Wavelet-TimesNet: Improving Long-Term Solar Power Forecasting via Adaptive Wavelet Transform and Multi-Scale Residual Networks

Unofficial AI-generated summary based on the public title and abstract. Not an official translation.

🔬研究者:Provides a novel hybrid architecture (wavelet + multi-scale CNN) that improves on Transformers for time-series forecasting, worth studying for other renewable prediction tasks.

🏢実務担当者:Directly applicable to solar farm operators and grid managers seeking higher forecast accuracy for better scheduling and reserve planning.

🏛政策担当者:Supports renewable energy integration targets by demonstrating how advanced forecasting can reduce grid balancing costs and curtailment.

Long-term photovoltaic power prediction is crucial for the optimal dispatch of energy systems and the stability of power grids. However, existing methods are limited in accuracy when dealing with non-stationary signals such as intermittent fluctuations in light due to issues like spectral leakage and the rigidity of fixed convolutional kernel feature extraction. To address these challenges, this paper proposes a novel model, Wavelet-TimesNet, which integrates adaptive wavelet transform and multi-scale residual networks, aiming to enhance the robustness of long-term predictions. This model dynamically adjusts the parameters of the wavelet basis function to achieve multi-resolution analysis of local periodic features, effectively suppressing noise interference. It constructs a multi-scale residual network to capture local details such as hourly irradiance mutations and global trends such as seasonal power variations using parallel convolutional kernels of different sizes. An adaptive wavelet attention mechanism is introduced to dynamically weight and fuse frequency-domain and time-domain features, enhancing the focus on key information. Experiments were conducted based on photovoltaic datasets from Xinjiang's temperate continental climate and a subtropical monsoon climate region in China. The results show that in 96-hour predictions, Wavelet-TimesNet reduces the mean absolute error (MAE) by 4.97% and 3.29% in Xinjiang and China, respectively, and the mean squared error (MSE) by 7.80% and 5.75%. In 192-hour predictions, the mean squared percentage error (MSPE) of the Xinjiang dataset is reduced by 36.42%. Compared with advanced models such as Transformer and Informer, this model demonstrates significant advantages in handling non-stationary signals and capturing long-term trends, especially in extreme weather scenarios like sudden sandstorms and continuous rainy days, where prediction accuracy is notably improved. The research results provide an efficient solution for the precise dispatch of photovoltaic power stations, which is of great significance for reducing peak shaving costs in power grids, promoting the consumption of renewable energy, and facilitating the transformation of energy structures.

• Zenodo https://zenodo.org/records/20456139first seen 2026-05-31 04:14:01 · last seen 2026-06-02 04:14:41