Dynamic Multi-Region Division Framework

• Dynamic multi-region division is a paradigm that allocates spatial units and tasks based on data-driven, flexible regional decomposition rather than fixed boundaries.
• It integrates region, context, and time specialization to dynamically tailor expert routing, as exemplified by the TriForecaster model in electric load forecasting.
• Empirical results demonstrate significant improvements, with up to a 22.4% reduction in forecasting errors across multiple datasets, highlighting its practical impact.

A dynamic multi-region division framework denotes a modeling and inference paradigm in which spatial units, expert capacity, or routing resources are allocated through data-dependent regional decomposition rather than a fixed partition. In the most explicit formulation among the supplied works, TriForecaster defines such a framework for Multi-Region Electric Load Forecasting (MRELF) by treating regions as tasks in a multi-task learning setting and dynamically specializing computation along region, context, and time dimensions (Zhu et al., 13 Aug 2025). Related literatures use the same idea in adjacent senses: latent expert partitions over fixed urban patches, nonparametric contiguous spatial regionalization from time series, multi-objective region generation for transportation services, adaptive region resizing in LEO satellite networks, and region-wise online map management for dynamic-object removal (Liu et al., 30 Jan 2026, Weng et al., 6 May 2026, Alkafri et al., 2023, Song et al., 14 Jul 2025, Yan et al., 2023).

1. Conceptual scope and domain-specific meanings

The term “region” is not uniform across the literature. In MRELF, a region may be a city, a subarea within a city, or a bus in a distribution network, and the central problem is short-term forecasting across multiple such spatial units under shared and region-specific structure (Zhu et al., 13 Aug 2025). In urban profiling, the basic regions are fixed $200\times 200$ Sentinel-2 patches at 10 m resolution, and the dynamic element lies not in re-drawing boundaries but in sparsely routing each region–task pair to a subset of experts (Liu et al., 30 Jan 2026). In spatial time-series regionalization, regions are contiguous connected subsets of a spatial graph, inferred jointly with representative driver time series under a minimum-description-length principle; the partition is static in space but dynamic in the sense that regions are defined by temporal signatures (Weng et al., 6 May 2026). In transportation service management, regions are generated by clustering atomic spatial elements under predictability, specificity, connectivity, and area constraints, and the resulting system is re-run in a daily $T+1$ pipeline rather than continuously online (Alkafri et al., 2023). In LEO networks, a region is a contiguous rectangular block of satellites whose size is adaptively adjusted according to workload status (Song et al., 14 Jul 2025). In RH-Map, regions are coarse 3D spatial cells in a two-layer hash map that support real-time region-wise dynamic-object removal (Yan et al., 2023).

This suggests that “dynamic multi-region division” is best understood as a family of mechanisms for allocating structure at region granularity while allowing the region-level organization, regional interactions, or region-specific pathways to depend on data, workload, or observed spatiotemporal state.

Setting	Region unit	Dynamic mechanism
TriForecaster	Cities, subareas, buses	Similarity-based expert fusion across region, context, and time
UrbanMoE	Fixed urban patches	Sparse task-specific expert routing induces latent region groups
MDL regionalization	Contiguous graph clusters	Joint inference of partition and temporal drivers
RegionGen	Atomic spatial elements	Multi-objective clustering with periodic re-optimization
DDRO for LEO	Rectangular satellite blocks	GA-based region resizing from workload status
RH-Map	3D hash-map regions	Online region creation and region-wise dynamic removal

2. Formal multi-task formulation in TriForecaster

TriForecaster casts MRELF as a multi-task learning problem with $T$ regions and per-region datasets

$$\mathcal{D}_t = \{(\mathbf{X}^t_i, \mathbf{y}^t_i)\}_{i=1}^{N_t},$$

where $\mathbf{X}^t_i = (\mathbf{S}^t_i,\mathbf{Z}^t_i)$, $\mathbf{S} \in \mathbb{R}^{L \times C}$ contains historical load and covariates over look-back window length $L$, $\mathbf{Z} \in \mathbb{R}^{H \times C_z}$ contains future covariates over horizon $H$, and $\mathbf{y}^t_i \in \mathbb{R}^{H \times 1}$ is the future load target (Zhu et al., 13 Aug 2025). The multi-region objective shares parameters across tasks while retaining region-specific parameters: $T+1$0

The architectural answer to this formulation is a TSMixer-based all-MLP backbone combined with tri-dimensional specialization. Input for region $T+1$1 is embedded into $T+1$2, propagated through $T+1$3 RegionMixer layers and $T+1$4 CTSpecializer layers, and then mapped by a region-specific head $T+1$5. The latent dimension is $T+1$6; optimization uses Adam with learning rate $T+1$7 and batch size $T+1$8; all regions are trained jointly under a loss combining MSE with a contrastive term weighted by $T+1$9 (Zhu et al., 13 Aug 2025).

