The large-scale access of new energy vehicles to distribution grids in recent years has affected the safe operation of traditional distribution grids, while electric vehicles, as highly flexible mobile energy storage units, and the demand response of electric vehicle users under the influence of various incentives in the electricity market will change the load characteristics and affect load forecasting. A combined CNN-AM-BiLSTM forecasting method based on response feature decomposition is proposed for the characteristics of high participation of demand response and volatility of charging load of users in public areas. According to the predictability of demand response signals and the independence of seasonal signals, wavelet decomposition and other methods are used to decompose charging loads in public areas at different levels to form seasonal base loads and demand response dominant loads. The seasonal base load is analysed for smoothness and the time-series forecasts are made based on the analysis results. To address the problem that demand response dominant load has many influencing factors and is difficult to obtain, convolutional neural networks are used to extract depth features and an attention mechanism is introduced to make the Bi-LSTM model focus more on the features that are most relevant to the load information, thus improving the forecasting effect. Finally, the two parts of the prediction results are superimposed to obtain the public area EV charging load prediction results.