Abstract: Aiming at the problem of low accuracy of PV output prediction, a multiple convolutional combination of large models is proposed, i.e., the combination prediction model of triple convolutional neural networks, weighted fully connected regres-sion networks and improved bidirectional encoder representations networks (TCNNs-WFRN-IBERT). The triple convolu-tional neural networks (TCNNs) use multiple sizes of convolutional kernels to efficiently mine the feature information of PV data from shallow to deep; the weighted fully-connected regression network (WFRN) optimizes the weight coefficients of the prediction outputs of the two deep neural networks (DNNs) using the particle swarm optimization (PSO) to improve the prediction accuracy; and integrates the prediction results of the TCNNs and the WFRN and inputs them into the im-proved Bidirectional Encoder Representation Network (IBERT) improved large model for training, and utilizing the atten-tion mechanism of IBERT to achieve interpretable feature analysis to determine the final PV output prediction value. TCNNs-WFRN-IBERT was used to predict the hourly PV output power for 1 day ahead in the city of Natal, Brazil, and simulation experiments were performed and compared with 38 algorithms using actual PV output and meteorological data. The results show that the average absolute error, mean square error and root mean square error of the TCNNs-WFRN-IBERT model are 22.61 W, 1818.20 W2 and 42.64 W. The evaluation indexes of TCNNs-WFRN-IBERT are lower than those of the other comparative models and the value of the MAE is at least 2.71% smaller than that of the other 38 comparative models, which validates the effectiveness of the proposed model. , verifying the validity of the pro-posed model.

Key words: triple convolutional neural network, weighted fully connected regression network, improved bidirectional encoder repre-sentations from transformers, photovoltaic output forecasting.