基于机器学习算法的城区建筑能耗不确定性及敏感性研究

Reliable and accurate building energy prediction is the key foundation of energy system planning, system configuration, operating economy and carbon emission calculation for community buildings. The uncertainty of variables in energy systems leads to the difficulty of estimating energy use for community buildings. Therefore, this project will implement forward/inverse uncertainty and sensitivity analysis by coupling ensemble and deep machine learning methods to quantify influential factors, calibrate energy models, propagate uncertainty, and optimize energy saving measures at community scale. The construction of variable uncertainty database for community building energy systems can be very useful to further analyze energy performance of community buildings. The probabilistic energy models can be created by extracting coordinates and attributes from geographical information system based on approximate Bayesian computation to solve the problems of unknown variables in energy analysis models. The mechanics of aleatory and epistemic uncertainty can be explored using the two-dimension Monte Carlo simulation to identify the patterns of change of energy use in community buildings. The global Sobol sensitivity analysis method will be used to determine the key variables influencing energy use and find out effective energy saving measures. This project would create a comprehensive theory of uncertainty analysis of energy patterns of community buildings and also realize the automation of calibrating and creating building energy models. The results from this research can provide accurate and reliable energy use information to form the theoretical foundation of the design and optimization of community building energy systems. This project will promote the development of ecological city community, low-carbon town, and green campus.

准确可靠的建筑能耗预测是城区建筑能源规划、能源系统配置、运行经济性和碳排放评估的重要基础。针对能源系统变量具有高度不确定性，导致城区建筑能耗预测困难的问题，采用正向和逆向不确定性及敏感性分析理论，耦合集成和深度机器学习算法，进行影响因素定量化、能耗模型标定、能耗不确定性传递和节能方案优化四方面研究。建立城区建筑能源系统变量不确定性数据库，为城区建筑能耗分析提供坚实数据基础；提取地理信息系统中建筑属性数据，基于贝叶斯算法，建立动态能耗模型，解决模型标定中未知参数确定难题；利用二维蒙特卡洛模拟，阐明随机和认知不确定能耗机理，揭示城区建筑能耗变化规律；采用全局敏感性法，解析影响能耗关键因素，确定有效节能措施。本研究将建立城区建筑能耗不确定性分析的系统理论，实现城区建筑能耗模型自动建立，提供精准城区建筑能耗预测，为城区能源系统设计及优化奠定理论基础，促进我国生态城区、低碳城镇和绿色校园等的发展。

针对缺少城区建筑能耗系统化不确定性及敏感性分析方法的问题，本项目以建筑热物理为基础，结合机器学习算法，根据正向和逆向不确定性分析方法，研究城区建筑能耗不确定性和敏感性。从5个方面进行深入研究：建筑输入不确定性数据库的建立；基于地理信息系统的城市建筑能耗模型建立；城区建筑能耗的正向不确定传递；基于近似贝叶斯的建筑能耗模型标定；基于序贯抽样的城区建筑能耗敏感性分析。以提供可靠的城区建筑能耗输入参数不确定性数据为目的，选择轻量型SQLite作为不确定性数据库工具，主要包括建筑围护、暖通空调、人行为、天气参数四类参数约440条数据，建立了一个全面、通用、开放、易用的建筑能耗输入参数不确定性数据库；基于地理信息系统，结合建筑能耗模拟软件与R语言环境，以天津科技大学河西校区建筑群为例，提出一套适用于城区规模建筑能耗模拟研究与节能潜力评估的方法，包括地理信息系统数据处理、自动化建模、机器学习代理模型建立、节能方案可视化评估等；基于二维蒙特卡洛不确定性分析理论，结合地理信息系统和机器学习算法，以天津科技大学泰达校区为研究对象，提出一整套基于城区建筑群能耗模拟的正向不确定性研究方法；耦合近似贝叶斯计算和机器学习标定建筑能耗模型，近似贝叶斯计算可在不计算似然函数的情况下得到模型未知参数的分布，机器学习方法通过建立可靠的建筑能耗机器学习模型，解决贝叶斯反演中建筑能耗模型计算成本过高的问题；提出基于序贯抽样的建筑能耗敏感性研究，动态调整建筑能耗模拟抽样次数，在满足敏感性结果收敛要求前提下显著减少模型计算时间，并提供准确稳定的敏感性结果。本研究提出了城区建筑能耗的不确定性和敏感性分析系统性理论方法，在建筑能耗领域有重要的学术价值和理论意义，对我国建筑节能减排工作中绿色城区、生态小镇、绿色校园等的发展有着非常现实的应用价值。