Experiments of DSAEF_LTP Model with Two Improved Parameters for Accumulated Precipitation of Landfalling Tropical Cyclones over Southeast China

Tropical cyclones (TCs) develop over the tropical or subtropical waters and have an organized circulation. Located in eastern Asia on the west coast of the Pacific Ocean, China is frequently impacted by the rainfall induced by TCs (Chen and Ding ^[1]; Chen and Meng ^[2]; Li et al. ^[3]; Chen et al. ^[4]). TC rainstorms can cause severe disasters (Chen and Xu ^[5]; Ren et al. ^[6]; Liang et al. ^[7]; Zhao et al. ^[8]). For example, in 2006, severe tropical storm Bilis caused an accumulated precipitation of more than 400 mm over southern China, resulting in 612 people dead or missing. The direct economic losses exceeded 266 billion yuan during this TC-related rainstorm event (Ma et al. ^[9]; Ye and Li ^[10]; Wang et al. ^[11]; Ren et al. ^[12]; Liu and Cui ^[13]; Gao et al. ^[14]). In 2016, super typhoon Meranti (1614) made landfall in Xiamen, causing 29 people dead or missing and direct economic losses of about 16.9 billion yuan (Zhao et al. ^[15]; Chen et al. ^[16]; Liu et al. ^[17]).

The forecast performance for TC-related precipitation by numerical weather prediction (NWP) models still needs further improvement (Zhu et al. ^[18]; Xu et al. ^[19]; Reddy et al. ^[20]; Hsiao et al. ^[21]; Hong et al. ^[22]). In addition to research for the improvement of NWP models, the combined dynamical-statistical approach is an important way of improving TC-related precipitation forecasts (Chou ^[23]). Research on dynamical-statistical forecasting methods for landfalling TC precipitation prior to 2018 can be divided into three main categories: (1) the TC forecasting by numerical models and the observed historical TC precipitation being applied to gain the climatological average precipitation of target TCs (Marks et al. ^[24]; Lee et al. ^[25]; Lonfat et al. ^[26]), (2) the TC forecasting by numerical models and the distribution of the rainfall rate at the initial time being used to get the accumulative precipitation during the forecasting period (Kidder et al. ^[27]; Ebert et al. ^[28]), and (3) the dynamic similarity method based on the forecasting element field by the numerical models being adopted to do TC precipitation forecast (Li and Zhao ^[29]; Zhong et al. ^[30]).

Ren et al. ^[31] developed the Dynamical-Statistical-Analog Ensemble Forecast model for landfalling TC precipitation (the DSAEF_LTP model). The underlying concept of the DSAEF model is that the operational NWP model is used for the forecast track. The precipitation forecast is obtained by finding analog cyclones, and then making a precipitation forecast from an ensemble of the analogs. The generalized initial value (GIV) determines the choice of analogs and consists of physical factors that affect TC precipitation. The original version included two physical factors (TC track and landfall season). Ding et al. ^[31], Jia et al. ^[33] and Jia et al. ^[34] then improved the model by introducing the TC intensity, new variations of the similarity region for track analog, and five new ensemble methods, respectively.

The above studies focus on landfalling TCs over South China or on case studies. It is also important to test the simulation and forecast performance of the DSAEF_LTP model for landfalling TCs over Southeast China, which is severely affected by TC rainstorm damage (Chen et al. ^[35]). In this study, large-sample prediction experiments on landfalling TCs over Southeast China are carried out using four different versions of the DSAEF_LTP model. Section 2 presents the data and methods. Section 3 describes the target TC samples and the design of the experiments. Section 4 provides the results of the experiments. A summary is described in Section 5.

Daily (20: 00-20: 00) precipitation data used are from 141 stations in Southeast China (Zhejiang Province, Fujian Province, Shanghai, and Taiwan) during 1960-2019 (Fig. 1). The data for the Chinese mainland are obtained from the National Meteorological Information Center (NMIC) of the China Meteorological Administration (CMA), while those for Taiwan Island are from the meteorological authority of Taiwan.

Figure 1.  Geographical distribution of the 141 meteorological stations in Southeast China.

