Prediction and control method of local carbon emission hotspots in cities based on generative adversarial network

Abstract

The present invention discloses a method for predicting and regulating local carbon emission hotspots in cities based on a generative adversarial network, comprising the following steps: S1, data collection and preprocessing; S2, using hourly carbon emission data to construct a short-term memory module and perform short-term spatiotemporal feature encoding; S3, based on daily and weekly carbon emission data, using a multi-scale convolutional network to extract periodic change features and perform medium-term spatiotemporal feature encoding; S4, using a long-term time series dependency model to analyze monthly and quarterly carbon emission data, obtain long-term spatiotemporal features, and globally encode the overall carbon emission distribution of the city; S5, constructing and training a generative adversarial network; S6, using a spatiotemporal attention mechanism to weightedly fuse spatiotemporal features of different levels, and generate a regulation strategy in combination with real-time data; S7, real-time monitoring of the regulation effect, and adaptive adjustment and optimization. The present invention adopts a generative adversarial network and a spatiotemporal feature encoding method to achieve accurate prediction and dynamic regulation of urban carbon emission hotspots.

Description

Technical Field

[0001] The present invention relates to the fields of environmental protection and urban management, and in particular to a method for predicting and regulating local carbon emission hotspots in cities based on a generative adversarial network. Background Art

[0002] In the process of modern urbanization, the monitoring and control of carbon emissions has become an important issue for environmental protection and sustainable development. The main sources of urban carbon emissions include transportation, industrial production, building energy consumption, and residents' lives, which have significant temporal and spatial heterogeneity and multi-scale variation characteristics. However, due to the dynamic characteristics and complexity of urban carbon emissions, existing technologies still have many deficiencies in the accurate prediction and regulation of carbon emission hotspots.

[0003] Existing carbon emission prediction methods are usually based on traditional statistical models and machine learning methods, such as regression analysis, time series analysis, and support vector machines. These methods mainly rely on the modeling of historical carbon emission data to predict future emission trends. However, traditional statistical models usually assume a linear relationship between data and are difficult to capture nonlinear dynamic changes in carbon emission data. Although machine learning-based methods have strong nonlinear modeling capabilities, they show great limitations when dealing with multi-scale spatiotemporal characteristics. For example, existing methods often fail to fully consider the multi-level changes in carbon emissions at different time and spatial scales, and it is difficult to accurately reflect short-term emergencies and long-term trends. At the same time, existing methods lack comprehensive analysis of multi-dimensional data (such as traffic flow, meteorological conditions, etc.) during the prediction process, resulting in low accuracy and stability of prediction results.

[0004] For the regulation of carbon emissions, existing technologies usually rely on preset strategies or rules, such as traffic restrictions and industrial emission reductions. Although these measures can achieve certain emission reduction effects in some cases, the formulation of their regulation strategies is often based on experience or static rules and lacks the ability to dynamically respond to real-time data. In addition, the adaptability of regulatory measures is poor. When the carbon emission situation in a city changes suddenly or significantly, it is difficult to adjust the strategy in time to cope with the new environment. With the development of Internet of Things technology, more and more cities have begun to deploy intelligent sensing networks for real-time monitoring of carbon emissions and other environmental data. However, how to use these real-time data to optimize regulatory strategies is still a challenge facing existing technologies.

[0005] In terms of multi-level spatiotemporal feature encoding, most current methods analyze a single time scale (such as daily or weekly level) or spatial scale (such as regional or city level), and cannot take into account the feature extraction of different time and space scales at the same time. This method is prone to prediction bias when dealing with short-term carbon emission fluctuations and long-term trend changes. In addition, for the correlation analysis between multiple regions, the existing technology also has certain limitations and cannot accurately capture the mutual influence between different regions, thus affecting the precise positioning and prediction of carbon emission hotspots. Although the existing multi-scale convolutional neural network (MSCNN) or long short-term memory network (LSTM) has improved the spatiotemporal feature extraction capability of the model in some aspects, there is still the problem of insufficient fusion of multi-level features, which makes it impossible to form a comprehensive understanding of carbon emission changes.

[0006] In terms of the generation and optimization of carbon emission control strategies, existing methods usually use fixed optimization algorithms and lack the ability to adaptively learn from real-time feedback. Although some methods have introduced reinforcement learning (RL) technology to optimize control strategies, the exploration and utilization mechanism in the reinforcement learning process is often difficult to maintain the optimal control effect in a dynamically changing urban environment. In addition, the design of multi-objective control strategies involves the problem of trade-offs and priority selection among multiple objectives. Existing technologies lack an efficient multi-objective decision-making mechanism and cannot achieve the optimal comprehensive balance between carbon emission reduction, traffic smoothness, and economic impact.

[0007] As a powerful generative model, generative adversarial network (GAN) technology has achieved remarkable results in the fields of image synthesis, video generation, and natural language processing. However, there are still many challenges in applying GAN to the prediction and regulation of urban carbon emission hotspots. First, the generator and discriminator of GAN need to model complex spatiotemporal features, and the existing GAN architecture is difficult to directly handle multi-level spatiotemporal feature encoding. Secondly, the training process of GAN often relies on a large amount of sample data. In the actual application scenarios of urban carbon emissions, the availability and quality of data may be greatly limited. In addition, how to design a suitable loss function to simultaneously optimize the performance of GAN in short-term emergency response and long-term trend prediction is also a problem that existing technologies need to solve.

[0008] Therefore, how to provide a method for predicting and regulating urban local carbon emission hotspots based on a generative adversarial network is an urgent problem that technicians in this field need to solve. Summary of the invention

[0009] One purpose of the present invention is to propose a method for predicting and regulating urban local carbon emission hotspots based on a generative adversarial network. The present invention adopts a generative adversarial network and a multi-level spatiotemporal feature encoding method to achieve accurate prediction and dynamic regulation of urban carbon emission hotspots. Through hierarchical spatiotemporal feature fusion and adaptive spatiotemporal attention mechanism, multi-scale data is comprehensively analyzed, and a control scheme is dynamically generated in combination with a real-time multi-objective optimization strategy. It has the advantages of high-precision prediction, rapid response, strong emergency handling capabilities and high adaptability, which significantly improves the intelligent level of urban carbon emission management and solves the shortcomings of existing methods in multi-level feature modeling and real-time regulation.

