This paper proposes a technique to estimate the equivalent reflected impedance on the transmit side of an inductive power transfer (IPT) system, produced by an inductively coupled load. This technique consists of analysing changes in the drain voltage waveform of the switching devices to estimate the reflected impedance, and hence not require feedback information from the receive side or the coupling between the coils. The correlation between the drain voltage waveform and the reflected impedance is done by training a model with machine learning techniques. The proposed impedance estimation method is demonstrated using circuit simulations and is verified experimentally, with an IPT transmitter driven by a load independent Class EF inverter operating at 13.56 MHz. The IPT receiver consists of an ac-load which allows changes in the residual reactance. The model trained from experimental data is capable of estimating the equivalent reflected impedance with an accuracy (coefficient of determination) of 0.9899 for the real part and 0.9743 for the imaginary part.