Electricity load forecasting using a deep neuralnetwork

• Main Author: Chawalit Jeenanunta
• Format: Article
• Language: English
• Published: Khon Kaen University 2019-03-01
• Series: Engineering and Applied Science Research
• Subjects: Deep neural network; Artificial neural network; upport vector machines; Forecasting; Short-term load forecasting
• Online Access: https://www.tci-thaijo.org/index.php/easr/article/view/116025/128513

Forecasting the daily load demandofan electric utility provideris a complex problem as it is nonlinear andinfluenced byexternal factors. Deep learning, machine learning and artificial intelligence techniques have been successfully employed in electric consumption load, financial market, and reliability predictions. In this paper, we propose the use ofa deep neural network(DNN) for short-term load forecasting (STLF) to overcome nonlinearity problems and to achieve higher forecasting accuracy. Historical data wascollected every 30 minutes for 24 hourperiods from the Electricity Generating Authority of Thailand (EGAT). Theproposed techniques were tested with cleaned data from 2012 to 2013,where holidays, bridging holidays, and outliers were replaced. The forecasting accuracy is indicatedbythemean absolute percentage error (MAPE). In this paper, there are two different training datasets,everyday training datasetwhich is arranged by day sequentiallyand same day training datasetwhich is separated seven groups of day (for e.g., only Monday training is used to forecast Monday). The outcomes of a deep neural network (DNN)are compared with an artificial neural network (ANN) and support vector machines (SVM) with an everyday training dataset. The empirical results reveal that the proposed DNN model outperforms the ANN and SVM models. Moreover, the DNN model trained with same day training datasets provides better performance than a DNN trained with an everyday training dataset for weekends, bridging holidays, and Mondays. Additionally, the DNN using a same day training datasets provides higher accuracies for December and January.