Forecasting Energy Consumption Using Deep Learning in Smart Cities

Abstract

Global energy demand is increasing continuously due to growth in the world population and industrial developments. In a parallel dimension, the problem of decreasing CO2 emissions in smart cities is becoming a priority. Forecasting energy consumption is essential for implementing a decarbonization plan in a smart city. The energy consumption forecasting problem has some challenges because of lacking appropriate data, including energy consumption patterns in the energy sector. In such a context, in this study, we focus on short-term time series forecasting for energy consumption tasks with comprehensive data. We employed LSTM, Transformer, XGBoost, and hybrid models to predict energy consumption via time series. The models were tested on the JERICHO-E-usage Germany dataset for Berlin, Düsseldorf, and the whole of Germany. We executed an energy consumption forecasting pipeline in our experiments to summarize Information and Communication Technology and Lighting energy types. Finally, we presented a comparative analysis between state-of-art deep learning and machine learning models (e.g., LSTM, Transformer, XGBoost), and a hybrid model. The proposed energy consumption forecasting pipeline can be applied to various countries and cities based on geographical distributions.