Pub Date : 2026-02-01Epub Date: 2026-01-16DOI: 10.1016/j.jastp.2026.106735
Teddy Miller Samo, Calvine Ominde, Justus Maithya, James Munyithya
This study explores the application of Artificial Neural Network (ANN) architectures, particularly Long Short-Term Memory (LSTM) and Gated Recurrent Units (GRU), for forecasting CO2 concentration trends in geothermal fields. The research is motivated by the growing need to quantify and predict emissions from geothermal power generation—an important renewable energy source whose environmental impacts are often overlooked. Geothermal field measurements were combined with meteorological variables to train and validate the models. Performance was evaluated using mean absolute error (MAE), root mean square error (RMSE), mean absolute percentage error (MAPE), coefficient of determination (R2), and bias. The ANN (LSTM–GRU) model demonstrated superior predictive ability, achieving lower RMSE (0.0952 vs. 0.0989), MAE (0.0809 vs. 0.0828), and MAPE (0.20 % vs. 0.21 %), with a higher R2 (0.8912 vs. 0.8361) compared to the SARIMA model. Although ANN recorded a slightly higher bias (0.0797 vs. 0.002), its overall performance underscores its effectiveness in modeling complex, non-linear, and temporal patterns of CO2 concentration. The findings confirm that ANN-based models are more accurate and adaptable than conventional statistical approaches. Their application in geothermal fields provides a robust tool for forecasting emissions, enabling better planning, monitoring, and implementation of environmental strategies to mitigate the contribution of geothermal energy production to greenhouse gas emissions.
本研究探讨了人工神经网络（ANN）架构，特别是长短期记忆（LSTM）和门控循环单元（GRU）在地热田二氧化碳浓度趋势预测中的应用。地热发电是一种重要的可再生能源，其对环境的影响往往被忽视。结合地热场实测数据和气象变量对模型进行训练和验证。使用平均绝对误差（MAE）、均方根误差（RMSE）、平均绝对百分比误差（MAPE）、决定系数（R2）和偏倚来评估性能。与SARIMA模型相比，ANN （LSTM-GRU）模型的预测能力更强，RMSE （0.0952 vs. 0.0989）、MAE （0.0809 vs. 0.0828）和MAPE （0.20% vs. 0.21%）均较低，R2 （0.8912 vs. 0.8361）较高。尽管人工神经网络记录的偏差略高（0.0797 vs. 0.002），但其总体性能强调了其在模拟复杂、非线性和时间模式的CO2浓度方面的有效性。研究结果证实，基于人工神经网络的模型比传统的统计方法更准确，适应性更强。它们在地热领域的应用为预测排放提供了一个强有力的工具，能够更好地规划、监测和实施环境战略，以减轻地热能生产对温室气体排放的贡献。
{"title":"Artificial Neural Network (ANN) modeling for CO2 concentration prediction in geothermal fields","authors":"Teddy Miller Samo, Calvine Ominde, Justus Maithya, James Munyithya","doi":"10.1016/j.jastp.2026.106735","DOIUrl":"10.1016/j.jastp.2026.106735","url":null,"abstract":"<div><div>This study explores the application of Artificial Neural Network (ANN) architectures, particularly Long Short-Term Memory (LSTM) and Gated Recurrent Units (GRU), for forecasting CO<sub>2</sub> concentration trends in geothermal fields. The research is motivated by the growing need to quantify and predict emissions from geothermal power generation—an important renewable energy source whose environmental impacts are often overlooked. Geothermal field measurements were combined with meteorological variables to train and validate the models. Performance was evaluated using mean absolute error (MAE), root mean square error (RMSE), mean absolute percentage error (MAPE), coefficient of determination (R<sup>2</sup>), and bias. The ANN (LSTM–GRU) model demonstrated superior predictive ability, achieving lower RMSE (0.0952 vs. 0.0989), MAE (0.0809 vs. 0.0828), and MAPE (0.20 % vs. 0.21 %), with a higher R<sup>2</sup> (0.8912 vs. 0.8361) compared to the SARIMA model. Although ANN recorded a slightly higher bias (0.0797 vs. 0.002), its overall performance underscores its effectiveness in modeling complex, non-linear, and temporal patterns of CO<sub>2</sub> concentration. The findings confirm that ANN-based models are more accurate and adaptable than conventional statistical approaches. Their application in geothermal fields provides a robust tool for forecasting emissions, enabling better planning, monitoring, and implementation of environmental strategies to mitigate the contribution of geothermal energy production to greenhouse gas emissions.</div></div>","PeriodicalId":15096,"journal":{"name":"Journal of Atmospheric and Solar-Terrestrial Physics","volume":"279 ","pages":"Article 106735"},"PeriodicalIF":1.9,"publicationDate":"2026-02-01","publicationTypes":"Journal Article","fieldsOfStudy":null,"isOpenAccess":false,"openAccessPdf":"","citationCount":null,"resultStr":null,"platform":"Semanticscholar","paperid":"146034950","PeriodicalName":null,"FirstCategoryId":null,"ListUrlMain":null,"RegionNum":4,"RegionCategory":"地球科学","ArticlePicture":[],"TitleCN":null,"AbstractTextCN":null,"PMCID":"","EPubDate":null,"PubModel":null,"JCR":null,"JCRName":null,"Score":null,"Total":0}
Pub Date : 2026-02-01Epub Date: 2026-01-02DOI: 10.1016/j.jastp.2025.106714
Tong Zhang , Xing Wang , Hao Li , Qiliang Wu , Dan Geng
The accuracy of precipitation forecasting exerts significant impacts on socio-economic activities. However, numerical weather prediction (NWP) models often exhibit systematic biases due to the chaotic nature of atmospheric dynamics and localized topographic effects. This study proposes M3F-U-Net, a multi-stream multi-level multi-physics Fusion U-Net designed to decouple and reconstruct atmospheric dynamic, thermodynamic, moisture transport, and topographic forcing processes. A dynamic weighted fusion module is employed to adaptively adjust the importance of physical features, while a residual learning framework is incorporated to achieve precipitation forecast correction. Experiments based on ECMWF model forecasts and ERA5 reanalysis data demonstrate that the model reduces the Mean Squared Error (MSE) and Mean Absolute Error (MAE) by 20.04 % and 14.42 %, respectively. Notably, the Equitable Threat Score (ETS) improves by 6.19 % overall, driven by a significant reduction in false alarms. Subgroup analysis confirms the model's robustness across diverse terrains, achieving an 8.56 % ETS improvement even in complex mountainous regions. Furthermore, the model effectively mitigates nonlinear error accumulation in long-term forecasts (up to 72 h). This research establishes a physically consistent deep learning framework for mitigating systematic biases in NWP-based precipitation forecasts, providing an innovative technical paradigm for numerical model post-processing.
