Photovoltaic power prediction based on quadratic decomposition and deep learning

Abstract A collaborative deep learning prediction model based on the secondary decomposition of high-frequency data is presented to increase the prediction accuracy and robustness of photovoltaic power generation. The fully adaptive noise ensemble empirical modal algorithm (CEEMDAN) first decomposes the original power sequence. Next, the variational modal algorithm (VMD) decomposes the high-frequency randomly perturbed signals in it twice. Finally, the CNN-LSTM model predicts each component, and reconstruction yields the final power prediction. The CEEMDAN-VMD-CNN-LSTM model is compared and analyzed with other models using the Australian DKASC dataset as an example. The results demonstrate that the target method has high prediction accuracy and that the MAE is reduced by 11.49%, 30.98%, and 14.92%, respectively, in comparison to the primary decomposition prediction models, VMD-CNN-LSTM and CEEMDAN-CNN-LSTM. With an R 2 of up to 0.99, the model prediction results also show the best fit to the real power, demonstrating how much the method increases the accuracy of the PV power forecast.