Urban carbon emission intelligent prediction method based on multi-modal data and machine learning integration

Abstract

The invention relates to the technical field of environment monitoring, in particular to an urban carbon emission intelligent prediction method based on multi-modal data and machine learning integration, which comprises the following steps: feature acquisition and preprocessing: acquiring multi-source heterogeneous data and processing the multi-source heterogeneous data into a standardized data set; constructing an LSTM (Long Short Term Memory) and XGBoost hybrid prediction model; predicting future data on the basis of existing data, and predicting input features of future time steps of the weight integration prediction model on the basis of historical data in a standardized data set; and predicting carbon emission in the future. Industrial structure indexes, environment variables and social and economic substitution indexes are integrated through time characteristic engineering, and emission driving factors in different urban environments can be comprehensively captured; a recursive prediction mechanism is established, development trajectories of different cities are considered, reliable emission prediction can be carried out, meanwhile, time consistency is kept, and prediction uncertainty is quantified; interference influences under different policies are quantified, and a quantitative basis is provided for carbon management decisions.

Description

Technical Field

[0001] This invention relates to the field of environmental monitoring technology, specifically to an intelligent prediction method for urban carbon emissions based on the integration of multimodal data and machine learning. Background Technology

[0002] With the increasing severity of global climate change, accurate carbon emission prediction and management have become crucial for cities to achieve carbon neutrality goals. However, existing technologies have the following problems:

[0003] Limitations of a single model: Traditional methods often employ a single statistical regression model or neural network model, which cannot effectively capture time series dependencies and complex nonlinear feature interactions simultaneously, resulting in limited prediction accuracy.

[0004] Feature engineering is inadequate: Existing technologies lack a systematic multi-dimensional feature construction mechanism and fail to fully utilize the synergistic effect of multi-source heterogeneous data from industries, environments, and socio-economic sectors.

[0005] Insufficient consideration of urban differences: Existing methods lack personalized prediction mechanisms for different urban development characteristics, making it difficult to accurately reflect the development differences and emission patterns between cities.

[0006] Weak long-term forecasting ability: Traditional methods suffer from error accumulation in multi-step forecasting, and cannot provide reliable medium- and long-term forecasting results.

[0007] Lack of policy scenario analysis: Existing technologies lack an effective mechanism for evaluating the effects of policy interventions, and cannot provide quantitative support for policy formulation. Summary of the Invention

[0008] To address the shortcomings of existing technologies, the purpose of this invention is to provide an intelligent prediction method for urban carbon emissions based on the integration of multimodal data and machine learning.

[0009] The technical solution adopted by this invention to solve its technical problem is:

[0010] A smart prediction method for urban carbon emissions based on multimodal data and machine learning integration, the method comprising:

[0011] S1. Feature Acquisition and Preprocessing: Collect multi-source heterogeneous data and process it into a standardized dataset, specifically:

[0012] Data sources include carbon emission data, nighttime light data, socioeconomic statistics, and normalized difference vegetation index (NDVI) data.

[0013] Data processing: The collected data sources are cleaned, time features are constructed, spatial features are encoded, and Z-score is standardized. The training set and test set are divided using a time series cross-validation strategy, and a standardized dataset is output.

[0014] The S2, LSTM, and XGBoost hybrid prediction model is constructed by training the hybrid prediction model based on the standardized dataset output by S1, specifically as follows:

[0015] Model architecture construction: A city-embedded enhanced LSTM network and an XGBoost regressor are built. The LSTM network is input with continuous sequence data to capture time dependencies, and the XGBoost regressor is trained on flattened sequence features to model nonlinear feature interactions.

[0016] Differentiated feature strategy: LSTM: uses 6-dimensional city embedding (continuous vector), XGBoost uses 11-dimensional One-Hot encoding (discrete binary); Feature composition: 60-dimensional flattened temporal features (3 years × 20 features), 11-dimensional city One-Hot, total dimension: 71; Hyperparameters: n_estimators=100, max_depth=3, learning_rate=0.1;

[0017] Training optimization: The weighted mean squared error loss function is adopted, the Adam optimizer is used, and L2 weight decay regularization, ReduceLROnPlateau learning rate scheduling strategy, Epoch patience value early stopping mechanism and batch training mechanism are combined to optimize training.

[0018] Dynamic weighted ensemble: The weights of the LSTM network and the XGBoost regressor are calculated based on the R² scores of the LSTM network and the XGBoost regressor to obtain the weighted ensemble prediction model;

[0019] S3. Predict future data based on existing data. Using historical data from the standardized dataset in S1 as a basis, predict the weighted integration of the input features of the prediction model for future time steps, specifically:

[0020] Initialization: Traverse all unique cities in the standardized dataset, sort the historical basic feature data of each city by year, select the historical year data for each basic feature, and calculate the historical trend, including linear trend, volatility, average value and annual growth rate, and finally store it as a city trend dictionary;

[0021] Recursive prediction: Receives the target city, the scaled feature value of the basic features at the current time step, the feature scaler, the prediction step size, and the scenario factor. It inversely scales the scaled feature value of the basic features at the current time step to the true value. Based on the scaled true value, it obtains the corresponding historical trend information from the city trend dictionary, adjusts the linear trend using the scenario factor, and calculates the trend component, growth component, and random noise to finally obtain the predicted true value. The predicted true value is then rescaled using the feature scaler, and the time features are updated based on the rescaled predicted true value to obtain the future contextualized input feature sequence.

