Productivity prediction of a multilateral-well geothermal system based on a long short-term memory and multi-layer perceptron combinational neural network

Geothermal energy is one of renewable and clean energy resources. Predicting geothermal productivity is an essential task for managing a continuable geothermal system, which is a huge challenge due to the highly nonlinear relationship between the productivity and constraint conditions, such as reservoir properties and operational conditions. Using numerical simulation to predict the geothermal productivity is computationally expensive and very time consuming. Therefore, this study proposes a novel Long Short-Term Memory (LSTM) and Multi-Layer Perceptron (MLP) combinational neural network to effectively forecast the geothermal productivity considering constraint conditions. In the LSTM and MLP combinational neural network, MLP is trained to learn the non-linear relationship between the geothermal productivity and constraint conditions, while LSTM is used to memorize sequential relations within the production data. We comprehensively evaluate the geothermal productivity prediction performance of the LSTM and MLP combinational network. It indicates that the LSTM and MLP combinational neural network could accurately and stably predict the geothermal productivity and has a good generalization ability. Compared with original LSTM, MLP, Simple Recurrent Neural Network (RNN), the LSTM and MLP combinational network demonstrates the best geothermal productivity prediction accuracy, stability and generalization ability. This study provides a high precision and efficiency forecasting method for the geothermal productivity prediction.

Automated summary

Geothermal productivity prediction is crucial for managing geothermal systems. This study proposes a novel LSTM and MLP combinational neural network, which shows good accuracy and generalization ability in forecasting geothermal productivity.