Deterministic Forecasts Research Articles

Probabilistic medium-range forecasts of extreme heat events over East Asia based on a global ensemble forecasting system

This study suggests a methodology for probabilistic forecasts of the extreme heat events in East Asia based on an operational global ensemble prediction used by the Korea Meteorological Administration (KMA). It focuses on the medium range of up to 11 days, providing probabilities of heatwave and tropical night occurrence each day. Forecast validation in the summer from 2016 to 2021 shows that the deterministic heatwave forecast provides 5 days of optimal forecast range, while the probabilistic forecast can extend the practically predictable range up to 10 days in the Korean Peninsula and 7 days in Japan, respectively. Comparing prediction skills for heatwave and tropical night, the skills for tropical night tend to be inferior, presumably due to complex mechanisms of the tropical night and large uncertainty in the numerical model, such as microphysics and radiation. In addition, the coarse resolution of the operational system does not seem to resolve temperature variability at night. As a case study, this study also examines the forecast of the onset and offset of the 2018 South Korean heatwave event. The temporal evolution of the heatwave matches well with the changes in the upper-level atmospheric circulation pattern, which can be used for useful forecast guidance. This probabilistic forecast based on the global ensemble forecasting system is expected to provide reliable prediction information for heatwaves in advance, reducing exposure to extreme events.

Performance Evaluation of a National Seven-Day Ensemble Streamflow Forecast Service for Australia

The Australian Bureau of Meteorology offers a national operational 7-day ensemble streamflow forecast service covering regions of high environmental, economic, and social significance. This semi-automated service generates streamflow forecasts every morning and is seamlessly integrated into the Bureau’s Hydrologic Forecasting System (HyFS). Ensemble rainfall forecasts, European Centre for Medium-Range Weather Forecasts (ECMWF), and Poor Man’s Ensemble (PME), available in the Numerical Weather Prediction (NWP) suite, are used to generate these streamflow forecasts. The NWP rainfall undergoes pre-processing using the Catchment Hydrologic Pre-Processer (CHyPP) before being fed into the GR4H rainfall–runoff model, which is embedded in the Short-term Water Information Forecasting Tools (SWIFT) hydrological modelling package. The simulated streamflow is then post-processed using Error Representation and Reduction In Stages (ERRIS). We evaluated the performance of the operational rainfall and streamflow forecasts for 96 catchments using four years of operational data between January 2020 and December 2023. Performance evaluation metrics included the following: CRPS, relative CRPS, CRPSS, and PIT-Alpha for ensemble forecasts; NSE, PCC, MAE, KGE, PBias, and RMSE; and three categorical metrics, CSI, FAR, and POD, for deterministic forecasts. The skill scores, CRPS, relative CRPS, CRPSS, and PIT-Alpha, gradually decreased for both rainfall and streamflow as the forecast horizon increased from Day 1 to Day 7. A similar pattern emerged for NSE, KGE, PCC, MAE, and RMSE as well as for the categorical metrics. Forecast performance also progressively decreased with higher streamflow volumes. Most catchments showed positive performance skills, meaning the ensemble forecast outperformed climatology. Both streamflow and rainfall forecast skills varied spatially across the country—they were generally better in the high-runoff-generating catchments, and poorer in the drier catchments situated in the western part of the Great Dividing Range, South Australia, and the mid-west of Western Australia. We did not find any association between the model forecast skill and the catchment area. Our findings demonstrate that the 7-day ensemble streamflow forecasting service is robust and draws great confidence from agencies that use these forecasts to support decisions around water resource management.

