Toward low-carbon and sustainable power systems, hydro–wind–solar complementarity provides an important pathway for enhancing renewable energy accommodation and operational flexibility, while accurate characterization of uncertainty and cross-energy dependencies is essential for system optimization, operational risk assessment, and sustainable utilization of renewable resources. This study proposes a conditional Wasserstein generative adversarial network with gradient penalty integrating one-dimensional multi-scale channel attention (MSA) and an exponential moving average (EMA) mechanism (MSA-cWGAN-GP) for joint runoff–wind–photovoltaic (PV) scenario generation. The generator employs parallel depthwise 1D convolutions with multiple temporal receptive fields to capture multi-timescale variations, while an EMA shadow generator is used for model validation and scenario generation. Conditional labels are obtained by clustering joint 24 h runoff–wind–PV profiles, enabling generation under typical resource states. Case studies using historical runoff observations from Shuibuya Hydropower Station and wind and PV power series derived from ERA5 reanalysis data show overall absolute errors of the autocorrelation function (ACF) and Kendall coefficient of 0.0113 and 0.0495, respectively. The proposed model achieves the best average performance among the evaluated models in preserving intraday temporal dependence, cross-energy dependencies, and distributional characteristics, providing representative scenarios for uncertainty analysis and subsequent optimization of hydro–wind–solar complementary systems.

The large-scale integration of wind and solar photovoltaic (PV) power is a cornerstone of low-carbon, sustainable energy systems. However, the uncertainty of the output brings great challenges to the operation and dispatching of power systems. To clearly describe the fluctuation characteristics of wind–PV power output and the spatial–temporal coupling relationship, a two-stage wind–PV scenario-generation method is proposed. This method is based on Difference-Constrained Variational Autoencoder and Mixture of Factor Analyzers (DVAE-MFA). In the first stage, a differential constraint term is added to the reconstruction loss of the Variational Autoencoder (VAE) to build the Difference-Constrained Variational Autoencoder (DVAE) model. This helps the model better learn the fluctuation characteristics of output sequences. In the second stage, to solve the problem of the posterior distribution of the DVAE latent variables deviating from the standard normal prior, the Mixture of Factor Analyzers (MFA) model is introduced for secondary probability modeling of the latent space. The simulation experiment results show that the proposed DVAE-MFA model outperforms comparison models in terms of the scenario temporal fluctuation characteristics, spatial–temporal correlations, and statistical distribution similarity. The generated output scenarios can reproduce the features of historical data, providing high-quality data support for the stochastic optimization scheduling of sustainable power systems.