Boosting Arctic Air Quality Forecasts with Deep Learning

The golden rule of creating a machine learning algorithm is to handle as much data as possible under consistent conditions. In the field of air quality, these models can be used for short-term air quality forecasting, helping predict high pollution peaks and alerting residents in cases where pollutant concentrations exceed legal or World Health Organization (WHO) limits.For daily air quality predictions and short-term forecasts, the Slovenian Environment Agency uses the CAMx dispersion model, which is frequently upgraded with new input data. Based on the model results and expert opinions, we inform the public about air quality levels and issue alerts when concentrations reach the limit values. The Copernicus Atmosphere Monitoring Service (CAMS) has developed Model Output Statistics (MOS), which downscale air quality forecasts produced by regional ensemble models and incorporate measurement values from observation sites across Europe.In our study, we evaluated and compared CAMS MOS forecasts with forecasts from the Slovenian operational CAMx model and raw CAMS ensemble forecasts at Slovenian observational sites. The evaluation was conducted using time series plots, scatter plots, and Taylor diagrams, alongside statistical metrics such as the fraction of forecasts within a factor of two (FAC2), mean bias (MB), mean gross error (MGE), normalized mean bias (NMB), normalized gross error (NMGE), root mean square error (RMSE), correlation coefficient (r), coefficient of efficiency (COE), and index of agreement (IOA).Our findings indicate that CAMS MOS forecasts are almost always more accurate than raw CAMS or CAMx model predictions, particularly for the MOS day-1 forecast. However, for PM10 and PM2.5, longer-term MOS forecasts (day-2, day-3, and day-4) were less accurate than RAW and/or CAMx predictions based on certain statistical measures. When analysing individual stations, we observed occasional instances where MOS forecasts were less accurate for specific pollutants.CAMS MOS is a useful tool for improving the national alert system, helping to warn residents about poor air quality conditions and enabling timely protective measures. Consequently, our results suggest that the MOS day-1 forecast can serve as a valuable additional tool for issuing daily PM10 forecasts and, for some stations, even O3 daily maxima predictions. However, all MOS forecasts should be used with an awareness of their limitations—for example, the time lag in predicting high PM10 episodes and the systematic underestimation of O3 daily maxima at many Slovenian stations.

Error characteristics of model output statistics (MOS) temperature forecasts are calculated for over 200 locations around the continental United States. The forecasts are verified on a station-by-station basis for the year 2001. Error measures used include mean algebraic error (bias), mean absolute error (MAE), relative frequency of occurrence of bias and MAE values, and the daily forecast errors themselves. A case study examining the spatial and temporal evolution of MOS errors is also presented. The error characteristics presented here, together with the case study, provide a more detailed evaluation of MOS performance than may be obtained from regionally averaged error statistics. Knowledge concerning locations where MOS forecasts have large errors or biases and why those errors or biases exist is of great value to operational forecasters. Not only does such knowledge help improve their forecasts, but forecaster performance is often compared to MOS predictions. Examples of biases in MOS forecast errors are illustrated by examining two stations in detail. Significant warm and cold biases are found in maximum temperature forecasts for Los Angeles, California (LAX), and minimum temperature forecasts for Las Vegas, Nevada (LAS), respectively. MAE values for MOS temperature predictions calculated in this study suggest that coastal stations tend to have lower MAE values and lower variability in their errors, while forecasts with high MAE and error variability are more frequent in the interior of the United States. Therefore, MAE values from samples of MOS forecasts are directly proportional to the variance in the observations. Additionally, it is found that daily maximum temperature forecast errors exhibit less variability during the summer months than they do over the rest of the year, and that forecasts for any one station rarely follow a consistent temporal pattern for more than two or three consecutive days. These inconsistent error patterns indicate that forecasting temperatures based on recent trends in MOS forecast errors at an individual station is usually not a good strategy. As shown in earlier studies by other authors and demonstrated again here, MOS temperature forecasts are often inaccurate in the vicinity of strong temperature gradients, for locations affected by shallow cold air masses, or for stations in regions of anomalously warm or cold temperatures. Finally, a case study is presented examining the spatial and temporal distributions of MOS temperature forecast errors across the United States from 13 to 15 February 2001. During this period, two surges of cold arctic air moved south into the United States. In contrast to error trends at individual stations, nationwide spatial and temporal patterns of MOS forecast errors could prove to be a powerful forecasting tool. Nationwide plots of errors in MOS forecasts would be useful if made available in real time to operational forecasters.

