Low-carbon strategy generation method and system based on carbon emission factor prediction

Abstract

The invention provides a low-carbon strategy generation method based on carbon emission factor prediction. The method belongs to the technical field of power system low-carbon strategies. The method comprises the steps that a terminal layer collects power generation side output data, user side load data and environmental meteorological data of a power system; the cloud layer inputs power generation side output data, user side load data and environmental meteorological data into a dynamic carbon emission factor prediction model to generate a carbon emission factor prediction curve; the carbon emission factor prediction curve is input into a low-carbon regulation and control optimization model with the goal of carbon emission minimization, and a long-term low-carbon control strategy is obtained; and based on the peak valley characteristics of the carbon emission factor prediction curve, the long-term low-carbon control strategy and the user load characteristic spectrum, determining a low-carbon strategy meeting the power grid carbon emission reduction target and the user energy consumption demand. According to the method, the dynamic prediction of the carbon emission factor can be realized, the low-carbon energy consumption strategy is generated in combination with the user side demand, and the accuracy of carbon emission metering and the effectiveness of participation in emission reduction by the user are improved.

Description

Technical Field

[0001] This invention belongs to the field of low-carbon strategy technology for power systems, and in particular relates to a method and system for generating low-carbon strategies based on predicted carbon emission factors. Background Technology

[0002] The measurement and application of carbon emission factors mainly involve the energy, power, and environmental protection sectors. Under the strategic background of dual-carbon goals, the power industry is one of the key sources of carbon emissions; therefore, carbon emission monitoring and management on the electricity consumption side has become the core of research and application. The carbon emission factor is an important parameter for measuring the carbon emission level per unit of electricity. Traditionally, my country's power system has mostly adopted the average carbon emission factor method based on annual fuel consumption data from provincial power grids. Due to its ease of calculation, it has been applied in many regions, providing fundamental support for energy statistics and carbon accounting.

[0003] While the widely adopted method based on average carbon emission factors is simple to calculate, it also has shortcomings. Firstly, this method is rather crude in both time and space, failing to comprehensively reflect the differences in grid operation and carbon emissions across different regions and time periods. Secondly, the emission reduction measures available to users under this method are relatively limited, typically relying solely on reducing electricity consumption, lacking more diversified approaches and scientific guidance. This results in limited user motivation to participate in emission reduction and hinders effective coordination with the grid's low-carbon dispatching.

[0004] To address the aforementioned issues, several improvements have been proposed in existing technologies, exploring methods for calculating carbon emission factors and user-side participation mechanisms. These attempts have, to some extent, improved the applicability of carbon emission measurement and increased user participation in emission reduction.

[0005] However, overall, existing technologies still have limitations. For example, related methods often remain at a relatively macroscopic level, making it difficult to reflect the complexities of power grid operation in detail; user-side participation is also limited, relying more on passive responses and lacking an effective interactive mechanism. Furthermore, in the face of complex and volatile power system operating environments, existing solutions still fall short in terms of adaptability and stability. Therefore, existing technologies still fall short of actual needs and require further improvement. Thus, there is an urgent need for a technical solution that can dynamically predict carbon emission factors and generate low-carbon energy consumption strategies based on user-side needs, in order to improve the accuracy of carbon emission measurement and the effectiveness of user participation in emission reduction. Summary of the Invention

[0006] The purpose of this invention is to provide a method and system for generating low-carbon strategies based on predicted carbon emission factors, which can realize dynamic prediction of carbon emission factors and generate low-carbon energy use strategies in combination with user-side needs, thereby improving the accuracy of carbon emission measurement and the effectiveness of user participation in emission reduction.

[0007] To achieve the above objectives, the present invention adopts the following technical solution: In a first aspect, the present invention provides a method for generating low-carbon strategies based on predicted carbon emission factors, comprising: The terminal layer collects power output data from the power generation side, load data from the user side, and environmental meteorological data of the power system. The cloud layer inputs power generation output data, user load data, and environmental meteorological data into the dynamic carbon emission factor prediction model to generate carbon emission factor prediction curves. The cloud layer inputs the carbon emission factor prediction curve into a low-carbon regulation and optimization model aimed at minimizing carbon emissions to obtain a long-term low-carbon control strategy. Based on the peak and valley characteristics of the carbon emission factor prediction curve, long-term low-carbon control strategies, and user load characteristic maps, the cloud layer determines low-carbon strategies that meet the grid carbon emission reduction targets and user energy needs.

[0008] Furthermore, the low-carbon strategy generation method based on predicted carbon emission factors provided by this invention also includes: The terminal layer sends the power generation output data, user load data, and environmental meteorological data of the power system to the edge layer; The edge layer preprocesses the power generation output data, user load data, and environmental meteorological data of the power system and uploads them to the cloud layer. Preprocessing includes the following steps for filling in missing data: When complete historical data for the same period exists, it is filled in based on the statistical average of the same period over the past k days; When historical data for the same period is missing, the weighted average of the data before and after the missing point is used to fill the gap, with the weights set to decrease according to the time distance. Preprocessing also includes the following steps to unify the time resolution of data from different sampling frequencies: Raw data of different frequencies are aggregated to the minute level using a sliding window method; The unified data is then standardized and quantile normalized.

[0009] Furthermore, the low-carbon strategy generation method based on predicted carbon emission factors provided by this invention also includes: The edge layer inputs the preprocessed power generation output data, user load data and environmental meteorological data of the power system into the short-term carbon emission factor prediction model to generate the carbon emission factor prediction value for the first time limit in the future. The short-term carbon emission factor prediction model is a moving average or exponential smoothing model. When the predicted value of carbon emission factor in the first time limit exceeds the short-term threshold, a corresponding control instruction is sent to the control module. During communication anomalies between the edge layer and the cloud layer, the edge layer inputs preprocessed power generation output data, user load data, and environmental meteorological data into the emergency carbon emission factor prediction model to generate predicted carbon emission factor values ​​for the second time limit in the future. The emergency carbon emission factor prediction model is a simplified linear regression model. When the predicted carbon emission factor value for the second future time limit exceeds the emergency period threshold, a corresponding emergency period load reduction control command is sent to the control module.

[0010] Furthermore, the cloud layer includes a central cloud and an edge cloud. The dynamic carbon emission factor prediction model and the low-carbon regulation optimization model are deployed in the central cloud. The low-carbon strategy generation method based on predicted carbon emission factors provided by this invention also includes: The edge layer sends the pre-processed power generation output data, user load data, and environmental meteorological data of the power system to the edge cloud, and then the edge cloud sends them to the central cloud. The edge layer sends heartbeat packets to the edge cloud every set period. If the edge cloud does not receive heartbeat packets for n consecutive periods, it is determined that an abnormal communication between the edge layer and the cloud layer has been detected. After determining that an abnormality in communication between the edge layer and the cloud layer has been detected, if the edge cloud continuously receives heartbeat packets for m consecutive cycles, then it is determined that communication between the edge layer and the cloud layer has been restored. During communication anomalies between the edge layer and the cloud layer, the edge layer temporarily stores the pre-processed power generation output data, user load data, and environmental meteorological data. After communication is restored, the temporarily stored power generation output data, user load data, and environmental meteorological data are sent to the edge cloud, where the edge cloud performs data integrity verification before forwarding it to the central cloud.

