Nowcasting By Blending Techniques Using Cosine Similarity and Lead Time Correction

Radar-based extrapolation (EXP) and the storm-scale numerical weather prediction (NWP) models are two major deterministic methods for very short-range forecasting. Both models have individual advantages and limitations. Radar-based extrapolation is more likely to be more accurate at shorter time scales. The predictive accuracy of radar-based forecasting decreases rapidly within the several hours of severe weather development because the growth and decay of storms are not considered. Low-quality results are obtained during the first few hours of NWP-based predictions because of difficulties in depicting small-scale convective features in the model during the initial conditions, which is followed by the evolution of these features leading to correct predictions. The storm-scale NWP model using the rapid update cycle has improved performance over initial lead times. In this study, these different deterministic forecast (EXP and NWP) fields with high spatial–temporal resolution (1 km, 1 hour) were merged and evaluated. To merge the short-range forecasting model and the storm-scale NWP model, cosine similarity and lead time correction were applied to increase the predictive performance of precipitation within the short range of 1–6 h. The new blending method merges extrapolation-based forecasts with numerical weather prediction-based forecasts, heavily weighting the extrapolation forecasts at lead times of 0–3 h and transitioning emphasis to the NWP-based forecasts at later lead times through lead time correction using the hyperbolic tangent function. As shown by the root mean square error, correlation coefficient, bias, and critical success index, the blending method was generally better than EXP and NWP for lead times of up to 6 h. As a result of structure, amplitude, and location (SAL) verification, a dense assemblage of points in the top right-hand corner of the A–S diagram indicated that the three schemes produced precipitation amounts that were too large and significantly overestimated precipitation objects. A method of combining two or more prediction results with precipitation prediction using a single product may yield better results. However, the blending method depends on the predicted results of the products used in the merging. Research on developing position correction and the utilization of a larger data set are required to improve the performance of the blending technique further.