1. Introduction
Precipitation forecasting is one of the most challenging and complex weather prediction tasks. Accurately predicting the occurrence and amount of precipitation is challenging for meteorologists and forecasting experts. These challenges stem from factors such as nonlinear interactions between meteorological variables, spatial and temporal variability, complexity of physical processes, limitations in data availability, and observational errors.
However, advancements in technology, data-driven methodologies, and improved computing capabilities have contributed to the progress in precipitation forecasting. Certain techniques such as the use of high-resolution models to simulate realistic spatial patterns or postprocessing methods to correct systematic biases, have shown promise. Particularly, studies have attempted to enhance prediction models using data analysis techniques such as machine learning. Recent studies have explored diverse topics related to numerical weather and climate models. These include parameterizing physical processes, utilizing neural networks for short-term precipitation forecasting, and employing machine learning interpretability methods to predict large-scale precipitation patterns (Espeholt et al. 2022; Gibson et al. 2021; Ham et al. 2019). These studies cover various aspects, including the prediction of extreme rainfall, a trend and change in precipitation, and probability forecasting (Nayak and Ghosh 2013; Praveen et al. 2020; Jose et al. 2022). However, these studies reported limitations in their ability to accurately forecast precipitation.
The genetic algorithm (GA), a machine learning technique inspired by the concept of genes in problem solving, has been employed to address weight optimization challenges and has demonstrated potential across diverse applications. Katoch et al. (2021) offer a comprehensive review of the advancements in GAs. GAs have been widely utilized in the optimization of coefficients for the physical processes in NWP models. For example, model parameter optimization to simulate the hydrological cycle over the Han River basin in South Korea was observed to be potentially beneficial (Hong et al. 2014). Through GA experiments for physical parameter tuning, Lee et al. (2006) demonstrated improved performance for a heavy rainfall event that occurred on the Korean Peninsula.
Efforts to enhance numerical weather prediction models have led to substantial computational costs due to their large size considerable number of individuals or conducting repetitive executions. Consequently, there was a pressing need for efficient optimization methods that take into account factors like initial convergence speed and search range. The microGA (MGA), as proposed by Krishnakumar (1990), proved to be effective in handling small populations. This method was utilized by Park and Park (2021) to optimize the parameterization of physical processes within a grid using a global numerical weather prediction model.
A multimodel ensemble forecasting technique was developed to fully exploit information from each model and compensate for any deficiencies in the physical processes captured by the individual models (Ebert 2001). Ensemble members are typically created by combining the outputs from various operational centers. Previous studies have consistently shown that ensemble forecasting outperforms individual forecasting, particularly in terms of accuracy, in short- and medium-range weather predictions (Wei et al. 2022; Zhi et al. 2012). These studies demonstrated that calculating weight coefficients in multimodel ensemble prediction using multiple general circulation models yielded superior results compared to using a simple arithmetic average (Raftery et al. 2005; Liu and Xie 2014). GA has also been applied to determine optimal model weights in multimodel averaging. For instance, Ahn and Lee (2016) utilized GA for a multimodel mean, aiming to enhance the forecasting accuracy of temperature and precipitation. Similarly, Ratnam et al. (2019) demonstrated an improved seasonal temperature forecast by applying GA.
There are several ways of improving prediction occurrence. Given the additional challenge of combining precipitation from multiple general circulation models, we decided to develop our own model. We aimed to maximize the prediction accuracy of precipitation occurrence by optimizing weights among models. By increasing the weights of the models that demonstrated superior prediction performance in precipitation occurrence and decreasing the weights of underperforming models, we constructed a model with a more optimized performance than that of the operational model of the Korea Meteorological Administration (KMA). In the operational model, the precipitation was determined based on the uniformity of precipitation predictions in at least two or more models. Although this simple definition is effective in predicting precipitation occurrence, there is room for improvement.
In this study, we aim to identify postprocessing methods that improve the accuracy of forecasting models considering precipitation uncertainty and demonstrate enhanced rainfall prediction performance. For instance, we seek to develop techniques for creating an optimal forecast model by calculating the weights of merged numerical models using the MGA. Furthermore, we aim to evaluate the fitness functions for different components to identify the optimal rainfall model by maximizing their respective values to improve rainfall prediction performance in South Korea.
