Transfer precipitation learning via patterns of dependency matrix-based machine learning approaches

Accurate precipitation prediction is very significant for urban, environmental, and water resources management as well as mitigating the negative effects of drought and flood. However, precipitation prediction is a complex and challenging task which involves meteorological parameters that contain uncertainty. This study attempts to ease the complexity of the problem via proposing a correlation matrix approach. Covariance and correlation matrices are analytical tools that are widely used to identify the interrelationships and possible dependencies throughout the data. Correlation matrices have some advantages over covariance matrices. The main drawback of covariance matrices is their sensitivity to the measurement units of variables. The variables with relatively large variances will dominate the results of multivariate analysis when the covariance matrix is used. Accordingly, the covariance matrix fails to provide useful information when there exist large differences between variances of variables. On the other hand, besides their easy interpretable features, the results of different analyses obtained from correlation matrices can effectively be compared. Therefore, in this study, in order to improve the performances of the predictive models, interrelationships and possible dependencies among data obtained from eighteen precipitation observation stations located in the Upper Euphrates Basin of Turkey (1980–2010) is investigated using correlation matrix approach. Relatedly, dependencies between the stations are resolved by means of examining the correlation matrix and optimal model inputs (data of particular stations) are selected for each prediction scenario. The transfer precipitation learning was performed throughout the period from 1980 to 2010 for eighteen precipitation observation stations located in the Upper Euphrates. Three different data-driven models Fuzzy, K-nearest neighbors (KNN), and multilinear regression (MR) are developed based on the patterns of correlation matrix. Predictive powers of the models are compared by means of performance evaluation criteria, i.e., Nash–Sutcliffe efficiency, mean square error, mean absolute error, and coefficient of determination (R²). Results of this study show that all developed correlation matrix patterns-based Fuzzy, KNN, and MR models have high precipitation prediction performance. However, even though all model results are close to each other, Fuzzy model provided more accurate results with requiring data from a relatively low number of stations. Therefore, patterns of correlation matrix-based Fuzzy model is the most efficient and well-suited approach for precipitation prediction among all the developed models.