Anomaly detection of power quality is crucial for the reliable operation of the distribution network，and its accuracy is directly related to the stability and security of the power grid system.In practice，it is not only time-consuming and laborious to analyze the anomaly in the field by manpower but also impossible to manage the anomaly in advance.To this end，a prediction model for power quality anomaly of distribution networks，which combines feature screening，signal decomposition，and an attention-based deep neural network，is proposed.First，the fast correlation-based filter （FCBF） algorithm based on maximal information coefficient （MIC） is combined with a feature screening method to select features of the input data.Then，the selected input features and corresponding components are fed into the temporal fusion transformer （TFT） model for prediction，and the prediction results of power quality anomalies such as voltage deviation，frequency deviation，and harmonic distortion rate are output.The complexity and computation time are significantly reduced，and the accuracy of fault diagnosis is potentially improved by the model，compared to conventional models.The real-time monitoring of power quality is realized，and the intelligence and visualization of the daily management of distribution network operation and maintenance are promoted.