A Behaviour-Aware Federated Forecasting Framework for Distributed Stand-Alone Wind Turbines

Bowen Li; Xiufeng Liu; Maria Sinziiana Astefanoaei

cs.LGMar 5, 2026

A Behaviour-Aware Federated Forecasting Framework for Distributed Stand-Alone Wind Turbines

Authors: Bowen Li, Xiufeng Liu, Maria Sinziiana Astefanoaei

Organizations: IT University of Copenhagen, Denmark. · Technical University of Denmark (DTU), Denmark

Abstract

Accurate short-term wind power forecasting is essential for grid dispatch and market operations, yet centralising turbine data raises privacy, cost, and heterogeneity concerns. We propose a two-stage federated learning framework that first clusters turbines by long-term behavioural statistics using Double Roulette Selection (DRS) initialisation with recursive Auto-split refinement, and then trains cluster-specific LSTM models via FedAvg. Experiments on 400 stand-alone turbines in Denmark show that DRS-auto discovers behaviourally coherent groups and achieves competitive forecasting accuracy while preserving data locality. Behaviour-aware grouping consistently outperforms geographic partitioning and matches strong k-means++ baselines, suggesting a practical privacy-friendly solution for heterogeneous distributed turbine fleets.

Figures & tables

Panel of Figure 1 — Figure 1: (a) 400 turbines selected via nearest-neighbour sampling. (b) Daily generation variation of 400 turbines over one day. (c) Federated clustering followed by cluster-specific federated LSTM forecasting.

Group	Features
Meteorological	wind_speed , wind_dir_sin/cos , temperature
Temporal	hour_sin/cos (diurnal encoding)
Static	Capacity_kw , age
Autoregressive	power_output_lag_1 to lag_24 ( $L{=}24$ h window)

Table 1: Prediction model input features.

Cluster	Training time (min)
Cluster 0	27
Cluster 1	43
Cluster 2	3
Cluster 4	45
Cluster 5	5
Cluster 6	7

Table 2: Wall-clock training time in the prediction stage (per-cluster FedAvg LSTM).

Panel of Figure 4 — Figure 4: Comparison of behaviour- and geo-based grouping strategies.

Cluster	Count	mean_power_scaled_mean	std_power_scaled_mean	cv_scaled_mean	zero_ratio_mean	ramp_mean_scaled_mean	ramp_std_scaled_mean
0	51	2.3383	2.3408	-0.1774	0.0672	-1.1569	2.3493
1	2	-0.7299	-0.8549	13.2276	0.9987	0.2984	-0.7925
2	4	1.7114	1.8085	-0.1317	0.1102	4.7725	1.6934
3	251	-0.3049	-0.2922	-0.0830	0.1465	0.0177	-0.3018
4	76	-0.6317	-0.7048	0.0607	0.3157	0.2485	-0.6684
5	6	0.5266	0.8178	-0.0903	0.0900	0.7233	0.7399

Table 3: Cluster-wise turbine counts and mean standardized feature values.

Cluster	mean_power_scaled_std	std_power_scaled_std	cv_scaled_std	zero_ratio_std	ramp_mean_scaled_std	ramp_std_scaled_std
0	0.9186	0.8308	0.0209	0.0207	1.9099	0.8828
1	0.0000	0.0018	6.5609	0.0011	0.0000	0.0083
2	1.5020	1.0891	0.0810	0.0611	1.2589	0.9328
3	0.0971	0.1045	0.0484	0.0820	0.3357	0.0973
4	0.0736	0.1091	0.1783	0.1942	0.1157	0.0919
5	0.0610	0.1199	0.0168	0.0183	0.5113	0.1440

Table 4: Within-cluster standard deviations of standardized features.

ID	Type	Size	Key Characteristics
0	High power, high variability	51	Above-avg power ( $\approx$ 2.34), low shutdown, high ramps
1	Faulty / shutdown	2	Near-zero output, $\approx$ 100% shutdown ratio (excluded)
2	Ramp-dominated	4	Extreme ramping ( $\approx$ 4.77), high heterogeneity
3	Baseline stable	251	Near-average features, high homogeneity (main group)
4	Mid-risk, low-output	76	Below-avg power, elevated shutdown ( $\approx$ 0.32)
5	Promising but volatile	6	Above-avg power, low shutdown, high variability

Table 5: Summary of cluster behavioural types from DRS-auto.

Panel of Figure 5 — Figure 5: DRS-auto clustering results.

Table 11

Method	Split	MSE	RMSE	MAE	$R^{2}$
DRS-auto (federated)	Train	0.028091	0.166485	0.122107	0.678116
DRS-auto (federated)	Test	0.028822	0.168435	0.121883	0.711436
Centralized LSTM	Train	0.006223	0.078890	0.050925	0.887562
Centralized LSTM	Test	0.006159	0.078483	0.049881	0.905096

Table 8: Comparison between federated DRS-auto and centralized LSTM on the same 51-turbine cluster. 2 2 2 Due to training-time and resource constraints, the centralized baseline was not run on all 400 turbines; instead, both models are trained and evaluated on the same representative subset of 51 turbines. DRS-auto keeps the federated constraint (multiple logical clients), whereas the centralized LSTM trains a single global model on all available data within this subset.

