LSTM with shallow NNs for indoor temperature long-term predictions in refrigeration systems

Machacuay J.; Manrique-Silupu J.; Ipanaqué W. · Neural Computing and Applications (2025) · DOI: 10.1007/s00521-024-10477-8

Resumen

Refrigeration systems account for a significant percentage of the energy consumption in buildings and industrial cold storage. Model-based control approaches, such as model predictive control (MPC), overcome classical control shortcomings and optimize energy consumption and expensive operating costs. However, they require an accurate long-term indoor temperature prediction model, and finding such a model is still challenging. This paper modifies the standard long short-term memory (LSTM) model used in the literature by adding shallow neural networks (NNs), named ShallowLSTM, that are able to learn different hidden representations from high-dimensional embeddings to improve the predictions in longer prediction horizons. Two versions of the ShallowLSTM with and without input gate of are presented. Experimental results on a 120 min-ahead prediction horizon show that the ShallowLSTM achieves better results than the LSTM model. Specifically, we found a model that achieves a mean absolute error of 0.0762 over the 0.2149 of the standard LSTM model, even with less parameters, thus providing empirical evidence of our proposal to be a candidate for improving long-term indoor temperature modeling and a candidate model for model-based control approaches. © The Author(s), under exclusive licence to Springer-Verlag London Ltd., part of Springer Nature 2024.

Instituciones: Universidad de Piura