{"title":"M3F-U-Net: A hybrid deep learning model for precipitation forecast correction","authors":"Tong Zhang , Xing Wang , Hao Li , Qiliang Wu , Dan Geng","doi":"10.1016/j.jastp.2025.106714","DOIUrl":"10.1016/j.jastp.2025.106714","url":null,"abstract":"<div><div>The accuracy of precipitation forecasting exerts significant impacts on socio-economic activities. However, numerical weather prediction (NWP) models often exhibit systematic biases due to the chaotic nature of atmospheric dynamics and localized topographic effects. This study proposes M<sup>3</sup>F-U-Net, a multi-stream multi-level multi-physics Fusion U-Net designed to decouple and reconstruct atmospheric dynamic, thermodynamic, moisture transport, and topographic forcing processes. A dynamic weighted fusion module is employed to adaptively adjust the importance of physical features, while a residual learning framework is incorporated to achieve precipitation forecast correction. Experiments based on ECMWF model forecasts and ERA5 reanalysis data demonstrate that the model reduces the Mean Squared Error (MSE) and Mean Absolute Error (MAE) by 20.04 % and 14.42 %, respectively. Notably, the Equitable Threat Score (ETS) improves by 6.19 % overall, driven by a significant reduction in false alarms. Subgroup analysis confirms the model's robustness across diverse terrains, achieving an 8.56 % ETS improvement even in complex mountainous regions. Furthermore, the model effectively mitigates nonlinear error accumulation in long-term forecasts (up to 72 h). This research establishes a physically consistent deep learning framework for mitigating systematic biases in NWP-based precipitation forecasts, providing an innovative technical paradigm for numerical model post-processing.</div></div>","PeriodicalId":15096,"journal":{"name":"Journal of Atmospheric and Solar-Terrestrial Physics","volume":"279 ","pages":"Article 106714"},"PeriodicalIF":1.9,"publicationDate":"2026-02-01","publicationTypes":"Journal Article","fieldsOfStudy":null,"isOpenAccess":false,"openAccessPdf":"","citationCount":null,"resultStr":null,"platform":"Semanticscholar","paperid":"145940739","PeriodicalName":null,"FirstCategoryId":null,"ListUrlMain":null,"RegionNum":4,"RegionCategory":"地球科学","ArticlePicture":[],"TitleCN":null,"AbstractTextCN":null,"PMCID":"","EPubDate":null,"PubModel":null,"JCR":null,"JCRName":null,"Score":null,"Total":0}
Pub Date : 2026-02-01Epub Date: 2026-01-07DOI: 10.1016/j.jastp.2026.106730
Abhimanyu Kumar Gond, Aarif Jamal, Tarun Verma
This study employs a hybrid Temporal Convolutional Network (TCN) and Transformer model to forecast air quality trends in Korba, Chhattisgarh, a critically polluted industrial hub dominated by extensive coal mining (including Gevra, Dipka, and Kusmunda mines), multiple thermal power plants, aluminium smelters, and cement production, from 2025 to 2047. Utilizing 44 years of satellite-derived data (1980–2024), the model integrates meteorological variables, vegetation indices (NDVI and NDBI), and coal mining metrics. Under a moderate policy change scenario, it predicts gradually rising pollutant levels: PM10 (32–84 μg/m3), PM2.5 (10–33 μg/m3), SO2 (7.25–12 μg/m3), NO2 (5.25–9 μg/m3), and AQI (45–110, Moderate per CPCB Delhi standards), with seasonal patterns showing reduced concentrations during monsoon due to rainfall washout and elevated levels in summer and winter owing to limited atmospheric dispersion. The model demonstrated strong performance (R2 = 0.75–0.91; RMSE = 1.09–11.91), effectively capturing both short- and long-term trends driven by industrial emissions and environmental factors. A sensitivity analysis further revealed the model's robust response to ±10–20 % variations in key drivers, with the most decisive influence from coal production increases, which could be elevated by 20–30 % and add 10–15 AQI points, while reduced rainfall, higher temperatures, and lower NDVI amplified dust resuspension and secondary pollutant formation. Uncertainty analysis identified high-risk periods, including elevated PM2.5 variability in 2025–2026 and AQI in 2033–2038. Validation against ground-truth data from Urja Nagar, Rampur stations, and MODIS Satellite-derived AQI (From January–September 2025; R2: 0.72, 0.61, and 0.58) confirmed forecasted AQI (30–164, mostly Moderate), posing potential respiratory risks to vulnerable groups upon prolonged exposure. These projections highlight escalating public health threats, particularly respiratory and cardiovascular diseases, underscoring the need for urgent interventions, such as stricter emission controls, a transition to renewable energy sources, the adoption of an air quality health index, and enhanced reforestation for dust mitigation. This work offers a robust, data-driven baseline and scalable framework for sustainable air quality management in industrial regions, aligning with India's vision for balanced development by 2047.
