Abstract: The development trend of large-scale electric vehicles (EVs) in the future adapts to the current national call for "New infrastructure", and provides new opportunities for alleviating the crisis of fossil energy and environmental problems. However, the charging randomness and uncertainty of a large number of EVs have brought new challenges to the safe and stable operation of the power grid. The charging disorder of household users is stronger. It is of great significance to monitor and extract the real-time household charging loads, formulate corresponding demand response strategy or energy efficiency management and guide their charging behaviors. Therefore, this paper proposes a non-intrusive charging load extraction method using the low-frequency characteristics of electric vehicle charging loads. Firstly, the low frequency components of smart meters are extracted by the two-stage decomposition technique. On this basis, the event monitoring and dynamic time warping (DTW) are used to estimate the charging intervals and amplitudes. Then, taking the decomposed electric vehicle charging loads and the aggregated smart meter data as the inputs, the CNN-Attention-LSTM neural network algorithm is used for the data training in order to forecast the short-term charging of electric vehicles for the household users. The effectiveness of the algorithm is verified on numerical cases.