Long-Term Energy Usage Prediction in Public Buildings Using Aggregated Modal Decomposition and GRU

Abstract

According to the Global State of Building and Construction Report 2024, the building sector accounts for one-fifth of global greenhouse gas (GHG) emissions. High energy consumption in buildings is destroying the environment; causing air pollution, the greenhouse effect, and the urban heat island effect; and causing great harm to social and economic development. Public buildings are of great concern due to their high energy consumption per unit area, low energy efficiency, and prominent energy waste. By accurately predicting energy consumption, energy use strategies can be optimized to improve energy efficiency and reduce energy consumption, which helps to reduce carbon emissions from buildings. This is of great significance in addressing global climate change and realizing sustainable development goals. Building energy consumption, as typical time series data, is affected by various factors such as dew point temperature, barometric pressure value, and wind speed. Therefore, how to construct accurate and reliable energy consumption prediction models is an important area of research in the field of construction worth further investigation. This study proposes a method for predicting energy usage using aggregated modal decomposition and gated recurrent units (GRUs). The model is developed by creating a number of smooth component sequences from the original random energy usage time series data, clustering them by the K-shape method, and, in order to predict each internal modal function, the GRU prediction method is adopted. Last, the total prediction is produced by combining the predictions made by each component. In order to demonstrate the accuracy of the prediction algorithm chosen in this study, several comparative studies were conducted, and to verify the generalization of the model, five buildings with different uses were used for the tests. Compared to other models, the model predicted values with minimum values for the error metrics root mean square error (RMSE), mean absolute error (MAE), mean absolute percentage error [MAPE (%)], and maximum accuracy (R2). To evaluate the scalability of the model in real-time applications, we conducted several experiments to test the model’s performance with other data sets, different lengths of time (quarterly, half-yearly), and computational resources. The results show that the model has good scalability and can maintain high prediction accuracy and response speed with increasing data volume and computational resources. In the future practical operational environment, we can deploy the building energy consumption prediction model in a simulated real-time system. First, export the trained model into an appropriate format; package it into an application programming interface (API) using a framework such as Flask or FastAPI; design a data preprocessing module to handle the real-time data streams; and collect and process the building operation data, such as the environmental conditions (temperature, humidity) and equipment status (HVAC, lighting), by using a tool such as Apache Kafka or Flink. Then, build the system architecture, including data reception, preprocessing, model inference (using TensorFlow Serving or TorchServe, etc.), and result output (e.g., presentation via Dash or Grafana). Finally, perform system integration testing and monitor performance with tools such as Prometheus.