The remainder of this paper is organized as follows. The data used in this study and MGA methods are discussed in section 2. The performance of the optimal model, sensitivity experiments, and a case study are discussed in section 3. Section 4 constitutes a summary of the results.
2. Data and methods
a. Observation
First, 1-h accumulated precipitation was obtained from the Automated Weather Stations (AWS) of the KMA. Of the 741 stations in South Korea, 247 representative validation points were used (Fig. 1). Calculation of the distances between each location and its adjacent points showed that the average distance to adjacent points was 8.5 km (ranging from 1.0 to 18.3 km) without considering distant islands. This indicates a relatively high horizontal resolution, reflecting fine-scale precipitation patterns in the data. The results were compared with those of the operational models by interpolating the model grid points to the AWS locations. The training period spanned October 2020–July 2022 and the test period August 2022–May 2023. The training was performed seasonally, and the evaluation was monthly.
Study area with location of automated weather stations (AWS) in South Korea.
Citation: Artificial Intelligence for the Earth Systems 3, 1; 10.1175/AIES-D-23-0069.1
b. Models
The optimal system for precipitation prediction comprises three models: the Korean Integrated Model (KIM), Unified Model (UM), and Integrated Forecasting System (IFS). First, the KIM consists of a spectral-element nonhydrostatic dynamical core on a cubed-sphere grid, and a state-of-the-art physics parameterization package with initial conditions based on advanced hybrid four-dimensional ensemble–variational data assimilation (4DEnVar) (Hong et al. 2018). The horizontal resolution was NE360 (∼12 km), and the vertical level had 91 layers. No significant changes were observed during the study period. Second, we used the operational global NWP model of the KMA UM, version 10.8, known as Parallel Suite, version 5 (PS05). The horizontal resolution was N1280 (∼10 km) and the vertical level had 70 layers. The model constitutes hybrid 4DVAR data assimilation. Walters et al. (2017) provide related details. Finally, the IFS was updated from Cycle47r1 to Cycle47r3 during the study period (Maass 2022). However, we do not expect these changes to have any relevant effects on our results. The horizontal resolution was O1280 (∼9 km), and the vertical resolution was 137 levels.
c. MGA
A GA constitutes an initialization process that defines the genetic representation of a given problem. This process involves generating random individuals while defining the population size. A population consisting of a predetermined number of chromosomes, each of which represents a solution, is called a population pool. The genetic algorithm can accommodate multiple population pools. Although the GA are considered more traditional compared to deep learning techniques like long short-term memory, recurrent neural networks, and convolutional neural networks, it is still highly effective in finding optimal solutions within complex data, even with limited data availability. In a study conducted by Ahn and Lee (2016), they compared temperature and precipitation forecasts based on genetic algorithms with those not utilizing them, demonstrating the superior performance of the genetic algorithm-based multimodel average. In a more recent study, Kim et al. (2021) conducted a comparison among various weighting methods. This study explored different multimodel ensemble (MME) methods to enhance seasonal predictions using six World Meteorological Organization designated long-range models. The evaluated methods included simple composite, linear regression, the best selection anomaly, artificial neural network, and the genetic algorithm. The GA emerged as the most successful MME method, exhibiting superior performance in predicting global 2-m temperature and precipitation across all seasons.
The MGA is an efficient version of the conventional GA (Krishnakumar 1990). It employs smaller population sizes to reduce computational time, a necessity due to the large population sizes in numerical models. MGA addresses the convergence challenge in small populations by incorporating elitism, preserving the best individual in each generation. It also maintains diversity through reinitialization rather than mutation, making it particularly suitable for optimal parameter estimation and scheme-based optimization in meteorological applications.
The MGA database was created by optimizing the open-source Fortran Genetic Algorithm Front-End Drive Code, version 1.7.1 (Carroll 1996). See the data availability statement section for more details. Key settings applied to the code are listed in Table 1. The setting not mentioned here follows the default configuration proposed by Carroll (1996).
MGA settings used in the study.
d. Precipitation verification
Threshold values for precipitation detection were set at 0.25 mm h−1 for the rainy summer and autumn seasons and 0.1 mm h−1 for the dry winter and spring seasons.