{"title":"Long-term air quality forecasting in Korba, India (2025–2047): A hybrid model using 44-year satellite data","authors":"Abhimanyu Kumar Gond, Aarif Jamal, Tarun Verma","doi":"10.1016/j.jastp.2026.106730","DOIUrl":"10.1016/j.jastp.2026.106730","url":null,"abstract":"<div><div>This study employs a hybrid Temporal Convolutional Network (TCN) and Transformer model to forecast air quality trends in Korba, Chhattisgarh, a critically polluted industrial hub dominated by extensive coal mining (including Gevra, Dipka, and Kusmunda mines), multiple thermal power plants, aluminium smelters, and cement production, from 2025 to 2047. Utilizing 44 years of satellite-derived data (1980–2024), the model integrates meteorological variables, vegetation indices (NDVI and NDBI), and coal mining metrics. Under a moderate policy change scenario, it predicts gradually rising pollutant levels: PM<sub>10</sub> (32–84 μg/m<sup>3</sup>), PM<sub>2</sub>.<sub>5</sub> (10–33 μg/m<sup>3</sup>), SO<sub>2</sub> (7.25–12 μg/m<sup>3</sup>), NO<sub>2</sub> (5.25–9 μg/m<sup>3</sup>), and AQI (45–110, Moderate per CPCB Delhi standards), with seasonal patterns showing reduced concentrations during monsoon due to rainfall washout and elevated levels in summer and winter owing to limited atmospheric dispersion. The model demonstrated strong performance (R<sup>2</sup> = 0.75–0.91; RMSE = 1.09–11.91), effectively capturing both short- and long-term trends driven by industrial emissions and environmental factors. A sensitivity analysis further revealed the model's robust response to ±10–20 % variations in key drivers, with the most decisive influence from coal production increases, which could be elevated by 20–30 % and add 10–15 AQI points, while reduced rainfall, higher temperatures, and lower NDVI amplified dust resuspension and secondary pollutant formation. Uncertainty analysis identified high-risk periods, including elevated PM<sub>2</sub>.<sub>5</sub> variability in 2025–2026 and AQI in 2033–2038. Validation against ground-truth data from Urja Nagar, Rampur stations, and MODIS Satellite-derived AQI (From January–September 2025; R<sup>2</sup>: 0.72, 0.61, and 0.58) confirmed forecasted AQI (30–164, mostly Moderate), posing potential respiratory risks to vulnerable groups upon prolonged exposure. These projections highlight escalating public health threats, particularly respiratory and cardiovascular diseases, underscoring the need for urgent interventions, such as stricter emission controls, a transition to renewable energy sources, the adoption of an air quality health index, and enhanced reforestation for dust mitigation. This work offers a robust, data-driven baseline and scalable framework for sustainable air quality management in industrial regions, aligning with India's vision for balanced development by 2047.</div></div>","PeriodicalId":15096,"journal":{"name":"Journal of Atmospheric and Solar-Terrestrial Physics","volume":"279 ","pages":"Article 106730"},"PeriodicalIF":1.9,"publicationDate":"2026-02-01","publicationTypes":"Journal Article","fieldsOfStudy":null,"isOpenAccess":false,"openAccessPdf":"","citationCount":null,"resultStr":null,"platform":"Semanticscholar","paperid":"145979506","PeriodicalName":null,"FirstCategoryId":null,"ListUrlMain":null,"RegionNum":4,"RegionCategory":"地球科学","ArticlePicture":[],"TitleCN":null,"AbstractTextCN":null,"PMCID":"","EPubDate":null,"PubModel":null,"JCR":null,"JCRName":null,"Score":null,"Total":0}
This study compared three forecasting approaches: Classical time-series, ARIMA variants, and Artificial Neural Network (ANN)-based Long Short-Term Memory (LSTM) models for PM2.5 prediction at two Delhi stations (Bawana and Okhla). All models produced strong forecasts and revealed consistently high predictive performance, with accuracy metrics and error values within acceptable ranges, R2 values above 0.82 and RMSE between 15 and 30 μg/m3. Okhla generally outperformed Bawana, achieving the highest Classical model R2 (0.957) and lowest RMSE (15.70 μg/m3). ARIMA delivered reliable results (Okhla: R2 ≈ 0.913, RMSE ≈ 22.49 μg/m3), but performance declined at Bawana (R2 ≈ 0.826, RMSE ≈ 29.18 μg/m3), highlighting its sensitivity to station-specific parameters and stationarity assumptions. LSTM architectures outperformed both classical and ARIMA models. ConvLSTM achieved the best overall accuracy (Okhla: R2 = 0.891, RMSE = 20.10 μg/m3; Bawana: R2 = 0.865, RMSE = 23.74 μg/m3), closely followed by Stacked LSTM (Okhla: R2 = 0.889, RMSE = 20.78 μg/m3) and Bi-directional LSTM (Okhla: R2 = 0.901, RMSE = 23.31 μg/m3). These models demonstrated stable convergence, minimal overfitting, and robustness against noise. While classical and ARIMA methods remain interpretable and computationally efficient, they struggle with complex nonlinear dependencies and require substantial reconfiguration for new sites. LSTM models, by contrast, adapt easily to other monitoring locations with minor hyperparameter tuning (e.g., learning rate, sequence length, layer depth), making them more scalable for city-wide deployment. In conclusion, although all approaches are viable for PM2.5 forecasting, LSTM-based models, particularly ConvLSTM and Stacked LSTM provide the highest accuracy, robustness, and adaptability, offering a reliable framework for operational air quality management in Delhi. From an applied perspective, the improved short-term predictability offered by LSTM-based models can support early-warning systems, targeted emission-control strategies, and data-driven decision-making for urban air-quality management.
