摘要 : The prediction for indoor temperature as an essential reference for system tuning has attracted great concern in the HVAC field. The existing researches mostly focused on transplanting novel machine learning algorithms, especially... 展开 The prediction for indoor temperature as an essential reference for system tuning has attracted great concern in the HVAC field. The existing researches mostly focused on transplanting novel machine learning algorithms, especially recurrent neural network (RNN), from the neuro-linguistic programming (NLP) like text generation and speech recognition into HVAC systems. But unlike semantics, the large-scale data processing during the training phase requires the high-speed processor and large memory, which restricts their online application in HVAC systems. In this work, we develop a strategy of improving gated RNN prediction for indoor temperature by estimating time delays of HVAC systems. The suggested strategy can lighten the computational burden and decrease memory capacity to favor online prediction and control, where the traceback range of error is narrowed down to the nearest-neighbors delay periods and the parameter learning time is reduced to about 2% of the original. Among them, we introduce transfer entropy (TE) into the HVAC field to estimate its time delays by mining monitoring data. Compared with correlation coefficients, it can filter redundant information between variables and discern the nonlinearity. Subordinately, we reconstruct time series by embedding time delays as inputs of the prediction model, and clip gradients within backpropagation through time (BPTT) with multivariate TE index, which promote the correlation between inputs and output and absorb the most valuable information in the parameter learning. Lastly, the proposed methods are validated with simulated and real time series. This work is enlightening and instructive to improving time series prediction in HVAC systems. 收起