Performance-Driven Time-Adaptive Stochastic Unit Commitment Based on Neural Network

The low-efficiency and power imbalance risk have challenged the aging fixed time resolution scheduling, especially when facing largely penetrated renewable energies. Time-adaptive unit commitment (T-UC) is recently advanced to solve the issues. However, existing T-UC methods are subjective open-looped, thus may be still far from optimality. To further improve the T-UC, a performance-driven time-adaptive stochastic UC (T-SUC) based on neural network (NN) is proposed. It firstly leverages k-means++ on multivariate forecasts to settle dispatch resolution for SUC. Then, the SUC performances, involving computing efforts and power imbalance risks (PIRs) at the finest horizon, are encoded by neural network. The analyzing for the NN further allows us to feedback the performances to control dispatch resolution. Numerical studies justify that, compared to recent T-UC rivals, our method reduces over 40% of the PIR on the finest intraday time resolution, with the fastest elapsed time.