Abstract
This paper evaluates a range of neural network (NN) architectures for predicting short-term unit commitment (UC) decisions in a hydro-dominated power system. A dataset is generated using an operational hydro scheduling tool for a real Norwegian watercourse. We compare simple, convolutional, recurrent, and hybrid NN models for multilabel UC prediction over 24-h and 168-h scheduling horizons. All models achieve high predictive performance, with F1-scores above 97.35% and clear improvements over the relaxed Mixed-Integer Programming (No_MIP) baseline. Convolutional architectures exhibit the most consistent results across both horizons. When embedded into the scheduling pipeline, machine-learning-predicted UC decisions reduce median runtime by up to 64.4% while maintaining objective values close to those obtained with an exact MIP solution.