[0022] S4. Predict future carbon emissions, specifically:

[0023] The future contextualized input feature sequence obtained in S3 is input into the weighted ensemble prediction model trained in S2, and the target city carbon emission prediction results for the future time step are output.

[0024] As a preferred embodiment, a further technical solution of the present invention is:

[0025] Preferably, carbon emission data is obtained by estimating city-level carbon dioxide emissions within a specified year range based on the Global Atmospheric Research Emissions Database.

[0026] Nighttime light data: By comprehensively utilizing satellite imagery data and employing remote sensing data fusion and calibration methods, continuous and calibrated nighttime light data is constructed.

[0027] Socioeconomic statistics include the output value of the secondary industry and its share of GDP. The output value of the secondary industry is a direct indicator of industrial scale and is closely related to energy consumption and carbon emissions. The share of the secondary industry in GDP reflects the regional economic structure and is used to capture the long-term impact of industrial restructuring on carbon emission intensity.

[0028] Normalized Difference Vegetation Index (NDVI): Based on surface reflectance data from satellite imagery, the city-level NDVI is calculated using a median composite method to mitigate the impact of seasonal fluctuations and to exclude cloud-contaminated pixels using quality control bands.

[0029] Preferably, data cleaning involves using interpolation to fill in missing values, detect and handle anomalies in multi-source heterogeneous data, and ensure data integrity and accuracy.

[0030] Time feature construction: In order to capture short-term fluctuations and long-term trends in carbon emissions, including lag features, 1-year and 2-year lag values ​​are used to reflect short-term dependence; first-order difference features are used to capture interannual variability; and 3-year moving average features are used to smooth the data and reveal long-term trends.

[0031] Spatial feature encoding: City identity is encoded using city embedding technology, mapping city ID to a six-dimensional learning space vector, enabling the model to learn emission patterns of specific cities while maintaining cross-city generalization ability;

[0032] Z-score standardization: Scaling the feature values ​​of the data to a range with a mean of 0 and a standard deviation of 1 to eliminate dimensional differences.

[0033] Preferably, the weighted mean square error loss function in S2 is expressed as follows:

[0034] ;

[0035] in, This represents the weighted loss value. Represents the total number of samples. This represents the weight assigned to sample i. This represents the actual carbon emissions of sample i. This represents the predicted carbon emission value of sample i. In the case of a high emission scenario where the predicted carbon emission value after normalization is greater than 0.5, the weight assigned to the sample is set to 2, otherwise it is 1, to ensure that the model focuses on the high emission scenario.

[0036] Preferably, in S3, the historical year data is selected from the most recent min(3, len(city_data)) years for analysis to capture recent development dynamics, where len(city_data) represents the number of years the data has been statistically available;

[0037] Linear trend calculation is performed using the following formula:

[0038] ;

[0039] in, This represents the linear trend of sample i. This indicates that the latest value of sample i is typically the average of the past 3 years. Indicates the initial value. and For the corresponding year;

[0040] Volatility calculation: used to measure the dispersion of characteristic value changes, standardized by standard deviation / (absolute value of mean + 1e-6).

[0041] Preferably, in S3, the predicted true value is calculated using the following expression:

[0042] ;

[0043] Where k represents the prediction step size; Represents the observation value at time t; Indicates trend components, The linear trend adjusted based on scenario factors is expressed as follows:

[0044] , Indicates situational factors;

[0045] The growth component is represented as:

[0046] , The growth rate of the index is determined based on historical data;

[0047] Noise is represented as:

[0048] , For volatility, the noise is a normally distributed random noise with a mean of 0 and a standard deviation that is related to volatility and the latest observation.

[0049] Preferably, the scenario factors are set with ranges based on different fundamental characteristics to reflect development trends under different policy interventions, including:

[0050] Slowing Industrial Growth Scenario: The secondary industry output scenario factor is 0.8-0.9, simulating a slowdown in industrial growth;

[0051] The scenario factor for the proportion of secondary industry output is 0.96-0.98, simulating an average annual decrease in the share of industrial structure.

[0052] Ecological enhancement scenario: The normalized vegetation index scenario factor is 1.05-1.1, simulating ecological enhancement;

[0053] The nighttime light scene factor is 1.02-1.05, simulating an increase in nighttime light intensity caused by urban renewal, infrastructure improvement, and increased economic activity;

[0054] Baseline Scenario: The baseline scenario factor is 1, simulating a continuation of historical trends;

[0055] Green transition scenario: A comprehensive adjustment of at least two scenario factors;

[0056] Deep decarbonization scenario: emission reduction measures based on multi-dimensional scenario factors.

[0057] Preferably, the carbon emission prediction results under different scenario factors are compared with the carbon emission prediction results under the baseline scenario factors. By calculating the percentage or absolute amount of reduction in carbon emission prediction results under different scenario factors relative to the carbon emission prediction results under the baseline scenario factors, the emission reduction effect of various policy interventions can be accurately evaluated.

[0058] Hierarchical clustering and K-means algorithms were used to classify cities, and principal component analysis was used for verification. Finally, the cities were divided into city groups with different carbon emission characteristics.

[0059] For each city cluster, multi-dimensional decision-making objectives are set, and optimization algorithms are used to explore the optimal policy combination for each city cluster to achieve the decision-making objectives. The decision-making objectives include, but are not limited to: minimizing carbon emissions under a specific economic growth target and minimizing economic impact under a given carbon emission reduction target. The optimization algorithms include, but are not limited to: genetic algorithm and simulated annealing algorithm.

