Estimation of precipitation induced by tropical cyclones based on machine‐learning‐enhanced analogue identification of numerical prediction

Abstract

A tropical cyclone (TC) is an extremely hazardous weather event. These events include heavy rain as an important hazard factor, which poses a serious threat to the public safety of coastal cities. Presently, numerical weather prediction (NWP) is an important and commonly used method to support the forecasting of the impact of TCs. However, relatively high uncertainty still remains in quantitative precipitation predictions provided by NWP, which makes it difficult to meet the demands of public safety management in regions affected by TCs. To remediate this deficiency, the study combines machine learning (ML) techniques with NWP and proposes a new analogue identification method for TC precipitation estimation. The method consists of three parts. First, the output data of historical NWPs, including wind, temperature, humidity of 850 hPa and sea-level pressure field, are deposited into a massive sample library. Second, the dimensionality of the sample library is reduced by using the locally linear embedding (LLE) method, and a characteristic subspace containing the features of the original sample library is formed. Finally, when a new NWP prediction appears, it can be projected into the characteristic subspace, and the historical NWP sample having the greatest similarity with the current NWP prediction can be identified. The observational precipitation corresponding to the most similar historical NWP sample can then be used to estimate the potential impact of current TCs. This method is verified using observation results from the coastal city of Shenzhen, China. The results show that the prediction of the method proposed in the current study exhibited significant improvement compared with prediction results provided by both the NWP's direct output and a traditional method used in the city.

Understanding the Influence of Urban Form on the Spatial Pattern of Precipitation

“…Compared with the 500 hPa data which determine large‐scale conditions, the 850 hPa circulation conditions can more clearly display the specific weather processes (Y.‐Y. Liu et al., 2021) and thus can provide information about the movement of weather systems. In addition, the 850 hPa represents the top (or close to the top) of the planetary boundary layer for the locations close to sea level, which can reflect the surface impacts while having a more stable value.…”

Research on the usability of different machine learning methods in visibility forecasting

“…At present, the most widely applied machine learning methods include algorithms like the support vector machine, the k-nearest neighbor, and the random forest. The process of machine learning can be summarized as follows: the collected training data are input, all possible hypotheses (expressed as functions) are tested using the learned algorithm, and the hypothesis which is closer to the actual pattern is identified (Liu et al, 2019;Liu et al, 2020;Liu et al, 2021).This same process is shown in Figure 1.…”

Densely Connected Neural Networks for Nonlinear Regression

“…In recent years, machine learning has been successfully applied in atmospheric and environmental science, see [ 41 , 42 , 43 , 44 , 45 , 46 , 47 , 48 , 49 ] for more details. In this section, DenseNet regression is also employed on climate modeling.…”