Spectral-aware decomposition with multi-scale temporal modeling for photovoltaic power forecasting

Abstract

Photovoltaic (PV) power forecasting plays a critical role in power system operation and renewable energy integration, and has attracted increasing research attention. Owing to their strong temporal modeling capability, deep learning models have achieved promising performance in PV power prediction. However, existing methods often struggle to effectively characterize the complex temporal dynamics caused by the coexistence of long-term trends and short-term fluctuations in PV power series. In addition, most models treat PV power signals as a single entangled sequence, which limits their ability to capture heterogeneous temporal patterns at different scales. To address these challenges, we propose a spectral-aware decomposition with multi-scale temporal modeling (SD-MTM) framework for PV power forecasting. We aim to apply spectral-aware decomposition to the entangled sequences based on the frequency-domain spectral amplitude distribution and obtain the trend and residual parts. Shifting from the unified modeling to the differentiated modeling, we leverage a covariate-aware module to capture global evolving patterns. Meanwhile, local stochastic dynamics are captured by a multi-scale architecture fused with an adaptive mixture-of-experts mechanism. Experimental results across eight real-world datasets demonstrate that SD-MTM can capture both local fluctuations and global trends within power patterns, achieving the lowest MAE in 7 out of 8 stations and the lowest RMSE in 6 out of 8 stations, with a mean MAE of 1.4318 compared with 1.4773 for the best-performing competing baseline selected separately for each station, corresponding to a 3.1% improvement.