Computer‐assisted demand‐side energy management in residential smart grid employing novel pooling deep learning algorithm

Abstract

Demand-side energy management increases the unpredictability and ambiguity of forecasting the load profiles of residential energy management. The energy management accuracy seems to be low by employing a traditional residential energy forecasting algorithm. This research work emphasizes on design and development of computer-assisted residential energy management by forecasting employing a deep learning algorithm. Hankel matrix is formed using copula function to process the collected automatic metering infrastructure (AMI) load data in the smart grid. From this data processing, model optimization was obtained by the proposed novel pooling-based deep neural network (PDNN). Moreover, this proposed PDNN avoid overfitting problem in training and testing by increasing AMI data variety and data size. The proposed PDNN is implemented in the TensorFlow platform. Based on real-time AMI southern grid data onto Tamil Nadu Electricity Board, India testing case studies was carried. Compared to traditional residential energy management techniques the proposed deep learning model outperforms support vector machine by 9.5% and 12.7%, deep belief network by 6.5% and 9.5%, and neural network auto aggressive integral moving average by 20.5% and 8.5% in terms of accuracy of energy forecasting and mean absolute error, respectively. Overall, the obtained results proved the effectiveness of the proposed deep learning algorithm for residential short-term load forecasting and management over other traditional methods. Highlights In this research work: Novel pooling-based deep neural network is applied for residential energy management in a smart grid. The copula fusion theory is adopted to improve the accuracy of load management in a smart grid. Day-ahead and week-ahead prediction load on Tamil Nadu Electricity Board dataset in summer and winter season was used to validate the performance of the proposed method with other data-driven methods.