Carbon emission prediction method based on adaptive weight optimization

Abstract

The invention provides a carbon emission prediction method based on adaptive weight optimization, and relates to obtaining power output and consumption data from a regional power grid, calculating carbon emission reduction, and generating a plurality of execution strategies based on an optimization algorithm. A machine learning model of a self-adaptive feature selection and weight updating mechanism is adopted through historical data training, the carbon emission rate under each strategy is predicted, and dynamic updating of the model is achieved in combination with incremental learning. Finally, a prediction result is presented in a visual mode, a user is assisted in optimizing a decision, and accurate prediction and intelligent regulation and control of carbon emission are achieved.

Description

Technical Field

[0001] The present invention relates to the technical field of carbon emission management and optimization, and in particular to a carbon emission prediction method based on adaptive weight optimization, which belongs to the intersection of smart grid, energy management and environmental protection technology. Background Art

[0002] The role of regional power grids in the transformation of energy structure is becoming increasingly critical. In order to effectively control the total amount of carbon emissions, more and more power grid operation and management systems have begun to pay attention to carbon emissions in the process of electricity production and consumption, trying to predict and optimize the impact of power activities on the environment through technical means. In existing technologies, they mainly rely on regular statistics of carbon emissions, use fixed algorithms for load scheduling, or formulate emission reduction strategies based on manual experience. However, due to the complexity and dynamics of power grid operation data, these static or semi-static methods are often difficult to adapt to the rapidly changing load fluctuations and user behaviors in actual scenarios.

[0003] Traditional carbon emission management methods mostly rely on static emission factor calculation models, which are usually estimated based on information such as power generation type and fuel consumption, and have obvious lags in time granularity and response speed. Especially in the context of the rapid development of multi-source power generation and distributed power consumption, the distribution of carbon emissions in the power grid no longer has a simple and predictable linear relationship, and the existing estimation methods cannot fully capture the dynamic coupling characteristics between energy consumption behavior and carbon emissions, thus affecting the prediction accuracy and scheduling reliability.

[0004] In addition, some studies have begun to try to use data-driven methods to improve carbon emission prediction capabilities. Models based on linear regression, support vector machines or shallow neural networks have gradually been used for carbon emission trend analysis. Although such methods have improved modeling flexibility to a certain extent, in the face of uncertain factors such as frequent adjustments to the grid structure, changes in the electricity market mechanism, and diversified user behavior, conventional models are difficult to achieve dynamic management and structural adaptation of input features, resulting in a rapid decline in model performance after long-term operation, and even the accumulation of prediction errors.

[0005] More seriously, most current solutions lack a feedback mechanism for the effectiveness of strategy execution, and are unable to evaluate and compare the emission reduction efficiency of multiple scheduling or energy-saving solutions based on predictions. This "prediction island" design limits the system's ability to expand from intelligent perception to intelligent decision-making. At the same time, models that lack a continuous learning mechanism are difficult to iterate and optimize based on the latest operating data, and thus show poor adaptability in emergencies or boundary scenarios.

[0006] In summary, the existing carbon emission prediction and optimization technologies have multiple bottlenecks, and there is an urgent need to promote the integrated development of prediction, evaluation and optimization of carbon emission reduction plans through more intelligent, dynamic and strategy-oriented technical paths to meet the increasingly complex green power management needs. Summary of the invention

[0007] In order to solve the above problems in the prior art, the present invention proposes a carbon emission prediction method based on adaptive weight optimization, the prediction method comprising: Step S1, collecting power output data of the generator and power consumption data of multiple energy consuming users; Step S2, comparing the power consumption data with a preset energy consumption reference value to calculate the carbon emission reduction; Step S3, based on the carbon emission reduction, a plurality of execution strategies are generated using a multi-objective optimization algorithm; Step S4, training a supervised learning model based on historical power generation data, energy consumption data and corresponding carbon emission results; during the training process, dynamically selecting input features and adjusting feature weights according to model evaluation results; Step S5, inputting the execution strategy and current power data into the model to predict the carbon emission rate under each strategy; Step S6, incrementally updating the model based on the newly received data; Step S7: Output the predicted carbon emission rate results corresponding to each execution strategy in a visual manner to assist users in making decisions.