[0060] Preferably, the city embedding enhanced LSTM network includes a city embedding layer, a feature fusion layer, and a two-layer unidirectional LSTM structure. The city embedding layer maps the city ID to a 6-dimensional continuous vector space, and the embedding matrix weights are automatically optimized during training. The feature fusion layer concatenates the city embedding vector with the original features at each time step, so that the LSTM input dimension = the original feature dimension + 6. In the two-layer unidirectional LSTM structure, each LSTM layer contains 256 hidden units. A Dropout rate of 0.3 is applied between layers and before the fully connected output layer. The LSTM network inputs continuous three-year sequence data to capture temporal dependence.

[0061] The XGBoost regressor employs differential feature allocation, using 60-dimensional flattened temporal features and 11-dimensional One-Hot encoded city features to form 71-dimensional input features. The decision tree algorithm can directly construct splitting conditions through One-Hot encoded city features.

[0062] The LSTM network was trained using an Adam optimizer with an initial learning rate of 0.0005, combined with a 1×10⁻⁶... -5 The L2 weight decay regularization, the ReduceLROnPlateau learning rate scheduling strategy with a learning rate decay factor of 0.5 and a patience value of 3 epochs, the early stopping mechanism with a patience value of 10 epochs, and the batch training mechanism with a batch size of 64.

[0063] The XGBoost regressor is configured with hyperparameters n_estimators = 100, max_depth = 3, and learning_rate = 0.1.

[0064] Preferably, the process of updating time features based on the rescaled predicted true values ​​in S3 includes:

[0065] 1-year lag characteristic = predicted true value at the previous time step;

[0066] 2-year lag characteristic = predicted true value two time steps ago;

[0067] Differential feature = predicted true value at the current time step - predicted true value at the previous time step;

[0068] Moving average feature = the mean calculated using a sliding window with a maintenance time step of 3;

[0069] The updated time features are standardized and a sequence window is slidable to add the new feature vector to the end of the sequence, while the earliest time step feature sequence is removed, keeping the feature sequence length at 3.

[0070] Compared with the prior art, the present invention, which adopts the above technical solution, has the following outstanding features: it integrates industrial structure indicators, environmental variables, and socio-economic alternative indicators through time feature engineering, which can comprehensively capture emission driving factors in different urban environments; it establishes a recursive prediction mechanism that considers the development trajectory of different cities, enabling reliable emission prediction while maintaining time consistency and quantifying prediction uncertainty; and it quantifies the intervention impact under different policies, providing a quantitative basis for carbon management decisions. Attached Figure Description

[0071] Figure 1 This is a schematic diagram of the method flow in an embodiment of the present invention. Detailed Implementation

[0072] The present invention will be further illustrated below with reference to specific embodiments. The purpose of this illustration is solely to provide a better understanding of the invention. Therefore, the examples given do not limit the scope of protection of the present invention.

[0073] like Figure 1 This embodiment presents a smart prediction method for urban carbon emissions based on multimodal data and machine learning integration. The method includes:

[0074] S1. Feature Acquisition and Preprocessing: Collect multi-source heterogeneous data and process it into a standardized dataset, specifically:

[0075] Data sources include carbon emission data, nighttime light data, socioeconomic statistics, and normalized difference vegetation index (NDVI) data.

[0076] Data processing: The collected data sources are cleaned, time features are constructed, spatial features are encoded, and Z-score is standardized. The training set and test set are divided using a time series cross-validation strategy, and a standardized dataset is output.

[0077] The S2, LSTM, and XGBoost hybrid prediction model is constructed by training the hybrid prediction model based on the standardized dataset output by S1, specifically as follows:

[0078] Model architecture construction: A city-embedded enhanced LSTM network and an XGBoost regressor are built. The LSTM network is input with continuous sequence data to capture time dependencies, and the XGBoost regressor is trained on flattened sequence features to model nonlinear feature interactions.

[0079] Differentiated feature strategies: LSTM: uses 6-dimensional city embedding (continuous vector); XGBoost uses 11-dimensional One-Hot encoding (discrete binary). Feature composition: 60-dimensional flattened temporal features (3 years × 20 features), 11-dimensional city One-Hot encoding, total dimension: 71 dimensions; hyperparameters: n_estimators=100, max_depth=3, learning_rate=0.1;

[0080] Training optimization: The weighted mean squared error loss function is adopted, the Adam optimizer is used, and L2 weight decay regularization, ReduceLROnPlateau learning rate scheduling strategy, Epoch patience value early stopping mechanism and batch training mechanism are combined to optimize training.

[0081] Dynamic weighted ensemble: The weights of the LSTM network and the XGBoost regressor are calculated based on the R² scores of the LSTM network and the XGBoost regressor to obtain the weighted ensemble prediction model;

[0082] S3. Predict future data based on existing data. Using historical data from the standardized dataset in S1 as a basis, predict the weighted integration of the input features of the prediction model for future time steps, specifically:

[0083] Initialization: Traverse all unique cities in the standardized dataset, sort the historical basic feature data of each city by year, select the historical year data for each basic feature, and calculate the historical trend, including linear trend, volatility, average value and annual growth rate, and finally store it as a city trend dictionary;

[0084] Recursive prediction: Receives the target city, the scaled feature value of the basic features at the current time step, the feature scaler, the prediction step size, and the scenario factor. It inversely scales the scaled feature value of the basic features at the current time step to the true value. Based on the scaled true value, it obtains the corresponding historical trend information from the city trend dictionary, adjusts the linear trend using the scenario factor, and calculates the trend component, growth component, and random noise to finally obtain the predicted true value. The predicted true value is then rescaled using the feature scaler, and the time features are updated based on the rescaled predicted true value to obtain the future contextualized input feature sequence.

