Disentangled Representation Learning Based Distributionally Robust Joint Chance Constrained Voltage Control for Distribution Networks

This paper proposes a data-driven hierarchical framework for distributionally robust joint chance constrained voltage control in distribution networks (DNs). Disentangled representation learning technology is employed to model the distribution of photovoltaic (PV) power at different buses under given predictions. Benefiting from the disentangled model, the correlations implied in the historical dataset are effectively captured. These include correlations between predicted errors and given predicted values, as well as the correlations between PV power at different buses. The KL-metric ambiguity set for PV power distribution is constructed with the reference distribution introduced by a disentangled conditional decoder. Based on the KL-metric ambiguity set, distributionally robust joint chance constraints are adopted. To transform the distributionally robust joint chance constraint into a tractable form, a reconstruction solution method with KL divergence is presented, and a corresponding acceleration technique for selecting support constraints is developed to reduce computation time while ensuring control effectiveness. The proposed method is validated using distribution networks of different scales.