In current applications of heating tower heat pumps (HTHPs), the systems tend to run with constant speed or fixed set points, which can be inefficient under varying weather data and building loads. To address this issue, this study proposes a model-based optimal operation of the HTHPs to achieve energy savings in both cooling and heating modes. Firstly, a physics-based model for an existing HTHP system was developed. Then, artificial neural network (ANN) models were developed and trained with vast amount of operational data generated by the physics-based model. The ANN models were found to be highly accurate (average relative error less than 1%) and computationally efficient (about 300 times faster than the physics-based model). After that, three optimal approaches were proposed to minimize the total energy consumption of the HTHP system. Approach 1 optimizes the load distribution between different heat pump units. Approach 2 optimizes the speed of fans and pumps by fixed approach and range of the condenser water (or evaporator solution). Approach 3 optimizes both the load distribution and the speed of fans and pumps. The optimization is implemented by using the ANN models, proposed approaches, and a genetic algorithm via a case study. The results show that the energy savings in the cooling season are 2.7%, 11.4%, and 14.8% by the three approaches, respectively. In the heating season, the energy savings of the three approaches are 1.6%, -1.4%, and 4.7%, respectively. Moreover, the thermodynamic performance in typical days was analyzed to investigate how energy savings could be achieved.