Enhanced Electricity Forecasting for Smart Buildings Using a TCN-Bi-LSTM Deep Learning Model

Abstract

Integration of sensor technology and advanced software empowers consumers to manage energy usage proactively. This proactive approach yields positive impacts at both micro and macro levels, benefiting individuals and contributing to broader environmental conservation efforts. By leveraging predictive models, consumers can make informed decisions that serve their interests and promote a greener and more sustainable future for all. Thus, energy consumption (EC) prediction is crucial for effective resource management. In this study, we propose an innovative deep-learning approach to predict EC, focusing specifically on smart buildings. Our model utilises a hybrid deep learning architecture to effectively capture low and high information patterns present in multivariate time series data of various sensors deployed in smart buildings and numerous influencing factors. To address the nonlinear and dynamic nature of this data, our model combines a deep neural network (DNN) with a deep learning sequential model (DLS). Specifically, temporal convolutional networks (TCN) within the DNN family are employed to extract various trends from the data, while the DLS model, which consists of Bi-directional Long Short-term Memory Networks (Bi-LSTM), is employed to learn and capture these trends effectively. Consequently, we present a hybrid deep learning framework that leverages for learning multivariate time series data related to EC with shared feature representation. To validate our approach, we extensively evaluate our model using a dataset from an office building in Berkeley, California. Experimental results demonstrate that our model achieves satisfactory accuracy in EC prediction. For the 7-h horizon and on multivariate TS data, an R² of 0.97 is realised for the proposed model. This is confirmed by the 1.65% improvement in transiting from univariate to multivariate data, which supports using multiple modalities.