PaperlayerView exact revision on arXiv

Source-linked AI summary

Assessing Covariate-Informed Grid Load Forecasting with a Time-Series Foundation Model

Varsha Pendyala, Yiwei Fu, Weizhong Yan, Nurali Virani

arXiv:2609.06656v1cs.LGcs.AIeess.SY

TL;DR

The paper addresses whether a time-series foundation model can support multivariate, covariate-informed grid load forecasting and how it compares with task-specific models. It evaluates Chronos-2 across zero-shot and adapted regimes on two utility datasets, finding that fine-tuning improves short-horizon performance while zero-shot accuracy lags and errors grow faster with horizon.

  • Problem

    It remains insufficiently understood how Chronos-2 performs for multivariate and covariate-informed grid load forecasting relative to established task-specific deep learning models, especially when task-specific historical data are available.

  • Method

    The study evaluates Chronos-2 on ISO New England and ENTSO-E, benchmarking zero-shot and adapted configurations against task-specific deep learning models and MMMF-based baselines.

  • Results

    Fine-tuned Chronos-2 achieves superior short-horizon performance, but zero-shot Chronos-2 underperforms task-specific models and its errors increase more rapidly with horizon length.

  • Takeaways & Limitations

    Chronos-2 can be useful for operational load forecasting when adapted with task-specific data, while its advantages are horizon-dependent and limited in zero-shot use.

  • Takeaways & Limitations

    Foundation-model computational overhead may be prohibitive for frequent updates across thousands of nodes without substantial hardware acceleration.

Abstract

from arXiv · show

Modern power systems are growing increasingly complex as they integrate diverse generation sources to meet rising demand, making accurate load forecasting challenging. Recent advances in time-series foundation models (TSFMs) resulted in promising performance in zero-shot univariate load forecasting tasks. However, real-world load forecasting often involves multiple target variables and requires the integration of exogenous variables, raising important questions about the utility of TSFMs in realistic settings. In this study, we position Chronos-2, a recently developed model by Amazon, as a representative multi-channel TSFM that supports univariate, multivariate, and covariate-informed forecasting, and conduct a systematic investigation of how such models can be used for real-world load forecasting. While prior work has evaluated Chronos-2 on a limited number of energy-related tasks in a zero-shot setting, its performance relative to established task-specific deep learning models and its behavior when adapted using task-specific historical data remains insufficiently understood. In this work, we evaluate Chronos-2 on two real-world utility datasets, ISO New England and ENTSO-E, and benchmark it against widely used task-specific deep learning models. Our results show that Chronos-2 benefits substantially from task-specific fine-tuning and achieves strong short-horizon forecasting performance, but its zero-shot accuracy lags behind task-specific models and its forecasting error grows more rapidly with increasing forecast steps. Overall, this study provides a detailed characterization of the strengths and limitations of TSFMs such as Chronos-2 in grid load forecasting and offers practical insights into how a pretrained TSFM can be effectively adapted for operational load forecasting applications.

I. INTRODUCTION

The paper examines whether Chronos-2 can support realistic multivariate and covariate-informed grid load forecasting beyond zero-shot univariate tasks. It evaluates zero-shot and adapted regimes against task-specific deep learning models, finding benefits from fine-tuning but faster error growth across forecast horizons.

  • Motivation: Real-world utility forecasting commonly requires simultaneous prediction across regions and integration of weather and calendar covariates.This contrasts with the predominantly univariate zero-shot setting emphasized in earlier TSFM studies.
  • Study design: Chronos-2 is evaluated on ISO New England and ENTSO-E against established task-specific deep learning forecasting models.The two utility datasets contain multiple target variables and diverse covariates.
  • Study design: The study compares zero-shot Chronos-2 with task-specific models and examines full fine-tuning, LoRA adaptation, calendar incorporation, and horizon-dependent error.These analyses target practical conditions in which utilities possess task-specific historical data.
  • Findings: Zero-shot Chronos-2 has substantially higher forecasting errors, whereas fine-tuned Chronos-2 performs better at short horizons than deep learning models.Full fine-tuning outperforms LoRA-based adaptation, and calendar variables improve fine-tuned performance.
  • Findings: Chronos-2’s forecasting error increases more rapidly with forecast steps, while task-specific deep learning models show more stable error trajectories.The paper frames this horizon dependence as relevant to distinct short-term and mid-term grid decisions.

II. BACKGROUND

Chronos-2 is a multi-channel encoder-only Transformer designed for univariate, multivariate, and covariate-informed forecasting. Its architecture combines temporal and cross-series processing for direct multi-step probabilistic prediction.

