.png){kind=icon}
.png){kind=icon}
Return
Load forecasting for smart grid based on continuous-learning neural network
.png){kind=icon}
PRE
AI
DOI：10.1016/j.epsr.2021.107545.png)
.png){kind=icon}
Original request for help.png){kind=icon}
Share.png){kind=icon}
Save
Abstract
En 中文
Smart grids (SG) are concepts based on the composition of several sources of electric power generation (synchronous, wind, photovoltaic generation, among others) operating in the same system forming a complex operational arrangement. In order to make its operation feasible, the implementation of modern technologies, especially regarding artificial intelligence, is necessary. In this new scenario, the load forecasting will have to offer new solutions. In this sense, it is proposed a load forecasting system applied in the SG environment. The proposed solution consists of using a Fuzzy-ARTMAP (FAM) artificial neural network (ANN). The training of this ANN is performed through the use of historical databases in order to extract the initial (basilar) knowledge. Parallel to the load forecasting, in FAM-ANN a routine named continuous-learning (CL) is implemented that will take care of the extraction of knowledge in an incremental way using the data that are provided by the real-time measurement system. As this information becomes available, the CL device performs the necessary calculations aiming to improve the FAM ANN synaptic matrix (of weights), continuously, without the need to restart the entire process, when new information is available. It is a solution that improves over time, including situations that are not usually foreseen. This design is properly aligned with the needs of SG systems. This formulation has been possible to be developed, mainly due to the property of the ANNs of the ART family, which is plasticity. Certainly, it is a differential in relation to most publications in the literature. In order to test the proposed methodology, a historical database of a company in the electric sector was used, producing a MAPE of around 5% without CL performance and, in most cases, less than 2% when considering the execution of the CL.
Keywords:
Electric power systems
Global Load forecasting
Artificial neural networks
Adaptive resonance theory
Continuous-Learning
Journal
IF:
4.2
Papers:
1.1W
Citations:
2.2W