Gated Recurrent Unit Neural Networks for Wind Power Forecasting based on Surrogate-Assisted Evolutionary Neural Architecture Search

Abstract

Wind energy has become an important part of national power systems due to its wide distribution, low cost, and non-polluting characteristics. However, the intermittence, randomness, and fluctuating of wind energy make it extremely difficult to connect wind power to the grid, which in turn affects the normal dispatch of power resources. Therefore, accurate wind power forecasting is crucial for power systems. Deep neural networks (DNNs) can efficiently capture high-dimensional nonlinear spatiotemporal features and are employed. The architectures of state-of-the-art DNNs are usually hand-designed by users with extensive expertise. In this paper, a gated recurrent unit neural networks for wind power forecasting approach based on surrogate-assisted evolutionary neural architecture search (SA-ENAS) is proposed. Firstly, SA-ENAS uses gated recurrent unit neural networks (GRU) to capture high-dimensional nonlinear spatiotemporal features, while incorporating delay variables into ENAS. Secondly, the GRU architecture is jointly encoded with delay variables. Then, the architecture search and delay variable selection are achieved using a surrogate model based ENAS approach. Finally, the effectiveness and superiority of the proposed method are verified through the case study of an actual wind farm dataset.