The contingency table represents the predictive ability of the forecasting model and the types of errors that occur in the predictions (Table 2). The case where the model predicted the occurrence of rainfall and rainfall was recorded in the actual observations was classified as a hit (H) and a false alarm (F) if rainfall did not occur. Furthermore, when the model predicted the absence of rainfall but rainfall occurred, the case was classified as a miss (M), whereas the absence of rainfall in both the model prediction and observation was classified as a correct negative (C).
Contingency table of the precipitation.
e. Fitness
The fitness function we employed adopts a balanced approach among CSI, POD, and FBI to ensure a thorough evaluation of the model’s performance. We use a weighted sum of these metrics for balanced optimization. The evaluation involves applying the fitness function to both training and test datasets. During training, the model is iteratively refined to enhance its ability to identify pertinent precipitation patterns and relationships compared to observations. The test set is subsequently utilized to validate the model’s generalization capabilities. This methodical evaluation strategy ensures that the fitness function accurately reflects the model’s overall prediction accuracy.
We evaluated the fitness function seasonally for the 0000 and 1200 UTC initial cycles. Every day within the season was considered during the training period when evaluating the fitness function for each hour in the prediction time. This comprehensive approach allowed for an accurate assessment of the model’s performance for each time.
In this study, the MGA has a fixed population size and generation count: 20 and 50, respectively. In other words, 20 individuals were randomly generated and underwent stages of selection, crossover, and mutation to produce a new solution, thus completing one generation. This process was repeated 50 times to apply the proposed algorithm. After performing these iterations until a specified termination condition is met, the optimal population among the remaining individuals in the generation is considered the optimal coefficient.
The two key parameters in our study, the population size (20) and the number of generations (50), were established by optimizing the fitness function to achieve maximum performance. While this choice might appear arbitrary, we conducted sensitivity experiments to validate it. Figure 2 displays how CSI, POD, ACC, and FAR evolved across generations in response to varying population sizes. The colors purple, blue, black, and red represent population sizes of 5, 10, 20, and 30, respectively. A consistent trend observed in all figures is the continuous improvement in the model’s performance up to the initial 10 generations. However, post the 10th generation, there is a notable divergence in performance dependent on the population size. For instance, populations of 5 and 10 show inconsistent performance. In contrast, when the population size exceeds this range, the fluctuations in evaluation metrics with generational progression become less pronounced and tend to stabilize. Given that an excessively large population would lead to extended training times, a population size of 20 was considered optimal.
Evolution of performance metrics: CSI, POD, ACC, and FAR as a function of generation. Populations of 5, 10, 20, and 30 are shown in purple, blue, black, and red colors, respectively. The displayed results depict a 72-h forecast starting from the 0000 UTC initial time in summer.
Citation: Artificial Intelligence for the Earth Systems 3, 1; 10.1175/AIES-D-23-0069.1
The process of calculating the model weights for the optimal precipitation model was investigated using the MGA, and the fitness variations were examined. Figure 3 illustrates the maximum, minimum, and average fitness values for each generation and the corresponding model weights during the execution of the MGA up to the 50th generation. Although the maximum fitness value was relatively unchanged and converged, the minimum fitness value exhibited significant fluctuations. Additionally, the average fitness value showed minor variations over generations but generally converged to a stable value. By observing the convergence and divergence of fitness over 50 generations, we found that fitness began to converge after the 10th generation. Similar to the fitness convergence, the model weights exhibited a converging trend after the 10th generation. The progression of individual evaluation metrics, including CSI, POD, ACC, and FAR shows a convergence in their indices for the first 10 generations, similar to the fitness function. This is followed by gradual improvements in these indices up to 50 generations. Notably, the evolution of CSI closely mirrors that of the fitness function, underscoring its role in accurately predicting precipitation.
(top left) Evolution of fitness as a function of generation with a population size of 20. Two thin lines and a thick line between them represent the maximum, minimum, and average fitness values at each generation, respectively. (top right) The corresponding weight evolution of three models to the fitness: KIM (orange), UM (light blue), and IFS (dark blue). The sum of the model weights is 1. The values on the right side indicate the optimal model weights. (middle),(bottom) Evolution of CSI, POD, ACC, and FAR as a function of generation. Two thin lines and a thick line between them represent the maximum, minimum, and average values of metrics at each generation, respectively. The displayed results depict a 72-h forecast starting from the 0000 UTC initial time during the summer.
