Microgrid energy supply end carbon emission prediction method and system based on time series data

Abstract

The present invention discloses a method and system for predicting carbon emissions at a microgrid energy supply end based on time series data. The method includes: extracting feature subsequences strongly correlated with carbon emission changes from multidimensional original time series data during the operation of the microgrid energy supply end to generate a correlation feature set; inputting the correlation feature set into a hybrid attention network to generate a current weighted feature representation; based on the historical weighted feature representation, using the generative-discriminative collaborative learning carbon emission simulation data enhancement technology to obtain a stable historical enhanced carbon emission data set; training a dual-weighted multi-objective prediction model based on the historical enhanced carbon emission data set, combining a combined gated recurrent unit and a dynamic uncertainty quantification technology for prediction, and outputting multi-objective prediction results of carbon emissions at the microgrid energy supply end on different time scales. By using the embodiments of the present invention, accurate prediction of carbon emissions can be achieved with high precision and flexibility.

Description

Technical Field

[0001] The present invention belongs to the technical field of carbon emission prediction, and in particular to a method and system for predicting carbon emissions at a microgrid energy supply end based on time series data. Background Art

[0002] With the rapid development of the global economy, the problem of carbon emissions has become increasingly prominent and has become an important challenge facing global environmental protection and sustainable development. As an emerging energy supply model, microgrids have become a key solution to achieve low-carbon economic goals due to their high efficiency, flexibility and renewable characteristics. Microgrids are composed of multiple distributed power sources, energy storage systems and load managers, and their operation process involves complex dynamic changes, such as the volatility of renewable energy and the time-varying nature of loads. These factors directly affect the carbon emissions of microgrids. Therefore, accurate prediction of carbon emissions at the energy supply end of microgrids is of great significance for optimizing operation management, reducing carbon emissions and improving energy efficiency.

[0003] However, the carbon emissions of microgrids are affected by many factors, including the volatility of renewable energy and the time-varying nature of loads. Traditional carbon emission prediction methods are often unable to effectively capture these dynamic characteristics and their interrelationships when processing complex multidimensional time series data, resulting in limited prediction capabilities. Summary of the invention

[0004] The purpose of the present invention is to provide a method and system for predicting carbon emissions at the energy supply end of a microgrid based on time series data to address the deficiencies in the prior art. The method and system can achieve accurate prediction of carbon emissions through efficient feature extraction, dynamic modeling and data enhancement technology with high precision and flexibility.

[0005] An embodiment of the present application provides a method for predicting carbon emissions at a microgrid energy supply end based on time series data, the method comprising:

[0006] For the multi-dimensional original time series data during the operation of the microgrid energy supply end, the feature subsequences that are strongly correlated with the changes in carbon emissions are extracted to generate a related feature set;

[0007] The associated feature set is input into a hybrid attention network, which combines the temporal self-attention mechanism and the channel interaction attention mechanism to capture the global temporal dependencies in the time series features and the potential interactions between key channels, and generates a current weighted feature representation to retain the information of the original features while enhancing the correlation with carbon emission changes;

[0008] Generate historical weighted feature representations corresponding to historical multi-dimensional original time series data. Based on the historical weighted feature representations, use the generative-discriminative collaborative learning carbon emission simulation data enhancement technology to obtain a stable historical enhanced carbon emission data set.

[0009] A double-weighted multi-objective prediction model is trained based on the historical enhanced carbon emission dataset. The model dynamically adjusts the historical data and key features of the time series through the time weight mechanism and the associated feature weight mechanism respectively, and combines the combined gated recurrent unit and dynamic uncertainty quantization technology for prediction. Based on the current weighted feature representation, the trained double-weighted multi-objective prediction model is used to output the multi-objective prediction results of carbon emissions at the microgrid energy supply end on different time scales.

[0010] Optionally, the multi-dimensional original time series data during the operation of the microgrid energy supply end is extracted with respect to feature subsequences that are strongly correlated with changes in carbon emissions to generate a correlation feature set, including:

[0011] According to the multi-dimensional original time series data in the operation process of the microgrid energy supply end, the variational mode decomposition technology is used to decompose the multi-dimensional original time series into multiple intrinsic mode functions of different time scales;

[0012] Based on the mutual information quantification technology, the correlation analysis of the decomposed intrinsic mode function components is carried out to extract the characteristic subsequences that are strongly correlated with the changes in carbon emissions and generate a correlation feature set.

[0013] Optionally, the method further comprises: performing dimensionality reduction processing on the weighted feature representation based on a differentiable sparse dimensionality reduction technique to obtain a low-dimensional feature representation with a compact feature structure and significant expressiveness;

[0014] According to the historical weighted feature representation, the carbon emission simulation data enhancement technology of generative-discriminative collaborative learning is used to obtain a historical enhanced carbon emission data set with stable distribution, including:

[0015] According to the historical weighted feature representation, the corresponding historical low-dimensional feature representation is generated. Based on the historical low-dimensional feature representation, the carbon emission simulation data enhancement technology of generative-discriminative collaborative learning is used to obtain a historical enhanced carbon emission dataset with stable distribution.

[0016] Optionally, the carbon emission simulation data enhancement technology based on historical low-dimensional feature representation and generative-discriminative collaborative learning is used to obtain a historical enhanced carbon emission dataset with stable distribution, including:

[0017] Input the historical low-dimensional feature representation into a generator, and generate simulated carbon emission data through a deep random nested residual generation technique;

[0018] An adaptive distributed adversarial loss function is introduced through the discriminator to improve the distribution consistency between the simulated carbon emission data and the real observed carbon emission data. After repeated iterative optimization, an enhanced carbon emission data set with stable distribution is obtained to avoid the sparse and uneven distribution of carbon emission data at the microgrid energy supply end.

[0019] Optionally, the multi-objective prediction results of carbon emissions at the energy supply end of the microgrid on different time scales are outputted using the trained double-weighted multi-objective prediction model based on the current weighted feature representation, including:

[0020] Based on the current low-dimensional feature representation, the trained dual-weighted multi-objective prediction model is used to output the multi-objective prediction results of carbon emissions at the microgrid energy supply end on different time scales.

[0021] Optionally, the mutual information quantization technology is used to perform correlation analysis on the decomposed intrinsic mode function components, extract feature subsequences that are strongly correlated with carbon emission changes, and generate a correlation feature set, including:

[0022] For each decomposed intrinsic mode function of different time scales, the mutual information value between the mode function and the actual carbon emission value of the corresponding time scale is calculated;

[0023] The intrinsic mode functions whose mutual information values ​​are greater than a preset significance threshold are screened, and the time series portion corresponding to the screened intrinsic mode functions is extracted as a feature subsequence strongly correlated with carbon emission changes, and all extracted feature subsequences are integrated into a correlation feature set.

