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.
Figures & tables
Dataset
Weather covariates
Train split
Test split
ISO‑New England (8 zones)
Dew point and dry bulb temperature
2011–2020 (3,653 days)
2021 (365 days)
ENTSO-E (6 countries)
Temperature; direct and diffuse horizontal radiation
TABLE III: Average MAPE of models trained for different forecast horizons over ISO New England ∗ .
Method
Covariates
Model type
H=30
H=60
H=90
MMMF
✓
TCN
4.52
4.51
4.43
LSTM
5.21
5.22
5.39
Transformer
4.65
4.7
4.82
Chronos-2
✓
Zero-shot
5.63
6.57
7.67
LoRA fine-tune
4.54
4.91
5.32
Full fine-tune
4.31
4.66
4.86
TABLE IV: Average MAPE of models trained for different forecast horizons over ENTSO-E ∗ .
Encoding
Zero-shot
Full fine-tune
No date features
10.15
5.58
Ordinal
10.33
5.3
Cyclical
10.61
5.19
∗ Table shows average MAPE across horizon H=60 .
TABLE V: Ablation study over ISO New England on the impact of calendar feature encoding ∗ .
Fig. 1: Weekly forecast performance. Columns correspond to model types, and rows correspond to datasets. Each model is trained with forecasting horizons H∈{30,60,90} . The y-axis shows average MAPE by week (x-axis), and error bars denote the standard deviation across five random seeds.
Fig. 2: Forecast performance of zero-shot Chronos-2 in cold-start scenarios, with and without ICL in (a) ISO New England (b) ENTSO-E. X-axis represents the context length of each target zone. Each ICL line plot represents a cross-learning scenario with the context length of non-target zone ∈{30,60,90} Note: Y-axes are scaled differently across subplots to best visualize the trends within each dataset.
Karlsruhe Institute of Technology (KIT), Germany and Netze BW GmbH, Germany · Netze BW GmbH, Germany · Karlsruhe Institute of Technology (KIT), Germany