[0008] The step S1 further comprises: Deploy monitoring equipment on the power generation and consumption sides of the regional power grid to collect power output data of power generators and power consumption data of energy users; transmitting said data to a central data processing platform via a communication link; The data is preprocessed and formatted for use in the training of carbon emission prediction models and the calculation of control strategies.

[0009] The step S2 further comprises: The collected user electricity consumption data is compared with the preset energy consumption reference value to determine whether there is energy-saving behavior; when there is energy-saving behavior, the corresponding carbon emission reduction is calculated based on the carbon emission factor model and combined with the energy saving; the carbon emission reduction is used as the input parameter for generating the control strategy.

[0010] The step S3 further comprises: Based on the carbon emission reduction assessment results, an optimization model is constructed with the goal of maximizing carbon emission reduction; The load regulation parameters on the user side and the output regulation parameters on the power supply side are used as decision variables of the optimization model; Set grid operation constraints including supply and demand balance constraints and unit operation restrictions; A multi-objective optimization algorithm is used to solve the optimization model and generate multiple execution strategies, each of which includes a user load response suggestion and a power generation unit output configuration suggestion; The multiple execution strategies are used as inputs to a subsequent carbon emission prediction model.

[0011] The step S4 further comprises: A feature weight update mechanism is embedded in the training process to periodically evaluate the marginal impact of each input feature on the prediction error and adjust the input variable weights. The following weight adjustment function is used to update each feature weight:

[0012] in: Indicates Input features in round iteration The weight value of is the learning rate; Indicates that the feature is removed The change in the post-model error; is the sum of the absolute values ​​of all feature error changes; is a smoothing term used to prevent the denominator from being zero.

[0013] When a certain characteristic After being removed, the error increases, that is, , then the corresponding weight was raised; when When the weight was downgraded; : Indicates that under the weight update mechanism, the feature In the next iteration ( The new weight value in round).

[0014] The step S5 further comprises: Obtain power output data under the current grid operation status and power consumption data of energy users; Combining a plurality of generated execution strategies with current standardized power grid data to form an input vector; The input vector is input into the trained carbon emission prediction model to obtain the carbon emission prediction value corresponding to each strategy.

[0015] The step S6 further comprises: Continuously receive power grid operation data to build new sample data sets; Use the newly added samples for incremental model training according to the preset update frequency; Adopt online learning mechanism to fine-tune the parameters of the trained model to adapt to the changes in the grid operation status; When the grid operation status changes dramatically, a sliding window update mechanism is used to retain only samples from the most recent time period for training; When abnormal fluctuations in model performance are detected, the fallback mechanism is triggered to restore the model parameters to a stable state.

[0016] The step S7 further comprises: The carbon emission prediction results corresponding to each execution strategy are displayed on the interactive terminal through a graphical user interface; The graphical user interface supports filtering and sorting execution strategies and displays parameter information of the strategies; The above interface assists users in selecting emission reduction strategies and making optimal decisions.

[0017] Beneficial effects: The present invention uses a machine learning model to dynamically predict carbon emissions and combines it with an adaptive optimization strategy to achieve intelligent management of carbon emissions in regional power grids. The method can continuously update model parameters based on real-time data to improve the accuracy and timeliness of predictions. At the same time, by visually displaying the carbon emission results under a variety of strategies, users can be assisted in making decisions with greater emission reduction benefits, thereby effectively promoting green and low-carbon scheduling and improving the environmental adaptability of power grid operation and the level of carbon emission control. BRIEF DESCRIPTION OF THE DRAWINGS

[0018] The drawings described herein are used to provide a further understanding of the present invention and constitute a part of the present application, but do not constitute an improper limitation of the present invention. In the drawings: Figure 1 The overall flow chart of a carbon emission prediction method based on adaptive weight optimization of the present invention is shown. DETAILED DESCRIPTION

[0019] The present invention will be described in detail below in conjunction with the accompanying drawings and specific embodiments, wherein the illustrative embodiments and descriptions are only used to explain the present invention but are not intended to limit the present invention.

