针对光伏发电功率因天气变化导致的间歇性和波动性从而导致预测精度较低的问题，提出一种基于自组织映射（self-organizing map，SOM）网络和K均值（K-means）算法（S-Kmeans）的协同聚类、改进的人工旅鼠算法（improved artificial lemming algorithm，IALA）优化的变分模态分解（variational mode decomposition，VMD）与时序卷积网络（temporal convolutional networks，TCN）-双向门控循环单元（bidirectional gated recurrent unit，BiGRU）联合构建的多层级短期光伏功率预测方法。首先，根据相关性分析选出关键气象因子，采用S-Kmeans的协同聚类将光伏数据划分为晴天、多云、阴雨3种典型天气类型。在此基础上，利用IALA对VMD参数组合进行自适应优化，实现光伏功率序列的最优分解，从而更好地捕捉信号的局部特征。最后，对每个子序列构建TCN-BiGRU模型，通过分量预测与全局重构得到预测结果，提升预测精度。实验结果表明，提出的优化模型在各项性能指标上均优于对比模型，验证了其在提高短期光伏功率预测精度方面的有效性。

To address the issue of low prediction accuracy in photovoltaic power generation caused by the intermittency and volatility resulting from weather changes， this paper proposes a multi-level short-term photovoltaic power forecasting method. The method is based on collaborative clustering using self-organizing map and K-means algorithm （S-Kmeans）， and an improved artificial lemming algorithm （IALA）-optimized variational mode decomposition （VMD）， combined with a temporal convolutional network （TCN） and bidirectional gated recurrent unit （BiGRU）. First， key meteorological factors are selected through correlation analysis， and photovoltaic data is classified into three typical weather conditions - sunny， cloudy， and rainy by using the S-Kmeans co-clustering method. Then， the IALA is employed to adaptively optimize the VMD parameters， enabling optimal decomposition of the photovoltaic power series and capturing local signal features more effectively. Finally， a TCN-BiGRU model is constructed for each subsequence， and the prediction result is obtained through component forecasting and global reconstruction， thereby improving prediction accuracy. Experimental results show that the proposed model outperforms the comparison models across all performance metrics under various weather conditions， validating its effectiveness in short-term photovoltaic power forecasting.