Downscaled numerical weather predictions can improve forecasts of sugarcane irrigation indices

Efficient irrigation reduces energy and water costs, increases profit margins and delivers better environmental outcomes. Whilst many growers rely on weather forecasts to make decisions, few studies have sought to incorporate weather forecast uncertainty into the optimisation of irrigation management or to evaluate weather forecasts through the lens of irrigation indices. Therefore, in this study, we seek to generate and evaluate ensemble forecasts of irrigation indices produced by coupling numerical weather prediction (NWP) forecasts with a biophysical process model (APSIM). We investigate a case study application for sugarcane in northeastern Australia. As a first step, three and a half years of forecasts from the Australian Bureau of Meteorology’s ACCESS-G3 model are statistically post-processed to generate 7-day forecasts that are downscaled and calibrated to local climate zones. In addition, the forecast post-processor converts the deterministic forecasts into an ensemble, thus quantifying forecast uncertainty. The generated forecasts are then used as forcing for the APSIM crop model to produce ensemble forecasts of soil water deficit (SWD), crop water use (CWU) and crop stress (Stress) for a simulated sugarcane crop. Through cross-validation, the post-processed weather forecasts demonstrate improved forecast accuracy compared to naive climatology and raw NWP forecasts for daily rainfall, maximum and minimum temperature and solar radiation; with the added benefit of providing a reliable uncertainty estimate. Improvement to an even greater degree is observed for the derived irrigation indices, particularly CWU and Stress, for which the forecasts are also reliable. The developed irrigation indices based on NWP can be used directly for decision making or, alternatively, may be used further in machine learning for optimisation of irrigation schedules in conjunction with other remotely sensed variables.

AI‐Based Ensemble Flood Forecasts and Its Implementation in Multi‐Objective Robust Optimization Operation for Reservoir Flood Control

AbstractInforming reservoirs with forecasts is highly important for real‐time flood control. This study proposed a forecast‐informed methodology framework for reservoir flood control operation under uncertainty. A new combination of two post‐processing methods, that is, the Cloud model and error‐based copula functions, were developed to merge individual AI‐based forecasts to ensemble flood forecasts, so called stochastic errors‐based Cloud (SE‐Cloud). A multi‐objective robust optimization model (MRO) integrating the risk, resilience, and vulnerability was then proposed to tackle flood control problems under ensemble forecasts; for comparison, a two‐objective stochastic optimization model (TSO) was developed to minimize the expected highest reservoir level and peak release. The proposed methodology was applied to the Lishimen reservoir in the Shifeng River subbasin, China, aiming to comprehensively verify the relationships among deterministic forecasts, ensemble forecasts, and flood control performance. Results showed that the Cloud model could effectively integrate different models and improve forecast accuracy. But a higher deterministic forecast quality did not consistently result in improved flood control performance. SE‐Cloud could capture the peak flow and effectively characterize forecast uncertainties and increased hypervolume values by 13.14%–39.65% compared to the Cloud model, indicating the superiority of ensemble forecasts in generating robust solutions over individual deterministic forecasts. MRO released more inflow than TSO, decreasing the expected highest water level by 0.05 m and incrementing the expected peak release by 4.29%. However, with downstream resilience value remaining at zero, it is demonstrated that MRO improving upstream vulnerability did not necessarily diminish resilience. The enhanced robustness highlights the potential of AI‐based ensemble forecasts in flood control.

Evaluation of Probabilistic Forecasts of Binary Events with the Neighborhood Brier Divergence Skill Score

Abstract A procedure for evaluating the quality of probabilistic forecasts of binary events has been developed. This is based on a two-step procedure: pooling of forecasts on the one hand and observations on the other hand, on all the points of a neighborhood in order to obtain frequencies at the neighborhood length scale and then to calculate the Brier divergence for these neighborhood frequencies. This score allows the comparison of a probabilistic forecast and observations at the neighborhood length scale, and therefore, the rewarding of event forecasts shifted from the location of the observed event by a distance smaller than the neighborhood size. A new decomposition of this score generalizes that of the Brier score and allows the separation of the generalized resolution, reliability, and uncertainty terms. The neighborhood Brier divergence skill score (BDnSS) measures the performance of the probabilistic forecast against the sample climatology. BDnSS and its decomposition have been used for idealized and real cases in order to show the utility of neighborhoods when comparing at different scales the performances of ensemble forecasts between themselves or with deterministic forecasts or of deterministic forecasts between themselves. Significance Statement A pooling of forecasts on the one hand and observations on the other hand, on all the points of a neighborhood, is performed in order to obtain frequencies at the neighborhood scale. The Brier divergence is then calculated for these neighborhood frequencies to compare a probabilistic forecast and observations at the neighborhood scale. A new decomposition of this score generalizes that of the Brier score and allows the separation of the generalized resolution, reliability, and uncertainty terms. This uncertainty term is used to define the neighborhood Brier divergence skill score which is an alternative to the popular fractions skill score, with a more appropriate denominator.