[0024] Another embodiment of the present application provides a microgrid energy supply end carbon emission prediction system based on time series data, the system comprising:

[0025] An extraction module is used to extract feature subsequences that are strongly correlated with carbon emission changes from the multi-dimensional original time series data during the operation of the microgrid energy supply end, so as to generate a correlation feature set;

[0026] A generation module, used for inputting the associated feature set into a hybrid attention network, which combines the temporal self-attention mechanism and the channel interaction attention mechanism to capture the global temporal dependencies in the time series features and the potential interactions between key channels, and generates a current weighted feature representation to retain the information of the original features while enhancing the correlation with the change of carbon emissions;

[0027] The enhancement module is used to generate historical weighted feature representations corresponding to historical multi-dimensional original time series data. Based on the historical weighted feature representations, the carbon emission simulation data enhancement technology of generative-discriminative collaborative learning is used to obtain a historical enhanced carbon emission data set with stable distribution;

[0028] The prediction module is used to train a dual-weighted multi-objective prediction model based on the historical enhanced carbon emission data set. The model dynamically adjusts the historical data and key features of the time series through the time weight mechanism and the associated feature weight mechanism, respectively, and combines the combined gated recurrent unit and dynamic uncertainty quantification technology for prediction. Based on the current weighted feature representation, the trained dual-weighted multi-objective prediction model is used to output the multi-objective prediction results of carbon emissions at the microgrid energy supply end on different time scales.

[0029] Yet another embodiment of the present application provides a storage medium, wherein the storage medium stores a computer program, wherein the computer program is configured to execute any of the above methods when running.

[0030] Yet another embodiment of the present application provides an electronic device, including a memory and a processor, wherein the memory stores a computer program, and the processor is configured to run the computer program to execute any of the methods described above.

[0031] Compared with the prior art, the present invention provides a method for predicting carbon emissions at the energy supply end of a microgrid based on time series data. For the multi-dimensional original time series data during the operation of the energy supply end of the microgrid, feature subsequences that are strongly correlated with carbon emission changes are extracted to generate an associated feature set; the associated feature set is input into a hybrid attention network to generate a current weighted feature representation; based on the historical weighted feature representation, a generative-discriminative collaborative learning carbon emission simulation data enhancement technology is used to obtain a historical enhanced carbon emission data set with stable distribution; a dual-weighted multi-objective prediction model is trained based on the historical enhanced carbon emission data set, and predictions are performed in combination with a combined gated recurrent unit and dynamic uncertainty quantification technology, and multi-objective prediction results of carbon emissions at the energy supply end of the microgrid on different time scales are output, thereby enabling accurate prediction of carbon emissions through efficient feature extraction, dynamic modeling and data enhancement technology, with high precision and flexibility. BRIEF DESCRIPTION OF THE DRAWINGS

[0032] Figure 1 A hardware structure block diagram of a computer terminal for a method for predicting carbon emissions at a microgrid energy supply end based on time series data provided in an embodiment of the present invention;

[0033] Figure 2 A schematic flow chart of a method for predicting carbon emissions at a microgrid energy supply end based on time series data provided by an embodiment of the present invention;

[0034] Figure 3 A structural schematic diagram of a carbon emission prediction system for a microgrid energy supply end based on time series data provided in an embodiment of the present invention. DETAILED DESCRIPTION

[0035] The embodiments described below with reference to the accompanying drawings are exemplary and are only used to explain the present invention, but should not be construed as limiting the present invention.

[0036] The embodiment of the present invention first provides a method for predicting carbon emissions at the energy supply end of a microgrid based on time series data. The method can be applied to electronic devices, such as computer terminals, specifically ordinary computers, etc.

[0037] The following describes it in detail by taking running on a computer terminal as an example. Figure 1 The hardware structure block diagram of a computer terminal for a method for predicting carbon emissions at a microgrid energy supply end based on time series data provided by an embodiment of the present invention. Figure 1 As shown, the computer device includes a processor, a memory, and a network interface connected via a system bus, wherein the memory may include a non-volatile storage medium and an internal memory.

[0038] The non-volatile storage medium can store an operating system and a computer program. The computer program includes program instructions, and when the program instructions are executed, the processor can execute any method for predicting carbon emissions at the energy supply end of a microgrid based on time series data.

[0039] The processor is used to provide computing and control capabilities and support the operation of the entire computer equipment.

[0040] The internal memory provides an environment for the operation of the computer program in the non-volatile storage medium. When the computer program is executed by the processor, the processor can execute any carbon emission prediction method for the microgrid energy supply end based on time series data.

[0041] The network interface is used for network communication, such as sending assigned tasks, etc. Those skilled in the art will understand that Figure 1 The structure shown in the figure is only a block diagram of a part of the structure related to the solution of the present application, and does not constitute a limitation on the computer device to which the solution of the present application is applied. The specific computer device may include more or fewer components than those shown in the figure, or combine certain components, or have a different arrangement of components.

[0042] It should be understood that the processor may be a central processing unit (CPU), and the processor may also be other general-purpose processors, digital signal processors (DSP), application-specific integrated circuits (ASIC), field-programmable gate arrays (FPGA) or other programmable logic devices, discrete gate or transistor logic devices, discrete hardware components, etc. Among them, the general-purpose processor may be a microprocessor or the processor may also be any conventional processor, etc.

[0043] See also Figure 2 , an embodiment of the present invention provides a method for predicting carbon emissions at a microgrid energy supply end based on time series data, which may include the following steps:

[0044] S201, extracting feature subsequences that are strongly correlated with carbon emission changes from the multi-dimensional original time series data during the operation of the microgrid energy supply end, to generate a correlation feature set;

[0045] The operation of the microgrid power supply involves the collection of multi-dimensional raw time series data, including power generation, load demand, environmental factors, and other variables that affect carbon emissions. This step generates a set of associated features by extracting feature subsequences that are strongly correlated with changes in carbon emissions. The selection and extraction of feature subsequences are key because they directly affect the predictive ability of subsequent models. By analyzing these time series data, the system can identify features that have a significant impact on changes in carbon emissions, thereby providing support for subsequent modeling and analysis. The goal of this process is to ensure that the selected features truly reflect the dynamic changes in carbon emissions, thereby improving the accuracy of the prediction.

[0046] The importance of extracting feature subsequences that are strongly correlated with carbon emission changes and generating associated feature sets is that it can significantly improve the performance of carbon emission prediction models. Selecting accurate features can filter out noise and irrelevant information, allowing the model to focus more on key factors, thereby improving the accuracy and stability of predictions. In addition, by focusing on these features, the model can better identify the potential connection between carbon emissions and the operating status of the energy supply end, providing a more effective data basis for the design of decision support systems. This method not only provides a new perspective for the study of carbon emissions in microgrids, but also has practical application value, and can provide support for environmental management and optimized scheduling strategies.

[0047] Specifically, the multidimensional original time series data during the operation of the microgrid energy supply end can be decomposed into multiple intrinsic mode functions of different time scales using variational mode decomposition technology;

[0048] In the prediction of carbon emissions at the microgrid power supply end, variational mode decomposition technology is used to decompose multidimensional original time series data into multiple intrinsic mode functions of different time scales. The core of this technology is that it can decompose complex time series signals into a series of physically meaningful intrinsic mode functions, which respectively reflect the different frequency components in the data and their changing characteristics. Through this decomposition, the potential patterns in the time series can be deeply explored, which is convenient for subsequent correlation analysis and feature extraction. Variational mode decomposition not only improves the interpretability of features, but also reduces the impact of high-frequency noise on the analysis results, making data processing more accurate and reliable.

[0049] The use of variational mode decomposition technology for data processing is of great significance. It can effectively extract the periodic and non-periodic components in time series data, allowing more complex data characteristics to emerge. This method helps to improve the understanding of carbon emission changes and provides multi-level information support for subsequent feature extraction. By analyzing the intrinsic modes at different time scales, the key factors related to carbon emission changes can be better identified. This decomposition not only provides rich feature information for the model, but also provides a more effective data basis for actual carbon emission monitoring and prediction, thereby providing reliable support for more accurate environmental management and decision-making.