Citation: Artificial Intelligence for the Earth Systems 3, 1; 10.1175/AIES-D-23-0069.1
3. Results
The model weights for the optimal precipitation model as a function of forecast time across all seasons are illustrated in Fig. 4. The optimal weights of the models varied from zero to one for each forecast time. The IFS (dark blue) consistently had a higher weight than the other two models. This indicates that the IFS, on average, yielded more optimized precipitation predictions over the Korean Peninsula.
Time series of optimal weights from three models: KIM (orange), UM (light blue), and IFS (dark blue). The weights are shown for the (left) 0000 UTC and (right) 1200 UTC initial cycles in each season: (first row) spring, (second row) summer, (third row) autumn, and (fourth row) winter. At each forecast time, the sum of the weights is 1. Vertical dotted lines are placed at 24-h intervals. The local time, in Korea standard time (KST), is indicated below in parentheses. The x axis commences at the 6-h forecast, corresponding to 1500 KST (0300 KST) for 0000 UTC (1200 UTC) in KST.
Citation: Artificial Intelligence for the Earth Systems 3, 1; 10.1175/AIES-D-23-0069.1
The weight of the UM (light blue) exhibited periodic peaks during the early morning hours in spring (top panels in Fig. 4). UM is presumed to have performed remarkably in predicting rainfall during this period. Rainfall is primarily associated with a strong inflow of low-level jet streams into the Korean Peninsula. The time series data of the UM weight exhibited periodicity, and power spectrum analysis confirmed the presence of such periodicity (top panels in Fig. 5). This emphasizes significant frequency components that occur daily in the early morning.
Power spectrum of the model weight: KIM (orange), UM (light blue), and IFS (dark blue). The (left) 0000 UTC and (right) 1200 UTC initial cycles are shown in each season: (first row) spring, (second row) summer, (third row) autumn, and (fourth row) winter. Vertical dotted lines are placed at 12- and 24-h intervals.
Citation: Artificial Intelligence for the Earth Systems 3, 1; 10.1175/AIES-D-23-0069.1
We also investigated the characteristics of the model weights during summer (second row in Fig. 4). For predictions up to three days in advance, the IFS exhibited the highest weight. Interestingly, KIM’s weight peaked early in the morning. These two models dominated most weights, whereas UM’s weights remained relatively low for all prediction times within three days. The power spectrum of model weights effectively captured the daily cycle pattern (second row in Fig. 5). In the case of the 0000 UTC predictions, although the power magnitude was relatively small, a 12-h periodicity was clearly observed. Considering the weight trends beyond three days, both IFS and UM exhibited significant weights. Therefore, these characteristics appear to be reflected in the 12-h periodicity.
The UM weight tended to increase in the prediction performance beyond three days. The UM produces heavy rainfall during the summer, particularly convective precipitation (Han et al. 2021). This characteristic was particularly pronounced in predictions beyond three days. The tendency of UM to predict abundant precipitation is presumably consistent with the characteristics of frequent and heavy summer rainfall on the Korean Peninsula.
During autumn, in contrast with summer, UM exhibited the highest weight in predictions up to three days ahead (third row in Fig. 4), indicating its relatively optimized performance in predicting rainfall within the first three days. However, for predictions beyond three days, the IFS yielded the highest weight. The power spectrum analysis did not reveal any distinct periodicity (third row in Fig. 5). In winter, among the three models, the IFS consistently demonstrated the highest weight throughout the prediction period (bottom rows in Fig. 4).
Independently setting the model weights for each prediction time can lead to unexpected issues in the generation of an optimal rainfall model. If the weights change with each prediction time, the precipitation forecasts from the optimal model may become inconsistent and lack persistence. For example, in the 0000 UTC forecast during spring, the weights of the UM are most significant every 24 h. However, as time progresses, the UM weights diminish rapidly, and the IFS weights become predominant. The issue arises when these weight transitions are abrupt. Given the differing precipitation prediction patterns of each model, during periods of rapid weight changes, the precipitation patterns in the optimal model may suddenly shift. This is a significant concern for operational forecasts in KMA.