ST-TransNet: A Spatiotemporal Transformer Network for Uncertainty Estimation from a Single Deterministic Precipitation Forecast

Abstract The forecast uncertainty, particularly for precipitation, serves as a crucial indicator of the reliability of deterministic forecasts. Traditionally, forecast uncertainty is estimated by ensemble forecasting, which is computationally expensive since the forecast model is run multiple times with perturbations. Recently, deep learning methods have been explored to learn the statistical properties of ensemble prediction systems due to their low computational costs. However, accurately and effectively capturing the uncertainty information in precipitation forecasts remains challenging. In this study, we present a novel spatiotemporal transformer network (ST-TransNet) as an alternative approach to estimate uncertainty with ensemble spread and probabilistic forecasts, by learning from historical ensemble forecasts. ST-TransNet features a hierarchical structure for extracting multiscale features and incorporates a spatiotemporal transformer module with window-based attention to capture correlations in both spatial and temporal dimensions. Additionally, window-based attention can not only extract local precipitation patterns but also reduce computational costs. The proposed ST-TransNet is evaluated on the TIGGE ensemble forecast dataset and Global Precipitation Measurement (GPM) precipitation products. Results show that ST-TransNet outperforms both traditional and deep learning methods across various metrics. Case studies further demonstrate its ability to generate reasonable and accurate spread and probability forecasts from a single deterministic precipitation forecast. It demonstrates the capacity and efficiency of neural networks in estimating precipitation forecast uncertainty.

Steering Flow Sensitivity in Forecast Models for Hurricane Ian (2022)

Abstract Hurricane Ian made landfall along Florida’s west coast at 1905 UTC 28 September 2022 near the Fort Myers area as a high-impact storm. Here, we examine the potential link between track forecast errors near the time of landfall and errors in both the synoptic-scale upper-level flow and a shortwave moving within that flow. Five days before the actual landfall (0000 UTC 23 September), most model guidance indicated landfall would occur close to where Ian eventually came ashore. But by 0000 UTC 25 September, model forecasts were all forecasting landfall in the Florida Panhandle. One day later, the models again agreed with each other but for a landfall 100–200 km north of Tampa, Florida. By 0000 UTC 27 September, forecast models indicated landfall would occur near Tampa. Model forecasts continued shifting to the right and finally converged on Punta Gorda, Florida, as the landfall location, less than 24 h before landfall. In this short article, we hypothesize that the track of Ian depended on subtle interactions with an extratropical wave in the middle and upper atmosphere. Deterministic and ensemble model forecasts reveal that the interactions were very sensitive to the characteristics of this wave and the synoptic-scale flow in which the wave was embedded. A 1–2-dam difference in the geopotential heights played a major role in whether Ian moved north into the Panhandle or toward the east, making landfall in central Florida. Significance Statement The purpose of this work is to look at possible causes of error in forecasts of Hurricane Ian’s landfall with the intent of saving lives and reducing damage, which can run to the billions of dollars. The path that hurricanes take is strongly influenced by the larger-scale weather patterns around them. We find that small errors in the upper-level flow patterns north of the hurricane appear to have played an important role in the forecast. These errors were present in both the general flow and a wave moving through that flow, resulting in small but significant changes to the forecast path. Understanding the cause of forecast errors is a key step both in improving forecasts and in identifying types of events that are more difficult to forecast.

How well can global ensemble forecasts predict tropical cyclones in the southwest Indian Ocean?

