A High Accuracy Short-Term Precipitation Forecasting Method Using Improved VIL Nowcast Based on Three-dimensional Observation Data of MP-PAWR

In recent years, demand for weather forecasting technology has increased due to an increase in sporadic, localized weather disasters such as heavy rains and tornadoes. To meet this demand, we have developed a high accuracy short-term precipitation forecasting method.

The developed forecasting method is based on VIL (Vertically Integrated Liquid water) Nowcast 1) by using three-dimensional radar observation data from the MP-PAWR (Multi-Parameter Phased Array Weather Radar). The MP-PAWR is able to get fast and highly accurate three-dimensional observation data, which allows us to estimate the VIL. VIL Nowcast utilizes the VIL for rainfall intensity forecast by using the mass-balance equation where the total derivative of the VIL is related by the difference of the precipitation (rainfall) term and the source term corresponding to net effects of vapor condensation, coalescence, and melting, etc.

This paper focuses on improving the accuracy of VIL Nowcast based short-term rainfall forecast. To this end, the following four improvements have been conducted.

The first one is the improvement of the advection vector estimation in VIL Nowcast. In terms of stable estimation of the advection vectors, we propose to use an FFT-based cross-correlation method with multivalued rainfall intensity to estimate the advection vectors, instead of using the conventional direct cross-correlation method with binarized rainfall intensity. The direct cross-correlation method is commonly used for nowcasting because of its simple principle and often used with binarized rainfall intensity due in part to high computational burden. On the other hand, FFT-based cross-correlation method utilizes the Fast Fourier Transform (FFT) for speeding up calculation time. In addition, we have confirmed that more stable estimation of the advection vector can be achieved by the FFT-based cross-correlation method possibly due to the use of multivalued rainfall intensity, which contributes to improving the forecast accuracy.

The second one is the mesh size optimization of the radar observation data to estimate the source term used in VIL Nowcast. We propose to use downsized averaged data obtained from coarser mesh than the original 250meters mesh to reduce the influence of the error appeared in the difference calculation between the aligned past VIL distribution and the current VIL distribution that is used to estimate the source term.

The third one is the application of a post-filter to the forecasting results. We have applied not only a simple mean value filter to the forecast results, but also a special filter that prevents value reduction by restoring the original value for meshes whose values have been reduced by filtering.

The final one is to incorporate an additional correction term into the source term by resorting a machine learning technique. To improve the forecast accuracy in rapidly growing rainfall events such as localized heavy rainfall, we have used machine learning to extract rainfall features and incorporated the correction term that increases the source term based on these features.

The effectiveness of the proposed improvements is illustrated by evaluating the forecast accuracy. If the baseline is defined as the condition where only the direct cross-correlation method with binary rainfall intensity is applied, the change from the direct cross-correlation method to the FFT cross-correlation method improved forecast accuracy by 4.7 percentage points, adding an optimization of the mesh size when calculating the source term to the former improved forecast accuracy by an additional 2.1 percentage points, applying post-filters to the former improved by an additional 0.8 percentage points and adding a new term obtained by machine learning to the former improved by 0.1 point.

In addition to the increase in forecast accuracy, the correlation between observed and forecast rainfall amounts over 10 minutes periods was also greatly improved, the baseline RMSE (Root Mean Square Error) was 1.21 mm, whereas the four measures improved the RMSE to 0.65 mm and we obtained a 46% improvement in RMSE, and the correlation coefficient also improved from 0.75 to 0.86.

This work was supported by Council for Science, Technology and Innovation (CSTI), Cross-ministerial Strategic Innovation Promotion Program (SIP), “Enhancement of societal resiliency against natural disasters” (Funding agency: JST).

1) Kohin Hirano，Masayuki Maki：Imminent Nowcasting for Severe Rainfall Using Vertically Integrated Liquid Water Content Derived from X-Band Polarimetric Radar，pp.201-220，Journal of the Meteorological Society of Japan，Vol.96，2018．