Permutation Importance Research Articles

Robot Ground Media Classification Based on Hilbert–Huang Transform and Attention-Based Spatiotemporal Coupled Network

With the development of technology, mobile robots are increasingly deployed in real-world environments. To enable robots to work safely in a variety of terrain environments, we proposed a ground-type detection method based on the Hilbert–Huang transform (HHT) and attention-based spatiotemporal coupled network. Taking a dataset containing multiple sets of robot signals from a Kaggle competition as an example; we use the proposed method to classify the signals and thus achieve a terrain classification of the robot’s location. Firstly, the signal data were processed using the discrete wavelet transform for noise reduction, and all channels in the dataset were ranked by importance using the permutation importance method. Next, the instantaneous frequencies of the two most important channels were extracted using the HHT and added to the original dataset to expand the feature dimension. Then the features in the expanded dataset were extracted by the convolutional neural network, long short-term memory, and attention module. Afterward, the fully extracted features were passed into the fully connected layer for classification, and an average classification accuracy of 83.14% was obtained. The effectiveness of each part in our method was demonstrated using ablation experiments. Finally, we compared our method with some common methods in the field and found that our method obtained the highest classification accuracy, proving the superiority of the proposed method.

Constructing a Model of Poplus spp. Growth Rate Based on the Model Fusion and Analysis of Its Growth Rate Differences and Distribution Characteristics under Different Classes of Environmental Indicators

Poplar (Poplus spp.) is an important forest species widely distributed in China of great significance in identifying factors that clearly influence its growth rate in order to achieve effective control of poplar growth. In this study, we selected 16 factors, including tree size, competition, climate, location, topography, and soil characteristics, to construct linear regression (LR), multilayer perceptron (MLP), k-nearest neighbor regression (KNN), gradient boosting decision tree (GBDT), extreme gradient boosting (XGB), random forest (RF), and deep neural network (DNN) models based on the poplar growth rate. Using model fusion methods, the fitting accuracy and estimation capability were improved. The relative importance of each variable in estimating the poplar growth rate was analyzed using the permutation importance evaluation. The results showed the following: (1) the model fusion approach significantly improved the estimation accuracy of the poplar growth rate model with an R2 of 0.893; (2) the temperature and precipitation exhibited the highest importance in poplar growth; (3) the forest stand density, precipitation, elevation, and temperature had significant variations in growth rates among different-sized poplar trees within different ranges; (4) low-forest stand density, high-precipitation, low-elevation, and high-temperature environments significantly increased the poplar growth rate and had a larger proportion of large-sized individuals with high growth rates. In conclusion, environmental factors significantly influence poplar growth, and corresponding planting and protection measures should be tailored to different growth environments to effectively enhance the growth of poplar plantations.

Investigating the mental health of university students during the COVID-19 pandemic in a UK university: a machine learning approach using feature permutation importance

Mental wellbeing of university students is a growing concern that has been worsening during the COVID-19 pandemic. Numerous studies have gathered empirical data to explore the mental health impact of the pandemic on university students and investigate factors associated with higher levels of distress. While the online questionnaire survey has been a prevalent means to collect data, regression analysis has been observed a dominating approach to interpret and understand the impact of independent factors on a mental wellbeing state of interest. Drawbacks such as sensitivity to outliers, ineffectiveness in case of multiple predictors highly correlated may limit the use of regression in complex scenarios. These observations motivate the underlying research to propose alternative computational methods to investigate the questionnaire data. Inspired by recent machine learning advances, this research aims to construct a framework through feature permutation importance to empower the application of a variety of machine learning algorithms that originate from different computational frameworks and learning theories, including algorithms that cannot directly provide exact numerical contributions of individual factors. This would enable to explore quantitative impact of predictors in influencing student mental wellbeing from multiple perspectives as a result of using different algorithms, thus complementing the single view due to the dominant use of regression. Applying the proposed approach over an online survey in a UK university, the analysis suggests the past medical record and wellbeing history and the experience of adversity contribute significantly to mental wellbeing states; and the frequent communication with families and friends to keep good relationship as well as regular exercise are generally contributing to improved mental wellbeing.

