Short-term photovoltaic power forecasting using parameter-optimized variational mode decomposition and attention-based neural network

Photovoltaic power generation is impacted by various meteorological factors leading to significant intermittent and volatile, so dispatch of photovoltaic power plants and safe operation of power systems hinge on accurate prediction of PV power output. Researchers have proposed a variety of ways to improve the performance of predictions, and a hybrid model often performs better than a single model. Considering that the sequence decomposition method can alleviate the volatile nature of the original sequence, we propose a new hybrid model VMD-GA-Conv-A-LSTM, design a method to determine the optimal parameters of the VMD and utilize the parameter-optimized VMD for sequence decomposition, combining with a novel deep learning model for more accurate prediction. The model first calculates the optimal parameters for the variational mode decomposition (VMD) using a search algorithm over a specified parameter range, and uses these parameters to decompose the photovoltaic power sequence into several sub-sequences. Then, the sub-sequences and preprocessed historical meteorological data are input into several long short-term memory (LSTM) integrated with 1D convolution and attention mechanism (Conv-A-LSTM) separately. The predictions corresponding to each sub-sequence are accumulated to get the predictions of the hybrid model. The hybrid model was validated on the dataset generated from the 5.20 kW Photovoltaic site in Alice Springs, Australia, and ERA5 data, respectively. Compared with baseline models, the proposed hybrid model achieves the best prediction accuracy. The RMSE, MAE, and R2 of the 2-hour prediction performed on the Australia dataset are 0.1884 kW, 0.0758 kW and 0.9876, respectively. Therefore, the hybrid model proposed in this study is able to provide statistical data support for photovoltaic plant operation and scheduling.