UC Irvine

In response to the growing challenge of energy and power management caused by increasing

implementation of volatile sustainable energy sources, a case study of forecasting building

electric loads using statistical and machine learning methods is conducted for a multi-purpose

LEED-certified institutional building on the UC Irvine campus. Four data-driven methods,

which require no detailed building information and strong building energy knowledge, are

employed and compared, including the polynomial regression, Autoregressive Integrated

Moving Average (ARIMA), TBATS (Trigonometric seasonal formulation, Box-Cox

transformation, ARMA errors, Trend, and Seasonal components), and backpropagation Artificial

Neural Networks (ANN). These models are investigated to satisfy the ASHRAE standards and

optimize the prediction performance. A full year of hourly electric load and meteorological data

of the building in 2019 was obtained using the existing meters for this data-driven study. Root

Mean Square Error (RMSE), Coefficient of Variance (CV), Mean Absolute Percentage of Error

(MAPE) and R2 are calculated as evaluation criteria to compare the performances of these datadriven

methods in terms of prediction accuracy. Akaike’s Information Criteria (AIC) is

introduced as a guideline to determine the optimal model for several prediction models. The polynomial regression is performed using MATLAB and is shown capable of only data fitting

instead of forecasting when the total hour is used as the independent variable. When using the

daily and weekly data, the polynomial regression method fails for forecasting. For the wholemonth

data, ARIMA, TBATS, and ANN methods are used to predict hourly power load in the

next month with Python. The ARIMA model shows relatively low accuracy, indicating that it is

unable to handle multiple seasonlities in the data. TBATS shows a substantially improved

accuracy and satisfactory prediction. The backpropagation ANN is also conducted with its

configuration, including inputs, number of hidden layers and neurons, optimizer, and activation

functions, optimized after extensive testing. Different sets of training data are examined for both

TBATS and ANN. The ANN’s forecasting accuracy is found to be about 5~20% better than

TBATS’ when only using one month’s data for training. The residuals of these forecasting

methods show there could be information uncaptured in forecasting. It is speculated that

operation and activity schedules can serve as additional inputs for the ANN to achieve better

forecasting accuracy.