AbstractThe southwest Indian Ocean (SWIO) recently experienced its most active, costliest and deadliest cyclone season on record (2018–2019). The anticipation and forecasting of natural hazards, such as tropical cyclones, are crucial to preparing for their impacts, but it is important to understand how well forecasting systems can predict them. Despite the vulnerability of the SWIO to tropical cyclones, comparatively little research has focused on this region, including understanding the ability of numerical weather prediction systems to predict cyclones and their impacts in southeast Africa. In this study, we evaluate ensemble probabilistic and high‐resolution deterministic forecasts of tropical cyclones in the SWIO from 2010 to 2020, using two state‐of‐the‐art global forecasting systems: one from the European Centre for Medium‐Range Weather Forecasts (ECMWF) and the other from the U.K. Met Office. We evaluate predictions of the track, assessing the location of the centre of each storm and its speed of movement, as well as its intensity, looking at maximum wind speeds and minimum central pressure, and discuss how the forecasts have evolved over the 10‐year period. Overall, ECMWF typically provides more accurate forecasts, but both systems tend to underestimate translation speed and intensity. We also investigate the impact of the Madden‐Julian Oscillation (MJO) on tropical cyclones and their forecasts. The MJO impacts where and when tropical cyclones form, their tracks and intensities, which in turn impacts forecast skill. These results are intended to provide an increased understanding of the ability of global forecasting systems to predict tropical cyclones in the SWIO, for the purpose of decision making and anticipatory action.

The role of soil moisture-temperature coupling for the 2018 Northern European heatwave in a subseasonal forecast

This study investigates the predictability of the 2018 Northern Europe heatwave using the GloSea5 forecast model from the perspective of land-atmosphere interactions. We focus on an inverse relationship wherein soil drying leads to increased temperatures and the model's ability to simulate this hypersensitivity in the soil moisture-temperature coupling on the dry side of a breakpoint defined as the soil moisture threshold below which land feedbacks nonlinearly amplify extreme heat. When evaluating forecast model performance in predicting this heatwave, we compare deterministic forecast scores (Hit Rate (HR) and True Skill Score (TSS)) for whether model Surface Soil Moisture (SSM) falls within the hypersensitive regime. GloSea5 exhibits enhanced prediction skill for the extreme heat event when the modelled soil moisture is within the hypersensitive regime. To understand the skill of the heatwave forecast for hit and missed cases of capturing SSM below the breakpoint, we first evaluate the climatological model performance for the water- and energy-limited processes, and then perform a comparison classified by whether SSM verifies on the dry side of the wilting point. The composite analysis demonstrates that the reproducibility of the breakpoint is tied to an improvement in climatological land coupling processes, mainly for classification in the water-limited coupling regime. Therefore, the results suggest that the process-based connection between soil moisture and temperature is a potential source for improving heatwave forecasts on subseasonal to seasonal (S2S) time scales.

КОНСТРУКТИВНО-ПРОДУКЦІЙНЕ МОДЕЛЮВАННЯ ДЕТЕРМІНОВАНИХ ЧАСОВИХ РЯДІВ З ФРАКТАЛЬНИМИ ВЛАСТИВОСТЯМИ

This study represents the usage of the constructive-synthesizing modeling for deterministic fractal time series forecasting by finding the appropriate model. The constructive-synthesizing modeling was already successfully used for generation various time series using L-system and normal distribution. The main purpose of this research is the opposite process – find the model, which is related to the given values. The composite constructor has been developed and it consists of four interconnected constructors, which are performing actions from generating the time series to finding a corresponding model using genetic algorithm. The software solution has been developed to achieve the automatization of the testing process. There are four main stages of the experiments with different complexity of the time series models, starting with one-rule with four characters in the right part and ending with one-to-three rules with four-to-ten characters in the right part and complex axiom. The practical application of this solution is requiring further studying with using stochastic real-life time series.

