Partial least squares regression (PLSR) is used to load forecasting with considering several related factors, and it has a strong ability to explain the forecasting model. Because it selects the principal components of the independent variable set which are related to the load, and it overcomes the negative influence of the multiple relativity between the independent variables on the load modeling. But PLSR also has some weakness, such as there is useless orthogonal information between the independent variables and dependent variables, which may decrease the model’s forecasting accuracy. Based on the characteristics of PLSR and BP neural network, a PLSR-BP neural network was established, which could modified the weight and regression coefficient in original PLSR model. The practical example result shows that this method is correct and effective.