Forecast Data Research Articles

Predictability of Severe Convective Storm Environments in Global Ensemble Forecast System, Version 12, Reforecasts

Abstract Prediction of severe convective storms at time scales of 2–4 weeks is of interest to forecasters and stakeholders due to their impacts on life and property. Prediction of severe convective storms on this time scale is challenging, since the large-scale weather patterns that drive this activity begin to lose dynamic predictability beyond week 1. Previous work related to severe convective storms on the subseasonal time scale has mostly focused on observed relationships with teleconnections. The skill of numerical weather prediction forecasts of convective-related variables has been comparatively less explored. In this study over the United States, a forecast evaluation of variables relevant in the prediction of severe convective storms is conducted using Global Ensemble Forecast System, version 12, reforecasts at lead times of up to 4 weeks. We find that kinematic and thermodynamic fields are predicted with skill out to week 3 in some cases, while composite parameters struggle to achieve meaningful skill into week 2. Additionally, using a novel method of weekly summations of daily maximum composite parameters, we suggest that the aggregation of certain variables may assist in providing additional predictability beyond week 1. These results should serve as a reference for forecast skill for the relevant fields and help inform the development of convective forecasting tools at time scales beyond current operational products. Significance Statement Prediction of severe weather beyond 1 week in advance is of interest to stakeholders given their impacts on life and property. In this study, we evaluate 20 years of forecast data generated by a numerical model ensemble to determine whether variables relevant in severe weather forecasts can be predicted in weeks 2–4. The variables with the best skill measures generally represent large-scale weather patterns that are more predictable on longer time scales, although some manipulation of other severe weather parameters yielded additional results. We suggest that the results found in this study can inform future work that assesses the predictability of severe weather in weeks 2–4 using more complex methods such as machine learning.

Development and Applicability Assessment of an Hourly Water Level Prediction Model for Agricultural Reservoirs Using Automated Machine Learning

Objectives:This study aims to develop and evaluate hourly water level prediction models for agricultural reservoirs using an automated machine learning (AutoML) approach, specifically, employing TPOT.Methods:The study focuses on the Baekrok and Baekryeon reservoirs using rainfall forecast data from the Korea Meteorological Administration, along with observed rainfall and reservoir water level data. TPOT, which utilizes genetic algorithms to automate the generation of optimal machine learning pipelines, was used to build models. Additionally, Random Forest (RF) was implemented as a benchmark to evaluate TPOT’s applicability. Predictions were generated with lead times of 1, 3, 6, and 12 hours, and model accuracy was evaluated using the Nash-Sutcliffe Efficiency (NSE), coefficient of determination (R2), and root-mean-square error (RMSE).Results and Discussion:The predictions from both TPOT and RF showed high accuracy for shorter lead times, with NSE values exceeding 0.99 for the 1-hour predictions. However, predictive accuracy decreased as lead time increased, likely due to greater uncertainty in rainfall forecast data, particularly for the 12-hour predictions. Despite this trend, TPOT-derived models maintained more stable and accurate performance compared to RF models across all lead times.Conclusion:This study demonstrates the applicability of TPOT-based AutoML techniques in predicting hourly water levels in agricultural reservoirs. Future work should explore strategies to mitigate the decline in accuracy for longer lead times.

Integrating Thermal Infrared Imaging and Weather Data for Short-Term Prediction of Building Envelope Thermal Appearance

This study presents a novel deep-learning framework for predicting the thermal appearance of building envelopes under varying weather conditions based on a new dataset collected using a thermal infrared camera at 10 min intervals over a one-and-a-half-year period. Unlike existing studies that rely on simulated data or physical models that do not always accurately reflect the complex heat transfer processes in real buildings, we have collected a large dataset showing how a building behaves under different climatic conditions. We propose a novel deep-learning approach that integrates weather data and thermal imagery to predict the temperature distribution on the building façade for the next 24 and 48 h. The model uses a state-of-the-art recurrent neural network architecture, PredRNN V2, with an action conditioning mechanism to incorporate weather forecasting data into the prediction process. We evaluate this approach in terms of average accuracy, prediction accuracy in specific regions, and visual-perceptual performance of the images. The proposed framework achieves a prediction accuracy of 1.5 °C (root mean square error—RMSE) for the 24 h prediction and 2.04 °C (RMSE) for the 48 h prediction, outperforming baseline models in terms of temperature prediction accuracy and structural similarity of the predicted images.