Continuity throughout the entire prediction time is essential for accurate rainfall prediction. This issue can be mitigated by developing a strategy that ensures continuous model weights, regardless of prediction time. Such an approach allows the model to maintain continuity in rainfall predictions while reducing temporal variability. One practical method is to maintain model weights based on past prediction results as an average value over a certain period. Therefore, for the continuity and reliability of rainfall prediction, it is crucial to adopt a constant weight strategy rather than setting weights for each prediction time separately.
In this study, we use the average weights for the entire prediction time. Considering the overall results across all seasons, the IFS consistently showed the highest average weights throughout the prediction period (Table 3). For example, the seasonal average weight of the IFS at 0000 UTC was 68% and 61% at 1200 UTC. This implies that the IFS generally performs optimally in rainfall prediction on the Korean Peninsula. However, although the average weights of the two other models did not exceed that of the IFS, there were seasons when these models exhibited relatively high weights. This suggests that the two models have specific strengths in rainfall prediction. For instance, UM showed the highest average weight during autumn compared with the other seasons (35% and 33% for 0000 and 1200 UTC, respectively). Conversely, KIM exhibited the highest average weights during summer compared to other seasons (12% and 20% for 0000 and 1200 UTC, respectively).
Model weights according to season (%).
Figure 6 shows the performance of the optimal model created using the average weights over the periods and compares the rainfall prediction performances of the three models, specifically for August 2022. The CSI ranges from 0.15 to 0.45, with higher values indicating more accurate models and lower values indicating underperforming models (top row in Fig. 6). Expectedly, the CSI values of the three models decreased as the prediction time increased. The results obtained using the MGA showed similar or more optimized performance throughout the prediction time compared to the three models. In other words, the model created using MGA enables more accurate predictions. The POD of the optimal model is very high. As the MGA-based model incorporates weights for predicting rainfall from the three models, it tends to detect rainfall in all areas where the three models predict rainfall. Therefore, the optimal rainfall prediction model created using the MGA exhibited a high POD, indicating a tendency to overestimate the rainfall area. Efforts to reduce this overestimation will be discussed later.
Performance metrics: CSI, POD, ACC, and FAR as a function of lead time starting from August 2022. (left) The 0000 UTC initial cycle and (right) the 1200 UTC initial cycle. Models KIM, UM, IFS, and MGA are color coded as orange, light blue, dark blue, and green, respectively. Black dotted lines are placed at 24-h intervals. The numbers displayed in each panel indicate the corresponding performance values for the models.
Citation: Artificial Intelligence for the Earth Systems 3, 1; 10.1175/AIES-D-23-0069.1
The ACC of the optimal model was similar to that of the three models. The FAR was comparable to or more optimized than that of the other three models. Specifically, it demonstrated more optimized performance than the other three models at 0000 UTC, and outperformed two models, excluding the IFS, at 1200 UTC.
Distribution of CSI, POD, ACC, and FAR as a function of parameters a and b, with a fixed value of c = 0.2, in August 2022. The operational performance of multimodel mean by the KMA is indicated in the upper and right sides of the panel and is highlighted in red in the CSI panel. The MGA model is represented at the intersection of the horizontal and vertical dotted lines.
Citation: Artificial Intelligence for the Earth Systems 3, 1; 10.1175/AIES-D-23-0069.1
POD values are displayed according to changes in the coefficients of CSI and POD in the fitness index in the top-right panel of Fig. 7. When b is below 0.3, increasing the value of a leads to continuous improvement in POD performance. However, when b exceeds 0.3, increasing the value of a leads to a decline in POD performance. However, the trend for ACC and FAR was opposite that of POD. Increasing the values of a or b resulted in continuous performance improvement.
The MGA for August 2022 generally yielded more optimized results for all metrics than the three other models. However, determining whether the results are limited to a single month or season is necessary. To address this, we extended the evaluation period to include all available data periods. Figure 8 illustrates the performance variations in rainfall prediction for the three models and MGA from August 2022 to May 2023, spanning 10 months. First, the MGA demonstrated higher CSI values than the other three models for most months by up to 13% [=(0.307 − 0.271)/0.271 for the KIM]. Notably, the MGA achieves this without applying additional physics to describe precipitation. The POD values were significantly higher than those of the other three models by up to 10% [=(0.636 − 0.579)/0.579 for the KIM]. The tendency to overestimate precipitation was observed across all seasons (second row in Fig. 8). However, the MGA does not appear to provide any advantages in terms of ACC. The FAR yielded more optimized performance than the three other models by up to 5% [=(0.624 − 0.655)/0.655 for the KIM]. Overall, the MGA demonstrated improved rainfall prediction performance.
