GNGTS 2024 - Atti del 42° Convegno Nazionale

Session 3.1 GNGTS 2024 We did this with a numerical simulaton approach using COMSOL 6.1. However, numerical simulatons can sufer high computatonal tme , which can become a limitaton.Machine learning, specifcally Long-Short Term Memory (LSTM) neural networks, is explored for predictng long-term producton temperatures based on the mentoned variables, ofering a faster alternatve to numerical simulatons. To do that, we used the COMSOL results of each DCHE as a training set for the LSTM system. So far, the tested LSTM system is a very shallow network with three hidden layers with 24, 14, and 7 neurons, respectvely. Given the low number of training instances, we were interested in testng diferent learning strategies to observe which was faster in convergence and determine the best in temperature forecastng afer ten years of producton. These strategies are the curriculum and the non-curriculum learning (standard). We applied the proposed concepts in a specifc case study in Cesano, integratng available geological and geothermal informaton. The simulaton assessed the producton temperature afer 10 years for a 3-branch DCHE confguraton in the Cesano area, known for its abundance of abandoned or depleted wells and a high geothermal gradient. The results from the numerical simulaton indicate that the producton temperature in such a DCHE confguraton would be 139.57 °C. Our analysis using a Long-Short Term Memory (LSTM) neural network trained with curriculum and non-curriculum learning strategy predicted a value of 136.8 °C and 95.7 °C with an absolute error of 2.77 °C and 41.1°C, respectvely. This suggests that the LSTM can provide accurate predictons for long-term producton temperatures, ofering an efcient alternatve to more tme-consuming numerical simulatons. Figure 1 3D view of the Deep Closed-Loop Heat Exchanger (DCHE) model in the Cesano area made up of three reused geothermal wells (2 injectons and 1 producton boreholes). The surfaces delineatng the lithothermal units as well as the numerical mesh are displayed.