Water demand forecasting is the key to allocating water resources, saving energy, and reducing water age of water distribution network. Existing research focuses on the forecasting models but ignores the pre-processing steps such as abnormal detection, which restricts the accuracy of the models. A local outlier factor (LOF) model based on density was proposed to identify abnormal values of water demand data. The LOF was then combined with light gradient boosting machine (LightGBM) to form a hybrid water demand forecasting model LOF+LightGBM. The model was tested through actual cases. Results show that the root-mean-square error of the forecasting model based on data processed by LOF reduced by about 10% on average, compared with the forecasting model based on raw data. The mean absolute error of LightGBM on different datasets was 24.7% lower than artificial neural network (ANN) and support vector machine (SVR) on average. Overall, LOF+LightGBM showed the best prediction performance and the Nash-Sutcliffe model efficiency coefficients for three district metered areas (DMAs) were 0.6,0.951, and 0.942, respectively. The training and computational time of all the models was less than 0.7 s. In conclusion, LOF model, LightGBM model, and LOF+LightGBM model are conducive to improving the accuracy of the water demand forecasting model.