Abstract
High-resolution precipitation data is crucial for modern hydrological and building hygrothermal performance simulation models. In Australia, historical observations are inadequate, as half-hourly recordings only replaced daily observations at many stations from the early 2000s. Moreover, existing machine learning approaches are limited to generating hourly time series data. This paper presents a recurrent neural network using long short-term memory to disaggregate daily precipitation observations into half-hourly intervals. The model leverages temporal dependencies and hourly weather measurements. Our results, based on stations across five Australian climate zones, demonstrate that the model effectively preserves key half-hourly precipitation statistics, including variance and the quantity and distribution of wet half-hours. When aggregated to hourly intervals, our model outperforms other models in most measured metrics.