Behind-the-Meter Photovoltaic Disaggregation via Deep Learning: Deployment Regimes and Performance Insights

Summary:This research investigates deep learning-based disaggregation of behind-the-meter photovoltaic (PV) generation data, addressing the challenge of limited visibility from net smart meters. The study evaluates model performance across four deployment regimes (known households, new households, future periods, and new regions) using Ausgrid and Pecan Street datasets for cross-dataset adaptation. Results show significant performance variations across regimes, with the lowest error observed for known h

---

Background and Motivation

Smart meters provide only net electricity data, lacking direct visibility into behind-the-meter photovoltaic (PV) generation, posing challenges for grid operators. Deep learning offers a promising solution, but its applicability across different scenarios requires thorough investigation.

Methodology and Experiments

The study employs deep learning models for PV disaggregation and evaluates their performance across four deployment regimes:

1. Known Households: The model has seen the household data during training.
2. New Households: The model needs to process data from unseen households.
3. Future Periods: The model predicts PV generation for future time periods.
4. New Regions: The model is transferred across geographical regions, adapting to different data characteristics.

The experiments use Ausgrid smart meter data for training and Pecan Street data for cross-dataset adaptation. The evaluation is conducted with a fixed 48-hour context length, focusing on daytime performance.

Key Findings

1. Significant Performance Variations: The model's mean absolute error (MAE) is 0.104 kW for known households in the test period, rising to 0.185 kW for unseen households.
2. Deployment Regime Dominance: The deployment regime has a far greater impact on error than the choice of model under a fixed training protocol.
3. Effective Fine-Tuning: Fine-tuning on a small number of target-region households reduces transfer error by 28%-36%.
4. Long Input Windows: Long input windows benefit known household extrapolation but show varying benefits for new households depending on the model.

Engineering Trade-offs and Recommendations

1. Regime-Specific Error Budgets: Error budgets should be tailored to the deployment regime.
2. Limited Target Supervision: For new region deployments, limited target supervision is recommended to enhance model adaptability.
3. Context Length Selection: The choice of context length should balance computational cost and performance based on the specific use case.

Conclusion

This study provides a comprehensive performance evaluation and deployment guidelines for deep learning-based PV disaggregation, highlighting the critical role of deployment regimes and suggesting future research directions such as cross-region transfer and long-term prediction.

---

Source: ArXiv Machine Learning (cs.LG) (2026-10-12)

— END —

Tags: #Deep Learning Follow #Photovoltaic Follow #Smart Meters Follow #Data Disaggregation Follow #Model Transfer Follow