[0011] Furthermore, the steps by which the edge layer sends the preprocessed power system generation-side output data, user-side load data, and environmental meteorological data to the edge cloud include: The edge layer utilizes a deployed hardware encryption engine to perform AES-128 block encryption on the pre-processed power generation output data, user load data, and environmental meteorological data, generating encrypted data and a checksum. Simultaneously, it uses an RSA public key to encrypt the session key used for AES-128 data encryption, obtaining the session key ciphertext. The session key ciphertext, encrypted data, and checksum are then packaged and sent to the edge cloud. The edge cloud uses the corresponding RSA private key to decrypt the session key ciphertext to obtain the session key, and decrypts the encrypted data based on the session key. Integrity verification is then performed based on the checksum. Low-carbon strategy generation methods based on predicted carbon emission factors also include: The terminal layer sets a collection timestamp T1 for the collected data, including power generation output data, user load data, and environmental meteorological data of the power system. After the edge layer preprocesses the collected data, it calibrates the collection timestamp T1. When the deviation between the collection timestamp T1 and the edge layer's local clock T0 is less than or equal to 100ms, the collection timestamp T1 is used as the synchronization timestamp T3. When the deviation is greater than 100ms, the synchronization timestamp T3 is determined according to the network reference time T2, where T3 = T1 + (T2 - T0). The collection timestamp T1 of the collected data is then updated to the synchronization timestamp T3. After receiving the collected data, the edge cloud calibrates the synchronization timestamp T3. When the deviation between the synchronization timestamp T3 and the edge cloud standard clock T4 is less than or equal to 500ms, the synchronization timestamp T3 is used as the final timestamp T5. When the deviation is greater than 500ms, the final timestamp T5 is determined according to the edge cloud standard clock T4, where T5 = T3 + (T4 - T3). The synchronization timestamp T3 is then updated to the final timestamp T5, and the collected data with the final timestamp T5 is sent to the central cloud.

[0012] Furthermore, the objective function of the low-carbon regulation optimization model is:

[0013]

[0014]

[0015] in, For total energy carbon emissions, Energy consumption As a dynamic carbon emission factor, This refers to the amount of material consumed. As a carbon emission factor of materials, In order to control the cost coefficient, Adjusting power costs for equipment To adjust the duration, For clean energy unit price, For green electricity replacement volume, These are the weighting coefficients; The constraints of the low-carbon regulation optimization model are as follows:

[0016]

[0017]

[0018] in, This is the minimum power requirement for the equipment. Rated power of the equipment For actual output, For the target output, For carbon emission allowances.

[0019] Furthermore, the dynamic carbon emission factor prediction model includes: a spatiotemporal feature embedding layer, a dynamic graph structure generator, and a multimodal gating fusion layer. The spatiotemporal feature embedding layer takes power generation output data, user load data, and environmental meteorological data as input. The spatiotemporal feature embedding layer extracts local periodic features of the data through a convolutional neural network and inputs the output of the convolutional network into a long short-term memory network to capture time series features, thereby forming power data features and meteorological data features of each power grid node respectively. The dynamic graph structure generator is composed of a dynamic graph attention network. Its input is the power data features output by the spatiotemporal feature embedding layer. It is used to dynamically update the edge weights between power grid nodes based on real-time transmission power, and calculate the correlation coefficient between nodes through the graph attention mechanism. It weights the importance of nodes in different time periods and spaces, and outputs the updated power data features. The multimodal gating fusion layer takes updated power data features and meteorological data features as input, and uses them to concatenate the updated power data features and meteorological data features into a hybrid feature. It dynamically adjusts the contribution of power mode and meteorological mode through sigmoid gating mechanism, and outputs the fused feature after linear transformation to generate the carbon emission factor prediction curve for the next 24 hours. The training steps for the dynamic carbon emission factor prediction model include: Historical power generation output data, user load data, environmental meteorological data, and corresponding historical carbon emission factor real values ​​are obtained as training samples. The training samples are divided into training set and validation set. The training set is used to learn the parameters of the dynamic carbon emission factor prediction model, and the validation set is used to evaluate the prediction performance of the dynamic carbon emission factor prediction model during the training process. The parameters of the dynamic carbon emission factor prediction model are iteratively optimized using the backpropagation algorithm until the loss function of the prediction result converges to within a preset threshold.

[0020] Furthermore, the training steps for the dynamic carbon emission factor prediction model also include: Perturbations are applied to the training samples to obtain adversarial examples; The training samples and adversarial samples are input into the dynamic carbon emission factor prediction model for training. The mean squared error is used as the loss function, and the parameters of the dynamic carbon emission factor prediction model are updated using an adaptive moment estimation optimization algorithm.

[0021] Furthermore, the low-carbon strategy generation method based on predicted carbon emission factors provided by this invention also includes: The predicted average value of carbon emission factors is calculated based on the carbon emission factor prediction curve. The continuous time period that is 30% higher than the predicted average value is marked as the peak period, and the continuous time period that is 20% lower than the predicted average value is marked as the valley period. The steps to determine a low-carbon strategy include: Based on long-term low-carbon control strategies and user load characteristic maps, during peak periods, flexible loads will be subject to one or more operations, including reducing electric vehicle charging power, increasing air conditioning temperature, or reducing lighting brightness. During off-peak periods, flexible loads will be subject to one or more operations, including pre-cooling or preheating, restoring electric vehicle charging power, or restoring lighting operation.

[0022] In a second aspect, the present invention provides a low-carbon strategy generation system based on predicted carbon emission factors, comprising: a terminal layer and a cloud layer, wherein the terminal layer collects power generation output data, user load data and environmental meteorological data of the power system. The cloud layer inputs power generation output data, user load data, and environmental meteorological data into the dynamic carbon emission factor prediction model to generate carbon emission factor prediction curves. The cloud layer inputs the carbon emission factor prediction curve into a low-carbon regulation and optimization model aimed at minimizing carbon emissions to obtain a long-term low-carbon control strategy. Based on the peak and valley characteristics of the carbon emission factor prediction curve, long-term low-carbon control strategies, and user load characteristic maps, the cloud layer determines low-carbon strategies that meet the grid carbon emission reduction targets and user energy needs.

