Abstract: Developing a stochastic hydrometeorological field with spatiotemporal correlations and a clear physical coherence is critical for hydrological simulations. This study uses a coupled model of multivariate first-order autoregressive (MAR1) model, a first-order Markov chain, and a K-nearest neighbors (KNN) to develop a multisite, multivariate weather generator that can reflect spatiotemporal dependencies, inter-variable correlations, and low-frequency interannual oscillations inherent in hydrometeorological processes. We have applied this generator to the random simulations of daily precipitation and maximum and minimum air temperatures across 12 secondary water resource divisions in the Yangtze River basin, and achieved physically meaningful meteorological fields that are characterized by temporal and spatial correlations. The model is evaluated comprehensively using several metrics, such as basic statistical characteristics, correlation features, and interannual variability. The results demonstrate the multisite, multivariate weather generator effectively reconstructs a range of characteristic indicators of the observed meteorological fields. However, it does show certain underestimated durations of the maximum drought and wet periods at a few gauge stations, and similar errors in the first-order autocorrelation coefficients for daily maximum and minimum air temperatures. The findings of this study provide valuable insights for distributed stochastic hydrological simulations.

Key words: multisite multivariate weather generator, spatial correlation, inter-variable correlation, inter-annual variability, Yangtze River basin