Predicting Photovoltaic Power Output with Convolutional Neural Networks: A Case Study in Cepu, Central Java, Indonesia

• Published in: Journal Europeen des Systemes Automatises Vol. 56; no. 4; p. 583
• Main Authors: Muhammad K. Asy’ari, Nugraha, Azizah L, Sahrin, Alfin, Rafi, Tajuddin A, Indriawati, Katherin, Musyafa, Ali
• Format: Journal Article
• Language: English; French
• Published: Edmonton International Information and Engineering Technology Association (IIETA) 01.09.2023
• Subjects: Alternative energy sources; Atmosphere; Electric currents; Electric power; Electricity; Electricity distribution; Energy industry; Energy transition; Industrial plant emissions; Light; Neural networks; Per capita; Power plants; Radiation; Renewable resources; Solar energy; Indonesia
• ISSN: 1269-6935; 2116-7087
• DOI: 10.18280/jesa.560408

The prediction of the general national electricity plan has stated that Indonesia's electricity needs in 2038 will increase to 1,000 TWh or the equivalent of 3.3 MWh per capita. If this trend continues, then by 2050, per capita energy consumption is expected to reach 7.7 MWh or 2,600 TWh. This prediction is important because it is to prepare power plant infrastructure throughout the region for the coming period.By applying and adding learning algorithms based on Convolutional Neural Networks (CNN) to developing solar power plants, CNN-based Forecast can increase accuracy by up to 30% and has the ability to train models up to 2 times faster than currently available algorithms. This new algorithm can more accurately detect leading demand indicators, such as pre-order information, increased product demand, price changes, and promotional spikes, to build more accurate forecasts.That before the discovery of CNN there was no algorithm that provided the most accurate estimates for all types of data. Traditional statistical models have been useful in predicting the demand for products that have regular demand patterns, such as summer or winter electricity production. However, statistical models cannot provide accurate forecasts for more complex scenarios, such as frequent changes in energy prices, differences between regional versus national demand, products with different selling speeds, and additions of new products.As for CNN, it is appropriate to predict the power from the panel, with a MAPE performance value of 18.7633%, MAE of 0.0176 and RMSE of 0.0466 so that the prediction system built has fulfilled the prediction of the targeted system.