Efficient Demand-Response Prediction in Smart Grids using Deep Learning

Smart grids (SG) envision efficient use of information and communication technologies for automated, efficient and sustainable electricity production and distribution. A network of grid distribution lie at the heart of these smart grids. Efficient, proactive energy usage estimation is an important research challenge in smart grid networks, as any discrepancy between electricity demand and supply raises cost for electricity providers and can potentially lead to outage. Successful electricity usage prediction holds the key for designing efficient and sustainable SG demand response (DR) programs. In this paper we explore different deep learning techniques, like Convolution Neural Network (CNN), Long Short Term Memory (LSTM) network and Gated Recurrent Unit (GRU) models to propose new efficient electricity usage prediction across major power grids of the United States. After pre-processing publicly available electricity usage data of some major electricity distributors in the United States, we perform extensive training and validation of our deep learning models. Our test results demonstrate that our designed CNN, LSTM and GRU models are capable of predicting accurate electricity usage with very low loss (< 0.01).