[0020] As attached Figure 1 As shown, a carbon emission prediction method based on adaptive weight optimization of the present invention specifically includes the following steps: Step S1: Data collection and input like Figure 1As shown in the figure, the system collects the power output data of the generator and the power consumption data of multiple energy-consuming users in real time through the monitoring equipment deployed in the regional power grid. The power output data includes the real-time power, frequency, voltage and other operating parameters of various generator sets; the power consumption data includes the user's power load, daily cycle power consumption behavior, etc. The above data is transmitted to the central data processing platform through the communication interface as the basic input for subsequent processing.

[0021] Step S2: Calculation of carbon emission reduction The system presets energy consumption reference value libraries for different types of users (such as industrial users, commercial users, and residential users). The collected electricity consumption data of each user is compared with the reference value, and the carbon emission reduction of each user per unit time is calculated based on the carbon emission factor model. The calculation result is used to evaluate the carbon benefits of the user's energy-saving behavior and serve as the input basis for strategy generation.

[0022] Step S3: Execution strategy generation Based on the carbon reduction assessment results, the system uses a multi-objective optimization algorithm, such as the non-dominated sorting genetic algorithm NSGA-II or the improved particle swarm optimization algorithm, to generate multiple execution strategies. Each strategy includes load response adjustment suggestions, power configuration suggestions on the power supply side, etc., aiming to maximize the overall carbon reduction under the premise of meeting the grid security constraints.

[0023] Step S4: Machine learning model training The system builds a carbon emission prediction model, and the training data includes historical power generation data, user energy consumption data and their corresponding carbon emission results. The adaptive feature selection mechanism is used in the training process to dynamically select the most representative input variables at the current stage, such as power fluctuation rate, user type, time node, etc.

[0024] The model adopts a supervised learning framework, selects regression algorithms such as gradient boosted decision tree GBDT or deep neural network DNN, and embeds a feature weight update mechanism during the training process. The system periodically adjusts the weight distribution of input variables based on the cross-validation evaluation results, thereby improving the model's adaptability to the changing power grid environment and prediction stability.

[0025] Step S5: Strategy effect prediction After completing the model training, the system calls the current power output and consumption data, inputs the generated multiple execution strategies into the model in turn, and predicts the future carbon emission rate of each strategy under a given operating state. Each predicted value is a joint function of the strategy variables, power data and model parameters, reflecting the actual impact of the strategy on carbon emissions after implementation.

[0026] Step S6: Model incremental update The system continuously receives new grid operation data and conducts incremental learning on the trained model in combination with actual carbon emission feedback results. It uses mechanisms such as online learning and sliding window updates to continuously optimize the model weights and achieve rapid response to dynamic changes in time and sudden load events.

[0027] Step S7: Visualization output and decision support Finally, the system will graphically display the predicted carbon emission rates corresponding to each execution strategy on the operator interaction screen, including visual elements such as bar charts and line trend charts. Users can intuitively compare strategy options based on the prediction results to assist in making the optimal carbon emission reduction scheduling decision.

[0028] This embodiment builds a carbon emission prediction system that integrates data perception, intelligent analysis, and adaptive optimization, and realizes a closed-loop optimization process from data collection, model training, strategy evaluation to user decision-making. It has good real-time, accuracy, and scalability, and is suitable for carbon emission control scenarios in regional power grids, microgrids, and energy management platforms.

[0029] The implementation methods of each step are as follows: The data collection and input process described in step S1 is the basic link for the present invention to achieve carbon emission prediction and optimization. It mainly realizes real-time data collection on the power generation side and the power consumption side by deploying monitoring equipment at key nodes of the regional power grid. The monitoring equipment on the power generation side is installed at the outlet of various types of generator sets, and is equipped with a power parameter acquisition unit with high-precision sampling capability, which can obtain power output related data at a frequency of seconds or higher. Specific acquisition parameters include but are not limited to: three-phase active power, reactive power, voltage, current, frequency and equipment operation status signals. The data is used to reflect the actual output capacity of the generator set and its contribution to the power supply structure of the power grid.

