Predicting realizations of daily weather data for climate

Abstract

Weather is one of the primary driving variables that prominently impacts agricultural production and<br>associated disciplines, such as resource management. Lack of daily weather data for many locations along with many<br>prognosis requirements for weather for various applications has resulted in continuous efforts to determine the best possible<br>approach for weather sequence prediction. The goal of this study was to verify the k-nearest neighbours (k-NN) approach<br>for the prediction of daily weather sequences. This method can be employed on the assumption that the weather during<br>the target year is analogous to the weather recorded in the past. We used the nearest-neighbour re-sampling method for the<br>simultaneous prediction of daily radiation, maximum and minimum temperature, and precipitation for multiple locations.<br>A vector of weather variables, including precipitation, radiation, maximum and minimum temperature, on day (t + 1) is<br>re-sampled from historical data by conditioning on the vector of the same variables for the preceding day (t ). Observed<br>historical weather data for ten different sites located in Georgia were used for evaluation. The selected sites represent<br>different climatic conditions and the number of daily records varied from 46 to 97 years. The predicted daily and monthly<br>data were compared with both the observed daily and monthly average historical weather data and the target year of<br>2005 for all ten study sites. The statistical analysis included summary statistics, mean square difference (MSD) and its<br>components, and the Kolmogorov-Smirnov (KS) test. The results showed that the k-NN approach was able to reproduce<br>a similar pattern of the target year 2005 from the observed historical weather data. For all weather variables, both the<br>lower and upper quartiles (Q1 and Q3) showed a very good agreement with the data of the observed target year. The<br>cumulative distribution functions (CDFs) for the observed and predicted data were not significantly (P &gt;0.05) different<br>across all sites for precipitation, except for the minimum temperature of seven study sites, radiation for five study sites,<br>and maximum temperature for one study site. Our investigation to determine the minimum number of historical observed<br>weather data required for obtaining reliable prediction revealed that 25 years of data were sufficient to find similar patterns<br>compared to when all available weather data were used across all sites. It can be concluded from this study that the k-NN<br>approach on the basis of pattern recognition can be considered as a reliable method to predict daily weather sequences<br>based on historical weather data.