The motivation is explicitly threefold. Regional variation implies that global models wash out differences while purely local models overfit. Contextual variation means that covariate importance depends on joint weather–holiday–season conditions. Temporal variation means the relevance of those covariates changes across positions within the forecast horizon. The framework therefore treats region, context, and time as distinct but interacting axes of heterogeneity.

3. RegionMixer and dynamic region-level division

RegionMixer is the component that most directly instantiates dynamic multi-region division in the region dimension. Each RegionMixer layer maintains one region-specific expert $T$0 for each region $T$1 and one shared expert $T$2. For input $T$3, the layer produces region-specific and shared outputs $T$4 and $T$5, both in $T$6 (Zhu et al., 13 Aug 2025).

The defining mechanism is similarity-based interaction rather than an explicit gating network. For each time index $T$7 and feature index $T$8, pairwise distance between expert $T$9 and expert $$\mathcal{D}_t = \{(\mathbf{X}^t_i, \mathbf{y}^t_i)\}_{i=1}^{N_t},$$0 is

$$\mathcal{D}_t = \{(\mathbf{X}^t_i, \mathbf{y}^t_i)\}_{i=1}^{N_t},$$1

which is converted into an affinity tensor through a Softmax over negative distances. This acts as a gating distribution over experts per position, but it is computed from expert outputs rather than by a separate router. The layer then performs stochastic pooling: for each $$\mathcal{D}_t = \{(\mathbf{X}^t_i, \mathbf{y}^t_i)\}_{i=1}^{N_t},$$2, it samples one expert index according to the affinity distribution and uses that activation in fusion.

Three design properties distinguish this mechanism. First, there is no gate per region per layer. Second, gating is data-driven via expert similarity, not a separate network. Third, there is no explicit load-balancing loss; exploration is encouraged by stochastic pooling. The consequence is a soft, fine-grained division in which each region can remain self-reliant, borrow from the shared expert, or follow other regions’ experts at per-time-step and per-feature resolution.

This is not static clustering. The framework does not predefine regional groups or a graph. Instead, it repeatedly recomputes similarity-based borrowing across stacked layers, which yields a dynamic region structure that changes with learned activations and input patterns. In the paper’s own formulation, regions that are similar under certain patterns share experts more, whereas regions that diverge rely more heavily on self or shared experts (Zhu et al., 13 Aug 2025).

4. ContextMoE, TimeMoE, and tri-dimensional specialization

On top of region-specialized embeddings, CTSpecializer adds two further axes of division. Each layer applies a context-specializing MoE followed by a time-specializing MoE with a residual connection: $$\mathcal{D}_t = \{(\mathbf{X}^t_i, \mathbf{y}^t_i)\}_{i=1}^{N_t},$$3 ContextMoE uses $$\mathcal{D}_t = \{(\mathbf{X}^t_i, \mathbf{y}^t_i)\}_{i=1}^{N_t},$$4 experts $$\mathcal{D}_t = \{(\mathbf{X}^t_i, \mathbf{y}^t_i)\}_{i=1}^{N_t},$$5, implemented as MLPs along the feature dimension at each timestamp. Expert activations themselves generate a probability tensor through Softmax, and stochastic fusion selects context experts per region, time, and feature. TimeMoE operates analogously after transposing to $$\mathcal{D}_t = \{(\mathbf{X}^t_i, \mathbf{y}^t_i)\}_{i=1}^{N_t},$$6, with experts $$\mathcal{D}_t = \{(\mathbf{X}^t_i, \mathbf{y}^t_i)\}_{i=1}^{N_t},$$7 specializing along the horizon axis (Zhu et al., 13 Aug 2025).

The context dimension is explicitly regularized. For each raw sample $$\mathcal{D}_t = \{(\mathbf{X}^t_i, \mathbf{y}^t_i)\}_{i=1}^{N_t},$$8, the method constructs a contextually similar positive sample and contextually dissimilar negative samples using Euclidean distances in relevant covariates. A contrastive loss on the context-specialized embeddings encourages similar contexts to produce similar representations and dissimilar contexts to diverge. This strengthens the interpretation of ContextMoE as a learned decomposition over weather regimes, calendar regimes, and other covariate configurations.

The combined outcome is tri-dimensional specialization: RegionMixer decides which regional expert behavior to follow, ContextMoE decides which contextual sub-expertise to use, and TimeMoE decides which temporal specialization is appropriate for each horizon position. The paper summarizes this as learning which experts to use for which region, in which context, at which time (Zhu et al., 13 Aug 2025).

5. Static partitioning, empirical evidence, and deployment

TriForecaster is explicitly contrasted with static strategies. It is not static clustering of regions, because there is no fixed partition or prior graph. It is not fixed time segmentation, because TimeMoE decides expert usage per horizon index rather than relying on hand-crafted day parts. It is not fixed context regime design, because context experts are discovered through data-driven softmax routing and contrastive learning. Routing is therefore data-dependent, fine-grained, and stochastic (Zhu et al., 13 Aug 2025).