{"title":"Geospatial time series forecasting of air quality across Delhi using ARIMA and LSTM models","authors":"Faizan Tahir Bahadur , Shagoofta Rasool Shah , Rama Rao Nidamanuri","doi":"10.1016/j.jastp.2026.106723","DOIUrl":"10.1016/j.jastp.2026.106723","url":null,"abstract":"<div><div>This study compared three forecasting approaches: Classical time-series, ARIMA variants, and Artificial Neural Network (ANN)-based Long Short-Term Memory (LSTM) models for PM2.5 prediction at two Delhi stations (Bawana and Okhla). All models produced strong forecasts and revealed consistently high predictive performance, with accuracy metrics and error values within acceptable ranges, R<sup>2</sup> values above 0.82 and RMSE between 15 and 30 μg/m<sup>3</sup>. Okhla generally outperformed Bawana, achieving the highest Classical model R<sup>2</sup> (0.957) and lowest RMSE (15.70 μg/m<sup>3</sup>). ARIMA delivered reliable results (Okhla: R<sup>2</sup> ≈ 0.913, RMSE ≈ 22.49 μg/m<sup>3</sup>), but performance declined at Bawana (R<sup>2</sup> ≈ 0.826, RMSE ≈ 29.18 μg/m<sup>3</sup>), highlighting its sensitivity to station-specific parameters and stationarity assumptions. LSTM architectures outperformed both classical and ARIMA models. ConvLSTM achieved the best overall accuracy (Okhla: R<sup>2</sup> = 0.891, RMSE = 20.10 μg/m<sup>3</sup>; Bawana: R<sup>2</sup> = 0.865, RMSE = 23.74 μg/m<sup>3</sup>), closely followed by Stacked LSTM (Okhla: R<sup>2</sup> = 0.889, RMSE = 20.78 μg/m<sup>3</sup>) and Bi-directional LSTM (Okhla: R<sup>2</sup> = 0.901, RMSE = 23.31 μg/m<sup>3</sup>). These models demonstrated stable convergence, minimal overfitting, and robustness against noise. While classical and ARIMA methods remain interpretable and computationally efficient, they struggle with complex nonlinear dependencies and require substantial reconfiguration for new sites. LSTM models, by contrast, adapt easily to other monitoring locations with minor hyperparameter tuning (e.g., learning rate, sequence length, layer depth), making them more scalable for city-wide deployment. In conclusion, although all approaches are viable for PM2.5 forecasting, LSTM-based models, particularly ConvLSTM and Stacked LSTM provide the highest accuracy, robustness, and adaptability, offering a reliable framework for operational air quality management in Delhi. From an applied perspective, the improved short-term predictability offered by LSTM-based models can support early-warning systems, targeted emission-control strategies, and data-driven decision-making for urban air-quality management.</div></div>","PeriodicalId":15096,"journal":{"name":"Journal of Atmospheric and Solar-Terrestrial Physics","volume":"279 ","pages":"Article 106723"},"PeriodicalIF":1.9,"publicationDate":"2026-02-01","publicationTypes":"Journal Article","fieldsOfStudy":null,"isOpenAccess":false,"openAccessPdf":"","citationCount":null,"resultStr":null,"platform":"Semanticscholar","paperid":"145940741","PeriodicalName":null,"FirstCategoryId":null,"ListUrlMain":null,"RegionNum":4,"RegionCategory":"地球科学","ArticlePicture":[],"TitleCN":null,"AbstractTextCN":null,"PMCID":"","EPubDate":null,"PubModel":null,"JCR":null,"JCRName":null,"Score":null,"Total":0}
Estimating agricultural drought is hindered by limited data availability. Conversely, assessing meteorological drought is very easy as it requires only meteorological data that are easy to obtain. Thus, the study aimed to find meteorological drought indices suitable for assessing agricultural drought in a tropical river basin using remote sensing satellite data. Seven meteorological drought indices, i.e., Standardised Precipitation Index (SPI), Deciles Index (DI), Percent of Normal Index (PNI), Rainfall Anomaly Index (RAI), China-Z Index (CZI), Modified China-Z Index (MCZI), and Z-score Index (ZSI) were compared with an efficient agricultural drought index, viz., Standardised Precipitation Evapotranspiration Index (SPEI) for the Subarnarekha River basin using the NASA POWER satellite data. It was noticed that the CZI had the highest correlation with the 3-month SPEI (R2 = 0.88–0.94, followed by the RAI (R2 = 0.85–0.90). The meteorological drought indices were divided into three groups based on the ability to replicate SPEI: ‘excellent’ for SPI, CZI, and MCZI; ‘good’ for ZSI and RAI; and ‘poor’ for PNI and DI. Relying solely on correlation coefficients to evaluate drought indices provided erroneous results. However, integrating correlation, RMSE, and statistical distribution enhanced the robustness of the assessment. This approach reclassified MCZI and SPI from “good” to “excellent”. SPI and CZI showed better alignment with SPEI distributions throughout both dry and wet periods, establishing them as the most reliable alternatives for evaluating agricultural drought. The findings emphasize the importance of comprehensive evaluation criteria that incorporate several statistical metrics for reliable drought assessment in tropical and data-scarce regions.
