Scenario forecasting for wind power using flow-based generative networks

Wind power prediction is an integral part of power system operations and planning. Due to rising penetrations of wind turbines, fluctuation and intermittence of wind powers seriously limit the accuracy of power forecasts. A popular way to mitigate this challenge is to provide a range of possible scenarios instead of deterministic point forecasting values, so operators can account for the uncertainties. This paper proposes a model-free scenario forecasting approach for wind powers using flow-based generative networks, which generate a set of high-quality scenarios to represent possible behaviors based on historical wind powers and deterministic point forecasting values. Firstly, an unsupervised deep learning framework is proposed to learn the latent patterns in historical wind power curves. Then, a large number of possible future scenarios are obtained by solving an optimization problem. Simulation results show that the proposed approach has better performance than popular baselines such as variational auto-encoder and generative adversarial networks. (C) 2021 The Author(s). Published by Elsevier Ltd.