Groundwater heat pumps are gaining importance for resource-saving heating and cooling due to their low energy consumption. These heat pump systems consist of an extraction well and an injection well. Water is drawn from the extraction well and pumped through a heat exchanger. Afterwards, the water, now at a different temperature, is discharged into the injection well. Due to pressure differences in the groundwater-bearing layers, the introduced heat spreads through the subsurface. This creates heat plumes that align with the pressure gradient in the ground and can extend to the system’s own extraction well or to boreholes of other heat pumps. The temperature change at the borehole influences the efficiency of the affected heat pump. In areas with multiple heat pumps, it is crucial to predict the spread of heat plumes in advance to improve the overall system’s efficiency. Existing machine learning models are limited to a fixed spatial area or struggle to encode spatial features within composite inputs. This work introduces an approach based on sequential modeling networks, particularly a convolutional long short-term memory (ConvLSTM) architecture, to enhance heat plume predictions beyond these limitations. The encoder-decoder framework leverages the sequencing of input data, enabling more accurate predictions under varying subsurface conditions. Two model configurations with different prediction horizons were evaluated. While both models performed similarly in terms of Root Mean Square Error (RMSE) and Mean Absolute Error (MAE) on the test dataset, the model with the shorter prediction horizon excelled in predicting more pronounced heat plumes. Compared to previous UNet-based approaches, the ConvLSTM model shows improvements, especially in predicting the ends of the heat plumes and capturing spatial relationships.