Abstract
Severe rainfall has become increasingly frequent and intense in the Taipei metropolitan area. A complex thunderstorm in the Taipei Basin on 14 June 2015 produced an extreme rain rate (>130 mm h−1), leading to an urban flash flood. This paper presents storms’ microphysical and dynamic features during the organizing and heavy rain stages, mainly based on observed polarimetric variables in a Doppler radar network and ground-based raindrop size distribution. Shallower isolated cells in the early afternoon characterized by big raindrops produced a rain rate > 10 mm h−1, but the rain showers persisted for a short time. The storm’s evolution highlighted the behavior of merged convective cells before the heaviest rainfall (exceeding 60 mm within 20 min). The columnar features of differential reflectivity (Z DR) and specific differential phase (K DP) became more evident in merged cells, which correlated with the broad distribution of upward motion and mixed-phase hydrometeors. The K DP below the environmental 0°C level increased toward the ground associated with the melted graupel and resulted in subsequent intense rain rates, showing the contribution of the ice-phase process. Due to the collision–breakup process, the highest concentrations of almost all drop sizes and smaller mass-weighted mean diameter occurred during the maximum rainfall stage.
Abstract
Severe rainfall has become increasingly frequent and intense in the Taipei metropolitan area. A complex thunderstorm in the Taipei Basin on 14 June 2015 produced an extreme rain rate (>130 mm h−1), leading to an urban flash flood. This paper presents storms’ microphysical and dynamic features during the organizing and heavy rain stages, mainly based on observed polarimetric variables in a Doppler radar network and ground-based raindrop size distribution. Shallower isolated cells in the early afternoon characterized by big raindrops produced a rain rate > 10 mm h−1, but the rain showers persisted for a short time. The storm’s evolution highlighted the behavior of merged convective cells before the heaviest rainfall (exceeding 60 mm within 20 min). The columnar features of differential reflectivity (Z DR) and specific differential phase (K DP) became more evident in merged cells, which correlated with the broad distribution of upward motion and mixed-phase hydrometeors. The K DP below the environmental 0°C level increased toward the ground associated with the melted graupel and resulted in subsequent intense rain rates, showing the contribution of the ice-phase process. Due to the collision–breakup process, the highest concentrations of almost all drop sizes and smaller mass-weighted mean diameter occurred during the maximum rainfall stage.
Abstract
This paper presents an innovational way of assimilating observations of clouds into the icosahedral nonhydrostatic weather forecasting model for regional scale (ICON-D2), which is operated by the German Weather Service (Deutscher Wetterdienst) (DWD). A convolutional neural network (CNN) is trained to detect clouds in camera photographs. The network’s output is a grayscale picture, in which each pixel has a value between 0 and 1, describing the probability of the pixel belonging to a cloud (1) or not (0). By averaging over a certain box of the picture a value for the cloud cover of that region is obtained. A forward operator is built to map an ICON model state into the observation space. A three-dimensional grid in the space of the camera’s perspective is constructed and the ICON model variable cloud cover (CLC) is interpolated onto that grid. The maximum CLC along the rays that fabricate the camera grid, is taken as a model equivalent for each pixel. After superobbing, monitoring experiments have been conducted to compare the observations and model equivalents over a longer time period, yielding promising results. Further we show the performance of a single assimilation step as well as a longer assimilation experiment over a time period of 6 days, which also yields good results. These findings are proof of concept and further research has to be invested before these new innovational observations can be assimilated operationally in any numerical weather prediction (NWP) model.
