Supervised based machine learning models for short, medium and long-term energy prediction in distinct building environment

Abstract

The substantial measure of energy usage connected to the building atmosphere supports and sustains power usage modeling diligence. Amongst the numerous strategies to elaborate energy methods, supervised based machine learning approaches are immeasurable alternative to circumvent the inconvenience correlated to various engineering and data mining approaches when measured/observed data are accessible. This research depicts an analysis of electricity requirement forecasting by supervised based machine learning models with the limited data information. The power usage or energy consumption data is collected from power transmission and distribution networked organization independent system operator New England for one-year ahead energy forecasting. Moreover, energy consumption data is categorized into monthly, seasonally and yearly basis to foresee the performance for short-term, medium-term and long-term as well. Four-supervised based machine learning models employed for energy forecasting which are: i) Binary Decision Tree; ii) Compact Regression Gaussian Process; iii) Stepwise Gaussian Processes Regression; iv) Generalized Linear Regression Model. The input variables comprise the limited external environmental data, day-type/hour-type and the net energy consumption of various types of load. The output is the total energy demand of the building power usage. Modeling studies are escorted for expected energy demand in future perceptive based on Independent System Operator New England data. The performance evaluation indices applied in evaluating the model's forecasting error are coefficient of variation and mean absolute percentage error. In autumn season, the best MAPE and CV of the binary decision tree is 0.809% and 1.359% respectively for seasonal forecasting, and is 0.989% and 1.601% respectively for yearly forecasting. It is observed that the accuracy in forecasting is modest in the autumn season. In yearly prediction, the MAPE and CV of compact regression Gaussian process, stepwise Gaussian processes regression and generalized linear regression are 3.245% and 3.650%, 4.039% and 4.860%, 5.118% and 5.927% respectively. The machine learning model's performance compared and validated with the actual energy consumption, existing artificial neural network model and the mean absolute percentage error and coefficient of variation found 2.416% and 3.290% respectively for yearly prediction. It is depicted that including the utilization of limited energy usage and environmental data as one of the model's input variables, the electricity forecasting precision is more accurate, precise and can be improved.