Wind speed prediction of complex terrain based on multi-dimensional time series decomposition

Complex terrain areas have fast wind speed changes and strong destructive power, posing a massive threat to wind power generation, transportation, and production safety. However, due to the local, random, intermittent, and fluctuating nature of wind speed changes in complex terrain areas, this poses a challenge for accurate local wind speed prediction. In this paper, we propose a Decomposition-based Multiple Sparse Autoencoders enhanced Kolmogorov–Arnold Networks, so-called DMeKAN scheme, aiming to resolve the challenge of small-scale regional wind speed prediction. Specifically: 1) by introducing a multi-scale sparse autoencoder, the model effectively captures fluctuation characteristics of wind speed and related meteorological factors across multiple frequency domains; 2) an enhanced Kolmogorov–Arnold network (eKAN) is designed to explicitly model complex nonlinear relationships among multivariate variables, incorporating a temporal-aware mechanism to enhance responsiveness to dynamic changes. Furthermore, the model has been deployed on a Linux-based system, enabling real-time sensor data acquisition and synchronous training to achieve high-precision online prediction of wind speed for the upcoming hour. Experiments were carried out based on real complex terrain datasets, and the results show that the mean square error (MSE) of the proposed model in multivariate time series prediction is reduced by 16.85%, and its performance is significantly better than that of the baseline methods.