<strong class="journal-contentHeaderColor">Abstract.</strong> Air quality (AQ) forecasting systems are usually built upon physics-based numerical models that are affected by a number of uncertainty sources. In order to reduce forecast errors, first and foremost the bias, they are often coupled with model output statistics (MOS) modules. MOS methods are statistical techniques used to correct raw forecasts at surface monitoring station locations, where AQ observations are available. In this study, we investigate the extent to which AQ forecasts can be improved using a variety of MOS methods, including moving average, quantile mapping, Kalman filter, analogs and gradient boosting machine methods, and consider as well the persistence method as a reference. We apply our analysis to the Copernicus Atmospheric Monitoring Service (CAMS) regional ensemble median O<span class="inline-formula">₃</span> forecasts over the Iberian Peninsula during 2018–2019. A key aspect of our study is the evaluation, which is performed using a comprehensive set of continuous and categorical metrics at various timescales, along different lead times and using different meteorological input datasets. Our results show that O<span class="inline-formula">₃</span> forecasts can be substantially improved using such MOS corrections and that improvements go well beyond the correction of the systematic bias. Depending on the timescale and lead time, root mean square errors decreased from 20 %–40 % to 10 %–30 %, while Pearson correlation coefficients increased from 0.7–0.8 to 0.8–0.9. Although the improvement typically affects all lead times, some MOS methods appear more adversely impacted by the lead time. The MOS methods relying on meteorological data were found to provide relatively similar performance with two different meteorological inputs. Importantly, our results also clearly show the trade-offs between continuous and categorical skills and their dependencies on the MOS method. The most sophisticated MOS methods better reproduce O<span class="inline-formula">₃</span> mixing ratios overall, with the lowest errors and highest correlations. However, they are not necessarily the best in predicting the peak O<span class="inline-formula">₃</span> episodes, for which simpler MOS methods can achieve better results. Although the complex impact of MOS methods on the distribution of and variability in raw forecasts can only be comprehended through an extended set of complementary statistical metrics, our study shows that optimally implementing MOS in AQ forecast systems crucially requires selecting the appropriate skill score to be optimized for the forecast application of interest.

Model Output Statistics (MOS) approaches relying on machine learning algorithms were applied to downscale regional air quality forecasts produced by CAMS (Copernicus Atmosphere Monitoring Service) at hundreds of monitoring sites across Europe. Besides the CAMS forecast, the predictors in the MOS typically include meteorological variables but also ancillary data. We explored first a “local” approach where specific models are trained at each site. An alternative “global” approach where a single model is trained with data from the whole geographical domain was also investigated. In both cases, local predictors are used for a given station in predictive mode. Because of its global nature, the latter approach can capture a variety of meteorological situation within a very short training period and is thereby more suited to cope with operational constraints in relation with the training of the MOS (frequent upgrades of the modelling system, addition of new monitoring sites). Both approaches have been implemented using a variety of machine learning algorithms: random forest, gradient boosting, standard and regularized multi-linear models. The quality of the MOS predictions is evaluated in this work for four key pollutants, namely particulate matter PM10 and PM2.5, ozone O3 and nitrogen dioxide NO2, according to scores based on the predictive errors and on the detection of pollution peaks (exceedances of the regulatory thresholds). Both the local and the global approaches significantly improve the performances of the raw Ensemble forecast. The most important result of this study is that the global approach competes with and can even outperform the local approach in some cases. This global approach gives the best RMSE scores when relying on a random forest model, for the prediction of daily mean, daily max and hourly concentrations. By contrast, it is the gradient boosting model which is better suited for the detection of exceedances of the European Union regulated threshold values for O3 and PM10.

