Abstract: The wind power prediction for wind farm clusters is important for optimal scheduling of regional wind farms. The existing cluster prediction algorithms and cluster division methods do not consider the differential fluctuations among stations' numerical weather prediction (NWP) along time scale, and the cluster isn't divided reasonably. Consequently, we put forward a short-term wind power prediction method based on weather-process-dynamic wind farm cluster division. First, the 96-hour-time-scale prediction sample is equally divided into 16 sub-samples, followed by separate cluster division judging for each sub-sample. Then, the training set is organized for each sub-sample's sub-cluster according to cluster division result. Finally, the bidirectional long-short term memory (BLSTM) neural network is used for power prediction for each sub-cluster. The results show that, using the proposed method, the prediction accuracy can increase by 1.69%, 0.77%, and 0.59% compared to using the statistical upscaling method in 4 h-urtra-short-term, 24 h-day-ahead, and 96 h-short-term wind power prediction, respectively. The research can provide a reference for the topic of wind farm cluster division and short-term power prediction.