Data science, finding knowledge and insight from a vast amount of data is the essential tool for the success of the intelligent energy management and ESCO (Energy Service Company) industries. The energy performance of an energy intensive facility and total energy consumption of a whole building are two representative and typical cases for the prediction modeling because of their wide applicability.
This thesis is intended to develop models for the energy performance and load prediction applying the data analysis techniques with an open source language “R”, and focused on a ground source heat pump system (GSHP) with 450 RT capacity composed of ten heat pump units and the entire hospital building, where the GSHP provides the heating and cooling energy.
The seasonal cooling and heating performance and operation properties of the GSHP system were analyzed using in situ monitoring data from Jun. 2016 to Mar. 2017. On this basis, hourly performance prediction models using the statistical models of a multiple linear regression (MLR) and an autoregressive distributed lag (ARDL) model, and the machine learning model of a deep neural network (DNN) were developed and compared.
The quantitative effects of influencing variables on the system performance, including the entering source temperature (EST) and the entering load temperature (ELT) were analyzed applying the structured ARDL model with statistical significance. The EST variation of 9 °C and ELT variation of 17 °C during this heating season can cause the performance changes by 29.3 ~ 34.0 % and 29.5 ~ 33.8 % respectively, based on the seasonal heating performance in the condition that other variables are kept constant.
The seasonal performance of the GSHP system was 3.21 for heating and 3.43 for cooling. The prediction accuracy of hourly heating performance was 3.4 % by ARDL, and 1.7 % by DNN, based on the coefficient of variation of root mean squared error (CVRMSE) without overall bias. The hourly performance prediction models can be used as a baseline for the measurement and verification (M&V) of possible future energy conservation measures (ECMs) and real-time performance monitoring to check malfunction of the system.
The prediction accuracy of the hourly heating load for the hospital building was 28.3 % CVRMSE by ARDL, and 28.9 % by DNN without overall bias. There was an accuracy improvement of 21.2 % by the ARDL model in comparison with the MLR model, not having the time lagged terms of variables. Unlike the system performance prediction model using detailed measurement data, the information on the outside weather obtained from the meteorological office, and occupancy was only used for the load prediction. The ARDL model can be applied as the baseline for evaluating the future changes of the heating load of the building.
Because the overall energy savings uncertainty is determined by combining the RMSE of a baseline model and measurement error, the level of measurement and the prediction model for the baseline should be decided in consideration of both cost for measurement and benefit from low uncertainty.
Accuracy and interpretation are two important values for a prediction model, but generally, there is a tradeoff between them. In this thesis, it is demonstrated that the use of both statistical and machine learning model is a worthwhile approach for the system performance and load prediction to achieve high prediction accuracy and interpretation of the coefficients. The MLR or ARDL is first applied to select relevant explanatory variables and interpret their quantitative effects, and then DNN is used to improve prediction accuracy by capturing nonlinear relationship among explanatory and response variables.
The method and procedure for the performance and load prediction modeling in this thesis can be widely applied to the various types of energy intensive facilities, buildings, plants and sub-systems.