  • The Chronos-2 model: Chronos-2 is a 120M-parameter encoder-only Transformer trained on univariate, multivariate, and covariate-informed forecasting tasks.For D targets, M covariates, history T, and horizon H, its input dimensions are (D + M) × (T + H).
  • The Chronos-2 model: Each input series represents either a target or covariate, while categorical features use target or ordinal encoding.The model also incorporates time-index and observed-value mask metadata at each time step.
  • Input processing: Chronos-2 standardizes and transforms input series, adds metadata masks, divides series into non-overlapping patches, and tokenizes them into fixed-size embeddings.The mask distinguishes missing future targets and unknown future covariates from known future covariates.
  • Attention mechanism: Alternating time-attention and group-attention layers capture within-series temporal dependencies and information sharing across grouped series.Grouping targets and covariates supports cross-series dependencies and covariate–target interactions.
  • Training objective: Quantile regression produces direct multi-step forecasts for multiple quantiles, with pretraining horizons increasing to 1,024 in the second stage.The first pretraining stage uses a shorter horizon whose exact length is unspecified.

B. Masked Multi-step Multivariate Forecasting (MMMF)

MMMF formulates multivariate forecasting as masked reconstruction using historical targets and known future covariates. It directly predicts the masked future segment, avoiding recursive error propagation.

  • Framework: MMMF is a self-supervised framework that formulates multi-step forecasting as a masked reconstruction task incorporating known future information.It is used as the primary baseline because prior work found it superior to several traditional and direct forecasting approaches on real-world utility datasets.
  • Training: During training, the model receives historical targets, full covariate sequences, and future targets replaced by random masks.The sequence spans the history and forecast period, with future target values masked for reconstruction.
  • Inference: The loss is computed only on masked target segments, and inference masks the desired horizon before generating multi-step predictions directly.This avoids the error propagation inherent in recursive strategies that feed predictions back into the model.

III. EVALUATION SETUP

The evaluation compares Chronos-2 and MMMF-based models across zero-shot and fine-tuned configurations using two utility datasets. Experiments vary covariate use, adaptation method, and forecast horizon under a fixed 30-day history.

  • Model configurations: Chronos-2 is evaluated zero-shot and after task-specific fine-tuning across target-history-only and covariate-informed configurations.Fine-tuning also compares full parameter updates with LoRA and varies the forecast horizon.
  • Evaluation protocol: All models use a fixed history of 30 days during inference, while Chronos-2 performance is reported using MAPE from its median prediction.Chronos-2 and MMMF models are trained on an NVIDIA H200 GPU.
  • Datasets: The study uses ISO New England electricity demand and weather data for eight zones, training on 2011–2020 and evaluating on 2021.The dataset covers Connecticut, Maine, Northeast Massachusetts and Boston, New Hampshire, Rhode Island, Southeast Massachusetts, Vermont, and West/Central Massachusetts.
  • Datasets: ENTSO-E combines electricity load and weather covariates for six European countries in the Entsoe-1H multivariate forecasting task.The countries are Austria, Belgium, Germany, Hungary, Luxembourg, and the Netherlands.

B. Chronos-2 fine-tuning

Chronos-2 fine-tuning uses validation-selected checkpoints and repeated training, with separate optimization settings for full and LoRA adaptation.

  • Fine-tuning procedure: Each Chronos-2 configuration is fine-tuned for 20 epochs, selecting the checkpoint with the best validation MAPE for test evaluation.Validation uses 20% of each dataset's training set.
  • Adaptation settings: The fine-tuning setup compares full adaptation with parameter-efficient LoRA adaptation under the same validation-based model-selection procedure.
  • Fine-tuning procedure: Each fine-tuning run is repeated with five random seeds, and results are reported using means and standard deviations.
  • Adaptation settings: Full fine-tuning uses an initial learning rate of 10^-6, whereas LoRA uses 10^-5 because it learns substantially fewer parameters.

C. Baseline models

The study benchmarks Chronos-2 against MMMF-based TCN, LSTM, and Transformer models across datasets and forecast horizons. Zero-shot Chronos-2 trails the baselines, while task-specific fine-tuning often matches or surpasses them.

  • Baseline architectures: MMMF baselines comprise Temporal Convolutional Network, Long Short-Term Memory, and Transformer models.These models are trained within the MMMF framework and encode categorical variables with five-dimensional embeddings.
  • Evaluation: Tables III and IV compare average MAPE across regions and forecast steps for three training horizons, with and without covariates.For each dataset, MAPE is averaged over all regions and forecast steps from day 1 through day H.
  • Zero-shot comparison: Across both datasets, MMMF models outperform pretrained Chronos-2 in the zero-shot setting, and their performance gap widens as forecast horizon H grows.
  • Fine-tuned comparison: Task-specific fine-tuning substantially improves Chronos-2 accuracy, with fine-tuned models often outperforming or matching MMMF baselines.Full fine-tuning consistently performs better than LoRA, especially when covariates are available.

