CIPformer: an intra-day PV power forecasting model integrating adaptive time-series decomposition and channel-interaction-aware relational modeling

Abstract

To improve the accuracy of intra-day PV power forecasting, this paper proposes CIPformer (Channel Interaction Perception-former), an enhanced iTransformer model that integrates neural STL-inspired learnable decomposition, label-dependency modeling, and joint time–frequency optimization. Firstly, a self-learning neural STL-inspired learnable decomposition is developed to adaptively extract trend, seasonal, and residual components from photovoltaic (PV) power sequences using multi-scale depthwise separable convolutions. These local representations complement the global modeling capability of iTransformer. Secondly, a channel interaction perception attention mechanism is introduced to capture dependencies among variables with different physical attributes. This mechanism combines hybrid multi-head attention with pairwise modulation to enhance variable-specific interaction modeling. Finally, a joint time–frequency loss is constructed based on the fast Fourier transform to compensate for the limited sensitivity of purely time-domain losses to periodic phase and amplitude variations. Experiments on two real-world datasets demonstrate that the proposed CIPformer consistently outperforms the baseline model. Specifically, compared with its iTransformer backbone, CIPformer reduced RMSE by 34.5 and 15.6%, decreased MAE by 38.4 and 21.4%, and lowered MAPE by 16.8 and 15.4% on the two datasets, respectively.