A Hybrid Model for Ultra-short-term PV Prediction Using SOM Clustering and ECA

Abstract

For the grid to operate safely and steadily and for PV electricity to be connected on a broad scale, the precision of the ultra-short-term PV power prediction is crucial. A combination model of ultra-short-term PV prediction based on an attention mechanism is proposed to increase the prediction accuracy of PV output power under various weather circumstances. First, using a Pearson correlation coefficient analysis, important climatic variables that are closely associated with PV power generation are selected and normalized on a monthly basis. The Sky Condition Factor (SCF), a classification index, is then obtained by computing a weighted summation. This reduces the dimensionality of the input variables and eliminates seasonal influence on weather classification and the coupling interactions among various meteorological elements. Second, an unsupervised clustering of SCFs using a Self-Organizing Map (SOM) neural network is used to classify three types of weather. After that, CNN prediction models are built for each of the three types of weather. The Efficient Channel Attention (ECA) module is then added, allowing the model to focus on key feature information and increase prediction accuracy by adaptively assigning phase weights to each of the multiple channels of feature information that the CNN has extracted. Lastly, the efficacy of the suggested prediction model is verified by simulations run on historical observed data, which demonstrate an improvement in the prediction model's accuracy under various weather conditions when compared to the model without the ECA module.