Abstract
This paper presents an innovational way of assimilating observations of clouds into the icosahedral nonhydrostatic weather forecasting model for regional scale (ICON-D2), which is operated by the German Weather Service (Deutscher Wetterdienst) (DWD). A convolutional neural network (CNN) is trained to detect clouds in camera photographs. The network’s output is a grayscale picture, in which each pixel has a value between 0 and 1, describing the probability of the pixel belonging to a cloud (1) or not (0). By averaging over a certain box of the picture a value for the cloud cover of that region is obtained. A forward operator is built to map an ICON model state into the observation space. A three-dimensional grid in the space of the camera’s perspective is constructed and the ICON model variable cloud cover (CLC) is interpolated onto that grid. The maximum CLC along the rays that fabricate the camera grid, is taken as a model equivalent for each pixel. After superobbing, monitoring experiments have been conducted to compare the observations and model equivalents over a longer time period, yielding promising results. Further we show the performance of a single assimilation step as well as a longer assimilation experiment over a time period of 6 days, which also yields good results. These findings are proof of concept and further research has to be invested before these new innovational observations can be assimilated operationally in any numerical weather prediction (NWP) model.
Abstract
Improving estimation of snow water equivalent rate (SWER) from radar reflectivity (Ze), known as a SWER(Ze) relationship, is a priority for NASA’s Global Precipitation Measurement (GPM) mission ground validation program as it is needed to comprehensively validate spaceborne precipitation retrievals. This study investigates the performance of eight operational and four research-based SWER(Ze) relationships utilizing Precipitation Imaging Probe (PIP) observations from the International Collaborative Experiment for Pyeongchang 2018 Olympic and Paralympic Winter Games (ICE-POP 2018) field campaign. During ICE-POP 2018, there were 10 snow events that are classified by synoptic conditions as either cold low or warm low, and a SWER(Ze) relationship is derived for each event. Additionally, a SWER(Ze) relationship is derived for each synoptic classification by merging all events within each class. Two new types of SWER(Ze) relationships are derived from PIP measurements of bulk density and habit classification. These two physically based SWER(Ze) relationships provided superior estimates of SWER when compared to the operational, event-specific, and synoptic SWER(Ze) relationships. For estimates of the event snow water equivalent total, the event-specific, synoptic, and best-performing operational SWER(Ze) relationships outperformed the physically based SWER(Ze) relationship, although the physically based relationships still performed well. This study recommends using the density or habit-based SWER(Ze) relationships for microphysical studies, whereas the other SWER(Ze) relationships are better suited toward hydrologic application.
Abstract
Improving estimation of snow water equivalent rate (SWER) from radar reflectivity (Ze), known as a SWER(Ze) relationship, is a priority for NASA’s Global Precipitation Measurement (GPM) mission ground validation program as it is needed to comprehensively validate spaceborne precipitation retrievals. This study investigates the performance of eight operational and four research-based SWER(Ze) relationships utilizing Precipitation Imaging Probe (PIP) observations from the International Collaborative Experiment for Pyeongchang 2018 Olympic and Paralympic Winter Games (ICE-POP 2018) field campaign. During ICE-POP 2018, there were 10 snow events that are classified by synoptic conditions as either cold low or warm low, and a SWER(Ze) relationship is derived for each event. Additionally, a SWER(Ze) relationship is derived for each synoptic classification by merging all events within each class. Two new types of SWER(Ze) relationships are derived from PIP measurements of bulk density and habit classification. These two physically based SWER(Ze) relationships provided superior estimates of SWER when compared to the operational, event-specific, and synoptic SWER(Ze) relationships. For estimates of the event snow water equivalent total, the event-specific, synoptic, and best-performing operational SWER(Ze) relationships outperformed the physically based SWER(Ze) relationship, although the physically based relationships still performed well. This study recommends using the density or habit-based SWER(Ze) relationships for microphysical studies, whereas the other SWER(Ze) relationships are better suited toward hydrologic application.
Abstract
The second part of this series presents results from verifying a precipitation forecast calibration method discussed in part 1, based on quantile mapping (QM), weighting of sorted members, and dressing of the ensemble. NOAA’s Global Ensemble Forecast System, version 12 (GEFSv12) reforecasts were used in this study. The method was validated with pre-operational GEFSv12 forecasts from the between December 2017 and November 2019. The method is proposed as an enhancement for GEFSv12 precipitation postprocessing in NOAA’s National Blend of Models.
