Abstract
The increasing integration of weather-dependent renewable energy sources into Virtual Power Plants (VPPs) introduces significant uncertainty in short-term dispatch planning. This paper develops a comprehensive decision-making framework that jointly optimizes credible capacity allocation and operational dispatch under forecast uncertainty. We first compare three probabilistic forecasting methods-Bayesian Neural Networks (BNN), Quantile Regression Forests (QRF), and Gradient Boosted Decision Trees (GBDT)-to quantify wind and solar variability, using metrics such as root mean square error (RMSE), continuous ranked probability score (CRPS), and empirical coverage. BNN is selected as the primary forecasting tool due to its superior calibration and robustness across both resource types. A distributionally robust optimization (DRO) model is then formulated using a Wasserstein ambiguity set and a Conditional Value-at-Risk (CVaR) objective to hedge against worst-case renewable output scenarios. Empirical data from a stylized VPP system comprising 340 kW of installed renewables and urban mixed loads is used to evaluate the framework. Results indicate that the proposed DRO model reduces expected shortfall by up to 78% compared to deterministic baselines, and outperforms quantile-based models in both cost consistency and energy-not-supplied metrics. Scenario-based dispatch simulations reveal that increasing the CVaR confidence level from 90 to 99% improves system reliability from 95.1 to 99.7%, albeit with a 5.7% decline in expected profit. The analysis also quantifies the relative contribution of wind, solar, and load forecast errors to overall dispatch uncertainty. This work highlights the value of integrated probabilistic modelling and risk-aware optimization in enabling reliable and economically efficient VPP operations.