Random Forests and Artificial Neural Network for Predicting Daylight Illuminance and Energy Consumption Muhammad Waseem Ahmad1, Jean-Laurent Hippolyte2,Monjur Mourshed3, Yacine Rezgui4 BRE Centre for Sustainable Engineering, School of Engineering, Cardiff University, Cardiff, CF24 3AA, UK 1AhmadM3@Cardiff.ac.uk; 2HippolyteJ@Cardiff.ac.uk; 3MourshedM@Cardiff.ac.uk; 4RezguiY@Cardiff.ac.uk Abstract ing performance of occupied spaces. One example of Predicting energy consumption and daylight illumi- this research work is Ahmad et al. (2015), the authors nance plays an important part in building lighting proposed a genetic algorithm based method to adjust control strategies. The use of simplified or data- the window blind position. On the other hand, energy driven methods is often preferred where a fast re- prediction strategies are one of the core components sponse is needed e.g. as a performance evaluation en- of building energy control and operational strategies gine for advanced real-time control and optimization (Li and Wen, 2014). applications. In this paper we developed and then In recent years, a number of prediction approaches, compared the performance of the widely-used Arti- either detailed or simplified, have been proposed and ficial Neural Network (ANN) with Random Forest applied for predicting building energy consumption (RF), a recently developed ensemble-based algorithm. and daylight illuminance. These approaches can be The target application was predicting the hourly en- broadly classified into three categories i.e. numerical, ergy consumption and daylight illuminance values of analytical and predictive. Numerical approaches (e.g. a classroom in Cardiff, UK. Overall, RF performed EnergyPlus, DAYSIM, RADIANCE, TRNSYS, etc.) better than ANN for predicting daylight illuminance; often enable the user to evaluate designs with reduced with coefficients of determination (R2) of 0.9881 and uncertainties. However, these methods do not per- 0.9799 respectively. On the energy consumption test- form well in predicting the energy use and daylight- ing dataset, ANN performed marginally better than ing illuminance of occupied buildings as it is difficult RF with R2 values of 0.9973 and 0.9966 respectively. to model occupants' behavior and how they interact RF performs internal cross-validation and is relatively with their buildings. On the other hand, predictive easy to tune as it has few tuning parameters. The models (e.g. artificial neural networks, support vec- paper also highlighted possible future research direc- tor machines, etc.) have been successfully applied tions. to predict the energy consumption of occupied build- ings. These models quickly perform predictions and Introduction thus are more suitable for real{time control purposes. Buildings are responsible for 40% of the total global In a recent review, Ahmad et al. (2016) discussed energy use and account for 30% of the total emission several computational intelligence (CI) techniques for of CO2, one of the greenhouse gases responsible for HVAC systems. It was mentioned that significant ad- anthropogenic climate change (Ahmad et al., 2016). vances has been made in the past decades on the ap- To mitigate this, building regulations have been de- plication of CI techniques for building energy appli- veloped or updated to reduce the impact of climate cations. Most of these techniques use historical data change and enhance the performance of buildings. to train a model or develop expert rules. With the With sustained reductions in a building's heating and evolution towards Internet of Things (IoT), there is cooling demands, the energy used by artificial lighting an abundance of data available from buildings, and increases in relative terms (Ahmad et al., 2015). Day- therefore these CI techniques can easily be applied to light is an essential part of our life, and building oc- enhance building energy performance. Random For- cupants tend to prefer daylight over artificial lighting. est (RF) has been less explored for building energy It also provides a comfortable and effective learning and daylight illuminance predictions. RF does not environment in schools. An appropriate lighting level require much fine-tuning of their hyper-parameters, is necessary to satisfy both psychological and visual and default parameters often give better results than comfort conditions. Typically Venetian blinds have being fine-tuned. On the other hand, artificial neural been used in buildings to control daylighting. In the networks (ANNs) have been extensively used for en- literature, various automatic control strategies have ergy and daylight predictions because of their fault{ been developed to enhance the thermal and daylight- tolerant and robust nature. The objective of this study is to develop ANN and RF based models to luminance in buildings. Hu and Olbina (2011) pro- predict the daylight illuminance and energy consump- posed an illuminance-based Venetian blind control tion of a classroom. This paper offers an alternative method and used ANNs to predict illuminance val- methodology