The k-day extreme precipitation depths (k=1, 2, 3, 4, 5, 7, 10, 15, 20, 25 and 30) at the climatological network of Belgium (165 stations) are analysed to assess the regional GEV growth curves and to determine the at-site fractiles. The calendar year and the hydrological summer and winter are considered separately. The method proposed combines regional L-moment estimates of the GEV parameters and tends to take advantage of a few long-term well-documented series. Therefore, a data extension procedure based on the fractiles method has been used to extend the 1951-1995 observation period to the 1910-1995 reference period. This ensures the temporal homogeneity of the series by assessing the possible missing extremes and it places all the series in a reference period where the stationarity of the extreme precipitation has been verified. Using the 9 historical series and generating randomly located missing values the efficiency of three data extension methods has been evaluated. This comparison indicates that a procedure using the regional growth curve satisfies this task. It shows that the residual mean square error of the at-site means is reduced when the mean correlation between the reference station and the series presenting gaps exceeds 0.52 but that the corresponding error on high order fractiles is reduced for all the observed correlation and for large numbers (40-50) of missing values. A practical estimation of the confidence intervals is proposed.