[0023] Compared with the prior art, the beneficial effects of the present invention are as follows: 1. The low-carbon strategy generation method proposed in this invention collects multi-source data, including power generation output, user load, and meteorological data, and performs missing data filling and unified temporal resolution processing to ensure data integrity and consistency. Based on this, a spatiotemporal dual-stream fusion structure is introduced, utilizing a long short-term memory network to extract time-series features, combining this with a convolutional neural network to obtain spatial distribution features, and incorporating a dynamic graph attention mechanism at the output to dynamically adjust the correlation between grid nodes. This design enables carbon emission factor prediction to take into account both temporal and spatial variations, significantly improving prediction accuracy and robustness. Compared to the traditional average factor method, this method can realistically reflect the carbon emission characteristics under different operating conditions, providing reliable support for the low-carbon operation of the power grid, and laying a solid foundation for user-side strategy generation.

[0024] 2. In the model training phase, this invention uses historical power generation data, load data, and meteorological data as training samples, and achieves parameter learning and performance evaluation by dividing the training set and validation set. Iterative optimization using the backpropagation algorithm ensures that the loss function of the prediction results converges within a reasonable range. Simultaneously, mean squared error is introduced as a loss function, and the Adam optimization algorithm is used for parameter updates to improve convergence speed and stability. Furthermore, adversarial examples are generated by applying perturbations to the training samples to enhance the model's robustness to extreme operating conditions and abnormal data. This design ensures that the carbon emission factor prediction model maintains high accuracy and reliability even under complex scenarios such as new energy fluctuations and load abrupt changes, thus providing robust data support for low-carbon strategy generation.

[0025] 3. In the strategy generation stage, this invention proposes differentiated control schemes for flexible loads based on the peak-valley characteristics of the carbon emission factor prediction curve and user load characteristic maps. During peak periods, the charging power of electric vehicles can be reduced, the air conditioning temperature can be increased, or the lighting brightness can be decreased. During off-peak periods, pre-cooling / pre-heating or restoration of flexible load operation can be implemented, achieving coordinated carbon reduction on the user side and the grid side. This scheme enhances user participation and energy flexibility, enabling low-carbon strategies to meet both grid emission reduction targets and user needs. The research results not only have promotional value for industrial parks and enterprise users, but also provide a reference for formulating carbon emission metering, management, and emission reduction policies, promoting the coupled development of the electricity market and the carbon market, and helping to achieve the strategic goals of carbon peaking and carbon neutrality. Attached Figure Description

[0026] The accompanying drawings, which form part of this application, are used to provide a further understanding of the invention. The illustrative embodiments of the invention and their descriptions are used to explain the invention and do not constitute an undue limitation of the invention. In the drawings: Figure 1 This is a flowchart of a low-carbon strategy generation method based on predicted carbon emission factors, as described in an embodiment of the present invention. Figure 2 This is a flowchart illustrating the implementation of the low-carbon strategy generation method based on predicted carbon emission factors in an embodiment of the present invention. Figure 3 This is a flowchart illustrating the low-carbon interaction strategy generation in an embodiment of the present invention. Figure 4 This is a schematic diagram of a low-carbon strategy generation system based on predicted carbon emission factors, according to an embodiment of the present invention. Detailed Implementation

[0027] The present invention will now be described in detail with reference to the accompanying drawings and embodiments. It should be noted that, unless otherwise specified, the embodiments and features described in this application can be combined with each other.

[0028] The following detailed description is exemplary and intended to provide further detailed explanation of the invention. Unless otherwise specified, all technical terms used in this invention have the same meaning as commonly understood by one of ordinary skill in the art to which this application pertains. The terminology used in this invention is for the purpose of describing particular embodiments only and is not intended to limit the scope of exemplary embodiments according to the invention.

[0029] Example 1 This invention provides a method for generating low-carbon strategies based on predicted carbon emission factors. This method employs a three-tier architecture: terminal layer, edge layer, and cloud layer. Figure 1 and Figure 2 As shown, the method includes the following steps: Step S1: The terminal layer collects power generation output data, user load data, and environmental meteorological data of the power system.

[0030] Step S2: The cloud layer inputs the power generation output data, user load data, and environmental meteorological data into the dynamic carbon emission factor prediction model to generate a carbon emission factor prediction curve.

[0031] Step S3: The cloud layer inputs the carbon emission factor prediction curve into the low-carbon regulation and optimization model aimed at minimizing carbon emissions to obtain a long-term low-carbon control strategy.

[0032] Step S4: Based on the peak and valley characteristics of the carbon emission factor prediction curve, long-term low-carbon control strategies, and user load characteristic maps, the cloud layer determines low-carbon strategies that meet the grid carbon emission reduction targets and user energy needs.

[0033] In step S1, the terminal layer collects data curves from different sources: power generation output curves (including different types of power sources such as thermal power, wind power, and photovoltaic power), user-side load curves (time-of-use load curves and adjustable load curves), and environmental meteorological data (weather forecast data obtained from external APIs, including information such as rainfall and temperature).

[0034] The terminal layer sends power generation output data, user load data, and environmental meteorological data from the power system to the edge layer. The edge layer preprocesses the power generation output data, user load data, and environmental meteorological data and then uploads them to the cloud layer. Preprocessing includes the following steps to fill in missing data: When complete historical data for the same period exists, it is filled in based on the statistical average of the same period over the past k days.

[0035] Specifically, in multi-source data acquisition for power systems, factors such as communication interruptions and equipment failures often lead to partial data loss (e.g., the output data of a thermal power plant from 00:00 to 02:00 on May 10, 2023 was lost). To achieve high-precision training of the carbon factor prediction model, adaptive imputation of missing data is required. The specific method is as follows: When historical data for the same period is missing, the weighted average of the data before and after the missing point is used to fill the gap, with the weights set in descending order of time distance.

[0036] Extract complete data from the historical same period (e.g., the same time period in the past 7 days) corresponding to the missing time period, and calculate their arithmetic mean as the imputation value. For a missing time period t, select complete data from the historical same period (the same time period in the past k days) and calculate its mean, as follows:

[0037] in The length of the cycle (e.g., daily cycle) =1 day).

[0038] For example, if a power plant's power generation data is missing at 12:00 on May 5, 2024, data from the same period over the past four days will be used to fill the gap. The calculation method is as follows:

[0039] When historical data from the same period is invalid, select local data from N hours before and after the missing period, and calculate the weighted mean based on time decay weights:

[0040] in, For a moment The weights, The attenuation coefficient is set to 0.5.

[0041] For example, if a power plant's power generation data for May 5, 2024, at 12:00 is missing, and historical data for the same period is also lost, then data from the previous three hours will be used to fill the gaps. The calculation method is as follows:

[0042] When historical data for the same period is missing, the weighted average of the data before and after the missing point is used to fill the gap, with the weights set in descending order of time distance.

