A Multi-step ahead photovoltaic power forecasting model based on TimeGAN, Soft DTW-based K-medoids clustering, and a CNN-GRU hybrid neural network

Abstract

Accurate photovoltaic (PV) power generation forecasting is very important for making economic and reliable power dispatching plans. This study proposes a multi-step ahead PV power forecasting (PPF) model, which combines time-series generative adversarial networks (TimeGAN), soft dynamic time warping (DTW)-based K-medoids clustering algorithms, and a hybrid neural network model computed by a convolutional neural network (CNN) and gated recurrent units (GRU). First, a new PV power data augmentation method based on TimeGAN is proposed to realize the expansion of the historical PV power data. Second, the combination of soft-DTW and K-medoids clustering is developed to extract and distinguish the commonalities and differences between time series characteristics of PV power. Finally, a hybrid neural network model is developed to integrate CNN and GRU into a unified framework for an accurate PPF under different weather conditions. The performance of the proposed model was tested using real datasets of two PV power stations with an installed capacity of 50MW. The comparative analysis of the proposed model with five reference models confirms its superiority in terms of forecasting accuracy. For the sunny days and cloudy days, the range of root mean square error (RMSE) values are produced by the proposed model distributed in [0.927–2.523] MW and [1.781–6.173] MW respectively with respect to the multi-step ahead forecasting from 15-min up to 24-hour. The experimental results of the hybrid model proposed in this study indicate that data augmentation-based clustering method is very helpful for improving the accuracy of PV power forecasting.