PERAMALAN HARGA GULA PASIR MENGGUNAKAN VARIASI KALENDER REGARIMA DENGAN MOVING HOLIDAY EFFECT (PERIODE JANUARI 2018 SAMPAI DENGAN DESEMBER 2022 DI PASAR KOTA SEMARANG)

National sugar production is decreasing day by day with the pattern of people's consumption of sugar continuing to increase, thus encouraging the government to import sugar from other countries to meet food needs. This is because the import of granulated sugar causes the price of local granulated sugar to fluctuate which can rise up to 2 times the highest retail price. Eid al-Fitr is determined based on the Islamic calendar, this will cause a shift in the date each year on the Maseh calendar, the date of the celebration of Eid al-Fitr which moves from year to year is known as the "moving holiday effect". One of the calendar variation models used to eliminate the Holiday Effect Transfer and has a simple processing flow is the RegARIMA model. The RegARIMA model method is a modeling technique that combines the ARIMA model with the regression model. In the regression model, the weight matrix is used as the independent variable and the price of granulated sugar is used as the dependent variable. The weight value is obtained based on the number of days that affect Eid, which is 14 days. Based on an analysis of the price of granulated sugar in the Semarang City market for the period January 2018 to December 2022, the best model is obtained, namely SARIMA (1,0,0)12 with the smallest AIC value and the sMAPE value obtained from forecasting data for 2021 is of 19.04%, which means that the forecasting is still at a reasonable level.

A novel modelling framework for real-time prediction of path travel times using both traffic and weather data

The problems of path travel time predictions have been studied over decades. Traditional approaches to this challenge have relied on different traffic data. However, in metropolitan cities with frequent rainfall, there is a need to consider the weather effects on real-time prediction of path travel times. Hence, this research integrates traffic data with weather data, including rainfall intensity data and weather forecast data to consider the temporal relationships between weather data and path travel times. The proposed modelling framework is evaluated with data from a major expressway and an urban arterial road in Hong Kong. Results unequivocally demonstrate that incorporating weather data significantly boosts prediction accuracy. This study also examines the impact of different frequencies on weather data, as well as weather forecast correctness, on prediction accuracy. Finally, the applicability of the proposed modelling framework has been verified without and with the input of ground truth on path travel times.

A Novel WTG Method for Predicting Ship Trajectories in the Fujian Inshore Area Based on AIS Data

The increasing congestion in major global maritime routes poses significant threats to international maritime safety, exacerbated by the proliferation of large, high-speed vessels. To improve the detection of abnormal ship behavior, this research employed automatic identification system (AIS) data for ship trajectory forecasting. Traditional methods primarily focus on spatial and temporal correlations but often lack accuracy and reliability. In this study, ship path predictions were enhanced using the WTG model, which combines wavelet transform, temporal convolutional networks (TCN), and gated recurrent units (GRU). Initially, wavelet decomposition was applied to deconstruct the input trajectory time series. The TCN and GRU modules then extracted features from both the time series and the decomposed data. The predicted elements were reassembled using a multi-head attention mechanism and a pooling layer to produce the final predictions. Comparative experiments demonstrated that the WTG model surpasses other models in the accuracy of ship trajectory prediction. The model proposed in this study proves to be reliable for forecasting ship paths, which is crucial for marine traffic management and ensuring safe navigation.

Observation impact explanation in atmospheric state estimation using hierarchical message-passing graph neural networks

Abstract The impact of meteorological observations on weather forecasting varies with the sensor type, location, time, and other environmental factors. Thus, the quantitative analysis of observation impacts is crucial for the effective and efficient development of weather forecasting systems. However, existing impact analysis methods are dependent on specific forecast systems, because system-specific adjoint models are used and the sensitivity of the observation to the forecast is measured. This study investigates the impact of observations on atmospheric state estimation in weather forecasting systems by developing a novel graph neural network (GNN) model specialized for analyzing the heterogeneous relations between observations and atmospheric states. The observation impact can then be assessed by applying explainable methods to the proposed GNN model, which is independent of forecasting systems. Further, we develop a novel application called `CloudNine,' a system that provides impact analysis for individual observations with visualization. Our GNN model comprises hierarchical message-passing modules that separately analyze spatial correlations between observations at close locations and atmospheric states at close locations and then examine correlations between observations and atmospheric states. To consider the different factors influencing these correlations, we utilized geo-coordinates and types of observations in the attention mechanism of the modules with their feature vectors. We then applied gradient-based explainability methods to quantify the significance of the different observations in the estimation. Evaluated using data from 11 satellites and land-based observations, the results highlight the effectiveness of the proposed model and the visualization of observation impacts, enhancing the understanding and optimization of observational data in weather forecasting.