The historical best-track dataset during 1960-2019, including position, maximum surface wind speed near TC center, and pressure of the TCs at 6 hourly intervals, is obtained from the Shanghai Typhoon Institute (Ying et al. ^[36]). The operational NWP model forecast tracks were obtained from the NMIC for the period 2004-2019.

To compare the precipitation forecast results of the DSAEF_LTP model with that from NWP models, we also include corresponding rainfall forecast historical data from the European Centre for Medium-Range Weather Forecasts (ECMWF) model, the National Centers for Environmental Prediction Global Forecast System (GFS) model, the Global/ Regional Assimilation Prediction System (GRAPES) model of the China Meteorological Administration and the Shanghai Meteorological Service WRF ADAS Real-Time Modeling System (SMS-WARMS), with horizontal resolutions being 0.125° × 0.125°, 0.25° × 0.25°, 0.25° × 0.25°, and 0.09° × 0.09°, respectively. To process the data uniformly, we have converted the horizontal resolutions of the four models to 0.1° × 0.1° by bilinear interpolation.

Application of the DSAEF_LTP model (Ren et al. ^[37]; Ding et al. ^[32]; Jia et al. ^[33-34]) consists of four steps: (1) obtaining the TC track forecast from the NWP model, (2) constructing the GIV, (3) identifying analogs, and (4) finding the ensemble LTP of the analogs.

Table 1 lists the eight parameters (P1-P8) of the original DSAEF_LTP model. The third column gives the values of the parameters in the experiments. The first step is to combine the observed track of the target TC before the initial time and the forecast track after the initial time with its complete track. The observed track prior to the initial time (P1) and the operational NWP model forecast track after the initial time are combined as the target TC track.

The second step is to construct the GIV. GIV includes TC internal factors and environmental factors of TC. So far, there are three TC internal factors in the GIV: TC track, landfall season, and tintensity of the TC.

The third step is to determine the similarity of GIVs. P2, P3, and P4 (i.e., similarity region, threshold of the segmentation ratio of a latitude extreme point, and overlapping percentage threshold of two TC tracks) are used to determine historical TCs whose tracks to that of the target TC. TC track Similarity Area Index (TSAI) (Ren et al. ^[37]) is calculated in this process. It represents the degree of TC track similarity. The more similar the two TC tracks are, the closer the index is to 0. After determining the track similarity, the seasonal similarity and the intensity similarity are determined by P5 and P6, respectively. After that, N historical TCs (analogs) with GIV most similar to the target TCs are selected where N, the number of analogs, is specified by P7. In the fourth step, the TC precipitation of analogs are used to obtain the ensemble precipitation forecast of the target TC using an ensemble method as specified by P8.

As described, the analog-ensemble forecast technique is specified by a set of parameters, each having a range of values. The combinations of different parameter values are called schemes. In this paper, the schemes of each experiment are different, which are introduced in detail in Section 3.1.

Based on Table 1, Table 2 provides the five new values for the improved parameter P2 (Jia et al. ^[33]) and the five new values for the improved parameter P8 (Jia et al.^[34]). The serial numbers for the five new values of P2 are 16th-20th, respectively. Meanwhile, these for the five new values of P8 are 3rd-7th, with the names being the optimal percentile (90th percentile in this study), fuse, probability matching mean, equal difference-weighted mean, and TSAI-weighted mean, respectively.

(a) OSAT

The observed TC precipitation are obtained from daily precipitation data over Southeast China by the Objective Synoptic Analysis Technique (OSAT) for partitioning TC precipitation (Ren et al. ^[38-39]; Wang et al. ^[40]). This method can identify the precipitation generated by TCs from daily observed precipitation data based on the distance between the stations and the precipitation centers.

(b) Threat score (TS)

The index to evaluate the forecast performance in this paper is the TS, which is widely used in meteorological operational forecasts and is calculated as follows:

The TS ranges from 0 to 1 and the closer it is to 1, the higher the hit rate.

