基于增强特征提取与新型损失函数的TimesNet光伏功率多步短期预测方法

Abstract The instability of weather conditions often causes photovoltaic power generation to exhibit randomness and volatility, making accurate and reliable photovoltaic power forecasting crucial for the stable scheduling of integrated energy systems . Multi-step forecasting remains a challenge due to the difficulty in capturing temporal dependencies among neighboring discrete time points, which is attributable to the limited expressiveness of time-series features using one-dimensional modeling methods. Hence, this paper proposes a methodological framework tailored for multi-step short-term forecasting of photovoltaic power generation . The framework is based on the TimesNet architecture, which models meteorological features in two dimensions to enhance feature expressiveness. Additionally, a new feature extraction module is introduced to replace the Inception module in the original TimesNet, mitigating issues of feature redundancy and convolution kernel sharing associated with standard convolution. This enhancement aims to improve TimesNet's ability to recognize critical information. Considering the inevitable presence of outliers in datasets and the drawbacks of traditional loss functions, which are sensitive to outliers or struggle to fit nonlinear relationships, this paper proposes a novel loss function to overcome these limitations. To validate the performance of the proposed method, it was tested on three datasets across four prediction horizons (1 h, 3 h, 6 h, and 12 h ahead). Compared to the original TimesNet, it reduces the average RMSE and MAPE by 3.21 % and 9.36 % for the 12-h prediction. Compared to LightTS, Informer, and DLinear, it reduces the average MAE by 16.45 %, 24.62 %, and 11.41 % for the 12-h prediction, respectively. The proposed loss function also outperforms traditional loss functions (MAE, MSE , Huber, Log-Cosh) with an optimal metrics rate averaging 77 %. These results demonstrate that the proposed model and loss function achieve excellent accuracy in multi-step photovoltaic power forecasting, guiding the stable integration of renewable energy into the grid.