Sensitivity of Dust Deposition for Parabolic Trough Collector Mirrors to Different Meteorological Drivers

This abstract presents a soiling forecasting tool (SFT) for assessing the deposition of dust on parabolic trough collector (PTC) mirrors and its cumulative impact on their reflectivity. The SFT was developed by the University of Patras in the frame of the Smart Solar System (S3) project (Horizon 2020 Solar-Era.net). An initial version of the model was presented at the SolarPACES 2022 conference, where the adaptation of the model to dust transport phenomena was demonstrated. The subsequent step in the evolution of the algorithm concerned the calibration of the SFT under different atmospheric conditions. In addition, the model was modified to incorporate meteorological and aerosol forecast data as inputs. The atmospheric predictions of dust and aerosol optical depth are from the global atmospheric forecasts of the Copernicus Atmosphere Monitoring Service. Regarding the meteorological data, three independent sources are used. Those are the Norwegian Meteorological Institute's meteorological forecasts (YR), the METAR meteorological observations from the closest airport, and lastly, the data from the automatic weather station at the location of the PTC system at the KEAN Soft Drinks Ltd factory in Limassol, Cyprus. The aim of this paper is to assess the impact of the three distinct meteorological input sources on the modelled reflectivity and its comparison to measurements taken during the experimental campaign.

HIGH-DIMENSIONAL FORECASTING WITH KNOWN KNOWNS AND KNOWN UNKNOWNS

Abstract Forecasts play a central role in decision-making under uncertainty. After a brief review of the general issues, this article considers ways of using high-dimensional data in forecasting. We consider selecting variables from a known active set, known knowns, using Lasso and One Covariate at a time Multiple Testing, and approximating unobserved latent factors, known unknowns, by various means. This combines both sparse and dense approaches to forecasting. We demonstrate the various issues involved in variable selection in a high-dimensional setting with an application to forecasting UK inflation at different horizons over the period 2020q1–2023q1. This application shows both the power of parsimonious models and the importance of allowing for global variables.

Development of a DNI Forecast Tool Based on Machine Learning for the Smart Operation of a Parabolic Trough Collector System with Thermal Energy Storage

In the research project Smart Solar System (S3), the Solar-Institut Jülich (SIJ) developed forecast tools to predict the direct normal irradiance (DNI) in hourly resolution for the current day. The aim of the daily DNI forecast is to use it as input to enable a smart operation of a parabolic trough collector (PTC) system with a concrete thermal energy storage (C-TES) located at the company KEAN Soft Drinks Ltd in Limassol, Cyprus. The main focus in this work is on the development of a DNI forecast tool based on long short-term memory (LSTM), which is a recurrent neural network (RNN), and its comparison with a DNI forecast tool based on analytical algorithms (non-machine learning). Only the non-machine learning DNI forecast tool with hourly update was tested in real-life PTC plant operation since end of 2022. The comparisons between the three DNI forecast tools show that the potential of using machine learning is very high. Different comparisons were made including an evaluation of the accuracy of the tool (i.e. comparison of the DNI forecast data with DNI measurement data). The DNI forecast based on the LSTM network proved to be more accurate than the non-machine learning DNI forecasts when considering errors greater than ±100 W/m2. The error of the LSTM network compared to DNI measurement data was as follows: 43.6 % of the data was within ±100 W/m2, 74.8 % was within ±200 W/m2, 89.8 % was within ±300 W/m2 and 95.8 % was within ±400 W/m2.

