acceptodds
Under review as a conference paper at ICLR 2027

STGFM: A Spatio-Temporal Global Factor Model for Multi-Enterprise Electricity Load Forecasting

Abstract

Accurate multi-series multi-step load forecasting requires exploiting shared dynamics across series while preserving series-specific heterogeneity. However, cross-series associations evolve across forecasting instances, making fixed shared representations insufficient for these dual objectives. We propose a spatio-temporal global factor model for multi-series load forecasting, termed STGFM. A cross-series Transformer derives input-conditioned entmax aggregation weights that project historical loads onto compact time-indexed factors. A Mamba state-space module encodes historical factor dynamics, whose final representation is combined with known future covariates to generate multi-step factor forecasts. A conditional decoder then reconstructs series-specific trajectories from the predicted factors, series context, and forecast-step information. Experiments on the Real-world Load Dataset and the UCL Electricity Dataset show that STGFM reduces forecasting errors across multiple horizons. On the Real-world Load Dataset, STGFM reduces RMSE by 2.0%–7.0% across four horizons (12, 24, 48, and 96 hours) compared with TimesNet. On the UCL Electricity Dataset, the corresponding reductions are 8.2%–21.3% at 24, 48, and 96 hours relative to the strongest baseline at each horizon. Ablation studies further show that conditional reconstruction consistently reduces both RMSE and MAE compared with shared MLP decoding across both datasets and all four horizons.

Primary area:
learning on time series and dynamical systems
Keywords:
Multi-entity load forecasting, Global factor model, Time-varying associations, Cross-series Transformer, Mamba state-space

Reviews

Among:
Shape:
1234this paper, Bottom 2%
Figure 1. Mean review rating of the reviewed papers at ICLR 2027. This paper averages 1.33 over 3 reviews, above 2% of the others.
Reviewer_Tyn7OpenReview ↗
Rating
2
Confidence
3
Informativeness
2
Presentation
3
Soundness
2
Read review
Paper summary

The paper proposes STGFM for multi-series electricity load forecasting. A cross-series Transformer constructs input-dependent weights that compress historical loads into time-indexed global factors. Mamba encodes their history, and known future calendar covariates help generate factor forecasts. A conditional decoder then combines these forecasts with historical series context, static statistics, and forecast-step information to recover individual load trajectories. Experiments cover two datasets and four horizons, with five runs per stochastic setting. The reported RMSE gains are promising, particularly at longer horizons on the selected UCL panel.

Critical strengths
  1. The design addresses a useful tension between shared dynamics and differences among individual load series. Factor construction preserves the historical time axis, while reconstruction uses series context and forecast-step information to adapt the common representation (Sections 3.2–3.3). The paper also clearly distinguishes weights that adapt across forecasting windows from weights that remain fixed within each window. This makes the approach easy to understand and gives a concrete way to combine shared and individual information.

  2. The forecasting results are encouraging and the comparison tables are reasonably transparent. STGFM has lower mean RMSE than the evaluated learned baselines at all eight dataset–horizon settings. On UCL, its RMSE reductions relative to the strongest reported baseline at 48 and 96 hours are 21.3% and 16.4% (Tables 4–5). The paper reports five-seed variation, includes seasonal forecasting, and acknowledges the settings where another method has lower MAE or Seasonal Naive performs better. The conditional decoder also improves mean RMSE and MAE over the shared MLP in all eight ablation settings.

Critical weaknesses
  1. The closest comparisons and the baseline protocol are incomplete. Section 2 discusses Deep Factors, DeepGLO, and SOFTS, which also combine shared structure with series-specific prediction, but none is evaluated in Tables 4–5. The experiments therefore leave the benefit over a close factor or series-core alternative unresolved. Sections 4.2–4.4 also do not specify the method-specific input mappings or explain whether calendar covariates, observation masks, and static statistics are available to each baseline. The scope of context-length selection, the training objective, and architecture-specific tuning is unclear. Using the same datasets, horizons, and chronological test protocol does not establish comparability on these points. A complete configuration account and a comparison with a close alternative would help determine whether the gain follows from the proposed decomposition, additional information, or model selection. This is a central gap because the paper's contribution is the way it organizes shared and individual information.

