Forecasting hydraulic head changes in injection wells using LSTM network

Abstract

Monitoring of well's specific capacity is commonly used to plan maintenance of injection wells in open-loop GSHP and standing column well systems. However, this method does not consider the effect of temperature on hydraulic conductivity. A first step towards an alternative approach that does include the effect of temperature is proposed in this work. We present a long short-term memory network capable of predicting the water level in the injection well of an operating GSHP system. The methodology consists of building a training set using a numerical model. A total of 500 simulations were conducted to evaluate hydraulic head signals under various inlet temperatures and flow rates along with hydraulic and thermal parameters drawn from a uniform distribution. Predictive performance of the artificial neural network is tested on an operational data set. The resulting RMSE between the forecasted and operational data set is 14.8 cm.