[0043] Preprocessing also includes the following steps to unify the time resolution of data from different sampling frequencies: Raw data at different frequencies were unified to the minute level using a sliding window aggregation method. The unified data was then standardized and normalized to quantiles. After imputing missing values ​​to obtain complete data, standardization and quantile normalization were performed to improve the solution speed and generalization ability of subsequent deep learning models. Simultaneously, data at different time resolutions (hourly, minute-level, and second-level) were uniformly adjusted to the minute level using a sliding window aggregation method.

[0044] In this embodiment of the invention, the edge layer provides a short-term forecasting function. The edge layer inputs the preprocessed power generation output data, user load data and environmental meteorological data of the power system into the short-term carbon emission factor prediction model to generate the carbon emission factor prediction value within the first time limit in the future (e.g., the next 1-4 hours). The short-term carbon emission factor prediction model is a moving average or exponential smoothing model.

[0045] When the predicted carbon emission factor for the first time period exceeds the short-term threshold, a corresponding control command is sent to the control module. For example, when the short-term predicted value is greater than 1.05 times the hourly carbon allowance, a control command is automatically sent to the control module.

[0046] In this embodiment of the invention, the cloud layer includes a central cloud and an edge cloud. The dynamic carbon emission factor prediction model and the low-carbon regulation optimization model are deployed in the central cloud. The edge layer sends the preprocessed power generation output data, user load data and environmental meteorological data of the power system to the edge cloud, and then the edge cloud sends them to the central cloud.

[0047] The edge layer has a network outage self-governance mechanism. During communication anomalies between the edge layer and the cloud layer, the edge layer inputs preprocessed power generation output data, user load data, and environmental meteorological data into the emergency carbon emission factor prediction model to generate carbon emission factor prediction values ​​for the next time limit (e.g., the next hour). The emergency carbon emission factor prediction model is a simplified linear regression model deployed in the local SRAM algorithm area.

[0048] In this embodiment of the invention, peak and valley identification can be performed on the ultra-short-term carbon emission factor prediction value for the next hour output by the prediction model. Based on the peak and valley identification results, when the predicted carbon emission factor value for the second time limit in the future exceeds the emergency period threshold, a corresponding emergency period load reduction control command is sent to the control module. For example, when the predicted value is greater than the carbon quota hourly value × 110%, a load reduction command is automatically sent (such as reducing motor power by 15% or turning off non-essential lighting).

[0049] Communication anomaly detection: The edge layer sends heartbeat packets to the edge cloud at set intervals. If the edge cloud does not receive heartbeat packets for n consecutive intervals, a communication anomaly between the edge layer and the cloud layer is detected. For example, if a heartbeat packet is sent to the edge cloud every 30 seconds and there is no response for 3 consecutive times (timeout 10 seconds / timeout), a network outage is determined, the Flash network outage flag is triggered, and the system switches to local autonomous mode.

[0050] After detecting an anomaly in communication between the edge layer and the cloud layer, if the edge cloud continuously receives heartbeat packets for m consecutive cycles, then it is determined that communication between the edge layer and the cloud layer has been restored. n and m are preset positive integers.

[0051] During communication anomalies between the edge layer and the cloud layer, the edge layer temporarily stores pre-processed power generation output data, user load data, and environmental meteorological data. After communication is restored, it switches back to the cloud-edge collaborative mode. The temporarily stored power generation output data, user load data, and environmental meteorological data are sent to the edge cloud, where data integrity is verified before being forwarded to the central cloud. The central cloud then updates its local model and carbon factor table.

[0052] In the cloud-edge collaborative network architecture, the terminal layer sets a collection timestamp T1 for the collected data, including the power generation output data, user load data, and environmental meteorological data of the power system. For example, the collection timestamp T1 is a local timestamp with a precision of 1ms.

[0053] After preprocessing the collected data, the edge layer calibrates the collection timestamp T1. When the deviation between the collection timestamp T1 and the edge layer's local clock T0 is less than or equal to 100ms, the collection timestamp T1 is used as the synchronization timestamp T3. When the deviation is greater than 100ms, the synchronization timestamp T3 is determined based on the network reference time T2 (i.e., corrected based on the NTP protocol), where T3 = T1 + (T2 - T0), and the collection timestamp T1 of the collected data is updated to the synchronization timestamp T3.

[0054] After receiving the collected data, the edge cloud performs a secondary calibration on the synchronization timestamp T3. When the deviation between the synchronization timestamp T3 and the edge cloud standard clock T4 is less than or equal to 500ms, the synchronization timestamp T3 is used as the final timestamp T5. When the deviation is greater than 500ms, the final timestamp T5 is determined according to the edge cloud standard clock T4, where T5 = T3 + (T4 - T3). The synchronization timestamp T3 is then updated to the final timestamp T5, and the collected data with the final timestamp T5 is sent to the central cloud.

[0055] The edge layer utilizes a deployed hardware encryption engine to perform AES-128 block encryption on the pre-processed power generation output data, user load data, and environmental meteorological data, generating encrypted data and a checksum. Simultaneously, it uses an RSA public key to encrypt the session key used for AES-128 data encryption, obtaining the session key ciphertext. The session key ciphertext, encrypted data, and checksum are then packaged and sent to the edge cloud. The edge cloud uses the corresponding RSA private key to decrypt the session key ciphertext to obtain the session key, and decrypts the encrypted data based on the session key. Integrity verification is then performed based on the checksum.

[0056] In step S2, a dynamic carbon emission factor prediction model is deployed in the cloud layer. This model is based on a deep learning dynamic graph neural network model and is used to dynamically predict carbon emission factors. The model adopts a spatiotemporal dual-stream-dynamic fusion architecture and includes the following core modules: a spatiotemporal feature embedding layer, a dynamic graph structure generator, and a multimodal gating fusion layer. The spatiotemporal feature embedding layer takes power generation output data, user load data, and environmental meteorological data as input. The spatiotemporal feature embedding layer extracts local periodic features of the data through a convolutional neural network and inputs the output of the convolutional network into a long short-term memory network to capture time series features, thereby forming the power data features and meteorological data features of each grid node.

[0057] The dynamic graph structure generator is composed of a dynamic graph attention network. Its input is the power data features output by the spatiotemporal feature embedding layer. It is used to dynamically update the edge weights between power grid nodes based on real-time transmission power, and calculate the correlation coefficient between nodes through the graph attention mechanism. It weights the importance of nodes in different time periods and spaces, and outputs the updated power data features.

[0058] The multimodal gating fusion layer takes updated power data features and meteorological data features as input, and uses them to concatenate the updated power data features and meteorological data features into a hybrid feature. It dynamically adjusts the contribution of power mode and meteorological mode through a sigmoid gating mechanism, and outputs the fused feature after linear transformation to generate the carbon emission factor prediction curve for the next 24 hours.

