Gas turbine intelligent regulation and control method, system and equipment based on operation data analysis and medium

Abstract

The invention discloses a gas turbine intelligent regulation and control method, system and equipment based on operation data analysis and a medium, and belongs to the field of gas turbine operation regulation and control. The method comprises the steps that multi-dimensional monitoring data in gas turbine operation is collected, and a data set is constructed; extracting relevance between time sequence characteristics and variables based on a sliding time window, and identifying a current running state; fusing the state classification result and the data set, constructing a trend modeling input structure, executing trend prediction and outputting a prediction sequence of the target index; selecting a control strategy path based on a state and trend combination result, generating control parameters and constructing an adjustment instruction; and collecting execution feedback data, performing prediction deviation comparison, optimizing control parameter generation logic based on a comparison result, and realizing iterative adjustment of a control strategy. According to the method, a closed-loop regulation and control chain of state recognition, trend prediction, control decision and feedback correction is constructed, and the control response precision and the regulation stability of the gas turbine under complex working conditions are improved.

Description

Technical Field

[0001] The present invention relates to the technical field of gas turbine operation regulation, and particularly relates to an intelligent regulation method, system, device and medium for gas turbines based on operation data analysis. Background Art

[0002] As a typical thermal mechanical device, the operating state of a gas turbine is affected by the complex coupling of multiple parameters such as temperature, speed, pressure, and fuel supply. During the actual operation process, the gas turbine faces various state switching and non-linear disturbance problems. Conventional control methods mostly rely on fixed operating condition judgment thresholds and static parameter adjustment strategies, and it is difficult to effectively adapt to the operating conditions with high-speed changes and multi-stage characteristics.

[0003] Traditional gas turbine control systems mainly operate based on preset operating condition models and expert experience rules, lacking the ability to dynamically analyze real-time operation data and build trend models. In terms of state recognition, existing methods mostly use fixed parameter interval judgments and cannot capture the subtle trend changes and multi-factor linkage behaviors during the operation process; in terms of trend prediction, most methods only rely on time series modeling and fail to combine the current operating condition state for dynamic structure scheduling; in terms of control parameter generation, conventional methods are difficult to achieve the parameter logic combination driven by state-trend joint, and lack an effective deviation evaluation and adaptive adjustment mechanism during the control feedback process, which is prone to adjustment lag or control instability.

[0004] In summary, the existing gas turbine regulation technologies still have obvious deficiencies in aspects such as the accuracy of state recognition, the adaptability of trend prediction, the intelligent generation of control parameters, and feedback iterative optimization, and it is difficult to meet the intelligent regulation requirements of high performance and high robustness. Summary of the Invention

[0005] In view of the above problems, the present invention is proposed.

[0006] Therefore, the technical problem solved by the present invention is: how to construct a gas turbine operation regulation method with dynamic state recognition ability and trend structure modeling ability based on multi-dimensional time series monitoring data, realize the real-time discrimination of the current operating state of the gas turbine and the structured description of the future operating trend, and construct the control parameter generation logic to improve the intelligent control accuracy and response stability of the gas turbine under multi-stage operating conditions.

[0007] To solve the above technical problems, the present invention provides the following technical solution: an intelligent regulation method for gas turbines based on operation data analysis, which includes,

[0008] Collect multi-dimensional monitoring data from the operation process of the gas turbine, and construct it into an initial operation data set in chronological order and perform preprocessing operations to generate a data set with a unified time base and feature fields;

[0009] The data set is input into the operation status identification process. Based on the time-series change characteristics and correlation patterns of the monitoring data, the current operation status of the gas turbine is identified, and the corresponding status classification results are output.

[0010] The state classification results are merged with the dataset, and trend prediction processing based on the time series change patterns is performed.

[0011] In trend prediction processing, target monitoring indicators related to gas turbine operation control are selected, a prediction model is constructed to characterize the changing trend within the target time period in the future, and the prediction results are output.

[0012] Control parameters are generated based on the prediction results and state classification results, and adjustment instructions are constructed.

[0013] The system executes adjustment commands to obtain operational feedback data during the control execution process, compares the deviation with the prediction results, and optimizes and iterates the control commands based on the comparison results.

[0014] As a preferred embodiment of the intelligent control method for gas turbines based on operational data analysis described in this invention, the operational status identification process includes:

[0015] The dataset is used to construct a sliding time window sample according to a set sampling period. For each window sample, the slope, magnitude of change and segment mean of the monitoring data in time sequence are extracted as time series features. The correlation between the fields in the structured dataset is combined to form a state recognition input vector. The data is then processed by a classification model to output the state classification result corresponding to the current window.

[0016] As a preferred embodiment of the intelligent control method for gas turbines based on operational data analysis described in this invention, the step of performing trend prediction processing based on time series variation patterns includes:

[0017] The status classification results are synchronized and aligned with the dataset according to the timestamp. The status type field is appended to the record data of the corresponding time period to construct a trend modeling input sample sequence containing monitoring parameter field and status label field. The sample sequence is then used as the input content of the trend prediction model.

[0018] As a preferred embodiment of the intelligent control method for gas turbines based on operational data analysis described in this invention, the step of performing trend prediction processing based on time series variation patterns includes:

[0019] For the target monitoring indicators in the input sample sequence of trend modeling, the numerical change trajectory and parameter structure of the current indicator in a continuous time period are extracted, and a prediction model input path is constructed to represent the direction and magnitude of trend change. The predicted value sequence covering each time point in the target prediction period is output through the current path.

[0020] As a preferred embodiment of the intelligent control method for gas turbines based on operational data analysis described in this invention, the step of generating control parameters based on prediction results and state classification results includes:

[0021] Based on the trend output structure of the prediction results and the state information in the state classification results, the corresponding adjustment condition structure is selected in the control logic structure to generate a set of control parameters.

[0022] The set of control parameters is constructed into adjustment commands, which are then sent to the control system for execution.

[0023] During the execution of the adjustment instructions, operational feedback data is collected, and the operational feedback data is time-aligned with the prediction results to form a feedback comparison data structure.

[0024] The control deviation index is determined based on the feedback comparison data structure, and the set of control parameters is adjusted according to the control deviation index.

[0025] As a preferred embodiment of the intelligent control method for gas turbines based on operational data analysis described in this invention, the trend prediction processing further includes:

[0026] The state type in the state classification results is used as the structural control condition.

[0027] Multiple modeling structural paths are set in the prediction model, each corresponding to a different state type.

[0028] During model execution, the corresponding path is selected based on the current state type, and the trend modeling input sample sequence is input into the selected path.

[0029] Based on the generated trend prediction sequence, the change direction segments and trend intensity indicators covering the target prediction time period are extracted to construct a trend combination condition structure, which serves as the input basis for the control parameter generation logic.

