A fine-grained frequency decomposition framework for long-term photovoltaic and wind power forecasting

Abstract

Accurate forecasting of solar and wind energy is critical for achieving efficient grid integration. However, existing machine learning and deep learning methods face several challenges when handling complex and varying time series data, such as limited universality, insufficient generalization, and difficulty balancing computational efficiency and prediction accuracy. To address these challenges, this study proposes a fine-grained frequency decomposition framework (FDF) and designs a sequence decomposition scheme based on wavelet transform and down-sampling strategy (continuous sampling and interval sampling). The framework aims to deeply explore intricate temporal patterns in time series and fully capture long-range dependencies. More concretely, FDF applies wavelet transform to break down the original time series into multiple components of different frequencies. Then, each component undergoes continuous and interval sampling separately, which further extracts short-term fluctuations and long-term trends within each frequency band. Extensive experiments were conducted on two wind power datasets and one photovoltaic dataset for long-term forecasting. Results indicate that FDF achieved average reductions of 11.42 % in MSE and 7.65 % in MAE, with an average of 0.0015 G FLOPs and 0.2873 M parameters. It not only demonstrated excellent predictive performance and generalization capability but also outstanding lightweight characteristics.