Multi-dimensional carbon emission prediction method, system, equipment and medium

Abstract

The invention relates to the technical field of carbon emission prediction, in particular to a multi-dimensional carbon emission prediction method, system and device and a medium, and the method comprises the steps: carrying out the parallel training of a plurality of prediction models based on a core feature subset, and dynamically triggering the model retraining or switching based on a model performance detection result, and generating an initial prediction result; based on the error index feedback of the initial prediction result, automatically optimizing the hyper-parameter configuration of each model, and generating an optimized prediction model; and performing multi-index comprehensive evaluation on the initial prediction result and the optimized prediction model, and outputting a target prediction result through a weighted integration or rule selection mechanism. The method has the advantages that the problem of data and model separation in gas-electricity enterprise carbon emission prediction is solved, and high-precision and high-reliability prediction of complex nonlinear emission behaviors is achieved by means of a dynamic self-adaption and interpretability decision-making mechanism.

Description

Technical Field

[0001] This invention relates to the field of carbon emission prediction technology, and in particular to a multi-dimensional carbon emission prediction method, system, device and medium. Background Technology

[0002] As a key sector for carbon emissions, the energy industry is facing increasingly stringent emission reduction requirements. Gas-fired power plants, as an important component of the thermal power industry, are affected by multiple factors, including equipment structure, energy efficiency, and the external environment, resulting in emission patterns that are significantly nonlinear and complex. However, current mainstream carbon emission monitoring methods are still mainly based on static accounting, generally relying on annual emission factors or fuel consumption inventories, which cannot meet the needs of real-time, accurate and intelligent prediction. In addition, traditional methods are difficult to effectively integrate multi-source heterogeneous data, lack a unified data preprocessing and feature extraction mechanism, and most existing carbon emission prediction studies are based on a single modeling method, lacking a multi-model integration strategy for high spatiotemporal resolution, resulting in poor stability and interpretability of prediction results. Summary of the Invention

[0003] To solve the above-mentioned technical problems, the present invention provides the following technical solution: In a first aspect, the present invention provides a multi-dimensional carbon emission prediction method, which includes acquiring multi-source heterogeneous data in the gas-fired power generation process and generating a standardized dataset after preprocessing. Based on the dataset, time-series statistical features are constructed using sliding windows across multiple time scales, and a subset of core features is selected based on feature importance evaluation. Multiple prediction models are trained in parallel based on a subset of core features, and model retraining or switching is dynamically triggered based on model performance detection results to generate initial prediction results. Based on the error index feedback of the initial prediction results, the hyperparameter configuration of each model is automatically optimized to generate the optimized prediction model. The initial prediction results and the optimized prediction model are comprehensively evaluated using multiple indicators, and the target prediction results are output through weighted integration or rule selection mechanisms.

[0004] As a preferred embodiment of the multi-dimensional carbon emission prediction method of the present invention, the method is deployed through a microservice architecture and supports operation on local servers, edge gateways or cloud platforms. The system provides a visual user interface and data interface, and uses incremental learning to correct the prediction model in real time based on user feedback.

[0005] As a preferred embodiment of the multi-dimensional carbon emission prediction method of the present invention, the preprocessing includes spatiotemporal alignment of multi-frequency collected data through a sliding window mechanism; For high-frequency data, statistical values ​​are calculated within a window to generate reference frequency data points; For low-frequency data, interpolation methods are used to extend it to the reference frequency.

[0006] As a preferred embodiment of the multi-dimensional carbon emission prediction method of the present invention, the core feature subset is selected based on feature importance assessment, including: At least one of the following methods, namely correlation coefficient analysis, rank correlation analysis, or mutual information calculation, is used to score the importance of each feature in the feature set; Based on cross-validation stability assessment and adaptive threshold, the core feature subset for final input model training is determined from the feature set.

[0007] As a preferred embodiment of the multi-dimensional carbon emission prediction method of the present invention, the model performance detection logic includes: When the continuous prediction error exceeds the threshold calculated based on the historical error mean and standard deviation, the model retraining or switching mechanism is automatically triggered. Each time a model is trained, a feature analysis report is generated, including a SHAP value plot and an impact factor ranking plot.

[0008] As a preferred embodiment of the multidimensional carbon emission prediction method of the present invention, the hyperparameter configuration of each model is automatically optimized using a hybrid strategy of Bayesian optimization and grid search. The optimized hyperparameters include training period, learning rate, and regularization strength, which are optimized using a two-stage parameter tuning method.

[0009] As a preferred embodiment of the multidimensional carbon emission prediction method of the present invention, the baseline window length is 1 hour.

[0010] Secondly, the present invention provides a multi-dimensional carbon emission prediction system, including: an acquisition module for acquiring multi-source heterogeneous data in the gas-fired power generation process, and generating a standardized dataset after preprocessing; The filtering module is used to construct time-series statistical features based on the dataset through a sliding window at multiple time scales, and to filter out a subset of core features based on feature importance evaluation. The first generation module is used to train multiple prediction models in parallel based on a subset of core features, and dynamically trigger model retraining or switching based on model performance detection results to generate initial prediction results. The second generation module is used to automatically optimize the hyperparameter configuration of each model based on the error index feedback of the initial prediction results, and generate the optimized prediction model. The output module is used to comprehensively evaluate the initial prediction results and the optimized prediction model using multiple indicators, and outputs the target prediction results through weighted integration or rule selection mechanisms.