Part 1 described adaptations to the methodology to leverage the ~ 20-year GEFSv12 reforecast data. As shown in this part 2, when compared to probabilistic quantitative precipitation forecasts (PQPFs) from the raw ensemble, the adapted method produced downscaled, high-resolution forecasts that were significantly more reliable and skillful than raw ensemble-derived probabilities, especially at shorter lead times (i.e., < 5 days) and for forecasts of events from light precipitation to > 10 mm 6 h−1. Cool-season events in the western US were especially improved when the QM algorithm was applied, providing a statistical downscaling with realistic smaller-scale detail related to terrain features. The method provided less value added for forecasts of longer lead times and for the heaviest precipitation.
Abstract
The second part of this series presents results from verifying a precipitation forecast calibration method discussed in part 1, based on quantile mapping (QM), weighting of sorted members, and dressing of the ensemble. NOAA’s Global Ensemble Forecast System, version 12 (GEFSv12) reforecasts were used in this study. The method was validated with pre-operational GEFSv12 forecasts from the between December 2017 and November 2019. The method is proposed as an enhancement for GEFSv12 precipitation postprocessing in NOAA’s National Blend of Models.
Part 1 described adaptations to the methodology to leverage the ~ 20-year GEFSv12 reforecast data. As shown in this part 2, when compared to probabilistic quantitative precipitation forecasts (PQPFs) from the raw ensemble, the adapted method produced downscaled, high-resolution forecasts that were significantly more reliable and skillful than raw ensemble-derived probabilities, especially at shorter lead times (i.e., < 5 days) and for forecasts of events from light precipitation to > 10 mm 6 h−1. Cool-season events in the western US were especially improved when the QM algorithm was applied, providing a statistical downscaling with realistic smaller-scale detail related to terrain features. The method provided less value added for forecasts of longer lead times and for the heaviest precipitation.
Abstract
Data from a dual-polarized, solid-state X-band radar and an operational C-band weather radar are used for high-resolution analyses of two hailstorms in the Vienna, Austria, region. The combination of both radars provides rapid-update (1 min) polarimetric data paired with wind field data of a dual-Doppler analysis. This is the first time that such an advanced setup is used to examine severe storm dynamics at the eastern Alpine fringe, where the influence of local topography is particularly challenging for thunderstorm prediction. We investigate two storms transitioning from the pre-Alps into the Vienna basin with different characteristics: 1) A rapidly evolving multicell storm producing large hail (5 cm), with observations of an intense Z DR column preceding hail formation and the rapid development of multiple pulses of hail; and 2) a cold pool–driven squall line with small hail, for which we find that the updraft location inhibited the formation of larger hailstones. For both cases, we analyzed the evolution of different Z DR column metrics as well as updraft speed and size and found that (i) the 90th percentile of Z DR within the Z DR column was highest for the cell later producing large hail, (ii) the peak 90th percentile of Z DR preceded large hailfall by 20 min and highest updraft size and speed by 10 min, and (iii) sudden drops of the 90th percentile of ZH within the Z DR column indicated imminent hailfall.
Significance Statement
Thunderstorm evolution on the transition from complex terrain into the Vienna basin in northeastern Austria varies strongly. In some instances, thunderstorm cells intensify once they reach flat terrain, while in most cases there is a weakening tendency. To improve our process understanding and short-term forecasting methods, we analyze two representative cases of hail-bearing storms transitioning into the Vienna basin. We mainly build our study on data from a new, cost-efficient weather radar, complemented by an operational radar, lightning observations, and ground reports. Our results show which radar variables could be well suited for early detection of intensification, and how they relate to thunderstorm updraft speeds and lightning activity.
