Temporal Fusion Transformer and transfer learning techniques applied to predict steam enthalpy with limited data in geothermal power plants

Japanese geothermal power plants are expected to be significant renewable energy sources owing to the abundance of geothermal sources in Japan. The plant capacity factor in geothermal power plants is low. One reason is frequent sudden pressure drops in production well corresponding to the change of subsurface condition. To obtain a stable steam quantity, it is necessary to observe the subsurface conditions in real-time and perform appropriate operations. The use of a model to predict steam enthalpy in real-time has potential to monitor changes in subsurface conditions and contribute to the composition of plant operational strategies. However, training a model requires a large amount of data. The purpose of this study is to evaluate the effectiveness of transferring the knowledge of a pretrained model for predicting steam enthalpy in one plant to another plant with limited data. This study proposes a methodology based on the combination of the temporal fusion transformer (TFT) architecture and transfer learning (TL). This approach represents a novel way to share knowledge from a pretrained model based on historical data from a plant, which helps reduce the need for large amounts of data when dealing with a new plant. A pretrained TFT model (PM) enables the prediction of rapid steam enthalpy decreases in the source plant. Transfer learning using a PM was confirmed to enhance the performance of steam enthalpy prediction in another plant compared to using a model without pretraining. The effectiveness of transfer techniques has the potential to contribute to improving the operational efficiency of geothermal power plants. The transfer learning strategies proposed in this study heavily rely on the similarity of the source data. In the future, we aim to compute data correlations between plants and the effectiveness of transfer learning.