[0011] Thirdly, the present invention provides a computer device, including a memory and a processor, wherein the memory stores a computer program, and the processor executes the computer program to implement the steps of the method described above.

[0012] Fourthly, the present invention provides a computer-readable storage medium having a computer program stored thereon, wherein the computer program, when executed by a processor, implements the steps of the method described above.

[0013] Compared with existing technologies, the beneficial effects of this invention are as follows: By standardizing and preprocessing multi-source heterogeneous data and aligning it with time series, it effectively solves the shortcomings of traditional methods such as single data dimension and lagging updates; by using a multi-timescale sliding window to extract dynamic statistical features and combining importance assessment to screen core feature subsets, it enhances the interpretability and anti-interference ability of input variables; by using a parallel training mechanism and data-driven multi-class prediction models, combined with dynamic weight allocation and hyperparameter adaptive optimization mechanisms, it achieves improved accuracy and stable switching under different operating scenarios; by integrating and evaluating the prediction results based on sliding error window weighted fusion and trend change identification, it can provide real-time early warning of emission anomalies and preliminary identification of causes, improving prediction transparency and decision-making confidence, and providing methodological support for gas-fired power companies to achieve refined prediction and dynamic management of carbon emissions. Attached Figure Description

[0014] To more clearly illustrate the technical solutions of the embodiments of the present invention, the drawings used in the following description of the embodiments will be briefly introduced. Obviously, the drawings described below are only some embodiments of the present invention. For those skilled in the art, other drawings can be obtained based on these drawings without creative effort.

[0015] Figure 1 This is a flowchart illustrating a multi-dimensional carbon emission prediction method. Detailed Implementation

[0016] To make the above-mentioned objects, features, and advantages of the present invention more apparent and understandable, specific embodiments of the present invention will be described in detail below with reference to the accompanying drawings. Obviously, the described embodiments are only a part of the embodiments of the present invention, and not all of them. Based on the embodiments of the present invention, all other embodiments obtained by those skilled in the art without creative effort should fall within the protection scope of the present invention.

[0017] Example 1, referring to Figure 1 This is the first embodiment of the present invention, which provides a multi-dimensional carbon emission prediction method, including: S100: Acquire multi-source heterogeneous data during the gas-fired power generation process, and generate a standardized dataset after preprocessing; S200: Based on the dataset, time-series statistical features are constructed through sliding windows at multiple time scales, and a subset of core features is selected based on feature importance evaluation; S300: Trains multiple prediction models in parallel based on a subset of core features, and dynamically triggers model retraining or switching based on model performance detection results to generate initial prediction results; S400: Based on the error index feedback of the initial prediction results, it automatically optimizes the hyperparameter configuration of each model and generates the optimized prediction model. S500: Performs a comprehensive evaluation of the initial prediction results and the optimized prediction model using multiple indicators, and outputs the target prediction results through weighted integration or rule selection mechanisms.

[0018] It should be noted that gas-fired power plants, as the main force in peak shaving of the power system, are affected by a combination of factors such as fuel composition fluctuations, equipment performance degradation, and variable dispatch conditions, exhibiting significant nonlinear and multi-scale dynamic characteristics in their carbon emissions. Existing carbon emission monitoring relies on static accounting algorithms based on annual emission factors or fuel inventories, which cannot meet the needs of hourly real-time prediction. At the same time, the heterogeneity of data from different production, fuel, and monitoring systems, the different sampling frequencies, and the lack of a unified preprocessing mechanism lead to low information utilization. Furthermore, single-mechanism or data-driven models cannot adapt to complex scenarios such as hydrogen-blended combustion and frequent start-stop operations, resulting in poor prediction stability and interpretability.

[0019] Therefore, to address the aforementioned problems, this method, through steps S100-S500, first performs standardized preprocessing on multi-source heterogeneous data to break down information silos. Then, it mines temporal dependencies through multi-timescale sliding windows to select dynamic feature subsets sensitive to emissions. Furthermore, by employing a multi-model parallel training and dynamic switching mechanism, it can adapt to fuel fluctuations and changes in operating conditions. Even further, it automatically optimizes hyperparameters based on prediction error feedback to continuously improve model accuracy and generalization ability. Finally, it outputs high-precision prediction results through weighted integration or rule selection and drives emission reduction decisions in reverse, achieving a closed-loop iteration of prediction-optimization.

[0020] Example 2, refer to Figure 1 As an embodiment of the present invention, a multi-dimensional carbon emission prediction method is provided based on the above embodiment.

