Abstract: In order to further improve the accuracy of short-term load forecasting, normalized the weather and holidays data, the paper respectively uses three methods of Radial Basis Function, General Regression Neural Network and Probabilistic Neural Network to do modeling and forecasting, and do neural network ensemble forecasting based on genetic algorithm. By using the right combination of genetic algorithm to dynamically optimize the value of that time by optimizing the weights, the results show that, the resulting prediction accuracy by optimizing the combination of more than a single method to predict has been significantly improved. Through the two-week data to predict load, show that the method has prediction accuracy, stable performance, high precision and good practicality.