<strong class="journal-contentHeaderColor">Abstract.</strong> Model Output Statistics (MOS) approaches relying on machine learning algorithms were applied to downscale regional air quality forecasts produced by CAMS (Copernicus Atmosphere Monitoring Service) at hundreds of monitoring sites across Europe. Besides the CAMS forecast, the predictors in the MOS typically include meteorological variables but also ancillary data. We explored first a &ldquo;local&rdquo; approach where specific models are trained at each site. An alternative &ldquo;global&rdquo; approach where a single model is trained with data from the whole geographical domain was also investigated. In both cases, local predictors are used for a given station in predictive mode. Because of its global nature, the latter approach can capture a variety of meteorological situation within a very short training period and is thereby more suited to cope with operational constraints in relation with the training of the MOS (frequent upgrades of the modelling system, addition of new monitoring sites). Both approaches have been implemented using a variety of machine learning algorithms: random forest, gradient boosting, standard and regularized multi-linear models. The quality of the MOS predictions is evaluated in this work for four key pollutants, namely particulate matter PM₁₀ and PM_2.5, ozone O₃ and nitrogen dioxide NO₂, according to scores based on the predictive errors and on the detection of pollution peaks (exceedances of the regulatory thresholds). Both the local and the global approaches significantly improve the performances of the raw Ensemble forecast. The most important result of this study is that the global approach competes with and can even outperform the local approach in some cases. This global approach gives the best RMSE scores when relying on a random forest model, for the prediction of daily mean, daily max and hourly concentrations. By contrast, it is the gradient boosting model which is better suited for the detection of exceedances of the European Union regulated threshold values for O₃ and PM_10.

Reliable aerosol optical depth (AOD) data with high spatial and temporal resolutions are needed for research on air pollution in China. AOD products from the Advanced Himawari Imager (AHI) onboard the geostationary Himawari-8 satellite and reanalysis datasets make it possible to capture diurnal variations of aerosol loadings. However, due to the different retrieval methods, their applicability may vary with different space and time. Thus, in this study, taking the measured AOD at the Aerosol Robotic NETwork (AERONET) stations as the gold standard, the performance of the latest AHI hourly AOD product (i.e., L3 AOD) was evaluated and then compared with that of two reanalysis AOD datasets offered by Modern-Era Retrospective Analysis for Research and Applications, Version 2 (MERRA-2) and Copernicus Atmosphere Monitoring Service (CAMS), respectively, covering from July 2015 to December 2017 over China. For all the matchups, AHI AOD shows the highest robustness with a high correlation (R) of 0.82, low root-mean-square error (RMSE) of 0.23, and moderate mean absolute relative error (MARE) of 0.56. Although MERRA-2 and CAMS products both have lower R values (0.74, 0.72) and higher RMSE (0.28, 0.26), the former is slightly better than the latter. Accuracy of AOD products could be mainly affected by the pollution level and less affected by particle size distribution. Comparisons among these AOD products imply that AHI AOD is more reliable in regions with high pollution levels, such as central and eastern China, while in the northern and western part, MERRA-2 AOD seems more satisfying. The performance of all the three AOD products presents a significant diurnal variety, as indicated by the highest accuracy in the morning for AHI and at noon for reanalysis data. Moreover, due to various pollution distribution patterns and meteorological conditions, there are distinct seasonal characteristics in the performance of AOD products for different regions.

Ensemble approach based on bagging, boosting and stacking for short-term prediction in agribusiness time series

Globally, wind energy if properly harnessed, could serve as a source of energy generation in Africa. This study compared the performance of two Machine Learning (ML) algorithms (Linear regression and Random Forest) in predicting wind speed in five major cities in Africa (Yaoundé, Pretoria, Nairobi, Cairo and Abuja). Wind data were collected between January 1, 2000, and December 31, 2022, using the Solar Radiation Data Archive. The data preprocessing was carried out with 80% of the data used for training and 20% for validation. The performance of these ML algorithms was evaluated using Mean Square Error (MSE), Root Mean Square Error (RMSE), Mean Absolute Error (MAE), Mean Absolute Percentage Error (MAPE) and coefficient of determination (R2). The result shows that Nairobi (3.814795 m/s) closely followed by Cairo (3.606453 m/s) has the highest mean wind speed while Yaoundé (1.090512 m/s) has the lowest. Based on the performance metrics used, the two Machine Learning algorithms were competitive. Still, the Linear Regression (LR) algorithm outperformed the Random Forest Algorithm in predicting wind speed in all the selected major African cities. In Yaoundé (RMSE = 0.3892, MAE= 0.3001, MAPE =0.5030), Pretoria (RMSE=1.2339, MAE=0.9480, MAPE=0.7450) Nairobi (RMSE= 0.4223, MAE =0.6499, MAPE =0.1872), Nairobi (RMSE=0.6499, MAE=0.5171, MAPE =0.1872), Cairo (RMSE =1.0909, MAE =0.8544, MAPE =0.3541) and Abuja (RMSE = 0.70245, MAE =0.5441, MAPE= 0.4515) the Linear regression algorithms was found to outperformed Random Forest Regression. Therefore, the Linear regression algorithm is more reliable in predicting wind speed compared with the Random Forest regression.