Generative Ensemble Deep Learning Severe Weather Prediction from a Deterministic Convection-Allowing Model

Abstract An ensemble postprocessing method is developed for the probabilistic prediction of severe weather (tornadoes, hail, and wind gusts) over the conterminous United States (CONUS). The method combines conditional generative adversarial networks (CGANs), a type of deep generative model, with a convolutional neural network (CNN) to postprocess convection-allowing model (CAM) forecasts. The CGANs are designed to create synthetic ensemble members from deterministic CAM forecasts, and their outputs are processed by the CNN to estimate the probability of severe weather. The method is tested using High-Resolution Rapid Refresh (HRRR) 1–24-h forecasts as inputs and Storm Prediction Center (SPC) severe weather reports as targets. The method produced skillful predictions with up to 20% Brier skill score (BSS) increases compared to other neural-network-based reference methods using a testing dataset of HRRR forecasts in 2021. For the evaluation of uncertainty quantification, the method is overconfident but produces meaningful ensemble spreads that can distinguish good and bad forecasts. The quality of CGAN outputs is also evaluated. Results show that the CGAN outputs behave similarly to a numerical ensemble; they preserved the intervariable correlations and the contribution of influential predictors as in the original HRRR forecasts. This work provides a novel approach to postprocess CAM output using neural networks that can be applied to severe weather prediction. Significance Statement We use a new machine learning (ML) technique to generate probabilistic forecasts of convective weather hazards, such as tornadoes and hailstorms, with the output from high-resolution numerical weather model forecasts. The new ML system generates an ensemble of synthetic forecast fields from a single forecast, which are then used to train ML models for convective hazard prediction. Using this ML-generated ensemble for training leads to improvements of 10%–20% in severe weather forecast skills compared to using other ML algorithms that use only output from the single forecast. This work is unique in that it explores the use of ML methods for producing synthetic forecasts of convective storm events and using these to train ML systems for high-impact convective weather prediction.

Uncertainty forecasting system for tropical cyclone tracks based on conformal prediction

Tropical cyclone tracks are one of the most critical factors in tropical cyclone forecasting, but there is an inherent uncertainty in their forecasts, but there are no relevant machine learning methods to carry out uncertainty studies. This study proposes an uncertainty forecasting system using machine learning models within a conformal forecasting framework, aiming to provide reliable forecast regions to improve the ability of decision makers to prepare for and respond to potential hazards. This study jointly examines the path forecasting performance of 10 major machine learning models and 10 major conformal forecasting methods through a comparative study. The research work models forecast timescales of 6, 12 and 24 h, and the study covers hurricanes from 1975 to 2021. The experimental results show that the deterministic forecast performance of the model is comparable to the skill of the operated benchmark model, demonstrating that the machine learning model possesses forecasting skill, while also providing tight uncertainty intervals about the path forecast. The method has high prediction accuracy and reliability and is expected to be widely used in the field of tropical cyclone track forecasting and risk communication in the future.

Slow Slip Events in New Zealand: Irregular, yet Predictable?

AbstractCurrent earthquake forecasting approaches are mainly based on probabilistic assumptions, as earthquakes seem to occur randomly. Such apparent randomness can however be caused by deterministic chaos, rendering deterministic short‐term forecasts possible. Due to the short historical and instrumental record of earthquakes, chaos detection has proven challenging, but more frequently occurring slow slip events (SSE) are promising candidates to probe for determinism. Here, we characterize the SSE signatures obtained from GNSS position time series in the Hikurangi Subduction Zone (New Zealand) to investigate whether the seemingly random SSE occurrence is governed by chaotic determinism. We find evidence for deterministic chaos for stations recording shallow SSEs, suggesting that short‐term deterministic forecasting of SSEs, similar to weather forecasts, might indeed be possible over timescales of a few weeks. We anticipate that our findings could open the door for next‐generation SSE forecasting, adding new tools to existing probabilistic approaches.