[0050] Before implementing the variational mode decomposition technology, it is necessary to first preprocess the multidimensional raw time series data collected from the microgrid power supply end. This process includes data cleaning, missing value filling and anomaly detection to ensure the quality and integrity of the input data. Next, the preprocessed data is organized into a format suitable for decomposition, usually a multidimensional array or matrix, containing the sampled values ​​of different variables (such as power generation, load, grid frequency, ambient temperature, etc.) in the same time period. After the preprocessing is completed, select the appropriate variational mode decomposition algorithm and set the necessary parameters, such as the number of decomposition levels and stop conditions. The choice of these parameters directly affects the quality of the decomposition results, so they need to be adjusted and optimized according to the specific data characteristics.

[0051] Next, the preprocessed multidimensional time series data is decomposed using the variational mode decomposition (VMD) algorithm. The core of the VMD algorithm is its ability to decompose complex signals into a set of intrinsic mode functions (IMFs), each of which corresponds to a different frequency range and time scale. Specifically, the algorithm first sets a fixed number of iterations, and then gradually adjusts the center frequency and bandwidth of each intrinsic mode through an iterative optimization process, so that each mode in the frequency domain retains the characteristics of the time series as much as possible and is independent of other modes. In each iteration, the algorithm uses the Lagrange multiplier method to handle constraints and updates the mode function by minimizing the reconstruction error. This adaptive decomposition process ensures that the final generated IMFs can accurately reflect the time-varying characteristics and frequency distribution of the input signal.

[0052] Finally, after obtaining multiple intrinsic mode functions, they are further evaluated and screened to determine which IMFs are more relevant to carbon emission changes. This process requires evaluating the relationship between each IMF and the actual carbon emission data through mutual information quantification analysis, correlation analysis and other methods. A significance threshold can be set to screen out intrinsic mode functions with higher mutual information values. These related IMFs will be integrated into the final set of associated features for subsequent feature extraction and carbon emission prediction models. Through this series of refined steps, the variational mode decomposition technology not only improves the accuracy of data analysis, but also provides strong data support for carbon emission monitoring at the energy supply end of the microgrid.

[0053] Based on the mutual information quantification technology, the correlation analysis of the decomposed intrinsic mode function components is carried out to extract the characteristic subsequences that are strongly correlated with the changes in carbon emissions and generate a correlation feature set.

[0054] In the carbon emission prediction method at the energy supply end of the microgrid, the correlation analysis of the decomposed intrinsic mode function components is performed based on the mutual information quantification technology, aiming to identify the characteristic subsequences that are strongly correlated with the changes in carbon emissions. Mutual information is a statistical method used to measure the dependence between two random variables. By calculating the mutual information value between the intrinsic mode function and the actual carbon emission value, the mutual dependence between the characteristic function and the carbon emission can be evaluated. This process can accurately identify the key factors that affect the changes in carbon emissions by quantifying the amount of information shared between the feature and the target variable (carbon emissions), thereby providing a theoretical basis for subsequent feature extraction and model training.

[0055] Through correlation analysis based on mutual information quantification technology, the accuracy and effectiveness of feature selection can be significantly improved. Extracting feature subsequences that are strongly correlated with carbon emission changes can not only enhance the quality of model training, but also reduce the complexity of the model and improve its effectiveness. By focusing on important features, the impact of data noise on the prediction results can be reduced, thereby improving the interpretability and stability of the model. This method can provide more accurate predictions in carbon emission prediction of microgrids and provide a solid theoretical basis for practical applications, thereby contributing to sustainable energy management and decision support.

[0056] Specifically, the correlation analysis of the decomposed intrinsic mode function components is performed based on the mutual information quantification technology, and the characteristic subsequences strongly correlated with the carbon emission changes are extracted to generate a correlation feature set. For each decomposed intrinsic mode function of different time scales, the mutual information value between the mode function and the actual carbon emission value of the corresponding time scale can be calculated;

[0057] In this step, it is first necessary to perform statistical analysis on each intrinsic mode function (IMF) to quantify its relationship with the actual carbon emission value at the corresponding time scale. The calculation of the mutual information value can reveal the dependency between the two variables and reflect the amount of information carried in the modal function. In this way, we can obtain the mutual influence between the modal function and carbon emissions at different time scales, thereby providing a basis for subsequent feature screening. The core significance of calculating the mutual information value is that it can help screen out features that are significantly related to changes in carbon emissions. This step not only reduces noise interference, but also provides high-quality input data for subsequent analysis and modeling. By understanding the specific degree of correlation between the modal function and carbon emissions, researchers can more accurately grasp the potential laws in the time series data, thereby laying a solid foundation for carbon emission prediction.

[0058] In the first step of realizing the calculation of mutual information value, it is necessary to fully prepare and clean the multi-dimensional original time series data collected during the operation of the microgrid power supply end. These data may include power load, climate change, equipment status, and renewable energy power generation. In order to ensure data quality, missing values ​​and outliers are first detected and processed. Missing values ​​can be filled by interpolation, while outliers can be identified and eliminated by statistical methods such as Z-score method. After data cleaning is completed, the variational mode decomposition (VMD) technology is applied to decompose the multi-dimensional time series data into multiple intrinsic mode functions (IMFs). Each IMF represents a specific frequency component to ensure that information within different time scales is captured.

[0059] The key to quantitatively analyzing the mutual information of the decomposed IMF is to calculate the joint probability distribution between each IMF and the actual carbon emission value. In this process, the numpy and scipy libraries in Python can be used to calculate the joint probability distribution by constructing histograms and performing probability density estimation. In addition, the marginal probability distribution is also indispensable. The marginal probability is the probability distribution of a single variable. In this analysis, it is the basis for understanding the relationship between a single IMF and carbon emissions. After completing the calculation of the joint probability distribution and the marginal probability distribution, the mutual information value between each IMF and the carbon emission value is calculated using the formula of mutual information, and these values ​​are stored in a data frame for subsequent screening.

[0060] After completing the calculation of the mutual information value, it is very important to conduct a preliminary analysis of the results, which can help us understand the strength of the relationship between each IMF and carbon emissions. Visualization methods, such as heat maps, can be used to display the mutual information values ​​between different IMFs and actual carbon emissions values, so as to quickly identify features with strong correlations. This step not only lays the foundation for subsequent feature screening, but also provides researchers with a clear perspective to better understand the potential relationships contained in the data and provide strong data support for subsequent analysis and modeling.

[0061] The intrinsic mode functions whose mutual information values ​​are greater than a preset significance threshold are screened, and the time series portion corresponding to the screened intrinsic mode functions is extracted as a feature subsequence strongly correlated with carbon emission changes, and all extracted feature subsequences are integrated into a correlation feature set.

[0062] In this step, by setting a predefined significance threshold, the calculated mutual information values ​​are screened to ensure that there is a significant relationship between the selected intrinsic mode function and the change in carbon emissions. The time series portion corresponding to the qualified mode function will be extracted as an important feature subsequence. Finally, these qualified feature subsequences are integrated into a correlation feature set for subsequent modeling and prediction. The implementation of this step can effectively filter out information redundancy and ensure that the data based on which the subsequent analysis is based is highly relevant and valid. By focusing on the extraction of significant features, researchers can focus on factors that play an important role in carbon emission changes, thereby improving the prediction accuracy and reliability of the model. In addition, the formed correlation feature set will provide a solid data foundation for subsequent algorithm design and model training.