Abstract
Data from a dual-polarized, solid-state X-band radar and an operational C-band weather radar are used for high-resolution analyses of two hailstorms in the Vienna, Austria, region. The combination of both radars provides rapid-update (1 min) polarimetric data paired with wind field data of a dual-Doppler analysis. This is the first time that such an advanced setup is used to examine severe storm dynamics at the eastern Alpine fringe, where the influence of local topography is particularly challenging for thunderstorm prediction. We investigate two storms transitioning from the pre-Alps into the Vienna basin with different characteristics: 1) A rapidly evolving multicell storm producing large hail (5 cm), with observations of an intense Z DR column preceding hail formation and the rapid development of multiple pulses of hail; and 2) a cold pool–driven squall line with small hail, for which we find that the updraft location inhibited the formation of larger hailstones. For both cases, we analyzed the evolution of different Z DR column metrics as well as updraft speed and size and found that (i) the 90th percentile of Z DR within the Z DR column was highest for the cell later producing large hail, (ii) the peak 90th percentile of Z DR preceded large hailfall by 20 min and highest updraft size and speed by 10 min, and (iii) sudden drops of the 90th percentile of ZH within the Z DR column indicated imminent hailfall.
Significance Statement
Thunderstorm evolution on the transition from complex terrain into the Vienna basin in northeastern Austria varies strongly. In some instances, thunderstorm cells intensify once they reach flat terrain, while in most cases there is a weakening tendency. To improve our process understanding and short-term forecasting methods, we analyze two representative cases of hail-bearing storms transitioning into the Vienna basin. We mainly build our study on data from a new, cost-efficient weather radar, complemented by an operational radar, lightning observations, and ground reports. Our results show which radar variables could be well suited for early detection of intensification, and how they relate to thunderstorm updraft speeds and lightning activity.
Abstract
This study revisits the issue of why tropical cyclones (TCs) develop more rapidly at lower latitudes, using ensemble axisymmetric numerical simulations and energy diagnostics based on the isentropic analysis, with the focus on the relative importance of the outflow-layer and boundary-layer inertial stabilities to TC intensification and energy cycle. Results show that although lowering the outflow-layer Coriolis parameter and thus inertial stability can slightly strengthen the outflow, it does not affect the simulated TC development, whereas lowering the boundary-layer Coriolis parameter largely enhances the secondary circulation and TC intensification as in the experiment with a reduced Coriolis parameter throughout the model atmosphere. This suggests that TC outflow is more likely a passive result of the convergent inflow in the boundary layer and convective updraft in the eyewall.
The boundary-layer inertial stability is found to control the convergent inflow in the boundary layer and depth of convection in the eyewall and thus the temperature of energy sink in the TC heat engine, which determines the efficiency and overall mechanical output of heat engine and thus TC intensification. It is also shown that the hypothesized isothermal and adiabatic compression legs at the downstream end of the outflow in the classical Carnot cycle is not supported in the thermodynamic cycle of the simulated TCs, implying that the assumed TC Carnot cycle is not closed. It is the theoretical maximum work of heat engine, not the energy expenditure following the outflow downstream, that determines the mechanical work used to intensify a TC.
Abstract
This study revisits the issue of why tropical cyclones (TCs) develop more rapidly at lower latitudes, using ensemble axisymmetric numerical simulations and energy diagnostics based on the isentropic analysis, with the focus on the relative importance of the outflow-layer and boundary-layer inertial stabilities to TC intensification and energy cycle. Results show that although lowering the outflow-layer Coriolis parameter and thus inertial stability can slightly strengthen the outflow, it does not affect the simulated TC development, whereas lowering the boundary-layer Coriolis parameter largely enhances the secondary circulation and TC intensification as in the experiment with a reduced Coriolis parameter throughout the model atmosphere. This suggests that TC outflow is more likely a passive result of the convergent inflow in the boundary layer and convective updraft in the eyewall.
The boundary-layer inertial stability is found to control the convergent inflow in the boundary layer and depth of convection in the eyewall and thus the temperature of energy sink in the TC heat engine, which determines the efficiency and overall mechanical output of heat engine and thus TC intensification. It is also shown that the hypothesized isothermal and adiabatic compression legs at the downstream end of the outflow in the classical Carnot cycle is not supported in the thermodynamic cycle of the simulated TCs, implying that the assumed TC Carnot cycle is not closed. It is the theoretical maximum work of heat engine, not the energy expenditure following the outflow downstream, that determines the mechanical work used to intensify a TC.
