A Robust Prediction Method Based on Artificial Neural Network for Electric Power Production of Organic Rankine Cycle in Lahendong Geothermal Field

Abstract

The industry has adopted Artificial Intelligence (AI), and it is a significant opportunity for the industry; especially the application of Machine Learning (ML) to improve prediction results, which will affect decision making. The Lahendong Organic Rankine Cycle (ORC) power plant offers a unique cycle in power output optimization, with an additional cycle called the intermediate cycle from using the hot water. As the most potential renewable energy in Indonesia, predicting performance power plant output is a crucial task, by implementing ML along with the Real-Time Monitoring System (RTMS). The power prediction model follows the application of the Artificial Neural Network (ANN). A dataset with the Multilayer Perceptron (MLP) has been developed to find the most robust solution in prediction. It contains the historical measured variables from; brine temperature, brine pressure, hot water temperature, hot water pressure, and power output. The dataset is split into two parts; 70% for the data training and 30% for the method’s validation. Meanwhile, the accuracy between the power actual data and its prediction is evaluated by using normalized Root Mean Square Error (RMSE), Mean Absolute Error (MAE), and R-squared (R2) Overall, the prediction results from ANN are excellent. The R2 score around 0.9954 for the power data training, with a validation score of 2.9 % for MAE and 9.9 % for RMSE. The power validation data also shows an impressive result of 0.9955 R2, with a validation score of 4.7 % for MAE and 9.9 % for RMSE. Both results were categorized as excellent results, which R2 score close to one and a validation score under 10 %. Furthermore, this study surely can improve fast and accurate decision-making in Lahendong ORC power plant, especially for power plant management.