Application Of Deep Reinforcement Learning Techniques In Office Building HVAC System Energy Consumption Forecasting

Building sector has become a significant energy consumer in China.As an important building service system which is responsible for maintaining indoor air quality and thermal comfort,building HVAC system accounts for around half of the building total energy consumption.In this context,accurate energy consumption forecasting is becoming increasing vital in taping the energy-conservation potential of building HVAC system,since it plays a fundamental role in many energy management tasks,e.g.,system fault detection and diagnosis and optimal operation strategy control.This thesis utilizes data-driven techniques and transform energy consumption prediction problem into sequential decision-making problem,then proposes a novel deep reinforcement learning(DRL)based methodology for building HVAC system energy consumption prediction.This thesis mainly investigates the effectiveness and performances of DRL techniques in office building HVAC system energy consumption prediction.Firstly,the energy consumption data of an office building situated in Henan as well as local meteorological data are collected,an expert variable is also introduced to help models better identify the system status.In order to enhance data quality and model performances,LOF algorithm,ACF and PACF are respectively adopted for data outlier detection and feature extraction.This thesis deploys three common DRL algorithms based models for both single-step ahead forecasting and multi-step ahead forecasting,of which the key hyper-parameters are respectively identified and optimized.The models are also comprehensively analyzed from three perspectives,i.e.prediction accuracy,convergence speed and computation time.The results demonstrate that the prediction accuracy of proposed DDPG and RDPG models measured by mean absolute error can be improved by 16%-24% for single-step ahead prediction,and 19%-32% for multi-step ahead prediction compared to common supervised models,which indicates the DDPG and RDPG models can obviously enhance the accuracy in office building HVAC system energy consumption forecasting.However,A3C model performs poor prediction accuracy in the tasks.A3C also shows much slower convergence speed than DDPG and RDPG,which indicates that the training process of A3C model needs more iterations than the other two models.In term of model computation time,DRL models are more time-consuming than supervised models due to their complicated structures and training schemes.Among three DRL models,A3C is the most efficient model,whereas the RDPG model with LSTM neural network embedded is the most time-consuming one.To sum up,the proposed DRL models can evidently enhance accuracy in office building HVAC system energy consumption prediction,while accounting for more computation time.With the upgrade of hardware configuration and the development of artificial intelligence technologies,the proposed DRL models have certain research and application values in energy consumption forecasting and some other associated fields.