A novel multi-gradient evolutionary deep learning approach for few-shot wind power prediction using time-series GAN

Abstract

Due to the lack of historical data, accurate prediction is a great challenge for newly constructed wind farms (NWFs). How to guarantee satisfactory prediction accuracy with limited data has become a bottleneck issue in wind power prediction (WPP). To solve the few-shot learning problem in NWFs, a novel multi-gradient evolutionary deep learning neural network (EATDLNN) prediction model is proposed, which incorporates the time-series GAN (TimeGAN) and multivariate variational mode decomposition (MVMD) method. TimeGAN adds an auto-encoder network and a supervised loss function to GAN, which enables it to capture the dynamic time series correlation characteristics of wind data and generate high-quality samples. MVMD is then employed to simultaneously decompose nonstationary sequences. In the EATDLNN prediction model, an evolution-based multi-gradient training approach is proposed to train the DLNN by integrating three gradient descent methods (i.e., Adam, MBGD-Momentum, RMSprop). Compared with the traditional single-gradient training approach, the multi-gradient training approach provides different gradient descent paths and the gradient can adaptively choose to update towards the path with the smallest error based on the evolutionary framework, which can help the model jump out of the local optimum and enhance the generalization ability. Massive experiments prove the excellent performance of the proposed method, especially in one-step prediction, the root mean square errors reduce by 79.40%∼91.05% compared with other state-of-the-art methods.