{"title":"Assessing meteorological drought indices for monitoring agricultural drought using SPEI: a remote sensing approach","authors":"Kanhu Charan Panda , Pradosh Kumar Paramaguru , Ram Mandir Singh , Sudhir Kumar Singh","doi":"10.1016/j.jastp.2026.106726","DOIUrl":"10.1016/j.jastp.2026.106726","url":null,"abstract":"<div><div>Estimating agricultural drought is hindered by limited data availability. Conversely, assessing meteorological drought is very easy as it requires only meteorological data that are easy to obtain. Thus, the study aimed to find meteorological drought indices suitable for assessing agricultural drought in a tropical river basin using remote sensing satellite data. Seven meteorological drought indices, i.e., Standardised Precipitation Index (SPI), Deciles Index (DI), Percent of Normal Index (PNI), Rainfall Anomaly Index (RAI), China-Z Index (CZI), Modified China-Z Index (MCZI), and Z-score Index (ZSI) were compared with an efficient agricultural drought index, viz., Standardised Precipitation Evapotranspiration Index (SPEI) for the Subarnarekha River basin using the NASA POWER satellite data. It was noticed that the CZI had the highest correlation with the 3-month SPEI (R<sup>2</sup> = 0.88–0.94, followed by the RAI (R<sup>2</sup> = 0.85–0.90). The meteorological drought indices were divided into three groups based on the ability to replicate SPEI: ‘excellent’ for SPI, CZI, and MCZI; ‘good’ for ZSI and RAI; and ‘poor’ for PNI and DI. Relying solely on correlation coefficients to evaluate drought indices provided erroneous results. However, integrating correlation, RMSE, and statistical distribution enhanced the robustness of the assessment. This approach reclassified MCZI and SPI from “good” to “excellent”. SPI and CZI showed better alignment with SPEI distributions throughout both dry and wet periods, establishing them as the most reliable alternatives for evaluating agricultural drought. The findings emphasize the importance of comprehensive evaluation criteria that incorporate several statistical metrics for reliable drought assessment in tropical and data-scarce regions.</div></div>","PeriodicalId":15096,"journal":{"name":"Journal of Atmospheric and Solar-Terrestrial Physics","volume":"279 ","pages":"Article 106726"},"PeriodicalIF":1.9,"publicationDate":"2026-02-01","publicationTypes":"Journal Article","fieldsOfStudy":null,"isOpenAccess":false,"openAccessPdf":"","citationCount":null,"resultStr":null,"platform":"Semanticscholar","paperid":"145940740","PeriodicalName":null,"FirstCategoryId":null,"ListUrlMain":null,"RegionNum":4,"RegionCategory":"地球科学","ArticlePicture":[],"TitleCN":null,"AbstractTextCN":null,"PMCID":"","EPubDate":null,"PubModel":null,"JCR":null,"JCRName":null,"Score":null,"Total":0}
Accurate forecasting of tropospheric surface ozone (O3) concentrations is increasingly vital due to rapid urbanization and industrial growth, which have intensified photochemical smog episodes and associated health and environmental risks. Traditional deterministic models often struggle to capture the complex nonlinear relationships among meteorological and precursor variables influencing O3 formation. Addressing this limitation, the present study introduces a Radial Basis Function Network (RBFN) model for seasonal forecasting of surface O3 concentrations using real-time ground monitoring data (2014–2017) from an urban air-quality station in Hyderabad, India (TIFR–NBF: Tata Institute of Fundamental Research–National Balloon Facility). The novelty of this work lies in applying RBFNs to capture seasonal and diurnal O3 variations with high precision, using input parameters including NO, NO2, NOx, CO, VOCs, and NMHCs. The RBFN model was trained with 80 % of the dataset and validated using 20 %, and its performance was assessed through Mean Square Error (MSE), Mean Absolute Error (MAE), Root Mean Square Error (RMSE), and the Willmott's index(WI)Comparative analysis with the nonlinear regression-based Response Surface Methodology (RSM) revealed that the RBFN achieved superior predictive accuracy with lower RMSE (0.0038 vs. 0.0185), MSE (0.000014 vs. 0.000342), MAE (0.0021 vs. 0.0094), and (WI = 0.9999 vs. 0.9812), The model effectively captured seasonal and diurnal O3 peaks, particularly during summer and pre-monsoon periods. Overall, this study demonstrates the robustness, adaptability, and real-time forecasting potential of RBFNs, underscoring their application in early-warning systems, air-quality management, and policy-driven pollution mitigation strategies for rapidly urbanizing regions.
由于快速的城市化和工业增长加剧了光化学烟雾事件和相关的健康和环境风险，对流层表面臭氧（O3）浓度的准确预报变得越来越重要。传统的确定性模型往往难以捕捉影响臭氧形成的气象和前兆变量之间复杂的非线性关系。为了解决这一限制，本研究引入了径向基函数网络（RBFN）模型，利用印度海德拉巴城市空气质量站（TIFR-NBF：塔塔基础研究所-国家气球设施）的实时地面监测数据（2014-2017年）对地表O3浓度进行季节性预测。这项工作的新颖之处在于，利用输入参数，包括NO、NO2、NOx、CO、VOCs和NMHCs，应用rbfn以高精度捕获季节和日O3变化。RBFN模型使用80%的数据集进行训练，20%的数据集进行验证，并通过均方误差（MSE）、平均绝对误差（MAE）、均方根误差（RMSE）和Willmott指数（WI）对其性能进行评估。与基于非线性回归的响应面方法（RSM）的对比分析表明，RBFN模型在RMSE （0.0038 vs. 0.0185）、MSE （0.000014 vs. 0.000342）、MAE （0.0021 vs. 0.0094）、RBFN模型预测精度较低的情况下取得了较好的预测精度。和（WI = 0.9999 vs. 0.9812），该模式有效捕获了季节和日O3峰值，特别是在夏季和季风前期。总体而言，本研究证明了rbfn的鲁棒性、适应性和实时预测潜力，强调了其在快速城市化地区的预警系统、空气质量管理和政策驱动的污染缓解战略中的应用。
{"title":"Forecasting seasonal surface ozone trends using radial basis function networks and ground level monitoring data","authors":"Sharanya Suraboyina , Gangagni Rao Anupoju , Anand Polumati","doi":"10.1016/j.jastp.2026.106725","DOIUrl":"10.1016/j.jastp.2026.106725","url":null,"abstract":"<div><div>Accurate forecasting of tropospheric surface ozone (O<sub>3</sub>) concentrations is increasingly vital due to rapid urbanization and industrial growth, which have intensified photochemical smog episodes and associated health and environmental risks. Traditional deterministic models often struggle to capture the complex nonlinear relationships among meteorological and precursor variables influencing O<sub>3</sub> formation. Addressing this limitation, the present study introduces a Radial Basis Function Network (RBFN) model for seasonal forecasting of surface O<sub>3</sub> concentrations using real-time ground monitoring data (2014–2017) from an urban air-quality station in Hyderabad, India (TIFR–NBF: Tata Institute of Fundamental Research–National Balloon Facility). The novelty of this work lies in applying RBFNs to capture seasonal and diurnal O<sub>3</sub> variations with high precision, using input parameters including NO, NO<sub>2</sub>, NO<sub>x</sub>, CO, VOCs, and NMHCs. The RBFN model was trained with 80 % of the dataset and validated using 20 %, and its performance was assessed through Mean Square Error (MSE), Mean Absolute Error (MAE), Root Mean Square Error (RMSE), and the Willmott's index(WI)Comparative analysis with the nonlinear regression-based Response Surface Methodology (RSM) revealed that the RBFN achieved superior predictive accuracy with lower RMSE (0.0038 vs. 0.0185), MSE (0.000014 vs. 0.000342), MAE (0.0021 vs. 0.0094), and (WI = 0.9999 vs. 0.9812), The model effectively captured seasonal and diurnal O<sub>3</sub> peaks, particularly during summer and pre-monsoon periods. Overall, this study demonstrates the robustness, adaptability, and real-time forecasting potential of RBFNs, underscoring their application in early-warning systems, air-quality management, and policy-driven pollution mitigation strategies for rapidly urbanizing regions.