Developing a reliable and robust algorithm for accurate energy demand prediction is indispensable for utility companies for various applications, e.g., power dispatching, market participation and infrastructure planning. However, this is challenging because the performance of a forecasting algorithm may be affected by various factors, such as data quality, geographic diversity, forecast horizon, customer segmentation and the forecast origin. Furthermore, an approach that performs well in one region may fail in other regions, and similarly, a model that forecasts accurately in one horizon may fail to produce an accurate prediction for other horizons. To overcome the above challenges such as rough data quality, different forecasting horizons, different kinds of loads and forecasting for different regions, this study proposes four machine learning/supervised learning models. These models are applied to improve the generalization of the network and reduce forecasting. These models are intended to simplify or demystify terms, complex concepts and data granularity used in energy forecasting. Two different data sites and four forecasting horizons are used to validate the proposed models. The coefficient of variation and mean absolute percentage error are 50% higher as compared with the existing model. The proposed supervised learning models ensure a generalization ability, robustness and high accuracy for building and utilities energy consumption forecasting. The forecasting results help to improve and automate the predictive modeling process while covering the knowledge-gaps between machine learning and conventional forecasting models.