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

Highlights

• This study proposes an ML model optimisation method for building energy loads.

• The research carries out specific tuning for ML model using simulated building energy data generated in EnergyPlus.

• Feature importance method is used to highlight the relevance of input variables on the prediction of energy loads.

• The paper suggests a powerful decision-making tool for building design and retrofit planning.

• The optimisation method reduces the time complexity and increases the accuracy of energy modelling

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 decision-making 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.

Introduction

There have been several approaches proposed to enhance the energy efficiency of buildings in many countries in recent decades. For instance in Europe, it was estimated in 2010 that 60 billion Euros could be saved annually by improving EU buildings' energy performance by 20% [1].

Every attempt to optimise the energy performance of buildings involves a series of calculations to estimate the energy consumption and create an index, such as an ‘energy performance indicator’ or ‘use intensity’ from the measured data [2,3]. Most prevailing optimisation methods are simulation-based where the energy-related objectives (i.e. energy consumption or gas emissions) are calculated by a Building Performance Simulation (BPS) tool such as EnergyPlus, TRNSYS and ESP-r. This approach restricts the computing complexity of the algorithms to BPSs' calculation time. As such, when a vast range of solutions are defined, the calculation and optimisation process may become extremely costly and cumbersome [4]. For this reason, most of the studies which focused on decision making for energy performance improvement of buildings either investigated basic and simple optimisation models or targeted retrofitting only one or two parts of envelopes to pare-down total calculation time and cost. It should also be noted that the majority of studies targeted residential buildings, and there are only a few examples of research related to tertiary buildings. A key component of achieving global development and meeting climate change mitigation targets is the optimisation of the entire building stock. This process requires significant testing and planning to deliver.

With the tremendous growth in the amount of valid and attainable datasets of buildings and collection of Big Data from smart buildings, there is an increasing interest in the employment of Artificial Intelligent (AI) methods specifically Machine Learning (ML) techniques for analysing, modelling, and predicting building data [4,5].

The precision and suitability of the data and the relationships inferred from it become a critical fact in the successful application of ML models. As ML methods build s model over a historical dataset, the main and most important step for having accurate predictions is the extraction of relevant features. Depending on the nature of predictions (the energy indicator and forecasting period), this variable could include simple basic weather indices (e.g. temperature and humidity) or complex building characteristics and climate parameters. Previous research has demonstrated that rather than feature extraction the process of tuning a model itself not only increases the predictive accuracy but also reduces model complexity, ease of use, and consistency of predictions [6]. It has been argued that considering occupancy in retrofit decision-making, particularly in populated real estate properties such as higher education buildings, could leverage energy efficiency [7].

ML techniques have been widely used for modelling building energy loads and performance. Traditionally, the default values for hyper-parameters have been used in this field. However, in recent years researchers have started to tune the ML models to have more accurate predictions of energy metrics [[8], [9], [10], [11]]. Tuning ML model hyper-parameters using a grid search can be time-consuming when a complex method is chosen such as Artificial Neural Networks (ANN) [13] or models based on decision trees.

When MLs are utilised for forecasting multiple measures such as heating and cooling loads, models need to be optimised for both the targets [12,13]. This procedure, in turn, increases the time required for processing and improves the usability of MLs.

In the proposed method, evolutionary-based multi-objective optimisation (MOO) algorithm was employed to smartly explore the ML model's configuration parameters space and suggest a set of packages for maximising ML accuracy for both heating and cooling load predictions. This study applied a Random-Forest (RF) model because a python implementation is capable of providing the multivariate forecasting.

Section 2 provides an overview of the preceding studies with regards to tuning ML models with the purpose of building energy indicators forecasting. Afterwards, the RF method and the studied dataset are described in Section 3. Section 4 presents the results of the proposed ML optimisation approach. The final section provides detailed discussions and recommendations for future work.