^{10}$9

with weekly budgets $|T|$0 satisfying

$|T|$1

The paper proves the equivalence

$|T|$2

(Jacobson et al., 2023). This equivalence is the mathematical basis for temporally splitting the operational problem while preserving exact enforcement of the annual policy constraint.

The resulting multi-cut TSBD master is

$|T|$3

Each weekly subproblem is an LP: $|T|$4

This formulation is designed for planning models with detailed operations and time-coupling constraints, including storage dynamics, ramping, transmission, and aggregated unit commitment within each week (Jacobson et al., 2023). The key point is that the policy coupling across weeks is moved into the MP, while intraweek chronology remains fully represented in each SP. The paper states that the method “decouples operational timesteps using budgeting variables in the master model,” and that this device permits modeling “relevant constraints coupling decisions across time periods (e.g. policy constraints) within a decomposed framework” (Jacobson et al., 2023).

A common misconception is that temporal splitting necessarily destroys exactness for year-level policy constraints. In this formulation, that is not the case: exactness follows from the equivalence theorem above (Jacobson et al., 2023). Another potential misconception is that all intertemporal structures can be handled equally well by such budgets. The same source explicitly limits the exact result to couplings of the form $|T|$5, while noting that extension to other constraints “should be extensible” rather than established in the reported experiments (Jacobson et al., 2023).

4. Linking-variable TSBD for stochastic models with storage

The stochastic formulation in (Sasanpour et al., 14 Aug 2025) generalizes temporal splitting to two-stage stochastic capacity expansion planning with hourly resolution and storage. Here, classical scenario decomposition is insufficient because each scenario subproblem remains large, and storage links all hours within a year. TSBD resolves this by partitioning time into blocks and introducing additional master variables $|T|$6 to represent storage energy at block boundaries.

The base two-stage stochastic program is

$|T|$7

with scenario recourse $|T|$8 defined over hourly operations, transmission, unserved demand, and storage (Sasanpour et al., 14 Aug 2025).

Time is partitioned into blocks $|T|$9, and the MP is augmented with storage boundary variables

$|\Omega|$0

Within each block, storage obeys the same intertemporal law, but block boundaries are enforced through

$|\Omega|$1

If periodic operation is desired, then $|\Omega|$2 (Sasanpour et al., 14 Aug 2025).

For each $|\Omega|$3, the LP subproblem produces block-level subgradients with respect to investment capacities and boundary storage values. The TSBD master is

$|\Omega|$4

subject to capacity bounds, $|\Omega|$5, and block-scenario cuts

$|\Omega|$6

(Sasanpour et al., 14 Aug 2025).

This formulation is exact for the storage coupling considered. The source explicitly states that the original two-stage recourse with storage dynamics over $|\Omega|$7 is equivalent to a block-based program if end-of-block storage states are elevated to first-stage linking variables $|\Omega|$8 and enforced at block boundaries (Sasanpour et al., 14 Aug 2025). That distinguishes TSBD from naive time decomposition, which would typically break storage feasibility.

Feasibility is maintained by storage slack variables $|\Omega|$9, penalized in the recourse and inserted into the storage balance equation. As a result, only optimality cuts are needed (Sasanpour et al., 14 Aug 2025). The paper further combines TSBD with a level bundle stabilization method and inactive-cut deactivation, stating that these do not affect optimality because the unstabilized MP provides the true lower bound and inactive cuts can be regenerated later (Sasanpour et al., 14 Aug 2025).

5. Cut generation, stabilization, and clustering-enhanced variants

The standard TSBD cut is a supporting hyperplane to a convex recourse function. In deterministic week-indexed form this appears as

$\bm{q}_w$0

(Jacobson et al., 2023). In stochastic block-scenario form it appears as

$\bm{q}_w$1

(Sasanpour et al., 14 Aug 2025). In the electricity-sector TSBD benchmark of (Law et al., 29 May 2026), the per-segment cut is

$\bm{q}_w$2

The 2026 paper (Law et al., 29 May 2026) focuses on a specific TSBD weakness: temporal splitting increases the number of subproblems and thus the number of cuts added to the MP. The resulting cut accumulation can make the MP the computational bottleneck. To mitigate this, the paper develops clustering-enhanced grouped cuts and representative-subproblem scheduling.

