Tuning-free stochastic optimization for neural network training in building energy prediction

Building energy prediction is nowadays a crucial tool for reducing overall building energy consumption. Machine learning models can be used efficiently with this aim; however, performing feature selection and setting up a model is computationally demanding as it requires a large amount of training. Thus, it is crucial to adopt an efficient stochastic optimization method that reduces the computational burden associated with the single training phase as much as possible. In this article, a stochastic first-order trust-region method that does not require learning rate and batch size tuning is considered. Numerical experiments on a real problem show that the proposed algorithm can be used efficiently for training artificial neural network models in energy prediction, reducing the computational effort owing to the grid search hyperparameters tuning phase.