[0021] In this embodiment of the application, step S100 involves acquiring multi-source heterogeneous data from the gas-fired power generation process, preprocessing it to generate a standardized dataset, and including the following steps A1-A3: It should be noted that the multi-source heterogeneous data in the gas-fired power generation process includes real-time monitored energy consumption indicators, historical operation logs, external meteorological data, and regional environmental information. This can be understood as including, but not limited to, real-time monitored energy consumption indicators such as unit load, steam flow, gas flow, heat rate, power generation, grid-connected power, boiler efficiency, combustion temperature, and real-time data from various sensors. These energy consumption indicators provide the system with necessary input information for real-time monitoring and accurate prediction of carbon emissions from gas-fired power plants, thereby supporting refined management and dynamic optimization of carbon emissions. Historical operation logs include, but are not limited to, unit start-up and shutdown records, equipment fault logs, maintenance records, historical emission monitoring data, and operator operation logs. Historical data, such as records, can be used to analyze equipment operating status and failure modes, thereby helping the system to more accurately predict future carbon emission behavior and optimize management strategies; external meteorological data, including but not limited to temperature, humidity, wind speed and direction, air pressure, precipitation, and air quality index (AQI), provides a basis for carbon emission prediction based on external environmental conditions and helps analyze the impact of meteorological changes on carbon emissions; regional environmental information, including but not limited to regional emission restriction policies, power grid dispatch load, regional energy supply and demand status, and data from surrounding environmental monitoring stations, helps the system understand regional carbon emission policies and energy supply and demand, thereby providing decision support for enterprises' carbon emission management.

[0022] A1: Preprocessing includes spatiotemporal alignment of multi-frequency acquired data using a sliding window mechanism, wherein the reference window length is preferably 1 hour.

[0023] Understandably, the sliding window mechanism in this method is mainly used to solve the problem of inconsistent acquisition frequencies of multi-source heterogeneous data. The purpose is to unify data streams from all sources and at different frequencies into the same time coordinate system, thus providing the possibility for subsequent construction of consistent feature windows and model training.

[0024] It should be noted that setting a unified benchmark time window (preferably 1 hour is based on the real-time requirements of gas power plant production monitoring and industry practice) and using this window as a benchmark to resample all input data is to force all variables to be strictly aligned on the timestamp, and eliminate information asynchrony caused by misalignment of data collection time.

[0025] A2: For high-frequency data, calculate statistical values ​​within the window to generate reference frequency data points.

[0026] It should be noted that for high-frequency data such as energy consumption monitoring at the second or minute level, the system slides within a fixed one-hour window, calculating the average, maximum, or cumulative value within each window to generate stable hourly data points. For example, gas flow data occurring once per minute is calculated as the average flow or cumulative consumption over one hour.

[0027] A3: For low-frequency data, interpolation methods are used to extend it to the reference frequency.

[0028] It should be noted that for low-frequency data such as daily updated environmental quality indices and monthly updated equipment efficiency parameters, methods such as linear interpolation or forward imputation are used to extend the data to one data point per hour. For example, for data with only one AQI value at midnight each day, interpolation can be used to estimate the approximate AQI value for each hour of that day.

[0029] Furthermore, it should be noted that preprocessing also includes: performing normalization transformations for different variable types, retaining data transformation parameters for restoration processing during the prediction stage, ensuring the continuity and controllability of model input through real-time data quality scoring and dynamic data validity detection, and providing a local buffering strategy to support fault-tolerant operation in scenarios of brief data interruptions. Specifically, for any original variable... The standard normalization conversion process is as follows: In the formula: Represents the original variable value; Representing variables Standard deviation; This represents the standardized variable input value.

[0030] Furthermore, after the model predicts the output, in order to perform inverse normalization to recover the true physical quantities, the following is adopted: Maintaining a consistent input scale during operation facilitates the conversion of prediction results into actual carbon emissions or related physical indicators during post-processing.

[0031] Preferably, step S100 transforms multi-source heterogeneous data into a spatiotemporally aligned standardized dataset, ensuring that the features contained in each sample during model training are observations or estimates within the same time segment, which significantly improves the reliability and prediction accuracy of the subsequent model.

[0032] In one optional implementation, the multi-source heterogeneous data acquired in step S100 during the gas-fired power generation process can be collected in real time by edge computing nodes and transmitted hierarchically via message queues. Specifically, lightweight edge computing devices are deployed on the data source sides such as gas turbine control cabinets, fuel blending stations, and environmental monitoring cabins. Parameters such as unit load, fuel calorific value, and flue gas composition are subscribed to in real time via OPC UA protocol or Modbus TCP. After the edge nodes complete data cleaning and preliminary compression, second-level key parameters are sent to the production network message queue cluster via MQTT protocol according to priority. Minute-level statistical data is aggregated into the data platform cache area via Kafka channel. The entire process uses the national cryptographic SM4 algorithm for encrypted transmission and achieves one-way physical isolation based on industrial control firewalls. Finally, the data is uniformly extracted to the preprocessing module by the data consumption service.

