FreeGNN: Continual Source-Free Graph Neural Network Adaptation for Renewable Energy Forecasting

Abderaouf Bahi; Amel Ourici; Ibtissem Gasmi; Aida Derrablia; Warda Deghmane; Mohamed Amine Ferrag

cs.LGMar 2, 2026

FreeGNN: Continual Source-Free Graph Neural Network Adaptation for Renewable Energy Forecasting

Authors: Abderaouf Bahi, Amel Ourici, Ibtissem Gasmi, Aida Derrablia, Warda Deghmane, Mohamed Amine Ferrag

Organizations: Computer Science and Applied Mathematics Laboratory (LIMA), Faculty of Science and Technology, Chadli Bendjedid University, P.O. Box 73, El Tarf 36000, Algeria · Mathematical Modeling and Numerical Simulation Laboratory (LAM2SIN), Faculty of Technology, Badji Mokhtar University, P.O. Box 12, Annaba 23000, Algeria · College of Information Technology, United Arab Emirates University, Al Ain, United Arab Emirates

Abstract

Accurate forecasting of renewable energy generation is essential for efficient grid management and sustainable power planning. However, traditional supervised models often require access to labeled data from the target site, which may be unavailable due to privacy, cost, or logistical constraints. In this work, we propose FreeGNN, a Continual Source-Free Graph Domain Adaptation framework that enables adaptive forecasting on unseen renewable energy sites without requiring source data or target labels. Our approach integrates a spatio-temporal Graph Neural Network (GNN) backbone with a teacher--student strategy, a memory replay mechanism to mitigate catastrophic forgetting, graph-based regularization to preserve spatial correlations, and a drift-aware weighting scheme to dynamically adjust adaptation strength during streaming updates. This combination allows the model to continuously adapt to non-stationary environmental conditions while maintaining robustness and stability. We conduct extensive experiments on three real-world datasets: GEFCom2012, Solar PV, and Wind SCADA, encompassing multiple sites, temporal resolutions, and meteorological features. The ablation study confirms that each component memory, graph regularization, drift-aware adaptation, and teacher--student strategy contributes significantly to overall performance. The experiments show that FreeGNN achieves an MAE of 5.237 and an RMSE of 7.123 on the GEFCom dataset, an MAE of 1.107 and an RMSE of 1.512 on the Solar PV dataset, and an MAE of 0.382 and an RMSE of 0.523 on the Wind SCADA dataset. These results demonstrate its ability to achieve accurate and robust forecasts in a source-free, continual learning setting, highlighting its potential for real-world deployment in adaptive renewable energy systems. For reproducibility, implementation details are available at: https://github.com/AraoufBh/FreeGNN.

Figures & tables

Panel of Table 1 — Table 1: State of the art summary

Fig. 2 : Spatio-temporal graph forecasting overall procedure

Dataset	#Sites	Samples	Duration	Freq.	Power Vars	Met. Vars
GEFCom2012	7 farms	$\sim$ 17,500/site	2 years	1h	1	4 (u, v, ws, wd)
Solar PV	2 plants	$\sim$ 3,200/plant	34 days	15 min	2 (DC, AC)	4–6
Wind SCADA	1 turbine	$\sim$ 52,000	1 year	10 min	1	2 (speed, direction)

Table 2: Summary of renewable energy datasets used in our experiments

Layer	Type	Dimension	Activation function	Hyperparameters
Input	Embedding nodes	128	–	–
1	Graph convolutional layer	64	ReLU	Learning rate: 0.01
2	Graph convolutional layer	64	ReLU	Dropout: 0.5
3	Attention layer	64	LeakyReLU ( $\alpha=0.2$ ) Attention heads: 4
Output	Fully connected	1	Sigmoid	–

Table 3: GNN’s architecture

Panel of Figure 6 — Fig. 3 : Baselines classification by category

	GEFCom				SolarPV				Wind SCADA
Model	MAE	RMSE	MAPE	sMAPE	MAE	RMSE	MAPE	sMAPE	MAE	RMSE	MAPE	sMAPE
GCN	6.432	8.512	8.341	8.604	1.412	1.923	5.241	5.534	0.462	0.631	6.312	6.584
GAT	6.118	8.247	8.021	8.295	1.326	1.812	4.923	5.207	0.441	0.603	6.118	6.401
GraphSAGE	5.987	8.031	7.812	8.087	1.298	1.734	4.812	5.094	0.432	0.592	6.021	6.287
SHOT	5.873	7.924	7.562	7.856	1.241	1.689	4.612	4.893	0.421	0.571	5.923	6.198
SFDA	5.812	7.842	7.498	7.784	1.214	1.643	4.534	4.812	0.409	0.562	5.842	6.134

Table 4: Performance comparison of FreeGNN against baselines. (The best result is bolded and the runner-up is underlined)

Dataset	Model	MAE	RMSE	MAPE	sMAPE
GEFCom	FreeGNN	5.237	7.123	6.842	7.018
	FreeGNN@1	6.143	8.012	8.034	8.203
	FreeGNN@2	5.812	7.821	7.523	7.719
	FreeGNN@3	5.623	7.512	7.234	7.412
	FreeGNN@4	5.894	7.893	7.612	7.812
SolarPV	FreeGNN	1.107	1.512	4.234	4.521

Table 5: Comparison of FreeGNN’s models

Fig. 4 : Ablation study results visualization

Fig. 5 : Performance comparison of evaluation metrics at all FreeGNN’s models

Panel of Figure 11 — Fig. 6 : Sensitivity of MAE to the loss weights $\lambda_{\text{graph}}$ and $\lambda_{\text{consistency}}$ .