[0010] According to an embodiment of the present invention, a method for predicting and regulating local carbon emission hotspots in an urban area based on a generative adversarial network comprises the following steps:

[0011] S1. Collect historical carbon emission data, traffic flow data and meteorological data, normalize the collected data, fill in missing values ​​and remove outliers, and perform multi-scale segmentation of time and space to form time series data at hourly, daily, weekly, monthly and quarterly levels;

[0012] S2. Use hourly carbon emission data to build a short-term memory module to extract the changing trend of short-term carbon emissions, and use a short-term spatiotemporal convolutional network to encode the short-term spatiotemporal features of block-level carbon emission data to detect sudden emission events;

[0013] S3. Based on daily and weekly carbon emission data, a multi-scale convolutional network is used to extract periodic change characteristics, encode medium-term spatiotemporal features, and cross-integrate the medium-term spatiotemporal features of different regions to analyze the correlation between regions.

[0014] S4. Use the long-term dependency model to analyze carbon emission data at the monthly and quarterly levels, obtain long-term spatiotemporal characteristics, and globally encode the overall carbon emission distribution of the city;

[0015] S5. Build and train a generative adversarial network. The generator integrates a multi-level spatiotemporal feature encoding module to generate a carbon emission hotspot distribution map for a specific time period in the future. The discriminator compares the generated distribution map with the real data, performs iterative training through an alternating optimization strategy, introduces a multi-level spatiotemporal feature loss function, and optimizes short-term emergency response and long-term trend prediction.

[0016] S6. Use the spatiotemporal attention mechanism to weightedly fuse spatiotemporal features at different levels, generate control strategies based on the prediction results of the generative adversarial network and combine real-time data to dynamically optimize urban carbon emission control;

[0017] S7. Monitor the control effect in real time, use feedback data to update the generative adversarial network and control strategy, and perform adaptive adjustment and optimization.

[0018] Optionally, the S2 specifically includes:

[0019] S21. Obtain hourly carbon emission time series data ,in represents the carbon emissions in the ith hour. The data is smoothed and denoised and frequency adjusted using a method based on nonlinear filtering and frequency domain transformation to obtain the smoothed carbon emission sequence. ;

[0020] S22. Construct a short-term memory module and use the improved long short-term memory network to analyze the smoothed carbon emission sequence. Modeling is performed, a multi-channel gating mechanism and a time decay function are introduced, and the hidden state is calculated:

[0021] ;

[0022] in, represents the hidden state at the current time t, L represents the number of nested memory layers, represents the activation function, and represents the weight matrix of the lth layer, represents the input data at the current time t, and represents the bias term of the lth layer, represents bit-wise multiplication of elements, represents the hyperbolic tangent function, Indicates The state of the memory unit in the lth layer at time , represents the time interval between the current time step and the previous time step, represents the parameter set of the time decay function of the lth layer, represents the time decay function;

[0023] S23. Perform high-order feature enhancement on the extracted short-term trend features, and perform feature transformation by combining feature adaptive amplification and orthogonal transformation:

[0024] ;

[0025] in, Indicates the enhanced short-term trend characteristics, Represents the original short-term trend characteristics, represents the adjustment parameter, The local gradient representing the short-term trend characteristics, represents the nonlinear enhancement index, Q and P represent orthogonal matrices;

[0026] S24. Use a short-term spatiotemporal convolutional network to perform convolution operations on the block-level carbon emission data matrix. Let the block-level carbon emission matrix be M, and use a multi-scale adaptive convolution kernel to extract local spatial features, perform tensor contraction and reweighting:

[0027] ;

[0028] in, represents the extracted spatial distribution features, represents the convolution kernel of the sth scale, represents the convolution operation, Tr represents the trace operation of the matrix, represents the scale adaptive adjustment factor;

[0029] S25. Enhanced short-term trend characteristics And the extracted spatial distribution features Perform nonlinear fusion and adopt adaptive projection fusion strategy:

[0030] ;

[0031] in, represents the short-term spatiotemporal features after fusion, represents the nonlinear mapping function, U and V represent the adaptive orthogonal matrix, represents the weight vector, W represents the bias tensor, represents bit-wise multiplication of elements, It means converting the vector into a diagonal matrix. represents the nonlinear index, Represents a tensor decomposition operation;

[0032] S26, based on the fused short-term spatiotemporal features , anomaly detection algorithms are applied to detect carbon emission data, identifying and marking abnormal areas and spatiotemporal information.

[0033] Optionally, the S3 specifically includes:

[0034] S31. Obtaining carbon emission time series data matrix at daily and weekly levels ,in represents the carbon emissions of the jth hour of the i-th day, and divides the matrix R into multiple sub-matrices , each sub-matrix represents the carbon emissions in a specific time period and region;

[0035] S32, using multi-scale convolutional neural network to sub-matrix Row convolution feature extraction, using an adaptive convolution kernel whose size is dynamically adjusted to obtain features of different convolution scales;

[0036] S33, fuse the features of different convolution scales, and use the adaptive nonlinear feature combination method:

[0037] ;

[0038] in, represents the fused features, P represents the number of scales, represents the fusion weight of the p-th scale, represents the local features extracted at the pth scale, Q represents the number of another set of scales, represents the local features extracted at the qth scale, and represents the nonlinear adjustment parameter, represents the characteristic frequency component, Indicates a cycle;

[0039] S34. The fused features Perform periodic analysis, use wavelet packet transform combined with multi-dimensional phase modulation method, perform multi-band decomposition and phase adjustment, and obtain the multi-frequency feature matrix:

[0040] ;

[0041] in, represents the multi-frequency feature matrix, represents the wavelet packet transform operation, represents bit-wise multiplication of elements, represents the phase modulation parameter matrix, represents the characteristic phase mapping matrix, represents the modulation frequency control matrix;

[0042] S35, multi-frequency feature matrix Perform inverse wavelet packet transform, combine multidimensional nonlinear projection and hierarchical feature reconstruction, and obtain the reconstructed feature representation ;

[0043] S36. Feature representation after reconstruction of different regions Perform adaptive graph convolutional network processing, analyze the complex correlations between regions through adaptive weighted connections between nodes, and generate the final mid-term spatiotemporal feature encoding .