Fast grading method based on data driven capacity prediction for high-efficient lithium-ion battery manufacturing

With the large-scale expansion of the battery market, the cost optimization of battery manufacturing has become a focus of attention. Among the complex production process of the battery, capacity grading requires a full discharge to measure the capacity and results in high cost. This study proposes a fast grading method in which the batteries are half discharged and graded according to the capacity predicted by a neural network. The prediction-based method takes half the time and saves about 37 % of the energy consumption. Twenty-three features are extracted as the initial features. The collinear features are screened-out, and three feature reduction methods are compared. Permutation importance can effectively clarify the nonlinear relationship and determine the critical features. The root-mean-square error of the testing set is 0.18 %. The method is feasible and has the potential for further evolution. The electrochemical mechanism involved in the features is analyzed by simulation. Under the guidance of the mechanism, the method can be transferred to other batteries.

Mid-Infrared Variable Selection for Soil Organic Matter Fractions Based on Soil Model Systems and Permutation Importance Algorithm.

In this research, an attempt was made to classify soil samples according to the different fractions of soil organic matter (SOM) using model systems in which the ratio of the fractions of SOM is chemically mimicked. A mixture of starch and nicotinamide was used for the labile organic matter model, while a standard of humic acid was used for the stabile organic matter. Changing the threshold value in the selected ranges after a permutation importance algorithm is conducted using train models and test data set, a list of selected important wavelengths and their importance scores were obtained. Three regions for the classification of soil fractions within the estimated probability density function are most prominent: 800-1200 cm-1, 0.48-0.55; 1800-2000 cm-1, 0.52-0.62; and 2500-3200 cm-1, 0.48-0.62, where the first component represents the spectral range while the second component covers the range of the importance score. Obtained wavelength ranges indicate the importance of the aliphatic stretching and bending vibration region, as well as the total soil reflectance (mineral content) for the characterization of organic matter fractions. A comparative evaluation with literature data found that the obtained wavelengths have a potential for application in methods of proximal and remote detection/calibration of existing and development of new sensors for Advanced Spaceborne Thermal Emission and Reflection Radiometer satellites, specifically in the shortwave infrared and thermal infrared ranges.

An extended car-following model considering backward-looking effect: A machine learning approach

Most car-following models have mainly focused on the effects of the lead vehicle on the target vehicle's behaviour or the driver's forward-looking effects, but not the effects of the vehicle behind the target vehicle (the following vehicle) or the driver's backward-looking effects. Therefore, this study proposes a data-driven car-following model that incorporates both backward- and forward-looking effects using a deep neural network (DNN). This model is called the “DNN with backward-looking effect (DNN-BE) model”. The DNN-BE model produced higher prediction accuracy than the DNN model with forward-looking effects only and a conventional mathematical car-following model that considers both forward- and backward-looking effects. It was found that the target vehicle is more likely to accelerate when the spacing with the following vehicle is shorter and the spacing with the lead vehicle is longer. The result of permutation importance also shows that variables related to the following vehicle are more important when the spacing with the following vehicles is shorter.

Unravelling the effects of dynamic urban thermal environment on utility-scale floating photovoltaic electricity generation

The utility-scale photovoltaic systems have been widely integrated for sustainable urban development. However, the varying photovoltaic conversion efficiency (PVCE) affected by dynamic urban thermal environment causes uncontrollable uncertainty in electricity generation, which challenges installed-capacity planning and load-balancing operations. To tackle this problem, this study develops a PVCE estimation model containing four hierarchical modules. First, photovoltaic surface temperatures (PVSTs) are retrieved from satellite imagery, and meteorological features are collected to represent the dynamic thermal environment. Second, the contribution of each feature to the PVST estimation is evaluated, referring to the Mean Decrease in Impurity, Permutation Importance, and SHapley Addictive exPlanations. Third, machine learning models using Support Vector Machine, Random Forest, and XGBoost are developed to establish robust regressions between PVSTs and the selected features, enabling an accurate estimation of PVST spatio-temporal heterogeneity. Finally, the spatio-temporally corresponding PVCEs are calculated, resulting in a refined estimation of annual electricity generation. The investigation of four floating PV systems in Singapore found that their PVCEs vary insignificantly throughout the year, which is probably because of its stable climate and the cooling effect of water. The proposed model is simple and effective, demonstrating its impact on PV potential estimation when the urban thermal effect becomes significant.

