This paper proposes a wind-solar joint scenario generation method to address source-side uncertainty under high renewable energy penetration. The proposed model integrates Temporal Convolutional Networks (TCN) and Variational Autoencoders (VAE), combining temporal dependency learning with latent distribution modeling to produce realistic and diverse scenarios. To evaluate the practical applicability of the generated scenarios, a performance-driven evaluation framework is developed based on a gravity energy storage (GES) system. The framework uses synthesized wind-solar scenarios and historical load data as inputs, simulating system behavior under real operational constraints. Experiments are conducted using 2019 historical data from Sardinia, Italy. Results show that the proposed TCN-VAE model effectively captures the temporal structure and seasonal characteristics of renewable generation, achieving low reconstruction error and strong statistical similarity with real data. Compared to simplified model variants, the full model demonstrates improved scenario diversity and better scheduling adaptability, as reflected in reduced curtailment, lower ramp rates, and higher dispatch efficiency under GES constraints. The method supports enhanced scenario-based decision-making for renewable-integrated power systems and offers a flexible foundation for applications such as optimal storage scheduling, grid planning, and uncertainty-aware dispatch strategies.