  2. The decoder ablation does not clearly isolate conditional modulation. The proposed decoder receives historical context, static statistics, calendar covariates, and step embeddings through Equations (11)–(14), whereas Section 4.5 only describes the replacement as a shared MLP. Its retained inputs and parameter count are unspecified. The consistent improvement over this variant supports the tested decoder configuration, but does not separate modulation from access to additional information or capacity. The dynamics results also support a narrower interpretation: the LSTM variant has lower mean RMSE in six of eight settings, including 981.4 ± 90.5 versus 1104.6 ± 69.2 at UCL 96 hours (Table 3). These are reported means and standard deviations, not evidence of statistical significance. The authors already acknowledge horizon-dependent component effects; the remaining issue is which controlled comparison supports each mechanism. A decoder comparison with the same conditioning inputs and comparable capacity would directly address the strongest attribution concern. These two evidence gaps keep my recommendation just below the acceptance threshold.

Questions for authors
  1. What calendar covariates, observation masks, and static series information does each baseline receive? Are context lengths, training objectives, and tuning budgets shared or selected separately? Please provide the configurations used in Tables 4–5.

  2. Which historical, static, calendar, and forecast-step inputs does the shared-MLP decoder retain, and what are its parameter count and that of the proposed decoder? Are the Mamba and LSTM ablations also matched in capacity and tuning?

  3. Were Deep Factors, DeepGLO, or SOFTS tested under the same protocol? If results already exist, which comparison best separates the contribution of adaptive time-indexed factors from shared structure and series-specific prediction?

  4. Are the raw mean and zero-value ratio in Section 4.2 computed from training data only? Does the selection of the 64 UCL series use training-only coverage and activity or the complete time span?

Reviewer_BhYvOpenReview ↗
Rating
1
Confidence
3
Informativeness
1
Presentation
2
Soundness
1
Read review
Paper summary

The paper proposes STGFM for multi-series, multi-step electricity load forecasting. Each series' look-back window (with an observation mask and a summary of calendar covariates) is embedded as one token; a Transformer attends across series; a linear layer followed by α-entmax over the series dimension gives K sets of input-dependent weights, and the K "global factors" are the corresponding weighted sums of the raw series at each historical time step. A Mamba block encodes the factor sequence; its final state is repeated over the horizon, concatenated with projected future calendar covariates, and mapped to K future factor values per step. A FiLM-style decoder rescales and shifts the factor vector per series and step (scale in (0.9, 1.1), shift in (−0.1, 0.1)) from series context, static statistics, calendar features and a step embedding, and a shared head outputs the forecast. Experiments use a private "Real-world Load Dataset" (1,257 industrial series, 4,080 hours) and a 64-series subset of the UCI electricity data (called "UCL" in the paper) at horizons of 12–96 hours, against 13 baselines, with five seeds and a five-variant ablation.

Critical strengths
  1. Honest and complete reporting. All 13 baselines, both naive references, five-seed means and standard deviations, and ablation results that are unfavourable to the method are reported in full, and the text concedes where Seasonal Naive or TimesNet is better. This is what makes the paper evaluable, and I credit it.

  2. A coherent factor formulation. Building time-indexed factors as window-specific convex combinations of the observed series, forecasting in factor space with known future covariates, and reconstructing per series is a reasonable way to connect classical dynamic factor models with learned aggregation.

Critical weaknesses
  1. The paper's own results do not support its claims about the method or its components.

