The increasing integration of renewable energies into modern electricity grids creates new challenges for grid operators and energy traders. Wind and solar energy are highly dependent on the weather and are therefore particularly characterized by spatial and temporal fluctuations. Therefore, precise and reliable power forecasts for hours and days into the future are essential for grid stability, energy trading, and cost reduction. Moreover, the increasing availability and quality of power measurements enables a variety of methods to adapt weather and power forecasts to these measurements, with combining these diverse approaches often leading to improved accuracy.

This work investigates an attention-based approach for the combination of power forecasts, in which soft-weights are dynamically assigned to individual forecasting models through a learned attention mechanism. In contrast to traditional methods like linear regression models, an attention mechanism computes context-sensitive weights at inference time, allowing the combination to respond dynamically to the specific characteristics of each forecast situation. Additionally, the learned weighting structure captures non-linear dependencies between base forecasters, which static combination rules cannot express.Unlike conventional combination approaches that rely on repeated retraining to keep model weights aligned with recent observations, the proposed method follows a semi-adaptive design: once trained, the attention mechanism adapts to the forecast combination implicitly through its learned weighting structure, without requiring continuous retraining cycles. This property could offer considerable practical advantages in operational deployment, as it reduces maintenance overhead and simplifies integration into existing forecasting pipelines. Whether this semi-adaptive, attention-based soft-weight combination can match or surpass the forecast quality of retraining-based baselines across both weather and power forecast settings is the central question this work seeks to explore.