[0044] Optionally, the S4 specifically includes:

[0045] S41. Obtain carbon emission data at the monthly and quarterly levels and set it as a time series matrix ,in represents the carbon emissions on the nth day of the mth month, and performs long-term segmentation on the matrix L to divide the data into multiple long-term dependent segments , each segment represents a continuous process of carbon emission change;

[0046] S42, using the long-term dependency model for each long-term dependency segment Modeling is performed to extract long-term dependent features of different time steps through a multi-head self-attention mechanism, and the long-term feature expression is enhanced by combining time weights and spatial relationships:

[0047] ;

[0048] in, represents the global self-attention output, T represents the total number of time steps, represents the time weight of the tth time step, exp represents the exponential function, represents the query matrix, represents the key matrix, represents the value matrix, represents the dimension of the key matrix, Represents the adaptive index adjustment parameter;

[0049] S43, multi-scale time fusion is performed on the extracted long-term spatiotemporal features. An adaptive time fusion strategy is used to decompose the long-term dependent features into feature subsets of multiple time scales, and the features of each scale are re-weighted and combined to obtain the fused features. ;

[0050] S44. Globally encode the fused long-term spatiotemporal features, using multi-layer nested nonlinear mapping and tensor decomposition strategies:

[0051] ;

[0052] in, represents the long-term spatiotemporal characteristics after global encoding, represents a tensor decomposition operation, means converting the result into a diagonal matrix, represents the feature reconstruction function, represents tensor product, represents the activation function;

[0053] S45. Long-term spatiotemporal features after global encoding Perform multi-layer nested feature enhancement processing, combine with adaptive feature reconstruction algorithm, reproject the features into different high-dimensional spaces, and generate enhanced global features ;

[0054] S46. The enhanced global features The short-term spatiotemporal features and the medium-term spatiotemporal features generated by S36 Multi-scale fusion is performed to form the final long-term spatiotemporal feature representation, which is used to analyze the overall carbon emission trend of the city.

[0055] Optionally, the S5 specifically includes:

[0056] S51, constructing a generative adversarial network, including two modules: a generator and a discriminator, wherein the generator integrates a multi-level spatiotemporal feature encoding module;

[0057] S52. Design the generator network structure, combine multi-layer convolution and deconvolution layers, encode and decode the input multi-level spatiotemporal features through the spatiotemporal convolution network, generate the carbon emission hotspot distribution map, and use the high-dimensional feature mapping strategy for the generator output to project the generated image into the carbon emission distribution space;

[0058] S53. Design the discriminator network structure, use a multi-scale convolutional network to extract features from the input carbon emission distribution map, and combine it with a multi-layer fully connected network to classify the spatiotemporal features of the image to determine whether the input data is real data or generated data;

[0059] S54. Introduce a multi-level spatiotemporal feature loss function, combine multi-dimensional feature mapping and adversarial learning, and optimize the training of the generator:

[0060] ;

[0061] in, represents the loss function of the generator, E represents the mathematical expectation operator, x represents the data sample, represents the distribution of real data, z represents the noise sample, represents the noise distribution, represents the output of the discriminator for real data, Represents the generated data output by the generator, exp represents the exponential function, and represents the adjustment parameter, represents the feature transformation function, represents the Frobenius norm, Tr represents the trace operation of the matrix;

[0062] S55, adopt the alternating optimization strategy for training, fix the parameters of the generator, update the weight of the discriminator; then fix the parameters of the discriminator, update the weight of the generator, and continuously optimize the quality of the generated results through multiple rounds of iterations;

[0063] S56. During the training process, the learning rate and momentum parameters are adaptively adjusted, and finally a generative adversarial network model for generating a carbon emission hotspot distribution map is obtained.

[0064] Optionally, the S6 specifically includes:

[0065] S61. Introduce the spatiotemporal attention mechanism to perform weighted fusion of spatiotemporal features at different levels, input short-term spatiotemporal features, medium-term spatiotemporal features and long-term spatiotemporal features into the attention module respectively, dynamically adjust the weights of features at each level, enhance the response capability to key time periods and important areas, and generate fused spatiotemporal features with comprehensive expression capabilities;

[0066] S62. Based on the prediction results of the generative adversarial network, the fused spatiotemporal features are combined with the real-time urban carbon emission data, traffic flow data and meteorological data to form multi-dimensional data input;

[0067] S63. Generate control strategies, consider the changes in real-time data and the dynamic weight distribution of spatiotemporal characteristics, and design preliminary control measures in combination with the predicted information of urban carbon emission hotspots;

[0068] S64. Adaptively adjust the control strategy and dynamically optimize the control parameters according to the real-time monitored changes in the urban environment.

[0069] The beneficial effects of the present invention are:

[0070] First, the present invention introduces multi-level spatiotemporal feature encoding. Through the hierarchical modeling method, the carbon emission characteristics are divided in detail according to multiple scales such as hours, days, weeks, months and quarters, and features are extracted at different time and space scales, which can not only accurately capture sudden emission events in the short term, but also reflect long-term trend changes. This multi-level spatiotemporal feature fusion breaks through the limitation that traditional methods can only analyze at a single scale, making the prediction of carbon emission hotspots more comprehensive and accurate. At the same time, the application of the spatiotemporal attention mechanism enables the model to adaptively adjust the feature attention weights of different time periods and regions, thereby effectively improving the response capabilities to key time periods and important areas, making the prediction results more flexible and accurate.

