Complex-valued neural network for short-term load forecasting of Yogyakarta hybrid renewable power plant

• Main Authors: Muthasyr, Ringga Prasetya Al, Setyawan, Novendra, Irfan, Muhammad
• Format: Conference Proceeding
• Language: English
• Subjects: Algorithms; Alternative energy sources; Datasets; Electric power demand; Electrical loads; Electricity distribution; Energy resources; Energy sources; Error analysis; Forecasting; Gusts; Mathematical models; Model testing; Neural networks; Parameters; Power plants; Renewable energy; Renewable resources; Solar radiation; Water flow
• ISSN: 0094-243X; 1551-7616
• DOI: 10.1063/5.0192080

Renewable Energy Power Plants, which are electric power generation using natural potential such as wind gusts, water flow, solar radiation, and so on, are environment-friendly power plants. Although environmentally friendly, renewable energy power plants have a very fluctuating level of electricity production depending on the natural phenomena around them, so proper planning is needed to prepare for better management of energy resources, one of which is forecasting short-term loads. This study uses the Complex-valued Neural Network (CVNN) method as short-term load forecasting at the Baron Technopark Yogyakarta renewable energy power plant in order to prepare for better management of energy resources to deal with fluctuating electrical energy production. The four CVNN networks architecture is tested, which are the input layer, hidden layer, and output layer will be varied for each network. This forecasting model uses Adam Optimizer as an optimization algorithm with a ratio of (80:20) and (70:30). Parameters for model training used include the number of samples that will be used to update the neural network model parameters (batch size), the number of times the entire dataset will be entered into the neural network model (epoch). The proportion of datasets that will be used as validation (validation split) and then the network performance will be evaluated using mean square error (MSE) and mean absolute percentage error (MAPE). This study shows that the best network model tested is a network model with 96-time units of input, 64 input layers, 32 hidden layers, 32 output layers, and a ratio (80:20) with MSE results of 0.001 and MAPE of 9.37%. The predictive ability of the network can be categorized as very good because the MAPE value is less than 10%.