High-quality power grid spatio-temporal data is essential for optimal scheduling decision-making in power system, yet real-world data is limited by privacy constraints and fails to cover scenarios from load fluctuations and renewable uncertainties. We propose SceDiff, a conditional diffusion model for generating power grid scenario spatio-temporal data, addressing the lack of joint modeling of multivariate time series, graph structures, and covariates in existing methods. The method employs a denoising network combining temporal and graph convolutional networks to capture temporal and spatial dependencies, and introduces a penalty loss to enforce scenario constraints. Experiments on real-world datasets show that SceDiff outperforms baselines in power interval width and coverage rate, producing values closest to real data with high coverage. Further ablation studies confirm the contribution of each module in the method.