[0071] In the optimization of control strategies, the present invention uses the prediction results of generative adversarial networks, combined with multi-source information such as carbon emissions, traffic flow and meteorological data acquired in real time, to generate control strategies with dynamic optimization capabilities, solving the problem that existing methods are difficult to respond to changes in carbon emissions in a timely manner. By introducing a multi-level spatiotemporal feature loss function, the present invention can continuously optimize the control strategy based on real-time monitoring data, and continuously adjust the strategy parameters using environmental feedback, so that the control measures can flexibly respond to complex urban environmental changes. In addition, the control strategy fully considers multiple goals such as carbon emission reduction, traffic smoothness and emergency response capabilities when designing it, and achieves a balance between multiple goals through an adaptive weight allocation mechanism, thereby ensuring the comprehensiveness and superiority of the control effect.

[0072] The present invention also designs a closed-loop dynamic control system, which monitors carbon emission changes and control effects in real time during the control execution process, and continuously updates and adaptively adjusts the control strategy through feedback data, so that the system can continuously adapt to the changing urban carbon emission environment. This closed-loop optimization mechanism not only improves the intelligence level of the control system, but also makes urban carbon emission control more flexible and robust, can significantly reduce the total amount of carbon emissions, and perform well in responding to sudden carbon emission events. Compared with traditional static rule control methods, the control system of the present invention has a higher level of adaptability and intelligence, and can effectively improve the efficiency of urban environmental management. BRIEF DESCRIPTION OF THE DRAWINGS

[0073] The accompanying drawings are used to provide a further understanding of the present invention and constitute a part of the specification. Together with the embodiments of the present invention, they are used to explain the present invention and do not constitute a limitation of the present invention. In the accompanying drawings:

[0074] Figure 1 This is a flow chart of the method for predicting and regulating local urban carbon emission hotspots based on a generative adversarial network proposed in the present invention;

[0075] Figure 2 This is a schematic diagram of the generative adversarial network structure of the urban local carbon emission hotspot prediction and regulation method based on the generative adversarial network proposed in the present invention. DETAILED DESCRIPTION

[0076] The present invention will now be described in further detail with reference to the accompanying drawings. These drawings are simplified schematic diagrams, which only illustrate the basic structure of the present invention in a schematic manner, and therefore only show the components related to the present invention.

[0077] refer to Figure 1 and Figure 2 , a method for predicting and regulating urban local carbon emission hotspots based on a generative adversarial network, including the following steps:

[0078] S1. Collect historical carbon emission data, traffic flow data and meteorological data, normalize the collected data, fill in missing values ​​and remove outliers, and perform multi-scale segmentation of time and space to form time series data at hourly, daily, weekly, monthly and quarterly levels;

[0079] S2. Use hourly carbon emission data to build a short-term memory module to extract the changing trend of short-term carbon emissions, and use a short-term spatiotemporal convolutional network to encode the short-term spatiotemporal features of block-level carbon emission data to detect sudden emission events;

[0080] S3. Based on daily and weekly carbon emission data, a multi-scale convolutional network is used to extract periodic change characteristics, encode medium-term spatiotemporal features, and cross-integrate the medium-term spatiotemporal features of different regions to analyze the correlation between regions.

[0081] S4. Use the long-term dependency model to analyze carbon emission data at the monthly and quarterly levels, obtain long-term spatiotemporal characteristics, and globally encode the overall carbon emission distribution of the city;

[0082] S5. Build and train a generative adversarial network. The generator integrates a multi-level spatiotemporal feature encoding module to generate a carbon emission hotspot distribution map for a specific time period in the future. The discriminator compares the generated distribution map with the real data, performs iterative training through an alternating optimization strategy, introduces a multi-level spatiotemporal feature loss function, and optimizes short-term emergency response and long-term trend prediction.

[0083] S6. Use the spatiotemporal attention mechanism to weightedly fuse spatiotemporal features at different levels, generate control strategies based on the prediction results of the generative adversarial network and combine real-time data to dynamically optimize urban carbon emission control;

[0084] S7. Monitor the control effect in real time, use feedback data to update the generative adversarial network and control strategy, and perform adaptive adjustment and optimization.

[0085] In this implementation, S2 specifically includes:

[0086] S21. Obtain hourly carbon emission time series data ,in represents the carbon emissions in the ith hour. The data is smoothed and denoised and frequency adjusted using a method based on nonlinear filtering and frequency domain transformation to obtain the smoothed carbon emission sequence. ;

[0087] S22. Construct a short-term memory module and use the improved long short-term memory network to analyze the smoothed carbon emission sequence. Modeling is performed, a multi-channel gating mechanism and a time decay function are introduced, and the hidden state is calculated:

[0088] ;

[0089] in, represents the hidden state at the current time t, L represents the number of nested memory layers, represents the activation function, and represents the weight matrix of the lth layer, represents the input data at the current time t, and represents the bias term of the lth layer, represents bit-wise multiplication of elements, represents the hyperbolic tangent function, Indicates The state of the memory unit in the lth layer at time , represents the time interval between the current time step and the previous time step, represents the parameter set of the time decay function of the lth layer, represents the time decay function;

[0090] S23. Perform high-order feature enhancement on the extracted short-term trend features, and perform feature transformation by combining feature adaptive amplification and orthogonal transformation:

[0091] ;

[0092] in, Indicates the enhanced short-term trend characteristics, Represents the original short-term trend characteristics, represents the adjustment parameter, The local gradient representing the short-term trend characteristics, represents the nonlinear enhancement index, Q and P represent orthogonal matrices;

[0093] S24. Use a short-term spatiotemporal convolutional network to perform convolution operations on the block-level carbon emission data matrix. Let the block-level carbon emission matrix be M, and use a multi-scale adaptive convolution kernel to extract local spatial features, perform tensor contraction and reweighting:

[0094] ;

[0095] in, represents the extracted spatial distribution features, represents the convolution kernel of the sth scale, represents the convolution operation, Tr represents the trace operation of the matrix, represents the scale adaptive adjustment factor;

[0096] S25. Enhanced short-term trend characteristics And the extracted spatial distribution features Perform nonlinear fusion and adopt adaptive projection fusion strategy:

[0097] ;

[0098] in, represents the short-term spatiotemporal features after fusion, represents the nonlinear mapping function, U and V represent the adaptive orthogonal matrix, represents the weight vector, W represents the bias tensor, represents bit-wise multiplication of elements, It means converting the vector into a diagonal matrix. represents the nonlinear index, Represents a tensor decomposition operation;

[0099] S26, based on the fused short-term spatiotemporal features , anomaly detection algorithms are applied to detect carbon emission data, identifying and marking abnormal areas and spatiotemporal information.

[0100] In this implementation, S3 specifically includes:

[0101] S31. Obtaining carbon emission time series data matrix at daily and weekly levels ,in represents the carbon emissions of the jth hour of the i-th day, and divides the matrix R into multiple sub-matrices , each sub-matrix represents the carbon emissions in a specific time period and region;

[0102] S32, using multi-scale convolutional neural network to sub-matrix Row convolution feature extraction, using an adaptive convolution kernel whose size is dynamically adjusted to obtain features of different convolution scales;

[0103] S33, fuse the features of different convolution scales, and use the adaptive nonlinear feature combination method:

[0104] ;

[0105] in, represents the fused features, P represents the number of scales, represents the fusion weight of the p-th scale, represents the local features extracted at the pth scale, Q represents the number of another set of scales, represents the local features extracted at the qth scale, and represents the nonlinear adjustment parameter, represents the characteristic frequency component, Indicates a cycle;

[0106] S34. The fused features Perform periodic analysis, use wavelet packet transform combined with multi-dimensional phase modulation method, perform multi-band decomposition and phase adjustment, and obtain the multi-frequency feature matrix:

[0107] ;

[0108] in, represents the multi-frequency feature matrix, represents the wavelet packet transform operation, represents bit-wise multiplication of elements, represents the phase modulation parameter matrix, represents the characteristic phase mapping matrix, represents the modulation frequency control matrix;

[0109] S35, multi-frequency feature matrix Perform inverse wavelet packet transform, combine multidimensional nonlinear projection and hierarchical feature reconstruction, and obtain the reconstructed feature representation ;

[0110] S36. Feature representation after reconstruction of different regions Perform adaptive graph convolutional network processing, analyze the complex correlations between regions through adaptive weighted connections between nodes, and generate the final mid-term spatiotemporal feature encoding .

[0111] In this implementation manner, the S4 specifically includes:

[0112] S41. Obtain carbon emission data at the monthly and quarterly levels and set it as a time series matrix ,in represents the carbon emissions on the nth day of the mth month, and performs long-term segmentation on the matrix L to divide the data into multiple long-term dependent segments , each segment represents a continuous process of carbon emission change;

[0113] S42, using the long-term dependency model for each long-term dependency segment Modeling is performed to extract long-term dependent features of different time steps through a multi-head self-attention mechanism, and the long-term feature expression is enhanced by combining time weights and spatial relationships:

[0114] ;

[0115] in, represents the global self-attention output, T represents the total number of time steps, represents the time weight of the tth time step, exp represents the exponential function, represents the query matrix, represents the key matrix, represents the value matrix, represents the dimension of the key matrix, Represents the adaptive index adjustment parameter;

[0116] S43, multi-scale time fusion is performed on the extracted long-term spatiotemporal features. An adaptive time fusion strategy is used to decompose the long-term dependent features into feature subsets of multiple time scales, and the features of each scale are re-weighted and combined to obtain the fused features. ;

[0117] S44. Globally encode the fused long-term spatiotemporal features, using multi-layer nested nonlinear mapping and tensor decomposition strategies:

[0118] ;

[0119] in, represents the long-term spatiotemporal characteristics after global encoding, represents a tensor decomposition operation, means converting the result into a diagonal matrix, represents the feature reconstruction function, represents tensor product, represents the activation function;

[0120] S45. Long-term spatiotemporal features after global encoding Perform multi-layer nested feature enhancement processing, combine with adaptive feature reconstruction algorithm, reproject the features into different high-dimensional spaces, and generate enhanced global features ;

[0121] S46. The enhanced global features The short-term spatiotemporal features and the medium-term spatiotemporal features generated by S36 Multi-scale fusion is performed to form the final long-term spatiotemporal feature representation, which is used to analyze the overall carbon emission trend of the city.

[0122] In this implementation manner, S5 specifically includes:

[0123] S51, constructing a generative adversarial network, including two modules: a generator and a discriminator, wherein the generator integrates a multi-level spatiotemporal feature encoding module;

[0124] S52. Design the generator network structure, combine multi-layer convolution and deconvolution layers, encode and decode the input multi-level spatiotemporal features through the spatiotemporal convolution network, generate the carbon emission hotspot distribution map, and use the high-dimensional feature mapping strategy for the generator output to project the generated image into the carbon emission distribution space;

[0125] S53. Design the discriminator network structure, use a multi-scale convolutional network to extract features from the input carbon emission distribution map, and combine it with a multi-layer fully connected network to classify the spatiotemporal features of the image to determine whether the input data is real data or generated data;

[0126] S54. Introduce a multi-level spatiotemporal feature loss function, combine multi-dimensional feature mapping and adversarial learning, and optimize the training of the generator:

[0127] ;

[0128] in, represents the loss function of the generator, E represents the mathematical expectation operator, x represents the data sample, represents the distribution of real data, z represents the noise sample, represents the noise distribution, represents the output of the discriminator for real data, Represents the generated data output by the generator, exp represents the exponential function, and represents the adjustment parameter, represents the feature transformation function, represents the Frobenius norm, Tr represents the trace operation of the matrix;

[0129] S55, adopt the alternating optimization strategy for training, fix the parameters of the generator, update the weight of the discriminator; then fix the parameters of the discriminator, update the weight of the generator, and continuously optimize the quality of the generated results through multiple rounds of iterations;

[0130] S56. During the training process, the learning rate and momentum parameters are adaptively adjusted, and finally a generative adversarial network model for generating a carbon emission hotspot distribution map is obtained.

