Short-term wind power forecasting approach based on Seq2Seq model using NWP data

Abstract Wind power is one of the main sources of renewable energy. Precise forecast of the power output of wind farms could greatly decrease the negative impact of wind power on power grid operation and reduce the cost of the power system operation. In this paper, a wind power output forecast model was proposed by integrating multivariate times series clustering algorithm with deep learning network. The NWP data and actual wind farm historical data were used as the input of the proposed model. 78 typical characteristic and statistical features were extracted from the inputs. Dimension reduction algorithm t-SNE was used to project the feature vectors into lower dimension and K-means algorithm was used to cluster the inputs into different clusters afterwards. At last, Seq2Seq with attention models were built for each cluster for power output prediction. The forecasting horizon is 1 day and the data resolution is 10 min. The results showed that the Seq2Seq model outperformed other existing forecasting methods such as Deep Belief Network and Random Forest. Clustering the input data into different clusters indeed improved the forecasting accuracy.