Multi-Behavior with Bottleneck Features LSTM for Load Forecasting in Building Energy Management System

With the wide use of the Internet of Things and artificial intelligence, energy management systems play an increasingly important role in the management and control of energy consumption in modern buildings. Load forecasting for building energy management systems is one of the most challenging forecasting tasks as it requires high accuracy and stable operating conditions. In this study, we propose a novel multi-behavior with bottleneck features long short-term memory (LSTM) model that combines the predictive behavior of long-term, short-term, and weekly feature models by using the bottleneck feature technique for building energy management systems. The proposed model, along with the unique scheme, provides predictions with the accuracy of long-term memory, adapts to unexpected and unpatternizable intrinsic temporal factors through the short-term memory, and remains stable because of the weekly features of input data. To verify the accuracy and stability of the proposed model, we present and analyze several learning models and metrics for evaluation. Corresponding experiments are conducted and detailed information on data preparation and model training are provided. Relative to single-model LSTM, the proposed model achieves improved performance and displays an excellent capability to respond to unexpected situations in building energy management systems.

Automated summary

This study proposes a novel multi-behavior with bottleneck features long short-term memory (LSTM) model for load forecasting in building energy management systems, combining the predictive behavior of long-term, short-term, and weekly feature models. The model offers improved performance and stability compared to single-model LSTM, displaying excellent adaptability to unexpected situations.