Research On Building Energy Consumption Prediction Method Based On Deep Learning

Global carbon emissions have always been an international hot topic.In China,the annual building energy consumption accounts for about 1/5 of the China’s total energy consumption.In foreign countries,developed economies have reduced energy consumption and emission reduction in building energy consumption,achieving a 3.2% reduction in total energy-related carbon emissions.Therefore,the proportion of building energy consumption is relatively large and has the potential to save energy.In 2021,the two sessions clearly pointed out "carbon peaking and carbon neutrality",and included it in the government work report for the first time and listed it as one of the key tasks.In the 14 th Five-Year Plan,it is pointed out that we need to accelerate the development of “green and low-carbon”,among which green buildings are one one of plans.During the period of the building’s working,the energy consumption appliances will be generated due to heating,cooling,lighting and other needs.The task of studying building energy consumption prediction can make a more accurate prediction of energy consumption for a specific building,which is based on the development of green buildings.For building energy consumption patterns,there is a relatively complex scale sensitivity,so it is difficult for a singlegranularity prediction model to achieve an ideal prediction effect.Therefore,a prediction model based on multi-granularity patterns is proposed in the thesis.In this paper,we use the LOF outlier detection algorithm to detect abnormal values in the dataset,since the dataset has the peak energy consumption value of the normal power consumption peak,we manually mark the nodes in peak power consumption period but have high LOF score.Then,we organized them into a group called normal nodes list.And potential outliers nodes compute the mean euclidean distance from the feature vectors of the nodes in the normal nodes list.It is judged as an normal node if the average distance is smaller than the threshold,and is judged as an abnormal node if the average distance is bigger than the threshold,so that the peak energy consumption node can be excluded from the abnormal nodes list.In this paper,we proposed a building energy consumption prediction model.The granularity segmentation,feedback method and parallel convlutional layers of the multigranularity patterns can capture many different scale sensitivity features.And the hybrid attention method and long-short term memory layer can capture the long-term features.The building energy consumption exists different scale sensitivity,we construct some baseline deep learning prediction model for experiments on the building energy consumption datasets used in this paper.According to the experimental results,the MSE of the Mg Ha-LSTM model in the IHEPC dataset was 0.2813,93.21% of the Ms C-LSTM model.And the MSE of the Mg HaLSTM model in the EERE dataset was 0.0161,49.08% of the LSTM model.Therefore,the prediction results of the Mg Ha-LSTM model show that the model is more accurate than other deep learning prediction models.