</div></div>","PeriodicalId":15096,"journal":{"name":"Journal of Atmospheric and Solar-Terrestrial Physics","volume":"279 ","pages":"Article 106725"},"PeriodicalIF":1.9,"publicationDate":"2026-02-01","publicationTypes":"Journal Article","fieldsOfStudy":null,"isOpenAccess":false,"openAccessPdf":"","citationCount":null,"resultStr":null,"platform":"Semanticscholar","paperid":"145979914","PeriodicalName":null,"FirstCategoryId":null,"ListUrlMain":null,"RegionNum":4,"RegionCategory":"地球科学","ArticlePicture":[],"TitleCN":null,"AbstractTextCN":null,"PMCID":"","EPubDate":null,"PubModel":null,"JCR":null,"JCRName":null,"Score":null,"Total":0}
Pub Date : 2026-02-01Epub Date: 2026-01-18DOI: 10.1016/j.jastp.2026.106738
Tongsen Yue
The author investigated the degradation behavior and mechanism of FACs perovskite solar cells under high-intensity (1–7.35 suns) and multi band illumination (ultraviolet, blue, visible light) conditions. The results indicate that the degradation rate of the device is linearly related to the light intensity within the range of 1–4 solar intensities; When there are ≥5 suns, the nonlinear acceleration phenomenon is significant, mainly due to the increase of interface defects and intensified ion migration caused by the photothermal coupling effect. The multi band illumination experiment showed that ultraviolet and blue light had the greatest impact on device stability. After 500 h of ultraviolet light irradiation, PCE decreased by 42.3 %, while blue light decreased by 28.6 %. Optimized packaging (Opt-Enc-M) combined with water cooling system significantly improves stability, with PCE retention rate exceeding 80 % after 1000 h under 5 solar intensities. The research provides a theoretical basis for standardizing accelerated aging testing and improving device stability.
{"title":"Degradation behavior of perovskite solar cells under high-intensity and multi band illumination conditions","authors":"Tongsen Yue","doi":"10.1016/j.jastp.2026.106738","DOIUrl":"10.1016/j.jastp.2026.106738","url":null,"abstract":"<div><div>The author investigated the degradation behavior and mechanism of FACs perovskite solar cells under high-intensity (1–7.35 suns) and multi band illumination (ultraviolet, blue, visible light) conditions. The results indicate that the degradation rate of the device is linearly related to the light intensity within the range of 1–4 solar intensities; When there are ≥5 suns, the nonlinear acceleration phenomenon is significant, mainly due to the increase of interface defects and intensified ion migration caused by the photothermal coupling effect. The multi band illumination experiment showed that ultraviolet and blue light had the greatest impact on device stability. After 500 h of ultraviolet light irradiation, PCE decreased by 42.3 %, while blue light decreased by 28.6 %. Optimized packaging (Opt-Enc-M) combined with water cooling system significantly improves stability, with PCE retention rate exceeding 80 % after 1000 h under 5 solar intensities. The research provides a theoretical basis for standardizing accelerated aging testing and improving device stability.</div></div>","PeriodicalId":15096,"journal":{"name":"Journal of Atmospheric and Solar-Terrestrial Physics","volume":"279 ","pages":"Article 106738"},"PeriodicalIF":1.9,"publicationDate":"2026-02-01","publicationTypes":"Journal Article","fieldsOfStudy":null,"isOpenAccess":false,"openAccessPdf":"","citationCount":null,"resultStr":null,"platform":"Semanticscholar","paperid":"146073907","PeriodicalName":null,"FirstCategoryId":null,"ListUrlMain":null,"RegionNum":4,"RegionCategory":"地球科学","ArticlePicture":[],"TitleCN":null,"AbstractTextCN":null,"PMCID":"","EPubDate":null,"PubModel":null,"JCR":null,"JCRName":null,"Score":null,"Total":0}
Pub Date : 2026-02-01Epub Date: 2026-01-14DOI: 10.1016/j.jastp.2026.106734
Rafath Samrin , Pundru Chandra Shaker Reddy , K. Arun Kumar , Natha Deepthi , C. Mithra , S Bhargavi Latha , Sucharitha Yadala , Gopal Kumar Thakur
India is the dominant player in the cultivation of rice around the world. Rice yield prediction can be considered as a problem that needs to be solved. Accurate and timely prediction of rice yield can provide meaningful benefits to crop yield. Time series models are widely used for rice yield prediction, but their accuracy remains inadequate. Despite their prominence, they often fail to deliver the required precision. This study considers one of the most practical machine learning (ML) methods for predicting rice yield, allowing forecasts for the next five years. The research carried out shows the rice yield prediction done using the hybrid framework which integrates multiple linear regression (MLR) with long-short-term memory (LSTM) and its performance is compared with state-of-the-art models. The yield is forecasted from the current year through the next five years, up to 2029. The data used for the prediction model will be 1998 to 2023 from four districts of West Bengal and Uttar Pradesh. An important finding of the study was that it is possible to predict rice harvest five years in advance of actual harvest, providing useful information for agricultural decision making and planning. Researchers, policy makers and farmers can all benefit from better food security planning and resource management thanks to the study's findings, which shed light on the possibilities of combining remote sensing with biophysical parameters using ML models. Measures used to assess the suggested model include R2, RMSE, MAE, MSE, accuracy (Acc), F1 score (F1), recall (Re) and precision (Pe), among others. The suggested approach yields improved accuracy, R2, RMSE, MAE, and MSE values of 0.9823, 0.956, 0.1436, 0.021, and 0.198, respectively.