Tropical cyclones (TCs) are destructive meteorological phenomena, necessitating accurate predictions of TC track and intensity to reduce risks to human life. This study evaluates three Transformer-based models – vanilla Transformer (Transformer), inverted Transformer (iTransformer), and temporal-variate Transformer (TVFormer) – which are trained, validated, and tested on best track data from 1980 to 2021 from the China Meteorological Administration for TC prediction, integrating temporal, variate, and hybrid token embeddings to analyze temporal and variate correlations. Comparative analysis with four recurrent neural network (RNN) models demonstrates the superiority of the refined Transformer models over RNNs: iTransformer reduces mean absolute error (MAE) and root mean square error (RMSE) by 29.55% and 25.80% (latitude), 50.31% and 46.18% (longitude), 8.71% and 9.98% (pressure), and 8.68% and 9.45% (wind speed), while TVFormer achieves MAE and RMSE reductions of 13.98% and 13.84% (latitude), 39.11% and 38.02% (longitude), 13.69% and 14.02% (pressure), and 12.84% and 12.94% (wind speed) on average. Among Transformer variants, iTransformer excels in track prediction, outperforming Transformer with 21.74% lower MAE and 18.26% lower RMSE for latitude, and 32.73% lower MAE and 24.01% lower RMSE for longitude. TVFormer dominates intensity prediction, reducing pressure errors by 4.42% (MAE) and 3.92% (RMSE) and wind speed errors by 19.21% (MAE) and 14.79% (RMSE) compared to Transformer, while outperforming iTransformer with 4.59% lower MAE and 3.68% lower RMSE for pressure and 3.83% lower MAE and 3.18% lower RMSE for wind speed. Notably, TVFormer also enhances track prediction, with 7.10% reduction in MAE and 7.02% reduction in RMSE for latitude, and 22.84% reduction in MAE and 17.09% reduction in RMSE for longitude compared to Transformer. These results highlight the superiority of iTransformer in track prediction and the efficacy of TVFormer in intensity prediction, thanks to their ability to exploit temporal and variate dependencies, offering potential for TC disaster preparedness systems.

This article uses machine learning techniques as fuzzy and neuro-fuzzy ANFISs, to develop and compare prediction models capable of relating pregnant women’s exposure to air pollutants, such as Nitrogen Dioxide and Particulate Matter, the mother’s age, and the number of prenatal consultations to the incidence of premature birth. In the current literature, studies can be found that relate prematurity to the exposure of pregnant women to NO2, O3, and PM10; to Toluene and benzene, mainly in the window 5 to 10 days before birth; and to PM10 in the week before birth. Both models used logistic regression to quantify the effects of pollutants as a result of premature birth. Datasets from Brazil—Departamento de Informatica do Sistema Único de Saúde (DATASUS) and Companhia Ambiental do Estado de São Paulo (CETESB)—were used, covering the period from 2016 to 2018 and comprising women living in the city of São José dos Campos (SP), Brazil. In order to evaluate and compare the different techniques used, evaluation metrics were calculated, such as correlation (r), coefficient of determination (R2), Root Mean Square Error (RMSE), Mean Absolute Percentage Error (MAPE), Mean Square Error (MSE), and Mean Absolute Error (MAE). These metrics are widely used in the literature due to their ability to evaluate the robustness and efficiency of prediction models. For the RMSE, MAPE, MSE, and MAE metrics, lower values indicate that prediction errors are smaller, demonstrating better model accuracy and confidence. In the case of (r) and R2, a positive and strong result indicates alignment and better performance between the real and predicted data. The neuro-fuzzy ANFIS model showed superior performance, with a correlation (r) of 0.59, R2 = 0.35, RMSE = 2.83, MAPE = 5.35%, MSE = 8.00, and MAE = 1.70, while the fuzzy model returned results of r = 0.20, R2 = 0.04, RMSE = 3.29, MSE = 10.81, MAPE = 6.67%, and MAE = 2.01. Therefore, the results from the ANFIS neuro-fuzzy system indicate greater prediction capacity and precision in relation to the fuzzy system. This superiority can be explained by integration with neural networks, allowing data learning and, consequently, more efficient modeling. In addition, the findings obtained in this study have potential for the formulation of public health policies aimed at reducing the number of premature births and promoting improvements in maternal and neonatal health.