(c) BIAS score

BIAS is used to describe how close the forecasted precipitation area coincides with the observed precipitation area, calculated as follows:

Take value 1 as the critical point. Bias greater than 1 means it is likely to be false alarms, and less than 1 means it is likely to miss.

According to the different TC factors and parameters, four experiments have been designed, the names of which are DSAEF_LTP_1, DSAEF_LTP_2, DSAEF_LTP_3, and DSAEF_LTP_4. The GIV in DSAEF_LTP_1 includes three factors: TC track, landfall season, and intensity of the TC. Eight of the parameters given in Table 1 are used. If all the parameter values can be obtained ideally, the number of schemes in DSAEF_LTP_1 will be 1, 620, 000 (= 3 × 15 × 3 × 6 × 5 × 4 × 5 × 10 × 2). However, some parameters cannot be used in some TC cases. For example, if the TC does not form at 1200UTC on the day before the day of the TC precipitation occurring on land, the 3rd value of P1 cannot be used. Thus, the number of schemes in the experiments is equal to or less than 1, 620, 000.

DSAEF_LTP_2 adds the improved five similarity regions (16th-20th) based on DSAEF_LTP_1 and ideally has 2, 160, 000 (= 3 × 20 × 3 × 6 × 5 × 4 × 5 × 10 × 2) forecast schemes. DSAEF_LTP_3 adds the improved five ensemble methods based on the parameters of DSAEF_LTP_1, for which ideally there exist 5, 670, 000 (= 3 × 15 × 3 × 6 × 5 × 4 × 5 × 10 × 7) forecast schemes. DSAEF_LTP_4 adopts both improvements, the similarity regions (as in DSAEF_LTP_2) and the ensemble methods (as in DSAEF_LTP_3), and ideally has 7, 560, 000 (= 3 × 20 × 3 × 6 × 5 × 4 × 5 × 10 × 7) forecast schemes.

From 2004 to 2019, there are 65 TCs with maximum daily precipitation ≥ 100 mm occurring in at least one station in Southeast China, and have been identified as target TCs for these experiments. The 47 TCs from 2004 to 2016 are selected as training samples for simulation experiments, while the 18 TCs from 2017 to 2019 are independent samples for forecast experiments. The tracks for the two samples are shown in Fig. 2.

Figure 2.  Tracks of (a) the 47 TCs used as training samples from 2004 to 2016 and (b) the 18 TCs used as independent samples from 2017 to 2019.

For the DSAEF_LTP model, the experiment procedure consists of the following steps. Firstly, simulation experiments are carried out for obtaining the best forecast schemes. This step has two sub steps: 1) Obtaining common schemes. As mentioned in Section 3.1, there are 1, 620, 000 schemes in DSAEF_LTP_1. The schemes that can be used in forecasting the precipitation of all the 47 TCs by DSAEF_LTP_1 are selected. Those schemes are called common schemes. 2) Choosing the best scheme from common schemes. Because TS for accumulated precipitation above 100 mm and 250 mm are of concern, the common scheme with the maximum value of the sum of the TS of the two thresholds (TSsum = TS250 + TS100) is defined as the best scheme. TS250 and TS100 are the TS of accumulated precipitation above 250 mm and 100 mm, respectively. Each of the four experiments has a best scheme.

The best schemes for each of the four model configurations are chosen from the application of the DSAEF model to the 47 training sample TCs (the simulation experiments). They are then applied to the 18 independent sample TCs (the forecast experiments). Finally, the results of the forecast experiments of the DSAEF_LTP model are compared with the prediction results of dynamical models.

Table 3 provides the best schemes of the four experiments. It shows that the values of the two improved parameters, P2 and P8, in the best schemes are all new values. P2 is the 17th value in DSAEF_LTP_2. P8 is the 4th value in DSAEF_LTP_3. In DSAEF_LTP 4, P2 is the 20th value and P8 is the 4th value. In the following paragraphs, we use the results of the best schemes to represent the performances of DSAEF_LTP_ 1, DSAEF_LTP_2, DSAEF_LTP_3, and DSAEF_LTP_4.

