There is a problem of series arc fault detection in low voltage lines. Thus a series arc fault detection method based on improved CEEMDAN decomposition and spatial-temporal features is proposed. First, the CEEMDAN algorithm is used to complete the decomposition of the current signal, and the rough selection of the high-frequency signal is realized based on the kurtosis index, margin index, energy feature and energy entropy feature of each IMF component. Then, a feature construction method combining spatial and temporal scales is proposed to capture the local feature of each high-frequency IMF component. This enhances the contrast and discriminants of the current feature. Finally, some subspace transformation algorithms are used to implement the second dimension reduction of the current spatial-temporal feature set, and the series fault arc detection is realized based on SVM. The actual test shows that the average fault arc detection accuracy of the proposed algorithm is 88.33%, which is efficient for series fault arc detection. This work is supported by the National Natural Science Foundation of China (No. 52077221).