Spatiotemporal Convolutional Approach for the Short-Term Forecast of Hourly Heavy Rainfall Probability Integrating Numerical Weather Predictions and Surface Observations

Abstract The accurate prediction of short-term rainfall, and in particular the forecast of hourly heavy rainfall (HHR) probability, remains challenging for numerical weather prediction (NWP) models. Here, we introduce a deep learning (DL) model, PredRNNv2-AWS, a convolutional recurrent neural network designed for deterministic short-term rainfall forecasting. This model integrates surface rainfall observations and atmospheric variables simulated by the Precision Weather Analysis and Forecasting System (PWAFS). Our DL model produces realistic hourly rainfall forecasts for the next 13 h. Quantitative evaluations show that the use of surface rainfall observations as one of the predictors achieves higher performance (threat score) with 263% and 186% relative improvements over NWP simulations for the first 3 h and the entire forecast hours, respectively, at a threshold of 5 mm h−1. Noting that the optical-flow method also performs well in the initial hours, its predictions quickly worsen in the final hours compared to other experiments. The machine learning model, LightGBM, is then integrated to classify HHR from the predicted hourly rainfall of PredRNNv2-AWS. The results show that PredRNNv2-AWS can better reflect actual HHR conditions compared with PredRNNv2 and PWAFS. A representative case demonstrates the superiority of PredRNNv2-AWS in predicting the evolution of the rainy system, which substantially improves the accuracy of the HHR prediction. A test case involving the extreme flood event in Zhengzhou exemplifies the generalizability of our proposed model. Our model offers a reliable framework to predict target variables that can be obtained from numerical simulations and observations, e.g., visibility, wind power, solar energy, and air pollution.

Uncertainty-Based Capacity Factors of Operational Wind Turbines Using the Generalized Likelihood Uncertainty Estimation (GLUE) Method

This study proposes a novel framework that couples the general likelihood uncertainty estimation (GLUE) method with a deterministic forecasting approach to conduct a new uncertainty analysis approach for assessing the energy production of operational wind turbines installed in the Jhongtun wind farm at Penghu (an island in the middle of Taiwan Strait). The 10-year measured data of wind speeds and energy output collected on these wind turbines is divided into two 5-year data sets for the present analysis framework of execution and validation to demonstrate the predictability of the GLUE method. The present study considers 15 scenario testing cases with various time periods, i.e., twelve months, the strong-wind (October-March) regime, the weak-wind (April-September) regime, and one year, for the framework to investigate the applicability of the GLUE method on long-term wind energy forecasting. In the execution framework, the 5-year measured data is used by the GLUE method to access the uncertainties involved in the deterministic approach (i.e., the shape and scale parameters of the Weibull wind speed distribution (WWSD), the performance curve, and the capacity factor) with two confidence intervals of 50% and 90%. The framework is then validated by the measured capacity factors in the last 5-year data and compared with the results of the uncertainty analysis approach by the Monte Carlo (MC) approach to discover the applicability of the new uncertainty analysis approach. From the simulated results, it is found that the proposed uncertainty analysis approach provides predictions of confidence intervals that match the measured data better than the MC-based uncertainty analysis approach. Specifically, the proposed approach can match the measured capacity factors in all the simulated scenarios. Conversely, the MC-based approach is found to create narrow confidence intervals that cannot completely capture the measured capacity factors, particularly for the strong-wind, weak-wind, and one-year scenarios. Therefore, this novel uncertainty analysis approach is proven to be useful in predicting the uncertainties of wind energy production.

Probabilistic Forecast of Ecological Drought in Rivers Based on Numerical Weather Forecast from S2S Dataset

Ecological droughts in rivers, as a new type of drought, have been greatly discussed in the past decade. Although various studies have been conducted to identify and evaluate ecological droughts in rivers from different indices, a forecast model for this type of drought is still lacking. In this paper, a numerical weather forecast, a hydrological model, and a generalized Bayesian model are employed to establish a new general framework for the probabilistic forecasting of ecological droughts in rivers, and the Daitou section in China is selected as the study area to examine the performance of the new framework. The results show that the hydrological model can accurately simulate the monthly streamflow with a Nash–Sutcliffe efficiency of 0.91 in the validation period, which means that the model can be used to reconstruct the natural streamflow from the impact of an upstream reservoir. Based on a comparison of ecological drought events from the observed and model-simulated streamflow series, the events from the observed series have a larger deficit volume and a longer duration of ecological droughts after 2014, indicating that human activities may lead to a more severe situation of ecological droughts. Furthermore, due to the uncertainty of precipitation forecasts, a probabilistic precipitation forecast is employed for probabilistic ecological drought forecasting. Compared to the deterministic forecast, the probabilistic ecological drought forecast has better performance, with a Brier score decrease of 0.35 to 0.18 and can provide more information about the risk of ecological droughts. In general, the new probabilistic framework developed in this study can serve as a basis for the development of early-warning systems and countermeasures for ecological droughts.