[0063] When screening feature subsequences, we first need to set the significance threshold of the mutual information value. Usually, we can choose a significance level of 95% as the standard. Next, we screen each IMF one by one for the mutual information value calculated in the first step, and extract those IMFs whose mutual information values ​​are higher than the preset threshold. This screening process can be implemented by writing a program, such as using Python's pandas library to traverse the result data frame of the mutual information value, screen out the IMF indexes that meet the conditions and record them. In this way, we can effectively identify features that are significantly associated with changes in carbon emissions, thereby reducing the noise in the data and improving the accuracy of subsequent analysis.

[0064] After screening out the IMFs with strong correlation, the next step is to extract the time series parts corresponding to these IMFs from the original time series. The extraction process must ensure the consistency of the time series to avoid introducing unnecessary interference in feature selection. In specific implementation, the recorded IMF index can be used to extract the corresponding time series data from the original data set. After the extraction is completed, these feature subsequences are integrated into a new data frame to construct a "correlated feature set". In this process, we also need to make formal adjustments to the integrated time series data, such as ensuring that the lengths of each time series are the same, to ensure that no errors will occur due to inconsistent data during subsequent model training.

[0065] Finally, further visualization and analysis of the obtained associated feature set is very important. By using visual charts (such as time series charts) to show the relationship between the extracted feature subsequences and the actual carbon emission values, researchers can intuitively understand the changing trend of the features. In addition, consider standardizing the feature subsequences to make the data more suitable for subsequent model training and improve the model's learning efficiency on the data. So far, through the steps of screening and extracting feature subsequences, we have successfully established a feature set containing valuable information, which can provide strong data support for subsequent carbon emission forecasts and further promote carbon emission management and optimization at the microgrid energy supply end.

[0066] S202, inputting the associated feature set into a hybrid attention network, which combines the temporal self-attention mechanism and the channel interaction attention mechanism to capture the global temporal dependency in the time series features and the potential interaction between key channels, and generates a current weighted feature representation to retain the information of the original features while enhancing the correlation with the change in carbon emissions;

[0067] In the carbon emission prediction method for the microgrid energy supply end, the input of the associated feature set to the hybrid attention network is a crucial step. The network combines the temporal self-attention mechanism and the channel interaction attention mechanism to capture the global temporal dependencies in the input time series features and the potential interactions between different feature channels. The core of this process is that through the self-attention mechanism, the hybrid attention network model can automatically learn which time points are more important in the sequence, so as to better understand the impact of past data on future carbon emission trends. At the same time, the channel interaction attention mechanism enables the information between different feature channels to be effectively integrated, helping the model to identify synergies and thus improve prediction accuracy.

[0068] The introduction of this step significantly improves the hybrid attention network model's ability to capture dynamic changes in carbon emissions. By considering the global perspective of time series data and the interaction between features, the hybrid attention network can not only maintain the information integrity of the original features, but also enhance the correlation with carbon emission changes. In practical applications, this can help decision makers grasp the operating status of microgrids in real time and accurately, and provide a scientific basis for formulating emission reduction strategies, optimizing power supply, and enhancing the use of renewable energy.

[0069] First, the associated feature set received by the hybrid attention network model is usually a multidimensional time series dataset, which contains multiple feature subsequences related to carbon emission changes. These features may include power load, climate change (such as temperature, humidity), renewable energy generation, etc. In the hybrid attention network, each dimension of the feature set is regarded as a channel.

[0070] First, through the temporal self-attention mechanism, the model calculates the correlation between each time point. Specifically, to achieve this, the model first linearly transforms the input data to obtain three matrices: query (Q), key (K), and value (V). By calculating the dot product of Q and K and subsequent softmax processing, the model obtains the weight distribution of different time points, which expresses the importance of different time points in information extraction at the current time point.

[0071] Subsequently, the channel interaction attention mechanism was introduced to focus on the relationship between different feature channels. It assigns channel weights to the original features, indicating that some features may be more influential in the overall carbon emission changes. For example, climate factors (such as temperature and humidity) may have a more significant impact on a specific load change. This mechanism can dynamically adjust the influence of feature channels through a similar attention mechanism.

[0072] After these two steps, the weighted feature representation obtained not only retains the temporal information of the time series, but also enhances the correlation between different features. The output of this process is a new feature representation, which will be used in conjunction with other model components (such as carbon emission simulation generators and forecasting models) to improve the accuracy and stability of the entire carbon emission forecasting system. In practical applications, for example, if the impact of weather factors on carbon emissions is significantly enhanced during a specific time period, the system can respond quickly and dynamically adjust the forecast results, providing a more accurate decision-making basis for the operation of the microgrid.

[0073] Furthermore, the weighted feature representation can be reduced in dimension based on the differentiable sparsification dimensionality reduction technology to obtain a low-dimensional feature representation with a compact feature structure and significant expressiveness.

[0074] Differentiable sparse dimensionality reduction technology reduces the dimensionality of weighted feature representations, aiming to reduce feature dimensions while retaining key information. This process streamlines the feature set by introducing sparsity constraints, so that the final low-dimensional feature representation is not only compact and convenient for subsequent modeling, but also significantly improves the expressiveness of the model. Specifically, sparse dimensionality reduction technology identifies the most critical features for carbon emission prediction, removes redundant or noisy features, and optimizes the feature space. This provides an efficient way to ensure that the model can focus on the most representative features, making it possible to maintain prediction accuracy and computational efficiency on large-scale data sets.

[0075] The introduction of differentiable sparse dimensionality reduction technology aims to improve the efficiency and accuracy of the model. High-dimensional data not only increases the computational complexity, but may also cause "dimensionality disaster", making it difficult for the model to capture effective patterns. By reducing the dimension of the data through dimensionality reduction, the model can operate in a simpler feature space, thereby reducing the risk of overfitting and improving generalization ability. In addition, the compact feature structure can significantly accelerate the training process of subsequent models, ensuring rapid response and output results in real-time carbon emission monitoring and decision-making.

[0076] When implementing the differentiable sparsification dimensionality reduction technique, first, the weighted feature representations from the hybrid attention network are accepted as input. These feature representations are usually high-dimensional and contain a lot of feature information related to carbon emission changes. The next step is to build a sparsification network that can guide feature selection by minimizing the loss function. The construction of the loss function usually combines the reconstruction error with the sparsity constraint, and encourages feature sparsification through L1 regularization. This process can be understood as dividing the features into two parts: important features and unimportant features. Through the optimization algorithm, the model automatically learns how to adjust the weights during training so that the weights of unimportant features approach zero, thereby achieving feature sparsification.

[0077] During the algorithm update process, the model will continuously iterate, update the feature representation, and ensure that effective features that are highly correlated with carbon emission changes can be effectively captured through back propagation. Taking a microgrid instance as an example, if the features that are significantly related to carbon emission changes include wind speed and solar radiation, these features will be retained after sparse processing, while some redundant features (such as power fluctuation range) will be removed. The final output is a low-dimensional feature representation, which not only retains important information related to carbon emissions, but also greatly reduces the dimension of the features. The feature representation processed in this way will provide a simplified and effective foundation for the subsequent generation of historical enhanced carbon emission data sets and the training of dual-weighted multi-objective prediction models. In this way, the model not only improves the response speed to dynamic changes in carbon emissions, but also reduces the computational burden when processing and analyzing data, ensuring that the carbon emission prediction task can still be effectively completed in a resource-constrained environment.

