基于Informer-LSTM-MLP的短期光伏发电功率预测

Short-Term Photovoltaic Power Forecasting Based on Informer-LSTM-MLP

摘要: 针对光伏发电功率预测精度较差的问题，提出一种基于Informer-LSTM-MLP的短期光伏发电功率预测混合模型。首先，利用3δ法则、箱线图法及孤立森林法检测异常数据，结合邻近窗口均值法与三次样条插值填补缺失值。通过Pearson相关系数筛选关键气象要素，并采用改进K-means算法将历史数据聚类为晴天、阴天及复杂天气3类。在此基础上，构建Informer-LSTM-MLP混合模型：采用LSTM网络提取环境特征的时序信息，生成隐含状态矩阵；采用Informer模型对功率序列进行深层时序依赖建模，提取全局特征；将两类特征输入Cross-Attention机制进行多源信息融合，最后通过MLP实现端到端功率预测。实验结果表明，该混合模型在不同天气条件下均能有效提升光伏发电功率预测精度。

Abstract: To address the problem of low prediction accuracy in photovoltaic (PV) power forecasting, this paper proposes a hybrid short-term PV power prediction model based on Informer-LSTM-MLP. First, the 3σ criterion, box plot method, and Isolation Forest algorithm are adopted to detect anomalous data, while the adjacent window mean method and cubic spline interpolation are used to impute missing values. Key meteorological factors are selected via the Pearson correlation coefficient, and an improved K-means algorithm is employed to categorize historical data into three weather types: sunny, cloudy, and complex. On this basis, the Informer-LSTM-MLP hybrid model is constructed: the Long Short-Term Memory (LSTM) network extracts temporal information from environmental features to generate hidden state matrices; the Informer model performs deep modeling of long-term dependencies in power sequences to capture global features; these two types of features are then fed into a Cross-Attention mechanism for multi-source information fusion, and finally, the Multi-Layer Perceptron (MLP) realizes end-to-end power prediction. Experimental results demonstrate that the proposed hybrid model can effectively improve the prediction accuracy of PV power generation under various weather conditions.