Combined adaptive DRN-GRU model based on ultra short-term PV generation
Traditional deep neural network model training process will deepen the
network to a certain number of layers will produce network degradation
phenomenon, and then affect the training effect and prediction accuracy.
To address this problem, a PV power prediction model based on deep
residual network (DRN) and gated recurrent unit (GRU) neural network is
proposed. Firstly, the Pearson correlation coefficient method is used to
filter out the meteorological variables with high correlation with PV
power from historical data and reduce the data dimensionality. Secondly,
the DRN-GRU prediction model is proposed to be trained using the
adaptive learning rate Adam optimization algorithm to obtain the optimal
parameters. Finally, a DRN-GRU rolling prediction model is built based
on the historical data series to derive the PV power prediction results.
The results of the algorithm show that the model can still maintain good
training effect in the network training of deep numbers, effectively
solve the network degradation problem, and have higher prediction
accuracy compared with models such as artificial neural networks and
traditional convolutional neural networks.