Short-term Photovoltaic Power Forecasting Using SOM-based Regional Modelling Methods

The inherent intermittency and uncertainty of photovoltaic (PV) power generation impede the development of grid-connected PV systems. Accurately forecasting PV output power is an effective way to address this problem. A hybrid forecasting model that combines the clustering of a trained self-organizing map (SOM) network and an optimized kernel extreme learning machine (KELM) method to improve the accuracy of short-term PV power generation forecasting are proposed. First, pure SOM is employed to complete the initial partitions of the training dataset; then the fuzzy c-means (FCM) algorithm is used to cluster the trained SOM network and the Davies-Bouldin index (DBI) is utilized to determine the optimal size of clusters, simultaneously. Finally, in each data partition, the clusters are combined with the KELM method optimized by differential evolution algorithm to establish a regional KELM model or combined with multiple linear regression (MR) using least squares to complete coefficient evaluation to establish a regional MR model. The proposed models are applied to one-hour-ahead PV power forecasting instances in three different solar power plants provided by GEFCom2014. Compared with other single global models, the root mean square errors (RMSEs) of the proposed regional KELM model are reduced by 52.06% in plant 1, 54.56% in plant 2, and 51.43% in plant 3 on average. Such results demonstrate that the forecasting accuracy has been significantly improved using the proposed models. In addition, the comparisons between the proposed and existing state-of-the-art forecasting methods presented have demonstrated the superiority of the proposed methods. The forecasts of different methods in different seasons revealed the strong robustness of the proposed method. In four seasons, the MAEs and RMSEs of the proposed SF-KELM are generally the smallest. Moreover, the $R^{2}$ value exceeds 0.9, which is the closest to 1.