Abstract: To improve the accuracy of photovoltaic power forecasting, a photovoltaic power forecasting method based on K-means clustering and an improved Long- and Short-term Time-series Network (LSTNet) is proposed to address the limitations of the traditional LSTNet model in modeling long-term dependency information and representing complex dynamic relationships in multivariate data. First, abnormal data in the original samples are removed, and correlation analysis is conducted to identify key meteorological features that have a significant impact on photovoltaic output. Then, the K-means algorithm is employed to cluster the samples so as to characterize the distribution differences of photovoltaic power data under different operating conditions, and the optimal number of clusters is determined using the Akaike Information Criterion (AIC) and the Bayesian Information Criterion (BIC). Furthermore, an encoder-decoder structure is introduced on the basis of LSTNet, and a Convolutional Neural Network (CNN) and a Transformer module are integrated to achieve the coordinated extraction of local temporal features, global dependency relationships, and long-term dynamic information of the sequence. The results show that the proposed Kmeans-ILSTNet model outperforms the comparison models in evaluation metrics such as RMSE and MAE, thereby verifying its effectiveness and superiority in photovoltaic power forecasting.