Neural Network–Based Short Term Forecasting of Water Levels in a Low-Temperature Geothermal Field

Abstract

Low-temperature geothermal fields are a critical component of district heating systems in Iceland, where short term fluctuations in heat demand and reservoir conditions can pose operational challenges. In such systems, the ability to reliably predict short term water level behavior in production wells is essential for ensuring secure heat delivery, particularly during periods of extreme weather. This study explores neural network–based time series models for short term prediction of water levels in a low-temperature geothermal reservoir. Using operational data from the Laugaland geothermal field in Iceland, multilayer perceptron (MLP) models are developed to predict water level dynamics based on production and injection rates. Model performance is evaluated under direct and recursive forecasting scenarios, with emphasis on robustness, uncertainty, and operational relevance. The results show that MLP models incorporating lagged inputs and target feedback accurately reproduce short term water level behavior and provide stable predictions across multiple initializations. Recursive predictions reveal increasing uncertainty with longer predictions horizons, and a scenario-based application illustrates how the models can be used to assess the risk of water levels approaching critical pump depth under sustained high-demand conditions. While not intended as a replacement for physics-based reservoir models, the findings suggest that neural network time series models can provide practical and computationally efficient decision support for short term management of low-temperature geothermal district heating systems.