Predicting Power Consumption Using Machine Learning Techniques

In modern society, power consumption plays a crucial role, since it is capable of influencing multiple sectors including residential, commercial, and industrial domains. It covers the electrical energy amount used by various devices, appliances, machinery, and systems within a specific time frame. The accurate prediction of power consumption is imperative for effective energy management, resource allocation, infrastructure planning, and cost optimization. In this study, we focused on predicting power consumption within DAEWOO Steel CO. Ltd, located in South Korea. We acquired, preprocessed, and analyzed a dataset containing information about the daily operations of the industrial facility, with data being sampled at 15-minute intervals. By leveraging machine learning techniques, we employed six tree-based algorithms and three ensemble learners to forecast the target variable. Our comparative analysis examined the performance of these regressors across three forecasting horizons: 15 minutes, 1 hour, and 1 day. We discovered that the efficacy of the regressors is complex and is linked to the forecasting horizon. Notably, the stacking ensemble learner outperformed others for the 15-minute horizon, achieving impressive metrics of 98.5% for D2, 99.9% for R2, 0.81 for MSE, and 0.37 for MAE. For 1-hour ahead predictions, XGBoost emerged as the most accurate model, attaining metrics of 96.3% for D2, 99.7% for R2, 53.07 for MSE, and 3.5 for MAE. Finally, for 1-day ahead forecasting, the Extra Tree regressor surpassed its counterparts, achieving metrics of 93.1% for D2, 99.3% for R2, 13944 for MSE, and 74.46 for MAE. These findings underscore the importance of tailoring predictive models to specific forecasting horizons and highlight the efficacy of ensemble learning techniques in enhancing power consumption predictions across varying time frames.