融合特征聚类与SFOA–TimeMixer的短期光伏功率预测

Short-Term Photovoltaic Power Prediction Based on Feature Fusion Clustering and Superb Fairy-wren Optimized TimeMixer

  • 摘要: 针对光伏短期功率预测中因气象条件非线性变化引起的预测误差显著增大的问题,提出了融合特征聚类与SFOA–TimeMixer的短期光伏功率预测方法. 首先,利用互信息(Mutual information, MI)筛选出关键气象特征. 其次,为降低数据异质性,提出特征聚类算法(Fusion multi-view hierarchical clustering, FMVHC),该算法通过变分模态分解(Variational mode decomposition, VMD)完成特征重构,并融合欧氏距离、余弦相似度和皮尔逊相关系数构建综合距离矩阵,改进层次聚类(Hierarchical clustering, HC),将数据划分为晴天、阴天和雨天三类. 随后,为强化模型的自适应建模能力,采用壮丽细尾鹩莺优化算法(Superb fairy-wren optimization algorithm, SFOA)分别对VMD参数和TimeMixer模型的超参数进行双阶段寻优. 最后,使用SFOA–TimeMixer模型对三类天气的数据集分别进行预测,以更好地捕捉复杂时序特征,提升预测精度. 基于江苏省某光伏电站2022年数据验证,本文所提方法在三类典型天气条件下均优于其他对比模型,显著提高了模型预测精度.

     

    Abstract: With the rapid development of renewable energy and the increasing integration of photovoltaic (PV) power generation into modern power systems, accurate short-term PV power forecasting has become essential for ensuring grid stability, improving energy utilization efficiency, and facilitating intelligent energy management. However, PV power output is affected by meteorological conditions, resulting in randomness, intermittency, and nonlinear fluctuations. The heterogeneity of PV generation data under different weather conditions makes it difficult for conventional forecasting models to achieve satisfactory accuracy and robustness. To address these challenges, this study proposes a short-term PV power forecasting framework that integrates fusion multi-view hierarchical clustering and the superb fairy-wren optimization algorithm (SFOA)-based TimeMixer model (FMVHC–SFOA–TimeMixer). First, mutual information was employed to evaluate the nonlinear dependency between meteorological variables and PV power output, thereby selecting the most informative meteorological factors and reducing redundant inputs. Subsequently, a fusion multi-view hierarchical clustering (FMVHC) method was developed to divide the historical PV data into different weather categories. In this method, variational mode decomposition (VMD) was introduced to decompose daily meteorological sequences into multiple intrinsic mode functions (IMFs), and the informative components are reconstructed to enhance feature representation while reducing noise interference. Furthermore, a multi-view distance measurement strategy combining the Euclidean distance, cosine similarity, and Pearson correlation coefficient was adopted to construct a comprehensive distance matrix based on which, hierarchical clustering was performed to improve the discrimination capability for different weather patterns, including sunny, cloudy, and rainy conditions. To further improve forecasting adaptability and reduce the influence of manual parameter selection, a dual-stage optimization strategy based on the SFOA is proposed. In the first stage, the SFOA was utilized to optimize the key parameters of the VMD including the number of decomposition modes and the penalty factor, ensuring the effective extraction of multiscale meteorological features. In the second stage, the SFOA was applied to optimize the critical hyper parameters of the TimeMixer forecasting model. Via this adaptive optimization mechanism, the proposed framework could automatically obtain more suitable parameter combinations for different weather conditions. The optimized TimeMixer model was adopted as the final forecasting module. Owing to its multiscale mixing architecture, TimeMixer model can effectively capture complex temporal dependencies, separate trends and seasonal information, and extract short-term fluctuation characteristics from PV power sequences. The proposed method was validated using measured operational data collected from a photovoltaic power station in Jiangsu Province, China. Experimental results demonstrate that the proposed FMVHC–SFOA–TimeMixer framework achieves superior forecasting performance under three typical weather conditions. Specifically, under sunny conditions, the proposed model achieved an RMSE, MAE, and R2 value of 4.0957 MW, 2.8125 MW, and 0.9752, respectively. Under cloudy conditions, the RMSE, MAE, and R2 reached 6.7490 MW, 4.9022 MW, and 0.9427, respectively. Under rainy conditions, the corresponding values were 6.0450 MW, 4.6174 MW, and 0.9216, respectively. Compared with other benchmark models, the proposed approach consistently reduces prediction errors and improves forecasting stability. The results indicate that the combination of weather pattern–oriented feature clustering, adaptive parameter optimization, and multiscale temporal modeling effectively enhances the capability of PV power forecasting under complex meteorological environments. The proposed framework provides a reliable and practical solution for short-term PV power prediction and offers valuable technical support for PV operation management and renewable energy integration.

     

/

返回文章
返回