Abstract
The atmospheric circulation response to global warming is an important problem that is theoretically still not well understood. This is a particular issue since climate model simulations provide uncertain, and at times contradictory, projections of future climate. In particular, it is still unclear how a warmer and moister atmosphere will affect midlatitude eddies and their associated poleward transport of heat and moisture. Here we perform a trend analysis of three main components of the global circulation—the zonal-mean state, eddies, and the net energy input into the atmosphere—and examine how they relate in terms of a moist static energy budget for the JRA-55 reanalysis data. A particular emphasis is made on understanding the contribution of moisture to circulation trends. The observed trends are very different between the hemispheres. In the Southern Hemisphere there is an overall strengthening and during boreal summer, also a poleward shifting, of the jet stream, the eddies, and the meridional diabatic heating gradients. Correspondingly, we find an overall strengthening of the meridional gradients of the net atmospheric energy input. In the Northern Hemisphere, the trend patterns are more complex, with the dominant signal being a clear boreal winter Arctic amplification of positive trends in lower-tropospheric temperature and moisture, as well as a significant weakening of both bandpass and low-pass eddy heat and moisture fluxes. Consistently, surface latent and sensible heat fluxes, upward and downward longwave radiation, and longwave cloud radiative fluxes at high latitudes show significant trends. However, radiative fluxes and eddy fluxes are inconsistent, suggesting data assimilation procedures need to be improved.
Significance Statement
We use a long-term reanalysis dataset to get an overall view of the changes in the global circulation and its role in transporting moist static energy from the equator to the poles. We do this by examining the trends in its three main components—the zonal means, the eddies, and the net energy input into the atmosphere. We find that in the Southern Hemisphere, there is an overall strengthening of the eddies, their poleward energy fluxes, and correspondingly the meridional gradients of the net atmospheric energy input. In the Northern Hemisphere, though the pattern is more complex, there is an overall weakening of the eddies and poleward eddy fluxes, and of the meridional gradients of the net atmospheric energy input, consistent with Arctic warming.
Abstract
The atmospheric circulation response to global warming is an important problem that is theoretically still not well understood. This is a particular issue since climate model simulations provide uncertain, and at times contradictory, projections of future climate. In particular, it is still unclear how a warmer and moister atmosphere will affect midlatitude eddies and their associated poleward transport of heat and moisture. Here we perform a trend analysis of three main components of the global circulation—the zonal-mean state, eddies, and the net energy input into the atmosphere—and examine how they relate in terms of a moist static energy budget for the JRA-55 reanalysis data. A particular emphasis is made on understanding the contribution of moisture to circulation trends. The observed trends are very different between the hemispheres. In the Southern Hemisphere there is an overall strengthening and during boreal summer, also a poleward shifting, of the jet stream, the eddies, and the meridional diabatic heating gradients. Correspondingly, we find an overall strengthening of the meridional gradients of the net atmospheric energy input. In the Northern Hemisphere, the trend patterns are more complex, with the dominant signal being a clear boreal winter Arctic amplification of positive trends in lower-tropospheric temperature and moisture, as well as a significant weakening of both bandpass and low-pass eddy heat and moisture fluxes. Consistently, surface latent and sensible heat fluxes, upward and downward longwave radiation, and longwave cloud radiative fluxes at high latitudes show significant trends. However, radiative fluxes and eddy fluxes are inconsistent, suggesting data assimilation procedures need to be improved.
Significance Statement
We use a long-term reanalysis dataset to get an overall view of the changes in the global circulation and its role in transporting moist static energy from the equator to the poles. We do this by examining the trends in its three main components—the zonal means, the eddies, and the net energy input into the atmosphere. We find that in the Southern Hemisphere, there is an overall strengthening of the eddies, their poleward energy fluxes, and correspondingly the meridional gradients of the net atmospheric energy input. In the Northern Hemisphere, though the pattern is more complex, there is an overall weakening of the eddies and poleward eddy fluxes, and of the meridional gradients of the net atmospheric energy input, consistent with Arctic warming.
