KAPNet: A physics-informed neural network with physical constraints modeled by Kolmogorov–Arnold network for wind speed prediction

Reliable wind speed prediction plays a crucial role in facilitating the efficient dispatch of wind energy, while its potential is limited by the inadequate integration of physical mechanisms in existing data-driven models. Accordingly, this paper develops a Kolmogorov–Arnold network (KAN) enabled physics-informed neural network (PINN) for wind speed prediction, called KAPNet, which incorporates the wind speed temporal evolution dynamics modeled by KAN into PINN. First, a predictor with a hierarchical architecture is designed to generate an initial wind speed prediction from the raw inputs comprising multivariate historical meteorological features and encoded target time. Second, the automatic differentiation mechanism is introduced to compute the partial derivatives of the initial prediction with respect to both the historical meteorological features and the target time. Third, a partial differential equation (PDE) modeled by KAN is constructed to dynamically capture the temporal evolution dynamics of wind speed by leveraging the initial prediction and the associated physical information. Finally, the KAN-modeled PDE serves as a physical constraint within PINN to dynamically refine the initial prediction into the final prediction, thereby enhancing both prediction precision and robustness. Two geographically contrasting datasets from both offshore and inland wind farms are employed to validate the effectiveness of KAPNet. The results demonstrate the superiority of KAPNet in prediction accuracy and stability over several state-of-the-art methods.