Integrated attention mechanism for GBDT building energy consumption prediction algorithm

Building energy consumption prediction plays an essential role in power demand response, real-time grid balancing, and determining the optimal operation mode of buildings. This paper proposes a building energy prediction algorithm (AM-GBDT) that combines the gradient-boosted decision tree (GBDT) with the attention mechanism to improve further the accuracy of the building energy prediction model. The AM-GBDT building energy consumption prediction model avoids the limitations of traditional GBDT, which cannot identify important information nodes by effectively highlighting the degree of influence of energy consumption features of different prediction nodes. The employed method can effectively replace the feature engineering aspect of the traditional prediction process of building energy consumption that relies on building background experience, reduces the time complexity of the algorithm, and improves the accuracy and training speed of the energy consumption prediction model. This study uses 20 office buildings in Shanghai to validate the specific type of building's sub-energy consumption. Simulation results show that compared with traditional machine learning-based prediction methods of building energy consumption (XGBoost, LSTM), the AM-GBDT method proposed in this paper improved the predicted building energy performance metrics MSE, MAE, and R2 by 65.19%, 2.83%, and 6.22% respectively. At the same time, the training time of the model has significantly reduced.