Enhancing short-term power forecasting of PV clusters: A statistical upscaling and CBAM-BiLSTM approach

Traditional short-term solar power forecasting primarily focuses on individual photovoltaic (PV) plants. Recently, there has been increasing demand for power forecasting of PV clusters. In this paper, a distributed PV cluster power prediction model based on statistical upscaling and convolutional block attention module (CBAM)–bi-directional long short term memory (BiLSTM) is proposed to strike a balance between prediction accuracy and model training time. To address the issue of inaccurate cluster partitioning caused by variations in distributed PV output curves, a K-shape based cluster partitioning method is proposed. This method effectively reduces the number of prediction scenarios. Furthermore, a soft-dynamic time warping based representative power plants selection method is introduced to accurately identify representative power plants from sub-clusters. This selection method takes into account the horizontal and vertical stretching of the output curves, ensuring a comprehensive reflection of curve similarity. To minimize the conversion error during cluster transformation, a real-time statistical upscale conversion method is proposed. This method considers multiple similar output days, resulting in more accurate sub-cluster output predictions. The hybrid model, which employs CBAM for initial feature extraction and BiLSTM for output forecasting, is introduced to predict the output of representative power stations. Finally, the effectiveness of the proposed model is verified through evaluation indices, with the mean absolute percentage error value of the representative power station being less than 5%. The accuracy of the results is further supported by the confidence interval analysis.