The effect of visibility on forecast and inventory management performance during the COVID-19 pandemic

During the COVID-19 pandemic, healthcare facilities faced significant shortages of critical supplies like personal protective equipment, with dire repercussions. This study evaluates the potential role of decreased data visibility on these shortages, analysing different forecasting methods integrated with a periodic review inventory system (i.e. a base stock policy) on semi-simulated data encompassing several visibility issues. The forecasting methods chosen pertain to different data types: Holt and naïve methods are used as demand-based predictors, while a modified epidemiological model utilises pandemic data for demand forecasting. We scrutinise three prevalent data visibility issues through specifically crafted scenarios examining the effects of delayed, temporally aggregated, and erroneous data on system performance. Generally, our research illustrates that data visibility issues have a detrimental impact on the healthcare supply chain's efficacy. Interestingly, the system performance sees an uptick when these issues spur significantly oversized over-forecasts, e.g. when employing an epidemiological compartmental model for predictions.

Cyanobacterial blooms prediction in China’s large hypereutrophic lakes based on MODIS observations and Bayesian theory

Cyanobacterial harmful algal blooms (HABs) pose a significant threat to aquatic ecosystems, water quality, and public health, particularly in large hypereutrophic lakes. Developing accurate short-term prediction models is essential for early warning and effective management of HABs. This study introduces a Bayesian-based model aimed at predicting HABs in three of China’s large hypereutrophic lakes: Lake Taihu, Lake Chaohu, and Lake Hulunhu. By integrating MODIS data from the Terra and Aqua satellites with meteorological data spanning from 2010 to 2018, the model forecasts HABs distributions 1, 4, and 7 days in advance. Validation with meteorological data from 2019 to 2020 showed high accuracy, with 0.83 at the pixel level, 0.74 for zonal predictions, and 0.64 for lake-wide HABs area forecasts. Further evaluation using 2023 weather forecast data yielded similar accuracies of 0.78, 0.57, and 0.62, respectively. In addition to predicting the spatial extent of HABs, the model provides binary HABs maps, outbreak areas, and HABs status within lake zones. This method for building prediction models significantly enhances early warning and management capabilities for HABs, providing a scalable framework that can be adapted to other regions facing similar threats from HABs.

Optimization of temperature prediction algorithms and simulation of future neighborhood-scale thermal environments in Chongqing

With climate change and urbanization, predicting the outdoor thermal environment of residential areas has become increasingly crucial, particularly during extreme weather conditions. Existing studies lack a comprehensive approach to optimizing temperature prediction algorithms for Chongqing and simulating extreme thermal environments at the neighborhood scale for future decades. This study employed various algorithms, including Levenberg-Marquardt optimization, Monte Carlo simulation, ARIMA modeling, and SARIMA prediction, to simulate and forecast temperature data for the central urban area of Chongqing from 2014 to 2023. By assessing metrics such as root mean square error (RMSE), R-squared (R2), and mean absolute percentage error (MAPE), the most effective prediction algorithm was identified. Results indicate that the Levenberg-Marquardt optimization algorithm, known for its nonlinear curve fitting capabilities, demonstrated superior performance in both the test and training sets, achieving an RMSE of 0.82 °C, MAPE of 1.81 %, and R2 of 0.75. This algorithm was subsequently used to forecast the outdoor temperature trend in Chongqing from 2024 to 2050, while also simulating the outdoor wind field, temperature distribution, building surface temperatures, and wind speeds in a residential district at neighborhood scale for 2030, 2040, and 2050 at a building density of 40 %. Findings reveal that peak pedestrian-level temperatures could reach 45.62 °C, and peak roof surface temperatures could reach 93.93 °C under extreme conditions. Furthermore, the study proposes practical recommendations for building layouts and cooling strategies. These findings are expected to improve residential planning in Chongqing, enhance residents' quality of life, and provide insights for other regions facing future temperature changes.

Station keeping control method based on deep reinforcement learning for stratospheric aerostat in dynamic wind field

Aiming at the problem of long-endurance station keeping of stratospheric aerostats under energy constraints in dynamic wind environment, a stratospheric aerostat station keeping controller was designed using the D3QN (Dueling Double Deep Q Network) algorithm of deep reinforcement learning in this article. The performance of the station keeping controllers are evaluated by the average station keeping radius and station keeping effective time ratio. Meanwhile, the effects of different reward functions on the performance of station keeping controller are studied. The simulation results of typical scenarios with European Centre for Medium-Range Weather Forecasts (ECMWF) data show that: under the task constraints of three days of station keeping time and 50 km of station keeping radius, the average station keeping radius of the stratospheric aerostat without power propulsion is 28.25 km, the station keeping time ratio is 90.56%, and the State of Charge (SOC) of the battery is more than 50%. Under the condition of propulsion, the control effect of the average station keeping radius of stratospheric aerostat is significantly improved. The control effect of adding dynamic propulsion in the north-south direction is better than that in the east-west direction, which is close to the dual-channel control effect. At the same time, this study verifies that the station keeping controller based on deep reinforcement learning has strong robustness, and demonstrates that the station keeping controller can be designed with different reward functions for the requirements of different station keeping task.

