Cabin: A collaborative and adaptive framework for wind power forecasting integrating ambient variables

Accurate wind power forecasting is paramount for maintaining grid stability and facilitating the efficient integration of renewable energy. However, the inherent variability of wind patterns and their complex dependencies on meteorological conditions pose significant forecasting challenges. This paper introduces Cabin, a novel framework designed for enhanced wind power prediction by effectively integrating historical wind power data with ambient variables such as temperature, wind speed, and direction. Cabin's architecture features two key modules: an Ambient Representation Module (ARM) for extracting multi-dimensional, context-aware features, and a Collaboration of Ambient Variables (CAV) module that synergistically integrates these features using temporal convolutions and Kolmogorov–Arnold Networks (KAN) for adaptive non-linear modeling. This collaborative and adaptive design allows Cabin to handle heterogeneous data fusion and accommodate varying data completeness through three distinct configurations. Comprehensive evaluations on two public benchmark wind power datasets (TWPF and GWPF) demonstrate Cabin's consistent superiority over a comprehensive suite of 34 state-of-the-art baseline models, achieving significant improvements in forecast error metrics, including reductions in Mean Squared Error (MSE) by up to 48.63% and Coefficient of Variation of Root Mean Squared Error (CV-RMSE) by up to 28.33% compared to the least performant baseline. These substantial gains are statistically validated by Diebold–Mariano tests, and new experiments further confirm Cabin's robust performance on hourly resolution data, underscoring its efficacy as a powerful tool for advancing predictive reliability in renewable energy systems.