Deep Learning-based Indoor Temperature Forecasting In Large--scale Public Buildings

Indoor Temperature Forecasting in large public buildings relies on mining historical temperature time series data to predict future temperature distribution and optimize energy management using advanced control techniques.It finds applications in energy optimization control of HVAC systems,demand response for building electricity consumption,and dynamic monitoring of integrated energy systems.However,indoor temperature time series data exhibit characteristics such as long and short-term periodicity,coupled multivariate features,and spatio-temporal correlations.Existing methods struggle to accurately extract complex temporal features and fully exploit coupling relationships between variables,resulting in suboptimal prediction performance.Significant progress has been made in the field of indoor temperature forecasting,particularly with the rapid development of artificial intelligence techniques,especially deep learning.This study focuses on indoor temperature scenarios in large public buildings and performs a detailed analysis of temporal features.It utilizes deep learning to extract features and designs an accurate temperature prediction model,enabling precise forecasting of indoor temperature distribution.The main research contents and innovations of this paper are as follows:(1)This thesis proposes a feature extraction and deep learning-based model,termed MS-ATCN(Multi-Scale Attention-based Temporal Convolutional Network),for single-zone(single-variable)indoor temperature prediction.Existing single-zone indoor temperature prediction models fail to adequately capture the inherent long and short-term periodicity in time series data.To address this issue,this thesis thoroughly analyzes the necessity of considering the underlying long and short-term periodic features in the feature extraction process for single-zone indoor temperature prediction.The MS-ATCN model is introduced,which utilizes a multi-scale input construction module to divide the data into inputs representing long and short-term periodicity.Additionally,an attention mechanism dynamically weights the long-term periodic input data,ensuring the accurate capture of long-term periodic features across different time scales.Subsequently,the TCN(Temporal Convolutional Network)model is employed to extract both long and short-term periodic features.After feature fusion and dimension matching,the model provides predictions of future indoor temperature for a given time horizon.Experimental results demonstrate that the proposed MS-ATCN model accurately extracts the long and short-term periodic features present in indoor temperature time series data.Through detailed comparisons with existing research in the field,the model achieves a 66.84%improvement in MAPE(Mean Absolute Percentage Error)on the test set compared to the best-performing baseline model,further validating the feasibility of the proposed model in the context of single-zone indoor temperature prediction.(2)The DL-GCN(Deep Learning-based Graph Convolutional Network)model is proposed for multi-zone(multi-variable)indoor temperature prediction,incorporating feature fusion and deep learning.Existing multi-zone indoor temperature prediction models often fail to fully capture the potential coupling features between variables and spatial correlations present in the data.To address this issue,this thesis extensively analyzes the necessity of considering the coupling features between variables and spatial correlations in the feature extraction process for multi-zone indoor temperature prediction.The DL-GCN model is introduced,which employs an LSTM network to extract temporal features and a GCN network to extract spatial features.Furthermore,a distributed fusion architecture is utilized to capture the coupling effects of other variables on indoor temperature variables.The extracted features are deeply fused to generate predictions of future indoor temperature for a given time horizon.Experimental results demonstrate that the proposed DL-GCN model accurately extracts the coupling features between variables and spatial features present in multi-zone indoor temperature time series data.The model achieves a high average correlation(CORR)of 95.4%between the actual and predicted values on the test set,and the mean squared error(MSE)metric shows a 66.84%improvement compared to the best-performing baseline model.This further confirms the advantages of the spatial,temporal,and distributed information fusion modules in multi-zone indoor temperature prediction in large public buildings,validating the feasibility of the proposed model in the context of multi-zone indoor temperature prediction scenarios.