[0033] In another optional implementation, the acquisition of multi-source heterogeneous data in the gas-fired power generation process in step S100 can also be achieved through a hybrid acquisition method combining incremental file synchronization and timed API retrieval. Specifically, for the historical database of the existing DCS system, the incremental data file is automatically generated at 5-minute intervals during off-peak scheduling periods using the CSV / XML export interface provided by the SCADA platform and pushed to the intermediate file server via the SFTP protocol. At the same time, for the component analysis data of the fuel supplier and the load command data of the power grid dispatch center, a RESTful API client is developed to timed retrieve JSON format data packets. After the data from both channels undergoes format verification and conflict detection, it is uniformly parsed and stored in the database by a batch ETL operation. The acquisition delay is controlled within 15 minutes to balance real-time performance and system stability.

[0034] In this embodiment of the application, step S200 involves constructing time-series statistical features based on the dataset using a sliding window across multiple time scales, and selecting a subset of core features based on feature importance evaluation, including the following steps B1-B3: B1: Set short-term, medium-term, and long-term windows, and calculate statistical indicators including mean, variance, maximum, minimum, skewness, kurtosis, and trend slope for the variables of the standardized dataset within each window to construct a feature set characterizing emission behavior at different time scales.

[0035] Understandably, this step aims to automatically extract features from aligned, standardized data that characterize carbon emission behavior patterns at different time granularities, providing highly informative input to the model. Specifically, the mean is calculated to reflect the trend level, variance to reflect the degree of fluctuation, maximum / minimum values ​​to reflect the extreme range, skewness / kurtosis to reflect the distribution pattern, and the trend slope to reflect the direction of change.

[0036] It should be noted that the short-term window is mainly used to capture rapid changes in the device, ranging from 1 to 6 hours or 5 to 20 sampling periods; the medium-term window is used to characterize intraday fluctuations and phased behavior, ranging from 6 to 24 hours or 20 to 100 sampling periods; and the long-term window is used to reflect periodic trends and cross-day changes, ranging from 1 to 7 days.

[0037] Preferably, this step constructs a high-dimensional, multi-resolution feature space, which enables the model to simultaneously learn short-term fluctuations, intraday cycles, and long-term trends in carbon emissions, greatly enhancing its ability to represent complex emission behaviors.

[0038] In one optional implementation, the feature set characterizing emission behavior at different time scales in step B1 can be constructed by variational mode decomposition and multi-scale entropy fusion. That is, variational mode decomposition is performed on the preprocessed core time-series data such as fuel flow and unit load to extract multiple intrinsic mode function components corresponding to hourly fluctuations, daily cycle adjustment, weekly scheduling patterns and seasonal fuel blending patterns. The multi-scale sample entropy of each component is calculated to quantify the complexity and randomness of emission behavior. At the same time, the mean, variance and trend slope of each component are statistically analyzed under three sliding windows of 15 minutes, 1 hour and 24 hours. Finally, the modal entropy value and window statistics are concatenated to form a multi-scale feature matrix that integrates frequency domain characteristics and time domain statistics.

[0039] In another optional implementation, the feature set representing emission behavior at different time scales in step B1 can be constructed using a dynamic graph neural network encoding method based on equipment, fuel, and operating conditions. Specifically, the gas turbine, waste heat boiler, and fuel blending device are abstracted as graph structure nodes. Operating parameters are extracted at 15-minute intervals to construct a temporal adjacency matrix to characterize the dynamic coupling relationship between heat flow, mass flow, and emission flow between nodes. A temporal graph attention network is used to encode the graph snapshot sequence at different time windows, automatically learning the hidden state vectors of each node at different scales. The cross-scale graph features representing system-level emission behavior are aggregated through pooling operations and used as input for subsequent models.

[0040] B2: Use at least one of the following methods to score the importance of each feature in the feature set: correlation coefficient analysis, rank correlation analysis, or mutual information calculation.

[0041] It should be noted that after step B1, the feature dimensions may expand dramatically, which may include a large number of redundant or irrelevant features. This step uses at least one of the following three methods: correlation coefficient analysis (to measure linear relationships), rank correlation analysis (such as Spearman's method, which measures monotonic relationships), and mutual information calculation (to measure any statistical dependency). This allows the system to comprehensively capture different types of relationships (linear, nonlinear, and non-monotonic) between features and target variables. For example, the mutual information method can discover those features that have complex nonlinear dependencies on carbon emissions but are very important (such as specific combinations of meteorological conditions). This avoids the limitations of a single evaluation method, thereby ensuring that key features are not missed and laying the feature foundation for building a high-performance prediction model.

[0042] In one optional implementation, the importance scoring of each feature in the feature set in step B2 can be achieved by incorporating Granger causality tests with physical constraints. Specifically, for candidate features such as fuel flow rate, gas turbine cylinder temperature, and load fluctuation rate, multivariate Granger causality tests are performed on carbon emission intensity at three lag orders: 15 minutes, 1 hour, and 24 hours, to obtain statistical significance p-values ​​and causality strength coefficients. Then, weights are assigned based on combustion chemical kinetics mechanisms—features directly related to carbon mass conservation, such as fuel carbon content and excess air coefficient, are assigned high confidence weights, while indirectly related features such as ambient temperature and humidity are assigned low weights. Finally, the top K-dimensional feature subsets that combine statistical causality and physical interpretability are selected by sorting the weighted comprehensive scores.