[0131] In this implementation manner, S6 specifically includes:

[0132] S61. Introduce the spatiotemporal attention mechanism to perform weighted fusion of spatiotemporal features at different levels, input short-term spatiotemporal features, medium-term spatiotemporal features and long-term spatiotemporal features into the attention module respectively, dynamically adjust the weights of features at each level, enhance the response capability to key time periods and important areas, and generate fused spatiotemporal features with comprehensive expression capabilities;

[0133] S62. Based on the prediction results of the generative adversarial network, the fused spatiotemporal features are combined with the real-time urban carbon emission data, traffic flow data and meteorological data to form multi-dimensional data input;

[0134] S63. Generate control strategies, consider the changes in real-time data and the dynamic weight distribution of spatiotemporal characteristics, and design preliminary control measures in combination with the predicted information of urban carbon emission hotspots;

[0135] S64. Adaptively adjust the control strategy and dynamically optimize the control parameters according to the real-time monitored changes in the urban environment.

[0136] Embodiment 1:

[0137] In order to verify the feasibility of the present invention in implementation, the present invention is applied to the prediction and regulation of carbon emission hotspots in City A. City A is a rapidly developing industrial city. Traffic congestion and industrial emissions are the main sources of carbon emissions. There are significant spatiotemporal differences in carbon emissions in different areas of the city, especially during weekdays and holidays, when carbon emissions fluctuate greatly. In order to improve urban air quality and achieve carbon reduction goals, the City of A hopes to accurately predict and regulate carbon emission hotspots by introducing intelligent carbon emission management solutions. The present invention provides a new solution for the prediction and dynamic regulation of carbon emission hotspots in City A by combining a generative adversarial network with a multi-level spatiotemporal feature coding method.

[0138] In the carbon emission management scenario of City A, the sources of carbon emission data include traffic flow data, industrial emission data and meteorological data (such as temperature, humidity and wind speed), covering emissions in different areas of the city. The data collection scope includes the main industrial areas, traffic arteries and residential areas in the city, and the data recording frequency is once an hour. In addition, historical data from the past two years are combined to support the analysis and prediction of long-term trends. According to the actual situation, the data preprocessing adopts normalization and denoising processing, and combined with the multi-level spatiotemporal feature coding method proposed in this invention, the data is divided into five scales: hour, day, week, month and quarter, so as to better capture the changing law of carbon emissions.

[0139] In the application process, the generative adversarial network model of the present invention is first used to model the multi-level spatiotemporal characteristics of the carbon emission data of City A, and a control strategy is generated based on the prediction results of the model combined with real-time data. Taking industrial area B and traffic-intensive area C as examples, the model identifies that 7:00 to 9:00 a.m. and 5:00 to 7:00 p.m. on summer working days are the peak hours of carbon emissions. The carbon emissions of industrial area B are mainly affected by the production and transportation activities of enterprises, while the carbon emissions of traffic-intensive area C are mainly due to vehicle emissions during the commuting peak. By identifying these carbon emission hotspots, the model can predict the carbon emission distribution in the next 24 hours in advance and generate a control strategy accordingly.

[0140] The formulation of control strategies mainly includes two aspects: traffic management and industrial emission control. In terms of traffic management, the present invention combines the prediction results of the generative adversarial network and real-time traffic flow data, and proposes to implement temporary restrictions on some roads in Zone C during peak hours, and guide vehicles to bypass low-emission areas. At the same time, it is recommended to increase the departure frequency of public transportation to divert commuters and reduce the use of private cars. In terms of industrial emission control, the model analyzes the production emission data of different enterprises and proposes to make short-term adjustments to the production activities of high-emission enterprises during the peak period of carbon emissions, such as delaying some production tasks or adjusting production shifts. In addition, it is also recommended to strengthen emission control in industrial areas and limit nighttime production emissions of some enterprises when meteorological conditions are not conducive to the diffusion of pollutants (such as windless or low wind speed weather).

[0141] In practical applications, the effectiveness of the present invention is proved by comparative analysis of the model prediction results. For example, in an application test in July 2024, the carbon emission hotspots in Industrial Zone B and Traffic-intensive Zone C in City A were significantly improved. Without taking regulatory measures, the average carbon emissions of Industrial Zone B from July 15 to July 21 were 98.2 tons per hour, and the average carbon emissions of Traffic-intensive Zone C were 65.4 tons per hour; after the method of the present invention was adopted for regulation, the carbon emissions of Industrial Zone B dropped to 81.7 tons per hour, and the carbon emissions of Traffic-intensive Zone C dropped to 53.8 tons per hour, which were 16.8% and 17.7% lower than those without regulation, respectively. In addition, the air quality test results show that after the implementation of regulatory measures, the concentrations of PM2.5 and PM10 in City A were significantly reduced, with the daily average PM2.5 concentration dropping from 45 micrograms / cubic meter to 37 micrograms / cubic meter, and the PM10 concentration dropping from 70 micrograms / cubic meter to 58 micrograms / cubic meter.

[0142] In terms of responding to sudden carbon emission events, the present invention demonstrates excellent flexibility and response speed. On August 1, 2024, due to a sudden traffic accident, a major traffic artery in District C was severely blocked, vehicles were stranded for more than 2 hours, and carbon emissions rose sharply. By generating real-time predictions and feedback from adversarial network models, the control system of the present invention responded immediately, adjusted the traffic management strategy in District C, and quickly diverted vehicles and guided detours. At the same time, after the accident, the investment in surrounding public transportation was increased to cope with sudden commuting needs. Through these measures, carbon emissions in District C returned to normal levels within 3 hours after the accident, while similar emergencies usually take more than 5 hours to be completely alleviated if no measures are taken.