[0059] This invention provides the following steps for constructing a dynamic carbon emission factor prediction model: A spatiotemporal dual-stream fusion structure is established, where the temporal stream employs a Long Short-Term Memory (LSTM) network to extract time-series features, and the spatial stream employs a Convolutional Neural Network (CNN) to extract spatial features. The spatiotemporal feature embedding layer, primarily used to extract local features from the input data, consists of a CNN+LSTM hybrid encoder. Its input is preprocessed and spatiotemporally aligned data. First, a 3x3 binary convolutional layer slides to extract local periodic features, outputting 64 channels. Then, the output of the convolutional layer is input into an LSTM to capture global features over longer time periods. Finally, these features are embedded into a unified feature space.

[0060] A dynamic graph attention mechanism is introduced at the output of the spatiotemporal dual-stream fusion structure. This mechanism dynamically adjusts the edge weights between nodes based on real-time grid power flow and updates the correlations of each grid node through graph attention calculation, thereby weighting the importance of nodes across different time periods and locations. A dynamic graph structure generator, composed of a dynamic graph attention network (GAT), updates the correlation weights between grid nodes in real time. Its input consists of features extracted from the spatiotemporal feature embedding layer. The dynamic graph is constructed as follows: nodes consist of grid regions or key nodes (e.g., thermal power plants, substations, industrial users). Edge weights are dynamically updated based on real-time transmission power, and the correlation strength between node i and node j is defined as:

[0061] in, Let be the power transmitted from node i to node j.

[0062] The attention coefficient of node i to its neighbor node j is calculated as follows:

[0063] in, , For node features, W and a are both learnable parameters with initial values ​​of random values.

[0064] A multimodal gating fusion layer is used to adaptively fuse data features from various modalities (e.g., power data, environmental meteorological data) obtained through a dynamic graph attention network. First, the power data feature He and the environmental meteorological data feature Hw are concatenated into a hybrid feature. The contribution of each mode is dynamically adjusted through a sigmoid gating mechanism, as follows: Gated vector generation:

[0065] in For the sigmoid function, and All of these are learnable parameters, with initial values ​​being random.

[0066] Feature dynamic fusion:

[0067] in This represents element-wise multiplication (Hadamard product).

[0068] Finally, the obtained fusion features The predicted carbon factor curve for the next 24 hours is obtained through linear transformation.

[0069] In terms of training strategy, the MSE loss function is adopted, the Adam optimizer is used, and adversarial training is employed to train the model. Historical power generation output data, user load data, and environmental meteorological data, along with corresponding historical real values ​​of carbon emission factors, are acquired as training samples. These samples are divided into training and validation sets. The training set is used to learn the parameters of the dynamic carbon emission factor prediction model, while the validation set is used to evaluate the predictive performance of the model during training. The backpropagation algorithm is used to iteratively optimize the parameters of the dynamic carbon emission factor prediction model until the loss function of the prediction result converges to within a preset threshold.

[0070] The specific loss function is as follows:

[0071] in Let be the actual data of the i-th node at time t. Let be the predicted data for the i-th node at time t.

[0072] Adam is a commonly used optimizer in deep learning, capable of dynamically adjusting the learning rate to improve model robustness and fast convergence. In this invention, the initial learning rate is set to 0.001.

[0073] Adversarial training involves adding perturbations to the training data so that the model can still maintain the stability of predictions under some extreme scenarios (such as sudden changes in new energy access or load changes), thereby improving the robustness of carbon factor prediction.

[0074] Specifically, perturbations are applied to the training samples to obtain adversarial examples. These training and adversarial examples are then input into the dynamic carbon emission factor prediction model for training. Mean squared error is used as the loss function, and an adaptive moment estimation optimization algorithm is employed to update the parameters of the dynamic carbon emission factor prediction model, thereby improving its stability and generalization ability under conditions of fluctuations in new energy output or sudden changes in user load.

[0075] In step S3, the central cloud deploys a low-carbon regulation and optimization model, whose objective function is:

[0076]

[0077]

[0078] in, For total energy carbon emissions, Energy consumption As a dynamic carbon emission factor, This refers to the amount of material consumed. As a carbon emission factor of materials, In order to control the cost coefficient, Adjusting power costs for equipment To adjust the duration, For clean energy unit price, For green electricity replacement volume, These are the weighting coefficients.

[0079] The constraints of the low-carbon regulation optimization model are: Equipment operating constraints:

[0080] Production schedule constraints:

[0081] Carbon quota constraints:

[0082] in, This is the minimum power requirement for the equipment. Rated power of the equipment For actual output, For the target output, For carbon emission allowances.

[0083] After the central cloud integrates industry and policy data with the carbon emission factor prediction curve and cloud data, it can generate long-term low-carbon control strategies (such as green electricity procurement ratio and equipment transformation suggestions).

[0084] In step S4, as Figure 3 As shown, the cloud layer determines low-carbon strategies that meet both grid carbon emission reduction targets and user energy needs based on the peak-valley characteristics of the carbon emission factor prediction curve, long-term low-carbon control strategies, and user load characteristic maps. First, the carbon emission factor prediction curve identifies peak and valley periods, obtaining peak-valley characteristics. The user load characteristic map then provides the adjustment capacity of flexible loads. Finally, the predicted average value of the carbon emission factor is calculated based on the carbon emission factor prediction curve. Continuous periods exceeding the predicted average by 30% are marked as peak periods, and continuous periods below the predicted average by 20% are marked as valley periods.

[0085] User loads can be divided into two categories: flexible loads, such as air conditioners, lighting, charging stations, heating, and cooling devices, which can be flexibly adjusted; and rigid loads, such as production lines or energy-consuming equipment that cannot be flexibly adjusted. For equipment that can be flexibly controlled, flexible load adjustment should be implemented according to long-term low-carbon control strategies and peak-valley characteristics, generally ensuring that the load increases during off-peak periods and decreases during peak periods. For equipment that can only be rigidly controlled, rigid control should be implemented without affecting its use.

[0086] The low-carbon interactive model's strategy is as follows: For flexible loads, air conditioning temperatures are uniformly increased by 2°C during peak periods. Charging pile power is uniformly reduced by 20%. Lighting brightness in non-production areas is uniformly reduced by 30%. During off-peak periods, air conditioning can preheat / precool, charging piles operate at full power, and lighting brightness is 100%. For rigid loads, unnecessary equipment operation is suspended during peak periods, and it is recommended to install green energy equipment such as photovoltaics to achieve green energy self-sufficiency. After the low-carbon strategy is output, it is implemented after manual confirmation and review. The strategy is then sent to the control module for implementation, and the control module sends the execution status and measurement point feedback back to the cloud for closed-loop updates, thereby optimizing model parameters. This achieves the dual satisfaction of carbon emission peak shaving and valley filling and user energy demand.

