Advanced techniques for electricity consumption prediction in buildings using comparative correlation analysis, data normalization, and Long Short-Term Memory (LSTM) networks: A case study of a U.S. commercial building

Abstract

This study introduces innovative assessment techniques to comprehend the effect of six data normalization methods, implemented through the LSTM algorithm, on predicting electricity consumption in commercial buildings. The focus lies on analyzing the relationship between various normalization process and its integration with LSTM method concerning building electricity consumption. The LSTM model incorporates input nodes from diverse sources, including weather data, plug load data, and occupancy ratio. Six distinct normalization process—Min-Max, Mean, Z-score, Gaussian, VSS, and IQR—are applied to assess the model’s evaluation on both training and test datasets. The study found that combining the LSTM method with Min-Max and IQR achieves lower figures, representing better performance and greater stability in comparison to alternative normalization techniques. These results described the critical role of data normalization in improving the performance of LSTM models, highlighting the importance of choosing suitable normalization techniques for specific applications while balancing improved accuracy against computational complexity. Furthermore, the study explores the correlation impact of variations in average input elements, particularly with a 5% increase and decrease in value, on electricity consumption in commercial buildings across different seasons. Plug loads emerge as dominant contributors to electricity consumption across various normalization methods and raw data, with temperature and humidity ratio exerting notable influence in specific seasons.