A wind power prediction model based on optimized N-BEATS network with multivariate inputs

The uncertainty of wind power has brought great challenges to the smooth operation and dispatch of the power system. Accurate short-term wind power prediction can provide a reliable basis for wind power grid connection and grid dispatch, and enhance the utilization of wind energy. Considering that neural basis expansion analysis (N-BEATS) only supports univariate input and ignores the short-term influence of meteorological factors such as wind speed and wind direction, a short-term wind power prediction model based on temporal convolutional network (TCN) optimized N-BEATS with multivariate inputs is proposed, named N-BEATS-TCN. The N-BEATS-TCN model designs a TCN stack consisting of TCN blocks, which enables the standard N-BEATS network to perform multivariate inputs. Specifically, the N-BEATS-TCN model adopts a novel architecture based on forward and backward residual links and three deep fully-connected layer stacks, namely trend stack, seasonal stack, and TCN stack, respectively, which can predict the trend, cycle, and external influences such as wind speed and wind direction of wind power. The proposed model is not only interpretable but also has good robustness and prediction accuracy. The model is applied to a data set obtained from the actual wind farm in northwest China. Compared with common statistical models, shallow neural networks and deep neural networks, the prediction accuracy of N-BEATS-TCN proposed in this paper is the highest, which provides a new technique for wind power prediction.