The empirical evidence is based on four real-world datasets of different granularities: EPC with 3 zones in Tétouan city at 10-minute interval and $$\mathcal{D}_t = \{(\mathbf{X}^t_i, \mathbf{y}^t_i)\}_{i=1}^{N_t},$$9 look-back/$\mathbf{X}^t_i = (\mathbf{S}^t_i,\mathbf{Z}^t_i)$0 horizon; CEESC with 2 areas in China at 15-minute interval and $\mathbf{X}^t_i = (\mathbf{S}^t_i,\mathbf{Z}^t_i)$1 look-back/$\mathbf{X}^t_i = (\mathbf{S}^t_i,\mathbf{Z}^t_i)$2 horizon; City-load with 17 cities at 15-minute interval and $\mathbf{X}^t_i = (\mathbf{S}^t_i,\mathbf{Z}^t_i)$3 look-back/$\mathbf{X}^t_i = (\mathbf{S}^t_i,\mathbf{Z}^t_i)$4 horizon; and Bus-load with 6 buses at 15-minute interval and $\mathbf{X}^t_i = (\mathbf{S}^t_i,\mathbf{Z}^t_i)$5 look-back/$\mathbf{X}^t_i = (\mathbf{S}^t_i,\mathbf{Z}^t_i)$6 horizon (Zhu et al., 13 Aug 2025). Across these datasets, TriForecaster outperforms all TSMixer-based baselines, achieving an average forecast error reduction of $\mathbf{X}^t_i = (\mathbf{S}^t_i,\mathbf{Z}^t_i)$7; more specifically, the paper reports $\mathbf{X}^t_i = (\mathbf{S}^t_i,\mathbf{Z}^t_i)$8 MSE reduction and $\mathbf{X}^t_i = (\mathbf{S}^t_i,\mathbf{Z}^t_i)$9 MAE reduction over the second-best method, with $\mathbf{S} \in \mathbb{R}^{L \times C}$0 MSE reduction on CEESC and $\mathbf{S} \in \mathbb{R}^{L \times C}$1 on Bus-load.

Ablation on the EPC dataset shows that each specialization axis is materially important.

Variant	MSE
Full TriForecaster	0.0795
w/o RegionMixer	0.0925
w/o ContextMoE	0.0966
w/o TimeMoE	0.0981

The qualitative case study on City-load reports that separate single-task models exhibit abnormal spikes and fluctuations, standard MTL predictions become too similar across cases, and TriForecaster produces smoother forecasts that better capture both scale and shape variation across cities and contexts (Zhu et al., 13 Aug 2025).

The framework also has an operational deployment. The paper states that TriForecaster has been deployed on the eForecaster platform in eastern China for city-level short-term load forecasts covering 17 cities, supporting a population exceeding 110 million and daily electricity usage over 100 gigawatt-hours (Zhu et al., 13 Aug 2025). The same section also states that the core ideas are not domain-specific and could plausibly transfer to traffic forecasting, multi-store demand forecasting, air-quality forecasting, or multi-building energy demand.

6. Related formulations, misconceptions, and limitations

A common misconception is that dynamic multi-region division necessarily means time-varying spatial boundaries. The literature does not support such a narrow reading. UrbanMoE keeps spatial regions fixed as city patches and does not explicitly re-draw or re-cluster them, yet its sparse task-specific routing induces latent functional groups through expert-usage signatures (Liu et al., 30 Jan 2026). The MDL regionalization framework infers contiguous spatial partitions and region-level driver time series from discretized temporal data, but the partition remains spatially static while the drivers encode the dynamics; the authors explicitly note that time-varying boundaries are a natural extension rather than part of the current model (Weng et al., 6 May 2026). RegionGen likewise implements dynamic behavior as periodic offline re-optimization in a daily $\mathbf{S} \in \mathbb{R}^{L \times C}$2 pipeline rather than continuous boundary motion, while DDRO in LEO networks more directly resizes rectangular regions according to workload status, and RH-Map dynamically creates and updates region-wise hash-map structures during online mapping (Alkafri et al., 2023, Song et al., 14 Jul 2025, Yan et al., 2023).

From TriForecaster’s perspective, the principal limitations and extension points are architectural rather than definitional. The paper’s explicit future work concerns broader deployment on the eForecaster platform, while the design discussion identifies region similarity modeling via output-based distances as a current choice that could be enhanced by learned region embeddings, spatial graphs, or relational information; RegionMixer’s similarity-based Softmax plus stochastic pooling could be replaced by hierarchical experts or more expressive gating; graph structure such as geographical distance or grid topology could regularize pairwise similarity; and extreme scaling to very many regions may require clustering, low-rank experts, or stronger parameter sharing (Zhu et al., 13 Aug 2025).

Taken together, these works indicate that a dynamic multi-region division framework is less a single algorithm than a recurrent architectural pattern: define region-level units appropriate to the domain, avoid static one-size-fits-all partitions, and allow routing, clustering, coupling, or resource allocation to adapt at the granularity required by the task. TriForecaster provides the clearest blueprint in predictive modeling by making that adaptation explicit across region, context, and time simultaneously (Zhu et al., 13 Aug 2025).