A hybrid PCA-LSTM predictive approach for enhancing multi-step indoor temperature prediction towards energy-efficient HVAC control

Abstract This study proposes a hybrid Machine Learning model called PCA-LSTM to improve the prediction accuracy of the indoor temperature using a minimal set of weather variables. In particular, this model integrates an LSTM layer dedicated to studying the variation in temperature time series, along with Principal Component Analysis (PCA) to reduce the dimensionality of the data. Moreover, to develop the n-step ahead PCA-LSTM predictive model, hourly observations of seven weather variables collected from the nearest weather station, including outdoor relative humidity, outdoor temperature, dewpoint temperature, wind speed, visibility, pressure, and electric heater consumption, collected from a passive house located in Belgium. As a result, the obtained results demonstrate a high-accuracy prediction in 15 min and 60min resolutions, compared to existing literature standards. This high predictive performance makes the model suitable for integration into an Energy Management System (EMS), enabling the control and management of the HVAC system. By utilizing the predicted indoor temperature along with a comfort temperature derived linearly from the outdoor temperature, the EMS can optimize energy consumption while maintaining indoor thermal comfort.