Figure 3 shows the distribution of TS for the four experiments. For the simulation experiments (Fig. 3a), the TSsum of DSAEF_LTP_4 (0.6352) is the highest among the four experiments. TS250 and TS100 are 0.2647 and 0.3710 respectively, which implies that the DSAEF_LTP model can significantly improve the simulation ability of the model after improving both parameters, similarity region and ensemble scheme. Moreover, the TS of DSAEF_LTP_2 (0.5874) is slightly better than that of DSAEF_LTP_3 (0.5788). This means that compared with the introduction of the new ensemble method parameter values, the performance of the DSAEF_LTP model improves more by introducing new similarity regions.

Figure 3.  Distribution of threat score TS for the four experiments. (a) simulation and (b) forecast experiment. (DSAEF_LTP_1: pink, DSAEF_LTP_2: green, DSAEF_LTP_3: red, DSAEF_LTP_4: blue)

Figure 3b shows the TS of the best scheme in the four forecast experiments. The performance of the DSAEF_LTP model in forecast experiment is slightly different from the simulation experiment. On the one hand, compared with DSAEF_LTP_1, the performance of DSAEF_LTP_2 has not improved. The TSsum of DSAEF_LTP_2 is 0.2639 higher than that of DSAEF_LTP_1 in the simulation experiment, while the TSsum is 0.0020 lower than that of DSAEF_LTP_1 in the forecast experiment. On the other hand, the forecast performance of DSAEF_LTP_3, not DSAEF_LTP_4, is the best among the four experiments, which indicates that the DSAEF_LTP model has the most obvious improvements in forecast performance after introducing just the new ensemble methods.

The tracks of training samples and independent samples are given in Fig. 2 in Section 3.2. If we pay attention to what we define as the typical TC track, i.e., TCs moving in a southeast towards northwest direction or making landfall over Southeast China, the numbers in the training samples and the independent samples are 28 (or 60% of tracks) and 4 (22%) respectively. The change in track characteristics between the training and independent samples may be the cause of the different model having optimum predictions from simulation to forecast experiments. This will be explored further in Section 4.3.

In conclusion, DSAEF_LTP_4 performs best in the simulation experiment and DSAEF_LTP_3 performs best in the forecast experiment. Thus, DSAEF_LTP_3 and DSAEF_LTP_4 are selected as representative experiments in subsequent analyses.

Figure 4 shows the TS of DSAEF_LTP_3, DSAEF_LTP_4 and the four dynamical models. The results show that the forecast performance of DSAEF_LTP_3 is best, whose TSsum, TS250, and TS100 scores are 0.4881, 0.1678, and 0.3203, respectively. DSAEF_LTP_4 ranks second, whose TSsum, TS250, and TS100 scores are 0.3504, 0.0950, and 0.2552, respectively. Among the dynamical models, the ECMWF model has the best forecast performance, and its TSsum (i. e., TS250 + TS100) is 0.2930 (i. e., 0.0583 + 0.2347). This is followed by the SMS-WARM, GFS, and GRAPES models. That is, DSAEF_LTP has an advantage in forecasting accumulated rainfall of ≥ 250 mm and ≥ 100 mm.

Figure 4.  Comparison of the TS of the DSAEF_LTP model with four numerical models (ECMWF, GRAPES, GFS, SMS-WARMS) in independent forecast experiments.

Figure 5 compares the average BIAS of DSAEF_LTP_3 and DSAEF_LTP_4 with four dynamical models in different magnitudes. BS250 (BS100) indicates the BIAS of accumulated precipitation ≥ 250 mm (≥ 100 mm). The BS250 and BS100 of the two improved DSAEF_LTP model (DSAEF_LTP_3 and DSAEF_LTP_4) are more than 1, implying that the model tends to have false alarms in accumulated forecasting precipitation above 250 mm and 100 mm. SMS-WARM, ECWMF, GFS, and GRPAPES are four dynamic models with the BS250 and BS100 being all less than 1, which means that the area of misses is wider than that of false alarms.

Figure 5.  Comparison of the BIAS of the DSAEF_LTP model with four numerical models (ECMWF, GRAPES, GFS, SMS-WARMS) in independent forecast experiments.

