Abstract:
The modern world is living in a digital age which demands more energy for continuous
operation of devices, gadgets, machines, household, etc. Consequently, there must be
some system for energy management in terms of harvesting, generation, reservation, and
balanced use. Which entails research for efficient machine learning forecasting models to
better predict future consumption. Since most of the modern machine learning models
are data-driven rather Engineering. Therefore for this study, we are focusing to predict
the building energy consumption using deep learning and other contemporary models
trained on real data captured for four years which was consumed by an educational
building situated in London, United Kingdom. These six machine learning classifiers
are, Bagging, Boosting, Random Forest, Deep Neural Network (DNN),Support Vector
Regression (SVR), Artificial Neural Network (ANN). The prediction models are fed the
same data for four years on various criteria such as outer temperature, solar radiation,
wind speed, humidity, and working day indicator. The last year data was used for testing
the predictive values of all the models. Results have shown that the last month in test
data seems to be an outlier as dropping it improves performance by 2 %. Furthermore,
the comparison also made for office day known as working day and non-working day
known as non office day using weekday indicator. The trained models are used to predict
electricity consumption units and all classifiers are compared with actual utilization units
of electricity for last year. Results reveal that ANN proves itself to the best of all five
approaches achieving a Mean Absolute Percentage Error (MAPE) of 6.41% where DNN,
SVR, Bagging, Boosting, Random Forest has achieved MAPE of 11.15%, 9%, 7.46%,
8.46% and 9.84% respectively. This work can be extended to other building energyrelated problems with respect to management, conservation, mitigation, and proper
utilization.