[0085] S4. Predict future carbon emissions, specifically:

[0086] The future contextualized input feature sequence obtained in S3 is input into the weighted ensemble prediction model trained in S2, and the target city carbon emission prediction results for the future time step are output.

[0087] In practice, carbon emission data is obtained by estimating city-level carbon dioxide emissions within a specified year range based on the Global Atmospheric Research Emissions Database.

[0088] Nighttime light NTL data: By comprehensively utilizing satellite imagery data and employing remote sensing data fusion and calibration methods, continuous and calibrated nighttime light data is constructed, ensuring high temporal continuity and high spatial resolution to reflect urbanization dynamics and the intensity of socio-economic activities;

[0089] Socioeconomic statistics include the output value of the secondary industry and its share of GDP. The output value of the secondary industry is a direct indicator of industrial scale and is closely related to energy consumption and carbon emissions. The share of the secondary industry in GDP reflects the regional economic structure and is used to capture the long-term impact of industrial restructuring on carbon emission intensity.

[0090] Normalized Difference Vegetation Index (NDVI): Based on surface reflectance data from satellite imagery, the city-level NDVI is calculated using a median composite method to mitigate the impact of seasonal fluctuations and to exclude cloud-contaminated pixels using quality control bands.

[0091] During implementation, data cleaning involves using interpolation to fill in missing values ​​and detect and process anomalies in multi-source heterogeneous data to ensure data integrity and accuracy.

[0092] Time features construction: In order to capture short-term fluctuations and long-term trends in carbon emissions, time features include lag features, using 1-year and 2-year lag values ​​to reflect short-term dependence; first-order difference features to capture interannual variability; and 3-year moving average features to smooth the data and reveal long-term trends.

[0093] Time series construction strategy: The sliding window method is adopted to construct the input series using data from three consecutive years to predict carbon emissions in the fourth year. The time step is selected based on capturing short-term time dependencies, balancing the number of samples and the amount of information in the series, avoiding gradient vanishing caused by excessively long series, and adapting to the interannual variation characteristics of urban carbon emissions.

[0094] A one-year lag characteristic refers to the value of a sample one year prior to the current time step. For example, if analyzing carbon emission data for 2023, the one-year lag value of secondary industry output is the secondary industry output value for 2022. If processing monthly data (e.g., extending to a monthly scenario), the one-month lag value of nighttime light is the nighttime light data for the month preceding the current month. Its purpose is to capture the impact of a sample on the current result in the most recent time unit, reflecting short-term historical dependencies.

[0095] First-order difference features are core derived features used to capture the rate of change of a sample at adjacent time steps. By calculating the difference between a sample at the current time step and the previous time step, the original absolute value sequence is transformed into a sequence of changes, thereby highlighting the short-term fluctuation trend of the variable.

[0096] Spatial feature encoding: City identity is encoded using city embedding technology, mapping city ID to a six-dimensional learning space vector, enabling the model to learn emission patterns of specific cities while maintaining cross-city generalization ability.

[0097] Z-score standardization: Scaling the feature values ​​of the data to a range with a mean of 0 and a standard deviation of 1 to eliminate dimensional differences.

[0098] In practice, parallel ensemble prediction is performed by integrating machine learning models such as LSTM and XGBoost, and the prediction results are optimized by combining dynamic weighting mechanism, which can simultaneously capture the complex time dependence and nonlinear feature interaction of carbon emissions.

[0099] LSTM Network Design: A city-embedded enhanced LSTM architecture is adopted, with each layer containing 256 hidden units. A dropout rate of 0.3 is applied between layers and before the final fully connected output layer to effectively prevent overfitting and capture the long-term temporal dependence of carbon emission patterns. The model input consists of three consecutive years of sequence data.

[0100] XGBoost Integration: An XGBoost regressor is introduced, trained on flattened sequence features to model complex nonlinear interactions between variables. XGBoost configuration parameters (such as tree structure, regularization parameters, etc.) are optimized for carbon emission prediction tasks.

[0101] In modeling time series data (such as annual carbon emission data for cities), time series cross-validation is a validation strategy specifically used to evaluate the generalization ability of a model. Its core logic is to strictly adhere to the chronological order to prevent future data from being leaked into historical training, ensuring that the model's performance in real-world prediction scenarios using historical data to predict the future is reliable. Time series cross-validation strictly follows the rule that the time range of the training set is always earlier than that of the test set, completely simulating real-world prediction scenarios; for example, training with data from 2010-2018 to predict 2019, and then training with data from 2011-2019 to predict 2020.

[0102] The original form of time series data is a multidimensional sequence arranged in chronological order; for example, the characteristic data of a city for three consecutive years includes variables such as secondary industry output, NDVI, and nighttime light value, with each variable forming a subsequence by year.

[0103] Flattened sequence features are created by expanding and concatenating time-dimension sequences into single feature vectors without a time dimension. The purpose of flattened sequence features is to adapt them to the XGBoost regressor; XGBoost is an ensemble model based on decision trees, whose input requires a "static feature matrix" (each row is a sample, each column is an independent feature, with no temporal order dependency), making it unable to directly handle the temporal sequence structures that LSTM excels at.

