Forecasting Global Temperature Variations by Neural Networks.
Takaya Miyano, Federico Girosi · 1994
Global temperature variations between 1861 and 1984 are forecast using regularization network, multilayer perceptrons, linear autoregression, and a local model known as the simplex projection method. The simplex projection method is applied to characterize complexities in the time series in terms of the dependence of prediction accuracy on embedding dimension and on prediction-time interval. Nonlinear forecasts from the library patterns between 1861 and 1909 reveal that prediction accuracies are optimal at the embedding dimension of 4 and deteriorate with prediction-time interval. Regularization network, backpropagation, and linear autoregression are applied to make short term predictions of the meteorological time series from 1910 to 1984. The regularization network, optimized by stochastic gradient descent associated with colored noise, gives the best forecasts. For all the models, prediction errors noticeably increase after 1965. These results are consistent with the hypothesis that...