[0087] The low-carbon strategy generation method based on predicted carbon emission factors provided by this invention first performs feature cleaning and spatiotemporal alignment preprocessing on multi-dimensional historical power system operation data. Then, a deep learning-based dynamic carbon factor prediction model is constructed. After the model converges through an adaptive training mechanism, dynamic prediction of the power grid carbon emission intensity curve is achieved. Based on this, a time-domain decomposition algorithm is used to extract peak and valley characteristic parameters of carbon emission time series, and user-side load characteristic maps are simultaneously fused to construct a low-carbon interactive optimization model with multi-objective constraints. Finally, an optimal low-carbon response strategy that balances power grid carbon emission absorption and user energy demand is generated, providing scientific decision support for demand-side resources to participate in the carbon market. This solution innovatively achieves synergistic optimization of carbon emission trend prediction and user behavior characteristics, effectively improving the accuracy and timeliness of low-carbon regulation of the power system.

[0088] Example 2 like Figure 4 As shown, based on the same inventive concept as the above embodiments, the present invention also provides a low-carbon strategy generation system based on predicted carbon emission factors. This system realizes full-process management of accurate carbon emission measurement, dynamic prediction, intelligent control, and autonomous operation during grid outages. The system includes a terminal layer, an edge layer, and a cloud layer.

[0089] The terminal layer collects power output data from the power generation side, load data from the user side, and environmental meteorological data. Module installation: Located near carbon emission sources (such as factory distribution boxes or boiler rooms), connecting to energy consumption sensors, CO2 concentration sensors, and load sensors.

[0090] The edge layer preprocesses power generation output data, user load data, and environmental meteorological data from the power system and uploads them to the cloud layer. The edge layer provides short-term forecasting capabilities. It inputs the preprocessed power generation output data, user load data, and environmental meteorological data into a short-term carbon emission factor prediction model to generate a predicted carbon emission factor value for the first time period (e.g., 1-4 hours). The short-term carbon emission factor prediction model is either a moving average or an exponential smoothing model. When the predicted carbon emission factor value for the first time period exceeds a short-term threshold, a corresponding control command is sent to the control module. For example, when the short-term predicted value is greater than 1.05 times the hourly carbon allowance, a control command is automatically sent to the control module.

[0091] The edge layer has a network outage self-governance mechanism. During communication anomalies between the edge layer and the cloud layer, the edge layer inputs pre-processed power generation output data, user load data, and environmental meteorological data into the emergency carbon emission factor prediction model to generate a predicted carbon emission factor value for the next second time limit (e.g., the next hour). The emergency carbon emission factor prediction model is a simplified linear regression model deployed in the local SRAM algorithm area. When the predicted carbon emission factor value for the next second time limit exceeds the emergency period threshold, a corresponding emergency period load reduction control command is sent to the control module. For example, when the predicted value is greater than 110% of the carbon quota hourly value, a load reduction command (such as reducing motor power by 15% or turning off non-essential lighting) is automatically sent to the control module.

[0092] The cloud layer is divided into a central cloud and an edge cloud. The cloud layer inputs power generation output data, user load data, and environmental meteorological data into a dynamic carbon emission factor prediction model to generate a carbon emission factor prediction curve. The cloud layer then inputs this carbon emission factor prediction curve into a low-carbon regulation optimization model aimed at minimizing carbon emissions to obtain a long-term low-carbon control strategy. Based on the peak-valley characteristics of the carbon emission factor prediction curve, the long-term low-carbon control strategy, and user load characteristic maps, the cloud layer determines a low-carbon strategy that meets both the grid's carbon emission reduction targets and users' energy needs.

[0093] The terminal layer sets a collection timestamp T1 for the collected data, including the power generation output data, user load data, and environmental meteorological data of the power system. For example, the collection timestamp T1 is a local timestamp with a precision of 1ms.

[0094] After preprocessing the collected data, the edge layer calibrates the collection timestamp T1. When the deviation between the collection timestamp T1 and the edge layer's local clock T0 is less than or equal to 100ms, the collection timestamp T1 is used as the synchronization timestamp T3. When the deviation is greater than 100ms, the synchronization timestamp T3 is determined based on the network reference time T2 (i.e., corrected based on the NTP protocol), where T3 = T1 + (T2 - T0), and the collection timestamp T1 of the collected data is updated to the synchronization timestamp T3.

[0095] After receiving the collected data, the edge cloud performs a secondary calibration on the synchronization timestamp T3. When the deviation between the synchronization timestamp T3 and the edge cloud standard clock T4 is less than or equal to 500ms, the synchronization timestamp T3 is used as the final timestamp T5. When the deviation is greater than 500ms, the final timestamp T5 is determined according to the edge cloud standard clock T4, where T5 = T3 + (T4 - T3). The synchronization timestamp T3 is then updated to the final timestamp T5, and the collected data with the final timestamp T5 is sent to the central cloud.

[0096] The edge layer utilizes a deployed hardware encryption engine to perform AES-128 block encryption on the pre-processed power generation output data, user load data, and environmental meteorological data, generating encrypted data and a checksum. Simultaneously, it uses an RSA public key to encrypt the session key used for AES-128 data encryption, obtaining the session key ciphertext. The session key ciphertext, encrypted data, and checksum are then packaged and sent to the edge cloud. The edge cloud uses the corresponding RSA private key to decrypt the session key ciphertext to obtain the session key, and decrypts the encrypted data based on the session key. Integrity verification is then performed based on the checksum.

[0097] The terminal layer, based on basic monitoring equipment, directly monitors direct carbon emissions and environmental meteorological parameters on the power generation side and the user side (distributed energy). The carbon metering module acquires sensor data through the downlink communication module, and then uses built-in data processing and edge short-term prediction algorithms to achieve short-term node prediction. In the event of a network outage, a local low-carbon control strategy is generated based on the short-term prediction stage, and the equipment is controlled for low-carbon operation through the downlink communication interface. Without network interruption, the local short-term prediction results and real-time data are uploaded to the edge cloud layer through the uplink interface. The edge cloud layer calibrates and temporarily stores the data, and uploads it to the cloud layer after all node data is complete. The cloud layer uses a built-in dynamic carbon emission factor prediction algorithm to perform long-term carbon emission prediction for all nodes in the power grid based on the uploaded data, calculation standards, and carbon factor library data. Based on the long-term prediction data, a low-carbon control strategy is generated and distributed to the low-carbon module, which then controls flexible load equipment according to the strategy.