[0104] By flattening the data, XGBoost can transform the "multi-time-step, multi-variable" information in time series into features that XGBoost can recognize. This enables it to capture the absolute value differences of variables at different time steps, such as the gap in industrial output between 2021 and 2023; learn nonlinear interactions across time steps, such as the combined impact of "high NDVI value in 2022 and low industrial output value in 2023" on carbon emissions; and complement LSTM—LSTM excels at capturing time series trends (such as year-on-year growth / decline patterns), while XGBoost captures the nonlinear correlations of specific values ​​through flattened features.

[0105] The city embedding-enhanced LSTM network consists of a city embedding layer, a feature fusion layer, and a two-layer unidirectional LSTM structure. The city embedding layer maps city IDs to a 6-dimensional continuous vector space, and the embedding matrix weights are automatically optimized during training. The feature fusion layer concatenates the city embedding vector with the original features at each time step, so that the LSTM input dimension = the original feature dimension + 6. In the two-layer unidirectional LSTM structure, each LSTM layer contains 256 hidden units. A Dropout rate of 0.3 is applied between layers and before the fully connected output layer. The LSTM network is input with three consecutive years of sequence data to capture temporal dependencies.

[0106] The XGBoost regressor employs differential feature allocation, using 60-dimensional flattened temporal features and 11-dimensional One-Hot encoded city features to form 71-dimensional input features. The decision tree algorithm can directly construct splitting conditions through One-Hot encoded city features.

[0107] The LSTM network was trained using an Adam optimizer with an initial learning rate of 0.0005, combined with a 1×10⁻⁶... -5 The L2 weight decay regularization, the ReduceLROnPlateau learning rate scheduling strategy with a learning rate decay factor of 0.5 and a patience value of 3 epochs, the early stopping mechanism with a patience value of 10 epochs, and the batch training mechanism with a batch size of 64.

[0108] The XGBoost regressor is configured with hyperparameters n_estimators = 100, max_depth = 3, and learning_rate = 0.1.

[0109] In the specific training process, the weighted mean squared error loss function is used to calculate the loss value, as shown below:

[0110] ;

[0111] in, This represents the weighted loss value. Represents the total number of samples. This represents the weight assigned to sample i. This represents the actual carbon emissions of sample i. Let represent the predicted carbon emissions of sample i. In high-emission scenarios where the predicted carbon emissions after normalization are greater than 0.5, the weight assigned to the sample is 2; otherwise, it is 1, ensuring the model focuses on high-emission scenarios. This design makes the model more attentive to high-emission scenarios during training, which is crucial for environmental policy and management decisions.

[0112] Dynamically Weighted Integration: R² Benchmark Dynamic Weighting: The final prediction results are dynamically weighted and fused using the R² scores of the LSTM and XGBoost models. The weight calculation formula is as follows:

[0113] ;

[0114] ;

[0115] ;

[0116] in, and These are the dynamic weights of the LSTM and XGBoost models, respectively. and These are the coefficients of determination for the LSTM and XGBoost models on the validation data, respectively. and These are the prediction results from the LSTM and XGBoost models, respectively. This is the final integrated prediction result.

[0117] Exception handling mechanism: When When using an equal-weight strategy ( = =0.5), to ensure the stability of the predicted values ​​and avoid division by zero errors.

[0118] This dynamic weighting mechanism can adjust the contribution based on the model's real-time performance on validation data, making full use of the complementary advantages of the two models to improve the robustness and accuracy of the overall prediction, especially when facing changes in data patterns.

[0119] In practice, the historical data in S3 is selected from the most recent min(3, len(city_data)) years for analysis to capture recent development trends. len(city_data) represents the number of years the data has been statistically available.

[0120] Linear trend calculation is performed using the following formula:

[0121] ;

[0122] in, This represents the linear trend of sample i. This indicates that the latest value of sample i is typically the average of the past 3 years. Indicates the initial value. and For the corresponding year;

[0123] Volatility calculation: used to measure the dispersion of characteristic value changes, standardized by standard deviation / (absolute value of mean + 1e-6).

[0124] In implementation, the predicted true value is calculated in S3 using the following expression:

[0125] ;

[0126] Where k represents the prediction step size; Represents the observation value at time t; Indicates trend components, The linear trend adjusted based on scenario factors is expressed as follows:

[0127] , Indicates situational factors;

[0128] The growth component is represented as:

[0129] , The growth rate of the index is determined based on historical data;

[0130] Noise is represented as:

[0131] , For volatility, the noise is a normally distributed random noise with a mean of 0 and a standard deviation that is related to volatility and the latest observation.

[0132] After predicting the baseline data features for the next time step, the lag features, difference features, and moving average features for the next time step are recalculated based on these new baseline data features. This ensures that all derived features maintain logical consistency with the currently predicted baseline data features during the rolling forecasting process, thereby constructing a composite data sequence for a rolling window as input for core carbon emission forecasting.

[0133] Scenario simulation defines and simulates different future development scenarios and evaluates the emission reduction effects of various policies on urban carbon emissions. The scenario design closely follows the key policy directions and objectives outlined in regional or national economic and social development plans, such as industrial transformation, ecological restoration, energy structure optimization, and urban development. Scenario factors are set for key basic characteristics (such as secondary industry output, secondary industry output share, NDVI, and total nighttime light) to reflect development trends under different policy interventions.

