融合时频分解与通道交互感知的多变量光伏功率预测OA

To improve the accuracy of multivariate photovoltaic(PV)power prediction,this study proposed a multivariate PV power prediction model based on time-frequency decomposition and channel interaction awareness.Aiming at the issue that exi-sting time-series decomposition methods relying on basic moving average kernels struggle to handle the nonlinear structures and complex trends of PV power data,it designed a dual decomposition mechanism integrating the time domain and frequency do-main to enhance the modeling capability for non-stationary sequences.To overcome the limitation that channel-independent methods ignored the potential correlations among multiple variables,it constructed a channel interaction-aware method.In ad-dition,addressing the shortcomings of traditional PV power prediction-such as neglecting the differences in time-step weights,changes in long-short-term correlations,and time-dependent features-it introduced a joint loss function.This func-tion combined mean squared error(MSE),signal attenuation loss,and first-order difference loss using an adaptive weight ad-justment scheme.Experiments on four actual PV datasets show that,compared with the optimal benchmark model,the pro-posed model reduces the MSE and mean absolute error(MAE)by an average of 5.59％and 5.01％,respectively,with the maximum reductions reaching 7.40％and 8.80％.The results demonstrate that the model significantly improves prediction ac-curacy and mitigates the cumulative effect of errors in the time dimension.