[Untitled]

Artificial time-delay feed-forward neural networks (NN) with one hidden layer and error back-propagation learning are used to predict surface air temperatures (SAT) for six hours up to one day. The networks were trained and tested with the use of data covering a total of 26280 hours (three years) monitored in the period 1998-2000 in location Spořilov (suburban area of Prague). The NN models provided a good fit with the measured data. Phase diagrams as well as the results of the variability studies of SAT revealed a fundamental difference between summer temperatures (April-September) and winter temperatures (October-March). Results of the trial runs indicated that NN models perform better when both periods are trained separately. The results show that relatively simple neural networks, with an adequate choice of the input data, can achieve reasonably good accuracy in one-lag as well as in multi-lag predictions. For the “summer” period the total errors give 0.055 and/or 0.044 mean accuracy of predicted values in training and testing sets, respectively. Similarly high mean accuracy of the simulated values of 0.057 and 0.065 was obtained for the training and testing sets in the winter season. Similarly good results with the mean error of 0.028 were obtained for the summer period of the year 2001, which were used for additional testing (see Appendix). Higher accuracy obtained for the year 2001 is due to the fact, that warm temperature extremes, which are generally predicted with less accuracy, did not occurred in the summer 2001.