APPLICATION OF DATA REDUCTION AND ENSEMBLE LEARNING TECHNIQUES FOR BUILDING ENERGY CONSUMPTION PREDICTION

• Main Author: Thoriq Ahmada, Marsa
• Format: Final Project
• Language: Indonesia
• Online Access: https://digilib.itb.ac.id/gdl/view/57202
• Institution: Institut Teknologi Bandung

The challenge in measuring energy efficiency is not an easy thing because many factors can influence it. This problem can be overcome by building a predictive model to predict future building needs. The prediction results will be compared with the energy consumption of the repaired building to calculate savings. However, with the massive amount of data, several problems arise that cause data quality to decrease, resulting in poor model accuracy in predicting and lack of scalability in terms of memory usage if the data gets bigger. These problems can be overcome by applying data reduction techniques and ensemble learning In this final project, research has been carried out on data reduction and ensemble learning to overcome the problem of lack of data quality and the size of the data. Data reduction is a technique to reduce the size of the data both from the number of samples and the number of features without reducing the accuracy of the model. The data reduction techniques that will be carried out are numerosity reduction and dimensionality reduction. The ensemble learning technique combines several predictive models to become a combined model. For ensemble learning, what will be done is using the LightGBM model which is based on boosting and will also add a bagging technique that will be compared with incremental learning. The existing data will be subjected to preprocessing and feature engineering stages to increase the accuracy of the model. The best technique will be selected based on an RMSE metric. The best model is optimized with tuning parameters to get a more optimal prediction accuracy. After the experiment, data reduction, namely numerosity reduction with stratified sampling technique, can increase the speed to 2.67 times faster by maintaining the accuracy of predictions. For dimensionality reduction, PCA and ICA can increase the speed to 1.15 times faster without causing a significant decrease in prediction accuracy. In addition, the model with ensemble learning, namely bagging, has the best performance in terms of RMSE and speed with RMSE 262.304 and 1.67 times faster than the model with incremental learning.