Neural network based on dynamic tunneling technique for weather forecast

In this paper, a method of short-term temperature forecasting based on artificial neural networks is presented. An improved learning algorithm of neural network, RPROP, combined with a new efficient computational technique, dynamic tunneling technique is used to train neural network, for short, GDT. These two techniques are repeated alternatively processed to avoid local minima and result into a global optimization. The proposed networks are trained with actual data of the past 24 months (1999-2000) and are tested with data of 6 months (2001.1∼2001.3,2001.7∼2001.9), which come from several meteorological stations around or in Chongqing, China. Since the average prediction error of network on the test set equals 1.4, the obtained results demonstrated the efficiency of proposed method and show that the scheme reaches global minimum soon and converges at high rate.