[0030] The beneficial effects of this preferred technical solution are as follows: By using the state type in the state classification results as the structural control condition, the modeling structure path that matches the current operating state of the gas turbine can be dynamically selected during the trend prediction process, thereby realizing state-based switching control of the predicted model structure. Compared with traditional single-structure trend prediction methods, the current technical solution, by pre-setting multiple modeling structure paths and scheduling paths based on state type, enables the trend modeling process to have structural adaptability, ensuring that the model structure matches the operating stage. Furthermore, the current technical solution extracts trend change direction and intensity indicators after trend prediction, constructing a trend combination condition structure. This provides input basis with state-trend joint characteristics for the subsequent control parameter generation logic, effectively establishing a structured connection between the trend prediction module and the control optimization module. This supports a differentiated generation mechanism for adjustment commands, improving the targeting of control path scheduling and the hierarchical capability of model organization.

[0031] As a preferred embodiment of the intelligent control method for gas turbines based on operational data analysis described in this invention, the generation of control parameters includes a control strategy path selection process based on the combination of state type and trend conditions in the state classification results.

[0032] The control strategy path selection process includes combining the state type in the state classification results with the trend change direction and change intensity indicators in the trend combination condition structure to form a state-trend joint identifier.

[0033] Several strategy paths are preset in the control strategy structure, and each strategy path corresponds to a state-trend joint identifier;

[0034] Match the corresponding strategy path based on the current state-trend joint identifier, and select control parameters from the matched strategy path to generate logic;

[0035] A set of control parameters is constructed using control parameter generation logic;

[0036] The control parameter set includes control execution module fields, parameter adjustment fields, and control priority fields, and is encapsulated as structured adjustment instructions to be sent to the control system.

[0037] The beneficial effects of this preferred technical solution are as follows: By combining the state type in the state classification results with the trend change direction and intensity indicators in the trend combination condition structure, a state-trend joint identifier is formed, which enables a structured expression of multi-dimensional operating condition characteristics during the control strategy path selection process. Based on the preset control strategy structure, this joint identifier is used to drive the strategy path matching operation, enabling the control system to accurately select the control path that best matches the current operating condition from multiple control logic branches. The control parameter generation logic associated with this path can output a structured set of control parameters containing execution module fields, adjustment parameter fields, and control priority fields according to the operating condition characteristics, thereby ensuring that the generation process of control commands has a clear rule mapping relationship and contextual relevance.

[0038] This invention provides a method and system for intelligent control of gas turbines based on operational data analysis.

[0039] To solve the above technical problems, the present invention provides the following technical solution: a gas turbine intelligent control method system based on operation data analysis, comprising: a data collection and preprocessing module, a state classification module, a data merging module, a model building module, a control command module, and an optimization iteration module;

[0040] The data collection and preprocessing module collects multi-dimensional monitoring data during the operation of the gas turbine, and constructs an initial operating dataset in chronological order. It then merges and preprocesses the data to generate a data set with a unified time reference and feature fields.

[0041] The state classification module inputs the data set into the operating state identification process, identifies the current operating state of the gas turbine based on the temporal change characteristics and correlation patterns of the monitoring data, and outputs the corresponding state classification results.

[0042] The data merging module merges the state classification results with the data set and performs trend prediction processing based on the time series change pattern.

[0043] The model building module selects target monitoring indicators related to gas turbine operation control in the trend prediction processing, builds a prediction model to depict the changing trend in the future target time period, and outputs the prediction results.

[0044] The adjustment instruction module generates control parameters and constructs adjustment instructions based on the prediction results and state classification results.

[0045] The optimization iteration module executes adjustment instructions to obtain operational feedback data during the control execution process, compares the deviation with the prediction results, and optimizes and iterates the control instructions based on the comparison results.

[0046] The present invention provides a computer device, including a memory and a processor, wherein the memory stores a computer program, characterized in that the processor executes the computer program to implement the steps of the intelligent control method for gas turbines based on operational data analysis.

[0047] The present invention provides 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 aforementioned intelligent control method for gas turbines based on operational data analysis.

[0048] The beneficial effects of this invention are as follows: By constructing a time-series structure and extracting change features from the collected multi-dimensional operational data, and combining this with the correlation patterns between operating parameters, this invention can accurately identify the current operating state of the gas turbine. This allows the control logic to be dynamically adjusted according to changes in operating conditions, thereby improving the pertinence and adaptability of strategy selection. By constructing a predictive structure based on state perception and multi-scale modeling, this invention can characterize the evolution trend of key monitoring parameters in advance over a future time period, providing a basis for feedforward adjustment of the control system and reducing the risk of overshoot, delay, or instability caused by parameter mutations. This invention sets up a process for comparing feedback data acquisition and prediction deviation after control execution, which can quantify the error of the adjustment effect and automatically correct the control parameter generation strategy based on the error results. This achieves feedback-based self-learning and adjustment optimization, enhancing the long-term operational stability and adaptability of the system. Attached Figure Description

[0049] To more clearly illustrate the technical solutions of the embodiments of the present invention, the accompanying drawings used in the description of the embodiments will be briefly introduced below. Obviously, the accompanying 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.

[0050] Figure 1 The above is a flowchart of an intelligent control method for gas turbines based on operational data analysis, provided as an embodiment of the present invention. Detailed Implementation

[0051] 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.

[0052] Example 1, referring to Figure 1This is one embodiment of the present invention, which provides a method for intelligent control of a gas turbine based on operational data analysis, comprising:

[0053] S1. Collect multidimensional monitoring data from the gas turbine during operation, and construct an initial operating dataset in chronological order. Combine and preprocess the data to generate a data set with a unified time reference and feature fields.

[0054] S2. Input the data set into the operation status identification process, identify the current operation status of the gas turbine based on the time-series change characteristics and correlation patterns of the monitoring data, and output the corresponding status classification results.

[0055] S3. Merge the state classification results with the dataset and perform trend prediction processing based on the time series change patterns.

[0056] S4. In the trend prediction processing, target monitoring indicators related to gas turbine operation control are selected, a prediction model is constructed to characterize the changing trend in the future target time period, and the prediction results are output.

[0057] S5. Generate control parameters based on the prediction results and state classification results, and construct adjustment instructions.

[0058] S6. Execute adjustment instructions to obtain operational feedback data during the control execution process, compare the deviation with the prediction results, and optimize and iterate the control instructions based on the comparison results.

[0059] Example 2 is an embodiment of the present invention, which provides a gas turbine intelligent control method based on operation data analysis based on the previous embodiment, including:

[0060] S1. Collect multidimensional monitoring data from the gas turbine during operation, and construct an initial operating dataset in chronological order. Combine and preprocess the data to generate a data set with a unified time reference and feature fields.

[0061] In this invention, the operating status monitoring data of the gas turbine serves as the input basis for the intelligent control system. First, high-frequency, continuous, and reliable data acquisition operations must be completed. Based on this, the raw data undergoes format standardization, noise removal, and structured processing. To this end, the system acquires multiple key physical parameters through the original gas turbine control platform or an external edge acquisition device, and constructs a set of raw operating data with sequential characteristics in the time dimension.

[0062] Data acquisition is carried out based on the following principles: the selected monitoring parameters should cover the core operating subsystems of the gas turbine, such as combustion, compression, and power output. Typical parameters include, but are not limited to, compressor inlet temperature, combustion chamber outlet temperature, combustion chamber fuel injection flow rate, turbine inlet pressure, fuel injection valve opening, main shaft speed, power generation load, and electrical output frequency.

[0063] The data sampling frequency must meet the time resolution requirements of the control algorithm. A sampling period of 100 to 200 milliseconds is recommended to ensure that sufficient dynamics are captured.

[0064] All collected data should be bound to a timestamp to construct the original time-series record, which will facilitate subsequent multi-parameter alignment and trend calculation.

[0065] After data collection is complete, the system performs the following preprocessing operations on the raw runtime data set:

[0066] The system first performs a physical boundary constraint check, which sets a reasonable range for each parameter. Data records exceeding this range are marked as invalid. In addition, a sliding window statistical analysis is introduced to calculate the mean, standard deviation, and short-term fluctuation amplitude of each parameter within the current window. If a single point change exceeds the dynamic tolerance range, it is identified as an abrupt change and removed.

[0067] Anomaly detection supports marking time periods, which can be skipped during subsequent modeling.

[0068] When certain parameters fail to be acquired at a specific sampling time, the system uses an interpolation strategy to correct them. Interpolation methods can include linear interpolation, nearest neighbor imputation, or moving average. To ensure synchronization between different signal sources, the system uses a unified sampling interval, resamples signals of different frequencies to the main time base, and aligns the timestamps to construct a complete and synchronized data structure.

[0069] To support trend analysis and modeling, the system generates derived fields based on the original monitoring parameters, including statistical descriptive indicators such as short-term mean, standard deviation, growth slope, and first-order difference. Simultaneously, the system normalizes the parameters of each dimension according to their global historical distribution, ensuring they fall within a uniform numerical scale range, thus improving the model's convergence and stability.

[0070] After completing the above processing, the system constructs a complete feature vector sequence from the data at each sampling time, forming a structured set of operational data in chronological order, for use by the subsequent state recognition, trend prediction and control calculation modules.

[0071] Taking a heavy-duty industrial gas turbine in a natural gas-steam combined cycle unit as an example, its main control system adopts the GEMark VIe control platform. This platform, through 18 analog acquisition channels and several digital control signals arranged on the main body, completes the full-process monitoring of the burner, compressor, fuel system, and power generation.

[0072] S2. Input the data set into the operation status identification process, identify the current operation status of the gas turbine based on the time-series change characteristics and correlation patterns of the monitoring data, and output the corresponding status classification results.

[0073] The dataset is used to construct a sliding time window sample according to a set sampling period. For each window sample, the slope, magnitude of change and segment mean of the monitoring data in time sequence are extracted as time series features. The correlation between the fields in the structured dataset is combined to form a state recognition input vector. The data is then processed by a classification model to output the state classification result corresponding to the current window.

[0074] Before performing trend forecasting based on time series variation patterns, the following steps are included:

[0075] The status classification results are synchronized and aligned with the dataset according to the timestamp. The status type field is appended to the record data of the corresponding time period to construct a trend modeling input sample sequence containing monitoring parameter field and status label field. The sample sequence is then used as the input content of the trend prediction model.

[0076] The process of performing trend forecasting based on time series variation patterns includes...

[0077] For the target monitoring indicators in the input sample sequence of trend modeling, the numerical change trajectory and parameter structure of the current indicator in a continuous time period are extracted, and a prediction model input path is constructed to represent the direction and magnitude of trend change. The predicted value sequence covering each time point in the target prediction period is output through the current path.

[0078] In a preferred embodiment of this invention, the identification of the current operating state of a gas turbine is as follows:

[0079] After completing the structured construction and cleaning process of the operation data, the system needs to intelligently identify the current operation state in order to provide accurate operating conditions information for subsequent trend prediction and regulation parameter generation. To achieve this function, the processed data set is first input into the operation state recognition process. This data set already contains multiple characteristic fields with engineering physical meanings and forms continuous records in chronological order, having a unified time base and structured characteristics that can be directly modeled. The operation state recognition is processed based on the sliding time window method. Specifically, the system slides forward every 100 milliseconds. Taking the current moment as the reference point, it intercepts the continuous operation data of the previous 3 seconds to construct a state recognition sample window. This window contains 30 sampling data, and each record contains multiple key characteristic fields, including original parameters such as T4, N1, F, P3, etc., and also derivative parameters such as the change rate of T4, the standard deviation of the fuel injection rate, the slope of the main shaft speed, the change range of the guide vane opening, etc. Overall, it forms a two-dimensional time series characteristic matrix of 30×n dimensions.

[0080] During the state recognition process, the system extracts the core judgment information based on the time series change characteristics and correlation rules of the monitoring data:

[0081] The so-called "time series change characteristics" refer to the continuous evolution trend and dynamic change pattern of the operation parameters in the time dimension. Different from the static threshold judgment, the present invention captures the subtle evolution of the operation behavior by characterizing the statistical and dynamic characteristics such as the fluctuation, slope, derivative, and extreme points of the key parameters in a short period. For example, the heating rate of the combustion chamber outlet temperature (T4) in the short term, the position of the maximum slope point, and the consistency of the continuous change direction can all be used as typical characteristic indicators during the temperature rise process; while the short-term amplitude range of the main shaft speed (N1) within the 3-second window and the zero-crossing frequency of the change rate can reflect the stability of the load state.

[0082] The correlation rule between parameters is the coupling relationship and linkage reaction mode between operation parameters. As a multi-variable strongly coupled object, there are clear physical and control correlations between the subsystems of the gas turbine system. For example, the change in the fuel injection rate should cause a response of T4 after a certain time delay; the adjustment of the guide vane angle is usually accompanied by a change in the exhaust pressure; a sudden increase in the electric power often synchronously pulls the adjustment of the gas flow rate and the air-fuel ratio. In the recognition process, the system刻画 these associations by calculating characteristics such as the sliding correlation coefficient, cross covariance function, and time delay peak between parameters.

[0083] The system will incorporate these "time series characteristics of independent variables" and "dynamic coupling modes between variables" into the characteristic input space of the state recognition model to form a multi-dimensional combined characteristic vector, which is used as the input of the state recognition classifier or clustering model. Finally, a model inference link for real-time state judgment is formed.

[0084] The identification model can adopt various structural forms, and depending on the scenario, rule-based logic, density clustering, K-means, support vector machine, or random forest models can be selected. In this implementation scheme, the system uses the K-means algorithm to construct a state identification model based on three months of historical operating data, setting four typical state categories: startup and heating state, stable operation state, slight disturbance state, and high temperature anomaly state. Each state category corresponds to a feature center vector, and real-time classification is performed using Euclidean distance or Mahalanobis distance.

[0085] During the online recognition phase, the system inputs the latest time window samples into the recognition model to complete state discrimination. Each recognition operation outputs a structured state classification result, including at least the following information:

[0086] Current running status type number (e.g., status 1 indicates stable running status); status identification confidence value (e.g., probability 0.87 or matching distance 0.34); current identification time point (consistent with the end of the data window); optional field: status explanation statement (e.g., "T4 is stable, N1 has small fluctuations, F is stable, determined to be a steady state").

[0087] The aforementioned status recognition results serve as the operating condition context input for the trend prediction and control parameter generation module, ensuring that the control process possesses real-time, accurate, and reliable operational status perception capabilities.

[0088] In one optional embodiment of this invention, identifying the current operating state of a gas turbine involves a system based on a supervised learning framework, using labeled samples to construct a state classification model. Before actual deployment, combining historical operating data and manually labeled information from the gas turbine, multiple typical operating state labels are defined, including but not limited to normal stable state, rapid temperature rise state, mild disturbance state, persistently high temperature state, and potential instability state. For each state category's sample window, the same temporal features and parameter correlation features as in the preferred embodiment are extracted to construct a multi-dimensional feature vector set. During the model training phase, a Support Vector Machine (SVM) algorithm is used to construct the classification boundary using the maximum margin criterion; when the state samples exhibit a non-linear feature distribution, a kernel function can be introduced for mapping improvement. The system deploys the trained classification model into the online recognition process, performing state recognition operations on each real-time window sample. This approach can improve classification accuracy and achieve higher-dimensional state interpretation capabilities in application scenarios with prior experience and operating condition labeling resources. Compared with unsupervised clustering, this implementation method has the advantages of stronger model convergence and controllable state partitioning boundaries, and is suitable for business needs in operation and maintenance scenarios that require manual review or expert intervention of state results.

[0089] Furthermore, it should be noted that by constructing state recognition samples using a sliding time window and extracting the slope, amplitude, and local statistical features of changes in operating parameters, combined with the dynamic correlation between various parameters, an input vector with time-series expressive capabilities is formed. The operating state is then categorized and identified using a K-means clustering model. This method can extract typical state patterns based on historical operating data without the need for preset thresholds or manual annotation, enabling autonomous discrimination of states such as startup heating, stable operation, slight disturbance, and high-temperature anomalies. It is suitable for practical working conditions with large amounts of operating data and ambiguous state boundaries, and has advantages such as simple model deployment, strong structural interpretability, and transparent recognition rules.

[0090] S3. Merge the state classification results with the dataset and perform trend prediction processing based on the time series change patterns.

[0091] After identifying the current operating status, the system needs to perform short-term trend prediction on key operating parameters of the gas turbine to determine their direction and rate of change in the near future, providing feedforward support for subsequent control parameter calculations. In this step, the system first integrates the operating status classification results identified in step two with the structured data set, and uses this as input to the trend prediction model to execute a prediction processing flow based on time series variation patterns.

[0092] In the structured operational dataset, a new "Current Operational Status Label" field is added to the data record for each time point. This field can be represented by an integer number (e.g., status 0 to status 3), or a confidence index can be attached, forming a status feature vector that can be directly used by the algorithm. This merging process is based on timestamp alignment principles, ensuring a one-to-one correspondence between the status label and its corresponding monitoring data in the time dimension. The final input sample possesses both the temporal dynamic characteristics of the operational data and the contextual information of the corresponding status, enhancing the condition awareness capability of the prediction model.

[0093] Based on a fixed-length sliding time window, multi-dimensional operational data from historical periods are extracted from the dataset as prediction input. Using this input, the changing trends of target parameters over several future time periods are predicted. Predicted parameters may include, but are not limited to, combustion chamber outlet temperature (T4), main shaft speed (N1), fuel flow rate, fuel injection rate, and active power load (MW).

[0094] Each prediction input sample consists of the following three parts:

[0095] The continuous value sequence of the original operating parameters in the time dimension; the derived dynamic indicators such as the derivative, fluctuation amplitude, slope, and change frequency of each parameter; the operating status label at the current moment, which serves as the operating condition identifier or model scheduling input;

[0096] After receiving samples, the model outputs a sequence of predicted values ​​for the target parameter within a certain time interval (e.g., 3 seconds), or outputs trend indicators (e.g., upward slope, maximum predicted value, future outlier determination results, etc.).

[0097] In this invention, to adapt to the high frequency, strong coupling and nonlinear variation characteristics of gas turbine operation data, a lightweight recurrent neural network structure is preferably used for trend prediction, especially a single-layer LSTM (Long Short-Term Memory) network or GRU (Gated Recurrent Unit) model, which has the ability to handle time series dependencies and dynamic trends. Moreover, under the condition of sufficient training samples, the prediction stability and computational efficiency are high, making it suitable for deployment on edge devices for real-time prediction.

[0098] During implementation, the model's input consists of sliding multidimensional time-series samples constructed within a fixed time window (e.g., 5 seconds). These samples include original operating parameters (e.g., temperature, pressure, injection rate) and their derived features such as rate of change, difference, and fluctuation amplitude. The operating state category identified in the previous stage is also incorporated as an auxiliary input to the model. Each sample forms a vector input in the form of "time step × feature dimension." The model output is a sequence of predicted values ​​for the target parameter (e.g., T4 temperature) over a preset future time period, or trend information derived from this, such as the slope of change and the predicted maximum value.

[0099] Model training is performed offline on the central server, using data from stable runtime periods and disturbance transition periods over the past few weeks to construct the training set. The Adam optimizer and MSE loss function are then used for optimization. After training, the model parameters are compressed and deployed to edge computing terminals, with rolling prediction cycles and anomaly triggering rules set to achieve low-latency, high-frequency predictions on the deployment side.

[0100] In contrast, traditional methods such as moving average and linear extrapolation are easily affected by short-term abnormal fluctuations under gas turbine operating disturbance conditions and cannot accurately reflect future trends; while ARIMA-type statistical modeling methods require the assumption of data stationarity and are difficult to adapt to the nonlinear time-varying characteristics in real-world scenarios, and therefore are not preferred solutions in this invention.

[0101] This invention achieves real-time prediction of the operating characteristics of high-frequency gas turbines by introducing an operating status perception mechanism, constructing input samples based on time series dynamic features, and combining a trend prediction method with a lightweight modeling structure.

[0102] S4. In the trend prediction processing, target monitoring indicators related to gas turbine operation control are selected, a prediction model is constructed to characterize the changing trend in the future target time period, and the prediction results are output.

[0103] In a preferred embodiment of this invention, the construction of a predictive model to depict the changing trend over a future target time period involves the following: After completing the identification of the operating status and the preparation of the predictive input data, the system enters the trend prediction modeling stage. The core of this step lies in selecting monitoring indicators closely related to the gas turbine control process from the structured operating data, constructing a predictive model based on these indicators, modeling the dynamic trend of the target variable over a specified future time period, and outputting the predicted results for control and regulation.

[0104] From all collected parameters, priority is given to physical quantities that directly affect the combustion safety, power output stability, and energy efficiency regulation of the gas turbine. Typical targets include: combustion chamber outlet temperature (T4), main shaft speed (N1), fuel injection valve displacement (F), power generation load (MW), and turbine inlet pressure (P3).

[0105] In this invention, to adapt to the complex behaviors of gas turbine operating data, such as cross-scale dynamic characteristics, short-term disturbances, and non-stationarity of state transitions, a state-aware gated recursive prediction network is preferably constructed as the trend modeling structure. This model has the ability to identify contextual operating conditions and introduces a multi-scale time window structure to enhance the model's ability to capture features at different evolution rates.

[0106] Regarding the model input structure, the system collects the following three types of features:

[0107] The target variable (e.g., T4) is a continuous time series over a fixed period of time (e.g., 5 seconds); covariates that are dynamically coupled with the target variable (e.g., F, P3, N1, etc.) and their derivative features such as derivatives, volatility, and moving standard deviation; and the corresponding time period's running status label sequence, which serves as auxiliary input or model scheduling signals.

[0108] The model structure consists of three processing modules: the first layer is a multi-scale sliding window extraction module, which constructs time series feature vectors at different resolutions with window lengths of 1 second, 3 seconds, and 5 seconds respectively; the second layer is a parallel recursive network based on gated units (GRU), which independently models features at each scale and extracts local trend representations; the third layer is a state-aware fusion module, which integrates running state labels and multi-scale trend embedding vectors to output a unified trend prediction sequence.

[0109] The model output includes not only a continuous sequence of predicted values ​​for the target time period in the future (such as the estimated value of T4 every 100ms in the next 3 seconds), but also a description of the trend structure derived from the predicted values, such as: the predicted maximum and minimum values; the maximum rate of temperature increase and the slope of change; the number and location of inflection points; and whether there is a "warning trend" (such as rising too fast or exceeding the set threshold).

[0110] The above output will serve as the feedforward input for the next control parameter generation module, improving the responsiveness and intervention accuracy of the control strategy.

[0111] Model training is based on historical operational data samples, covering multiple state labels such as startup state, stable state, perturbation state, and overheating risk state, constructing a training set containing state labels and trend evolution trajectories. The loss function design considers both prediction accuracy (e.g., MSE) and trend structure fidelity (e.g., the degree of consistency between predicted slope and measured slope). The final trained model is deployed on an edge computing platform, performing rolling predictions at 100ms intervals for real-time updates.

[0112] After completing the trend prediction of the target parameters, the system needs to combine the current operating status information with the prediction results to derive the control parameters required for adjustment, and construct them into structured adjustment instructions, which are then transmitted to the control execution unit to guide the fuel injection system, guide vane adjustment mechanism or load distribution equipment to make real-time dynamic adjustments.

[0113] In one optional embodiment of this invention, constructing a predictive model to characterize the changing trend within a future target time period involves employing a single-channel predictive structure that combines convolutional feature extraction and recursive modeling to improve the model's execution efficiency and deployment simplicity in resource-constrained scenarios. In this scheme, a single target monitoring indicator (such as T4, F, or P3) related to control behavior is first selected from a structured dataset. Its continuous time series is extracted within a fixed-length window, and a one-dimensional convolutional network (1D-CNN) is used to construct a local fluctuation feature representation. Subsequently, the convolutional output is used as the time series input and fed into a single-channel long short-term memory network (LSTM) to model the overall trend evolution path of the indicator.

[0114] Regarding the model input structure, only the original sequence of the target variable and its first-order difference sequence are collected, without introducing state labels or covariate channels. For the prediction output, the model directly outputs a continuous predicted value sequence of the indicator within the future target time period. The system extracts structural features such as the average rate of change, the magnitude of local peak changes, and trend inflection points from this predicted sequence to assist in determining the adjustment direction and magnitude of control commands.

[0115] The current optional implementation does not rely on external state recognition information, making it suitable for scenarios where state structure changes are not drastic or the operating mode is relatively simple. It can significantly simplify the trend modeling structure, reduce the burden of online inference, and facilitate deployment to edge control devices or low-power platforms. Model training uses samples with periodic or regular characteristics from historical data, and the evaluation metrics focus on prediction stability and mutation detection capabilities, and support a dynamic update mechanism based on a sliding window.

[0116] The preferred embodiment constructs a state-aware, multi-scale gated recursive prediction network to address the non-stationarity, multi-stage evolution, and short-term disturbances in gas turbine operating data. This is achieved through coordinated optimization at the model input structure, modeling path, and output feature representation levels. Introducing state labels and multi-source covariate features into the input structure enhances the model's adaptability to trend changes under different operating conditions. The modeling structure incorporates multi-scale sliding windows and parallel recursive networks to improve the ability to capture change patterns at different time granularities. The output structure is supplemented with structural indicators such as trend direction, intensity, and inflection point location, providing multi-dimensional feedforward information support for the control parameter generation module. The overall solution enhances the engineering value of trend prediction results in terms of structural clarity and the usability of control command generation, effectively improving the model's generalization ability and controllability in complex dynamic operating scenarios.

[0117] S5. Generate control parameters based on the prediction results and state classification results, and construct adjustment instructions.

[0118] The system performs joint logical judgment on trend characteristics in the prediction results (such as the rate of increase of target parameters, predicted maximum value, and trend inflection point judgment) and state identification results (such as stable operation state, disturbance state, and high temperature risk state), and generates physically meaningful adjustment quantities according to preset control rules. For example, if the prediction shows that T4 will rise by more than 4°C in the next 3 seconds, and the current state is stable, the system determines that it may enter a rising disturbance state and should reduce the fuel injection rate by 2% in advance; if the current state is a disturbance state, the guide vane angle can be adjusted by 0.5° to coordinate pressure balance control.

[0119] The control parameter generation methods include the following three strategies: rule-driven strategy table: defining corresponding adjustment actions for each state classification result and trend combination; multivariate mapping model: using empirical regression, interpolation table or small neural network to map the input trend vector into the adjustment vector; self-learning weight adjustment mechanism: introducing a parameter correction module based on feedback error to correct the adjustment amplitude coefficient online and improve adaptability.

[0120] The process of constructing adjustment instructions includes: encoding the adjustment object (such as the fuel injection system or guide vane control unit); filling in control parameter values ​​(such as the percentage change in opening and the amount of angle adjustment); specifying the execution time range or refresh cycle; packaging it into a standard format structure (such as binary command frame, OPC-UA control node call, or PLC register write structure); and marking the instruction priority and control mode (such as "predictive drive type").

[0121] The final generated adjustment command will be written to the control interface buffer and refreshed to the underlying control system according to a set period (e.g., every 100ms). The system also supports feedback acquisition after the adjustment action is executed (e.g., whether the change in T4 has reached the expected value), and generates a model or dynamically corrects the control strategy mapping rules based on the deviation correction control parameters.

[0122] S6. Execute adjustment instructions to obtain operational feedback data during the control execution process, compare the deviation with the prediction results, and optimize and iterate the control instructions based on the comparison results.

[0123] After generating control parameters and constructing adjustment commands, the system pushes these commands to the execution module in the gas turbine control system. During control execution, the system needs to simultaneously collect relevant operational feedback data to compare the execution effect with the original predicted trend and determine whether the adjustment behavior has achieved the expected goal. If there is a significant deviation, the control strategy needs to be optimized or iteratively adjusted to form a real-time self-closed-loop correction mechanism.

[0124] "Acquiring operational feedback data during the control execution process by executing adjustment commands" refers to the system continuously collecting the actual response values ​​of the target monitoring parameters at a fixed sampling period (e.g., every 100ms) within a duration (e.g., 3 seconds) after the adjustment command takes effect via a real-time data acquisition interface. These parameters generally correspond to the adjustment target; for example, after adjusting the fuel injection rate, the actual change value of T4 is mainly collected.

[0125] "Comparing the deviation with the prediction results" means that the system will compare the trend results output during the prediction phase (such as the T4 prediction sequence within the next 3 seconds) with the collected measured value sequence point by point, and calculate key trend deviation indicators, including:

[0126] The full sequence average error (such as mean squared error MSE, mean absolute error MAE); maximum value deviation (the difference between the measured maximum T4 and the predicted maximum T4); rise / fall rate deviation (slope difference); trend inflection point offset; whether abnormal trend reversal fluctuations are triggered, etc.

[0127] When the deviation exceeds the preset tolerance range (e.g., MSE > 2.0℃) 2 If the maximum error is greater than 1.5℃, the system will activate the optimization module to iteratively adjust the current control parameter generation logic.

[0128] Based on the comparison results, the system optimizes the iterative control commands by updating the adjustment strategy in one of the following ways: Adjusting the adjustment amplitude coefficient: If the predicted temperature rise is too fast and the actual response is insufficient, the adjustment amplitude of the injection rate is increased; Modifying the adjustment response rule mapping table: Introducing the current error sample and retraining or fine-tuning the adjustment rules; Updating the trend prediction model weights: Correcting part of the prediction path in reverse according to the deviation feedback to reduce future error accumulation; Adaptively selecting the combination of control objects: If adjusting the injection rate alone is insufficient, the guide vane angle is automatically added for fine-tuning.

[0129] Through the above closed-loop optimization process, the system achieves closed-loop control throughout the entire process from adjustment decision to execution feedback, which not only improves the accuracy of control command response but also enhances the system's stable control capability under dynamic disturbances.

[0130] Example 3 is an embodiment of the present invention, which provides a method for intelligent control of gas turbines based on operational data analysis. In order to verify the beneficial effects of the present invention, scientific demonstration is carried out through experiments.

[0131] During the deployment of this invention, the following monitoring parameters were selected as data acquisition objects: compressor inlet temperature (T1); combustion chamber outlet temperature (T4); main shaft first-stage speed (N1); total fuel injection rate of the burner (F); turbine inlet pressure (P3); actual load power (MW); fuel main valve displacement (Xv); intake guide vane opening (IGV); oxygen content feedback (O2), etc.

[0132] The system uses a 100-millisecond sampling period and transmits data in real time via the Modbus TCP protocol through the data link of the GE controller to the edge computing unit deployed on the station control server. The raw data record contains approximately 30 consecutive fields, generating approximately 360,000 records per hour.

[0133] After the data enters the processing flow, a boundary removal strategy is first executed. For example, the upper limit of T4 is set to 1400°C, the upper limit of N1 is 5400rpm, and the volatility of F must not exceed 10% per second. For missing sampling points, a forward interpolation strategy is used to repair them, and a sliding window (10 points) mean filter is used to process short-term noise.

[0134] The system then calculates the moving average, volatility, and incremental indicators for each monitored field, generating derived features. The total feature dimension expands from the original 10 signals to 32 processed fields.

[0135] All processed samples are constructed into a structured two-dimensional dataset, with each row representing a time point and columns representing standardized feature fields. The entire dataset can be used as direct input for the subsequent "state recognition model." This embodiment ensures the integrity and engineering feasibility of the entire process of cleaning, structuring, and modeling preparation of high-frequency operating data from gas turbines.

[0136] In step one, the multidimensional acquisition and standardized preprocessing of gas turbine operating data have been completed. Currently, we select the operating data from 9:00 to 9:15 on June 15, 2025 as the input sample for status identification. The data consists of a structure of 32 feature fields per record, sampled every 100 milliseconds, and a total of 9,000 records were collected.

[0137] The system processes the dataset using a sliding window approach, sliding once every 100 milliseconds. The first 3 seconds (30 records) form a sample window, constructing state identification samples. Each window ultimately extracts the following time-series change characteristics: the maximum rise slope, three-second fluctuation amplitude, and average heating rate of the combustion chamber outlet temperature (T4); the short-term maximum amplitude and first-order derivative change direction consistency of the main shaft speed (N1); the Pearson correlation coefficient between the injection valve opening (F) and T4; the slope comparison between the guide vane angle (IGV) and the turbine inlet pressure (P3); and the sliding hysteresis response time of the system active load (MW) change and fuel injection rate.

[0138] The above indicators are extracted through program encapsulation and combined into a 78-dimensional state feature vector.

[0139] During the model building phase, the system used the operating data from the past three months, selected samples from typical operating periods, and performed K-means clustering modeling. Finally, four state center clusters were obtained, corresponding to: State 0: Start-up and warming state (T4 rises rapidly, F fluctuates greatly, N1 is not stable); State 1: Stable operating state (T4 fluctuates <2℃, F and N1 are strongly correlated); State 2: Wave dynamics (MW disturbances are frequent, O2 content is abnormal); State 3: High temperature abnormal state (T4 is higher than 1320℃, P3 fluctuates frequently).

[0140] The model is permanently deployed in the edge computing unit and has an operating cycle of 100 milliseconds.

[0141] During the real-time operation phase, at 9:07:14 on June 15, 2025, the system detected that the current T4 heating rate in the sliding window was 0.3℃ / s, the spindle speed fluctuation was ±6rpm, the correlation between the oil injection rate and temperature was 0.93, the overall feature vector and the Euclidean distance of the center of state 1 were 0.42, and the corresponding recognition confidence was 0.88.

[0142] Therefore, the system outputs the following structured state classification results: Current state number: 1 (stable operating state); identification confidence: 0.88; state generation timestamp: 2025-06-15 09:07:14.300; state description: T4 fluctuation is low, N1 is stable, and fuel injection and temperature are highly synchronized.

[0143] The status recognition result is automatically cached in the status cache module and pushed to the trend prediction submodule as the basic input for judging the working condition during prediction.

[0144] At 09:07:14 on June 15, 2025, the system identified the gas turbine as being in "State 1 (Stable Operation)". This state number was output by the state identification process and included a confidence value of 0.88. The system then integrated this state number with the structured operational data set, adding a "run_state" field to the original data structure, so that each sampled record was accompanied by a corresponding operational state label.

[0145] The structure of the merged dataset is as follows: each record corresponds to a sampling time point, including the original collected fields (such as T4, N1, F, MW, etc.), derived features (such as T4 rate of change, N1 standard deviation, P3 fluctuation amplitude, etc.), and the current running state number (such as state 1). The entire dataset still maintains a time resolution of one record every 100ms, and samples are constructed using a sliding window method.

[0146] The trend prediction processing uses a sliding window approach to extract input samples. Currently, it's set to extract the most recent 5 seconds of data (50 records) per round as input samples, aiming to predict the trend of combustion chamber outlet temperature (T4) over the next 3 seconds. Each sample contains the following: a time dimension of length 50; a feature field dimension of 33 (32 original and derived features + 1 state number); an input tensor structure of 50×33; and the output is the predicted T4 value for the next 30 time points (one point every 100ms).

[0147] In this embodiment, a single-layer LSTM network is selected as the trend prediction model. The model consists of the following structure: Input layer: time step 50, input dimension 33; Hidden layer: LSTM unit number 64; Output layer: regression layer outputting the predicted T4 value for the next 30 points; Loss function: mean squared error (MSE); Optimizer: Adam; Training is performed using data from the past two months of steady-state operation, with corresponding samples selected, and the training set contains 60,000 samples.

[0148] After model training is complete, the model is exported in ONNX format and deployed to an edge computing server (Advantech MIC-770 series), with the runtime set to predict once every 100ms. During deployment, a model inference threshold is set: if the maximum predicted T4 value exceeds 1270℃ within the next 3 seconds, or its heating rate exceeds 1.5℃ / s, it will be marked as "heating too fast" and used as a trigger condition for adjustment commands.

[0149] In actual operation, the system performs trend prediction at 09:07:14.300. The model output shows that T4 will rise from 1262.3℃ to 1266.4℃ in the next 3 seconds, with a heating rate of 1.36℃ / s. The predicted maximum value is lower than the trigger threshold. The system records the result and proceeds to the next step of control parameter generation logic.

[0150] In step three, the system has completed the fusion processing of structured running data and running status labels within the current time window, and constructed a trend prediction sample in standard input format. The sample length is 5 seconds, with a total of 50 records. Each record contains 33-dimensional feature fields, including original and derived features such as T4, N1, F, MW, and P3, as well as a running status label (number 1, corresponding to the stable running state).

[0151] To enhance the predictive model's adaptability to trend behaviors under different operating states, the system employs the state-aware gated recursive network described in this invention to construct the predictive model. In this model: the input is a historical feature sequence (length 50) within a sliding time window. Each feature includes the historical value of the target parameter T4, its first derivative, variance, slope, and the changes in related covariates such as injection rate F, exhaust pressure P3, and spindle speed N1. Simultaneously, the operating state labels corresponding to the past 50 data points (all in state 1) are input. Furthermore, derived features at three time scales—1 second, 3 seconds, and 5 seconds—are introduced to form a multi-scale trend embedding structure.

[0152] The model network structure is configured as follows: Number of multi-scale input channels: 3 (corresponding to 1-second, 3-second, and 5-second sliding windows respectively); each channel is processed using GRU units, and the hidden layer dimension is 64; the state-aware fusion layer is a fully connected gated fusion module that generates a context weight vector by combining state labels; the output layer is a 30-dimensional fully connected regression layer that corresponds to the T4 prediction value for the next 3 seconds.

[0153] The training set was derived from the last 30 days of operation logs, with a total of 90,000 samples extracted by state, of which 52% were stable state samples. The model training optimization objective was a joint loss function, including prediction error (MSE) and trend structure matching degree (mean square of the difference between the slope of the predicted curve and the slope of the true curve).

[0154] After deployment, the system inputs the current window data sample for prediction at 09:07:14.300, and the output results are as follows: Current T4 value: 1262.3℃; Predicted maximum value (within the next 3 seconds): 1266.1℃; Predicted minimum value: 1261.7℃; Maximum rise slope (ΔT4 / Δt): 1.32℃ / s; Trend determination label: stable slight rise state; Whether the control threshold is triggered: No (heating rate <1.8℃ / s, temperature peak <1270°C); The system encapsulates this trend prediction sequence together with the above trend feature structure and uses it as the input basis for the control parameter generation module. The result is synchronously written to the trend buffer, and the original input window and prediction sequence are retained for use in the model's posterior accuracy evaluation.

[0155] The system successfully output the predicted curve of T4 (combustion chamber outlet temperature) for the next 3 seconds at 09:07:14.300 on June 15, 2025. The results show that the current T4 is 1262.3°C; the predicted maximum is 1266.1°C; the average heating rate is 1.32°C / s; and the overall trend of the predicted curve is "mild increase".

[0156] Meanwhile, the operating state identified in step two is "State 1: Stable operating state", with a state identification confidence level of 0.88.

[0157] Based on the current prediction results and operating status, the system calls the state trend mapping strategy table in the control parameter generation module, and matches the following control rules with reference to Table 1:

[0158] Table 1 Control Rules Table

[0159]

[0160] Based on the current monitoring values ​​and referring to Table 2, the system generates the following control parameter adjustment suggestions:

[0161] Table 2 Parameter Adjustment Suggestions

[0162]

[0163] The above parameters are written as control commands into the control register channel of the GE MarkVIe platform and marked as "predictive feedforward triggered regulation". The commands are executed immediately after being issued, with a duration of 3 seconds.

[0164] In addition, the system establishes a control feedback monitoring task, which collects the measured value of T4 every 100 milliseconds in the next 3 seconds and compares it with the original predicted value point by point: if the maximum measured value of T4 is ≤1266.1℃ and the heating rate drops to ≤1.0℃ / s, the adjustment is considered to be effective; otherwise, the system will add the current execution effect feedback to the next round of trend prediction and dynamically correct the fuel injection rate adjustment strategy (such as adding a -0.5% opening correction amount).

[0165] The adjustment history is fully recorded, including key fields such as control timestamp, adjustment parameters, execution feedback, response trend, and deviation, which are used for the accumulation of training data for subsequent trend prediction models and control mapping rules.

[0166] At 09:07:14.350 on June 15, 2025, the system issued a fuel injection rate adjustment command (reducing it from 78.2% to 76.7%, with an adjustment cycle of 3 seconds). After the control platform executed the command, the system began collecting the measured T4 value every 100ms from 09:07:14.400, continuously recording 30 points.

[0167] In the original predicted sequence, T4 should rise from 1262.3℃ to 1266.1℃ within 3 seconds, with a maximum rise rate of 1.32℃ / s. Referring to Table 3, the actual acquisition results are analyzed as follows:

[0168] Table 3. Data Collection Results Display Table

[0169]

[0170]

[0171] Because the MSE exceeds the set threshold of 2.0℃ 2 The system triggers a fine-tuning mechanism: adjusting the strategy coefficients, increasing the reduction ratio of the fuel injection rate from 1.5% to 2.0%; updating the adjustment parameters for the corresponding range of "steady state / heating rate 1.2~1.5℃ / s" in the control rule table; marking the current sample as a "trend deviation positive amendment example" and writing it into the strategy training database.

[0172] The next round of adjustments will be restarted after the predictive model is updated, and control accuracy will be improved by issuing new adjustment instructions.

[0173] Example 4 is an embodiment of the present invention, which provides a gas turbine intelligent control method system based on operational data analysis, including a data collection and preprocessing module, a state classification module, a data merging module, a model building module, a control command module, and an optimization iteration module;

[0174] The data collection and preprocessing module collects multidimensional monitoring data during the operation of the gas turbine and constructs an initial operating dataset in chronological order. It then merges and preprocesses the data to generate a data set with a unified time reference and feature fields.

[0175] The state classification module inputs the data set into the operating state identification process, identifies the current operating state of the gas turbine based on the temporal change characteristics and correlation patterns of the monitoring data, and outputs the corresponding state classification results.

[0176] The data merging module merges the state classification results with the data set and performs trend prediction processing based on the time series change patterns.

[0177] The model building module selects target monitoring indicators related to gas turbine operation control in the trend prediction processing, constructs a prediction model to depict the changing trend within a future target time period, and outputs the prediction results.

[0178] The adjustment instruction module generates control parameters and constructs adjustment instructions based on the prediction results and state classification results.

[0179] The optimization iteration module executes adjustment instructions to obtain operational feedback data during the control execution process, compares the deviation with the prediction results, and optimizes and iterates the control instructions based on the comparison results.

[0180] This embodiment also provides an electronic device applicable to a gas turbine intelligent control method based on operational data analysis, comprising: 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 realize the gas turbine intelligent control method based on operational data analysis as proposed in the above embodiment.

[0181] This embodiment also provides a storage medium storing a computer program that, when executed by a processor, implements a gas turbine intelligent control method based on operational data analysis as proposed in the above embodiments.

[0182] The storage medium proposed in this embodiment belongs to the same inventive concept as the method for intelligent control of gas turbines based on operational data analysis 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.

[0183] 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, but in many cases the former is a better implementation method. 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.

[0184] 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 method for intelligent control of gas turbines based on operational data analysis, characterized in that: include, Multidimensional monitoring data is collected during the operation of the gas turbine and constructed into an initial operating dataset in chronological order. Preprocessing operations are then combined to generate a dataset with a unified time reference and feature fields. The data set is input into the operation status identification process. Based on the time-series change characteristics and correlation patterns of the monitoring data, the current operation status of the gas turbine is identified, and the corresponding status classification results are output. The state classification results are merged with the dataset, and trend prediction processing based on the time series change patterns is performed. In trend prediction processing, target monitoring indicators related to gas turbine operation control are selected, a prediction model is constructed to characterize the changing trend within the target time period in the future, and the prediction results are output. Control parameters are generated based on the prediction results and state classification results, and adjustment instructions are constructed. The system executes adjustment commands to obtain operational feedback data during the control execution process, compares the deviation with the prediction results, and optimizes and iterates the control commands based on the comparison results.

2. The intelligent control method for gas turbines based on operational data analysis as described in claim 1, characterized in that: The operational status identification process includes, The dataset is used to construct a sliding time window sample according to a set sampling period. For each window sample, the slope, magnitude of change and segment mean of the monitoring data in time sequence are extracted as time series features. The correlation between the fields in the structured dataset is combined to form a state recognition input vector. The data is then processed by a classification model to output the state classification result corresponding to the current window.

3. The intelligent control method for gas turbines based on operational data analysis as described in claim 2, characterized in that: The process of performing trend prediction based on time series variation patterns includes, The status classification results are synchronized and aligned with the dataset according to the timestamp. The status type field is appended to the record data of the corresponding time period to construct a trend modeling input sample sequence containing monitoring parameter field and status label field. The sample sequence is then used as the input content of the trend prediction model.

4. The intelligent control method for gas turbines based on operational data analysis as described in claim 3, characterized in that: The trend prediction processing based on time series variation patterns includes... For the target monitoring indicators in the input sample sequence of trend modeling, the numerical change trajectory and parameter structure of the current indicator in a continuous time period are extracted, and a prediction model input path is constructed to represent the direction and magnitude of trend change. The predicted value sequence covering each time point in the target prediction period is output through the current path.

5. The intelligent control method for gas turbines based on operational data analysis as described in claim 4, characterized in that: The process of generating control parameters based on prediction results and state classification results includes, Based on the trend output structure of the prediction results and the state information in the state classification results, the corresponding adjustment condition structure is selected in the control logic structure to generate a set of control parameters. The set of control parameters is constructed into adjustment commands, which are then sent to the control system for execution. During the execution of the adjustment instructions, operational feedback data is collected, and the operational feedback data is time-aligned with the prediction results to form a feedback comparison data structure. The control deviation index is determined based on the feedback comparison data structure, and the set of control parameters is adjusted according to the control deviation index.

6. The intelligent control method for gas turbines based on operational data analysis as described in claim 4, characterized in that: The trend prediction process also includes, The state type in the state classification results is used as the structural control condition. Multiple modeling structural paths are set in the prediction model, each corresponding to a different state type. During model execution, the corresponding path is selected based on the current state type, and the trend modeling input sample sequence is input into the selected path. Based on the generated trend prediction sequence, the change direction segments and trend intensity indicators covering the target prediction time period are extracted to construct a trend combination condition structure, which serves as the input basis for the control parameter generation logic.

7. The intelligent control method for gas turbines based on operational data analysis as described in claim 6, characterized in that: The generation of control parameters includes a control strategy path selection process based on the combination of state type and trend conditions in the state classification results. The control strategy path selection process includes combining the state type in the state classification results with the trend change direction and change intensity indicators in the trend combination condition structure to form a state-trend joint identifier. Several strategy paths are preset in the control strategy structure, and each strategy path corresponds to a state-trend joint identifier; Match the corresponding strategy path based on the current state-trend joint identifier, and select control parameters from the matched strategy path to generate logic; A set of control parameters is constructed using control parameter generation logic; The control parameter set includes control execution module fields, parameter adjustment fields, and control priority fields, and is encapsulated as structured adjustment instructions to be sent to the control system.

8. A gas turbine intelligent control method system based on operational data analysis, employing the gas turbine intelligent control method based on operational data analysis as described in any one of claims 1 to 7, characterized in that, It includes: a data collection and preprocessing module, a state classification module, a data merging module, a model building module, an adjustment instruction module, and an optimization iteration module; The data collection and preprocessing module collects multi-dimensional monitoring data during the operation of the gas turbine, and constructs an initial operating dataset in chronological order. It then merges and preprocesses the data to generate a data set with a unified time reference and feature fields. The state classification module inputs the data set into the operating state identification process, identifies the current operating state of the gas turbine based on the temporal change characteristics and correlation patterns of the monitoring data, and outputs the corresponding state classification results. The data merging module merges the state classification results with the data set and performs trend prediction processing based on the time series change pattern. The model building module selects target monitoring indicators related to gas turbine operation control in the trend prediction processing, builds a prediction model to depict the changing trend in the future target time period, and outputs the prediction results. The adjustment instruction module generates control parameters and constructs adjustment instructions based on the prediction results and state classification results. The optimization iteration module executes adjustment instructions to obtain operational feedback data during the control execution process, compares the deviation with the prediction results, and optimizes and iterates the control instructions based on the comparison results.

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 intelligent control method for gas turbines based on operational data analysis as described in 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 the processor, it implements the steps of the intelligent control method for gas turbines based on operational data analysis as described in any one of claims 1 to 7.

Images