{"title":"A hybrid deep learning based framework for prediction of rice yield through integration of biophysical parameters and optical remote sensing data in India","authors":"Rafath Samrin , Pundru Chandra Shaker Reddy , K. Arun Kumar , Natha Deepthi , C. Mithra , S Bhargavi Latha , Sucharitha Yadala , Gopal Kumar Thakur","doi":"10.1016/j.jastp.2026.106734","DOIUrl":"10.1016/j.jastp.2026.106734","url":null,"abstract":"<div><div>India is the dominant player in the cultivation of rice around the world. Rice yield prediction can be considered as a problem that needs to be solved. Accurate and timely prediction of rice yield can provide meaningful benefits to crop yield. Time series models are widely used for rice yield prediction, but their accuracy remains inadequate. Despite their prominence, they often fail to deliver the required precision. This study considers one of the most practical machine learning (ML) methods for predicting rice yield, allowing forecasts for the next five years. The research carried out shows the rice yield prediction done using the hybrid framework which integrates multiple linear regression (MLR) with long-short-term memory (LSTM) and its performance is compared with state-of-the-art models. The yield is forecasted from the current year through the next five years, up to 2029. The data used for the prediction model will be 1998 to 2023 from four districts of West Bengal and Uttar Pradesh. An important finding of the study was that it is possible to predict rice harvest five years in advance of actual harvest, providing useful information for agricultural decision making and planning. Researchers, policy makers and farmers can all benefit from better food security planning and resource management thanks to the study's findings, which shed light on the possibilities of combining remote sensing with biophysical parameters using ML models. Measures used to assess the suggested model include R<sup>2</sup>, RMSE, MAE, MSE, accuracy (Acc), F1 score (F1), recall (Re) and precision (Pe), among others. The suggested approach yields improved accuracy, R<sup>2</sup>, RMSE, MAE, and MSE values of 0.9823, 0.956, 0.1436, 0.021, and 0.198, respectively.</div></div>","PeriodicalId":15096,"journal":{"name":"Journal of Atmospheric and Solar-Terrestrial Physics","volume":"279 ","pages":"Article 106734"},"PeriodicalIF":1.9,"publicationDate":"2026-02-01","publicationTypes":"Journal Article","fieldsOfStudy":null,"isOpenAccess":false,"openAccessPdf":"","citationCount":null,"resultStr":null,"platform":"Semanticscholar","paperid":"145979917","PeriodicalName":null,"FirstCategoryId":null,"ListUrlMain":null,"RegionNum":4,"RegionCategory":"地球科学","ArticlePicture":[],"TitleCN":null,"AbstractTextCN":null,"PMCID":"","EPubDate":null,"PubModel":null,"JCR":null,"JCRName":null,"Score":null,"Total":0}
Pub Date : 2026-02-01Epub Date: 2026-01-05DOI: 10.1016/j.jastp.2026.106724
Baris Baykant Alagoz , Cemal Keles , Burhan Baran
Global warming is a near-future major problem that has serious consequences on life on Earth. Accumulation of CO2 emission in the atmosphere becomes a dominating factor in the global warming process. This study introduces a global temperature uptrend model that considers effects of atmospheric CO2 buildup because of combustion of natural gas, oil and coal fuels. The model is based on earth's energy balance modeling and estimates impacts of fossil fuel-based CO2 emission on the average global temperature. To evaluate effects of fossil fuel reduction policies, this model is used to study yearly 3 %, 5 % and 7 % fossil fuel reduction scenarios on 2013–2022 global fuel consumption data. This backward projection indicated that the cumulative carbon dioxide emission of 10-year global fuel consumption has potential of about +0.023 °C/y global temperature rise between 2013 and 2022 years. Moreover, this model is also used for forward projections to discuss the effects of these 10-year-long fossil fuel reduction programs on future trends of global temperature anomaly.
