Aiming at the problems of strong periodicity, high volatility and low prediction accuracy of regional power load data, a non-stationary load forecasting method combining optimized variational mode decomposition(VMD) and Informer-Bidirectional Long Short-Term Memory (Informer-BiLSTM) is proposed. By introducing the crested porcupine optimizer (CPO), the number of modes and weight coefficients of VMD are optimized, and the complex load time series is effectively decomposed into multiple intrinsic mode functions to extract key time-frequency features. Subsequently, a parallel prediction model was constructed using Informer and BiLSTM to accurately predict each component after decomposition, and an integrated algorithm was introduced to further reduce the prediction error. The experimental results show that the prediction accuracy is significantly improved compared with other combined prediction models.