B. Calendar variables

The calendar-variable study compares ordinal and cyclical encodings for Chronos-2 fine-tuning. Calendar features help adapted models, with cyclical encoding producing the best performance, but not zero-shot models.

  • Encoding strategies: Calendar variables are converted to numerical inputs using either ordinal encoding with min-max scaling or cyclical encoding.Cyclical encoding represents each variable through sine and cosine transformations tied to its temporal period.
  • Evaluation: The calendar-feature ablation is conducted over ISO New England and summarizes results for H = 60.
  • Performance effects: Calendar features do not improve zero-shot performance, whereas fine-tuned models leverage them effectively.
  • Performance effects: Cyclical encoding yields the best fine-tuned performance because it preserves proximity between adjacent points in a temporal cycle.Ordinal encoding places the beginning and end of a cycle, such as December and January, far apart.

C. Performance at various horizons

Chronos-2 is strongest in the initial forecast weeks but its error rises with later forecast steps, eventually exceeding MMMF performance. Longer future-signal windows improve Chronos-2’s near-term forecasts but hurt MMMF models initially.

  • Forecast progression: Chronos-2 significantly outperforms MMMF in the initial weeks, indicating stronger use of recent historical information for short-term forecasting.
  • Forecast progression: Chronos-2’s weekly MAPE increases steadily across subsequent weeks and eventually exceeds the MAPE of MMMF models.
  • Effect of forecast horizon: Chronos-2 models trained with larger H consistently achieve lower weekly MAPEs than models trained with smaller H.The authors suggest that longer future signals refine near-term trajectory estimates.
  • Effect of forecast horizon: MMMF LSTM and Transformer models show the opposite pattern, with longer future-signal access worsening performance in the initial weeks.

D. Chronos-2’s cold-start ICL capability

Chronos-2’s in-context learning provides modest accuracy gains in zero-shot cold-start forecasting, especially when target-region history is scarce, but its practical benefit diminishes with more target history.

  • The experiment evaluates 30-day forecasts using target histories from 5 to 30 days and non-target histories of 30, 60, or 90 days.
  • ICL yields modest forecasting improvements in zero-shot cold-start scenarios.
  • The largest gains occur with only 5 days of target-region history, reducing MAPE by approximately 0.3 percentage points for ISO New England and 1 percentage point for ENTSO-E.
  • ICL improvements diminish as the target-region history increases toward 30 days.
  • With 30 days of target history, the best ICL MAPEs remain above both the best MMMF models and fine-tuned Chronos-2 on the same 30-day task.The best ICL MAPEs are 11.27 for ISO New England and 5.8 for ENTSO-E, compared with fine-tuned Chronos-2 values of 10.07 and 5.16, respectively.

E. Computational and operational implications

Chronos-2 offers accuracy advantages for short-horizon planning but introduces substantial computational and model-management costs that constrain high-frequency, high-volume deployment.

  • Chronos-2 has approximately three orders of magnitude more parameters than task-specific baselines and nearly two orders of magnitude longer inference time than MMMF-Transformer.Chronos-2 uses 120M parameters and requires 3.1 ms per sample, versus 100K parameters and 0.044 ms for MMMF-Transformer.
  • For real-time forecasting across thousands of nodes, foundation-model computational overhead may be prohibitive without substantial hardware acceleration.
  • Serving separate 120M-parameter models for hundreds or thousands of utility entities is impractical because of storage costs and GPU memory constraints.
  • LoRA can mitigate storage expansion by keeping one frozen backbone and swapping lightweight adapter weights for each target zone.The adapters typically contain less than 1% of the total parameters.
  • Foundation models are better suited to short-horizon planning when latency is less critical, whereas lightweight task-specific models remain operationally advantageous for high-frequency real-time dispatch.

V. CONCLUSIONS

The study evaluates Chronos-2 against task-specific models on real-world, multiregion utility data and finds that adaptation improves short-horizon accuracy, while zero-shot performance and long-horizon stability remain weaker.

  • The evaluation uses ISO New England and ENTSO-E datasets containing multiregion load data, calendar features, and weather covariates.
  • Zero-shot Chronos-2 does not match the best task-specific models, whereas task-specific fine-tuning enables superior short-horizon performance.
  • Chronos-2’s forecasting errors increase more rapidly with horizon length and eventually exceed those of task-specific models at longer horizons.
  • Full fine-tuning consistently outperforms LoRA, especially when covariates are available, while cyclical calendar encoding improves fine-tuned accuracy.
  • Cold-start ICL provides only marginal benefits, limiting its practical impact in the forecasting tasks studied.
  • Future work will examine additional foundation models, higher-resolution forecasting, probabilistic predictions, and quantile-based risk-sensitive evaluation.
Loading 2609.06656v1…