{"title":"A note on effects of fossil fuel reduction policies on atmospheric carbon dioxide buildup and global warming","authors":"Baris Baykant Alagoz , Cemal Keles , Burhan Baran","doi":"10.1016/j.jastp.2026.106724","DOIUrl":"10.1016/j.jastp.2026.106724","url":null,"abstract":"<div><div>Global warming is a near-future major problem that has serious consequences on life on Earth. Accumulation of CO<sub>2</sub> emission in the atmosphere becomes a dominating factor in the global warming process. This study introduces a global temperature uptrend model that considers effects of atmospheric CO<sub>2</sub> buildup because of combustion of natural gas, oil and coal fuels. The model is based on earth's energy balance modeling and estimates impacts of fossil fuel-based CO<sub>2</sub> emission on the average global temperature. To evaluate effects of fossil fuel reduction policies, this model is used to study yearly 3 %, 5 % and 7 % fossil fuel reduction scenarios on 2013–2022 global fuel consumption data. This backward projection indicated that the cumulative carbon dioxide emission of 10-year global fuel consumption has potential of about +0.023 °C/y global temperature rise between 2013 and 2022 years. Moreover, this model is also used for forward projections to discuss the effects of these 10-year-long fossil fuel reduction programs on future trends of global temperature anomaly.</div></div>","PeriodicalId":15096,"journal":{"name":"Journal of Atmospheric and Solar-Terrestrial Physics","volume":"279 ","pages":"Article 106724"},"PeriodicalIF":1.9,"publicationDate":"2026-02-01","publicationTypes":"Journal Article","fieldsOfStudy":null,"isOpenAccess":false,"openAccessPdf":"","citationCount":null,"resultStr":null,"platform":"Semanticscholar","paperid":"145940744","PeriodicalName":null,"FirstCategoryId":null,"ListUrlMain":null,"RegionNum":4,"RegionCategory":"地球科学","ArticlePicture":[],"TitleCN":null,"AbstractTextCN":null,"PMCID":"","EPubDate":null,"PubModel":null,"JCR":null,"JCRName":null,"Score":null,"Total":0}
For solar photovoltaic and solar thermal systems to be effectively integrated into the energy grid, accurate prediction of solar radiation in a given area is essential. This prediction facilitates more effective planning, management, and optimization of energy production by utilities and renewable energy providers. From this vantage point, this article aims to forecast the daily global solar irradiation data of five cities in Cameroon (Bamenda, Bertoua, Ebolowa, Maroua, and Yaounde), which primarily differ in terms of solar irradiation distribution. In the study, ten different machine learning algorithms (Artificial Neural Network (ANN), Linear Regression (LR) algorithms, K-Nearest Neighbors (K-NN), Convolutional Neural Networks (CNNs), Random Forest (RF), Gradient Boosting Machine (GBM), Decision Tree (DT), Long-Short Term Memory (LSTM), Feed forward Neural Network (FNN), and Recurrent Neural Network (RNN)) are used. In the training of these algorithms, date, UT time, temperature, relative humidity, pressure, wind speed, wind direction, rainfall and solar irradiation of these cities are used. The data originate from the National Aeronautics and Space Administration and span the forty-one years (January 1, 1980, to December 31, 2021). Seven distinct statistical indicators are discussed to evaluate the effectiveness of these algorithms: t-statistic, Mean Absolute Percentage Error (MAPE), Maximum Absolute Bias Error (MABE), Means Bias Error (MBE), Root Mean Squared Error (RMSE), R-squared (R2) and relative Root Mean Squared Error (rRMSE). According to the findings, the R2, MAPE, and RMSE values of every algorithm range from 0.718 to 0.937, from 12.2 % to 25.9 %, and from 232 to 978 kJ/m2, respectively. When it came to R2 and MAPE metrics, LR consistently showed the worst performance, and the algorithms that surpassed the t-critic value were KNN, RF, and ANN. The current study concludes that, although each of the machine learning techniques investigated in this research have the ability of reliably forecasting data on global solar radiation, the KNN algorithm proves to be the most suitable choice. Next in order of precedence are RF, LSTM, ANN, GBM, CNN, RNN, FNN, DT, and LR.
{"title":"Prediction of Cameroon's global solar radiation using deep learning and machine learning algorithms","authors":"Fodoup Cyrille Vincelas Fohagui , Yemeli Wenceslas Koholé , Clint Ameri Wankouo Ngouleu , Donald Noutchogouin Tedom , Ghislain Tchuen","doi":"10.1016/j.jastp.2026.106733","DOIUrl":"10.1016/j.jastp.2026.106733","url":null,"abstract":"<div><div>For solar photovoltaic and solar thermal systems to be effectively integrated into the energy grid, accurate prediction of solar radiation in a given area is essential. This prediction facilitates more effective planning, management, and optimization of energy production by utilities and renewable energy providers. From this vantage point, this article aims to forecast the daily global solar irradiation data of five cities in Cameroon (Bamenda, Bertoua, Ebolowa, Maroua, and Yaounde), which primarily differ in terms of solar irradiation distribution. In the study, ten different machine learning algorithms (Artificial Neural Network (ANN), Linear Regression (LR) algorithms, K-Nearest Neighbors (K-NN), Convolutional Neural Networks (CNNs), Random Forest (RF), Gradient Boosting Machine (GBM), Decision Tree (DT), Long-Short Term Memory (LSTM), Feed forward Neural Network (FNN), and Recurrent Neural Network (RNN)) are used. In the training of these algorithms, date, UT time, temperature, relative humidity, pressure, wind speed, wind direction, rainfall and solar irradiation of these cities are used. The data originate from the National Aeronautics and Space Administration and span the forty-one years (January 1, 1980, to December 31, 2021). Seven distinct statistical indicators are discussed to evaluate the effectiveness of these algorithms: t-statistic, Mean Absolute Percentage Error (MAPE), Maximum Absolute Bias Error (MABE), Means Bias Error (MBE), Root Mean Squared Error (RMSE), R-squared (R<sup>2</sup>) and relative Root Mean Squared Error (rRMSE). According to the findings, the R<sup>2</sup>, MAPE, and RMSE values of every algorithm range from 0.718 to 0.937, from 12.2 % to 25.9 %, and from 232 to 978 kJ/m<sup>2</sup>, respectively. When it came to R<sup>2</sup> and MAPE metrics, LR consistently showed the worst performance, and the algorithms that surpassed the t-critic value were KNN, RF, and ANN. The current study concludes that, although each of the machine learning techniques investigated in this research have the ability of reliably forecasting data on global solar radiation, the KNN algorithm proves to be the most suitable choice. Next in order of precedence are RF, LSTM, ANN, GBM, CNN, RNN, FNN, DT, and LR.</div></div>","PeriodicalId":15096,"journal":{"name":"Journal of Atmospheric and Solar-Terrestrial Physics","volume":"279 ","pages":"Article 106733"},"PeriodicalIF":1.9,"publicationDate":"2026-02-01","publicationTypes":"Journal Article","fieldsOfStudy":null,"isOpenAccess":false,"openAccessPdf":"","citationCount":null,"resultStr":null,"platform":"Semanticscholar","paperid":"146034949","PeriodicalName":null,"FirstCategoryId":null,"ListUrlMain":null,"RegionNum":4,"RegionCategory":"地球科学","ArticlePicture":[],"TitleCN":null,"AbstractTextCN":null,"PMCID":"","EPubDate":null,"PubModel":null,"JCR":null,"JCRName":null,"Score":null,"Total":0}
扫码关注我们
求助内容：
应助结果提醒方式：
京公网安备 11010802042870号