Short-Term PV Output Forecasting Approach Based on Deep Learning and Singular Spectrum Analysis

Currently, the uncertainty in meteorological conditions presents a challenge for accurately forecasting photovoltaic (PV) power output. Based on actual data from a PV plant, this manuscript proposes a PV power generation prediction method based on fuzzy c-means (FCM) clustering, singular spectrum analysis (SSA) and convolution neural network-bidirectional gated recurrent unit based on attention mechanism (CNN-BIGRU-ATTENTION). Firstly, the FCM clustering algorithm was used to cluster according to similar meteorological information. Secondly, SSA was used to decompose the relevant features to obtain new features that can help to analyse power generation characteristics. Among these, both FCM and SSA utilized the ant colony optimization (ACO) algorithm to search for optimal solutions for critical parameters. Thirdly, the CNN-BIGRU model was used to classify the weather types. Finally, the power generation was predicted based on the classification by using the CNN-BIGRU-ATTENTION model and the prediction results were evaluated. The mean absolute percentage error (MAPE) of the proposed prediction method for power generation prediction in sunny, cloudy and rainy was 4.69%, 6.08% and 6.60%, respectively. The presented methodology incorporated SSA to extract appropriate features, improving the prediction accuracy.