Based on the load data of meritorious power of some area power system,  three BP ANN models,  namely SDBP, LMBP and BRBP Model,  are established to carry out the short-term load forecasting work, and the results are compared.  Since the traditional BP algorithm has some unavoidable disadvantages,  such as the low training speed and the possibility of being plunged into minimums local minimizing the optimized function,  an optimized L-M algorithm, which can accelerate the training of neural network and improve the stability of the convergence,  should be applied to forecast to reduce the mean relative error.  Bayesian regularization can overcome the over fitting and improve the generalization of ANN.