For grouped cuts, the generic cluster-level form is

$\bm{q}_w$3

where

$\bm{q}_w$4

and the coefficients on $\bm{q}_w$5 are stacked from the corresponding $\bm{q}_w$6 (Law et al., 29 May 2026).

Three grouped-cut variants are distinguished.

First, fixed-grouping shared-recourse (fix-G-S) clusters SPs prior to the Benders iterations using exogenous time-series features such as normalized hourly load and wind/solar profiles, with Euclidean similarity in feature space and k-means clustering (Law et al., 29 May 2026).

Second, adaptive shared-recourse (adapt-G-S) clusters SPs using the current dual vectors $\bm{q}_w$7, typically at regrouping iterations. Historical grouped cuts are then reconstructed under the new grouping by re-aggregating stored per-SP coefficients $\bm{q}_w$8 for all $\bm{q}_w$9. The paper states that this guarantees mathematical validity of historical cuts after regrouping (Law et al., 29 May 2026).

Third, adaptive individual-recourse (adapt-G-I) retains $\ell_{b,r,s,\omega}$0 per subproblem but constrains grouped sums: $\ell_{b,r,s,\omega}$1 Because $\ell_{b,r,s,\omega}$2 remains indexed to SPs, historical grouped cuts remain valid without reconstruction (Law et al., 29 May 2026).

The same paper also uses regularization via an interior-point level-set method after each MP solve if not converged. The level-set MP is

$\ell_{b,r,s,\omega}$3

with $\ell_{b,r,s,\omega}$4 and $\ell_{b,r,s,\omega}$5 in the LP phase (Law et al., 29 May 2026).

These enhancements do not redefine TSBD; rather, they alter how much information is passed from temporally split SPs to the MP and how often all SPs must be solved.

6. Representative subproblems, computational behavior, and design rules

When the SP side dominates runtime because the number of temporal segments exceeds available CPUs, (Law et al., 29 May 2026) introduces a representative-subproblem strategy. The method alternates between full-SP iterations and representative-SP iterations. After a warm start with all SPs solved for $\ell_{b,r,s,\omega}$6 iterations, the algorithm clusters SPs by dual vectors, selects one representative $\ell_{b,r,s,\omega}$7 per group as the SP whose dual vector is nearest to the cluster centroid, and then solves only those representatives between regroupings (Law et al., 29 May 2026). On representative iterations, the lower bound is updated but the upper bound is held at the most recent full-SP value; if the gap falls below tolerance, a full-SP iteration is triggered to certify convergence (Law et al., 29 May 2026).

The computational behavior of TSBD depends strongly on whether runtime is concentrated in the MP or in the SPs, on the strength of inter-SP coupling, and on the available parallelization.

In (Jacobson et al., 2023), runtime grows linearly with the number of weeks, and TSBD consistently outperforms monolithic solves on MILPs with more than 2 weeks. The paper reports that 29 out of 70 monolithic MILP cases were intractable, while TSBD solved all within the time limit. In a 6-zone LP CO$\ell_{b,r,s,\omega}$8 case with 22 weeks, TSBD converges in 132 iterations at approximately 12 s per iteration, whereas full-operation Benders requires 630 iterations at approximately 241 s per iteration and becomes intractable at higher resolutions (Jacobson et al., 2023).

