Multiple linear regression, support vector machines and neural networks for prediction of commercial building energy consumption

Abstract

Considering the constant growth of interest in energy efficiency in the building sector, it is necessary to apply and improve existing and also to develop new methods for prediction and analysis of building energy consumption. In this paper cooling consumption of the model of a typical commercial building in Belgrade is analyzed. Detailed energy simulation is done using software HAP (Hourly Analysis Program). The influence of various building characteristics is investigated, and for creating building consumption database, three variables that most largely affect the cooling consumption are chosen: specific lighting power, window area and window shade coefficient. Those three parameters are varied and 245 simulations in total are used for creating and testing the prediction models. The multiple linear model is created and the obtained equation is used for cooling consumption evaluation taking these three building parameters as input. The artificial neural network and support vector machine (SVM) models are also developed for prediction and their results are compared with linear regression model. It has been shown that the statistical methods, such are neural networks and support vector machines can achieve much higher accuracy in prediction than the linear regression model, gaining almost perfect match with simulated values (mean absolute percentage error for testing the SVM model 0,26%).

S obzirom na stalni porast interesovanja za povećanje energetske efikasnosti u zgradarstvu, neophodno je primenjivati i unapređivati postojeće i razvijati nove metode za predviđanje i analizu potrošnje zgrada. Na modelu tipične poslovne zgrade u Beogradu ispitivan je uticaj različitih karakteristika zgrade. Simulacija potrošnje energije na časovnoj bazi urađena je korišćenjem programa HAP (Hourly Analysis Program). Za dalju analizu izabrana su tri faktora koja u najvećoj meri utiču na potrošnju energije za hlađenje: specifična instalisana snaga osvetljenja, udeo prozora u spoljašnjem zidu i koeficijent propustljivosti Sunčevog zračenja kroz prozore. Analiza je vršena za različite vrednosti ova tri parametra. Za kreiranje i testiranje višeparametarskog modela korišćeno je 245 simulacija. Predložen je višestruko linearni model koji može da se koristi za određivanje potrošnje energije za hlađenje, a koji kao ulazne veličine koristi pomenuta tri parametra. U cilju predviđanja potrošnje, razvijeni su modeli primenom metode potpornih vektora (support vector machine) i veštačkih neuronskih mreža i izvršeno je poređenje rezultata sa višestruko linearnim modelom. Pokazano je da modeli zasnovani na metodi potpornih vektora i neuronskim mrežama postižu veću tačnost predvidjanja u odnosu na linearni višeparametarski model.