Ultra-short term wind power prediction method based on LightGBM and transformer

Abstract Accurate ultra-short-term wind power prediction is very important for the safe and stable operation of the power system. At present, most wind power forecasting systems use numerical weather forecast data and the data collected by the Supervisory Control And Data Acquisition (SCADA) system. However, these data have the characteristics of high data dimension, feature redundancy, and feature correlation. Therefore, in the process of ultra-short power prediction, it is necessary to select features with high power correlation and low cross-correlation between features. In this paper, a feature selection method for wind power prediction based on LightGBM is proposed to calculate the importance of all features and determine the preliminary selected features. Then, the maximum information coefficient is used to construct the correlation discriminant matrix, according to which the features with similar importance in a single screening are evaluated, and the input features with high similarity are discarded. In addition, in order to improve the accuracy of the system’s wind power prediction, this paper uses the effective features extracted by LightGBM to design an ultra-short-term power prediction method based on Transformer. The results show that the RMSE of ultra-short-term wind power prediction is 7.3%.

Speed and Route Planning in Evolving Weather for Efficient Voyages

Abstract Fuel-saving potential by optimizing speed as well as route has been studied in a wide variety of operating conditions on a trans-Pacific route. Analysis using different seasonal weather patterns and various ship speeds allows a more robust estimation of possible fuel savings. Up to 18% reduction in fuel consumption was observed on eastbound voyages and up to 13% savings could be obtained on westbound voyages. It was observed that higher energy savings can be obtained for longer voyage times. Because the share of added wave resistance in total resistance is higher for lower speeds and there is greater flexibility to speed up and slow down during the voyage. A detailed comparison between the hindcast and forecast of metocean data has been presented since actual fuel savings would depend on the accuracy of the weather forecast. For future work on operational optimization, it is suggested to use weather forecast data instead of hindcast for realistic estimation of fuel saving potential.

A Deep–Learning Network for Wheat Yield Prediction Combining Weather Forecasts and Remote Sensing Data

Accurately predicting winter wheat yield before harvest could greatly benefit decision-makers when making management decisions. In this study, we utilized weather forecast (WF) data combined with Sentinel-2 data to establish the deep-learning network and achieved an in-season county-scale wheat yield prediction in China’s main wheat-producing areas. We tested a combination of short-term WF data from the China Meteorological Administration to predict in-season yield at different forecast lengths. The results showed that explicitly incorporating WF data can improve the accuracy in crop yield predictions [Root Mean Square Error (RMSE) = 0.517 t/ha] compared to using only remote sensing data (RMSE = 0.624 t/ha). After comparing a series of WF data with different time series lengths, we found that adding 25 days of WF data can achieve the highest yield prediction accuracy. Specifically, the highest accuracy (RMSE = 0.496 t/ha) is achieved when predictions are made on Day of The Year (DOY) 215 (40 days before harvest). Our study established a deep-learning model which can be used for early yield prediction at the county level, and we have proved that weather forecast data can also be applied in data-driven deep-learning yield prediction tasks.

Forecasting material flows of modern logistics centres of railway freight stations exemplifi ed by Xi’an (the People’s Republic of China)

The purpose of the study is to create a more accurate material fl ow forecasting model of Xi’an freight railway station in China. The combined forecasting model is more validated for forecasting freight flows of regional logistics compared to three methods: grey forecasting, Markov chains, entropy weighting. Through the creation of the combined model, the grey forecasting method is combined with Markov chain correction, and the projected data is compared with the actual data to obtain higher accuracy of the forecasting model. A combined model using the grey forecasting method combined with Markov chain correction is created, with the forecast data compared with the actual data to obtain high accuracy of the forecasting model. The practical significance is that in the context of the present post-pandemic economic development, the logistics enterprises that do not operate in accordance with the modern logistics methods may be displaced by competitors. If the railway does not improve its logistics infrastructure, logistics equipment, railway logistics network platform, etc., it will lose out to other modes of transport. In order to meet the needs of logistics and improve the market competitiveness, the main indicator of a freight station is loading and logistics flow. Therefore, exact prediction of future changes in the logistics flow of a freight station can help to determine whether the station needs to be upgraded as a railway station or transformed into a certain type of a logistics centre.