Abstract
Reconstructing the history of polar temperature from ice core water isotope (δ 18O) calibration has remained a challenge in paleoclimate research, because of our incomplete understanding of various temperature–δ 18O relationships. This paper resolves this classical problem in a new framework called the unified slope equations (USE), which illustrates the general relations among spatial and temporal δ 18O–surface temperature slopes. The USE is applied to the Antarctica temperature change during the last deglaciation in model simulations and observations. It is shown that the comparable Antarctica-mean spatial slope with deglacial temporal slope in δ 18O–surface temperature reconstruction is caused, accidentally, by the compensation responses between the δ 18O–inversion layer temperature relation and the inversion layer temperature itself. Furthermore, in light of the USE, we propose that the present seasonal slope of δ 18O–inversion layer temperature is an optimal paleothermometer that is more accurate and robust than the spatial slope. This optimal slope suggests the possibility of reconstructing past Antarctic temperature changes using present and future instrumental observations.
Significance Statement
This paper develops a new framework called the unified slope equations (USE) to provide, for the first time, a general relation among various spatial and temporal water isotope–temperature slopes. The application of the USE to Antarctic deglacial temperature change shows that the optimal paleothermometer is the seasonal slope of the inversion layer temperature.
Abstract
Reconstructing the history of polar temperature from ice core water isotope (δ 18O) calibration has remained a challenge in paleoclimate research, because of our incomplete understanding of various temperature–δ 18O relationships. This paper resolves this classical problem in a new framework called the unified slope equations (USE), which illustrates the general relations among spatial and temporal δ 18O–surface temperature slopes. The USE is applied to the Antarctica temperature change during the last deglaciation in model simulations and observations. It is shown that the comparable Antarctica-mean spatial slope with deglacial temporal slope in δ 18O–surface temperature reconstruction is caused, accidentally, by the compensation responses between the δ 18O–inversion layer temperature relation and the inversion layer temperature itself. Furthermore, in light of the USE, we propose that the present seasonal slope of δ 18O–inversion layer temperature is an optimal paleothermometer that is more accurate and robust than the spatial slope. This optimal slope suggests the possibility of reconstructing past Antarctic temperature changes using present and future instrumental observations.
Significance Statement
This paper develops a new framework called the unified slope equations (USE) to provide, for the first time, a general relation among various spatial and temporal water isotope–temperature slopes. The application of the USE to Antarctic deglacial temperature change shows that the optimal paleothermometer is the seasonal slope of the inversion layer temperature.
Abstract
Statistical downscaling (SD) of climate change projections is a key piece for impact and adaptation studies, due to its low computational expense compared to dynamical downscaling, which allows to explore uncertainties through the generation of large ensembles. SD has been extensively evaluated and applied in the extratropics, but few examples exist in tropical regions. In this study several state-of-the-art methods belonging to different families have been evaluated for maximum/minimum daily temperature and daily accumulated precipitation (both from the ERA5 reanalysis at 0.25°) in two regions with very different climates: Spain (Mid-latitudes) and Central America (Tropics). Some key assumptions of SD have been tested: the strength of the predictors/predictand links, the skill of different approaches and the extrapolation capability of each method. It has been found that relevant predictors are different in both regions, as well as the behavior of statistical methods. For temperature, most methods perform significantly better in Spain than in Central America, where Transfer Function methods present important extrapolation problems, probably due to the low variability of the training sample (present climate). In both regions, Model Output Statistics (MOS) methods have achieved the best results for temperature. In Central America Transfer Function (TF) methods have achieved better results than MOS methods in the evaluation in the present climate, but they do not preserve trends in the future. For precipitation, MOS methods and the machine learning method eXtreme Gradient Boost have achieved the best results in both regions. Additionally, it has been found that although the use of humidity indexes as predictors improve results for the downscaling of precipitation, future trends given by statistical methods are very sensitive to the use of one or another index. Three indexes have been compared: relative humidity, specific humidity and dew point depression. The use of the specific humidity has been found to seriously deviate trends given by the downscaled projections from those given by raw Global Climate Models in both regions.