Average performance indices over all lead times from August 2022 to May 2023 for the (left) 0000 UTC and (right) 1200 UTC initial cycles. Models KIM, UM, IFS, MGA, OPER, and MGA-N are color coded as orange, light blue, dark blue, green, black, and red, respectively. The values in each panel represent the average performance across all months.
Citation: Artificial Intelligence for the Earth Systems 3, 1; 10.1175/AIES-D-23-0069.1
Assessing whether the performance of the model created using the MGA was superior to that of the optimal operational rainfall model run by the KMA was important. The optimal operational rainfall model defined by the KMA is the arithmetic average of two or more precipitation prediction models. This approach aggregates common areas of precipitation predicted by multiple models, thereby enhancing the reliability of precipitation predictions in shared areas. The OPER represents the performance of the operational model used by the KMA, which outperformed the individual models in all metrics.
The performance level of the optimal model created using the MGA did not reach that of the OPER. That was primarily because the precipitation area of the model created using the MGA was very broad, consequently hindering accurate precipitation prediction. To address this, we introduced the definition used in the operational model to the MGA, which considers the precipitation commonly predicted by two or more models. This narrows down the precipitation area and improves the precipitation prediction accuracy. Consequently, the created model (MGA-N) outperformed the OPER in most metrics. For example, MGA-N showed up to 1.0% [=(0.331 − 0.328)/0.328], 0.4% [=(0.938 − 0.934)/0.934], and 6.8% [=(0.533 − 0.572)/0.572] improvement in the CSI, ACC, and FAR performance for the initial 0000 UTC cycle, respectively. Table 4 presents a summary of all metrics for MGA, OPER, and MGA-N models.
Average performance from August 2022 to May 2023.
The predictions of precipitation by three models were highly diverse (Fig. 9) owing to several physical and dynamic reasons. For example, dynamic conditions that dominate precipitation are often associated with atmospheric fronts and continuous ascent of moist air. Each model adopts a different approach. Accordingly, they have different optimal parameters, leading to diverse prediction patterns. The operational model used by the KMA is a simple arithmetic average of predictions from two or more models. The precipitation predicted by the operational model in the southern region mostly exceeded 2 mm h−1. The MGA-N, an enhancement of the MGA, illustrates precipitation exceeding 3 mm h−1, as recorded by the AWS. Moreover, the MGA-N emphasizes the spatial distribution of precipitation by assigning more weight to the IFS.
The 22-h forecast of precipitation (mm h−1) from the following models and observation for the 1200 UTC initial cycle on 13 Apr 2023: KIM, UM, IFS, OPER, MGA-N, and AWS.
Citation: Artificial Intelligence for the Earth Systems 3, 1; 10.1175/AIES-D-23-0069.1
4. Summary and discussion
Using machine learning methods, we developed an objective precipitation forecasting model. Combining the commonalities among individual prediction models, despite their limitations, allowed for the establishment of a relatively optimized multimodel. Furthermore, a model that outperforms the rest can be created by assigning higher weights to superior prediction models and lower weights to underperforming models. This process was achieved through an iterative approach using MGA.
Upon evaluating the performance of the developed precipitation optimized multimodel, we discovered that incorporating the MGA led to relatively precise predictions. Specifically, compared to the other models, the MGA-based model exhibited higher values in the CSI, POD, and FAR metrics by up to 13%, 10%, and 5%, respectively. However, the inclusion of MGA did not provide any advantage in the ACC metric. Overall, MGA plays a crucial role in enhancing the performance of precipitation predictions.
Additionally, sensitivity experiments were conducted using data from the 0000 UTC initial cycle in August 2022. The study examined the performance of the rainfall-optimal model based on the CSI:POD:FBI ratio. Increasing the CSI and POD coefficients improved CSI and POD performance, respectively. However, performance improvement was limited once a certain magnitude of the coefficients was reached. This suggested that CSI improved only with accurate precipitation predictions, and modifying the fitness components may not consistently result in continuous performance improvement.