[0030] The data collection device on the electricity consumption side is usually integrated into the user-side smart meter or load monitoring device to obtain the electricity consumption behavior information of various types of energy-consuming users. For different types of users, such as industrial enterprises, commercial complexes, and residential communities, the collection indicators include daily load curves, time-divided power factors, load fluctuation amplitudes, electricity density, etc., to characterize the load characteristics and energy consumption patterns of each user. For user scenarios with multiple load classifications, the collection system can also be refined to specific load modules such as lighting, air conditioning, and motor drives to achieve higher-resolution load behavior analysis.

[0031] The above monitoring equipment is connected to the power grid communication network, and the collected data is uploaded to the central data processing platform through high-speed communication links (such as optical fiber, 5G, NB-IoT, etc.). The edge preprocessing mechanism is used in the data transmission process to first perform basic processing such as noise filtering, missing value filling and timestamp synchronization to ensure the integrity and accuracy of the data. After receiving the data, the central platform performs unified formatting and structured processing, and converts it into the data structure required for model training and strategy calculation to ensure the consistency and operability of the data input of the subsequent processes of the system.

[0032] Through the above design, this step not only achieves full-link data coverage from the power generation end to the power consumption end, but also ensures the real-time, accuracy and scalability of the data, laying a solid data foundation for subsequent carbon emissions calculation and optimization strategy formulation.

[0033] The carbon emission reduction calculation process described in step S2 is based on the comparison between the user-side energy consumption data and the preset energy consumption reference value, and is quantified in combination with the carbon emission factor model. In order to meet the carbon emission reduction assessment needs of differentiated users, the system first establishes an energy consumption reference value database for multiple types of users. The database sets its baseline energy consumption level according to the user's type (such as industrial, commercial, residential), electricity consumption scale, historical load characteristics and the power grid operation standards in the region. The energy consumption reference value can be set in a hierarchical manner according to the statistical period (such as day, week, month), and supports dynamic updates to reflect energy consumption fluctuations caused by seasonality and changes in economic activities.

[0034] During operation, the system compares the real-time collected electricity consumption per user per unit time with the corresponding reference value. If the actual energy consumption is lower than the reference value, it is considered to be energy-saving behavior. To ensure the fairness of the evaluation, the system uses a time alignment mechanism and power standardization to eliminate the interference of data in statistical cycles and load fluctuations, ensuring that the comparison results are comparable and representative.

[0035] Based on the determination of the user's energy-saving behavior, the system calculates the carbon emission reduction per unit time according to the carbon emission factor model. The carbon emission factor is a coefficient value representing the carbon emission corresponding to each kilowatt-hour of electricity. This value can be set with reference to the energy structure of the regional power grid (such as the proportion of coal-fired power, gas-fired power, and renewable energy), and has time and regional differences. By multiplying the energy-saving electricity by the carbon emission factor of the corresponding period, the carbon emission reduction value can be obtained.

[0036] To improve computing efficiency and scalability, the system uses a batch processing mechanism to centrally calculate the data of all users within a unified statistical period and output the results in a structured manner. The calculated carbon emission reduction will be input into the subsequent strategy generation module as a key evaluation indicator to determine the environmental benefits of scheduling plans and user behavior adjustments, thereby achieving carbon-benefit-oriented optimization and regulation. The entire calculation process has good scalability while ensuring accuracy, and is suitable for the carbon emission management needs of large-scale user groups.

[0037] The execution strategy generation described in step S3 is based on the carbon emission reduction assessment results calculated in the previous stage, and a multi-objective optimization algorithm is used to model and solve the control scheme. The system first parameterizes the adjustment capabilities of the user side and the power supply side to construct a set of decision variables for the optimization problem. Specifically, the adjustable loads on the user side include interruptible loads, shiftable loads, and load-reducible loads, and their adjustment amplitudes and response periods are used as variable inputs; the adjustment parameters on the power supply side include the upper and lower output limits and adjustment slopes of each power generation unit to reflect the controllability of the unit.

[0038] When constructing the optimization goal, the system adopts a dual objective function: on the one hand, the core goal is to maximize the total carbon emission reduction of the system, and on the other hand, it takes into account auxiliary indicators such as operating costs, load balance or user comfort, so as to reflect the coordination of environmental benefits and operational stability in the optimization process. The above optimization goals are nonlinear, non-convex and uncertain, and are suitable for solving using multi-objective evolutionary algorithms.

