In wireless communication for Internet-of-Things (IoT), a base station (BS) is required to efficiently accommodate radio signals transmitted from massive IoT devices. We have developed and demonstrated an uplink non-orthogonal multiple access (UL-NOMA) system, in which signals from the devices are multiplexed in the power domain without any spatial multiplexing or spreading. In the UL-NOMA strategy, precise transmit power control (TPC) methods are required to compensate for various fading fluctuation. For example, NOMA that multiplex devices' signals in the power domain utilize the difference of received signal power at BS to successfully achieve successive interference cancellation (SIC). To maintain the difference of received power at BS, the implementation of precise TPC is crucial. In the paper, a prediction method on the future fading condition from current conditions is useful to realize precise TPC in UL-NOMA in the power domain. The prediction method at IoT device side with the autoregressive (AR) model is proposed which utilizes received OFDM signals in the downlink. The proposed TPC is evaluated by computer simulations. Simulation results show that TPC with the proposed fading prediction improves the signal separation success rate of UL-NOMA with SIC. The ratio that the signals of all devices in multiplexing 5 devices are separated is improved from 39% to 66%.