Abstract
Statistical downscaling (SD) of climate change projections is a key piece for impact and adaptation studies, due to its low computational expense compared to dynamical downscaling, which allows to explore uncertainties through the generation of large ensembles. SD has been extensively evaluated and applied in the extratropics, but few examples exist in tropical regions. In this study several state-of-the-art methods belonging to different families have been evaluated for maximum/minimum daily temperature and daily accumulated precipitation (both from the ERA5 reanalysis at 0.25°) in two regions with very different climates: Spain (Mid-latitudes) and Central America (Tropics). Some key assumptions of SD have been tested: the strength of the predictors/predictand links, the skill of different approaches and the extrapolation capability of each method. It has been found that relevant predictors are different in both regions, as well as the behavior of statistical methods. For temperature, most methods perform significantly better in Spain than in Central America, where Transfer Function methods present important extrapolation problems, probably due to the low variability of the training sample (present climate). In both regions, Model Output Statistics (MOS) methods have achieved the best results for temperature. In Central America Transfer Function (TF) methods have achieved better results than MOS methods in the evaluation in the present climate, but they do not preserve trends in the future. For precipitation, MOS methods and the machine learning method eXtreme Gradient Boost have achieved the best results in both regions. Additionally, it has been found that although the use of humidity indexes as predictors improve results for the downscaling of precipitation, future trends given by statistical methods are very sensitive to the use of one or another index. Three indexes have been compared: relative humidity, specific humidity and dew point depression. The use of the specific humidity has been found to seriously deviate trends given by the downscaled projections from those given by raw Global Climate Models in both regions.
Abstract
Based on the principle “learn from past errors to correct current forecasts,” statistical postprocessing consists of optimizing forecasts generated by numerical weather prediction (NWP) models. In this context, machine learning (ML) offers state-of-the-art tools for training statistical models and making predictions based on large datasets. In our study, ML-based solutions are developed to reduce forecast errors of 2-m temperature and 10-m wind speed of the ECMWF’s operational medium-range, high-resolution forecasts produced with the Integrated Forecasting System (IFS). IFS forecasts and other spatiotemporal indicators are used as predictors after careful selection with the help of ML interpretability tools. Different ML approaches are tested: linear regression, random forest decision trees, and neural networks. Statistical models of systematic and random errors are derived sequentially where the random error is defined as the residual error after bias correction. In terms of output, bias correction and forecast uncertainty prediction are made available at any point from locations around the world. All three ML methods show a similar ability to capture situation-dependent biases leading to noteworthy performance improvements (between 10% and 15% improvement in terms of root-mean-square error for all lead times and variables), and a similar ability to provide reliable uncertainty predictions.
Abstract
Based on the principle “learn from past errors to correct current forecasts,” statistical postprocessing consists of optimizing forecasts generated by numerical weather prediction (NWP) models. In this context, machine learning (ML) offers state-of-the-art tools for training statistical models and making predictions based on large datasets. In our study, ML-based solutions are developed to reduce forecast errors of 2-m temperature and 10-m wind speed of the ECMWF’s operational medium-range, high-resolution forecasts produced with the Integrated Forecasting System (IFS). IFS forecasts and other spatiotemporal indicators are used as predictors after careful selection with the help of ML interpretability tools. Different ML approaches are tested: linear regression, random forest decision trees, and neural networks. Statistical models of systematic and random errors are derived sequentially where the random error is defined as the residual error after bias correction. In terms of output, bias correction and forecast uncertainty prediction are made available at any point from locations around the world. All three ML methods show a similar ability to capture situation-dependent biases leading to noteworthy performance improvements (between 10% and 15% improvement in terms of root-mean-square error for all lead times and variables), and a similar ability to provide reliable uncertainty predictions.