Hedge Backpropagation Based Online LSTM Architecture for Ultra-Short-Term Wind Power Forecasting
Chunyang Pan, Shuli Wen, Miao Zhu, Huili Ye, Jianjun Ma, Sheng Jiang
Abstract
Owing to the global concern for serious energy crises and increased greenhouse gas emissions, the wind energy capacity has grown rapidly in recent years. However, due to weather variations, dramatic power fluctuations exacerbate the uncertainty of wind power generation. To enhance stability and maximize utilization, the precise prediction of wind power is critical for providing a reliable reference for economic dispatch and energy efficiency improvement. In this article, a novel hedge backpropagation-based online LSTM architecture is developed to predict ultra-short-term wind power outputs. The proposed architecture consists of a hedge backpropagation mechanism, an online deep learning algorithm and a concept drift detection scheme to adaptively enhance prediction accuracy by updating the capacity of the conventional LSTM model. Furthermore, several types of historical data from an offshore wind farm are used as inputs to extract the intrinsic relationships between wind power and environmental factors in a case study. The numerical results demonstrate that the proposed proactive model yields better prediction accuracy than traditional offline algorithms such as recurrent neural networks and gated recurrent unit networks and is expected to improve forecasting performance and support the operation of wind farms.