[0039] The system preferably uses the non-dominated sorting genetic algorithm NSGA-II or its improved algorithm. The algorithm constructs an initial population, iteratively performs selection, crossover and mutation operations, and uses a congestion sorting mechanism to control the distribution of the solution set, and finally obtains a set of Pareto optimal solutions that do not dominate each other in the target space. Each solution is a set of execution strategies, including load response recommendations for different users and output configuration plans for different power generation units. In order to improve the convergence speed and solution diversity of the algorithm solution, the system can adapt and set dynamic weight adjustment strategies and adaptive mutation operators during execution.

[0040] During the entire optimization process, the system sets the grid operation constraints as hard constraints, including supply and demand balance constraints, line power limits, unit ramp rate limits, and maximum adjustable capacity of users. The optimization algorithm performs constraint checks on candidate solutions in each iteration, and performs penalty function corrections or elimination operations on solutions that violate the constraints to ensure that the final generated strategy is feasible.

[0041] After completing the strategy generation, the system deconstructs the multiple sets of Pareto solutions obtained into a structured data format, and identifies the corresponding load response parameters, power output adjustment suggestions and expected carbon emission reduction benefits, which are used as the input content of the subsequent machine learning prediction module to further evaluate the actual carbon emission performance of each strategy under different operating conditions. Through the above mechanism, this step realizes the intelligent generation of multiple schemes for carbon emission reduction goals, providing flexible, controllable and low-carbon strategy support for power system operation and management.

[0042] The machine learning model training described in step S4 aims to build a carbon emission estimation model with high-precision prediction capabilities, which is used to quantitatively evaluate future carbon emission rates under different execution strategies. The model is based on a supervised learning framework, and the training data comes from historical power grid operation records, including power output data on the power generation side, energy consumption data on the user side, and actual carbon emission results corresponding to each data period. The carbon emission results can be indirectly calculated by the existing emission monitoring system or emission factor model, and paired with the input features in a timestamp-aligned manner to ensure the consistency and traceability of the training data.

[0043] In the preprocessing stage of model training, the system uses an adaptive feature selection mechanism to dynamically screen the input variables in the original feature set. This mechanism constructs a sliding time window or a rolling training set to evaluate the importance of each feature to the prediction result at the current stage in real time, such as information gain, correlation coefficient, or SHAP value, and retains the most representative variables accordingly. For example, for load change scenarios with obvious periodicity, the system may give priority to feature inputs such as time nodes (such as hours, weekend / weekday identification), power fluctuation rate, and user type classification code, thereby effectively reducing the interference of invalid variables on the model and improving the efficiency and generalization ability of model training.

[0044] The model body can be a regression learning model such as gradient boosting decision tree (GBDT), extreme gradient boosting (XGBoost) or deep neural network (DNN). The GBDT model is suitable for small and medium-sized structured data and has strong explanatory power; the DNN model is suitable for large-scale data structures with significant nonlinear relationships and has strong fitting capabilities. The selection of the model structure is flexibly configured based on the dimension of the training data, the sample size and the target accuracy requirements.

[0045] During the model training process, the system embeds a feature weight update mechanism to periodically adjust the importance of input variables. This mechanism uses cross-validation results as feedback, and dynamically optimizes feature weights by evaluating the prediction error performance of the model under different input variable combinations. For example, when an input variable does not significantly improve the prediction effect in multiple validation compromises, the system automatically reduces the participation of the variable in subsequent training, and vice versa, increases its weight to achieve the optimal adaptation between the input variable and the grid operation status.

[0046] Specifically, a set of learnable dynamic adjustment function weight vectors is adopted , where each weight Corresponding input features The system defines the weight update function after each iteration for:

[0047] The specific parameters are described as follows: : represents input features In the current iteration The weight value when . This value reflects the feature The importance of the model in the current structure is a dynamically learnable variable.

[0048] : Indicates that under the weight update mechanism, the feature In the next iteration ( The new weight value in round). Its value is determined by the current weight With features The marginal contribution to the model error is jointly determined.

[0049] : is the learning rate, which controls the sensitivity of weight adjustment in each iteration. Its value is usually set to a small positive number (such as 0.01 or 0.001) to avoid instability caused by too fast updates. Will make the weight more sensitive to error changes, suitable for high response scenarios; smaller It is suitable for stable scenarios that suppress oscillations.

