Accurately predicting the amount of construction waste is of great significance for carrying out the recycling treatment of construction waste and guiding the government to formulate relevant policies. However, the lack of reliable forecasting methods and historical data makes it difficult to predict the construction waste in the long- or short-term planning. On the basis of the univariate time series data of limited sample points, this paper puts forward a short and long memory(LSTM) time series prediction method to effectively solve the problem of construction waste prediction, which involves network structure with dropout layer and the algorithm of network training and prediction process. Taking Shanghai as a case，compared with other time series prediction models, numerical experiments were conducted to verify the effectiveness and accuracy of the LSTM prediction model in the filed of predicting construction waste generation.