Accurate short-term wind power forecasting is critical for grid stability and electricity trading. While transformer-based deep learning is state of the art, industry practice relies mostly on local models trained for individual assets. This study investigates whether a global temporal fusion transformer (TFT), trained jointly on a heterogeneous portfolio of wind farms across Germany, outperforms site-specific local models in 48-hour-ahead forecasting. Using numerical weather prediction (NWP) data from the regional ICON-D2 and the global ECMWF HRES models, we compare the global TFT against local models and a linear regression baseline. The global model leverages static covariates, such as farm age, altitude above sea level, hub height, rotor diameter and power curve parameters, to learn shared representations across sites. We evaluate on 100 synthetic wind farm profiles and a real portfolio of 11 German wind farms, across two seasonal test periods with intermediate retraining. The global TFT outperforms the local models for 87% of the synthetic and 77% of the real wind farms and clearly surpasses the linear baseline. Wilcoxon signed-rank tests confirm that these gains are significant across both seasons and datasets. Crucially, this superiority depends on static covariates: without asset-specific context, the global model fails to generalize. The TFT's variable selection network (VSN) reveals the learned feature selection. Altitude emerges as a relevant static discriminator on both datasets, while hub height is decisive on the real portfolio. For dynamic inputs, the encoder prioritizes historical power, whereas the decoder combines upper-level wind speeds from both NWP models, consistent with averaging out model-specific errors. These findings support a shift from single-site to scalable multi-site forecasting, reducing operational complexity while improving accuracy in large portfolios.
Time series forecasting is shifting from locally trained, per-series models to global models trained across many heterogeneous series, which requires learning what sets the series apart. When the raw data cannot be centralized, such a global model is trained with federated learning (FL), one client per data owner. State-of-the-art forecasters such as the Temporal Fusion Transformer (TFT) tell series apart through a static input that describes each series and does not change over time. In the special case of a federated setting where each client holds a single series, this static input is constant within each client and varies only across clients. We ask whether a federated TFT can learn from it. Per-batch FedAdam, which aggregates gradients after every batch, uses the static input almost as well as central training, whereas communication-efficient schemes with local epochs (FedAvg, FedAdam+Prox) extract far less from it and gain nothing significant on either application dataset. At the same local computation, an Adam-compatible SCAFFOLD keeps the static pathway responsive on wind power and gains significantly there. On building energy its error still drops, but the static pathway falls silent, so drift and communication frequency cannot be fully separated as the cause. Only per-batch FedAdam matches or beats the per-client local models across all datasets. We confirm the effect on building-energy meter data, a wind-power dataset, and a controlled synthetic dataset. The loss is not specific to the TFT: on the synthetic and the wind data two further architectures with different static-conditioning mechanisms show it as well.