[0143] Table 1 Carbon emission control effects of industrial area B and traffic-intensive area C

[0144]

[0145] Table 2 Air quality improvement effect table

[0146]

[0147] The data in Tables 1 and 2 above show in detail the changes in various indicators before and after the implementation of the method of the present invention. By comparing the data, it is clearly reflected that after the application of the generative adversarial network and the multi-level spatiotemporal feature encoding method, the carbon emissions and air pollution concentrations are significantly reduced, especially the carbon emissions in industrial areas and traffic-intensive areas are significantly controlled, and the air quality is significantly improved. Through practical application in City A, the present invention effectively solves the problems of insufficient prediction accuracy, poor adaptability of control strategies, and untimely response to emergencies in existing carbon emission management methods. The combination of multi-level spatiotemporal feature encoding and generative adversarial networks makes the prediction results more accurate, the control measures more flexible, and significantly improves the efficiency of urban carbon emission management. Practical application data show that the method of the present invention can achieve a significant reduction in carbon emissions, improve urban air quality, and perform well in sudden carbon emission events, effectively enhancing the city's ability to respond to environmental challenges.

[0148] The above description is only a preferred specific implementation manner of the present invention, but the protection scope of the present invention is not limited thereto. Any technician familiar with the technical field can make equivalent replacements or changes according to the technical scheme and inventive concept of the present invention within the technical scope disclosed by the present invention, which should be covered by the protection scope of the present invention.