[0134] Specifically, the scenario factors are set with ranges based on different fundamental characteristics to reflect development trends under different policy interventions, including:

[0135] Slowing Industrial Growth Scenario: The secondary industry output scenario factor is 0.8-0.9, simulating a slowdown in industrial growth;

[0136] The scenario factor for the proportion of secondary industry output is 0.96-0.98, simulating an average annual decrease in the share of industrial structure.

[0137] Ecological enhancement scenario: The normalized vegetation index scenario factor is 1.05-1.1, simulating ecological enhancement;

[0138] The nighttime light scene factor is 1.02-1.05, simulating an increase in nighttime light intensity caused by urban renewal, infrastructure improvement, and increased economic activity;

[0139] Baseline Scenario: The baseline scenario factor is 1, simulating a continuation of historical trends;

[0140] Green transition scenario: A comprehensive adjustment of at least two scenario factors;

[0141] Deep decarbonization scenario: emission reduction measures based on multi-dimensional scenario factors.

[0142] In practice, the carbon emission forecasts under different scenario factors are compared with those under the baseline scenario factor. By calculating the percentage or absolute amount of reduction in carbon emission forecasts under different scenario factors relative to those under the baseline scenario factor, the emission reduction effects of various policy interventions are accurately assessed. Furthermore, the forecast results include a 95% confidence interval, reflecting the uncertainty of future forecasts and providing policymakers with a basis for risk assessment.

[0143] In implementation, the process of updating time features in S3 based on the rescaled predicted true values ​​includes:

[0144] 1-year lag characteristic = predicted true value at the previous time step;

[0145] 2-year lag characteristic = predicted true value two time steps ago;

[0146] Differential feature = predicted true value at the current time step - predicted true value at the previous time step;

[0147] Moving average feature = the mean calculated using a sliding window with a maintenance time step of 3;

[0148] The updated time features are standardized and a sequence window is slidable to add the new feature vector to the end of the sequence, while the earliest time step feature sequence is removed, keeping the feature sequence length at 3.

[0149] Hierarchical clustering and K-means algorithms were used to classify cities, and principal component analysis was used for verification. Ultimately, cities were divided into city clusters with different carbon emission characteristics. Clustering features: differences in key variables such as total emissions, industrial output, industrial share, NDVI, and total nighttime light emissions were displayed for each city cluster using radar charts and other methods. For example, representative city clusters such as high-emission industrial centers, cities with medium emissions and mixed industrial and ecological environments, and low-emission cities could be identified.

[0150] For each city cluster, multi-dimensional decision-making objectives are set, and optimization algorithms are used to explore the optimal policy combination for each city cluster to achieve the decision-making objectives. The decision-making objectives include, but are not limited to: minimizing carbon emissions under a specific economic growth target and minimizing economic impact under a given carbon emission reduction target. The optimization algorithms include, but are not limited to: genetic algorithm and simulated annealing algorithm.

[0151] Ultimately, customized emission reduction strategies can be provided based on the unique emission characteristics and development stages of each urban agglomeration. For example: For cities dominated by heavy industry, it is recommended to implement policies such as industrial restructuring, energy efficiency improvement, and clean energy substitution. For rapidly urbanizing areas, it is recommended to strengthen land use control, investment in green infrastructure, and emission cap management for new industrial projects to decouple urban growth from carbon intensity.

[0152] In implementation, the technical solution of this invention can also provide user interaction and visualization functions. Specifically, the core functions include: providing a user-friendly interface to display carbon emission prediction results, scenario simulations, and policy evaluation results, and supporting interactive operations by users. The aim is to provide intuitive and comprehensive information display and convenient interactive functions; the interactive interface may include:

[0153] Historical and Forecast Trends: Line charts show historical CO2 emission data, forecast paths, and their 95% confidence intervals for each city and region.

[0154] Scenario Comparison Visualization: The carbon emission prediction paths under the baseline scenario and multiple policy scenarios are displayed in the same chart with different colors and styles, allowing for an intuitive comparison of the impact of different policies on emission reduction.

[0155] City-level differentiated presentation: Provides detailed forecast results at the city level, such as emission increase / decrease forecasts for major representative cities, and displays their confidence intervals, reflecting the development patterns of different cities.

[0156] Feature Importance Analysis: Visualizes the ranking of the basic feature importance of LSTM and XGBoost models and ensemble models, helping users understand which factors have the greatest impact on carbon emission prediction.

[0157] Model performance metrics: Display the model's evaluation metrics such as R², RMSE, MAE, and MAPE, and visually present the model's predictive performance and robustness through scatter plots, residual analysis plots, error distribution plots, and radar charts.

[0158] User input and configuration:

[0159] Scenario Factor Adjustment Interface: Provides an intuitive graphical interface (such as sliders and drop-down menus) that allows users to flexibly adjust scenario factors of basic characteristics such as secondary industry output, secondary industry output share, NDVI, and total nighttime light, based on policy objectives and actual conditions, and to observe the impact on carbon emission forecast results in real time.

[0160] City Selection and Group Viewing: Users can select a specific city to view its detailed predictions and assessments, or view the characteristics and policy recommendations of different city clusters.

[0161] Report generation and export: Supports generating customized reports that include forecast data, scenario analysis results, policy benefit assessments, and visualization charts, and supports exporting to PDF, Excel, and other formats.

[0162] The technical solution of this invention integrates industrial structure indicators, environmental variables, and socio-economic alternative indicators through time feature engineering, which can comprehensively capture emission driving factors under different urban environments; establishes a recursive prediction mechanism that considers the development trajectories of different cities, enabling reliable emission prediction while maintaining time consistency and quantifying prediction uncertainty; and quantifies the intervention impact under different policies, providing a quantitative basis for carbon management decisions.

