An Improved iTransformer for Day-Ahead Photovoltaic Power Forecasting Based on Frequency Domain and Periodic Modeling

Abstract Photovoltaic (PV) power day-ahead forecasting is crucial for grid dispatch and energy management, yet its accuracy is severely challenged by the nonstationarity and multiscale periodicity of the power series. To address the limitations of existing transformer models in explicit periodic modeling and handling nonstationary disturbances, this paper proposes an enhanced iTransformer model that integrates frequency-domain decoupling and periodic perception mechanisms. The framework features two key contributions. First, a multiscale contextual temporal query attention mechanism is introduced to enhance the explicit characterization of periodic structures through learnable cycle-aware queries and contextual convolution. Second, an attention-enhanced frequency-informed normalization module is incorporated into the framework to decouple the series into periodic and residual components via Fourier transform. By employing cross-attention for feature enhancement and fusion, this module improves robustness against nonstationary disturbances, such as abrupt weather changes. Extensive day-ahead forecasting experiments on multiple real-world PV plant datasets demonstrate that the proposed model achieves competitive performance against various baseline models across mean squared error, mean absolute error, and weighted absolute percentage error metrics. These results validate the effectiveness and robustness of the proposed approach in modeling complex and nonstationary PV power generation patterns.