To address the challenges of complex terrain, significant noise interference in wind speed and power data for small-scale distributed wind farms in rural areas, an improved WaveNet prediction algorithm integrating enhanced isolation forest data cleaning and spatiotemporal feature mining is proposed to enhance ultra-short-term power prediction accuracy. Firstly, an enhanced isolation forest （EIF） algorithm is employed for data cleaning. This algorithm incorporates dynamic quantile threshold optimization, path length window smoothing, and a multimodal correction strategy to effectively identify and eliminate anomalous data points. Subsequently, the maximal information coefficient （MIC） is utilized to screen highly correlated input features. A temporal attention mechanism combined with a Transformer module captures critical temporal dependencies, while a spatial attention mechanism integrated with a graph convolutional network （GCN） models spatial correlations among wind turbines for spatiotemporal feature fusion. Next, an ultra-short-term single-turbine power prediction model based on WaveNet architecture is constructed，which accurately captures the dynamic patterns of the fused feature sequence through its dilated convolution and skip connections, and outputs the single-turbine power prediction results. Finally, the total cluster power output is aggregated from single-turbine prediction results by considering the positional weights and spatial correlations of each turbine. A case study is conducted on three-small-turbine distributed wind farms with a single turbine capacity of 10 kW deployed in a certain rural area. The results demonstrate that the proposed method achieves high prediction accuracy.