[0163] The above description is merely a preferred embodiment of the present invention and is not intended to limit the scope of the invention. All equivalent changes made based on the description and drawings of the present invention are included within the scope of the present invention.

Claims

1. A smart prediction method for urban carbon emissions based on multimodal data and machine learning integration, characterized in that, The method includes: S1. Feature Acquisition and Preprocessing: Collect multi-source heterogeneous data and process it into a standardized dataset, specifically: Data sources include carbon emission data, nighttime light data, socioeconomic statistics, and normalized difference vegetation index (NDVI) data. Data processing: The collected data sources are cleaned, time features are constructed, spatial features are encoded, and Z-score is standardized. The training set and test set are divided using a time series cross-validation strategy, and a standardized dataset is output. The S2, LSTM, and XGBoost hybrid prediction model is constructed by training the hybrid prediction model based on the standardized dataset output by S1, specifically as follows: Model architecture construction: A city-embedded enhanced LSTM network and an XGBoost regressor are built. The LSTM network is input with continuous sequence data to capture time dependencies, and the XGBoost regressor is trained on flattened sequence features to model nonlinear feature interactions. Differentiated Feature Strategies: LSTM: uses 6-dimensional continuous vectors for city embeddings; XGBoost uses 11-dimensional One-Hot encoding. Feature composition: 60-dimensional flattened temporal features, 11-dimensional city One-Hot encoding, total dimension: 71 dimensions; Hyperparameters: n_estimators=100, max_depth=3, learning_rate=0.1 Training optimization: The weighted mean squared error loss function is adopted, the Adam optimizer is used, and L2 weight decay regularization, ReduceLROnPlateau learning rate scheduling strategy, Epoch patience value early stopping mechanism and batch training mechanism are combined to optimize training. Dynamic weighted ensemble: The weights of the LSTM network and the XGBoost regressor are calculated based on the R² scores of the LSTM network and the XGBoost regressor to obtain the weighted ensemble prediction model; S3. Predict future data based on existing data. Using historical data from the standardized dataset in S1 as a basis, predict the weighted integration of the input features of the prediction model for future time steps, specifically: Initialization: Traverse all unique cities in the standardized dataset, sort the historical basic feature data of each city by year, select the historical year data for each basic feature, and calculate the historical trend, including linear trend, volatility, average value and annual growth rate, and finally store it as a city trend dictionary; Recursive prediction: Receives the target city, the scaled feature value of the basic features at the current time step, the feature scaler, the prediction step size, and the scenario factor. It inversely scales the scaled feature value of the basic features at the current time step to the true value. Based on the scaled true value, it obtains the corresponding historical trend information from the city trend dictionary, adjusts the linear trend using the scenario factor, and calculates the trend component, growth component, and random noise to finally obtain the predicted true value. The predicted true value is then rescaled using the feature scaler, and the time features are updated based on the rescaled predicted true value to obtain the future contextualized input feature sequence. S4. Predict future carbon emissions, specifically: The future contextualized input feature sequence obtained in S3 is input into the weighted ensemble prediction model trained in S2, and the target city carbon emission prediction results for the future time step are output.

2. The intelligent urban carbon emission prediction method based on multimodal data and machine learning integration according to claim 1, characterized in that, Carbon emission data: Estimates of carbon dioxide emissions at the city level within a specified year range are obtained based on the Global Atmospheric Research Emissions Database; Nighttime light data: By comprehensively utilizing satellite imagery data and employing remote sensing data fusion and calibration methods, continuous and calibrated nighttime light data is constructed. Socioeconomic statistics include the output value of the secondary industry and its share of GDP. The output value of the secondary industry is a direct indicator of industrial scale and is closely related to energy consumption and carbon emissions. The share of the secondary industry in GDP reflects the regional economic structure and is used to capture the long-term impact of industrial restructuring on carbon emission intensity. Normalized Difference Vegetation Index (NDVI): Based on surface reflectance data from satellite imagery, the city-level NDVI is calculated using a median composite method to mitigate the impact of seasonal fluctuations and to exclude cloud-contaminated pixels using quality control bands.

3. The intelligent urban carbon emission prediction method based on multimodal data and machine learning integration according to claim 2, characterized in that, Data cleaning: Interpolation is used to fill in missing values ​​and detect and handle anomalies in multi-source heterogeneous data to ensure data integrity and accuracy; Construction of time features: In order to capture short-term fluctuations and long-term trends in carbon emissions, including lag features, 1-year and 2-year lag values ​​are used to reflect short-term dependence. First-order difference features to capture interannual variability; 3-year moving average features to smooth data and reveal long-term trends; Spatial feature encoding: City identity is encoded using city embedding technology, mapping city ID to a six-dimensional learning space vector, enabling the model to learn emission patterns of specific cities while maintaining cross-city generalization ability; Z-score standardization: Scaling the feature values ​​of the data to a range with a mean of 0 and a standard deviation of 1 to eliminate dimensional differences.

4. The intelligent urban carbon emission prediction method based on multimodal data and machine learning integration according to claim 1, characterized in that, The weighted mean squared error loss function in S2 is expressed as follows: ； in, This represents the weighted loss value. Represents the total number of samples. This represents the weight assigned to sample i. This represents the actual carbon emissions of sample i. This represents the predicted carbon emission value of sample i. In the case of a high emission scenario where the predicted carbon emission value after normalization is greater than 0.5, the weight assigned to the sample is set to 2, otherwise it is 1, to ensure that the model focuses on the high emission scenario.