[0050] :Indicates in In the iteration, remove the features The change in the model prediction error (such as mean square error MSE) caused by the model prediction error. The specific definition is:

[0051] in, is the error obtained by training the model using the full feature set, To remove features Then re-evaluate the error. Therefore, if , indicating that the error increases after removing the feature, indicating that the feature has a positive contribution to the model and the weight should be increased; if , it means that the feature may contain noise and its weight should be lowered.

[0052] : Represents the sum of the absolute values ​​of the error changes of all features in the current round, which is used for normalization to prevent the update stability from being affected by the magnitude differences of different features.

[0053] : is a regularization constant or smoothing term (such as ), which is used to prevent numerical instability or gradient explosion caused by the denominator being zero or too small. Its value should be small enough not to affect the normalized ratio, but sufficient to stabilize the calculation process.

[0054] This formula evaluates the marginal impact of each input feature on the prediction error in each iteration and dynamically adjusts its weight in the model, thereby achieving adaptive optimization of the model structure. Positive adjustment enhances the influence of features that are beneficial to the prediction, and negative adjustment suppresses noise features, effectively responding to input disturbances caused by frequent changes in the operating status of the power grid, and improving the robustness and accuracy of the model.

[0055] The training process uses Batch Gradient Descent or its improved form, and cooperates with the Early Stopping mechanism to prevent model overfitting. After the training is completed, the system saves the model weight parameters and outputs the standard model structure interface for the prediction module, so that the strategy effect can be evaluated in combination with real-time data. Through the above process, this step realizes a carbon emission prediction model that can adapt to the dynamic power grid environment changes, providing an accurate evaluation basis for strategy selection and optimization.

[0056] The strategy effect prediction described in step S5 is the inference stage after the machine learning model training is completed. The purpose is to evaluate the carbon emission performance of each execution strategy in the future period based on the current power grid operation data and the preset strategy parameters. The system first obtains the current power output data and the real-time power consumption data of energy users. After the data is standardized, it is used as the state quantity of the prediction input to ensure the consistency in structure and scale with the input features used in the model training stage.

[0057] Subsequently, the system calls the generated execution strategy solutions one by one. Each strategy contains decision variables such as user load adjustment suggestions and power output change parameters. The system combines these strategy variables with the current grid status data to form a complete input vector set, which is input into the trained carbon emission prediction model. The model uses the learned functional relationship to perform forward calculations on each set of inputs and outputs the corresponding carbon emission rate prediction value.

[0058] The predicted value is a scalar output, usually in grams per kilowatt-hour (g / kWh) or tons per hour (t / h), which is used to quantitatively characterize the carbon emission intensity of the strategy under the current operating state. The prediction process takes into account the nonlinear coupling relationship between input characteristics, and pays special attention to the amplification or suppression effect of strategy adjustment behavior on changes in power generation structure and load transfer effects, thereby achieving a refined assessment of carbon emission response.

[0059] To improve the stability and reliability of the prediction results, the system can use multiple rounds of redundant calculations in the prediction stage, that is, execute multiple sub-prediction processes for each strategy (such as using different time window averages, different data perturbation inputs, etc.), and take a weighted average of the prediction results to alleviate the error impact of single-point data fluctuations on the results.

[0060] The prediction results are output in the form of a structured array, recording the input parameter combination of each strategy and its corresponding carbon emission prediction value. This result is not only used for subsequent visualization, but also serves as a key basis for strategy optimization, supporting users to select the one with the best carbon emission reduction effect from multiple options, thereby achieving a close linkage between carbon emission control and operation strategy execution. Through the above design, the system effectively establishes an efficient closed loop from strategy generation to emission reduction effect prediction, improving the scientificity and controllability of carbon emission management.

[0061] The model incremental update mechanism described in step S6 is used to improve the dynamic adaptability of the carbon emission prediction model during long-term operation. The system constructs a new sample data set by continuously receiving real-time operation data from the power grid, including generator output, user-side load, and actual carbon emission values ​​monitored. The new data is stored in a time series manner and is processed by standardization, missing value repair, and feature alignment to ensure that it can be directly used in the incremental training process.