The MGA model alone does not surpass the performance of an operational model in KMA. However, adopting the definition used in the operation improved model performance. The operational definition involves using two or more models to predict precipitation and defining the overlapping areas of the predicted precipitation as actual precipitation. This definition helps narrow the overall area of predicted precipitation, leading to relatively accurate predictions. Consequently, the model that combined the MGA and operational definition outperformed the CSI performance of the operation by up to 1%.
This outcome is an effective combination of the strengths of the MGA model and operational definition. Establishing an operational definition and incorporating it into the MGA model enhanced prediction accuracy and model performance. Therefore, the model that applied both the MGA and operational definition demonstrated superior performance compared to the operational standard.
Several limitations were encountered in the development of an optimal model. First, we focused solely on the presence of precipitation and not its specific amount, a significant variable worth considering. However, we observed the potential for improvement in the prediction of precipitation amounts in our case study. Second, our model relies on an operational definition that incorporates the common areas from multiple models. Although we tested different coefficients of the fitness function, optimal values and result robustness are unclear, necessitating further research. We hope that this study will provide a reference for further research on the development of more optimal precipitation models. For example, future optimal models should incorporate additional models or eliminate operational definitions. Third, although this study primarily aimed to develop an optimal prediction model, different approaches, such as deep learning techniques, can be used to create a superior model that outperforms the operational model of the KMA (Ham et al. 2019; Espeholt et al. 2022; Gibson et al. 2021). Nonetheless, we firmly believe that our specific choices do not undermine the overall conclusion that our objective model using the MGA demonstrates notable and measurable improvement in precipitation prediction.
Acknowledgments.
The authors thank two anonymous reviewers for their helpful comments. This study was funded by the Numerical Modeling Center of the Korea Meteorological Administration Research and Development Program for the Development of Numerical Weather Prediction and Data Application Techniques under Grant KMA2018-00721.
Data availability statement.
The MGA code used in this study is available from https://cuaerospace.com/products-services/genetic-algorithm. Observational and model data sets used in this publication may be found at accessible at https://data.kma.go.kr. Data can also be retrieved from the author upon request.
REFERENCES
Ahn, J.-B., and J. Lee, 2016: A new multimodel ensemble method using nonlinear genetic algorithm: An application to boreal winter surface air temperature and precipitation prediction. J. Geophys. Res. Atmos., 121, 9263–9277, https://doi.org/10.1002/2016JD025151.
Carroll, D. L., 1996: Chemical laser modeling with genetic algorithms. AIAA J., 34, 338–346, https://doi.org/10.2514/3.13069.
Ebert, E. E., 2001: Ability of a Poor Man’s ensemble to predict the probability and distribution of precipitation. Mon. Wea. Rev., 129, 2461–2480, https://doi.org/10.1175/1520-0493(2001)129<2461:AOAPMS>2.0.CO;2.
Espeholt, L., and Coauthors, 2022: Deep learning for twelve hour precipitation forecasts. Nat. Commun., 13, 5145, https://doi.org/10.1038/s41467-022-32483-x.
Gibson, P. B., W. E. Chapman, A. Altinok, L. Delle Monache, M. J. DeFlorio, and D. E. Waliser, 2021: Training machine learning models on climate model output yields skillful interpretable seasonal precipitation forecasts. Commun. Earth Environ., 2, 159, https://doi.org/10.1038/s43247-021-00225-4.
Ham, Y.-G., J.-H. Kim, and J.-J. Luo, 2019: Deep learning for multi-year ENSO forecasts. Nature, 573, 568–572, https://doi.org/10.1038/s41586-019-1559-7.
Han, S.-U., S.-M. Kong, J. Woo, and Y. J. Kim, 2021: Why? How! (in Korean). Forecast Bureau, Korea Meteorological Administration, 120 pp., https://www.kma.go.kr/down/e-learning/hands/hands_51.pdf.
Hong, S., X. Yu, S. K. Park, Y.-S. Choi, and B. Myoung, 2014: Assessing optimal set of implemented physical parameterization schemes in a multi-physics land surface model using genetic algorithm. Geosci. Model Dev., 7, 2517–2529, https://doi.org/10.5194/gmd-7-2517-2014.