Claims

1. A method for predicting and regulating urban local carbon emission hotspots based on generative adversarial networks, characterized in that: The steps include: S1. Collect historical carbon emission data, traffic flow data and meteorological data, normalize the collected data, fill in missing values ​​and remove outliers, and perform multi-scale segmentation of time and space to form time series data at hourly, daily, weekly, monthly and quarterly levels; S2. Use hourly carbon emission data to build a short-term memory module to extract the changing trend of short-term carbon emissions, and use a short-term spatiotemporal convolutional network to encode the short-term spatiotemporal features of block-level carbon emission data to detect sudden emission events; S3. Based on daily and weekly carbon emission data, a multi-scale convolutional network is used to extract periodic change characteristics, encode medium-term spatiotemporal features, and cross-integrate the medium-term spatiotemporal features of different regions to analyze the correlation between regions. S4. Use the long-term dependency model to analyze carbon emission data at the monthly and quarterly levels, obtain long-term spatiotemporal characteristics, and globally encode the overall carbon emission distribution of the city; S5. Build and train a generative adversarial network. The generator integrates a multi-level spatiotemporal feature encoding module to generate a carbon emission hotspot distribution map for a specific time period in the future. The discriminator compares the generated distribution map with the real data, performs iterative training through an alternating optimization strategy, introduces a multi-level spatiotemporal feature loss function, and optimizes short-term emergency response and long-term trend prediction. S6. Use the spatiotemporal attention mechanism to weightedly fuse spatiotemporal features at different levels, generate control strategies based on the prediction results of the generative adversarial network and combine real-time data to dynamically optimize urban carbon emission control; S7, monitor the control effect in real time, use feedback data to update the generative adversarial network and control strategy, and make adaptive adjustments and optimizations; The S2 specifically includes: S21. Obtain hourly carbon emission time series data ,in represents the carbon emissions in the ith hour. The data is smoothed and denoised and frequency adjusted using a method based on nonlinear filtering and frequency domain transformation to obtain the smoothed carbon emission sequence. ; S22. Construct a short-term memory module and use the improved long short-term memory network to analyze the smoothed carbon emission sequence. Modeling is performed, a multi-channel gating mechanism and a time decay function are introduced, and the hidden state is calculated: ； in, represents the hidden state at the current time t, L represents the number of nested memory layers, represents the activation function, and represents the weight matrix of the lth layer, represents the input data at the current time t, and represents the bias term of the lth layer, represents bit-wise multiplication of elements, represents the hyperbolic tangent function, Indicates The state of the memory unit in the lth layer at time , represents the time interval between the current time step and the previous time step, represents the parameter set of the time decay function of the lth layer, represents the time decay function; S23. Perform high-order feature enhancement on the extracted short-term trend features, and perform feature transformation by combining feature adaptive amplification and orthogonal transformation: ； in, Indicates the enhanced short-term trend characteristics, Represents the original short-term trend characteristics, represents the adjustment parameter, The local gradient representing the short-term trend characteristics, represents the nonlinear enhancement index, Q and P represent orthogonal matrices; S24. Use a short-term spatiotemporal convolutional network to perform convolution operations on the block-level carbon emission data matrix. Let the block-level carbon emission matrix be M, and use a multi-scale adaptive convolution kernel to extract local spatial features, perform tensor contraction and reweighting: ； in, represents the extracted spatial distribution features, represents the convolution kernel of the sth scale, represents the convolution operation, Tr represents the trace operation of the matrix, represents the scale adaptive adjustment factor; S25. Enhanced short-term trend characteristics And the extracted spatial distribution features Perform nonlinear fusion and adopt adaptive projection fusion strategy: ； in, represents the short-term spatiotemporal features after fusion, represents the nonlinear mapping function, U and V represent the adaptive orthogonal matrix, represents the weight vector, W represents the bias tensor, represents bit-wise multiplication of elements, It means converting the vector into a diagonal matrix. represents the nonlinear index, Represents a tensor decomposition operation; S26, based on the fused short-term spatiotemporal features , apply anomaly detection algorithms to detect carbon emission data, identify and mark abnormal areas and spatiotemporal information; The S3 specifically includes: S31. Obtaining carbon emission time series data matrix at daily and weekly levels ,in represents the carbon emissions of the jth hour of the i-th day, and divides the matrix R into multiple sub-matrices , each sub-matrix represents the carbon emissions in a specific time period and region; S32, using multi-scale convolutional neural network to sub-matrix Row convolution feature extraction, using an adaptive convolution kernel whose size is dynamically adjusted to obtain features of different convolution scales; S33, fuse the features of different convolution scales, and use the adaptive nonlinear feature combination method: ； in, represents the fused features, P represents the number of scales, represents the fusion weight of the p-th scale, represents the local features extracted at the pth scale, Q represents the number of another set of scales, represents the local features extracted at the qth scale, and represents the nonlinear adjustment parameter, represents the characteristic frequency component, Indicates a cycle; S34. The fused features Perform periodic analysis, use wavelet packet transform combined with multi-dimensional phase modulation method, perform multi-band decomposition and phase adjustment, and obtain the multi-frequency feature matrix: ； in, represents the multi-frequency feature matrix, represents the wavelet packet transform operation, represents bit-wise multiplication of elements, represents the phase modulation parameter matrix, represents the characteristic phase mapping matrix, represents the modulation frequency control matrix; S35, multi-frequency feature matrix Perform inverse wavelet packet transform, combine multidimensional nonlinear projection and hierarchical feature reconstruction, and obtain the reconstructed feature representation ; S36. Feature representation after reconstruction of different regions Perform adaptive graph convolutional network processing, analyze the complex correlations between regions through adaptive weighted connections between nodes, and generate the final mid-term spatiotemporal feature encoding ; The S4 specifically includes: S41. Obtain carbon emission data at the monthly and quarterly levels and set it as a time series matrix ,in represents the carbon emissions on the nth day of the mth month, and performs long-term segmentation on the matrix L to divide the data into multiple long-term dependent segments , each segment represents a continuous process of carbon emission change; S42, using the long-term dependency model for each long-term dependency segment Modeling is performed to extract long-term dependent features of different time steps through a multi-head self-attention mechanism, and the long-term feature expression is enhanced by combining time weights and spatial relationships: ； in, represents the global self-attention output, T represents the total number of time steps, represents the time weight of the tth time step, exp represents the exponential function, represents the query matrix, represents the key matrix, represents the value matrix, represents the dimension of the key matrix, represents the adaptive index adjustment parameter; S43, multi-scale time fusion is performed on the extracted long-term spatiotemporal features. An adaptive time fusion strategy is used to decompose the long-term dependent features into feature subsets of multiple time scales, and the features of each scale are re-weighted and combined to obtain the fused features. ; S44. Globally encode the fused long-term spatiotemporal features, using multi-layer nested nonlinear mapping and tensor decomposition strategies: ； in, represents the long-term spatiotemporal characteristics after global encoding, represents a tensor decomposition operation, means converting the result into a diagonal matrix, represents the feature reconstruction function, represents tensor product, represents the activation function; S45. Long-term spatiotemporal features after global encoding Perform multi-layer nested feature enhancement processing, combine with adaptive feature reconstruction algorithm, reproject the features into different high-dimensional spaces, and generate enhanced global features ; S46. The enhanced global features Multi-scale fusion with short-term and medium-term spatiotemporal characteristics is performed to form the final long-term spatiotemporal characteristic representation, which is used to analyze the overall carbon emission trend of the city; The S5 specifically includes: S51, constructing a generative adversarial network, including two modules: a generator and a discriminator, wherein the generator integrates a multi-level spatiotemporal feature encoding module; S52. Design the generator network structure, combine multi-layer convolution and deconvolution layers, encode and decode the input multi-level spatiotemporal features through the spatiotemporal convolution network, generate the carbon emission hotspot distribution map, and use the high-dimensional feature mapping strategy for the generator output to project the generated image into the carbon emission distribution space; S53. Design the discriminator network structure, use a multi-scale convolutional network to extract features from the input carbon emission distribution map, and combine it with a multi-layer fully connected network to classify the spatiotemporal features of the image to determine whether the input data is real data or generated data; S54. Introduce a multi-level spatiotemporal feature loss function, combine multi-dimensional feature mapping and adversarial learning, and optimize the training of the generator: ； in, represents the loss function of the generator, E represents the mathematical expectation operator, x represents the data sample, represents the distribution of real data, z represents the noise sample, represents the noise distribution, represents the output of the discriminator for real data, Represents the generated data output by the generator, exp represents the exponential function, and represents the adjustment parameter, represents the feature transformation function, represents the Frobenius norm, Tr represents the trace operation of the matrix; S55, adopt the alternating optimization strategy for training, fix the parameters of the generator, update the weight of the discriminator; then fix the parameters of the discriminator, update the weight of the generator, and continuously optimize the quality of the generated results through multiple rounds of iterations; S56. During the training process, the learning rate and momentum parameters are adaptively adjusted, and finally a generative adversarial network model for generating a carbon emission hotspot distribution map is obtained.

2. The method for predicting and regulating urban local carbon emission hotspots based on generative adversarial networks according to claim 1 is characterized in that: The S6 specifically includes: S61. Introduce the spatiotemporal attention mechanism to perform weighted fusion of spatiotemporal features at different levels, input short-term spatiotemporal features, medium-term spatiotemporal features and long-term spatiotemporal features into the attention module respectively, dynamically adjust the weights of features at each level, enhance the response capability to key time periods and important areas, and generate fused spatiotemporal features with comprehensive expression capabilities; S62. Based on the prediction results of the generative adversarial network, the fused spatiotemporal features are combined with the real-time urban carbon emission data, traffic flow data and meteorological data to form multi-dimensional data input; S63. Generate control strategies, consider the changes in real-time data and the dynamic weight distribution of spatiotemporal characteristics, and design preliminary control measures in combination with the predicted information of urban carbon emission hotspots; S64. Adaptively adjust the control strategy and dynamically optimize the control parameters according to the real-time monitored changes in the urban environment.

Images