[0098] In the specific embodiments provided in this invention, the terminal layer consists of multiple types of sensors, including power generation side (thermal power / wind power / photovoltaic output sensors), user side (time-of-use load, adjustable load sensors), and environmental side (temperature, rainfall sensors).

[0099] The edge layer (with the carbon metering module at its core) adaptively fills in missing data and performs timestamp calibration (data with a deviation of ≤100ms between T1 and the module clock T0 is directly processed; otherwise, it is corrected based on the NTP protocol). The edge layer performs short-term forecasts, running a quantified short-term carbon emission forecast model (moving average / exponential smoothing) to predict carbon emissions for the next 1-4 hours. When the short-term forecast value is greater than 1.05 times the hourly carbon allowance, it automatically sends a control command to the control module. The edge layer uploads the processed data (with calibration timestamp T3) to the edge cloud via an encrypted pipeline (AES+RSA).

[0100] In specific implementations, the edge layer (carbon metering module) is deployed close to carbon emission sources (such as factory power distribution boxes and boiler rooms), connecting to energy consumption sensors, CO2 concentration sensors, and load sensors. For communication configuration, it supports SPI (communication with the IoT meter itself to obtain load data), 485 / RJ45 (controlling new energy equipment / energy storage), and 4G / 5G / Ethernet (cloud-edge data transmission), adapting to different network environments. If Ethernet is available in the application scenario, an RJ45 connection is used; in remote areas, a 4G / 5G module is added. The SPI interface connects to the IoT meter itself, and the 485 interface connects to new energy equipment / energy storage.

[0101] The edge layer carbon metering module is initialized with the following configurations: sensor acquisition frequency (energy consumption 1 time / second, CO2 1 time / 10 seconds), encryption key (AES-128 key preloaded into the SRAM secure area), and emergency threshold for network outage (carbon quota hourly value × 110%).

[0102] The MCU of the edge layer carbon metering module uses the HC32F460JETA / LQFP48 chip (ARM-M4 core, 1.8V-3.6V, maximum 200MHz main frequency, 512KBFlash+192KBSRAM) as its core, integrating data acquisition, encryption, communication and control modules.

[0103] The chip resource allocation scheme is as follows:

[0104] After the edge layer carbon metering module acquires data, the CPU sends an encryption request to the AES-128 accelerator. The accelerator reads the key from the SRAM secure area and encrypts it in blocks (ECB mode, 16 bytes / block), generating a checksum. At the same time, the RSA module encrypts the AES key. The final data packet format is [key ciphertext (256 bytes) + encrypted data (N bytes) + checksum (4 bytes)]. The CPU only needs 10μs to trigger the instruction, with an idle rate of ≥85%.

[0105] In the description of this specification, references to terms such as "an embodiment," "example," "specific example," etc., indicate that a specific feature, structure, material, or characteristic described in connection with that embodiment or example is included in at least one embodiment or example of the invention. In this specification, illustrative expressions of the above terms do not necessarily refer to the same embodiment or example. Furthermore, the specific features, structures, materials, or characteristics described may be combined in any suitable manner in one or more embodiments or examples.

[0106] Finally, it should be noted that the above embodiments are only used to illustrate the technical solutions of the present invention and not to limit it. Although the present invention has been described in detail with reference to the above embodiments, those skilled in the art should understand that modifications or equivalent substitutions can still be made to the specific implementation of the present invention. Any modifications or equivalent substitutions that do not depart from the spirit and scope of the present invention should be covered within the scope of protection of the claims of the present invention.

Claims

1. A method for generating low-carbon strategies based on predicted carbon emission factors, characterized in that, include: The terminal layer collects power output data from the power generation side, load data from the user side, and environmental meteorological data of the power system. The cloud layer inputs the power generation output data, user load data and environmental meteorological data into the dynamic carbon emission factor prediction model to generate a carbon emission factor prediction curve. The cloud layer inputs the carbon emission factor prediction curve into a low-carbon regulation and optimization model aimed at minimizing carbon emissions to obtain a long-term low-carbon control strategy. Based on the peak and valley characteristics of the carbon emission factor prediction curve, long-term low-carbon control strategies, and user load characteristic maps, the cloud layer determines low-carbon strategies that meet the grid carbon emission reduction targets and user energy needs.

2. The method for generating low-carbon strategies based on predicted carbon emission factors according to claim 1, characterized in that, Also includes: The terminal layer sends the power generation output data, user load data, and environmental meteorological data of the power system to the edge layer; The edge layer preprocesses the power generation output data, user load data, and environmental meteorological data of the power system and uploads them to the cloud layer. The preprocessing includes the following steps for filling in missing data: When complete historical data for the same period exists, it is filled in based on the statistical average of the same period over the past k days; When historical data for the same period is missing, the weighted average of the data before and after the missing point is used to fill the gap, with the weights set to decrease according to the time distance. The preprocessing also includes the following step of unifying the time resolution of data with different sampling frequencies: Raw data of different frequencies are aggregated to the minute level using a sliding window method; The unified data is then standardized and quantile normalized.

3. The method for generating low-carbon strategies based on predicted carbon emission factors according to claim 2, characterized in that, Also includes: The edge layer inputs the preprocessed power generation output data, user load data and environmental meteorological data of the power system into the short-term carbon emission factor prediction model to generate the carbon emission factor prediction value for the first time limit in the future. The short-term carbon emission factor prediction model is a moving average or exponential smoothing model. When the predicted value of the carbon emission factor within the first future time limit exceeds the short-term threshold, a corresponding control instruction is sent to the control module. During communication anomalies between the edge layer and the cloud layer, the edge layer inputs the preprocessed power generation output data, user load data, and environmental meteorological data into the emergency carbon emission factor prediction model to generate predicted carbon emission factor values ​​for the second time limit in the future. The emergency carbon emission factor prediction model is a simplified linear regression model. When the predicted value of carbon emission factor within the second future time limit exceeds the emergency period threshold, a corresponding emergency period load reduction control command is sent to the control module.

4. The method for generating low-carbon strategies based on predicted carbon emission factors according to claim 3, characterized in that, The cloud layer includes a central cloud and an edge cloud. The dynamic carbon emission factor prediction model and the low-carbon regulation optimization model are deployed in the central cloud. The method further includes: The edge layer sends the preprocessed power generation output data, user load data, and environmental meteorological data of the power system to the edge cloud, and then the edge cloud sends them to the central cloud. The edge layer sends a heartbeat packet to the edge cloud every set period. If the edge cloud does not receive the heartbeat packet for n consecutive periods, it is determined that an abnormal communication between the edge layer and the cloud layer has been detected. After determining that an abnormality in communication between the edge layer and the cloud layer has been detected, if the edge cloud continuously receives the heartbeat packet for m consecutive cycles, then it is determined that communication between the edge layer and the cloud layer has been restored. During communication anomalies between the edge layer and the cloud layer, the edge layer temporarily stores the pre-processed power generation output data, user load data, and environmental meteorological data. After communication is restored, the temporarily stored power generation output data, user load data, and environmental meteorological data are sent to the edge cloud, where the edge cloud performs data integrity verification before forwarding it to the central cloud.