Dataset	0	50	100	200	400
GEFCom2012 (Wind)	0.194	0.171	0.158	0.148	0.146
Solar PV (India)	0.128	0.117	0.110	0.106	0.104
Wind SCADA (Turkey)	0.165	0.152	0.143	0.136	0.134

Table 6: Impact of memory buffer size

|\mathcal{M}|

on MAE (lower is better).

Fig. 7 : Projection of node embeddings over time

Fig. 8 : Temporal evolution of node embeddings divided into consecutive time windows

Dataset	$h=1$	$h=3$	$h=6$	$h=12$	$h=24$
GEFCom2012 (Wind)	0.089	0.118	0.151	0.198	0.264
Solar PV (India)	0.061	0.084	0.116	0.162	0.228
Wind SCADA (Turkey)	0.075	0.099	0.132	0.179	0.241

Table 7: Effect of forecasting horizon

h

(MAE ↓).

Dataset	Time (ms/batch)	Memory (MB)
GEFCom2012 (Wind)	24	0.82
Solar PV (India)	18	0.48
Wind SCADA (Turkey)	15	0.36

Table 8: Average time per batch on GPU and memory footprint of the replay buffer

Source-Free Domain Adaptation

Source-Free Domain Adaptation transfers models to new data distributions without access to the original training examples.

See all 69 papers

Towards Continuous Power Forecasting: Practical Continual Learning for Real-World Energy Systems in Nonstationary Time Series
Jun 23, 2026Yujiang He, Frederic Uhrweiller, Bernhard SickContinual Learning Time Series Forecasting
Intelligent Embedded Systems, University of Kassel, 34121 Kassel, Germany
R $^2$ Energy: A Large-Scale Benchmark for Robust Renewable Energy Forecasting under Diverse and Extreme Conditions
Feb 17, 2026Zhi Sheng, Yuan Yuan, Guozhen Zhang +1Forecasting Benchmarks Time Series Forecasting
Center for Urban Science and Computation, Tsinghua University · Courant Institute of Mathematical Sciences, New York University
GraphToolbox: A Configurable Python Framework for Graph Neural Network Forecasting
Sep 21, 2026Eloi Campagne, Yvenn Amara-Ouali, Yannig Goude +1Time Series Forecasting Electricity Load Forecasting
CB · LMO, CELESTE · EDF R\&D +1

Jun 23, 2026J/K move · Enter open · S save

Towards Continuous Power Forecasting: Practical Continual Learning for Real-World Energy Systems in Nonstationary Time Series

Yujiang He, Frederic Uhrweiller, Bernhard Sick

Intelligent Embedded Systems, University of Kassel, 34121 Kassel, Germany

Continual Learning Time Series Forecasting Non-Stationary Time Series Forecasting Catastrophic Forgetting

Open paper AskPDF arXiv

Power forecasting models deployed in real-world energy markets must operate under nonstationary conditions, where data distributions continually evolve due to weather variability, infrastructure upgrades, and changing consumption behaviors. In practice, these models face strict operational constraints: historical data may be limited or unavailable for repeated retraining, and uninterrupted long-term service is often required. This paper addresses these challenges by proposing the paradigm of Continuous Power Forecasting, which views power forecasting as a continual learning problem rather than a static offline task. Based on an adaptive continual learning framework for regression, we systematically investigate the practical effectiveness of six representative continual learning approaches from three methodological categories. These approaches are evaluated under different realistic assumptions regarding data accessibility and update policies. Experimental validation on real-world power datasets demonstrates that continual learning enables forecasting models to self-adapt to distributional drift, accumulate knowledge over time, and mitigate catastrophic forgetting without relying on large-scale historical data storage. Beyond performance gains, our study provides practical insights into the stability and adaptation behaviors of different continual learning approaches under realistic operational constraints. Overall, this work illustrates how continual learning can be pragmatically integrated into industrial power forecasting pipelines, offering a scalable and sustainable solution for long-term deployment in dynamic environments.

Abstract

Source-Free Domain Adaptation

Explore similar work

Towards Continuous Power Forecasting: Practical Continual Learning for Real-World Energy Systems in Nonstationary Time Series

R2^22Energy: A Large-Scale Benchmark for Robust Renewable Energy Forecasting under Diverse and Extreme Conditions

GraphToolbox: A Configurable Python Framework for Graph Neural Network Forecasting

R $^2$ Energy: A Large-Scale Benchmark for Robust Renewable Energy Forecasting under Diverse and Extreme Conditions