to the existing energy consumption and ues at two set-points. The author concluded that the illuminance prediction techniques. proposed method has advantages for real-time blind control applications. Kazanasmaz et al. (2009) also Related work used an ANN to determine daylight illuminance for In the literature, a large number of studies have fo- an office building. The authors used building and cussed on using machine learning techniques to pre- weather parameters to predict illuminance values and dict energy consumption. For building energy predic- found that the prediction accuracy of the model was tion, ANNs are the most popular choice among other approximately 98%. Ahmad et al. (2015) proposed a computational intelligence techniques (Ahmad et al., method for controlling Venetian blinds by using Ener- 2016). In two different studies, both Gonz´alezand gyPlus as an evaluation engine. The authors stressed Zamarre~no(2005) and Nizami and Al-Garni (1995) the need for a surrogate model to reduce the com- used a simple neural network to predict hourly val- putational time required to run 1000s of simulations. ues of building energy consumption by using weather Mourshed et al. (2011) studied the optimum design and time stamp information as inputs to the mod- of artificial lighting and found that the search for an els. Nizami and Al-Garni (1995) compared the re- optimum design in a rugged solution space is a time sults with a regression model and it was found that consuming process, so there is a need to develop sur- ANN performed better. rogate models. ANNs were also used by Kalogirou and Bojic (2000) to predict the energy use of a passive solar build- Machine learning techniques ing. The authors developed different modules to pre- Random Forest dict outdoor and indoor air temperatures at next time step, as well as solar radiation and electrical Random forests (RFs) are ensemble-based decision heaters' state. Kreider et al. (1995) reported on the trees and were developed to overcome the shortcom- use of recurrent neural networks to predict cooling ings of traditional decision trees. In RF, like other and heating energy consumption. ANNs are also be- ensemble learning techniques, the performance of a ing used to predict energy consumption for different number of weak learners is boosted via a voting climate zones by using envelope performance parame- scheme. The main hallmarks of random forest in- ters, and heating and cooling degree days (Cheng-wen clude; 1) bootstrap sampling { randomly selecting and Jian, 2010). Manufacturing industries show high number of samples with replacement, 2) random fea- fluctuations in their energy use and modeling energy ture selection { randomly selecting only a small num- consumption for these buildings could be a challeng- ber of m features in the split of each node, 3) full ing task. Azadeh et al. (2008) predicted the annual depth decision tree growing, and 4) Out-of-bag error electricity consumption of this type of building by estimation { calculating error on the samples which using an ANN, where the results demonstrated their were not selected during bootstrap sampling (Jiang suitability for this purpose. et al., 2009). To the best of our knowledge, there are only a few In RF, a M number of decision trees are generated studies that focussed on the application of decision from a N number of training samples. For each trees for energy prediction.Tso and Yau (2007) and tree, bootstrap sampling is performed to create a new Yu et al. (2010) studied the use of decision trees for training set. The new training dataset is then used predicting energy demand and residential building en- to create a fully grown decision tree without prun- ergy performance respectively. Tso and Yau (2007) ing by using the 'classification and regression trees' compared the results from neural network, decision (CART) technique (Duda et al., 2012). Instead of trees, and regression analysis; and found that deci- using all available features at each split of the node, sion trees could be viable alternatives to understand only a small number of m features are randomly se- energy patterns. One key advantage of decision trees lected. This procedure is then repeated until M deci- is that the user can generate accurate models with- sion trees are created to form a randomly generated out having any computational knowledge. Yu et al. "Forest". For RF models, we used 1000 trees in the (2010) found that decision trees can produce accu- forest. One the hyper-parameters for RF is the num- rate models for predicting building energy use inten- ber of randomly selected variables mtry at each split sity levels. Ahmad et al. (2017) compared the results node; according to Breiman (2001), the recommended p of neural network and random forest (an ensemble- value of mtry is equal to p for classification prob- based method) for predicting hourly energy consump- lems (where p being the total number of predictors). tion, and found that ANN performed marginally bet- The author also mentioned that mtry << p should ter than random forest.