[0043] In another alternative implementation, the importance scoring of each feature in the feature set in step B2 can also be achieved through an adversarial temporal masking verification method. That is, a discriminator network is constructed to distinguish between normal carbon emission sequences and artificially disturbed sequences. Then, a random-length mask (simulating sensor disconnection or communication delay in an industrial field) is applied to each channel of the temporal features. The importance of the features is quantified by the degree of performance degradation of the discriminator. The more significant the performance degradation after masking, the higher the score of the feature. In this way, core features with strong robustness to missing data are selected.

[0044] B3: Based on cross-validation stability assessment and adaptive threshold, the core feature subset for final input model training is determined from the feature set.

[0045] It is important to emphasize that the adaptive threshold is not a pre-set fixed constant, but is determined jointly by the feature score distribution and the cross-validation results, which is equivalent to being determined autonomously during operation.

[0046] It should be noted that the system uses cross-validation stability assessment, which means running feature importance assessment multiple times on different subsets of the training data to obtain the importance score distribution of each feature and its fluctuation (stability) across different subsets. Then, instead of applying a preset formula to calculate the threshold, the system analyzes the overall shape and stability pattern of these score distributions and autonomously determines a reasonable dividing point (i.e., adaptive threshold) to filter out those features that not only score highly on most subsets but also have stable score performance, thus forming the final core feature subset.

[0047] In this embodiment of the application, step S300 involves training multiple prediction models in parallel based on a subset of core features, and dynamically triggering model retraining or switching based on model performance detection results to generate initial prediction results, including the following steps C1-C3: C1: Based on a subset of core features, multiple heterogeneous prediction models, including multiple linear regression, elastic network regression, and gradient boosting tree models, are trained in parallel.

[0048] Understandably, multiple linear regression models are used to capture linear global trends, elastic network regression models introduce regularization on the basis of linearity to prevent overfitting and automatically select features, and gradient boosting tree models such as XGBoost capture nonlinear interactions and complex local patterns.

[0049] Ideally, this step constructs a complementary set of predictors, providing a stable baseline through linear models and capturing complex relationships through tree models. Even if one type of model temporarily fails due to data characteristics, other models may still maintain good performance, enhancing the overall resilience of the system to data fluctuations and pattern changes.

[0050] C2: Monitor the prediction error of each model in real time. When the continuous prediction error exceeds the dynamic threshold set based on historical error statistics or a systematic shift in the prediction output is detected, it will automatically trigger incremental retraining of the corresponding model or switch the prediction path between different models.

[0051] It is understandable that incremental retraining of a model means updating the model with new data without discarding all historical knowledge, and switching prediction paths between different models means switching the prediction path to another model that is currently performing better.

[0052] It should be noted that each model supports independent training and evaluation, and can automatically select the activation path based on historical performance data. This automatic selection is achieved by evaluating each model's performance within a historical time window. Specifically: First, the system collects and calculates performance data for each model, including prediction error, stability, and volatility metrics. Then, weights are assigned to each model based on these metrics, with models exhibiting lower errors and stronger stability receiving higher weights. Next, the system selects the best-performing model within the current time window as the activation path based on the weighted results. When multiple models perform similarly, the system prioritizes the most suitable model based on data characteristics (such as nonlinearity and volatility). After each new data input, the system recalculates the historical performance of each model and dynamically adjusts the activation path, thereby achieving adaptive optimization for different data patterns. Furthermore, based on the characteristics of the prediction target, the system supports dynamically adjusting the model training data period, such as 7-day, 30-day, and 90-day sliding windows, to balance short-term responsiveness with long-term trend modeling.

[0053] Better yet, this ensures that the selection of activation paths is based on the model’s recent reliable overall performance, rather than accidental single-event performance.

[0054] It's important to further clarify that the threshold is determined based on the model's performance on historical data and the fluctuation range of prediction errors. Typically, the threshold can be set by calculating the average prediction error and standard deviation over a certain time window. When consecutive prediction errors exceed a certain multiple of the set standard deviation, the system will trigger a model retraining or switching mechanism. For example, assuming the average prediction error of historical data is 5% and the standard deviation is 2%, the threshold can be set to 5% + 2 × 2% = 9%. When consecutive prediction errors exceed 9%, the system will automatically trigger model retraining or switching to ensure the stability and reliability of the prediction results.

[0055] C3: After each model training or retraining, generate a model interpretation report including SHAP value plots and feature impact factor rankings.

[0056] It should be noted that after each (re)training session, the system not only outputs model parameters but also automatically generates a visual explanatory report. The SHAP value plot quantitatively displays the magnitude and direction of each feature's positive or negative contribution to a single prediction, while the feature influence factor ranking shows which features have the greatest global impact on the model output across the entire dataset. This not only achieves prediction but also provides diagnostic insights, upgrading the system from a simple prediction tool to an intelligent analysis platform that assists decision-making, thus enhancing the technology's practical value and acceptance.

