We model extreme river flow data from five UK rivers with distinct hydrological properties. The data exhibit significant and complex nonstationarity, which we model using a nonlinear function of hydrological covariates corresponding to soil saturation, latent flow of the river and rainfall. We additionally consider season as a covariate, although the hydrological covariates explain most of the seasonal effect directly. The standard approach to modelling data of this kind is to fix a threshold and to model exceedances of this threshold using the generalised Pareto distribution. We identify a number of problems with this approach in nonstationary cases. To overcome these issues, we propose the use of a censored generalised extreme value distribution for threshold exceedances. The data analysis illustrates a number of features of model fit and in particular the stability of the model parameters and return levels to threshold choice.