    • Naive baseline. On the public dataset at 12 hours, Seasonal Naive beats STGFM on every metric (RMSE 634.7 vs 725.4; MAE 75.0 vs 86.7; R² 0.926 vs 0.903); at 24 hours it has the lower MAE (101.2 vs 113.1). On the private dataset, STGFM has a worse MAE than TimesNet at 12 and 96 hours and than Seasonal Naive at 48 hours. The improvements are confined to squared-error metrics computed on the original scale and pooled over series, which are dominated by a few large consumers (RMSE is about eight times MAE on the public data); pooled R² likewise mostly reflects between-series scale. No scale-normalized or per-series metric is given.
    • Variance. STGFM's seed standard deviation on the public data is roughly ten times that of the baselines (RMSE 882 ± 85 at 24 h, 933 ± 133 at 48 h, versus ± 5–12 for TimesNet/DLinear). At 12 and 24 hours the intervals overlap with TimesNet and with Seasonal Naive, so the "8.2%" 24-hour gain is not established.
    • Ablation. On the private dataset, replacing Mamba with an LSTM is better at 12, 48 and 96 hours (0.1796, 0.2071, 0.2255 vs 0.1810, 0.2103, 0.2261 RMSE), and removing the dynamics module altogether is better at 48 hours. Removing the cross-series encoder is tied at 12 hours, and static linear loadings are tied at 12 and 24 hours; on the public data several variants beat the full model at 12 hours. Every one of these differences is within one standard deviation. The claims that "removing the cross-series encoder increases both RMSE and MAE at every horizon" and that input-conditioned aggregation and Mamba dynamics contribute are therefore not supported; the premise that cross-series associations change across windows is never examined (no analysis of how the weights vary, of sparsity, or of what the factors represent). For two of four horizons the selected α is 1.0, i.e. plain softmax, so the "sparse factors" of the conclusion do not exist there.
    • The one consistent effect, FiLM decoder versus "shared MLP", compares against a decoder that has no series-specific input and so cannot map K shared factors to N different series; it is a strawman, not evidence for this particular conditional design.
  2. The comparison is not adequate for the claims, and the work is not reproducible.

    • The baselines appear untuned or disadvantaged: PatchTST and iTransformer are worse than Seasonal Naive and DLinear on electricity load at most horizons, which contradicts their established behaviour on this type of data. STGFM's look-back (96–432 h) and α are selected per horizon on validation, and it receives future calendar covariates and static series statistics; the paper does not say what look-back, tuning or covariates the baselines get.
    • None of the factor-based or cross-series methods that the related-work section identifies as the closest prior art is compared: DeepGLO, Deep Factors, SOFTS, TFT, DeepAR, Crossformer, DUET, MTGNN. Covariate-aware recent models (e.g. TimeXer) and recent strong general baselines are also absent. Novelty is limited to assembling known blocks (series-as-token attention, entmax pooling, Mamba, FiLM).
    • One dataset is private and undescribed (source, region, missing rate, 170 days in total, so a test set of about 25 days). The public one is a non-standard 64-series subset chosen by coverage rather than the usual 321/370 series, and uses a 70/15/15 split, so no number can be checked against the literature. No code is offered.
    • The decoder bounds each series' deviation from the shared factor vector to ±10% scale and ±0.1 shift and has no local/idiosyncratic component (unlike Deep Factors or DeepGLO). Whether this limits heterogeneous series is not examined, although preserving heterogeneity is the stated goal.
Questions for authors

Clarifications that could change my assessment:

  1. What look-back, hyperparameter search and covariates did each baseline receive? Please give all baselines the same future calendar covariates and a per-horizon look-back search, and explain why PatchTST and iTransformer fall below Seasonal Naive.
  2. Please report scale-free metrics (per-series normalized MAE/RMSE averaged over series, MASE or WAPE) and the distribution of per-series improvement. Is the RMSE gain carried by a few high-load series?
  3. Why is STGFM's seed variance an order of magnitude above the baselines' on the public data? Please add paired significance tests for Tables 1–3.
  4. Given that LSTM and no-dynamics variants match or beat the full model, what is the justification for Mamba? Given that static loadings tie at short horizons, what is the evidence that the weights A vary meaningfully across windows?
  5. Please compare with DeepGLO, Deep Factors, SOFTS, TFT and TimeXer on the full, standard Electricity benchmark with the standard split, and state the correct name and source of the public dataset.
  6. Why does a 12-hour-ahead learned model lose to Seasonal Naive on all metrics on the public data?

