Data driven model improved by multi-objective optimisation for prediction of building energy loads

Abstract

Machine learning (ML) has been recognised as a powerful method for modelling building energy consumption. The capability of ML to provide a fast and accurate prediction of energy loads makes it an ideal tool for decisionmaking tasks related to sustainable design and retrofit planning. However, the accuracy of these ML models is dependent on the selection of the right hyper-parameters for a specific building dataset. This paper proposes a method for optimising ML models for forecasting both heating and cooling loads. The technique employs multi-objective optimisation with evolutionary algorithms to search the space of possible parameters. The proposed approach not only tunes single model to precisely predict building energy loads but also accelerates the process of model optimisation. The study utilises simulated building energy data generated in EnergyPlus to validate the proposed method, and compares the outcomes with the regular ML tuning procedure (i.e. grid search). The optimised model provides a reliable tool for building designers and engineers to explore a large space of the available building materials and technologies.

Classification of retrofit measures for residential buildings according to the global cost

“…Although widely used in other fields, RF and other ensemble methods still do not have extensive application in buildings-related research [55]. RF has a limited number of applications to predict the building energy load [13,14], indoor temperature [38], occupancy [41,56], etc.…”

“…Artificial neural networks (ANN), support vector machines (SMV), decision trees (DT), random forest (RF), gradient boost (GB), linear regression (LR), hybrid methods, etc., are applied to predict energy demand, solar radiation, wind power, prices, and other important quantities. ML can predict building electricity consumption [11], heating or cooling loads [12,13,14], occupancy and window-opening behavior [15], etc. The predictive performance of various ML methods is verified and compared in refs.…”

“…[13,16,17,18]. These methods can be further improved by hybridizing and implementing optimization [14,19].…”

Generative BIM workspace for AEC conceptual design automation: prototype development

“…These recurrent failures have also (in part) been exacerbated by integration and interoperability challenges held by different stakeholder parties (Pour Rahimian et al , 2019), which supports the intrinsic need to enhance the veracity of design information throughout the project life cycle from the outset (Goulding and Rahimian, 2012). In essence, the nature and complexity of communication within AEC projects have changed significantly over the last ten years, especially advances in information and communication technology (ICT) and the increased prevalence of: virtual reality-based collaboration technologies (Pour Rahimian et al , 2019); artificial-intelligence-based optimisation (Pilechiha et al , 2020); data-driven decision support (Seyedzadeh et al , 2019); smart data modelling (Seyedzadeh et al , 2020); blockchain and distributed ledger technologies (Elghaish et al , 2020a); and computer vision and graphics (Pour Rahimian et al , 2020). Within the AEC sector, ICT has revolutionised production and design (Abrishami et al , 2014a; Pour Rahimian et al , 2008), which has led to considerable changes in labour and skills (Fruchter et al , 2016), where, for example, these advancements are now able to assist decision-making to predict the cost and performance of optimal design proposals (Elghaish and Abrishami, 2020).…”

Predicting Primary Energy Savings of Building Retrofit Measures With Decision-Tree-Based Ensemble Methods

“…Cui et al [12] applied RF and GB (as well as other methods) as a part of a hybrid approach that predicts the indoor air temperature that combines the resistor-capacitor and black-box models. Seyedzadeh et al [39] predicted heating and cooling loads with RF in the frame of a multi-objective optimization approach and Smarra [27] used if for data-driven model predictive control. Wang et al [40] found extreme GB the best approach to predict cooling load one day ahead.…”