Abstract In the well tests in the geothermal reservoir, the non-isothermal behavior caused by the injection of cold water can affect the pressure response of the reservoir, which has been rarely considered in the existing analytical and numerical well testing models. In this work, we propose a well testing surrogate model to predict the pressure response of the geothermal reservoir during a non-isothermal injection fall-off test. First, a numerical well testing model of a geothermal reservoir with a horizontal well is built in MATLAB Reservoir Simulation Toolbox (MRST). Then, trained on the solutions of the numerical well testing model, two surrogate models based on long short term memory (LSTM) is developed. A multi-stage training scheme is employed to preserve the consistency by imposing the physics correlation between the pressure difference and pressure derivative. The surrogate models can accurately predict the pressure difference and derivative with mean absolute percentage errors (MAPE) of 5.90% and 5.51%, respectively. Moreover, the surrogate models can achieve a 2.90 × 105 speedup over the numerical well testing model, which contributes to efficient pressure response predictions and reservoir parameter evaluations.