The neural network approach has been extensively studied in the past years. Especially, the artificial neural network based on error back propagation algorithm (BP ANN) has played an important role in weather forecasting. But there is a problem for the BP ANN training: With the same sample model, the same network structure and the same input parameters, the weight and the final error results from each training are always not the same. Some final error results are good, but some poorer. So in the use of the BP ANN modeling training, we hope to find out a local optimal solution of the predictor model, helping the models estabish better generalization ability. In order to improve the BP ANN modeling and forecasting capability in forecasting operations, two year (2009-2011, each year from May 15 to September 15) forecast data from T639 and the maximum observed temperature from four stations in Beijing are used for numerical simulation. It is found by fitting modeling trials on four model samples that by using three improved methods, BP ANN’s random weights have a Gaussian distribution or initial weights field are initialized repeatedly, and with hidden layer neurons using dynamic network structure, hit rate from BP ANN fitting of the sample has improved to some extent. The last but the most important, taking the data of 115 days (from 15 May to 15 September 2011) as a forecast test data, three improved methods are integrated in BP ANN. Compared with the BP ANN before improvement, the result is found to be: hit rate from validation sample using the improved BP ANN is higher than the non improved BP ANN, and the optimized BP ANN has better generalization ability.