Deep learning-based power consumption and generation forecasting for demand side management

With the evolution of smart grid, the demand side management involving integration of renewable energy source and load management can be achieved by analyzing the data generated by monitoring devices. The objective of the work is to forecast the power consumption and generation for demand side management using deep learning techniques modelled in python. The deep learning models including recurrent neural network, three univariant models of long short-term memory and gate recurrent unit are used to forecast the power consumption, and, solar and wind energy generation and performance comparison is done. The error metrics such as mean absolute, mean square, root mean square and mean absolute percentage values are analyzed for all models and the one with least error values is selected as the best model. Stacked long short-term memory gave best results for prediction of power consumption and solar energy-based generation with a mean absolute error of 0.017 and 0.018, root mean square error of 0.24 and 0.25 and mean absolute percentage error of 1.39 and 1.91 respectively. Similarly for wind energy-based generation recurrent neural network turned out to be the best model with mean absolute error of 0.072, root mean square error of 0.38 and mean absolute percentage error of 2.68. For this work Pennsylvania-New Jersey-Maryland interconnection hourly power consumption dataset, solar power generation data and wind generation data from international renewable energy agency and entsoe dataset were used for training and testing of the models.