In the future smart cities, billions of communicating Internet of Things (IoT) devices are expected which communicate wirelessly in the limited spectrum offered by 5G and long-range technologies. This means that a huge amount of interferences must be overcome by new agile technologies without wasting power resources in the IoT nodes. In this paper, these challenges are addressed by a neural-network-based machine learning system that is based on frequency-domain features extracted from the communication channel. This machine learning system predicts the needed transmit power to overcome the interferences by a predefined margin. Extensive system simulations have been performed on a real-world dataset that shows power savings in the range of 35-83% and a packet receive-ratio of at least 95%. Similarly, it has been found that the system converts after approximately 50 supervised samples, which supports efficient tracking of parameter variations in the communication channel.