A reliable load history is crucial for a fatigue assessment of wind turbines. However, installing strain sensors on every wind turbine is not economically feasible. In this paper, a technique is proposed to reconstruct the thrust load history of a wind turbine based on high-frequency Supervisory Control and Data Acquisition (SCADA) data. Strain measurements recorded during a short period of time are used to train a neural network. The selection of appropriate input parameters is performed based on Pearson correlation and mutual information. Once the training is done, the model can be used to predict the thrust load based on SCADA data only. The technique is validated on two different datasets, one consisting of simulation data (using the software FAST v8, created by Jonkman and Jonkman, 2016) obtained in a controllable environment and one consisting of measurements taken at an offshore wind turbine. In general, the relative error between simulated or measured and predicted thrust load barely exceeds 15 % during normal operation.