[0078] S203, generating a historical weighted feature representation corresponding to the historical multi-dimensional original time series data, and using the generative-discriminative collaborative learning carbon emission simulation data enhancement technology to obtain a historical enhanced carbon emission data set with stable distribution according to the historical weighted feature representation;

[0079] In this method, firstly, a historical weighted feature representation generated based on historical multi-dimensional original time series data is used to provide an efficient way to extract features related to carbon emission changes. These weighted features can better reflect the dynamic change characteristics in time series data, thus providing a basis for subsequent data enhancement. Then, using the generative-discriminative collaborative learning carbon emission simulation data enhancement technology, a stable and representative historical enhanced carbon emission dataset is obtained by analyzing and reconstructing the historical weighted features. This dataset not only overcomes the problems of sparse and uneven distribution of original data, but also provides higher quality input data for model training, improving the accuracy and reliability of predictions.

[0080] By generating historical weighted feature representations and using generative-discriminative collaborative learning techniques to obtain historically enhanced carbon emission data sets, the carbon emission prediction performance of the microgrid energy supply end can be significantly improved. This process solves the sparsity and unevenness of time series data to a certain extent, ensures the distribution stability and consistency of the data set, and thus provides a solid foundation for subsequent training. This not only improves the learning ability of the model, but also effectively reduces the errors that may be caused by insufficient data during the prediction process, making carbon emission predictions more accurate.

[0081] Specifically, the corresponding historical low-dimensional feature representation can be generated according to the historical weighted feature representation. Based on the historical low-dimensional feature representation, the carbon emission simulation data enhancement technology of generative-discriminative collaborative learning is used to obtain a historical enhanced carbon emission data set with stable distribution.

[0082] The step of generating the corresponding historical low-dimensional feature representation based on the historical weighted feature representation involves further dimensionality reduction of the weighted features, aiming to retain the most informative and representative features while reducing the complexity of the data. In this process, the differentiable sparse dimensionality reduction technique used can effectively remove redundant features, thereby improving the interpretability of the data and the training efficiency of the model. The generated historical low-dimensional feature representation will provide a more compact feature space for the subsequent generative-discriminative collaborative learning process, allowing the generator and discriminator to focus on more important feature information. This transformation not only helps to improve the authenticity of the generated carbon emission simulation data, but also reduces the negative impact of data sparsity and uneven distribution, so that the generative model has stronger adaptability and accuracy when processing carbon emission data at the microgrid power supply end.

[0083] The use of generative-discriminative collaborative learning based on historical low-dimensional feature representation to obtain a stable historical enhanced carbon emission data set has important practical significance. First, this process can effectively alleviate the difficulty of model training caused by sparse or uneven distribution of original data, thereby improving the generalization ability of the prediction model. Secondly, the generated enhanced data set is not just a simple extension of the original data, but a rich data source for training the model by intelligently generating simulated samples that are consistent with the characteristics of real data. This highly integrated generation technology can ensure that the generated samples are consistent with the actual observed data in terms of statistical characteristics and distribution, thereby significantly improving the accuracy and reliability of subsequent models in carbon emission prediction tasks. Finally, the stable historical enhanced carbon emission data set provides solid data support for the management and optimization decision-making of the microgrid energy supply end, allowing relevant parties to make more accurate policy formulation and resource allocation, thereby further promoting the goal of low-carbon development and sustainable energy utilization. Through this technical solution, the carbon emission prediction of microgrids is not only scientific, but also enhances the operability in practice, which has a far-reaching impact on promoting future energy transformation.

[0084] Specifically, based on the historical low-dimensional feature representation, the carbon emission simulation data enhancement technology of generative-discriminative collaborative learning is used to obtain a historical enhanced carbon emission data set with stable distribution. The historical low-dimensional feature representation can be input into the generator to generate simulated carbon emission data through the deep random nested residual generation technology;

[0085] In this step, the historical low-dimensional feature representation is first input into the generator, which uses a deep random nested residual network structure to capture complex patterns and potential relationships in the input features. The network effectively solves the problem of information attenuation in inter-layer transmission through multi-level residual connections, thereby generating more realistic and diverse simulated carbon emission data. The implementation of this step ensures that the generated simulated carbon emission data can fully reflect the main features and trends in the historical data, and improves the diversity and authenticity of the data. The simulated data provides a rich data source for subsequent training, enhances the generalization ability of the model, and reduces the dependence on real data, which is particularly important when real data is sparse.

[0086] First, we need to clarify how to construct the historical low-dimensional feature representation. This feature representation is usually obtained by reducing the dimensionality of historical time series data. Methods such as principal component analysis (PCA), t-SNE, or autoencoders can be used to compress high-dimensional time series data into low-dimensional representative feature vectors. In this process, special attention should be paid to feature selection and transformation to ensure that the information that best reflects the changes in carbon emissions is retained. In addition, the features need to be standardized so that their mean is 0 and the variance is 1 to improve the convergence speed and stability of the subsequent generation model.

[0087] Next, a deep random nested residual generator model is constructed. The core of this model is to introduce residual connections (ResNet), which reduces the loss of information when it propagates in the network by establishing jump connections between different layers. Specifically, the generator can be designed to consist of multiple convolutional layers and fully connected layers alternating, and each convolutional layer is followed by an activation function (such as ReLU) and a batch normalization layer to maintain the fluidity and distribution consistency of the data. At the end of the network, a fully connected layer is used to map the feature dimension back to the dimension of the target carbon emission data to ensure that the generated data can reflect the characteristics of actual carbon emissions.

[0088] Finally, during the training process, the generator should be evaluated regularly to monitor its performance. The generated simulated carbon emission data can be compared with the actual carbon emission data, and indicators such as the mean square error (MSE) can be calculated to adjust the learning rate and training strategy of the generator in a timely manner. In order to further improve the diversity of simulated data, random noise can be introduced during training to increase the variation of generated results. This approach helps to enhance the robustness of the model, so that it can still maintain a high generation quality when facing complex inputs in the real world.

[0089] An adaptive distributed adversarial loss function is introduced through the discriminator to improve the distribution consistency between the simulated carbon emission data and the real observed carbon emission data. After repeated iterative optimization, an enhanced carbon emission data set with stable distribution is obtained to avoid the sparse and uneven distribution of carbon emission data at the microgrid energy supply end.

[0090] This step uses the discriminator to provide real-time feedback on the generated simulated data to evaluate its similarity to the real observed data. The discriminator adjusts the output of the generator through the adversarial loss function during the training process to make it closer to the real data in distribution. This adversarial learning mechanism promotes continuous optimization of the model, thereby generating an enhanced carbon emission data set with a more stable distribution. The key to this step is to ensure that the generated carbon emission data is not only realistic in form, but also consistent with the actual observed data in statistical characteristics. This can effectively reduce errors in the model training process and avoid the decline in prediction performance due to inconsistent input data distribution. In addition, the stability of the enhanced data set also enhances the adaptability of the model in different situations, providing more reliable data support for practical applications.

[0091] First, we need to build a discriminator model, whose task is to evaluate whether the input carbon emission data is real or generated. The discriminator is generally composed of multiple convolutional layers and fully connected layers to extract feature information from the data. In order to improve the discriminative ability, the last layer of the network usually uses a sigmoid activation function to output a probability value indicating the possibility that the input sample belongs to real data. During the training process, the discriminator will continue to receive real observation data and generated simulated data for learning and adjustment, thereby improving its ability to distinguish.

