Hybrid Kernel Function Enhanced Deep Gaussian Process for Wind Power Prediction

As the global demand for renewable energy continues to grow, wind power, as a clean and sustainable form of energy, the accuracy of its generation power prediction is of great significance for grid scheduling and energy management. In this paper, a probabilistic prediction method based on Improved Deep Gaussian Process (I-DGP) is proposed. The hybrid kernel function is constructed, combines Matérn and RQ kernel functions to accommodate multi-timescale features. Model-based sequential optimization (SMBO) is also applied to optimize the hyperparameters of I-DGP, and a three-dimensional search space covering the parameters of learning rate, maximum number of iterations, and weights of kernel functions is constructed, and the search process is guided by the agent model and expectation improvement function. The performance of the proposed method is systematically verified from the aspects of algorithm effectiveness, optimization strategy and kernel function mechanism through multi-dimensional experiments and three evaluation indexes.