Predicting Suitable Habitats for China’s Endangered Plant Handeliodendron bodinieri (H. Lév.) Rehder

Demarcating a plant species’ actual and potential biogeographical distribution is crucial for understanding the key environmental variables shaping its habitat conditions. We used MaxEnt and species distribution modeling to predict the likely range of China’s endangered species, Handeliodendron bodinieri (H. Lév.) Rehder, based on forty-four validated distribution records and eight selected environmental variables. Combined with percentage contribution and permutation importance, the jackknife statistical method was applied to test and evaluate pertinent factors restricting the potential distribution of H. bodinieri. The response curves of critical bioclimatic factors were employed to determine the potential species range. The generated MaxEnt model was confirmed to have excellent simulation accuracy. The current core potential distribution areas are concentrated in the Guangxi and Guizhou provinces of Southwest China, with a significant inter-regional difference. The precipitation of the warmest quarter (Bio18) and minimum temperature of the coldest month (Bio6) had the greatest impact on the distribution area of H. bodinieri. The findings could provide useful information and a reasonable reference for managers to enhance the protection of this declining species.

Exploring the variable importance in random forests under correlations: a general concept applied to donor organ quality in post-transplant survival

Random Forests are a powerful and frequently applied Machine Learning tool. The permutation variable importance (VIMP) has been proposed to improve the explainability of such a pure prediction model. It describes the expected increase in prediction error after randomly permuting a variable and disturbing its association with the outcome. However, VIMPs measure a variable’s marginal influence only, that can make its interpretation difficult or even misleading. In the present work we address the general need for improving the explainability of prediction models by exploring VIMPs in the presence of correlated variables. In particular, we propose to use a variable’s residual information for investigating if its permutation importance partially or totally originates from correlated predictors. Hypotheses tests are derived by a resampling algorithm that can further support results by providing test decisions and p-values. In simulation studies we show that the proposed test controls type I error rates. When applying the methods to a Random Forest analysis of post-transplant survival after kidney transplantation, the importance of kidney donor quality for predicting post-transplant survival is shown to be high. However, the transplant allocation policy introduces correlations with other well-known predictors, which raises the concern that the importance of kidney donor quality may simply originate from these predictors. By using the proposed method, this concern is addressed and it is demonstrated that kidney donor quality plays an important role in post-transplant survival, regardless of correlations with other predictors.

Can local explanation techniques explain linear additive models?

Local model-agnostic additive explanation techniques decompose the predicted output of a black-box model into additive feature importance scores. Questions have been raised about the accuracy of the produced local additive explanations. We investigate this by studying whether some of the most popular explanation techniques can accurately explain the decisions of linear additive models. We show that even though the explanations generated by these techniques are linear additives, they can fail to provide accurate explanations when explaining linear additive models. In the experiments, we measure the accuracy of additive explanations, as produced by, e.g., LIME and SHAP, along with the non-additive explanations of Local Permutation Importance (LPI) when explaining Linear and Logistic Regression and Gaussian naive Bayes models over 40 tabular datasets. We also investigate the degree to which different factors, such as the number of numerical or categorical or correlated features, the predictive performance of the black-box model, explanation sample size, similarity metric, and the pre-processing technique used on the dataset can directly affect the accuracy of local explanations.

Machine learning model for predicting late recurrence of atrial fibrillation after catheter ablation

Late recurrence of atrial fibrillation (LRAF) in the first year following catheter ablation is a common and significant clinical problem. Our study aimed to create a machine-learning model for predicting arrhythmic recurrence within the first year since catheter ablation. The study comprised 201 consecutive patients (age: 61.8 ± 8.1; women 36%) with paroxysmal, persistent, and long-standing persistent atrial fibrillation (AF) who underwent cryoballoon (61%) and radiofrequency ablation (39%). Five different supervised machine-learning models (decision tree, logistic regression, random forest, XGBoost, support vector machines) were developed for predicting AF recurrence. Further, SHapley Additive exPlanations were derived to explain the predictions using 82 parameters based on clinical, laboratory, and procedural variables collected from each patient. The models were trained and validated using a stratified fivefold cross-validation, and a feature selection was performed with permutation importance. The XGBoost model with 12 variables showed the best performance on the testing cohort, with the highest AUC of 0.75 [95% confidence interval 0.7395, 0.7653]. The machine-learned model, based on the easily available 12 clinical and laboratory variables, predicted LRAF with good performance, which may provide a valuable tool in clinical practice for better patient selection and personalized AF strategy following the procedure.