[0092] Secondly, it is critical to design an adaptive distribution adversarial loss function. This loss function can dynamically adjust the training target of the generator based on the feedback from the discriminator. This means that when the discriminator is more strict in judging the generated data, the generator will receive greater loss feedback and will be forced to generate more realistic carbon emission data. In specific implementations, the cross entropy loss function can be used as part of the adversarial loss and combined with other losses (such as the reconstruction loss of the generator) to form a comprehensive loss function. In this way, the generator and the discriminator form a dynamic optimization game during the training process, driving the improvement of their respective performance.

[0093] Finally, multiple iterations of optimization are performed to ensure that the generated data set is consistent with the real data in distribution. In each round of training, the generator and the discriminator are updated alternately to maintain their competitive relationship. To avoid overfitting of the model, an early stopping strategy can be used to set the upper limit of the number of training rounds and the performance indicators of the validation set. When the performance on the validation set no longer improves, the training can be terminated. In addition, applying data enhancement techniques (such as random cropping, rotation, etc.) to the input of the discriminator can further enhance the adaptability of the model. Through a series of optimizations and adjustments, a stable and high-quality enhanced carbon emission data set is finally obtained, which provides a solid data foundation for carbon emission prediction at the energy supply end of the microgrid.

[0094] S204, training a dual-weighted multi-objective prediction model based on the historical enhanced carbon emission data set. The model dynamically adjusts the historical data and key features of the time series through the time weight mechanism and the associated feature weight mechanism respectively, combines the combined gated recurrent unit and the dynamic uncertainty quantization technology for prediction, and based on the current weighted feature representation, uses the trained dual-weighted multi-objective prediction model to output the multi-objective prediction results of carbon emissions at the microgrid energy supply end on different time scales.

[0095] In the step of training a dual-weighted multi-objective prediction model based on a historical enhanced carbon emission dataset, the model uses a time weight mechanism and an associated feature weight mechanism to dynamically adjust the historical data and key features of the time series. This dynamic adjustment mechanism enables the model to more accurately capture the importance of features that change over time in the time series. For example, in a certain time period, electricity demand may increase significantly due to seasonal changes, and the corresponding carbon emissions will also increase. Through the time weight mechanism, the model can automatically strengthen the influence of data in that time period. At the same time, the associated feature weight mechanism ensures that features that are closely related to carbon emission changes are given priority in the prediction. Combined with the combined gated recurrent unit and dynamic uncertainty quantification technology, the model can effectively predict and output the prediction results of carbon emissions at the microgrid supply end at different time scales based on the current weighted feature representation, such as short-term (hourly level), medium-term (daily level) and long-term (monthly level) predictions. This enables microgrid operators to flexibly adjust energy management strategies for different time scales.

[0096] The core role of this dual-weighted multi-objective prediction model is to enhance the microgrid's predictive ability in carbon emission management, thereby supporting more scientific decision-making. Through accurate carbon emission forecasts, microgrid operators can formulate more effective energy allocation strategies. For example, during the expected peak period of electricity demand, timely adjust the power generation output of renewable energy and reduce the use of traditional fossil fuel power generation, thereby reducing the overall carbon emission level. In addition, the flexibility and adaptability of the model enable it to respond quickly and make adjustments under changing conditions such as policy environment, market demand and technological progress, thereby ensuring that microgrids play a key role in achieving carbon neutrality goals. This not only brings economic benefits to enterprises, but also makes positive contributions to social environmental protection and sustainable development.

[0097] When implementing the dual-weighted multi-objective prediction model, it is first necessary to build a comprehensive historical enhanced carbon emission data set, which should cover various operational data of the microgrid energy supply end, such as power demand, power generation sources (including wind power, solar power, traditional fuels), meteorological data, etc. Next, using these historical data, through the time series feature extraction process, generate current weighted feature representation and historical weighted feature representation, which will be used for model training.

[0098] In terms of model design, the combined gated recurrent unit (GRU) is used to process time series data because it excels in capturing dynamic time series information. The core of the model is the combination of the time weight mechanism and the associated feature weight mechanism. The time weight mechanism enables the model to dynamically adjust the importance of historical data at different time stages by introducing learnable parameters. For example, during peak periods of electricity demand, the model will give higher weights to data in this period to ensure its influence on carbon emission forecast results. At the same time, the associated feature weight mechanism ensures that features that are closely related to changes in carbon emissions are given priority in the forecast, thereby improving the forecast accuracy.

[0099] During the training process, dynamic uncertainty quantification technology will be integrated into the model architecture. This technology quantifies the uncertainty of the model's prediction results by establishing a framework for uncertainty estimation. This process includes using Bayesian methods or Monte Carlo dropout methods in deep learning to capture the model's confidence in future carbon emissions forecasts. For example, when the model outputs each prediction result, it also provides a confidence interval for the prediction to help operators understand the reliability of the prediction results. This information is crucial for developing response strategies, especially when facing a changing environment or policy. Operators can develop risk management measures based on the degree of uncertainty in the prediction.

[0100] Once the model is trained, the current low-dimensional feature representation can be input into the trained dual-weighted multi-objective prediction model to generate prediction results at different time scales. In practical applications, microgrid managers can use these prediction results to formulate daily operation strategies, such as deploying renewable energy in advance during expected peak power demand, reducing carbon emissions, or optimizing energy storage and distribution during low-demand periods. At the same time, the information provided by dynamic uncertainty quantification can help managers adjust their operating strategies based on the reliability of the prediction results, so as to respond to various uncertain situations more flexibly and effectively.

[0101] In this way, microgrids can not only accurately grasp carbon emissions in daily operations, but also provide a scientific basis for policy making, taking a solid step towards low-carbon transformation, thereby achieving sustainable energy supply and environmental protection goals.

[0102] Specifically, based on the current low-dimensional feature representation, the trained dual-weighted multi-objective prediction model can be used to output multi-objective prediction results of carbon emissions at the microgrid energy supply end on different time scales.

[0103] Specifically, the current low-dimensional feature representation is the feature extracted and reduced in the previous steps, which fully retains the key information related to carbon emission changes. The dual-weighted multi-objective prediction model combines the time weight mechanism and the associated feature weight mechanism, which can dynamically adjust the impact on historical data and its characteristics, so as to accurately predict carbon emissions on multiple time scales. This process not only focuses on instantaneous carbon emission data, but also captures long-term trends and cyclical changes to provide more comprehensive prediction results. Through the multi-objective prediction results output by the model, microgrid operators can intuitively understand the carbon emission levels at different time points in the future, thereby providing strong support for optimizing energy management and decision-making.

[0104] The significance of this prediction method is to provide a scientific basis and decision-making support for the sustainable development of microgrids. Through accurate carbon emission prediction, operators can grasp the carbon emission trend in real time and adjust energy production and consumption strategies in time, so as to optimize resource allocation and reduce carbon emissions. In addition, the ability of multi-objective prediction enables operators to formulate flexible operation plans on different time scales, such as giving priority to low-carbon energy during peak load periods, or storing and dispatching energy during low demand periods. This accurate prediction will help microgrids transform to low-carbon, improve energy efficiency and reduce environmental impact. At the same time, through quantitative analysis of uncertainty, operators can better cope with possible changes and challenges in the future and manage the energy supply of microgrids in a more intelligent and sustainable way.

