conditions is a key research question in the nuclear safety field. The analysis of the risk
associated to extreme temperatures has already been intensively studied in the literature.
Nevertheless, methods for estimating temperature return levels and constructing confidence
intervals were often used separately, without detailed comparison and in a stationary context.
The extreme value theory is often used to assess risks in a context of climate change. It provides
an accurate indication on marginal laws describing the frequency of occurrence of extreme
phenomena such as heat waves. However, in a non-stationary context, the notions of duration
and return periods are not easily interpretable. For example to evaluate the 100-year return level
in a future year, time-varying frequency models with different ways to compute the extreme
temperatures and different methods to evaluate associated uncertainties must be used and
This study examines the performance of three different ways to compute high temperature
return levels in a non-stationary context and three methods for constructing the associated
confidence intervals. To address this issue, we have used three explanatory variables (max daily
temperatures, anomalies and residuals) and we have used and compared the delta, the profile
likelihood and the bootstrap methods for computing confidence intervals. The analyses were
conducted assuming the General Extreme Value (GEV) distribution. The daily maximum
temperatures observed at the station of Orange in France are used as a case study.