Interpretable high-stakes decision support system for credit default forecasting

Methods for forecasting credit default have long been research focus for financial institutions. In this study, we propose an interpretable high-stakes decision support system for credit default forecasting called CDFS. Because of the high-stake nature of credit default prediction, the proposed CDFS adheres to the principle of "people in the loop." The proposed CDFS comprises six modules: data processing, feature selection, data balancing, forecasting, evaluation, and interpretation. A feature selection method (permutation importance method), nine resampling methods, and six high-performance forecasting methods were employed in the proposed CDFS. The China Taiwan credit card default dataset and South Germany credit dataset were used to test the interpretability and predictive performance of the proposed CDFS. Experiments showed that the feature selection and data balancing modules of the CDFS effectively improve the prediction performance. A comparison with traditional logistic regression models demonstrated that the CDFS can provide decision-makers with satisfactory explanations for prediction results. In summary, the CDFS proposed in this study exhibited excellent predictive performance and satisfactory interpretability. This study contributes to improving the accuracy of credit default forecasting and reducing credit risk in financial institutions.

Hybrid machine learning model for hourly ozone concentrations prediction and exposure risk assessment

With the development and improvement of optimization algorithms, machine learning models have become more important in air quality prediction. These models can be used to address problems associated with atmospheric environmental pollutants like ozone and facilitate the development of appropriate control policies. Firstly, the data sets of meteorological and pollution in Nanjing and four nearby cities from 2018 to 2021 were divided into four sections to build mixed models. The extreme gradient boosting (XGBoost) algorithm was improved by the permutation importance method (PIM) and Pearson correlation coefficient to acquire the most important features of the seasonal model and analysis of the effect on ozone prediction. The results reflect that the NO2, PM2.5, and PM10 parameters have an important influence, accounting for an average of 23% of prediction, and the average influence of the meteorological parameters of the solar radiation exceeds 11%. Secondly, the particle swarm optimization (PSO), grey wolf optimizer (GWO), and ant lion optimizer (ALO) algorithms were used to optimize the SVR model parameters to predict hourly ozone concentration in January, April, July, and September 2022. The R2 values are 0.95, 0.92, 0.85, and 0.84, respectively, and the RMSE values are 5.2, 12.4, 18.0 and 16.6 μg/m3, respectively. Finally, the impact of hourly ozone concentration on population health was studied by a risk assessment of the population density-weighted pollution exposure. The result shows that the influence on humans gradually decreased yearly from 2018 to 2021, the average variation from 2.6% to −1.1%, and health risks during the evening rush hours are still prevalent and require further attention.

Skin cancer classification using explainable artificial intelligence on pre-extracted image features

Skin cancer is the most common type of cancer worldwide, affecting a large population recently. To date, various machine learning techniques exploiting skin images have been applied directly to skin cancer classification, showing promising results in improving diagnostic accuracy. This study aims to develop a machine learning-based model capable of accurately classifying skin cancer by utilizing extracted features from preprocessed images in the publicly available PH² dataset. Preprocessed features are known to provide more significant information than raw image data, as they capture specific characteristics of the images that are relevant to the classification task. The proposed model of this study can identify the most pertinent information in the images more accurately, thereby improving the performance and interpretability of the machine learning classification. Our simulation results illustrate that employing XG-boost yields an accuracy of 94% and an area under the curve value of 0.9947, further indicating that the proposed technique effectively distinguishes between non-melanoma and melanoma skin cancer. Explainable artificial intelligence provides some explanations by leveraging model-agnostic methods such as partial dependence plot, permutation importance, and SHAP. Moreover, the explainable artificial intelligence results show that asymmetry and pigment network features are the most important feature in the classification of skin cancer. These specific characteristics emerge as the most influential factors in distinguishing between different types of skin cancer.

Coarse Aggregate Shape Classification Method Based on Per-Optuna-LightGBM Model

To improve the detection level of aggregate shape for automated road use, Per-Optuna-LightGBM model for aggregate shape classification is proposed. Collect aggregate images using industrial camera and extract 48 morphological feature parameters. A feature importance analysis method based on Spearman Correlation and Permutation Importance is proposed to remove redundant factors and select the feature parameters of aggregate morphology. Based on cross-validation, an optimized Optuna-LightGBM model is trained based on the constructed dataset. Compared with GS-XGBoost algorithm, the Optuna-LightGBM model can classify the shape of aggregates more accurately and efficiently. The accuracy value of the proposed model is 82.5%, which increased by 4% compared to before optimization. The proposed model can efficiently classify the shape of aggregates which meet the design requirements, also provide a certain foundation for automated classification of aggregate shapes.