[0062] To achieve efficient model updates, the system uses an online learning mechanism that allows the model to gradually receive and learn new data samples without retraining the entire data set. Under this mechanism, the system inputs the latest samples into the model according to the preset update frequency (such as every hour, every day, or every forecast cycle), and fine-tunes the existing model weight parameters to enable it to quickly adapt to changes in the current power grid operation status. For models that use gradient boosting algorithms, the system supports adding new weak learners to the original model; for neural network models, small batch training can be used to perform limited rounds of gradient backpropagation operations.

[0063] When some power grid environments change dramatically or sudden load fluctuations occur, the system can trigger a sliding window update mechanism, which only retains the training sample window in the most recent period as the current valid training set to remove outdated samples that do not contribute to the current state prediction and improve the time sensitivity of the model. The window size can be adjusted dynamically and set according to the changes in data distribution to achieve a balance between optimal forgetting of historical data and optimal retention of new data.

[0064] During the incremental update process, the system monitors whether the model performance indicators (such as prediction error, residual distribution, etc.) fluctuate within a reasonable range. If the model is found to be offset or performance degraded, the system can automatically trigger the rollback mechanism to restore the model weights to the previous stable state and annotate the new data for abnormalities to prevent erroneous samples from interfering with model evolution.

[0065] Through this incremental learning mechanism, the model can achieve continuous performance optimization without the need for frequent overall retraining, which not only improves the system's ability to track long-term trend changes, but also enhances its ability to respond to short-term emergencies, thereby ensuring the continued accuracy and robustness of carbon emission forecast results in a dynamic power grid environment.

[0066] The visualization output and decision-making assistance function described in step S7 is used to present the carbon emission rate results obtained in the strategy effect prediction stage in a graphical manner on the operator interaction terminal to enhance the intuitiveness of strategy comparison and the efficiency of user decision-making. The system constructs a graphical user interface (GUI) to centrally display the carbon emission prediction indicators corresponding to each execution strategy on the terminal display screen, and displays the strategy parameter summary information in conjunction to achieve a direct connection between the data results and the control behavior.

[0067] The system preferably uses a variety of graphical elements such as bar charts, line trend charts and tables to present information. The bar chart is used to display the predicted carbon emission values ​​of different strategies in the current period. The horizontal axis is the strategy number or brief description, and the vertical axis is the corresponding predicted carbon emission rate value; by comparing the column height, users can quickly identify strategies with excellent emission reduction effects. The line trend chart is used to display the carbon emission change trend of a certain strategy in multiple time periods in the future. It supports scrolling viewing and multi-strategy superposition comparison, which is convenient for identifying the stability and volatility of carbon emission trends.

[0068] The interface supports users to filter and sort the strategy sets, and the filtering conditions may include carbon emission thresholds, load adjustment range, power output change range, etc. Users can also obtain the specific parameter combination, applicable user type and associated operation constraint information of each strategy through floating prompts and drop-down menus to assist in understanding the strategy formation logic and its executability.

[0069] To further improve the interaction efficiency, the system supports dynamic refresh and historical comparison of chart content. Users can choose to compare the strategy effects under different forecast cycles, or compare the carbon emission performance of the current forecast results with historical operating data under similar conditions, so as to achieve experience feedback and strategy iteration. All graphical data comes from the output of the system's internal forecast module, with real-time and consistency, ensuring that the strategy effect data obtained by users in the decision-making process is authentic and reliable.

[0070] Ultimately, the system converts complex strategy prediction data into readable and comparable graphical content through a visual interface, allowing users to make scientific judgments based on intuitive results, thereby completing the optimal carbon emission reduction scheduling strategy selection and improving the green and intelligent level of the overall system operation.

[0071] The above description is only a preferred embodiment of the present invention, so all equivalent changes or modifications made according to the structure, characteristics and principles described in the scope of the patent application of the present invention are included in the scope of the patent application of the present invention.

