The first aim of this master thesis is to explore whether the normal force on the castor wheels can be estimated from IMU data using a machine learning approach. The second aim is to evaluate whether incorporating the changing load distribution due to trunk movement could improve the friction power estimation compared with neglecting changes in load distribution. Twenty-five subjects performed forward handrim wheelchair propulsions with no trunk, moderate and fast trunk movement on a treadmill in a wheelchair with six conditions regarding wheelchair mass and tire pressure. Two IMUs were placed on the trunk and wheelchair and two load pins in each castor wheel measured the normal force. After feature, model and hyperparameter selection, a model was trained to estimate the normal force on the castor wheels in percentage of the total weight from the IMU data. Accordingly, the predicted instantaneous normal force is used to calculate the friction power including changing mass distribution. When using the linear velocity and acceleration of the wheelchair and the, linear acceleration of forward movement of the trunk, adequate estimations (MAE of 3.69% total weight) of the normal force from an LSTM model can be obtained for unseen subjects. This model is robust for wheelchair settings regarding wheelchair mass and tire pressure and for propulsions with no, moderate and fast trunk movement. The instantaneous friction power prediction incorporating the changing load distribution is proven to more accurate during propulsions with moderate and especially fast trunk movement. Coaches, sport scientists, and athletes may find this model useful for analysing the effect of different propulsion techniques or wheelchair conditions on the friction power. As part of a larger context, this research will contribute to the process of filling the technological gap of in-field monitoring mechanical power. Future research must validate the robustness of the model during game situations