Wind power forecasting based on a novel gated recurrent neural network model

Abstract

Accurate wind power forecasting is crucial for upgrading the operation and maintenance of wind turbines and the effectiveness in wind power dispatching. In this work, density-based spatial clustering of applications with noise (DBSCAN) is adopted to exclude outliers, and is examined to yield top-quality data filtering for neural networks when implemented on supervisory control and data acquisition (SCADA) data from a 7-MW wind turbine. Apart from features, such as wind and rotor speeds, time is a vital feature involved with synchronous or asynchronous events. Thus, a model-agnostic vector representation for time, Time2Vec (T2V), is used. This can be embedded with neural network architectures, such as Long Short-Term Memory (LSTM) and Gated Recurrent Unit (GRU). The novel neural network model, T2V-GRU, concatenates T2V into vanilla GRU, replacing the notion of time vector by T2V. Forecasting results are evaluated by mean absolute error (MAE), root mean squared error (RMSE), and accuracy. In comparison to T2V-LSTM, LSTM and GRU, the proposed T2V-GRU model in combination with DBSCAN refinement has superior forecasting results for the 7-MW turbine, yielding maximum forecasting accuracy (93.12 %) and minimal prediction errors in terms of MAE (286.10 kW) and RMSE (489.54 kW), while simultaneously reducing the computational cost and time.