A Deep Learning-based Framework for Energy Consumption Prediction and Anomaly Detection in Residential Buildings

Accurate prediction of energy consumption and anomaly detection are critical for ensuring a stable power supply and for efficient and optimal planning to meet future energy needs. The complex dependencies and non-linearities in buildings’ systems make it challenging to accurately forecast energy consumption using traditional and classical machine learning techniques. In this paper, we proposed a data-driven strategy-based framework for energy consumption forecasting and anomaly detection in residential buildings. Towards this, we proposed two deep learning-based techniques: (a) Bi-LSTM based network, and (b) hybrid CNN with Bi-LSTM based network for energy usage or demand forecast in residential buildings. Further, we integrated a moving average-based filtering technique for anomaly detection in energy consumption patterns. Experimental results show the better performance of proposed models over the existing baseline techniques including SVR, LSTM, GRU, CNN-LSTM, and CNN-GRU, with the lowest error scores for MSE, RMSE, and MAPE metrics, while CNN-GRU has the lowest error score for MAE of 0.151, which is much closer to MAE score of 0.155 and 0.153 for proposed model 1 and proposed model 2 respectively.