A Different Forecast Without a Different Model

TL;DR for operators

A utility forecasting hundreds of heterogeneous customers and distribution transformers faces a model-management decision: maintain one shared forecaster, train separate local models, or specialize a common model without multiplying its full parameter set.

Li, Cheng, and Weng demonstrate a third architecture. Their method keeps a probabilistic forecasting backbone fixed and constructs customer-specific predictive mean and variance heads from a shared bank of low-rank adaptations. Customers can combine several adapters according to their load characteristics rather than receiving an independently trained model.

On 590 synthetic SMART-DS load profiles, normalized mean squared error (MSE) declines from $2.98 \times 10^{-3}$ to $2.72 \times 10^{-3}$, an 8.72% reduction relative to the unadapted backbone. The continuous ranked probability score (CRPS), which evaluates probabilistic forecast quality, falls from 1.32 to 1.18 kW.

The architectural scaling argument is attractive: with a fixed number of load clusters, the shared adapter parameter count does not increase with the number of customers. But the reported 0.58 MiB storage footprint is not reconciled with the paper’s 2,112,000 adapter parameters. The results justify prototyping shared-adapter personalization, not treating its measured accuracy or storage figures as established production performance.

The cost of knowing each customer’s demand

A utility needs tomorrow’s hourly electricity demand for residential customers, businesses, industrial sites, and distribution transformers. Their consumption patterns differ in magnitude, timing, variability, and exposure to external conditions.

The forecasts must also represent uncertainty. Two customers may have similar expected demand but different probabilities of unusually high consumption. For an operator making risk-sensitive capacity or scheduling decisions, the difference matters.

A single global forecaster can reuse information across customers and simplify model maintenance. Yet it may inadequately represent individual demand distributions. Independent local models offer specialization but multiply training, storage, updating, and management requirements.

The difficult question is whether local predictive distributions require customer-specific model instances at all.

In From Shared Demand Patterns to Local Uncertainty: Probabilistic Load Forecasting by Mixing Compact Adaptations, Haoran Li, Zhe Cheng, and Yang Weng propose an architecture that separates shared temporal knowledge from local distributional specialization.¹

Their approach, M²oE²-mLoRA, does not create an independent forecasting model for every customer. It assembles customer-specific output-head modifications from reusable components.

Personalize the prediction heads, not the whole forecaster

The starting point is M²oE², a variational probabilistic forecasting model that learns shared representations of historical electricity demand and contextual information.

The researchers first train this global backbone across the available load profiles. They then freeze its pretrained parameters.

Customer-specific adaptation is restricted to two output mappings:

• Predictive mean: the expected future electricity load.
• Predictive variance: the uncertainty associated with that prediction.

Each adaptation uses low-rank matrix factors, following the principle of low-rank adaptation (LoRA). Rather than training a complete output-head weight matrix, the model learns smaller matrices whose product represents a parameter update.

For cluster $k$, the modifications to the two predictive heads are:

The important architectural decision is where adaptation occurs. The global model continues to supply temporal and contextual representations, while the compact updates modify how those representations become customer-specific predictions.

This concentrates specialization on the output distributions rather than requiring each customer to relearn the forecasting backbone.

A customer can belong to several demand patterns

The next problem is selecting the appropriate adaptations.

Customers are rarely perfectly represented by one exclusive load category. A distribution transformer, for example, may aggregate several consumption patterns.

The method constructs customer-profile features describing average weekly demand shape, week-to-week variability, and absolute demand magnitude. Principal component analysis reduces the feature dimensions, after which a Gaussian mixture model (GMM) assigns probabilities of membership across load-pattern clusters.

Those probabilities initialize the combination of cluster-level adapters.

For customer $c$, the predictive-mean update becomes:

The predictive-variance head receives an analogous update. Here, $\alpha_{c,k}$ is the customer’s weight for cluster $k$.

A shared lightweight correction network subsequently adjusts the clustering-derived membership scores using the forecasting objective. Its corrections are bounded through a hyperbolic tangent function before softmax normalization. The bound constrains score adjustments, not the absolute change in final membership probabilities.

Training includes membership-weighted initialization of cluster adapters followed by joint optimization of the adapter bank and correction network. The forecasting backbone remains frozen.

The result is a shared collection of adaptation components that can be combined differently for each customer.

The adapted model improves both expected demand and uncertainty scores

The main evaluation uses 590 residential, commercial, and industrial load profiles from the synthetic SMART-DS distribution-system dataset.

The researchers aggregate 15-minute measurements into hourly observations, train on January–October 2018, validate during November, and test on December 1–30.

The December evaluation contains 17,700 daily forecasts and 424,800 hourly predictions, each using a 24-hour forecasting horizon.

The comparison with the unadapted M²oE² backbone is particularly informative because it tests the value of adding the proposed customer-aware adaptation within the same forecasting architecture.