A New Simheuristic Approach for Stochastic Runway Scheduling

We consider a stochastic, dynamic runway scheduling problem involving aircraft landings on a single runway. Sequencing decisions are made with knowledge of the estimated arrival times (ETAs) of all aircraft due to arrive at the airport, and these ETAs vary according to continuous-time stochastic processes. Time separations between consecutive runway landings are modeled via sequence-dependent Erlang distributions and are affected by weather conditions, which also evolve continuously over time. The resulting multistage optimization problem is intractable using exact methods, and we propose a novel simheuristic approach based on the application of methods analogous to variable neighborhood search in a high-dimensional stochastic environment. Our model is calibrated using flight tracking data for over 98,000 arrivals at Heathrow Airport. Results from numerical experiments indicate that our proposed simheuristic algorithm outperforms an alternative based on deterministic forecasts under a wide range of parameter values, with the largest benefits seen when the underlying stochastic processes become more volatile and also when the on-time requirements of individual flights are given greater weight in the objective function. Funding: This work was supported by the Engineering and Physical Sciences Research Council [Grant EP/M020258/1]. Supplemental Material: The online appendix and data files are available at https://doi.org/10.1287/trsc.2022.0400 .

Verification of the Global Forecast System, North American Mesoscale Forecast System, and High-Resolution Rapid Refresh Model Near-Surface Forecasts by Use of the New York State Mesonet

Abstract Global Forecast System (GFS), North American Mesoscale Forecast System (NAM), and High-Resolution Rapid Refresh (HRRR) 2-m temperature, 10-m wind speed, and precipitation accumulation forecasts initialized at 1200 UTC are verified against New York State Mesonet (NYSM) observations from 1 January 2018 through 31 December 2021. NYSM observations at 126 site locations are used to calculate standard error statistics (e.g., forecast error, root-mean-square error) for temperature and wind speed and contingency table statistics for precipitation across forecast hours, meteorological seasons, and regions. The majority of the focus is placed on the first 18 forecast hours to allow for comparison among all three models. A daily NYSM station-mean temperature error analysis identified a slight cold bias at temperatures below 25°C in the GFS, a cool-to-warm bias as forecast temperatures warm in the HRRR, and a warm bias at temperatures above 30°C in each model. Differences arise when considering temperature biases with respect to lead times and seasons. Wind speeds are overforecast at all ranges in each season, and forecast wind speeds ≥ 18 m s−1 are rarely observed. Performance diagrams indicate overall good forecast performance at precipitation thresholds of 0.1–1.5 mm, but with a high frequency bias in the GFS and NAM. This paper provides an overview of deterministic forecast performance across New York State, with the aim of sharing common biases associated with temperature, wind speed, and precipitation with operational forecasters and is the first step in developing a real-time model forecast uncertainty prediction tool.

