Bidirectional Gated Recurrent Unit-Based Lower Upper Bound Estimation Method for Wind Power Interval Prediction
Fang Liu, Qing Tao, Dechang Yang, Denis Sidorov
Abstract
High-quality interval prediction is helpful to accurately capture the uncertainty of wind power generation and provide support to grid dispatchers and operators. As an effective and reliable prediction interval (PI) construction framework, lower upper bound estimation (LUBE) method is used in forecasting tasks. This article proposes a new data-driven PI construction method for wind power prediction based on LUBE theory and a bidirectional gated recurrent unit (GRU) neural network, which integrates a loss function based on likelihood for model training. In this framework, the bidirectional GRU acts as the core predictor, and learns the long-term dependence of wind power time series in chronological and reverse chronological order. The learned features are merged by using a fully connected layer to generate the upper and the lower bounds of the target PI. The proposed method is evaluated on a real wind farm dataset. Numerical results show the superiority of the proposed method by comparing with the traditional LUBE method, long short-term memory-based LUBE method, mean-variance estimation method, and an emerging efficient gradient descent method.