[0057] Preferably, step S300 can not only independently model and mutually verify based on multiple methods, but also flexibly select the optimal prediction path according to data characteristics and scenario requirements, thus having higher accuracy and adaptability when facing seasonal, nonlinear and highly abrupt emission behaviors.

[0058] In this embodiment of the application, step S400 automatically optimizes the hyperparameter configuration of each model based on the error index feedback of the initial prediction results, generating an optimized prediction model, including the following steps D1-D2: D1: A two-stage hybrid optimization strategy is adopted. Based on the error index of the initial prediction results, the hyperparameters of each model are iteratively tuned. In the first stage, the Bayesian optimization algorithm is used for global exploration to quickly locate the approximate optimal range of key hyperparameters. In the second stage, within the approximate optimal range, a grid search strategy is used for local fine search to determine the final hyperparameter configuration.

[0059] Understandably, error metrics include RMSE, MAE, MAPE, etc.

[0060] It should be noted that in the first stage, the system uses the error index (such as RMSE) of the initial prediction results as the optimization objective. A probabilistic surrogate model of the objective function is constructed using a Bayesian optimization algorithm. By continuously selecting the most promising hyperparameter combinations for evaluation, the system quickly identifies the approximate optimal range of key hyperparameters (such as the number of ensemble trees n_estimators and the learning rate) with as few evaluations as possible. This avoids blind and intensive traversal of the global parameter space. Furthermore, in the second stage, within the approximate optimal interval identified in the first stage, a grid search strategy is employed. Since the search range has been significantly narrowed, a relatively intensive but computationally controlled traversal can be performed to fine-tune hyperparameters (such as tree depth max_depth and subsample rate) to find the optimal configuration within this local region.

[0061] The best approach is to use a coarse-to-fine hybrid strategy, which achieves the best balance between computational efficiency and optimization accuracy. This avoids the waste of resources in a wide space by pure grid search and overcomes the problem of slow convergence or potential miss of some flat optimal regions in the later stages of pure Bayesian optimization.

[0062] D2: The hyperparameters to be optimized include at least the training period, learning rate, and regularization strength. After optimization, a performance-improved, tuned prediction model is generated.

[0063] It should be noted that the system focuses on optimizing the core hyperparameters of each prediction model. For example, the XGBoost model is shown in Table 1 below: Table 1 XGBoost Model Tuning Strategies It should be noted that by automatically optimizing these key parameters, the system can ensure that each basic prediction model consistently maintains or approaches its optimal performance. The table below compares the prediction errors of typical gas-fired power plants after model tuning: Table 2. Comparison of prediction errors in typical gas-fired power plants after model optimization. As shown in Table 2, after optimization, the RMSE of the elastic regression model decreased from 49.92 (the baseline of the linear regression model before optimization) to 48.24, and the MAE decreased from 38.90 to 35.73, indicating a significant reduction in model prediction error. Furthermore, the elastic regression model performed best in terms of error metrics (RMSE and MAE), while XGBoost showed the best performance in terms of percentage error (MAPE) and prediction accuracy. Both models have their advantages, with an R² of 0.99, demonstrating their strong interpretability of gas and electricity data and excellent overall performance after optimization.

[0064] In this embodiment of the application, step S500 involves a multi-index comprehensive evaluation of the initial prediction results and the optimized prediction model, and outputting the target prediction results through a weighted integration or rule selection mechanism, including the following steps E1-E3: E1: Based on RMSE, MAE, R² and MAPE metrics, the predictive performance of each model within the sliding time window is comprehensively evaluated, and the dynamic fusion weights of each model are automatically calculated based on the evaluation results. A time decay function is introduced to reduce the impact of long-term historical samples on the current weight calculation.

[0065] It should be noted that the evaluation is not based on the entire history, but rather limited to a sliding time window (e.g., the most recent 24 hours or 7 days). More importantly, when calculating the overall score or weights, a time decay function (such as exponential decay) is introduced, ensuring that newer prediction errors have a greater impact on the current weight within the window, while older errors have a smaller impact. Specifically, when a model consistently performs well recently, its weight will increase rapidly; conversely, if its performance begins to decline, its influence will be quickly weakened. This ensures that the system always tends to trust the most reliable model at present.

[0066] E2: It adopts a hybrid selection framework that combines rule-based decision-making with data-driven learning. It prioritizes the determination and output of the optimal single-model prediction result based on the current data characteristics and the historical performance of each model. When multiple models are equally competitive, leading to decision conflicts, a soft decision-making mechanism based on a stacked ensemble model is activated to generate the final target prediction result.

[0067] It should be noted that the system first runs a rule-based rapid decision. For example, the rules might include: prioritizing the model type that has historically performed best under the current data characteristics (such as high nonlinearity or high volatility); or directly selecting the model with the highest evaluation weight in the recent period (within the sliding window). Furthermore, when the rule engine finds multiple models with similar competitiveness (e.g., very similar weights, or each leading on different metrics) and cannot make a clear decision, a stacked ensemble model is activated as an arbitrator. This model uses the prediction results of each candidate model as input features for secondary learning (i.e., soft decision), ultimately outputting a result that compromises or surpasses any single model.