5. The intelligent urban carbon emission prediction method based on multimodal data and machine learning integration according to claim 1, characterized in that, In S3, the historical year data is selected from the most recent min(3, len(city_data)) years for analysis to capture recent development trends. len(city_data) represents the number of years the data has been statistically available. Linear trend calculation is performed using the following formula: ； in, This represents the linear trend of sample i. This indicates that the latest value of sample i is typically the average of the past 3 years. Indicates the initial value. and For the corresponding year; Volatility calculation: used to measure the dispersion of characteristic value changes, standardized by standard deviation / (absolute value of mean + 1e-6).

6. The intelligent urban carbon emission prediction method based on multimodal data and machine learning integration according to claim 5, characterized in that, In S3, the predicted true value is calculated using the following expression: ； Where k represents the prediction step size; Represents the observation value at time t; Indicates trend components, The linear trend adjusted based on scenario factors is expressed as follows: , Represents situational factors; The growth component is represented as: , The index growth rate is determined based on historical data; Noise is represented as: , For volatility, the noise is a normally distributed random noise with a mean of 0 and a standard deviation that is related to volatility and the latest observation.

7. The intelligent urban carbon emission prediction method based on multimodal data and machine learning integration according to claim 6, characterized in that, Scenario factors are set with ranges based on different fundamental characteristics to reflect development trends under different policy interventions, including: Slowing Industrial Growth Scenario: The secondary industry output scenario factor is 0.8-0.9, simulating a slowdown in industrial growth; The scenario factor for the proportion of secondary industry output is 0.96-0.98, simulating an average annual decrease in the share of industrial structure. Ecological enhancement scenario: The normalized vegetation index scenario factor is 1.05-1.1, simulating ecological enhancement; The nighttime light scene factor is 1.02-1.05, simulating an increase in nighttime light intensity caused by urban renewal, infrastructure improvement, and increased economic activity; Baseline Scenario: The baseline scenario factor is 1, simulating a continuation of historical trends; Green transition scenario: A comprehensive adjustment of at least two scenario factors; Deep decarbonization scenario: emission reduction measures based on multi-dimensional scenario factors.

8. The intelligent urban carbon emission prediction method based on multimodal data and machine learning integration according to claim 7, characterized in that, Also includes: By comparing the carbon emission forecasts under different scenario factors with the carbon emission forecasts under the baseline scenario factors, and by calculating the percentage or absolute amount of reduction in carbon emission forecasts under different scenario factors relative to the carbon emission forecasts under the baseline scenario factors, the emission reduction effects of various policy interventions can be accurately assessed. Hierarchical clustering and K-means algorithms were used to classify cities, and principal component analysis was used for verification. Finally, the cities were divided into city groups with different carbon emission characteristics. For each city cluster, multi-dimensional decision-making objectives are set, and optimization algorithms are used to explore the optimal policy combination for each city cluster to achieve the decision-making objectives. The decision-making objectives include, but are not limited to: minimizing carbon emissions under a specific economic growth target and minimizing economic impact under a given carbon emission reduction target. The optimization algorithms include, but are not limited to: genetic algorithm and simulated annealing algorithm.

9. The intelligent urban carbon emission prediction method based on multimodal data and machine learning integration according to claim 3, characterized in that, The city embedding-enhanced LSTM network consists of a city embedding layer, a feature fusion layer, and a two-layer unidirectional LSTM structure. The city embedding layer maps city IDs to a 6-dimensional continuous vector space, and the embedding matrix weights are automatically optimized during training. The feature fusion layer concatenates the city embedding vector with the original features at each time step, so that the LSTM input dimension = the original feature dimension + 6. In the two-layer unidirectional LSTM structure, each LSTM layer contains 256 hidden units. A Dropout rate of 0.3 is applied between layers and before the fully connected output layer. The LSTM network is input with three consecutive years of sequence data to capture temporal dependencies. The XGBoost regressor employs differential feature allocation, using 60-dimensional flattened temporal features and 11-dimensional One-Hot encoded city features to form 71-dimensional input features. The decision tree algorithm can directly construct splitting conditions through One-Hot encoded city features. The LSTM network was trained using an Adam optimizer with an initial learning rate of 0.0005, combined with a 1×10⁻⁶... -5 The L2 weight decay regularization, the ReduceLROnPlateau learning rate scheduling strategy with a learning rate decay factor of 0.5 and a patience value of 3 epochs, the early stopping mechanism with a patience value of 10 epochs, and the batch training mechanism with a batch size of 64. The XGBoost regressor is configured with hyperparameters n_estimators = 100, max_depth = 3, and learning_rate = 0.

1.

10. The intelligent urban carbon emission prediction method based on multimodal data and machine learning integration according to claim 3, characterized in that, The process of updating time features based on the rescaled predicted true values ​​in S3 includes: 1-year lag characteristic = predicted true value at the previous time step; 2-year lag characteristic = predicted true value two time steps ago; Differential feature = predicted true value at the current time step - predicted true value at the previous time step; Moving average feature = the mean calculated using a sliding window with a maintenance time step of 3; The updated time features are standardized and a sequence window is slidable to add the new feature vector to the end of the sequence, while the earliest time step feature sequence is removed, keeping the feature sequence length at 3.

Images