[0105] It can be seen that for the multi-dimensional original time series data in the operation process of the microgrid energy supply end, the feature subsequences that are strongly correlated with the changes in carbon emissions are extracted to generate an associated feature set; the associated feature set is input into the hybrid attention network to generate the current weighted feature representation; according to the historical weighted feature representation, the carbon emission simulation data enhancement technology of generative-discriminative collaborative learning is used to obtain a historical enhanced carbon emission data set with stable distribution; based on the historical enhanced carbon emission data set, a dual-weighted multi-objective prediction model is trained, combined with the combined gated recurrent unit and dynamic uncertainty quantification technology for prediction, and the multi-objective prediction results of carbon emissions at the microgrid energy supply end on different time scales are output, so that accurate prediction of carbon emissions can be achieved through efficient feature extraction, dynamic modeling and data enhancement technology, with high precision and flexibility.

[0106] Another embodiment of the present invention provides a microgrid energy supply end carbon emission prediction system based on time series data, see Figure 3 , the system may include:

[0107] The extraction module 301 is used to extract the characteristic subsequences strongly correlated with the carbon emission changes from the multi-dimensional original time series data during the operation of the microgrid energy supply end, so as to generate a correlation feature set;

[0108] A generation module 302 is used to input the associated feature set into a hybrid attention network, which combines the temporal self-attention mechanism and the channel interaction attention mechanism to capture the global temporal dependencies in the time series features and the potential interactions between key channels, and generate a current weighted feature representation to retain the information of the original features while enhancing the correlation with the change of carbon emissions;

[0109] Enhancement module 303, used to generate historical weighted feature representation corresponding to historical multi-dimensional original time series data, and obtain a historical enhanced carbon emission data set with stable distribution by using the carbon emission simulation data enhancement technology of generative-discriminative collaborative learning based on the historical weighted feature representation;

[0110] The prediction module 304 is used to train a dual-weighted multi-objective prediction model based on the historical enhanced carbon emission data set. The model dynamically adjusts the historical data and key features of the time series through the time weight mechanism and the associated feature weight mechanism, respectively, and combines the combined gated recurrent unit and dynamic uncertainty quantization technology for prediction. Based on the current weighted feature representation, the trained dual-weighted multi-objective prediction model is used to output the multi-objective prediction results of carbon emissions at the microgrid power supply end on different time scales.

[0111] It can be seen that for the multi-dimensional original time series data in the operation process of the microgrid energy supply end, the feature subsequences that are strongly correlated with the changes in carbon emissions are extracted to generate an associated feature set; the associated feature set is input into the hybrid attention network to generate the current weighted feature representation; according to the historical weighted feature representation, the carbon emission simulation data enhancement technology of generative-discriminative collaborative learning is used to obtain a historical enhanced carbon emission data set with stable distribution; based on the historical enhanced carbon emission data set, a dual-weighted multi-objective prediction model is trained, combined with the combined gated recurrent unit and dynamic uncertainty quantification technology for prediction, and the multi-objective prediction results of carbon emissions at the microgrid energy supply end on different time scales are output, so that accurate prediction of carbon emissions can be achieved through efficient feature extraction, dynamic modeling and data enhancement technology, with high precision and flexibility.

[0112] An embodiment of the present invention further provides a storage medium, in which a computer program is stored, wherein the computer program is configured to execute the steps of any of the above method embodiments when running.

[0113] Specifically, in this embodiment, the above storage medium may be configured to store a computer program for performing the following steps:

[0114] S201, extracting feature subsequences that are strongly correlated with carbon emission changes from the multi-dimensional original time series data during the operation of the microgrid energy supply end, to generate a correlation feature set;

[0115] S202, inputting the associated feature set into a hybrid attention network, which combines the temporal self-attention mechanism and the channel interaction attention mechanism to capture the global temporal dependency in the time series features and the potential interaction between key channels, and generates a current weighted feature representation to retain the information of the original features while enhancing the correlation with the change in carbon emissions;

[0116] S203, generating a historical weighted feature representation corresponding to the historical multi-dimensional original time series data, and using the generative-discriminative collaborative learning carbon emission simulation data enhancement technology to obtain a historical enhanced carbon emission data set with stable distribution according to the historical weighted feature representation;

[0117] S204, training a dual-weighted multi-objective prediction model based on the historical enhanced carbon emission data set. The model dynamically adjusts the historical data and key features of the time series through the time weight mechanism and the associated feature weight mechanism respectively, combines the combined gated recurrent unit and the dynamic uncertainty quantization technology for prediction, and based on the current weighted feature representation, uses the trained dual-weighted multi-objective prediction model to output the multi-objective prediction results of carbon emissions at the microgrid energy supply end on different time scales.

[0118] It can be seen that for the multi-dimensional original time series data in the operation process of the microgrid energy supply end, the feature subsequences that are strongly correlated with the changes in carbon emissions are extracted to generate an associated feature set; the associated feature set is input into the hybrid attention network to generate the current weighted feature representation; according to the historical weighted feature representation, the carbon emission simulation data enhancement technology of generative-discriminative collaborative learning is used to obtain a historical enhanced carbon emission data set with stable distribution; based on the historical enhanced carbon emission data set, a dual-weighted multi-objective prediction model is trained, combined with the combined gated recurrent unit and dynamic uncertainty quantification technology for prediction, and the multi-objective prediction results of carbon emissions at the microgrid energy supply end on different time scales are output, so that accurate prediction of carbon emissions can be achieved through efficient feature extraction, dynamic modeling and data enhancement technology, with high precision and flexibility.

[0119] An embodiment of the present invention further provides an electronic device, including a memory and a processor, wherein the memory stores a computer program, and the processor is configured to run the computer program to execute the steps in any one of the above method embodiments.

[0120] Specifically, the electronic device may further include a transmission device and an input / output device, wherein the transmission device is connected to the processor, and the input / output device is connected to the processor.

[0121] Specifically, in this embodiment, the processor may be configured to perform the following steps through a computer program:

[0122] S201, extracting feature subsequences that are strongly correlated with carbon emission changes from the multi-dimensional original time series data during the operation of the microgrid energy supply end, to generate a correlation feature set;

[0123] S202, inputting the associated feature set into a hybrid attention network, which combines the temporal self-attention mechanism and the channel interaction attention mechanism to capture the global temporal dependency in the time series features and the potential interaction between key channels, and generates a current weighted feature representation to retain the information of the original features while enhancing the correlation with the change in carbon emissions;

[0124] S203, generating a historical weighted feature representation corresponding to the historical multi-dimensional original time series data, and using the generative-discriminative collaborative learning carbon emission simulation data enhancement technology to obtain a historical enhanced carbon emission data set with stable distribution according to the historical weighted feature representation;

[0125] S204, training a dual-weighted multi-objective prediction model based on the historical enhanced carbon emission data set. The model dynamically adjusts the historical data and key features of the time series through the time weight mechanism and the associated feature weight mechanism respectively, combines the combined gated recurrent unit and the dynamic uncertainty quantization technology for prediction, and based on the current weighted feature representation, uses the trained dual-weighted multi-objective prediction model to output the multi-objective prediction results of carbon emissions at the microgrid energy supply end on different time scales.