Hong, S.-Y., and Coauthors, 2018: The Korean Integrated Model (KIM) system for global weather forecasting. Asia-Pac. J. Atmos. Sci., 54, 267–292, https://doi.org/10.1007/s13143-018-0028-9.
Jose, D. M., A. M. Vincent, and G. S. Dwarakish, 2022: Improving multiple model ensemble predictions of daily precipitation and temperature through machine learning techniques. Sci. Rep., 12, 4678, https://doi.org/10.1038/s41598-022-08786-w.
Katoch, S., S. S. Chauhan, and V. Kumar, 2021: A review on genetic algorithm: Past, present, and future. Multimedia Tools Appl., 80, 8091–8126, https://doi.org/10.1007/s11042-020-10139-6.
Kim, G., J.-B. Ahn, V. N. Kryjov, W.-S. Lee, D.-J. Kim, and A. Kumar, 2021: Assessment of MME methods for seasonal prediction using WMO LC-LRFMME hindcast dataset. Int. J. Climatol., 41, E2462–E2481, https://doi.org/10.1002/joc.6858.
Krishnakumar, K., 1990: Micro-genetic algorithms for stationary and non-stationary function optimization. Proc. SPIE, 1196, 289–296, https://doi.org/10.1117/12.969927.
Lee, Y. H., S. K. Park, and D.-E. Chang, 2006: Parameter estimation using the genetic algorithm and its impact on quantitative precipitation forecast. Ann. Geophys., 24, 3185–3189, https://doi.org/10.5194/angeo-24-3185-2006.
Liu, J., and Z. Xie, 2014: BMA probabilistic quantitative precipitation forecasting over the Huaihe Basin using TIGGE multimodel ensemble forecasts. Mon. Wea. Rev., 142, 1542–1555, https://doi.org/10.1175/MWR-D-13-00031.1.
Maass, K., 2022: Implementation of IFS Cycle 47c3. ECMWF, https://confluence.ecmwf.int/display/FCST/Implementation+of+IFS+Cycle+47r3.
Nayak, M. A., and S. Ghosh, 2013: Prediction of extreme rainfall event using weather pattern recognition and support vector machine classifier. Theor. Appl. Climatol., 114, 583–603, https://doi.org/10.1007/s00704-013-0867-3.
Park, S., and S. K. Park, 2021: A micro-genetic algorithm (GA v1.7.1a) for combinatorial optimization of physics parameterizations in the weather research and forecasting model (v4.0.3) for quantitative precipitation forecast in Korea. Geosci. Model Dev., 14, 6241–6255, https://doi.org/10.5194/gmd-14-6241-2021.
Praveen, B., and Coauthors, 2020: Analyzing trend and forecasting of rainfall changes in India using non-parametrical and machine learning approaches. Sci. Rep., 10, 10342, https://doi.org/10.1038/s41598-020-67228-7.
Raftery, A. E., T. Gneiting, F. Balabdaoui, and M. Polakowski, 2005: Using Bayesian Model averaging to calibrate forecast ensembles. Mon. Wea. Rev., 133, 1155–1174, https://doi.org/10.1175/MWR2906.1.
Ratnam, J. V., H. A. Dijkstra, T. Doi, Y. Morioka, M. Nonaka, and S. K. Behera, 2019: Improving seasonal forecasts of air temperature using a genetic algorithm. Sci. Rep., 9, 12781, https://doi.org/10.1038/s41598-019-49281-z.
Walters, D., and Coauthors, 2017: The Met Office Unified Model Global Atmosphere 6.0/6.1 and JULES Global Land 6.0/6.1 configurations. Geosci. Model Dev., 10, 1487–1520, https://doi.org/10.5194/gmd-10-1487-2017.
Wei, X., X. Sun, J. Sun, J. Yin, J. Sun, and C. Liu, 2022: A comparative study of multi-model ensemble forecasting accuracy between equal- and variant-weight techniques. Atmosphere, 13, 526, https://doi.org/10.3390/atmos13040526.
Zhi, X., H. Qi, Y. Bai, and C. Lin, 2012: A comparison of three kinds of multimodel ensemble forecast techniques based on the TIGGE data. Acta Meteor. Sin., 26, 41–51, https://doi.org/10.1007/s13351-012-0104-5.