5. The method for generating low-carbon strategies based on predicted carbon emission factors according to claim 4, characterized in that, The steps by which the edge layer sends the preprocessed power generation output data, user load data, and environmental meteorological data of the power system to the edge cloud include: The edge layer utilizes a deployed hardware encryption engine to perform AES-128 block encryption on the preprocessed power generation output data, user load data, and environmental meteorological data, generating encrypted data and a checksum. Simultaneously, it uses an RSA public key to encrypt the session key used for AES-128 data encryption, obtaining the session key ciphertext. The session key ciphertext, the encrypted data, and the checksum are then packaged and sent to the edge cloud. The edge cloud uses the corresponding RSA private key to decrypt the session key ciphertext to obtain the session key, and decrypts the encrypted data based on the session key. Integrity verification is then performed based on the checksum. The method for generating low-carbon strategies based on predicted carbon emission factors also includes: The terminal layer sets a collection timestamp T1 for the collected data, including power generation output data, user load data, and environmental meteorological data of the power system. After preprocessing the collected data, the edge layer calibrates the collection timestamp T1. When the deviation between the collection timestamp T1 and the edge layer local clock T0 is less than or equal to 100ms, the collection timestamp T1 is used as the synchronization timestamp T3. When the deviation is greater than 100ms, the synchronization timestamp T3 is determined according to the network reference time T2, where T3 = T1 + (T2 - T0), and the collection timestamp T1 of the collected data is updated to the synchronization timestamp T3. After receiving the collected data, the edge cloud calibrates the synchronization timestamp T3. When the deviation between the synchronization timestamp T3 and the edge cloud standard clock T4 is less than or equal to 500ms, the synchronization timestamp T3 is used as the final timestamp T5. When the deviation is greater than 500ms, the final timestamp T5 is determined according to the edge cloud standard clock T4, where T5 = T3 + (T4 - T3). The synchronization timestamp T3 is then updated to the final timestamp T5, and the collected data with the final timestamp T5 is sent to the central cloud.

6. The method for generating low-carbon strategies based on predicted carbon emission factors according to claim 1, characterized in that, The objective function of the low-carbon regulation and optimization model is: in, For total energy carbon emissions, Energy consumption As a dynamic carbon emission factor, This refers to the amount of material consumed. As a carbon emission factor of materials, In order to control the cost coefficient, Adjusting power costs for equipment To adjust the duration, For clean energy unit price, For green electricity replacement volume, These are the weighting coefficients; The constraints of the low-carbon regulation optimization model are as follows: in, This is the minimum power requirement for the equipment. Rated power of the equipment For actual output, For the target output, For carbon emission allowances.

7. The method for generating low-carbon strategies based on predicted carbon emission factors according to claim 1, characterized in that, The dynamic carbon emission factor prediction model includes: a spatiotemporal feature embedding layer, a dynamic graph structure generator, and a multimodal gating fusion layer. The spatiotemporal feature embedding layer takes the power output data on the power generation side, the load data on the user side, and the environmental meteorological data as inputs. The spatiotemporal feature embedding layer extracts local periodic features of the data through a convolutional neural network and inputs the output of the convolutional network into a long short-term memory network to capture time series features, thereby forming the power data features and meteorological data features of each power grid node respectively. The dynamic graph structure generator is composed of a dynamic graph attention network. Its input is the power data features output by the spatiotemporal feature embedding layer. It is used to dynamically update the edge weights between power grid nodes based on real-time transmission power, and calculate the correlation coefficient between nodes through the graph attention mechanism. It weights the importance of nodes in different time periods and spaces, and outputs the updated power data features. The multimodal gating fusion layer takes the updated power data features and the meteorological data features as input, and uses them to concatenate the updated power data features and the meteorological data features into a hybrid feature. It also dynamically adjusts the contribution of the power mode and the meteorological mode through a sigmoid gating mechanism, and outputs the fused feature after linear transformation to generate the carbon emission factor prediction curve for the next 24 hours. The training steps for the dynamic carbon emission factor prediction model include: Historical power generation output data, user load data, environmental meteorological data, and corresponding historical carbon emission factor real values ​​are obtained as training samples. The training samples are divided into a training set and a validation set. The training set is used to learn the parameters of the dynamic carbon emission factor prediction model, and the validation set is used to evaluate the prediction performance of the dynamic carbon emission factor prediction model during the training process. The parameters of the dynamic carbon emission factor prediction model are iteratively optimized using the backpropagation algorithm until the loss function of the prediction result converges to within a preset threshold.

8. The method for generating low-carbon strategies based on predicted carbon emission factors according to claim 7, characterized in that, The training steps for the dynamic carbon emission factor prediction model also include: Perturbations are applied to the training samples to obtain adversarial examples; The training samples and the adversarial samples are input together into the dynamic carbon emission factor prediction model for training. The mean squared error is used as the loss function, and the parameters of the dynamic carbon emission factor prediction model are updated using an adaptive moment estimation optimization algorithm.

9. The method for generating low-carbon strategies based on predicted carbon emission factors according to claim 1, characterized in that, Also includes: Based on the carbon emission factor prediction curve, the predicted average value of the carbon emission factor is calculated. A continuous period of time that is 30% higher than the predicted average value is marked as a peak period, and a continuous period of time that is 20% lower than the predicted average value is marked as a trough period. The steps for determining a low-carbon strategy include: Based on the long-term low-carbon control strategy and the user load characteristic map, during the peak period, one or more operations are performed on the flexible load, including reducing the electric vehicle charging power, increasing the air conditioning temperature, or reducing the lighting brightness. During the valley period, one or more operations are performed on the flexible load, including pre-cooling or preheating, restoring the electric vehicle charging power, or restoring the lighting operation.

10. A low-carbon strategy generation system based on predicted carbon emission factors, characterized in that, include: The system consists of a terminal layer and a cloud layer. The terminal layer collects power output data from the power generation side, load data from the user side, and environmental and meteorological data. The cloud layer inputs the power generation output data, user load data and environmental meteorological data into the dynamic carbon emission factor prediction model to generate a carbon emission factor prediction curve. The cloud layer inputs the carbon emission factor prediction curve into a low-carbon regulation and optimization model aimed at minimizing carbon emissions to obtain a long-term low-carbon control strategy. Based on the peak and valley characteristics of the carbon emission factor prediction curve, long-term low-carbon control strategies, and user load characteristic maps, the cloud layer determines low-carbon strategies that meet the grid carbon emission reduction targets and user energy needs.

Images