Multicollinearity-reduced conditional variational autoencoder for low-carbon dispatch of industrial energy systems

Industrial integrated multi-energy systems must coordinate electricity, heat, gas, storage and flexible demand while limiting carbon emissions. Renewable uncertainty complicates this dispatch because wind and photovoltaic profiles are stochastic and correlated. This study develops a scenario-informed dispatch framework in which a multicollinearity-reduced conditional variational autoencoder (MCLRCVAE) generates joint wind–photovoltaic scenarios, a probabilistic-distance method retains representative scenarios with probabilities, and the reduced set is embedded directly in a low-carbon dispatch model. The multicollinearity-reduction step removes redundant linear dependence before latent-variable learning, while the conditional variational structure produces scenario distributions rather than a single forecast. To ensure that every numerical result is independently reproducible and to address concerns about parameter provenance, the numerical study is built entirely on transparent, fully public data sources with documented assumptions. On this benchmark, MCLRCVAE attains the lowest Wasserstein distance among the tested generators, indicating the closest match to the empirical joint distribution, while different baselines lead on individual point-error metrics. Scenario-based dispatch reduces total cost and expected CO₂ emissions relative to deterministic dispatch under the benchmark assumptions. The results support multicollinearity-aware scenario generation for low-carbon dispatch.