Answers are grounded in this paper and cite their sources.
Germany generated 54.5% of electricity from renewable energy in March 2019, according to the data collected by the Fraunhofer Institute for Solar Energy Systems. Forecasting power generation and consumption play an essential role in establishing a regional smart energy market. Numerous researches contributed to the field of power forecasting using machine learning and deep learning technologies. However, developing and perfecting energy markets lead to an unavoidable problem of adjusting the architectures of neural networks to adapt to new situations, e.g., new consumers or producers in the power grid. Another critical challenge is to learn new knowledge from the sequentially collected measurements efficiently and how to integrate the new information into the current neural network model. When retrained for a new task with a regular training process, neural network models could perform poorly on the previously learned tasks, which is referred to as the catastrophic forgetting problem. In this article, we design two real-world continuous learning scenarios for those challenges. The scenarios are based on the historical power data, which are obtained from a regional power grid in Germany. The results show that well-known continuous learning algorithms can be used to improve power forecasts with a sequential data stream in such scenarios. We believe that the work is the first step towards establishing an adaptively updating forecast system to assist the highly dynamic intelligent energy markets.
No takes yet. Share an insight, caveat, or question.
He et al. (2020) studied this question.
Synapse has enriched 10 related papers. Consider them for comparative context: