Biomass gas coupling unit optimization control method and system based on digital twinning

Abstract

The present application relates to the technical field of biomass gas utilization, and particularly relates to a biomass gas coupling unit optimization control method and system based on digital twinning, the method comprising: obtaining operation data of a biomass gas coupling gas turbine unit after entering a stable operation interval in a historical operation process, obtaining an original data set, based on the operation data, screening the collected original data set, eliminating data interference in start-stop and load mutation stages, forming an effective operation data set, based on the effective operation data set, constructing effective response training data, forming a training data set, and training a gas supply flow machine learning model. The present application selects only operation data of the gas turbine unit after entering the stable operation interval, and introduces a synchronous collection mechanism of the speed maintaining state and the output power, so that the formed data set can truly reflect actual response characteristics of the gas turbine to the gas quality and the gas supply capacity under the continuous stable working condition.

Description

Technical Field

[0001] This invention relates to the field of biomass gas utilization technology, specifically to an optimized control method and system for biomass gas coupled units based on digital twins. Background Technology

[0002] As the proportion of biomass energy in the renewable energy system continues to increase, biomass gasification power generation and its coupling operation with gas turbine units are gradually becoming important ways to achieve clean and efficient energy conversion. By introducing the combustible gas generated by the biomass gasification system into the gas turbine unit for energy utilization, the utilization efficiency of biomass resources can be improved to a certain extent, and dependence on fossil fuels can be reduced. However, in practical engineering applications, the biomass gasification process itself is significantly affected by differences in raw material composition, fluctuations in moisture content, and changes in reaction conditions, resulting in poor calorific value and component stability of the generated combustible gas, thus posing challenges to the safe and stable operation of the subsequent gas turbine unit.

[0003] In existing technologies, the control technology of biomass gas coupled gas turbine units mostly adopts open-loop or weak closed-loop control methods based on empirical thresholds or single operating parameters. It usually takes the output power of the gas turbine as the core regulation object, lacks modeling of the inherent coupling relationship between biomass gas quality fluctuations and gas turbine operation stability, and the training data often mixes data from start-up and shutdown stages and load change stages, resulting in insufficient model generalization ability and engineering applicability. Summary of the Invention

[0004] To address the aforementioned problems, this invention provides an optimized control method for biomass gas coupling units based on digital twins.

[0005] The present invention adopts the following technical solution: a biomass gas coupled unit optimization control method based on digital twins, which obtains the operating data of the biomass gas coupled gas turbine unit after it enters the stable operating range during the historical operation process, and obtains the original dataset. The original dataset includes the gas turbine output power, gas turbine speed, gas supply flow rate, biomass gas state parameters and biomass gas flow rate.

[0006] Based on operational data, the collected raw dataset is filtered to remove data interference from the start-up and shutdown phases and the load change phases, forming an effective operational dataset;

[0007] The gas turbine output power, biomass gas state parameters, biomass gas flow rate and gas supply flow rate in the effective operating dataset are used to construct the effective response training data, forming a training dataset, and a machine learning model for predicting gas supply flow rate is trained.

[0008] The system acquires the current output power of the gas turbine, biomass gas state parameters, and biomass gas flow rate. Based on the trained machine learning model, it predicts the gas supply flow rate and controls the gas supply flow rate of the gas turbine unit.

[0009] As a further description of the above technical solution: the biomass gas coupling unit optimization control method based on digital twin also includes:

[0010] Under experimental conditions, the stability risk index of the gas turbine was collected under different biomass gas state parameters and participation ratios, and a machine learning model for judging stable state was trained based on threshold labeling.

[0011] The current biomass gas state parameters, biomass gas flow rate, and predicted gas supply flow rate are input into the steady state discrimination machine learning model to obtain the discrimination result of whether the current biomass gas participation ratio meets the requirements for stable operation.

[0012] When the judgment result is not up to standard, the biomass gas participation ratio is searched in reverse along the decreasing direction, and the steady state discrimination machine learning model is used to judge the results of each participation ratio to obtain the maximum biomass gas participation ratio when the judgment result meets the standard, and the corresponding ideal biomass gas flow rate is obtained.

[0013] Based on the ideal biomass gas flow rate, biomass gas state parameters, and the current gas turbine output power, a machine learning model for predicting gas supply flow rate is invoked to obtain the gas supply flow rate prediction result under stability constraints.

[0014] The predicted gas supply flow rate and the corresponding biomass gas flow rate are combined and applied to the gas turbine unit.

[0015] As a further description of the above technical solution: the methods for filtering the collected raw dataset to form a valid operational dataset include:

[0016] The preset time window length is used to calculate the rate of change of the gas turbine output power within a continuous time window. The output power stability criterion is set as follows: when the rate of change of output power is lower than a preset threshold, the output power is determined to be in a stable state within the time window; when the rate of change of output power exceeds the preset threshold, it is determined to be a load adjustment or disturbance stage, and the corresponding data is not included in the stable operation range.

[0017] Within a preset time window, the speed maintenance coefficient of the gas engine is calculated, and the speed maintenance stability criterion is set as follows: when the speed maintenance coefficient is less than the preset speed maintenance coefficient threshold, the gas engine is determined to be in a stable maintenance state.

[0018] The above-mentioned output power stability criterion and speed maintenance stability criterion are jointly judged. Only when both criteria are met simultaneously within the same time window are the data corresponding to that time window determined as valid operating data.

[0019] As a further description of the above technical solution: the method for obtaining the speed holding coefficient includes: obtaining the deviation between the time average of the actual speed of the gas engine within a preset time window and the target speed set by the gas engine, obtaining the absolute value of the deviation, obtaining the standard deviation of the speed of the gas engine within the preset time window, obtaining the ratio of the absolute value of the deviation to the target speed set by the gas engine and the ratio of the standard deviation to the target speed set by the gas engine, and then weighted summing the two ratios to obtain the speed holding coefficient.

[0020] As a further description of the above technical solution: the effective response training data includes the gas turbine output power, biomass gas state parameters, biomass gas flow rate and the corresponding gas supply flow rate.

[0021] As a further description of the above technical solution: the biomass gas state parameters include the effective calorific value coefficient and the proportion of gas components;

[0022] The method for collecting the biomass gas state parameters includes:

[0023] At the gas outlet of the biomass gasification system, online data is collected on the gasification product before it enters the gas turbine unit. The collected data includes gas flow rate and volume fractions of carbon monoxide, hydrogen, and methane, as well as volume fractions of carbon dioxide and nitrogen, forming complete data on the proportion of gas components.

[0024] Based on the gas component proportion data, the effective calorific value coefficient of the gas is obtained by weighting and summing the standard lower heating value and volume fraction of each combustible component.

[0025] As a further description of the above technical solution: the training method of the machine learning model for predicting gas supply flow includes:

[0026] The gas turbine operating conditions are stratified, and stratified random sampling is adopted. The effective operating dataset is divided into training set, validation set and test set according to a preset ratio, so that the distribution of each operating condition is consistent in the training set, validation set and test set.

[0027] Gradient boosting regression tree was selected as the gas supply flow prediction model. Initial hyperparameters were set and the model was initialized. The initial predicted value of the gradient boosting regression tree model was set as the mean of the gas supply flow in the training set. For each sample in the training set, the difference between the gas supply flow predicted by the current model and the actual flow, i.e., the residual, was calculated. A new decision tree was constructed. With the goal of minimizing the residual, the optimal splitting feature and splitting point were selected based on the mean square error. The samples were divided into different child nodes until the preset stopping condition was met.

[0028] The Bayesian optimization method is used to search for the optimal combination of hyperparameters within a preset range;

[0029] For every 20 trees in each iteration, the root mean square error of the validation set is calculated. When the root mean square error of the validation set decreases by less than 0.001 for three consecutive iterations, training is stopped, and the model parameters with the smallest root mean square error of the validation set during training are saved.

[0030] As a further description of the above technical solution: when the judgment result meets the standard, the gas supply in the gas unit is controlled based on the predicted gas supply flow rate.

[0031] As a further description of the above technical solution: the method for obtaining the stability risk index includes:

[0032] Within a preset sliding time window, statistical analysis is performed on the gas engine speed data to calculate the characteristic quantity of speed fluctuation intensity.

[0033] Within a preset sliding time window, acquire time-varying data of combustion temperature and calculate the characteristic quantity of combustion response change rate;

[0034] After normalizing the characteristic quantities of speed fluctuation intensity and combustion response change rate, the stability risk index is obtained by weighted summation.

[0035] As a further description of the above technical solution: the threshold-based labeling method includes: when the stability risk index of the gas turbine unit is lower than the stability risk index threshold, it is set to meet the biomass gas participation ratio standard, and the label is set to 1;

[0036] When the stability risk index of the gas turbine unit is greater than or equal to the stability risk index threshold, it is set as a non-compliance of biomass gas participation ratio and the label is set to 0.

[0037] A digital twin-based optimization control system for biomass gas coupled units, used to implement the aforementioned digital twin-based optimization control method for biomass gas coupled units, includes:

[0038] The raw operating data acquisition module acquires the operating data of the biomass gas coupled gas turbine unit after it enters the stable operating range during the historical operation process, and obtains the raw dataset. The raw dataset includes the gas turbine output power, gas turbine speed, gas supply flow rate, biomass gas state parameters, and biomass gas flow rate.

[0039] The effective operation data generation module filters the collected raw dataset based on the operation data, eliminating data interference from the start-up and shutdown phases and the load change phases, and forming an effective operation dataset.

[0040] The training data construction module constructs the gas turbine output power, biomass gas state parameters, biomass gas flow rate and gas supply flow rate from the effective running dataset into effective response training data, forming a training dataset, and training a machine learning model to predict gas supply flow rate.

[0041] The gas supply flow prediction module obtains the current gas turbine output power, biomass gas state parameters and biomass gas flow, and predicts the gas supply flow based on the trained machine learning model to control the gas supply flow of the gas turbine unit.

[0042] The model training module collects the stability risk index of the gas turbine unit under experimental conditions, with different biomass gas state parameters and different biomass gas participation ratios. Based on the preset stability risk index threshold, it constructs the labeling rules for whether the biomass gas participation ratio meets or does not meet the standards, and trains a stability state discrimination machine learning model for judging the stability state of the gas turbine unit.

[0043] The stability discrimination module inputs the current biomass gas state parameters, biomass gas flow rate and predicted gas supply flow rate into the stability discrimination machine learning model to obtain the discrimination result of whether the current biomass gas participation ratio meets the requirements for stable operation.

[0044] The reverse search module performs a reverse search along the decreasing direction of the biomass gas participation ratio when the judgment result is not up to standard. It also judges the results of each participation ratio through the steady state discrimination machine learning model to obtain the maximum biomass gas participation ratio when the judgment result meets the standard, and obtains the biomass gas flow rate corresponding to the current maximum biomass gas participation ratio, which is marked as the ideal biomass gas flow rate.

[0045] The gas supply prediction module, based on the ideal biomass gas flow rate, biomass gas state parameters, and the current gas turbine output power, calls the predictive gas supply flow rate machine learning model to obtain the gas supply flow rate prediction result under stability constraints.

[0046] The control and execution module applies the predicted gas supply flow rate and the corresponding biomass gas flow rate to the gas turbine unit.

[0047] Beneficial effects:

[0048] In the above technical solution, the biomass gas coupled unit optimization control method based on digital twin provided by the present invention selects only the operating data after the gas turbine unit enters the stable operating range, and introduces a synchronous acquisition mechanism for speed maintenance and output power. This enables the resulting dataset to truly reflect the actual response characteristics of the gas turbine to gas quality and gas supply capacity under continuous stable operating conditions, avoids the masking of data patterns by start-up and shutdown processes and transient disturbances, and improves the representativeness and engineering effectiveness of the data from the source.

[0049] Furthermore, by constructing a machine learning model for judging the stable state of the gas turbine unit based on experimental data, when the proportion of biomass gas participation does not meet the requirements for stable operation, the model no longer adopts empirical or fixed proportion adjustment methods. Instead, it performs a reverse search based on the stability judgment results, moving in the direction of decreasing biomass gas participation. This avoids the risk of backfire, combustion oscillation, or flameout caused by continuing to increase the proportion of biomass gas under unstable operating conditions, and significantly improves the operational safety and reliability of the gas turbine unit under biomass gas blending conditions. Attached Figure Description

[0050] The present invention will be further explained below with reference to the accompanying drawings and embodiments:

[0051] Figure 1 A flowchart of the optimized control method for a biomass gas coupling unit based on digital twin provided in Embodiment 1 of the present invention;

[0052] Figure 2 The flowchart of the method for filtering the collected raw dataset to form an effective running dataset is provided in Embodiment 1 of the present invention.

[0053] Figure 3 A flowchart illustrating the method for obtaining the stability risk index provided in Embodiment 1 of the present invention;

[0054] Figure 4 This is a module connection diagram of the optimized control system for a biomass gas coupling unit based on digital twins provided in Embodiment 2 of the present invention. Detailed Implementation

[0055] To make the technical means, creative features, objectives, and effects of this invention readily understandable, the invention is further described below with reference to specific illustrations. It should be noted that, unless otherwise specified, the embodiments and features described in these embodiments can be combined with each other.

[0056] Example 1

[0057] Please see Figures 1-2This invention provides a technical solution: a biomass gas coupling unit optimization control method based on digital twins, comprising:

[0058] The operation data of the biomass gas coupled gas turbine unit after it entered the stable operation range during the historical operation process is obtained to obtain the original dataset. The original dataset includes the gas turbine output power, gas turbine speed, gas supply flow rate, biomass gas state parameters and biomass gas flow rate.

[0059] Based on the operational data, the collected raw dataset is filtered to remove data interference from the start-up and shutdown phases and the load change phases, forming an effective operational dataset for subsequent analysis.

[0060] Methods for filtering the collected raw datasets to form a valid running dataset for subsequent analysis include:

[0061] The preset time window length is used to calculate the rate of change of the gas turbine output power within a continuous time window. The output power stability criterion is set as follows: when the rate of change of output power is lower than a preset threshold, the output power is determined to be in a stable state within the time window; when the rate of change of output power exceeds the preset threshold, it is determined to be a load adjustment or disturbance stage, and the corresponding data is not included in the stable operation range.

[0062] Within a preset time window, the speed maintenance coefficient of the gas engine is calculated, and the speed maintenance stability criterion is set as follows: when the speed maintenance coefficient is less than the preset speed maintenance coefficient threshold, the gas engine is determined to be in a stable maintenance state.

[0063] The above-mentioned output power stability criterion and speed maintenance stability criterion are jointly judged. Only when both criteria are met simultaneously within the same time window are the data corresponding to that time window determined as valid operating data.

[0064] It should be noted that a time window length of 60 seconds and a period of 1 second are preferred.

[0065] The method for obtaining the speed maintenance coefficient includes: obtaining the deviation between the average time value of the actual speed of the gas engine within a preset time window and the target speed set by the gas engine, obtaining the absolute value of the deviation, obtaining the standard deviation of the speed of the gas engine within the preset time window, obtaining the ratio of the absolute value of the deviation to the target speed set by the gas engine and the ratio of the standard deviation to the target speed set by the gas engine, and then weighted summing the two ratios to obtain the speed maintenance coefficient.

[0066] Optionally, the formula for calculating the speed holding coefficient is:

[0067] ;

[0068] In the formula, The speed maintenance coefficient is used to comprehensively characterize the difficulty of maintaining the target speed stability of a gas engine under current gas supply conditions. The larger the value, the more difficult it is to maintain the speed. The target speed set for the gas engine, This represents the time average of the actual engine speed within a preset time window. The standard deviation of the gas engine speed within a preset time window. , These are the weighting coefficients.

[0069] It should be noted that the formulas mentioned above are all numerical calculations after removing dimensions. They are formulas that are closest to the real situation, obtained by software simulation based on a large amount of data. The weight coefficients in the formulas and the preset thresholds in the analysis process are set by those skilled in the art based on the actual situation or obtained by simulation based on a large amount of data.

[0070] The gas turbine output power, biomass gas state parameters, biomass gas flow rate and gas supply flow rate in the effective operating dataset are used to construct the effective response training data, forming a training dataset, and a machine learning model for predicting gas supply flow rate is trained.

[0071] The effective response training data includes gas turbine output power, biomass gas state parameters, biomass gas flow rate and corresponding gas supply flow rate;

[0072] The training method for the machine learning model for predicting gas supply flow includes:

[0073] The gas turbine is stratified according to its operating conditions, such as low power, medium power, and high power. Stratified random sampling is used to divide the effective operating dataset into training set, validation set, and test set according to a preset ratio (7:2:1), so that the distribution of each operating condition is consistent in the training set, validation set, and test set.

[0074] Gradient boosting regression tree was selected as the gas supply flow prediction model.

[0075] Set the initial hyperparameters, including: number of decision trees: 120-180; maximum depth of a single tree: 5-7; minimum number of samples for node splitting: 10-15; maximum number of features considered during splitting: 2 / 3 of the total number of features; learning rate: 0.05-0.1; loss function: mean squared error; regularization coefficient (L2): 0.1-0.2; single tree splitting criterion: mean squared error.

[0076] Model initialization involves setting the initial predicted values ​​of the gradient boosting regression tree model to the mean of the gas supply flow rate in the training set, providing a benchmark for subsequent iterative training. For each sample in the training set, the difference between the current model's predicted gas supply flow rate and the actual flow rate, i.e., the residual, is calculated. A new decision tree is constructed with the goal of minimizing the residual. Based on the mean squared error, the optimal splitting feature and splitting point are selected to divide the samples into different child nodes until a preset stopping condition is met, such as reaching the maximum depth or the number of samples in the child node being less than the minimum number of samples. The above residual calculation, fitting, and model update steps are repeated until a preset number of decision trees are constructed or an early stopping mechanism is triggered.

[0077] A Bayesian optimization method (more efficient than grid search and random search) is used to search for the optimal combination of hyperparameters within a preset range. Optimization range: Number of decision trees: 100-200; Maximum depth of a single tree: 4-8; Learning rate: 0.03-0.12; Minimum number of samples for node splitting: 8-20; Regularization coefficient (L2): 0.05-0.25.

[0078] Every 20 trees are trained in each iteration, the root mean square error of the validation set is calculated. When the root mean square error of the validation set decreases by less than 0.001 for three consecutive iterations, training is stopped to avoid the model overfitting the specific conditions of the training set.

[0079] Save the model parameters (including the splitting rules of all decision trees and the weights of leaf nodes) that minimize the root mean square error of the validation set during training.

[0080] The model is tested using a test set. If the root mean square error of the test set is less than or equal to 5% of the rated gas supply flow rate and the mean absolute error of the test set is less than or equal to 3% of the rated gas supply flow rate, the model is considered to be up to standard. The trained model is then exported in a lightweight format (such as pkl or ONNX) and integrated into the gas engine control system for use.

[0081] For example, a specific example of model training is as follows: The specific format of a single valid response training data is shown in Table 1:

[0082] Table 1 Effective Response Training Data

[0083]

[0084] For example, the stratified sampling by operating condition specifically includes: operating condition classification rules, which divide the operating condition levels according to the output power of the gas engine: low power condition: P < 350kW; medium power condition: 350 ≤ P < 550kW; high power condition: P ≥ 550kW;

[0085] The dataset contains 12,000 valid records, distributed as follows: 4,000 low-power records, 5,000 medium-power records, and 3,000 high-power records.

[0086] Stratified random sampling is performed in a 7:2:1 ratio. Taking the low-power condition as an example: training set: 2800 records, validation set: 800 records, test set: 400 records. The medium-power and high-power conditions are sampled in the same proportion to ensure that the proportion of the three types of conditions in each data subset is consistent.

[0087] For example, the model training process is as follows: First, the model is initialized: the mean of the gas supply flow rate in the training set is... The initial predicted value of the gradient boosting regression tree model was uniformly set to 128 Nm³ / h; in the first iteration, taking sample S3 as an example, the actual gas supply flow rate was... Initial predicted value The residual is 135-128=7. The model constructs the first decision tree based on the residuals of all samples, with the first layer based on the effective calorific value coefficient. ≤4.9 splits; the second layer is based on biomass gas flow rate Splitting rate ≤500 Nm³ / h; the final residual correction value given at the corresponding leaf node is +6.5 Nm³ / h;

[0088] The updated prediction value for the first tree is: When the learning rate When =0.08, ;

[0089] Repeat the above process, gradually adding 150 decision trees until the early stopping condition is met. During the Bayesian optimization process, a set of optimal hyperparameters are: 156 decision trees, maximum depth of 6, learning rate of 0.07, minimum number of samples per node of 12, and L2 regularization coefficient of 0.15. During the training process, the validation set RMSE is calculated every 20 trees added. The calculation results are shown in Table 2. Table 2 is the validation set RMSE calculation result table.

[0090] Table 2 RMSE Calculation Results for the Validation Set

[0091]

[0092] If the RMSE decreases by less than 0.001 for three consecutive times, an early stop is triggered, and the model corresponding to 140 trees is ultimately retained.

[0093] The system acquires the current gas turbine output power, biomass gas state parameters, and biomass gas flow rate. Based on the trained machine learning model, it predicts the gas supply flow rate and controls the gas supply flow rate of the gas turbine unit.

[0094] The biomass gas state parameters include the effective calorific value coefficient and the proportion of gas components.

[0095] The method for collecting the biomass gas state parameters includes:

[0096] At the gas outlet of the biomass gasification system, online data is collected on the gasification product before it enters the gas turbine unit. The collected data includes gas flow rate and volume fractions of carbon monoxide, hydrogen, and methane, as well as volume fractions of carbon dioxide and nitrogen, forming complete data on the proportion of gas components.

[0097] It should be noted that the gas component percentage data is based on the gas component detection principle, which measures the volume fractions of carbon monoxide, hydrogen, methane, carbon dioxide, and nitrogen in the gasification product. This is existing technology and will not be elaborated further here.

[0098] Based on the gas component proportion data, the effective calorific value coefficient of the gas is obtained by weighting and summing the standard lower heating value and volume fraction of each combustible component.

[0099] Optionally, the formula for calculating the effective calorific value coefficient is:

[0100] ;

[0101] In the formula, The effective calorific value coefficient of a gas is used to characterize the overall lower heating value per unit volume of a gas mixture. The number of combustible gas components involved in the calculation. For the first Volume fraction of each combustible gas component in the gas mixture. For the first The standard lower heating value corresponding to each combustible gas component.

[0102] In this embodiment, by selecting only the operating data after the gas turbine unit enters the stable operating range, and introducing a synchronous acquisition mechanism for speed maintenance and output power, the resulting dataset can truly reflect the actual response characteristics of the gas turbine to gas quality and gas supply capacity under continuous stable operating conditions. This avoids the obscuring of data patterns by start-up and shutdown processes and transient disturbances, thereby improving the representativeness and engineering effectiveness of the data from the source.

[0103] By performing time-series consistency verification and anomaly removal on the original stable operating data, atypical operating data caused by short-term load fluctuations, control lags, or local disturbances are systematically eliminated, so that each sample in the dataset has a unified physical meaning and comparable operating conditions, providing a high-quality input foundation for subsequent feature extraction and model training.

[0104] Furthermore, by introducing speed holding state as an operational response feature, the dataset not only contains power level information but also reflects the gas engine's sensitivity to changes in gas quality. This allows for the differentiation of operational difficulty corresponding to different gas calorific values ​​and stability under the same power conditions, providing richer criteria for subsequent refined gas replenishment control.

[0105] Example 2

[0106] Please see Figure 1 and Figure 3 Based on Example 1, this embodiment further discloses an optimized control method for biomass gas coupling units based on digital twins.

[0107] The biomass gas coupling unit optimization control method based on digital twins also includes:

[0108] Under experimental conditions, the stability risk index of the gas turbine unit was collected under different biomass gas state parameters and different biomass gas participation ratios. Based on the preset stability risk index threshold, a labeling rule for whether the biomass gas participation ratio meets or does not meet the standard was constructed, and a stability state discrimination machine learning model for judging the stability state of the gas turbine unit was trained.

[0109] The methods for obtaining the stability risk index include:

[0110] Within a preset sliding time window, statistical analysis is performed on the gas engine speed data to calculate the characteristic quantity of speed fluctuation intensity.

[0111] The formula for calculating the intensity characteristic of the rotational speed fluctuation is:

[0112] In the formula, This is a characteristic quantity of the intensity of rotational speed fluctuation. The standard deviation of the gas turbine speed within the sliding time window is given. The target speed set for the gas engine;

[0113] Within a preset sliding time window, acquire time-varying data of combustion temperature and calculate the characteristic quantity of combustion response change rate;

[0114] The formula for calculating the characteristic quantity of the combustion response change rate is:

[0115] ; This parameter represents the rate of change of combustion response, used to quantitatively characterize the relative intensity of change of combustion parameters over time. It is an important state indicator reflecting the stability of the combustion process. A larger value indicates more drastic changes in the combustion process and a higher risk to operational stability. This is a real-time sampled value of the combustion temperature. This serves as the baseline value for the corresponding parameter within its stable operating range. Indicates combustion state parameters Relative to time The first time derivative, i.e. the instantaneous rate of change of combustion parameters, is used to reflect the speed of change and dynamic response intensity of the combustion process.

[0116] After normalizing the characteristic quantities of speed fluctuation intensity and combustion response change rate, the stability risk index is obtained by weighted summation.

[0117] It should be noted that a time window length of 60 seconds and a period of 1 second are preferred.

[0118] The formula for calculating the stability risk index is as follows: ;

[0119] In the formula, This is a stability risk index. and These are the weighting coefficients. This is the normalized characteristic quantity of rotational speed fluctuation intensity. This is the normalized characteristic quantity of the combustion response change rate.

[0120] Furthermore, the normalized calculation formulas for the speed fluctuation intensity characteristic and the combustion response change rate characteristic are as follows:

[0121] ;

[0122] ;

[0123] This refers to the maximum safe value of the speed fluctuation intensity characteristic obtained from the statistical analysis of historical stable operation samples of gas turbine units; This refers to the maximum safe value allowed for the combustion response change rate characteristic quantity obtained from statistical analysis of historical stable operation samples of gas turbine units; This is a function that takes the minimum value.

[0124] It should be noted that the formulas mentioned above are all numerical calculations after removing dimensions. They are formulas that are closest to the real situation, obtained by software simulation based on a large amount of data. The weight coefficients in the formulas and the preset thresholds in the analysis process are set by those skilled in the art based on the actual situation or obtained by simulation based on a large amount of data.

[0125] The threshold-based labeling method includes: when the stability risk index of the gas turbine unit is lower than the stability risk index threshold, it is set to meet the biomass gas participation ratio standard and the label is set to 1;

[0126] When the stability risk index of the gas turbine unit is greater than or equal to the stability risk index threshold, it is set as a non-compliance of biomass gas participation ratio and the label is set to 0.

[0127] The training method for the stable state discrimination machine learning model includes:

[0128] Gradient boosting regression tree was selected as the machine learning model for determining steady state.

[0129] Set the initial hyperparameters, including: number of decision trees: 120-180; maximum depth of a single tree: 5-7; minimum number of samples for node splitting: 10-15; maximum number of features to consider during splitting: 2 / 3 of the total number of features; learning rate: 0.05-0.1; loss function: binary cross-entropy; regularization coefficient (L2): 0.1-0.2; splitting criterion: Gini impurity.

[0130] Model initialization involves setting the initial prediction value to the ratio of positive to negative samples in the training set. For example, if the proportion of qualified samples is 60%, the initial prediction probability is 0.6. For each sample in the training set, the negative gradient between the current model's predicted probability and the true label, i.e., the classification residual, is calculated to reflect the model's prediction bias. A new decision tree is constructed with the goal of "minimizing the residual". The optimal splitting feature is selected based on Gini impurity, such as prioritizing the proportion of biomass gas participation and the calorific value coefficient. The samples are then divided into different sub-nodes until the stopping condition is met. The stopping condition is that the maximum depth is reached or the number of sub-node samples is less than the minimum number of samples.

[0131] The weights of the leaf nodes of the new tree are optimized by gradient descent, and the contribution of the new tree to the final prediction is controlled by the learning rate. This avoids the dominance of a single tree and integrates it into the existing model to correct prediction bias.

[0132] The Bayesian optimization method is used to search for the optimal combination of hyperparameters within a preset range; optimization range: number of decision trees: 100-200; maximum depth of a single tree: 4-8; learning rate: 0.03-0.12; minimum number of samples for node splitting: 8-20; regularization coefficient (L2): 0.05-0.25;

[0133] Every 20 trees are trained in each iteration, the root mean square error of the validation set is calculated. When the root mean square error of the validation set decreases by less than 0.001 for three consecutive iterations, training is stopped to avoid the model overfitting the specific conditions of the training set.

[0134] Save the model parameters (including the splitting rules of all decision trees and the weights of leaf nodes) that minimize the root mean square error of the validation set during training; test the model using the test set, and when the F1 score is ≥96% and the precision is ≥95%, export the trained model in a lightweight format, such as pkl or ONNX, and integrate it into the gas engine control system for use.

[0135] For example, the training process of a steady-state discrimination model is as follows: Model initialization, the training set statistics show that the proportion of qualified samples is 60%, and the initial prediction probability is P(0)=0.6; First round of gradient boosting iteration, the true label y=0, the current prediction probability P(0)=0.6; Calculate the negative gradient (residual) under the binary classification cross-entropy loss. Decision tree splitting, the splitting rule for the first tree is that the proportion of biomass gas participation is ≤32%, and the effective calorific value coefficient is... If the energy content is ≤4.9 MJ / Nm³, the sample enters the leaf node of the "high participation rate + low calorific value" branch, with a corresponding leaf node weight of w = −1.1; the predicted probability is updated with a learning rate of η = 0.08, and the updated log probability is... The corresponding probability P(1) is approximately 0.56; with the gradual correction of the subsequent 140 trees, the predicted probability of this sample eventually drops to P(meeting the standard) = 0.24, which is lower than the set discrimination threshold of 0.5, and the output discrimination result is unqualified.

[0136] The current biomass gas state parameters, biomass gas flow rate, and predicted gas supply flow rate are input into the steady state discrimination machine learning model to obtain the discrimination result of whether the current biomass gas participation ratio meets the requirements for stable operation.

[0137] When the judgment result meets the standard, the gas supply in the gas unit is controlled based on the predicted gas supply flow rate;

[0138] When the judgment result is unsatisfactory, while keeping the current biomass gas state parameters unchanged, the biomass gas participation ratio is searched in reverse direction in the decreasing direction, that is, the biomass gas flow rate is reduced by a preset ratio, such as 1% of the total gas. The steady state judgment machine learning model is used to judge the results of each participation ratio to obtain the maximum biomass gas participation ratio when the judgment result meets the standard. The biomass gas flow rate corresponding to the current maximum biomass gas participation ratio is obtained and marked as the ideal biomass gas flow rate.

[0139] It should be noted that, under the condition that the state parameters of biomass gas remain unchanged, the total input energy required by the gas engine is constant. An increase in the input energy of biomass gas will lead to a proportional decrease in the input energy of gas. Thus, the volume of gas replaced by a unit volume of biomass gas is determined by the ratio of the lower heating values ​​of the two gases.

[0140] Based on the ideal biomass gas flow rate, biomass gas state parameters, and the current gas turbine output power, a machine learning model for predicting gas supply flow rate is invoked to obtain the gas supply flow rate prediction result under stability constraints.

[0141] The predicted gas supply flow rate and the corresponding biomass gas flow rate are combined and applied to the gas turbine unit.

[0142] It should be noted that biomass gas has a lower calorific value, lower methane content, and a higher proportion of inert components compared to fuel gas. Therefore, as the proportion of biomass gas increases, the equivalent calorific value of the mixed fuel decreases monotonically, the flame stability margin decreases monotonically, and the risks of combustion oscillation, backfire, and flameout increase monotonically. Thus, under given operating conditions and composition, stability exhibits a clear "monotonically constrained relationship" with the proportion of biomass gas involved. There is an upper limit to the proportion of biomass gas involved. Once this limit is exceeded, the system will inevitably enter an unstable range under that operating condition. When instability is detected, the only effective, controllable, and physically consistent adjustment direction is to reduce the proportion of biomass gas involved.

[0143] In this embodiment, the present invention constructs a stable state discrimination machine learning model for gas turbine units based on experimental data training. When it is detected that the biomass gas participation ratio does not meet the stable operation requirements, it no longer adopts empirical or fixed ratio adjustment methods. Instead, based on the stability discrimination results, it performs a reverse search along the direction of decreasing biomass gas participation ratio. This avoids the risk of backfire, combustion oscillation or flameout caused by continuing to increase the biomass gas ratio under unstable operating conditions, and significantly improves the operational safety and reliability of gas turbine units under biomass gas blending conditions.

[0144] Furthermore, this invention does not simply reduce the proportion of biomass gas involved, but rather, while keeping the biomass gas state parameters unchanged, it determines the maximum proportion of biomass gas involved that just meets the requirements for stable operation through a reverse search method. This ensures that the blending level of biomass gas always operates near the stability boundary, thereby maximizing the proportion of biomass gas replacing conventional gas without sacrificing combustion stability, and improving the utilization rate of renewable energy and the economic efficiency of the system.

[0145] Example 3

[0146] Please see Figure 4 This invention provides a technical solution: a biomass gas coupling unit optimization control system based on digital twins, which is used to implement the aforementioned biomass gas coupling unit optimization control method based on digital twins, including:

[0147] The raw operating data acquisition module acquires the operating data of the biomass gas coupled gas turbine unit after it enters the stable operating range during the historical operation process, and obtains the raw dataset. The raw dataset includes the gas turbine output power, gas turbine speed, gas supply flow rate, biomass gas state parameters, and biomass gas flow rate.

[0148] The effective operational data generation module filters the collected raw dataset based on operational data, eliminating data interference from the start-up and shutdown phases and the load mutation phase, and forming an effective operational dataset for subsequent analysis.

[0149] The training data construction module constructs the gas turbine output power, biomass gas state parameters, biomass gas flow rate and gas supply flow rate from the effective running dataset into effective response training data, forming a training dataset, and training a machine learning model to predict gas supply flow rate.

[0150] The gas supply flow prediction module obtains the current gas turbine output power, biomass gas state parameters and biomass gas flow, and predicts the gas supply flow based on the trained machine learning model to control the gas supply flow of the gas turbine unit.

[0151] The model training module collects the stability risk index of the gas turbine unit under experimental conditions, with different biomass gas state parameters and different biomass gas participation ratios. Based on the preset stability risk index threshold, it constructs the labeling rules for whether the biomass gas participation ratio meets or does not meet the standards, and trains a stability state discrimination machine learning model for judging the stability state of the gas turbine unit.

[0152] The stability discrimination module inputs the current biomass gas state parameters, biomass gas flow rate and predicted gas supply flow rate into the stability discrimination machine learning model to obtain the discrimination result of whether the current biomass gas participation ratio meets the requirements for stable operation.

[0153] The reverse search module performs a reverse search along the decreasing direction of the biomass gas participation ratio when the judgment result is not up to standard. It also judges the results of each participation ratio through the steady state discrimination machine learning model to obtain the maximum biomass gas participation ratio when the judgment result meets the standard, and obtains the biomass gas flow rate corresponding to the current maximum biomass gas participation ratio, which is marked as the ideal biomass gas flow rate.

[0154] The gas supply prediction module, based on the ideal biomass gas flow rate, biomass gas state parameters, and the current gas turbine output power, calls the predictive gas supply flow rate machine learning model to obtain the gas supply flow rate prediction result under stability constraints.

[0155] The control and execution module applies the predicted gas supply flow rate and the corresponding biomass gas flow rate to the gas turbine unit.

[0156] The foregoing has shown and described the basic principles, main features, and advantages of the present invention. Those skilled in the art should understand that the present invention is not limited to the above embodiments. The embodiments and descriptions in the specification are merely illustrative of the principles of the invention. Various changes and modifications can be made to the invention without departing from its spirit and scope, and all such changes and modifications fall within the scope of the present invention as claimed. The scope of protection of the present invention is defined by the appended claims and their equivalents.

Claims

1. A method for optimized control of biomass gas coupled units based on digital twins, characterized in that, include: The operation data of the biomass gas coupled gas turbine unit after it entered the stable operation range during the historical operation process is obtained to obtain the original dataset. The original dataset includes the gas turbine output power, gas turbine speed, gas supply flow rate, biomass gas state parameters and biomass gas flow rate. Based on operational data, the collected raw dataset is filtered to remove data interference from the start-up and shutdown phases and the load change phases, forming an effective operational dataset; The gas turbine output power, biomass gas state parameters, biomass gas flow rate and gas supply flow rate in the effective operating dataset are used to construct the effective response training data, forming a training dataset, and a machine learning model for predicting gas supply flow rate is trained. The system acquires the current gas turbine output power, biomass gas state parameters, and biomass gas flow rate. Based on the trained machine learning model, it predicts the gas supply flow rate and controls the gas supply flow rate of the gas turbine unit. Under experimental conditions, the stability risk index of the gas turbine was collected under different biomass gas state parameters and participation ratios, and a steady state discrimination machine learning model was trained based on threshold labeling. The current biomass gas state parameters, biomass gas flow rate, and predicted gas supply flow rate are input into the steady state discrimination machine learning model to obtain the discrimination result of whether the current biomass gas participation ratio meets the requirements for stable operation. When the judgment result is not up to standard, the biomass gas participation ratio is searched in reverse along the decreasing direction, and the steady state discrimination machine learning model is used to judge the results of each participation ratio to obtain the maximum biomass gas participation ratio when the judgment result meets the standard, and the corresponding ideal biomass gas flow rate is obtained. Based on the ideal biomass gas flow rate, biomass gas state parameters, and the current gas turbine output power, a machine learning model for predicting gas supply flow rate is invoked to obtain the gas supply flow rate prediction result under stability constraints. The predicted gas supply flow rate and the corresponding biomass gas flow rate are combined and applied to the gas turbine unit. The methods for obtaining the stability risk index include: Within a preset sliding time window, statistical analysis is performed on the gas engine speed data to calculate the characteristic quantity of speed fluctuation intensity. Within a preset sliding time window, acquire time-varying data of combustion temperature and calculate the characteristic quantity of combustion response change rate; After normalizing the characteristic quantities of speed fluctuation intensity and combustion response change rate, the stability risk index is obtained by weighted summation.

2. The optimized control method for biomass gas coupled units based on digital twins according to claim 1, characterized in that, Methods for filtering the collected raw datasets to form valid running datasets include: The preset time window length is used to calculate the rate of change of the gas turbine output power within a continuous time window. The output power stability criterion is set as follows: when the rate of change of output power is lower than a preset threshold, the output power is determined to be in a stable state within the time window; when the rate of change of output power exceeds the preset threshold, it is determined to be a load adjustment or disturbance stage, and the corresponding data is not included in the stable operation range. Within a preset time window, the speed maintenance coefficient of the gas engine is calculated, and the speed maintenance stability criterion is set as follows: when the speed maintenance coefficient is less than the preset speed maintenance coefficient threshold, the gas engine is determined to be in a stable maintenance state. The above-mentioned output power stability criterion and speed maintenance stability criterion are jointly judged. Only when both criteria are met simultaneously within the same time window are the data corresponding to that time window determined as valid operating data.

3. The optimized control method for biomass gas coupled units based on digital twins according to claim 2, characterized in that, The method for obtaining the speed maintenance coefficient includes: obtaining the deviation between the average time value of the actual speed of the gas engine within a preset time window and the target speed set by the gas engine, obtaining the absolute value of the deviation, obtaining the standard deviation of the speed of the gas engine within the preset time window, obtaining the ratio of the absolute value of the deviation to the target speed set by the gas engine and the ratio of the standard deviation to the target speed set by the gas engine, and then weighted summing the two ratios to obtain the speed maintenance coefficient.

4. The optimized control method for biomass gas coupled units based on digital twins according to claim 1, characterized in that, The effective response training data includes the gas turbine output power, biomass gas state parameters, biomass gas flow rate, and the corresponding gas supply flow rate.

5. The optimized control method for biomass gas coupled units based on digital twins according to claim 4, characterized in that, The biomass gas state parameters include the effective calorific value coefficient and the proportion of gas components; The method for collecting the biomass gas state parameters includes: At the gas outlet of the biomass gasification system, online data is collected on the gasification product before it enters the gas turbine unit. The collected data includes gas flow rate and volume fractions of carbon monoxide, hydrogen, and methane, as well as volume fractions of carbon dioxide and nitrogen, forming complete data on the proportion of gas components. Based on the gas component proportion data, the effective calorific value coefficient of the gas is obtained by weighting and summing the standard lower heating value and volume fraction of each combustible component.

6. The optimized control method for biomass gas coupled units based on digital twins according to claim 1 or 5, characterized in that, The training method for the machine learning model for predicting gas supply flow includes: The gas turbine operating conditions are stratified, and stratified random sampling is adopted. The effective operating dataset is divided into training set, validation set and test set according to a preset ratio, so that the distribution of each operating condition is consistent in the training set, validation set and test set. Gradient boosting regression tree was selected as the gas supply flow prediction model. Initial hyperparameters were set and the model was initialized. The initial predicted value of the gradient boosting regression tree model was set as the mean of the gas supply flow in the training set. For each sample in the training set, the difference between the gas supply flow predicted by the current model and the actual flow, i.e., the residual, was calculated. A new decision tree was constructed. With the goal of minimizing the residual, the optimal splitting feature and splitting point were selected based on the mean square error. The samples were divided into different child nodes until the preset stopping condition was met. The Bayesian optimization method is used to search for the optimal combination of hyperparameters within a preset range; For every 20 trees in each iteration, the root mean square error of the validation set is calculated. When the root mean square error of the validation set decreases by less than 0.001 for three consecutive iterations, training is stopped, and the model parameters with the smallest root mean square error of the validation set during training are saved.

7. The optimized control method for biomass gas coupled units based on digital twins according to claim 1, characterized in that, When the judgment result meets the standard, the gas supply in the gas unit is controlled based on the predicted gas supply flow rate.

8. The optimized control method for biomass gas coupled units based on digital twins according to claim 1, characterized in that, The threshold-based labeling method includes: when the stability risk index of the gas turbine unit is lower than the stability risk index threshold, it is set to meet the biomass gas participation ratio standard and the label is set to 1; When the stability risk index of the gas turbine unit is greater than or equal to the stability risk index threshold, it is set as a non-compliance of biomass gas participation ratio and the label is set to 0.

9. A biomass gas coupling unit optimization control system based on digital twins, used to implement the biomass gas coupling unit optimization control method based on digital twins as described in any one of claims 1-8, characterized in that, include: The raw operating data acquisition module acquires the operating data of the biomass gas coupled gas turbine unit after it enters the stable operating range during the historical operation process, and obtains the raw dataset. The raw dataset includes the gas turbine output power, gas turbine speed, gas supply flow rate, biomass gas state parameters, and biomass gas flow rate. The effective operation data generation module filters the collected raw dataset based on the operation data, eliminating data interference from the start-up and shutdown phases and the load change phases, and forming an effective operation dataset. The training data construction module constructs the gas turbine output power, biomass gas state parameters, biomass gas flow rate and gas supply flow rate from the effective running dataset into effective response training data, forming a training dataset, and training a machine learning model to predict gas supply flow rate. The gas supply flow prediction module obtains the current gas turbine output power, biomass gas state parameters and biomass gas flow, and predicts the gas supply flow based on the trained machine learning model to control the gas supply flow of the gas turbine unit. The model training module collects the stability risk index of the gas turbine unit under experimental conditions, with different biomass gas state parameters and different biomass gas participation ratios. Based on the preset stability risk index threshold, it constructs the labeling rules for whether the biomass gas participation ratio meets or does not meet the standards, and trains a stability state discrimination machine learning model for judging the stability state of the gas turbine unit. The stability discrimination module inputs the current biomass gas state parameters, biomass gas flow rate and predicted gas supply flow rate into the stability discrimination machine learning model to obtain the discrimination result of whether the current biomass gas participation ratio meets the requirements for stable operation. The reverse search module performs a reverse search along the decreasing direction of the biomass gas participation ratio when the judgment result is not up to standard. It also judges the results of each participation ratio through the steady state discrimination machine learning model to obtain the maximum biomass gas participation ratio when the judgment result meets the standard, and obtains the biomass gas flow rate corresponding to the current maximum biomass gas participation ratio, which is marked as the ideal biomass gas flow rate. The gas supply prediction module, based on the ideal biomass gas flow rate, biomass gas state parameters, and the current gas turbine output power, calls the predictive gas supply flow rate machine learning model to obtain the gas supply flow rate prediction result under stability constraints. The control and execution module applies the predicted gas supply flow rate and the corresponding biomass gas flow rate to the gas turbine unit.

Images