K-means clustering boosted hybrid CNN–LSTM models for indoor occupancy, temperature, and energy consumption prediction

This chapter proposes an unsupervised learning framework boosted by k-means clustering for predicting buildings’ indoor occupancy, temperature, and energy consumption. Unlike previous research, a new learning procedure is proposed by combining convolutional neural network (CNN) and long short-term memory (LSTM) models boosted by the K-means clustering technique to accurately predict indoor occupancy, temperature, and energy consumption. The K-mean clustering algorithm extracts features of multidimensional sensor data as input for a CNN–LSTM predicting the indoor occupancy and temperature. Another CNN–LSTM network outputs the daily energy consumption from the environment profiles and the expected occupancy and temperature. Three comparative models (CNN, LSTM, CNN–LSTM) with/without cluster features are investigated to verify the superior performance of the CNN–LSTM model compared to CNN and LSTM models, and the clustering algorithm significantly improves the prediction performance. The unsupervised learning method can boost the performance of the occupancy and temperature prediction at a mean square error (MSE) of 0.925 and 0.109 with cluster features (compared to an MSE of 0.961 and 0.191 without cluster features), respectively, to help the artificial neural network (ANN) predict the energy consumption more accurately up to an MSE of 1.859.