In (Sasanpour et al., 14 Aug 2025), temporal splitting into 8 blocks delivered up to approximately 60% total runtime reduction versus Benders without time splitting. With distributed memory via MPI, persistent subproblem instances through GAMS model instances (GMI), and TSBD, the total computing time reduction exceeded 80% relative to the baseline. The paper also reports reduced memory requirements and notes that too many blocks, such as 12, can make the stabilized MP dominant due to a surge in cuts, while 8 blocks was often best (Sasanpour et al., 14 Aug 2025).

In (Law et al., 29 May 2026), the computational picture is more conditional because the goal is not simply to validate TSBD but to choose among TSBD variants. The paper evaluates 11-, 20-, and 26-zone Eastern U.S. systems over a 52-week horizon using 48-hour SPs with $\ell_{b,r,s,\omega}$9 and 168-hour SPs with $q_s$0, under no CO$q_s$1 policy, a CO$q_s$2 price of $q_s$3, and a hard annual cap. For the 48-hour setup, the MP share of runtime in the multi-cut benchmark grows from 21.3% in the 11-zone case to 36.8% in the 20-zone case and 54.8% in the 26-zone case, motivating MP cut reduction (Law et al., 29 May 2026).

Under weak coupling with no CO$q_s$4 policy, adaptive grouped cuts outperform fixed grouping. For 48-hour SPs, runtime falls from 277 s to 192 s in the 11-zone case, from 1322 s to 753 s in the 20-zone case, and from 4006 s to 1891 s in the 26-zone case, corresponding to 31%, 43%, and 53% speedups for adapt-G-S with 32 groups (Law et al., 29 May 2026). For the 20-zone, 48-hour setup, adapt-G-I with 64 groups reduces average MP time from 9.9 s to 1.9 s, while the number of iterations drops from 49 to 43; however, very coarse grouping at 16 groups markedly increases iterations and renders some cases intractable (Law et al., 29 May 2026).

Under stronger coupling, the advantage of cut aggregation declines. With a CO$q_s$5 price, adaptive grouped cuts generally do not outperform multi-cut, apart from an 11-zone 48-hour case with 12% improvement. Under a hard annual CO$q_s$6 cap, runtimes increase substantially across all methods, adaptive grouped cuts show no advantage, and many low-group-count cases become intractable; in the 26-zone cap cases, all methods, including multi-cut, are intractable within the time limit (Law et al., 29 May 2026).

The representative-SP method is strongest when SP solution time dominates and CPUs are scarce. In a single-weather-year 20-zone 48-hour study, limiting cores to 61, 16, and 8 corresponds to SPs-to-CPU ratios of approximately 3, 11, and 23, and representative SPs increasingly outperform multi-cut as the ratio rises (Law et al., 29 May 2026). In the stochastic 3-weather-year case with 546 SPs and only 8 CPUs feasible due to memory, the best configurations use 16 to 32 representative SPs, and representative SPs remain feasible under CO$q_s$7 cap cases where adapt-G-I becomes intractable (Law et al., 29 May 2026).

The practical guidance given across the three papers is internally consistent. Moderate temporal partitioning is preferred: 4 to 8 blocks in (Sasanpour et al., 14 Aug 2025), and intermediate cluster counts such as 32 to 64 groups for 48-hour SPs in (Law et al., 29 May 2026). Adaptive grouping based on normalized dual vectors is preferred over fixed input-based grouping when grouped cuts are used (Law et al., 29 May 2026). Multi-cut TSBD is especially effective when rich cut information per iteration offsets MP growth, but if temporal splitting creates too many cuts, MP management becomes decisive (Sasanpour et al., 14 Aug 2025, Law et al., 29 May 2026). A plausible implication is that TSBD design should be driven by a bottleneck diagnosis: whether the main limitation is recourse size, MP cut accumulation, or insufficient parallelism.