[0068] Preferably, the hybrid framework of rule-based initial screening and machine learning final judgment adopted in step S500 combines the high transparency and low latency of the rule system with the high accuracy and strong fitting ability of the machine learning model. This further ensures that the final output is the optimal or near-optimal prediction under the current conditions, whether in a simple scenario or a complex conflict scenario, greatly improving the intelligence level and reliability of the system's decision-making.

[0069] E3: By performing trend line fitting, volatility analysis and mutation detection on continuous prediction results, it can identify abnormal emission behavior, such as sudden increases and decreases, and output preliminary cause-finding suggestions in combination with actual equipment operation records.

[0070] It should be noted that extreme value search and segmented slope analysis algorithms are embedded in the long-term prediction sequence to automatically determine whether the carbon emission level is approaching the set peak threshold, and to provide trend scores, peak prediction time points, and reliability assessment results. Furthermore, the system outputs an explanation report for each prediction result, using SHAP values ​​and local linear interpretation model technology to label the main influencing factors. It supports viewing the feature impact ranking and emission contribution rate through a graphical interface, and also provides an open strategy suggestion interface for managers to query the response solution library.

[0071] Furthermore, the method also includes deployment via a microservice architecture, supporting operation on local servers, edge gateways, or cloud platforms. Built on containerization technology, it supports integration with Linux, Windows, and mainstream industrial control systems, and possesses lightweight deployment capabilities. It supports online model retraining, offline model import, and horizontal migration, enabling rapid deployment of trained models to other gas power companies, improving system reusability and adaptability. It provides RESTful API interfaces, supporting integration with third-party energy management platforms, environmental monitoring platforms, and enterprise MES systems to achieve data interaction and prediction result feedback. It also features a built-in role-based access control mechanism, supporting multi-user hierarchical management, encrypted transmission of sensitive data, and audit log tracking, ensuring the system's secure and compliant operation in an enterprise-level environment. Furthermore, the system provides a visual operation interface and data interface, and based on user feedback (when users have questions or suggestions for correction regarding a prediction result or explanation report, they can submit manual annotations, and the system records the feedback data and contextual information), it uses incremental learning methods to correct the prediction model in real time. In particular, by constructing visualization tools such as error distribution evolution curves, accuracy time trend charts, and model comparison radar charts, it helps users understand the system's long-term prediction capabilities and identify model degradation trends. Based on feedback data, model performance, and changes in business needs during long-term use, it automatically adjusts feature selection thresholds, training window lengths, or enables model architecture update suggestions to improve the system's adaptability to environmental changes and data drift.

[0072] In summary, the beneficial effects of this multi-dimensional carbon emission prediction method are as follows: by standardizing and aligning multi-source heterogeneous data through preprocessing and time series alignment, it effectively solves the shortcomings of traditional methods such as single data dimension and lagging updates; by using a multi-timescale sliding window to extract dynamic statistical features and combining importance assessment to select a subset of core features, it significantly enhances the interpretability and anti-interference ability of input variables; through parallel training mechanism and data-driven multi-class prediction models, combined with dynamic weight allocation and hyperparameter adaptive optimization mechanism, it achieves accuracy improvement and stable switching under different operating scenarios; and by integrating and evaluating the prediction results based on sliding error window weighted fusion and trend change identification algorithms, it can provide real-time early warning of emission anomalies and preliminarily identify the causes, thereby improving prediction transparency and decision-making confidence.

[0073] Example 3 illustrates a multi-dimensional carbon emission prediction method. It should be noted that the technical solution of this multi-dimensional carbon emission prediction system and the technical solution of the aforementioned multi-dimensional carbon emission prediction method belong to the same concept. Details not described in detail in this example can be found in the description of the technical solution of the aforementioned multi-dimensional carbon emission prediction method.

[0074] This embodiment also provides a multi-dimensional carbon emission prediction system, including: The acquisition module is used to acquire multi-source heterogeneous data in the gas-fired power generation process and generate a standardized dataset after preprocessing. The filtering module is used to construct time-series statistical features based on the dataset through a sliding window at multiple time scales, and to filter out a subset of core features based on feature importance evaluation. The first generation module is used to train multiple prediction models in parallel based on a subset of core features, and dynamically trigger model retraining or switching based on model performance detection results to generate initial prediction results. The second generation module is used to automatically optimize the hyperparameter configuration of each model based on the error index feedback of the initial prediction results, and generate the optimized prediction model. The output module is used to comprehensively evaluate the initial prediction results and the optimized prediction model using multiple indicators, and outputs the target prediction results through weighted integration or rule selection mechanisms.

[0075] This embodiment also provides an electronic device suitable for high-temperature multidimensional carbon emission prediction, including: a memory and a processor; the memory is used to store computer-executable instructions, and the processor is used to execute the computer-executable instructions to implement the multidimensional carbon emission prediction method proposed in the above embodiment.