Sensitivity of horizontal resolution and land surface model in operational WRF forecast for Online Nuclear Emergency Response System (ONERS)

Accurate Meteorological forecasts are crucial for the assessment of plume dispersion and dose prediction in nuclear power plant (NPP) sites. In this work the forecast sensitivity of the Weather Research and Forecasting (WRF) model is tested by running a series of forecast simulations for horizontal resolution, and land surface models (LSM) in the context of Online Nuclear Emergency Response System (ONERS) for Indian NPP sites. 72 h forecast simulations are made for three seasons viz. summer, southeast and northeast monsoon using the Global Forecast data. Three simulation experiments, namely 2 km-NOAH, 3 km-NOAH and 3 km-NOAHMP are conducted using two different nested domain configurations (18–6–2 km and 9–3 km) and two LSM schemes (NOAH and NOAH-MP) and tested at four different NPP sites. Forecast comparison of surface winds, relative humidity, temperature, heat fluxes and planetary boundary layer heights with data from meteorological tower, radiosonde and the Modern-Era Retrospective analysis for Research and Applications 2 (MERRA-2) shows 3 km-NOAH is equally capable in predicting surface parameters as well as vertical profiles compared to 2 km-NOAH with marginal differences. 3 km-NOAHMP shows less mean bias and better correlation for boundary layer height and heat fluxes. Comparison of spatial flow-field with 5th generation European Centre for Medium-Range Weather Forecasts (ECMWF) reanalysis (ERA5) data shows synoptic scale seasonal winds, sea level pressure systems and temperature hot-spots are better captured by 3 km-NOAHMP compared to 6 km coarse domain in the 18–6–2 km configuration. The daily accumulated rainfall by all simulations is overestimated compared to ERA5 data. The predictions by 3 km-NOAHMP better agree with Integrated Multi-satellite Retrievals for Global Precipitation Measurement (GPM-IMERGE) data whereas 2 km-NOAH predicts delayed rainfall occurrence. Dispersion simulations of hypothetical plume release from a coastal NPP site with all three forecasts properly show the influence of local scale diurnal land-sea breeze and seasonal winds on the plume movement. Therefore the 9–3 km domain with NOAHMP LSM is found to be a suitable choice for operational weather forecast in ONERS for Indian NPP sites.

In-season maize yield prediction in Northeast China: The phase-dependent benefits of assimilating climate forecast and satellite observations

Various yield forecasting methods have been reported in literature, but the benefits of assimilating seasonal climate forecasts and satellite observations for in-season yield forecasting during different growth stages have rarely been examined using machine learning. By synthesizing census yields, seasonal climate forecasts (SCF) and satellite-based gross primary production (GPP), this study develops a machine learning (i.e., Random Forest)–based in-season maize yield forecasting model for Northeast China, which produces over 40 % of China's total maize production. Based on the dynamically trained model, 11 numerical experiments are conducted to investigate the benefits of assimilating SCF and GPP relative to the experiments without using SCF and GPP for yield prediction in four forecast phases (planting–three leaves, three leaves–jointing, jointing–tasseling, tasseling–milk). Generally, our yield forecasting model exhibits a promising skill with a low bias of 4.11 %−4.97 %, when the observed climate and GPP are used as inputs. When climate forecasts are assimilated into the model by sampling historical climate, satisfactory yield forecasting can be achieved one month before harvest with a bias of 5.50 %−5.63 %. Bias-correcting climate forecast data from dynamical weather forecast models has a larger benefit for yield prediction with a lower bias of 4.77 %−5.06 %. Furthermore, we found a better benefit of assimilating SCF when compared with GPP during the first three forecast phases, although its relative importance decreases substantially towards harvest. Finally, phase-dependent maps indicating the best model are produced for each county, with historical resampling methods performing best in 40 %, 32 %, 27 %, and 21 % of counties from Phase 1 to Phase 4 and dynamical weather forecast models showing greater predictability in 54 %, 58 %, 58 %, and 52 % of counties during Phases 1–4, respectively. This study provides a useful yield forecasting framework for the breadbasket of China, which can be extended to other crops.