One of the major challenges in grid-connected wind farms is the flicker emission caused by the reactive power extremely short time variations. A general used solution is installing a static volt-ampere reactive (VAR) compensator (SVC) as the most cost-saving device, which still holds the risk of incomplete reactive power compensation subject to the operating delay time. To this end, an accurate extreme short-term reactive power forecasting structure is essential. Although there are many studies related to the forecasting of wind power in long, medium, and short terms, however, the extremely short-term reactive power prediction of the wind farms is only addressed in few studies. The main aim in this article is to develop a method based on the convolutional neural network (CNN) to directly learn nonstationary and complex features from the raw wind farm reactive power time series and to contribute a predictive controller to mitigate the voltage flicker through an SVC connected to the wind farm. However, high sensitivity of typical mean-squared-based loss functions might lead to considerable errors in the forecasting. To resolve this issue, a modified loss function is proposed to enhance the performance and, consequently, an improved CNN is presented. The evaluation of the proposed method is based on the actual recorded data of a wind farm in Manjil, Iran. The data are directly used in the modeling of the wind farm. A current source, which is based on the actual records its amplitude and phase angle change every 0.01 s, is utilized to model the wind farm. The numerical results in terms of flicker sensation and short-term flicker perceptibility (Pst) measurement are used to verify the performance of the proposed method through comparison with the wind farm performance without SVC and SVC with a common control system. [ABSTRACT FROM AUTHOR]