Stochastic Seasonal Weather Forecasting

[摘要] References(25)The aim of this paper is to present a research of seasonal forecasting on a scientific basis.There is no difference between the stochastic formulation for short-range forecasting and that for long-range. For the latter, however, it is far more difficult to follow the physical relation than for the former, and it is inevitable to rely mainly on empirical or statistical relationships in the procedure of framing a forecasting scheme. Our method of stochastic numerical prediction may be summarized as follows: (1) The predictors should be empirically researchful by taking meteorological experiences into consideration. Significance of meteorological hypotheses must be statistically tested. Out of the factors considered and tested, we select some effective ones in the course of framing a stochastic prediction scheme. It is important to use the data of the recent stationary period. (2) The predictors should be synthesized into a stochastic prediction formula in which periodicities, multiple correlation and similarity methods are combined. In the case of the Gaussian process, into which we can transform the original series if it is not accepted as the Gaussian, the linear prediction has an optimum property. (3) Prediction should be probability-theoretically represented; in the case of the Gaussian process it may be expressed by the conditional mean value and conditional variance of the predictand. (4) When the conditional variance is smaller, the prediction scheme is better. Thus we can construct a stochastic prediction formula which may be best relative to an ‘extent of the consideration’ in the research for predictors. But the applicability of the prediction is of course dependent on the stationarity of the process ranging from the past to the future in question.According to the principle stated above, we practiced a seasonal prediction, as a trial, under the following conditions:Predictands: Monthly mean temperature for July and August 1953 in the region represented by Miyako, and monthly total amount of precipitation for June and July in the northern parts of the Kanto district including Maebashi, Kumagaya and Utsunomiya; the division of forecasting regions was performed by a stochastic criterion.Data and deadline time: Data before March 31, 1953 were used.The extent of consideration for each predictand is described in detail as much as possible in the present paper. Predictors adopted by the authors are as follows:July temperature: long period variation, periodic fluctuation with 13.0-year period, Pontadelgade-Malye Karmakouly pressure difference for January of the year and Malye Karmakouly-Stykkisholm pressure difference for January of the year.August temperature: long-period variation, periodic fluctuation with 7.4, 13.0 and 11.4-year periods, autoregressive periodicities, change in Chichizima-Shikka pressure difference for February of the year and of the last year, and Chichizima pressure for February of the year.June precipitation: total amount of precipitation for January and February of the year, total amount of precipitation for June 6 years before and that for the same month 10 years before.July precipitation: periodic fluctuation with 7/3-year period, Barrow-Dutchharbour pressure difference for January of the year, Naha-Chichizima pressure difference for January of the year and that for March of the year.If we adopt too many predictors the precision of forecasting may be reduced owing to the reduction of degrees of freedom.The large sample prediction formula is adopted here throughout; an example of exact sampling method of prediction is published in the Journal of Meteorological Research, Vol. 6 (1954), pp. 172-183.

[学科分类] 大气科学