Abstract:

To address the issue of low accuracy in photovoltaic output prediction， a multi-convolutional combined large model is proposed， integrating triple convolutional neural networks（TCNNs）， a weighted fully-connected regression network（WFRN），and improved bidirectional encoder representations from transformers（IBERT）. The TCNNs employed convolutional kernels of multiple sizes to efficiently extract feature information from photovoltaic data， progressing from shallow to deep layers. The weighted fully-connected regression network（WFRN） optimized the weight coefficients of the prediction outputs from two deep neural networks using particle swarm optimization algorithm， thereby enhancing prediction accuracy. The prediction results from TCNNs and WFRN were integrated and input into the IBERT for training. The attention mechanism of IBERT was then employed to perform interpretable feature analysis， determining the final photovoltaic output prediction value. The TCNNs-WFRN-IBERT model was applied to predict the hourly photovoltaic output power for the next day in Natal， Brazil. Simulation tests were conducted using actual photovoltaic output and meteorological data， and the results were compared with those of 38 algorithms. The results showed that the mean absolute error（MAE）， mean squared error（MSE）， and root mean squared error（RMSE） of the TCNNs-WFRN-IBERT model were 22.61 W， 1 818.20 W² and 42.64 W， respectively. Compared with other models， the evaluation metrics of TCNNs-WFRN-IBERT were lower than those of the other models， with its MAE value being at least 2.71% smaller than those of the other 38 comparative models， validating the effectiveness of the proposed model.

Key words: triple convolutional neural network, weighted fully connected regression network, improved model of bidirectional encoder representation from transformers, photovoltaic output prediction, multi-convolutional combined large model, attention mechanism