Uncertainty-Aware Transfer Learning for Cross-Building Energy Forecasting: Toward Robust and Scalable District-Level Energy Management

Abstract: Scaling data-driven energy forecasting to district level requires models that can be re-used across buildings with minimal target-domain data and honest uncertainty estimates. We present an uncertainty-aware transfer learning (TL) framework for cross-building energy forecasting based on the Temporal Fusion Transformer (TFT), evaluated on a newly released high-resolution real sub-meter dataset: an educational building at Aalborg University, Denmark (source) and the multi-typology NEST building at EMPA, Switzerland (target). We introduce the Transfer Robustness Index (TRI), an architecture-agnostic metric for quantifying generalization quality across domain gaps. A four-strategy layer-freezing ablation shows that Probe-Only fine-tuning, updating only 455 output-layer parameters out of 806K, achieves the best transfer quality (TRI = 3,097), outperforming full fine-tuning and suggesting that TFT encoders learn transferable temporal representations. Monte Carlo Dropout yields a prediction interval coverage probability of 93.2%, close to the nominal 95% target. A data-scarcity analysis further shows monotonic improvement with increasing target-domain data, providing practical guidance for district energy deployment.

• The paper introduces a novel uncertainty-aware transfer learning pipeline leveraging the Temporal Fusion Transformer for cross-building energy forecasting with minimal target data.
• It demonstrates that updating only the output layer (Probe Only fine-tuning) significantly reduces MAE and ensures robust performance across heterogeneous domains.
• The proposed Transfer Robustness Index provides a standardized metric for comparing transfer quality, while MC Dropout yields well-calibrated uncertainty estimates.

Uncertainty-Aware Transfer Learning for Cross-Building Energy Forecasting

Introduction and Motivation

The paper "Uncertainty-Aware Transfer Learning for Cross-Building Energy Forecasting: Toward Robust and Scalable District-Level Energy Management" (2605.29733) introduces a systematic framework leveraging the Temporal Fusion Transformer (TFT) for robust energy forecasting across structurally and operationally heterogeneous buildings. The work directly addresses the limitations of existing building energy forecasting solutions: their reliance on per-building training, the absence of calibrated uncertainty quantification (UQ), challenges in the presence of domain gaps (climate, typology, occupancy), and the overall data inefficiency crippling district-scale deployment.

Methodology

Data and Preprocessing

The study utilizes a high-resolution, real-world dataset spanning an educational building at Aalborg University (AAU, Denmark) as the source and the NEST building at EMPA (Switzerland) as the target. The dataset covers extensive sub-metered consumption channels (electricity, heating, DHW), with pronounced heterogeneity in climate, function, and building typology. Standardized preprocessing (hourly aggregation, interpolation, non-leakage min-max normalization) ensures comparability while preserving the essential temporal and structural characteristics in both domains.

Model Architecture

The underlying architecture is the TFT, which enables interpretable multi-horizon forecasts using gated residual networks, multi-head attention over LSTM-encoded temporal contexts, and explicit quantile regression for generating predictive distributions. The inputs include static building identifiers, time-based cyclical features, observed weather, and multi-energy sub-meter readings.

Transfer Learning Pipeline

The transfer learning protocol comprises four layer-freezing strategies, each controlling the fraction of parameters adapted during target-domain fine-tuning:

• Full Fine-Tuning (FF): All 806K parameters updated; allows for maximum task adaptation but at high risk of catastrophic forgetting given small target data.
• Partial Fine-Tuning (PF): Input embeddings frozen; encoder, decoder, and output head updated.
• Probe Only (PO): All parameters except the final 455-dimensional output projection are frozen; tests the reusability of learned temporal abstractions.
• Progressive Unfreezing (PU): Encoder and embedding layers frozen; decoder and head updated.

Sub-meter channel alignment across domains is handled by zero-padding as necessary.

Transfer Robustness Index

A key contribution is the Transfer Robustness Index (TRI), defined as the ratio of source-domain validation MAE to target-domain test MAE, both on normalized scales. This provides an architecture-agnostic, relative measure for generalization quality across diverse domain gaps, mitigating the lack of standardized evaluation criteria in the field.

Uncertainty Quantification

MC Dropout is employed for predictive uncertainty estimation. Stochastic forward passes at inference yield calibrated 95% prediction intervals. Performance is assessed via prediction interval coverage probability (PICP) and mean interval width (MIW), quantifying both calibration and informativeness.

Experimental Results

Forecasting and Transfer Performance

Zero-shot direct transfer without fine-tuning (MAE = 15.037, TRI = 1.05) fails catastrophically due to extreme domain gap, establishing the necessity of adaptation. All fine-tuned strategies reduce MAE by over 99.9%. Notably, Probe Only fine-tuning—updating only 455 output-layer parameters—achieves the best transfer quality (MAE = 0.0051, TRI = 3,097), outperforming full fine-tuning and indicating that the encoder's learned temporal features are highly transferable even across climate and typology divergences.

The negative $R^2$ values for all strategies (e.g., $R^2 = -0.046$ for PO) result from the difficulty of capturing stochastic spikes and demonstrate that relative error and robustness metrics (MAE, TRI) are better suited in these settings.

Uncertainty Calibration

The MC Dropout-based PO model reports PICP = 93.2% (nominal 95%) with tightly bounded mean prediction intervals (MIW = 0.028), signifying that the model provides well-calibrated and informative uncertainty estimates. Importantly, interval width increases around unfamiliar, high-variability regimes, reflecting appropriate epistemic uncertainty awareness.

Data Scarcity Analysis

A controlled data-scarcity protocol—fine-tuning with incremental windows (2 weeks to full data)—shows monotonic improvement in target-domain MAE and TRI as more data becomes available. Marked improvement is evident with at least 3 months of target data (MAE = 0.163, TRI = 96.8). This provides critical operational guidance: while the architecture allows meaningful transfer with minimal data, approximately three months of fine-tuning data is needed for robust deployment.

Implications and Future Work

The results have key theoretical and practical implications:

• Theoretical: The TRI metric establishes a reproducible baseline for comparative assessment of cross-domain transfer, agnostic of model architecture. Furthermore, the superiority of Probe Only adaptation is consistent with catastrophic forgetting literature and confirms that deep sequence encoders—when pretrained on high-frequency multi-vector energy domains—learn portable structural patterns transcending individual building idiosyncrasies.
• Practical: By decoupling the transferable encoder from the final calibrated target mapping, scalable district-level energy management systems can be built with minimal target data, low risk of overfitting, and high reliability. In real deployments, this allows for rapid expansion to new sites and supports demand-side flexibility and grid-interactive strategies with quantifiable risk.
• Future Directions: The framework opens pathways for widespread, uncertainty-aware forecasting across larger, stratified building portfolios. Refinements could incorporate domain alignment techniques, multi-target forecasting, conformal UQ, and integration with foundation or multitask time-series models.

Conclusion

The study presents a robust, uncertainty-aware transfer learning pipeline for district-scale cross-building energy forecasting, validated on large real-world datasets with extreme domain heterogeneity. It demonstrates that TFT encoders learn highly transferable representations, requiring only minimal adaptation on the output layer and a modest quantity of target data for robust performance. The introduction of the Transfer Robustness Index enables standardized benchmarking, and the use of MC Dropout ensures well-calibrated uncertainty estimates—a critical requirement for risk-aware, scalable energy management across smart districts.