Suggestion for future work (does not affect my rating): a per-series local component alongside the shared factors, and probabilistic forecasts as the authors propose.

Reviewer_HcznOpenReview ↗
Rating
1
Confidence
3
Informativeness
1
Presentation
2
Soundness
1
Read review
Paper summary

This paper focuses on how multivariate time series forecasting models can leverage cross-variable information while preserving sequence-specific characteristics, which has been extensively discussed in the field over the past few years. To address this, the authors propose the Spatio-Temporal Global Factor Model (STGFM) for multi-series electricity load forecasting. Specifically, STGFM uses a cross-series Transformer with α-entmax to compress multiple series into low-dimensional global factors, employs a Mamba state-space module alongside future temporal covariates to predict factor evolution, and utilizes a FiLM-based conditional decoder to reconstruct series-specific forecasts. While the paper attempts to balance shared dynamics and individual heterogeneity, it suffers from fundamental flaws regarding architectural novelty, contradictory ablation results and inadequate baseline comparisons.

Critical strengths
  1. The authors incorporated engineering considerations relevant to real-world industrial data and explicitly addressed missing values. This issue often bypassed in existing open-source time-series forecasting datasets through preprocessing steps like zero-padding or mean-padding.
Critical weaknesses
  1. Misleading word using in the title. While the title features the term "Spatio-Temporal", the methodology and formulation do not involve spatial coordinates, geographical topologies, or graph structures. The study is primarily focused on multivariate time-series forecasting, and aligning the terminology in the title with the actual scope would better reflect the work's primary contribution.

  2. Methodological Positioning and Novelty. The problem setting and the overall "aggregate-forecast-reconstruct" framework follow established paradigms in channel-mixing and global factor-based time-series forecasting, as reviewed in Section 2 (e.g., SOFTS, DeepGLO, and Deep Factors). Since the architecture largely combines existing components, further discussion on how these combinations introduce distinct theoretical or empirical advantages over prior factor models would strengthen the work's positioning.

  3. Mathematical Formalism and Notation Clarity. The mathematical formulations in Section 3 would benefit from a more rigorous and standard presentation. Expressing definitions at the instance level rather than carrying the batch dimension and array-slicing notation throughout the equations, alongside more formal definitions for tensor operations, would greatly improve the readability and clarity of the methodological design.

  4. Inconsistency between methodology and experiments. The ablation studies in Table 3 reveal that key proposed components are ineffective across multiple scenarios. Specifically, on the public UCL dataset, removing the cross-series encoder improves RMSE at the 12h and 96h horizons. More critically, replacing the Mamba dynamics module with a simple LSTM yields superior RMSE across most horizons on both datasets (e.g., 12h, 24h, and 96h on UCL, and 12h, 48h, and 96h on the private dataset), undermining the validity of the core architectural choices.

  5. Baseline Selection. The experimental evaluation primarily relies on traditional or general-purpose baselines. Given that several recent state-of-the-art cross-variate and factor-based models are thoroughly discussed in Section 2, including some of these directly relevant recent methods in the empirical comparison would provide a more comprehensive evaluation of the model's relative performance.

  6. Parameter Justification in Mechanism Design. Certain hyperparameter choices and design constraints, such as the specific modulation ranges for FiLM parameters and the fixed loss-component weighting ratios, are presented without detailed theoretical justification or sensitivity analyses. Providing a parameter sensitivity study would offer valuable insights into the robustness and stability of these design choices.

Questions for authors
  1. In the ablation study, why does replacing the Mamba module with a standard LSTM (w/ LSTM) yield lower RMSE across most horizons on both datasets, and why does removing the cross-series encoder (w/o CSE) improve RMSE at the 12h and 96h horizons on the UCL dataset? How do the authors justify the necessity of these proposed components given such contradictory empirical evidence?

  2. Can the authors provide theoretical justification or sensitivity analyses for tightly restricting the FiLM modulation ranges?

From OpenReview, CC BY 4.0. Authors as signed there; shown unedited.

Related papers

Loading the map…

Discussion (0)

Sign in to comment.