The accurate prediction of particulate matter ( ) levels, an indicator of natural pollutants such as those resulting from dust storms, is crucial for public health and environmental planning. This study aims to provide accurate forecasts of over Morocco for 5days. The analog ensemble (AnEn) and the bias correction (AnEnBc) techniques were employed to post-process forecasts produced by the Copernicus Atmosphere Monitoring Service (CAMS) global atmospheric composition forecasts, using CAMS reanalysis data as a reference. The results show substantial prediction improvements: the root mean squared error (RMSE) decreased from 63.83 in the original forecasts to 44.73 with AnEn and AnEnBc, while the mean absolute error (MAE) reduced from 36.70 to 24.30 . Additionally, the coefficient of determination ( ) increased more than twofold from 29.11 to 65.18%, and the Pearson correlation coefficient increased from 0.61 to 0.82. The integrating reanalysis data and the utilization of the AnEn substantially improved the accuracy of 5-day forecasting in Morocco. This is the first application of this approach in Morocco and the Middle East and North Africa (MENA) and has the potential for translation into early and more accurate warnings of pollution events. The application of such approaches in environmental policies and public health decision-making can minimize the health impacts of air pollution.

Machine learning techniques are widely used algorithms for predicting the mechanical properties of concrete. This study is based on the comparison of algorithms between individuals and ensemble approaches, such as bagging. Optimization for bagging is done by making 20 sub-models to depict the accurate one. Variables like cement content, fine and coarse aggregate, water, binder-to-water ratio, fly-ash, and superplasticizer are used for modeling. Model performance is evaluated by various statistical indicators like mean absolute error (MAE), mean square error (MSE), and root mean square error (RMSE). Individual algorithms show a moderate bias result. However, the ensemble model gives a better result with R2 = 0.911 compared to the decision tree (DT) and gene expression programming (GEP). K-fold cross-validation confirms the model’s accuracy and is done by R2, MAE, MSE, and RMSE. Statistical checks reveal that the decision tree with ensemble provides 25%, 121%, and 49% enhancement for errors like MAE, MSE, and RMSE between the target and outcome response.

Road inconsistency may cause dangerous driving experience that leads to carelessness while handling the vehicle. To determine the road inconsistency, development of models which are based on the 85th percentile of speed is adopted. The models which are taken into consideration are other geometric road characteristics such as superelevation, radius of curve, width of curve, tangent length, and road gradient that can help to minimize the risk of accidents at curve. To further analyze the effectiveness of these developed models, validation of the models using new set of data and comparing it with other established models is crucial. The validation process is conducted by calculating the mean squared error (MSE), mean absolute error (MAE), and mean absolute percentage error (MAPE). MSE value will indicate the differences between the empirical results with the predicted results, while MAE calculates the absolute difference between the empirical results and predicted results. MAPE evaluates the percentage errors between the empirical results with the predicted results. Small values of MSE, MAE, and MAPE is a proof that the models can predict good results.

Evaluation of artificial neurocomputing algorithms and their metacognitive robustness in predictive modeling of fuel consumption rates during tillage

Experiments have proven that traditional prediction research methods have limitations in practice. Proposing countermeasures for environmental changes is the key to optimal control of the cold chain environment and reducing the lag of control effects. In this paper, a cold chain transportation environment prediction method, combining k-means++ and a long short-term memory (LSTM) neural network, is proposed according to the characteristics of the cold chain transportation environment of agricultural products. The proposed prediction model can predict the trend of cold chain environment changes in the next ten minutes, which allows cold chain vehicle managers to issue control instructions to the environmental control equipment in advance. The fusion process for temperature and humidity data measured by multiple data sensors is performed with the k-means++ algorithm, and then the fused data are fed into an LSTM neural network for prediction based on time series. The prediction error of the prediction model proposed in this paper is very satisfactory, with a root-mean-square error (RMSE), mean absolute error (MAE), mean squared error (MSE), mean absolute percentage error (MAPE) and R-squared of 0.5707, 0.2484, 0.3258, 0.0312 and 0.9660, respectively, for temperature prediction, and with an RMSE, MAE, MSE, mean absolute percentage error and R-squared of 1.6015, 1.1770, 2.5648, 0.2736 and 0.9702, respectively, for humidity prediction. Finally, the LSTM neural network and back propagation (BP) neural network are compared in order to enhance the reliability of the results. In terms of the prediction effect of the temperature and humidity in cold chain vehicles transporting agricultural products, the proposed model has a higher prediction accuracy than that of existing models and can provide strategic support for the fine management and regulation of the cold chain transportation environment.