7. Theoretical properties, scope, and relation to adjacent methods

The theoretical status of TSBD in the cited literature is exact for the modeled couplings and convex recourse structures considered. In (Jacobson et al., 2023), exactness follows from the theorem showing equivalence between the original aggregated policy constraint and the budgeting-variable reformulation. In (Sasanpour et al., 14 Aug 2025), exactness follows from the compact reformulation that elevates end-of-block storage states to master-level linking variables, producing a convex LP recourse function in $q_s$8 whose cuts are valid supporting hyperplanes. In (Law et al., 29 May 2026), grouped cuts are proved to remain valid outer approximations, and adaptive regrouping preserves validity either through cut reconstruction or through historical group constraints with fixed $q_s$9 indexing.

All three works rely on LP subproblems for the cleanest convergence properties. In (Sasanpour et al., 14 Aug 2025), TSBD inherits Benders’ convergence guarantees with convex LP subproblems and exact optimality cuts. In (Jacobson et al., 2023), the master may contain integer investments while subproblems remain LPs, so the method follows standard Benders for MILPs with LP subproblems. In (Law et al., 29 May 2026), the empirical finding is that adaptive grouped cuts converge to the same optimal objective as multi-cut in tractable cases, while representative-SP iterations only approximate the full iteration and require a full-SP iteration to certify convergence (Law et al., 29 May 2026).

TSBD is distinct from several adjacent decomposition paradigms. Relative to standard Benders or the classical L-shaped method, its novelty is temporal recourse splitting rather than scenario-only decomposition (Jacobson et al., 2023, Sasanpour et al., 14 Aug 2025). Relative to Lagrangian temporal decomposition, TSBD retains exact primal feasibility through master-level coupling variables rather than relying on dualized linking constraints and subgradient coordination (Sasanpour et al., 14 Aug 2025). Relative to progressive hedging, it is cut-based rather than penalty-based and is presented as exact for the convex two-stage structures under study (Sasanpour et al., 14 Aug 2025). Relative to SDDP, it targets two-stage invest/operate models with long hourly resolution rather than multistage state-transition formulations (Sasanpour et al., 14 Aug 2025).

A recurrent misconception is that temporal decomposition in planning models is necessarily heuristic, comparable to rolling-horizon methods. The sources reject that characterization for the modeled structures: TSBD preserves exactness for annual policy couplings through budgets (Jacobson et al., 2023) and for long-term storage operation through linking variables $$\begin{aligned} \min_{y, \{q_s\}, \{x_{s,t}\} } \quad & c_I^\top y + \sum_{s \in \mathcal{S}} \sum_{t \in \mathcal{T}_s} c_{s,t}^\top x_{s,t} \ \text{s.t.} \quad & A_s x_s + B_s y \le d_s \quad && \forall s \in \mathcal{S} \ & \sum_{t \in \mathcal{T}_s} e_{s,t}^\top x_{s,t} \le q_s \quad && \forall s \in \mathcal{S} \ & \sum_{s \in \mathcal{S}} q_s = \bar{Q} \quad && \text{(annual CO}_2\text{ cap, if present)} \ & R y \le r \ & x_{s,t} \ge 0,\quad q_s \ge 0,\quad y \ge 0,\quad y_j \in \mathbb{Z}\;\; \forall j \in \mathcal{I}. \end{aligned}$$0 (Sasanpour et al., 14 Aug 2025). Another misconception is that finer temporal splitting is always beneficial. The later papers show the opposite once the MP becomes the bottleneck: excessive block counts or overly coarse grouped cuts can degrade performance or cause intractability (Sasanpour et al., 14 Aug 2025, Law et al., 29 May 2026).

Taken together, these works position TSBD as a family of exact decomposition formulations for large-scale energy planning models in which chronology is retained rather than aggregated away. The method’s significance lies in enabling higher temporal fidelity, richer policy modeling, and large stochastic expansions within feasible runtime and memory budgets, while leaving open a clear systems-level trade-off: temporal splitting increases parallelism and recourse granularity, but it also increases the number of cuts and the burden on the master problem (Jacobson et al., 2023, Sasanpour et al., 14 Aug 2025, Law et al., 29 May 2026).