LSTM-based estimation of energy consumption in energy-intensive facilities

Abstract With rising energy demand, it is crucial to diversify energy sources and enhance energy efficiency. This urgency is particularly evident in energy-intensive facilities, where efficient energy production and consumption management play a pivotal role in optimizing energy utilization. In this study, a long- short-term memory (LSTM) based energy consumption estimation model was developed for a steel production plant as a representative energy-intensive facility. LSTM models excel in handling time-series data because of their ability to store sequential data in memory using memory blocks. The proposed model was subjected to a performance comparison with a traditional back-propagation neural network (BPNN) and regression-based models. Utilizing three-year production and consumption data, along with heating and cooling degree day information specific to the plant, the models were developed and tested. We validated and analyzed the models using 8640 data entries from the fourth year, which were not included in the training data. The evaluation based on the root mean squared error parameter yielded error values of 0.0586, 0.0608, and 0.0879 for the LSTM, BPNN, and regression models, respectively. Notably, the LSTM model achieved approximately 3.6% and 33.3% lower error rates in predicting the fourth year’s energy consumption compared to the BPNN and regression methods, utilizing data from the previous three years in an energy-intensive facility. In conclusion, the LSTM model introduced in this study offers a systematic approach to energy consumption forecasting, proving its effectiveness for energy-intensive facilities due to its flexible structure and ability to model nonlinear relationships.