TSsum of the training samples and independent samples with the best scheme of DSAEF_LTP_3 are ranked from high to low, respectively. Figs. 6a and b respectively compare the intensity and precipitation characteristics of the five TCs with high TSsum ranking and the five TCs with low TSsum ranking in simulation experiment and forecasting experiment. For both the simulation and the forecast experiments the average maximum daily precipitation and average maximum accumulated precipitation of the five TCs with high TSsum are greater than that of the five TCs with low TSsum ranking. In contrast, the intensities of high performing versus low performing examples are almost the same. As for the issue of tracks discussed earlier, Fig. 7 compares the tracks of the five TCs with high TSsum ranking and the five TCs with low TSsum ranking in both the simulation experiment (7a, b) and the forecasting experiment (7c, d). In the simulation experiments, the tracks of the five TCs with high TSsum ranking all move in the southeast-northwest direction, i.e., they are TCs with typical tracks. Meanwhile, in the forecast experiment, over half of the TCs with high TSsum ranking have TC tracks of landfalling over Southeast China or southeast-northwestward direction. By comparison, the tracks of the five TCs with low TSsum ranking are not the typical TC track. The characteristics of TCs with good forecasts in DSAEF_LTP_4 are similar to those in the DSAE_LTP_3 model (figure omitted). Thus, the TCs with high TSsum ranking of DSAEF_LTP model are characterized by heavy precipitation and typical TC tracks over Southeast China.

Figure 6.  Comparison of intensity and precipitation characteristics of the five TCs with high TSsum ranking and the five TCs with low TSsum ranking of DSAEF_LTP_3 under the best scheme. ((a). simulation experiment, and (b) forecast experiment)

Figure 7.  Compare the tracks of the five TCs with high TSsum ranking and the five TCs with low TSsum ranking of DSAEF_LTP_3 under the best scheme. (a) and (b): simulation experiment. (c) and (d): forecast experiment. Left column high TSsum. Right column low TSsum.

Based on the above analyses, conclusions can be drawn as follows.

(1) The simulation ability of the DSAEF_LTP model with new parameter values of similarity region and ensemble method has been improved. The simulation ability of DSAEF_LTP_4 is the best among the four versions of the DSAEF_LTP model. Compared with DSAEF_LTP_1, the TS250 and TS100 of this model are improved by 14.7% and 15.6%, respectively. DSAEF_LTP_3 has the best performance among four versions of the DSAEF_LTP model in the independent sample forecast experiments, and its TS250 and TS100 are 0.1678 and 0.3203, respectively. The difference in forecast performance of the DSAEF_LTP model between simulation and forecast experiments may be due to the change in track characteristics between the two samples.

(2) The forecast performance of these two DSAEF_LTP model versions (DSAEF_LTP_3 and DSAEF_LTP_4) are better than that of the four dynamical models. The TSsum scores of DSAEF_LTP_3 and DSAEF_LTP_4 are 0.4881 and 0.3504, respectively. The TS scores of DSAEF_LTP_3 are higher than the best dynamical model, ECMWF, with 187% and 36% higher for TS250 and TS100, respectively. Moreover, the two improved DSAEF_LTP models (DSAEF_LTP_3 and DSAEF_LTP_4) tend to produce false alarms in accumulated forecasting precipitation above 250 mm and 100 mm, whereas the four dynamic models tend to miss.

(3) The forecast performance of the DSAEF_LTP model may change between TCs with different characteristics. In the current version of the model, heavy precipitation can be better forecast for TCs with heavy rainfall and typical TC tracks.

It has been found that the forecast performance of the DSAEF_LTP varies between TCs with different characteristics. In ongoing studies, we plan to address the variations in forecast performance for different stations over Southeast China. In addition, attention will be paid to the choice of experimental versus forecast samples, following on the fact that the results here are affected by changes in track characteristics. Further model improvements also can be explored by introducing other factors to determine analogs, including characteristics of the TC (e.g., the TC translation speed and structure) and environmental factors (e. g., the vertical wind shear, westerly trough, summer monsoon, and sub-height).