| With the rapid development in urbanization,building energy conservations and traffic flow have evolved to be the key in construction of smart city.The reliable and accurate prediction and diagnosis of multi field measured data helps to facilitate the intelligent decision-making of building energy management and traffic improvement.Conventional approaches rely on domain expertise and engineering experiences for multi field measured data analysis,which is neither efficient nor effective for practical applications.The recent development in information technologies has provided an ideal platform for developing data-driven solutions for multi field measured data predictions.By integrating with advanced algorithms in machine learning,such data-driven solutions can be fully automated and achieve very high accuracy.Nevertheless,the reliability of data-driven models is highly dependent on the quality of training data.In general,there are three main data problems.Firstly,a typical data collection system may have measurements with a large number of variables.The intrinsic multicollinearity embedded in such high-dimensional data imposes great challenges in reliable model development,such as the increase in model complexity and computation costs.Secondly,the implementation of advanced supervised learning algorithms typically requires a large amount training data,which may not be available for individual buildings and traffic due to the lack of data accumulation time or regular data collection activities.Thirdly,the ground truths or labels of multi field measured data are generally not available in practice,making it infeasible to directly adopt the manner of supervised learning for model development and evaluation.To tackle the above-mentioned problems,this study develops a specific data preprocessing framework for analyzing multi field measured data.To summarize,the framework consists of three functionalities.The first is to develop a deep learning-based solution for data dimensionality reduction and feature engineering.The method is based on the use of one-dimensional convolutional deep autoencoders.Compared with traditional dimensionality reduction methods,our solution can better preserve the data characteristics using nonlinear data representations.The research results indicate that the nonlinear features obtained could achieve better prediction performance.Secondly,this study develops a generative adversarial network-based methodology for data augmentation.The method is especially useful for buildings and traffic with limited measurements.In such a case,synthetic data are generated to simulate the true data distribution and used to enhance the reliability in implementing machine learning algorithms.Thirdly,a semi-supervised methodology is developed to tackle the practical data challenge of unlabeled data.The methodology utilizes the clustering analysis to learn the intrinsic data characteristics,based on which synthetic labels are generated in a data-driven approach and used for reliable supervised learning.This research provides theoretical solutions and practical tools for tackling typical data problems in the multi field.The research outcomes are helpful for the development of data-driven methods for intelligent building energy management and traffic improvement. |