The relationship and predictive value of dementia and frailty for mortality in patients with surgically managed hip fractures

BackgroundBoth dementia and frailty have been associated with worse outcomes in patients with hip fractures. However, the interrelation and predictive value of these two entities has yet to be clarified. The current study aimed to investigate the predictive relationship between dementia, frailty, and in-hospital mortality after hip fracture surgery.MethodsAll patients registered in the 2019 National Inpatient Sample Database who were 50 years or older and underwent emergency hip fracture surgery following a traumatic fall were eligible for inclusion. Logistic regression (LR) models were constructed with in-hospital mortality as the response variables. One model was constructed including markers of frailty and one model was constructed excluding markers of frailty [Orthopedic Frailty Score (OFS) and weight loss]. The feature importance of all variables was determined using the permutation importance method. New LR models were then fitted using the top ten most important variables. The area under the receiver-operating characteristic curve (AUC) was used to compare the predictive ability of these models.ResultsAn estimated total of 216,395 patients were included. Dementia was the 7th most important variable for predicting in-hospital mortality. When the OFS and weight loss were included, they replaced dementia in importance. There was no significant difference in the predictive ability of the models when comparing the model that included markers of frailty [AUC for in-hospital mortality (95% CI) 0.79 (0.77–0.81)] with the model that excluded markers of frailty [AUC for in-hospital mortality (95% CI) 0.79 (0.77–0.80)].ConclusionDementia functions as a surrogate for frailty when predicting in-hospital mortality in hip fracture patients. This finding highlights the importance of early frailty screening for improvement of care pathways and discussions with patients and their families in regard to expected outcomes.

Fracture toughness evaluation of ground granulated blast furnace slag concrete using experimental study and machine learning techniques

The relationship between mix design and fracture properties of ground granulated blast furnace slag (GGBFS) concrete is complex and influenced by many factors, such as mixture composition. This study evaluates the fracture toughness under mode-I, mode-II, and mixed-mode, performed on centrally straight-notched Brazilian discs. An experimental program generated an extensive database of the fracture toughness of GGBFS mortar, consisting of 143 specimens with different mix designs under different loading angle conditions. Then, a combination of support vector regression (SVR) algorithm and genetic algorithm (GA) as hyperparameter tuner was conducted to develop models for fracture toughness. The SVR algorithm was trained with Radial Basis Function (RBF), polynomial, sigmoid, and linear kernel functions. Based on the Taylor diagram, the models demonstrated robust performance in predicting fracture toughness. The performance metrics indicate that the RBF function is the most reliable model, with R2-score 0.95 and 0.94 for KIC and KIIC, respectively. The importance score of input features was determined with the permutation importance function. The angle between the crack and loading direction is the most influential parameter in the models.

Prediction of transportation energy demand in Türkiye using stacking ensemble models: Methodology and comparative analysis

The transportation sector accounts for 61.5% of global oil consumption and is responsible for 29% of the world’s total energy demand. Passenger transportation utilizes around 50%–60% of the energy used for transportation-related activities. Accurate prediction of future transportation energy consumption is essential for governments to make well-informed decisions regarding transportation infrastructure development and utilization, which supports the United Nations’ Sustainable Development Goals (SDGs) and advances the shift to a net-zero carbon economy. With the expected increase in population, vehicles, and economic growth, it is essential to predict the energy demand to ensure sustainable urban transportation. This is crucial not only for economic prosperity but also for promoting human health and mitigating carbon emissions. Therefore, transportation energy demand prediction plays a vital role in designing sustainable future urban transportation and making informed energy investment and policy decisions. This study proposes a novel methodology and investigates for the application of machine learning stacking ensemble method with hyperparameter tuning and multicollinearity removal to predict transportation energy demand in Turkey based on historic data from 1975–2019. The dataset includes GDP, year, vehicle miles traveled, population, oil price, passenger miles traveled, and ton-miles traveled as features. A performance evaluation and comparison of 19 machine learning algorithms is first carried out to find the best candidate for the stacking ensemble models, including eXtreme Gradient Boosting algorithm. This performance comparison uses all features and also only two of them during the training phase, and it takes into consideration a 4-fold cross-validation. A combination of permutation importance and hierarchical clustering algorithm on the Spearman rank-order correlations is used for dimensionality reduction of the dataset. Extra Tree Regressor and ADABoost Regressor, which are both placed in the second level of the suggested models, are two meta-regressors that are proposed for stacking ensembles because they perform better compared to single machine learning algorithm. In total, eight stacking ensemble models – four for each of the meta-regressors – were developed and investigated considering all features and only two of them separately. Six metrics – R-squared, MSE, MAE, RMSE, RMSLE, and MAPE – are used to assess all models. The Extra Trees Regressor can be used as a meta-regressor in the best proposed stacking ensemble model to predict the energy demand for transportation. This model achieves an R-squared value of approximately 0.99 when all the features are taken into consideration. When only two features from the dataset are considered the same stacking ensemble model can achieve an accuracy of 0.98. These findings have the potential to contribute to the development of more accurate models and results, which can, in turn, lead to improved strategies for managing future transportation energy demand. Additionally, this research can support the advancement of alternative technologies that promote sustainable urban development, ultimately helping to move towards a net-zero carbon economy.