Claims

1. A carbon emission prediction method based on adaptive weight optimization, characterized in that: The prediction method comprises: Step S1, collecting power output data of the generator and power consumption data of multiple energy consuming users; Step S2, comparing the power consumption data with a preset energy consumption reference value to calculate the carbon emission reduction; Step S3, based on the carbon emission reduction, a plurality of execution strategies are generated using a multi-objective optimization algorithm; Step S4, training a supervised learning model based on historical power generation data, energy consumption data and corresponding carbon emission results; during the training process, dynamically selecting input features and adjusting feature weights according to model evaluation results; Step S5, inputting the execution strategy and current power data into the model to predict the carbon emission rate under each strategy; Step S6, incrementally updating the model based on the newly received data; Step S7: Output the predicted carbon emission rate results corresponding to each execution strategy in a visual manner to assist users in making decisions.

2. A carbon emission prediction method based on adaptive weight optimization as claimed in claim 1, characterized in that: The step S1 further comprises: Deploy monitoring equipment on the power generation and consumption sides of the regional power grid to collect power output data of power generators and power consumption data of energy users; transmitting said data to a central data processing platform via a communication link; The data is preprocessed and formatted for use in the training of carbon emission prediction models and the calculation of control strategies.

3. The carbon emission prediction method based on adaptive weight optimization according to claim 1, characterized in that: The step S2 further comprises: The collected user electricity consumption data is compared with the preset energy consumption reference value to determine whether there is energy-saving behavior; when there is energy-saving behavior, the corresponding carbon emission reduction is calculated based on the carbon emission factor model and combined with the energy saving; the carbon emission reduction is used as the input parameter for generating the control strategy.

4. The carbon emission prediction method based on adaptive weight optimization according to claim 1, characterized in that: The step S3 further comprises: Based on the carbon emission reduction assessment results, an optimization model is constructed with the goal of maximizing carbon emission reduction; The load regulation parameters on the user side and the output regulation parameters on the power supply side are used as decision variables of the optimization model; Set grid operation constraints including supply and demand balance constraints and unit operation restrictions; A multi-objective optimization algorithm is used to solve the optimization model and generate multiple execution strategies, each of which includes a user load response suggestion and a power generation unit output configuration suggestion; The multiple execution strategies are used as inputs to a subsequent carbon emission prediction model.

5. The carbon emission prediction method based on adaptive weight optimization according to claim 1, characterized in that: The step S4 further comprises: A feature weight update mechanism is embedded in the training process to periodically evaluate the marginal impact of each input feature on the prediction error and adjust the input variable weights. The following weight adjustment function is used to update each feature weight: in: : Indicates that under the weight update mechanism, the feature In the New weight value in round; Indicates Input features in round iteration The weight value of is the learning rate; Indicates that the feature is removed The change in the post-model error; is the sum of the absolute values ​​of all feature error changes; is a smoothing term used to prevent the denominator from being zero.

6. A carbon emission prediction method based on adaptive weight optimization as claimed in claim 5, characterized in that: When the characteristics After being removed, the error increases, that is, When the corresponding weight was raised; when When the weight was downgraded.

7. The carbon emission prediction method based on adaptive weight optimization according to claim 1, characterized in that: The step S5 further comprises: Obtain power output data under the current grid operation status and power consumption data of energy users; Combining a plurality of generated execution strategies with current standardized power grid data to form an input vector; The input vector is input into the trained carbon emission prediction model to obtain the carbon emission prediction value corresponding to each strategy.

8. The carbon emission prediction method based on adaptive weight optimization according to claim 1, characterized in that: The step S6 further comprises: Continuously receive power grid operation data to build new sample data sets; Use the newly added samples for incremental model training according to the preset update frequency; Adopt online learning mechanism to fine-tune the parameters of the trained model to adapt to the changes in the grid operation status; When the grid operation status changes dramatically, a sliding window update mechanism is used to retain only samples from the most recent time period for training; When abnormal fluctuations in model performance are detected, the fallback mechanism is triggered to restore the model parameters to a stable state.

9. The carbon emission prediction method based on adaptive weight optimization according to claim 1, characterized in that: The step S7 further comprises: The carbon emission prediction results corresponding to each execution strategy are displayed on the interactive terminal through a graphical user interface; The graphical user interface supports filtering and sorting execution strategies and displays parameter information of the strategies; The above interface assists users in selecting emission reduction strategies and making optimal decisions.

Images