Source: Section V-B, Table II. Lower values indicate better performance. MSE is normalized using each customer’s training-period load range; the other listed metrics are expressed in kW.

The 8.72% normalized MSE reduction indicates improved point predictions within this test population. The approximately 10.61% CRPS reduction adds evidence of improved probabilistic forecasting quality. These two results address different operational concerns: expected demand accuracy and the usefulness of predictive distributions.

Across all eight reported deterministic and probabilistic forecasting metrics, the adapted model achieves the lowest error among the evaluated configurations. Against TimesFM-2.5 with LoRA, its normalized MSE is $2.72 \times 10^{-3}$ rather than $3.15 \times 10^{-3}$, a 13.65% reduction.

These are benchmark comparisons, not proof that each proposed architectural component independently produces a measured share of the improvement. The study does not present a complete ablation isolating soft membership, learned corrections, adapter initialization, and output-head placement.

A separate parameter-sensitivity experiment examines 14, 15, and 16 clusters. November validation MSE is respectively $2.97$, $2.96$, and $2.97 \times 10^{-3}$. Fifteen performs best among these three settings, but the narrow comparison cannot establish a generally optimal cluster count.

Shared adapters change the scaling equation

The central efficiency argument concerns how model parameters grow as a forecasting service adds customers.

With independent customer-level LoRA adapters, the required adaptation parameters grow approximately linearly with the customer count $|\mathcal C|$.

With the proposed shared adapter bank, the principal LoRA parameter term instead scales as:

Here, $K$ is the number of load clusters, $r$ the adaptation rank, and $d_x$ and $d_z$ the output and latent dimensions.

For fixed $K$, rank, and model dimensions, this term does not depend on the number of customers. The system maintains shared adaptation components and customer-specific combination weights rather than independent full adapters.

That is an analytical parameter-scaling result. It does not imply customer-independent costs for storing load histories, computing features, managing memberships, onboarding new customers, or refreshing models.

The reported inference measurements also require careful interpretation. M²oE²-mLoRA takes 0.11 ms per 24-hour forecast, compared with 0.08 ms for the unadapted backbone. Both are faster than the evaluated TimesFM-2.5 with LoRA configuration at 4.93 ms under the authors’ benchmark conditions.

The neural-model measurements use an NVIDIA Tesla T4, and the timing calculation covers a subset of December forecasts. They are not measurements of complete production forecasting workflows.

More consequentially, the storage figures are internally unresolved. Table II reports just 0.58 MiB of total model and routing-data storage for M²oE²-mLoRA, while Table III reports 2,112,000 adapter parameters at 15 clusters.

Conventional uncompressed representations of that parameter count would require considerably more memory. The source does not explain a precision, compression, or accounting scheme that reconciles the figures.

The shared-adapter scaling argument remains mathematically distinct from this discrepancy. Infrastructure budgeting, however, requires a verified accounting of both.

What a utility should validate before adopting the architecture

For a utility forecasting team or a forecasting-platform provider, the relevant decision is whether to prototype a shared backbone and adapter bank instead of expanding a fleet of independently maintained customer models.

The paper makes that architecture credible enough to test, but several questions remain outside its evidence.

First, the benchmark uses selected synthetic profiles from one year, with future observations drawn from the same profile population used in training. It does not establish transfer to entirely new customers or transformers, or robustness across different utilities, regions, seasons, and unusual demand regimes.

Second, aggregate probabilistic scores do not establish calibration for every customer class, transformer composition, or high-load event. An operator using prediction intervals for capacity-risk decisions should examine empirical interval coverage and error behavior separately across relevant segments.

Third, the reported efficiency excludes the full operational lifecycle. A realistic prototype should measure preprocessing, customer-profile construction, adapter assignment, updates, full inference latency, and actual resident and serialized memory.

A sensible validation sequence would compare the shared backbone, the proposed adaptation, and available local-model alternatives on operational load data; hold out new customers rather than only future dates; test temporal and geographic shifts; and reconcile stored parameter counts with measured resource consumption.

The acceptance decision should be based on forecast quality and uncertainty reliability under the intended operating conditions, alongside the actual cost of maintaining the forecasting service.

The model need not multiply with the customers

The study’s contribution is a concrete alternative to choosing between a globally shared forecaster and a growing collection of customer-specific models.

By separating shared representation learning from locally adjusted predictive distributions, and combining reusable low-rank components through soft memberships, M²oE²-mLoRA improves both point and probabilistic forecasting metrics on its selected SMART-DS benchmark.

For operators, the architecture is worth investigating where customer heterogeneity makes a single forecast model insufficient and independent specialization expensive.

Its unresolved questions are equally concrete: performance beyond the synthetic benchmark, reliability for new and changing loads, and the actual memory and maintenance costs of deploying the shared adapter bank.

The next meaningful result would be a reproducible operational comparison that verifies those costs while preserving the forecast improvements.

Cognaptus: Automate the Present, Incubate the Future.