Application, interpretability and prediction of machine learning method combined with LSTM and LightGBM-a case study for runoff simulation in an arid area

The runoff prediction can provide scientific basis for flood control, disaster reduction and water resources planning. Due to a large number of uncertainties in runoff prediction, it is difficult to make precise predictions. To improve the accuracy of runoff prediction, this study combines techniques of Long Short-Term Memory (LSTM) and Light Gradient Boosting Machine (LightGBM) in machine learning with reciprocal error method to develop an integrated data-driven model (i.e., LSTM-LightGBM) for runoff prediction. To demonstrate its applicability, the model is applied to the annual runoff prediction of the Caiqi hydrological monitoring station in the Shiyang River in an arid area. Indicators include Error of Peak (EP), Nash-Sutcliffe Efficiency (NSE), Root Mean Squared Error (RMSE), and Mean Absolute Error (MAE) are adopted to evaluate the prediction performance of the LSTM, LightGBM, and LSTM-LightGBM methods under the same hyperparameter combinations. Then, the interpretability of LSTM and LightGBM models is also explored based on the permutation importance method and Shapley Additive exPlanations (SHAP) values, respectively. Finally, future annual runoff at the Caiqi for the next 50 years (2025–2075) is predicted based on LSTM-LightGBM model under 12 climate scenarios. Therefore, results show that: 1. the integrated model (LSTM-LightGBM) has good performance than two single models in NSE (0.92), RMSE (0.075 million m3) and MAE (0.046 million m3) and EP value (i.e., for bridging the peak-valley runoff). 2. In this case, it is found that four feature variables have the greatest influence on the target variables through the interpretable analysis. 3. The 12 combined climate scenarios used in this investigation produced generally steady predictions. The scenarios with the highest and lowest mean values are GFDL RCP 6.0 (3.12 × 108m3) and IPSL RCP 2.6 (3.04 × 108m3), respectively, with a decrease of 24.09 % and 26.03 % compared to the mean annual runoff of 4.11 × 108m3 in the baseline period (1955–2017). These findings can provide scientific bases for future water resources planning in the downstream of the Shiyang River Basin.

Research on predicting the diffusion of toxic heavy gas sulfur dioxide by applying a hybrid deep learning model to real case data

Toxic heavy gas sulfur dioxide (SO2) is a specific life and environmental hazard. Predicting the diffusion of SO2 has become a research focus in fields such as environmental and safety studies. However, traditional methods, such as kinetic models, cannot balance precision and time. Thus, they do not meet the needs of emergency decision-making. Deep learning (DL) models are emerging as a highly regarded solution, providing faster and more accurate predictions of gas concentrations. To this end, this study proposes an innovative hybrid DL model, the parallel-connected convolutional neural network-gated recurrent unit (PC CNN-GRU). This model utilizes two CNNs connected in parallel to process gas release and meteorological datasets, enabling the automatic extraction of high-dimensional data features and handling of long-term temporal dependencies through the GRU. The proposed model demonstrates good performance (RMSE, MAE, and R2 of 20.1658, 10.9158, and 0.9288, respectively) with real data from the Project Prairie Grass (PPG) case. Meanwhile, to address the issue of limited availability of raw data, in this study, time series generative adversarial network (TimeGAN) are introduced for SO2 diffusion studies for the first time, and their effectiveness is verified. To enhance the practicality of the research, the contribution of drivers to SO2 diffusion is quantified through the utilization of the permutation importance (PIMP) and Sobol’ method. Additionally, the maximum safe distance downwind under various conditions is visualized based on the SO2 toxicity endpoint concentration. The results of the analyses can provide a scientific basis for relevant decisions and measures.