Accurate prediction of coal demand is essential for optimizing energy resources in long-term power system planning. This paper examines the coal demand in North China from 2007 to 2022 using econometric methods to identify key influencing factors as input variables. Then, the Sparrow Search Algorithm (SSA) is used to optimize the key parameters of the Least Squares Support Vector Machine (LSSVM) algorithm to enhance the prediction accuracy of coal demand. Case studies are conducted on actual data in North China, and the results show that the proposed hybrid SSA and LSSVM method outperforms traditional approaches in small-sample, multivariable forecasting, making it suitable for predictions in long-term power system planning.

Data-driven analytics are emerging as powerful tools to enhance the planning and operation of modern power systems characterized by the ever-increasing penetration level of renewable energy and flexible loads.However,their effectiveness is often constrained by imbalanced and limited real-world datasets. Consequently,conditional generative models are widely adopted to augment limited training data by synthesizing scenario-specific high-fidelity power profiles.However,they fundamentally rely on external conditional labels,the acquisition of which poses a significant practical challenge.Specifically, meteorological records frequently lack adequate spatiotemporal resolution,while manual annotations are costly and susceptible to inaccuracies that compromise downstream synthesis.To address these limitations,this paper proposes an end-to-end label-free conditional synthesis framework based on the Selective Kernel Gaussian Mixture Variational Autoencoder (SK-GMVAE).Within this framework,a Gaussian Mixture (GM)prior and a Selective Kernel (SK)mechanism operate synergistically.The GM prior establishes a highly interpretable latent space that autonomously disentangles discrete operational modes,thereby obviating the need for external labels.Concurrently,by dynamically adjusting receptive fields to capture multi-scale temporal dependencies,the SK mechanism prevents the degradation of synthesis quality and ensures high-fidelity profile generation.Extensive quantitative experiments demonstrate that the SK-GMVAE consistently outperforms state-of-the-art baselines across diverse datasets.Notably,it achieves up to 17.5%improvement in Fréchet Inception Distance (FID)and14.7%in Inter-class Variance (IV)compared to the best-performing baselines.This rigorously verifies its versatility and superiority in handling complex daily power profiles.Furthermore,downstream forecasting tasks confirm that augmenting training sets with our synthesized profiles effectively rectifies data imbalance and significantly enhances predictive robustness.