Hour-Ahead Photovoltaic Power Prediction Combining BiLSTM and Bayesian Optimization Algorithm, with Bootstrap Resampling for Interval Predictions

Photovoltaic (PV) power prediction plays a critical role amid the accelerating adoption of renewable energy sources. This paper introduces a bidirectional long short-term memory (BiLSTM) deep learning (DL) model designed for forecasting photovoltaic power one hour ahead. The dataset under examination originates from a small PV installation located at the Polytechnic School of the University of Alcala. To improve the quality of historical data and optimize model performance, a robust data preprocessing algorithm is implemented. The BiLSTM model is synergistically combined with a Bayesian optimization algorithm (BOA) to fine-tune its primary hyperparameters, thereby enhancing its predictive efficacy. The performance of the proposed model is evaluated across diverse meteorological and seasonal conditions. In deterministic forecasting, the findings indicate its superiority over alternative models employed in this research domain, specifically a multilayer perceptron (MLP) neural network model and a random forest (RF) ensemble model. Compared with the MLP and RF reference models, the proposed model achieves reductions in the normalized mean absolute error (nMAE) of 75.03% and 77.01%, respectively, demonstrating its effectiveness in this type of prediction. Moreover, interval prediction utilizing the bootstrap resampling method is conducted, with the acquired prediction intervals carefully adjusted to meet the desired confidence levels, thereby enhancing the robustness and flexibility of the predictions.

Improving 3-day deterministic air pollution forecasts using machine learning algorithms

Abstract. As air pollution is regarded as the single largest environmental health risk in Europe it is important that communication to the public is up to date and accurate and provides means to avoid exposure to high air pollution levels. Long- and short-term exposure to outdoor air pollution is associated with increased risks of mortality and morbidity. Up-to-date information on present and coming days' air quality helps people avoid exposure during episodes with high levels of air pollution. Air quality forecasts can be based on deterministic dispersion modelling, but to be accurate this requires detailed information on future emissions, meteorological conditions and process-oriented dispersion modelling. In this paper, we apply different machine learning (ML) algorithms – random forest (RF), extreme gradient boosting (XGB), and long short-term memory (LSTM) – to improve 1, 2, and 3 d deterministic forecasts of PM10, NOx, and O3 at different sites in Greater Stockholm, Sweden. It is shown that the deterministic forecasts can be significantly improved using the ML models but that the degree of improvement of the deterministic forecasts depends more on pollutant and site than on what ML algorithm is applied. Also, four feature importance methods, namely the mean decrease in impurity (MDI) method, permutation method, gradient-based method, and Shapley additive explanations (SHAP) method, are utilized to identify significant features that are common and robust across all models and methods for a pollutant. Deterministic forecasts of PM10 are improved by the ML models through the input of lagged measurements and Julian day partly reflecting seasonal variations not properly parameterized in the deterministic forecasts. A systematic discrepancy by the deterministic forecasts in the diurnal cycle of NOx is removed by the ML models considering lagged measurements and calendar data like hour and weekday, reflecting the influence of local traffic emissions. For O3 at the urban background site, the local photochemistry is not properly accounted for by the relatively coarse Copernicus Atmosphere Monitoring Service ensemble model (CAMS) used here for forecasting O3 but is compensated for using the ML models by taking lagged measurements into account. Through multiple repetitions of the training process, the resulting ML models achieved improvements for all sites and pollutants. For NOx at street canyon sites, mean squared error (MSE) decreased by up to 60 %, and seven metrics, such as R2 and mean absolute percentage error (MAPE), exhibited consistent results. The prediction of PM10 is improved significantly at the urban background site, whereas the ML models at street sites have difficulty capturing more information. The prediction accuracy of O3 also modestly increased, with differences between metrics. Further work is needed to reduce deviations between model results and measurements for short periods with relatively high concentrations (peaks) at the street canyon sites. Such peaks can be due to a combination of non-typical emissions and unfavourable meteorological conditions, which are rather difficult to forecast. Furthermore, we show that general models trained using data from selected street sites can improve the deterministic forecasts of NOx at the station not involved in model training. For PM10 this was only possible using more complex LSTM models. An important aspect to consider when choosing ML algorithms is the computational requirements for training the models in the deployment of the system. Tree-based models (RF and XGB) require fewer computational resources and yield comparable performance in comparison to LSTM. Therefore, tree-based models are now implemented operationally in the forecasts of air pollution and health risks in Stockholm. Nevertheless, there is big potential to develop generic models using advanced ML to take into account not only local temporal variation but also spatial variation at different stations.