基于增量贝叶斯随机配置网络的漂移环境概率风力预测

Standard data-driven probabilistic wind power prediction solutions typically assume a static exogenous environment, where historical data contains all information desired to predict the future. Unfortunately, the practical industrial data distribution is time-varying and unforeseen, known as concept drift (CD), violating the assumption. To address this challenge, an incremental Bayesian stochastic configuration network is proposed to enhance prediction accuracy under CD environments. Concretely, Stochastic Configuration Network (SCN), a lightweight and iteration-free random weight neural network, is explored to model latent relationships between variables and targets. Subsequently, Bayesian inference updates SCN's output layer parameters, approximating the parameter posterior distribution and generating the probabilistic prediction model BSCN. To manage CD, Maximum Mean Discrepancy and Continuous Ranked Probability Score are employed to detect virtual and real drifts. The real drift serves as a trigger for the incremental learning of the BSCN, with a specifically designed incremental update strategy to facilitate this process. Extensive experiments demonstrate that the proposed method can continuously learn new modes without forgetting and significantly improve probabilistic wind power prediction accuracy in dynamic CD environments.