[0076] This embodiment also provides a storage medium on which a computer program is stored. When the program is executed by a processor, it implements the multi-dimensional carbon emission prediction method proposed in the above embodiments.

[0077] The storage medium proposed in this embodiment belongs to the same inventive concept as the multi-dimensional carbon emission prediction method proposed in the above embodiments. Technical details not described in detail in this embodiment can be found in the above embodiments, and this embodiment has the same beneficial effects as the above embodiments.

[0078] Based on the above description of the implementation methods, those skilled in the art can clearly understand that the present invention can be implemented using software and necessary general-purpose hardware, and of course, it can also be implemented using hardware. Based on this understanding, the technical solution of the present invention, or the part that contributes to the prior art, can be embodied in the form of a software product. This computer software product can be stored in a computer-readable storage medium, such as a computer floppy disk, read-only memory (ROM), random access memory (RAM), flash memory, hard disk, or optical disk, etc., including several instructions to cause a computer device (which may be a personal computer, server, or network device, etc.) to execute the methods of the various embodiments of the present invention.

[0079] 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 preferred embodiments, those skilled in the art should understand that modifications or equivalent substitutions can be made to the technical solutions of the present invention without departing from the spirit and scope of the technical solutions of the present invention, and all such modifications or substitutions should be covered within the scope of the claims of the present invention.

Claims

1. A multi-dimensional carbon emission prediction method, characterized in that: include, Acquire multi-source heterogeneous data from the gas-fired power generation process, and generate a standardized dataset after preprocessing; Based on the dataset, time-series statistical features are constructed using a sliding window with multiple time scales, and a subset of core features is selected based on feature importance evaluation. Multiple prediction models are trained in parallel based on the core feature subset, and model retraining or switching is dynamically triggered based on model performance detection results to generate initial prediction results. Based on the error index feedback of the initial prediction results, the hyperparameter configuration of each model is automatically optimized to generate the optimized prediction model. The initial prediction results and the optimized prediction model are evaluated using multiple indicators, and the target prediction results are output through weighted integration or rule selection mechanisms.

2. The multi-dimensional carbon emission prediction method as described in claim 1, characterized in that: The method is deployed using a microservice architecture and supports operation on local servers, edge gateways, or cloud platforms. The system provides a visual user interface and data interface, and uses incremental learning to correct the prediction model in real time based on user feedback.

3. The multi-dimensional carbon emission prediction method as described in claim 1, characterized in that: The preprocessing includes spatiotemporal alignment of multi-frequency acquired data using a sliding window mechanism; For high-frequency data, statistical values ​​are calculated within a window to generate reference frequency data points; For low-frequency data, an interpolation method is used to extend it to the reference frequency.

4. A multi-dimensional carbon emission prediction method as described in claim 1 or 3, characterized in that: The core feature subset is selected based on feature importance assessment. include, The importance of each feature in the feature set is scored using at least one of the following methods: correlation coefficient analysis, rank correlation analysis, or mutual information calculation. Based on cross-validation stability assessment and adaptive threshold, the core feature subset for final input model training is determined from the feature set.

5. The multi-dimensional carbon emission prediction method as described in claim 4, characterized in that: The model performance detection logic includes, When the continuous prediction error exceeds the threshold calculated based on the historical error mean and standard deviation, the model retraining or switching mechanism is automatically triggered. Each time a model is trained, a feature analysis report is generated, including a SHAP value plot and an impact factor ranking plot.

6. The multi-dimensional carbon emission prediction method as described in claim 5, characterized in that: The automatic optimization of hyperparameter configuration for each model employs a hybrid strategy of Bayesian optimization and grid search. The optimized hyperparameters include training period, learning rate, and regularization strength, which are optimized using a two-stage parameter tuning method.

7. The multi-dimensional carbon emission prediction method as described in claim 3, characterized in that: The baseline window length is 1 hour.

8. A multi-dimensional carbon emission prediction system, employing the method described in any one of claims 1-7, characterized in that, include: The acquisition module is used to acquire multi-source heterogeneous data in the gas-fired power generation process and generate a standardized dataset after preprocessing. The filtering module is used to construct time-series statistical features based on the dataset through a sliding window with multiple time scales, and to filter out a subset of core features based on feature importance evaluation. The first generation module is used to train multiple prediction models in parallel based on the core feature subset, and dynamically trigger model retraining or switching based on model performance detection results to generate initial prediction results. The second generation module is used to automatically optimize the hyperparameter configuration of each model based on the error index feedback of the initial prediction results, and generate the optimized prediction model. The output module is used to comprehensively evaluate the initial prediction results and the optimized prediction model using multiple indicators, and output the target prediction results through weighted integration or rule selection mechanisms.

9. A computer device comprising a memory and a processor, wherein the memory stores a computer program, characterized in that, When the processor executes the computer program, it implements the steps of the method according to any one of claims 1 to 7.

10. A computer-readable storage medium having a computer program stored thereon, characterized in that, When the computer program is executed by a processor, it implements the steps of the method according to any one of claims 1 to 7.

Images