由于光伏发电具有间歇性和波动性，给电网运行的安全性和稳定性造成危害，对光伏功率进行准确预测可以有效解决这一问题。本文提出一种基于STLFormer的中短期光伏功率预测模型，该模型结合了季节趋势局部加权回归分解（STL分解）与神经网络模型。首先，STLFormer模型将光伏功率数据通过STL分解进行特征扩充，用于提取基于历史序列的周期项、趋势项特征。然后，拼接周期项、趋势项特征和原特征，进行数据预处理和特征编码并使用基于Informer模型的神经网络进行功率预测。最后，在真实数据集上进行大量实验。实验结果表明：STLFormer在中短期光伏功率预测任务中精度较高，其中在2 h光伏功率预测任务时，平均绝对值误差为0.176、均方误差为0.180；在28 h光伏功率预测任务时，平均绝对值误差为0.170、均方误差为0.154。

The intermittency and volatility of photovoltaic (PV) power generation poses risks to the safety and stability of power grid operations. Accurately predicting PV power can effectively address this issue. Therefore, this paper proposed a shortand mediumterm PV power prediction model based on STLFormer, which combined seasonal and trend decomposition using locally weighted regression (STL) and a neural network model. First, the STLFormer model utilized STL decomposition to expand the PV power data, extracting features based on historical sequences such as periodic and trend components. Then, the periodic and trend component features were concatenated with the original features, the data preprocessing and feature encoding were performed, and the power prediction was conducted using a neural network based on the Informer model. Finally, many experiments were carried out on real data sets. The experimental results show that STLFormer has high accuracy in shortand mediumterm photovoltaic power forecasting tasks. Among them, the average absolute value error is 0.176 and the mean square error is 0.180 in the task of PV power forecasting for 2 hours; the average absolute value error is 0.170 and the mean square error is 0.154 in the task of PV power forecasting for 28 hours.

TP391

山西省关键核心技术和共性技术研发攻关专项项目（2020XXX007）；山西省重点研发计划项目（202102020101006）