[0126] It can be seen that for the multi-dimensional original time series data in the operation process of the microgrid energy supply end, the feature subsequences that are strongly correlated with the changes in carbon emissions are extracted to generate an associated feature set; the associated feature set is input into the hybrid attention network to generate the current weighted feature representation; according to the historical weighted feature representation, the carbon emission simulation data enhancement technology of generative-discriminative collaborative learning is used to obtain a historical enhanced carbon emission data set with stable distribution; based on the historical enhanced carbon emission data set, a dual-weighted multi-objective prediction model is trained, combined with the combined gated recurrent unit and dynamic uncertainty quantification technology for prediction, and the multi-objective prediction results of carbon emissions at the microgrid energy supply end on different time scales are output, so that accurate prediction of carbon emissions can be achieved through efficient feature extraction, dynamic modeling and data enhancement technology, with high precision and flexibility.

[0127] The above describes in detail the structure, features and effects of the present invention based on the embodiments shown in the drawings. The above is only a preferred embodiment of the present invention, but the present invention is not limited to the scope of implementation shown in the drawings. Any changes made according to the concept of the present invention, or modifications to equivalent embodiments with equivalent changes, which still do not exceed the spirit covered by the description and drawings, should be within the protection scope of the present invention.

Claims

1. A method for predicting carbon emissions at the energy supply end of a microgrid based on time series data, characterized in that: The method comprises: For the multi-dimensional original time series data during the operation of the microgrid energy supply end, extract the feature subsequence that is strongly correlated with the carbon emission change to generate a related feature set; for the multi-dimensional original time series data during the operation of the microgrid energy supply end, extract the feature subsequence that is strongly correlated with the carbon emission change to generate a related feature set, including: According to the multi-dimensional original time series data in the operation process of the microgrid energy supply end, the variational mode decomposition technology is used to decompose the multi-dimensional original time series into multiple intrinsic mode functions of different time scales; Based on the mutual information quantification technology, the correlation analysis of the decomposed intrinsic mode function components is performed to extract the characteristic subsequences strongly correlated with the carbon emission changes and generate the associated feature set; wherein, for each decomposed intrinsic mode function of different time scales, the mutual information value between the modal function and the actual carbon emission value of the corresponding time scale is calculated; the intrinsic mode function whose mutual information value is greater than the preset significance threshold is screened, and the time series part corresponding to the screened intrinsic mode function is extracted as the characteristic subsequence strongly correlated with the carbon emission changes, and all the extracted characteristic subsequences are integrated into the associated feature set; The associated feature set is input into a hybrid attention network, which combines the temporal self-attention mechanism and the channel interaction attention mechanism to capture the global temporal dependencies in the time series features and the potential interactions between key channels, and generates a current weighted feature representation to retain the information of the original features while enhancing the correlation with carbon emission changes; Generate historical weighted feature representations corresponding to historical multi-dimensional original time series data. Based on the historical weighted feature representations, use the generative-discriminative collaborative learning carbon emission simulation data enhancement technology to obtain a stable historical enhanced carbon emission data set. A double-weighted multi-objective prediction model is trained based on the historical enhanced carbon emission dataset. The model dynamically adjusts the historical data and key features of the time series through the time weight mechanism and the associated feature weight mechanism respectively, and combines the combined gated recurrent unit and dynamic uncertainty quantization technology for prediction. Based on the current weighted feature representation, the trained double-weighted multi-objective prediction model is used to output the multi-objective prediction results of carbon emissions at the microgrid energy supply end on different time scales.

2. The method according to claim 1, characterized in that: The method further includes: based on a differentiable sparse dimensionality reduction technique, performing dimensionality reduction processing on the weighted feature representation to obtain a low-dimensional feature representation with a compact feature structure and significant expressiveness; According to the historical weighted feature representation, the carbon emission simulation data enhancement technology of generative-discriminative collaborative learning is used to obtain a historical enhanced carbon emission data set with stable distribution, including: According to the historical weighted feature representation, the corresponding historical low-dimensional feature representation is generated. Based on the historical low-dimensional feature representation, the carbon emission simulation data enhancement technology of generative-discriminative collaborative learning is used to obtain a historical enhanced carbon emission dataset with stable distribution.

3. The method according to claim 2, characterized in that The carbon emission simulation data enhancement technology based on historical low-dimensional feature representation and generative-discriminative collaborative learning is used to obtain a historical enhanced carbon emission dataset with stable distribution, including: Input the historical low-dimensional feature representation into a generator, and generate simulated carbon emission data through a deep random nested residual generation technique; An adaptive distributed adversarial loss function is introduced through the discriminator to improve the distribution consistency between the simulated carbon emission data and the real observed carbon emission data. After repeated iterative optimization, an enhanced carbon emission data set with stable distribution is obtained to avoid the sparse and uneven distribution of carbon emission data at the microgrid energy supply end.

4. The method according to claim 2, characterized in that: The multi-objective prediction results of carbon emissions at the energy supply end of the microgrid on different time scales are outputted based on the current weighted feature representation using the trained double-weighted multi-objective prediction model, including: Based on the current low-dimensional feature representation, the trained dual-weighted multi-objective prediction model is used to output the multi-objective prediction results of carbon emissions at the microgrid energy supply end on different time scales.

5. A carbon emission prediction system for microgrid energy supply end based on time series data, characterized in that: The system comprises: The extraction module is used to extract the feature subsequences strongly correlated with the carbon emission changes from the multi-dimensional original time series data during the operation of the microgrid energy supply end to generate a related feature set; the multi-dimensional original time series data during the operation of the microgrid energy supply end is used to extract the feature subsequences strongly correlated with the carbon emission changes to generate a related feature set, including: According to the multi-dimensional original time series data in the operation process of the microgrid energy supply end, the variational mode decomposition technology is used to decompose the multi-dimensional original time series into multiple intrinsic mode functions of different time scales; Based on the mutual information quantification technology, the correlation analysis of the decomposed intrinsic mode function components is performed to extract the characteristic subsequences strongly correlated with the carbon emission changes and generate the associated feature set; wherein, for each decomposed intrinsic mode function of different time scales, the mutual information value between the modal function and the actual carbon emission value of the corresponding time scale is calculated; the intrinsic mode function whose mutual information value is greater than the preset significance threshold is screened, and the time series part corresponding to the screened intrinsic mode function is extracted as the characteristic subsequence strongly correlated with the carbon emission changes, and all the extracted characteristic subsequences are integrated into the associated feature set; A generation module, used for inputting the associated feature set into a hybrid attention network, which combines the temporal self-attention mechanism and the channel interaction attention mechanism to capture the global temporal dependencies in the time series features and the potential interactions between key channels, and generates a current weighted feature representation to retain the information of the original features while enhancing the correlation with the change of carbon emissions; The enhancement module is used to generate historical weighted feature representations corresponding to historical multi-dimensional original time series data. Based on the historical weighted feature representations, the carbon emission simulation data enhancement technology of generative-discriminative collaborative learning is used to obtain a historical enhanced carbon emission data set with stable distribution; The prediction module is used to train a dual-weighted multi-objective prediction model based on the historical enhanced carbon emission data set. The model dynamically adjusts the historical data and key features of the time series through the time weight mechanism and the associated feature weight mechanism, respectively, and combines the combined gated recurrent unit and dynamic uncertainty quantification technology for prediction. Based on the current weighted feature representation, the trained dual-weighted multi-objective prediction model is used to output the multi-objective prediction results of carbon emissions at the microgrid energy supply end on different time scales.

6. A storage medium, characterized in that: The storage medium stores a computer program, wherein the computer program is configured to execute the method according to any one of claims 1 to 4 when executed.

7. An electronic device comprising a memory and a processor, characterized in that: A computer program is stored in the memory, and the processor is configured to run the computer program to perform the method according to any one of claims 1 to 4.

Images