Panel of Figure 6 — Figure 6: Forecast metrics per cluster under DRS-auto grouping.

Table 14

Panel of Figure 7 — Figure 7: 24-hour forecast cases for one representative turbine per cluster under DRS-auto grouping.

Panel of Figure 8 — Figure 8: Cluster 4 metrics before and after cluster-specific fine-tuning.

Appendix figures & tables2 assets

Supplementary material from the paper’s appendix.

Appendix

Symbol / Term	Definition
$C_{j}$	Local cluster centers on client $j$
$S_{j}$	Number of samples assigned to each cluster on client $j$
$C_{k}^{\text{global}}$	Global cluster center $k$ after aggregation
$k_{\text{local}}$	Number of clusters per client (local KMeans)
$k_{\text{global}}$	Number of global clusters
$T_{\text{cluster}}$	Number of communication rounds in the clustering phase

Appendix

Table 11: Notation used in the proposed federated clustering and prediction framework

Param	Meaning
$\mathbf{X}$	Behavioural feature matrix (standardised).
$\mathcal{G}$	Candidate grid $(n,k,c)$ : #clients, #clusters, #comm.rounds.
$\tau_{\text{sil}}$	Silhouette threshold for accepting a split.
$\tau_{\min}$	Min cluster ratio: nodes $\leq\tau_{\min}\|\mathbf{X}\|$ become outlier leaves.
$\tau_{\text{large}}$	Oversized ratio: forces split if node $>\tau_{\text{large}}\|\mathbf{X}\|$ .
$Q$ , $C_{\text{root}}$ , $C_{\text{node}}$	BFS queue, root node, current node.

Appendix

Table 12: Key parameters for the recursive auto-split procedure.

Wind Power Forecasting

Wind power forecasting predicts turbine and farm output from operational data, weather forecasts, and remote observations to support reliable energy planning.

See all 24 papers

Initialization Is Critical: Advancing Federated Short-Term Load Forecasting under Load Heterogeneity via Model Initialization
Aug 28, 2026Jianing Chen, Vajiheh Farhadi, Yan Li +1Electricity Load Forecasting Federated Learning
School of Electrical Engineering and Computer Science, The Pennsylvania State University, University Park, PA 16802, USA · Department of Electrical and Computer Engineering, Bucknell University, Lewisburg, PA 17837, USA
Federated Weather Modeling on Sensor Data
May 1, 2026Shengchao Chen, Guodong LongSpatiotemporal Forecasting Weather Forecasting
Australian AI Institute, Faculty of Engineering and IT, University of Technology Sydney, Australia
Tyan-WP: A Wind Power Foundation Model for Ultra-Short-Term Probabilistic Forecasting
Jun 7, 2026Jiahui Huang, Ao Luo, Lei Liu +6Zero-Shot Time Series Forecasting Time Series Forecasting
School of Information Science and Technology, University of Science and Technology of China, Hefei, 230022, China · China Electric Power Research Institute, Beijing, 100192, China

Aug 28, 2026J/K move · Enter open · S save

Initialization Is Critical: Advancing Federated Short-Term Load Forecasting under Load Heterogeneity via Model Initialization

Jianing Chen, Vajiheh Farhadi, Yan Li, Thomas La Porta

School of Electrical Engineering and Computer Science, The Pennsylvania State University, University Park, PA 16802, USA · Department of Electrical and Computer Engineering, Bucknell University, Lewisburg, PA 17837, USA

Electricity Load Forecasting Federated Learning Heterogeneous Federated Learning

Open paper AskPDF arXiv

Short-term load forecasting (STLF) provides essential information for numerous applications in modern power systems. However, accurate STLF often relies on fine-grained smart-meter data from distributed users, raising increasing concerns about data privacy. Federated learning (FL) has therefore emerged as a promising privacy-preserving paradigm for STLF. Nevertheless, this paper reveals structured heterogeneity in clients' load data. Specifically, clients exhibit different responses to exogenous factors and distinct temporal load profiles, which can degrade forecasting performance in FL. To mitigate these issues, this paper studies the role of model initialization in federated STLF, and proposes two initialization strategies from global and local perspectives. For global model initialization, when auxiliary public load data are available, a pretrained initialization strategy is developed to initialize the global model before federated training, thereby reducing client drift during the training process. For local model initialization, we propose SLIAvg, a sequential local initialization strategy that promotes a more consistent training process by allowing participating clients to start from progressively adapted models within each communication round. Since the proposed strategies only modify the initialization process, they are compatible with most existing FL frameworks and privacy-enhancing techniques. Experiments on real smart-meter data with two representative forecasting architectures demonstrate that the proposed strategies effectively improve forecasting performance, as evidenced by reduced client drift, improved convergence behavior, and lower forecasting errors.

Abstract

Wind Power Forecasting

Explore similar work

Initialization Is Critical: Advancing Federated Short-Term Load Forecasting under Load Heterogeneity via Model Initialization

Federated Weather Modeling on Sensor Data

Tyan-WP: A Wind Power Foundation Model for Ultra-Short-Term Probabilistic Forecasting