Abstract: Abstract: The accurate prediction of wind turbine load at crucial positions is of great importance for wind turbines' structural safety and scheme of operation and maintenance. Traditional load measurements using strain gauges are not suitable for long-term load monitoring because such measurements are labor-intensive, costly and time-consuming. Load simulation using software represented by GH bladed is not applicable to the evaluation of wind turbines' load because of the failure to get comprehensive information of wind in real time. Therefore, it is necessary to develop an economical and feasible assessment method for wind turbine load. The prediction of wind turbine load may be achieved by establishing the relationship between SCADA data and load measurements. In view of the complicated load condition and coupling relationship among loads, BP (back propagation) neural network is used to construct the relationship between SCADA data and load indicators. The load indicator used in this study is fatigue equivalent load. The SCADA data used as input parameters are selected by calculating the correlation coefficients between SCADA data and load indicators. The numbers of hidden layers and neurons are determined by trial-and-error approach. Taking blade edgewise blending moment as the example, when the rest of parameters of neural network remain fixed, the predicted outputs by the BP neural network with 1 hidden layer and 6 neurons are the most accurate. For other BP neural network models, 1 hidden layer is used and the number of neurons is determined by trial-and-error approach. In order to prove the validity of the model, load measurements at 7 crucial locations of a 2.5 MW wind turbine are carried out. The accuracy of the relationship model is judged by comparing the predicted outputs by the model with the measured values. The coefficient of determination and the arithmetic mean value of relative errors between model outputs and measured values are used to characterize the accuracy. The arithmetic mean values of relative errors are between 1.28% and 15.6%, and the coefficients of determination are between 0.882 and 0.951, which show that the predictions are in good agreement with measurements. Therefore, establishing the approximate relationship between SCADA data and load indicators by BP neural networks can be used as an effective means to achieve the long-term monitoring and evaluation of wind turbine load at crucial locations. In the consideration of the fact that the wind farm seldom has the wind measurement mast and anemometer mounted in nacelle is often affected by the rotating blades, this study has abandoned the wind speed in the load prediction. Using BP neural network, hub rotating speed, pitch angle and active power are used to predict wind speed. The predicted values agree well with the measured value, which indicates that wind speed is not indispensable for the accurate load prediction. The study also shows that better prediction accuracy can be achieved by increasing reasonable input parameters. When there are many SCADA data that can be used as model input, the correlation coefficient approach is more efficient than the trial-and-error approach in selecting model inputs.