The traditional electric power grid is moving toward smart grid, and with the advent of smart grids, a lot of data is generated in high volumes, velocity and variety. This brings several challenges with real time processing to get meaningful information for enhancing the benefits of smart grid. Furthermore, the application of short-term load forecasting poses additional challenges of being highly uncertain and volatile due to different load profiles. This paper conducts a study on the demand side management and load forecasting in electric power grids with help of the historical data obtained from smart grid. Machine learning models are developed using deep learning to ensure significant improvement in forecast accuracy when compared to benchmark Auto-Regressive Integrated Moving Averages ARIMA analysis. The short-term load forecast data has been merged with weather data from Application Program Interface (API). The discussed system uses an autonomous feedback loop to consider the historical load values as features for training and testing. This paper also applies deep learning methods like pooling based recurrent neural networks which could solve the curse of dimensionality that usually exists with increasing layers in traditional neural network methods. The paper proposes an averaging regression ensembles model for short-term load forecast.