Prediction of Heat Energy Consumption by LSTM Sequence- to-Sequence Models

The accurate estimation of heat energy performance in buildings is critical for optimizing energy demand and supply. Non-residential properties have predictable operating patterns in principle, incorporating these patterns into simulations of energy consumption can help estimate building energy use. In this work we develop Long-Short Term Memory (LSTM) Sequence to Sequence and Gated Recurrent Unit (GRU) architectures, which are composed of Dropout, Re- peat Vector, Time-distributed and Graph Convolution layers. We have conducted a rigor comparative study on the structures and hyper parameters using the na- tional grid data, then use the learnt models for the energy demand site manage- ment undertaken in a laboratory environment.