To reduce model accuracy caused by unknown uncertain pollutant sources, a water quality model based on a 4-dimensional variational assimilation method was established. Water quality modelling was conducted in the mainstream of the Nanfei River, Lake Chaohu Basin, in which water quality is highly unstable influenced by unmeasured pollutant sources. The proposed model takes the unsteady process of non-point source loads as the control variable. By assimilating the daily water quality monitoring data in different gauges along the river, the water quality process was successfully reproduced by identifying the daily pollution load process of each reach. It changes the premise that the traditional model simulation needs the right pollution load in advance. The numerical experiment shows that the results of water quality simulation using variational assimilation method are significantly improved, and the Nash efficiency coefficient of water quality simulation in key river reaches is increased from less than 0 to more than 0.5. The identified pollution process consistent with the observed rainfall process proves that the pollution into the river is closely related to the rainfall. The model can also identify some abnormal pollutant loads into the river and improve the diagnosis and analysis ability of the model for water environment problems.