A long-term water resource plan is driven by the need to establish strategies for mitigating the adverse impact of uncertain climate risks and their effects. For this purpose, stochastic generation models are widely used to generate plausible scenarios of weather variables that have statistically similar characteristics as that of observed variables. The stochastic model is only able to reproduce statistical properties of weather variables when the observed record is long enough to understand the underlying structure in a statistical way. However, these observed records, providing sufficient temporal and spatial coverage, are not readily available. A way to alleviate this is to generate artificial sequences of weather variables with finite length and use these as inputs to hydrologic models for risk and reliability assessment in the design and operation of water resource systems, given rare occurrences of these variables. There are a number of approaches to spatio-temporal downscaling modeling of precipitation for assessment of the impact of climate change and variability on water resources, mostly at a daily time-step. Although various approaches have been proposed for spatio-temporal modeling, the underestimation of extremes in multi-site daily rainfall simulation models has not been extensively studied. In view of the above mentioned limitations, this study proposes a machine learning based multisite daily simulation (ML-MRS) model with a mixture distribution to better reproduce rainfall statistics. For the multi-site model, this study employs a parametric model for spatially correlated rainfall occurrence and amount introduced by Wilks (1998).
ML-MRS approach improves the structural limitations on reproduction of daily rainfall statistics such as wet days and extreme rainfall, presented in the conventional multisite model. The proposed ML-MRL model almost perfectly preserves the inter-station correlations of precipitation occurrence and amount, as well as their dependences between the occurrence and amount. More specifically, the proposed ML-MRS approach has the advantage of connecting of occurrence patterns to their amount rainfall patters using a machine learning based clustering technique and ensemble decision tree technique method. In addition, the mixtures of gamma distributions was used to the estimation of precipitation amounts. As a results, most of rainfall statistics at different temporal scales (i.e. daily, monthly and yearly) were reasonably well reproduced. Therefore, it can be interpreted that the ML-MRS model can be regarded as an efficient stochastic daily rainfall generation model as compared to the existing models.