Low-rank kernel learning for electricity market inference

Recognizing the importance of smart grid data analytics, modern statistical learning tools are applied here to wholesale electricity market inference. Market clearing congestion patterns are uniquely modeled as rank-one components in the matrix of spatiotemporally correlated prices. Upon postulating a low-rank matrix factorization, kernels across pricing nodes and hours are systematically selected via a novel methodology. To process the high-dimensional market data involved, a block-coordinate descent algorithm is developed by generalizing block-sparse vector recovery results to the matrix case. Preliminary numerical tests on real data corroborate the prediction merits of the developed approach.