Gas turbine power plant time series prediction method and system based on deep learning

Abstract

The present application relates to the technical field of power generation capacity prediction of gas turbine power plants, and specifically discloses a time series prediction method and system for gas turbine power plants based on deep learning. The method is aimed at the technical problem of inaccurate power generation capacity prediction caused by sudden change of fuel characteristics during fuel type switching of the gas turbine. The fuel switching instruction signal and the operation time series data are received in real time, and are fused with the fuel chemical characteristic parameters to generate an input tensor. A Transformer deep learning model enhanced by an attention mechanism is used to quantify the fuel switching influence factor. The combustion stability index is dynamically calculated in combination with the temperature and pressure sequences. Based on the index, a dynamic correction value of the power generation capacity is generated, and is fused with the basic prediction value to output the final prediction result. The system includes data acquisition, data fusion, data processing and power generation capacity prediction modules. The present application can effectively improve the accuracy and real-time performance of power generation capacity prediction in the fuel switching scenario, and is suitable for operation optimization and dispatching management of gas turbine power plants.

Description

Technical Field

[0001] This invention relates to the field of power plant power generation prediction, and in particular to a deep learning-based time series prediction method and system for gas turbine power plants. Background Technology

[0002] As a key component of the power system, the accurate prediction of the power generation of gas turbine power plants is of great significance to the stability of the power grid, the efficiency of energy dispatch, and the economic operation of the power plants.

[0003] With the advancement of artificial intelligence technology, time series forecasting methods based on deep learning are increasingly being applied to the field of power plant operation optimization. These methods analyze historical operating data (such as temperature, pressure, and flow rate) and utilize neural network models (such as LSTM, GRU, and Transformer) to capture the nonlinear time-series characteristics of the data, enabling short-term or real-time forecasts of power generation. However, gas turbine power plants face dynamic fuel type switching conditions in actual operation, such as switching from the default fuel—natural gas—to alternative fuels like biomethane, leading to transient changes in the combustion process. This places higher demands on the real-time performance and adaptability of the forecasting models.

[0004] Existing power generation prediction schemes for gas turbine power plants mostly employ static deep learning models, whose input features are typically limited to fixed operating parameters (such as combustion chamber temperature, exhaust pressure, and fuel flow rate), failing to adequately consider the dynamic changes in fuel chemical properties. Specifically, these models (e.g., prediction systems based on LSTM or CNN) use historical data to learn the mapping relationship between power generation and input features during the training phase. However, in actual deployment, when fuel type changes, key parameters such as fuel calorific value, density, and composition ratio undergo abrupt changes, leading to alterations in combustion efficiency, temperature distribution, and pressure characteristics. Because existing models lack a real-time fuel parameter update mechanism, they cannot capture these abrupt changes at the moment of switching, resulting in prediction logic still based on assumptions about the old fuel and significant deviations between output results and actual values. Furthermore, traditional methods lack dedicated processing modules for fuel switching events, failing to quantify the impact of the switch or adjust the prediction curve through adaptive mechanisms, causing prediction errors to persist during the switching transition period.

[0005] The inventors of this application have discovered the following technical problems with the above-mentioned technology:

[0006] Existing deep learning models do not incorporate dynamic parameters of fuel chemical properties in power generation prediction. When fuel type is switched, the model input cannot reflect key changes such as fuel calorific value and composition ratio, resulting in significant deviations in power generation prediction output near the switching point. Summary of the Invention

[0007] This invention provides a time series prediction method and system for gas turbine power plants based on deep learning. The technical problem to be solved is: to construct a real-time adaptive power generation prediction model in the scenario of dynamic switching of gas turbine fuel type, and to solve the prediction inaccuracy problem caused by sudden changes in fuel characteristics.

[0008] To achieve the above-mentioned objectives, the technical solution adopted by this invention is as follows:

[0009] A deep learning-based time series forecasting method for gas turbine power plants includes the following steps:

[0010] Step S10: Receive fuel switching command signal from the gas turbine control system; acquire gas turbine operation time series data in real time, the time series data including combustion chamber temperature sequence, exhaust pressure sequence, fuel calorific value sequence and fuel flow sequence, and all data are stored in an industrial database with a unified timestamp;

[0011] Step S20: The time series data is fused with the corresponding parameters in the fuel chemical property parameter database to generate an input tensor; the input tensor is input into a deep learning model with attention mechanism enhancement. The model is based on the Transformer architecture and calculates the adaptive correlation weights between time steps to quantify the fuel switching impact factor. The fuel switching impact factor is calculated based on the abrupt deviation between the temperature sequence and the calorific value sequence in the input tensor.

[0012] Step S30: Receive the fuel switching influencing factor through the fuel type identification module, and calculate the combustion stability index in real time by combining the temperature sequence and pressure sequence in the time series data. The combustion stability index is dynamically generated based on the weighted sum of the standard deviation of the temperature sequence fluctuation and the rate of change of the pressure sequence. The standard deviation of the fluctuation is calculated from the sliding window variance of the temperature sequence, and the rate of change is obtained by the differential operation of the pressure sequence.

[0013] Step S40: Generate a dynamic correction value for power generation after fuel switching based on the combustion stability index. The dynamic correction value is used to adjust the slope of the prediction curve near the time point triggered by the fuel switching command signal through a linear interpolation method. The dynamic correction value is then fused with the output value of the basic power generation prediction model to output the real-time power generation prediction result.

[0014] The beneficial effects of this invention are as follows:

[0015] 1. In the scenario of dynamic switching of gas turbine fuel type, by fusing fuel chemical characteristic parameters and operating time series data in real time, and quantifying the influencing factors of fuel switching based on a deep learning model enhanced by attention mechanism, the problem of inaccurate power generation prediction caused by sudden changes in fuel characteristics is effectively solved.

[0016] 2. By constructing a knowledge graph of gas turbine fuel switching and performing reasoning based on graph neural networks, a deep integration of historical switching experience and real-time data was achieved, which enhanced the model's generalization ability and interpretability for complex switching scenarios, reduced dependence on external labeled data, and improved the system's adaptability and robustness under varying operating conditions.

[0017] 3. The wolf pack algorithm is introduced to dynamically optimize the model hyperparameters, enabling adaptive adjustment of key parameters such as learning rate, regularization coefficient and number of attention heads. This effectively overcomes the problems of slow convergence and easy getting trapped in local optima in traditional fixed parameter models during fuel switching, and improves the model's response speed and stability under sudden change conditions. Attached Figure Description

[0018] Figure 1 This is a logical schematic diagram of the prediction method of the present invention;

[0019] Figure 2 This is a schematic diagram of the prediction system of the present invention. Detailed Implementation

[0020] First, an overview of the inventive concept is provided: This invention addresses the technical challenge of existing deep learning prediction models failing to respond in real time and resulting in severely inaccurate power generation predictions during dynamic fuel type switching in gas turbine power plants. It proposes a time-series prediction method and system that integrates multi-source data and an adaptive learning mechanism. The scheme collects fuel switching commands and operating parameters in real time, constructing an input tensor using a fuel chemical characteristic database. It employs an attention-enhanced Transformer model to quantify fuel switching influencing factors and dynamically calculates the combustion stability index based on temperature and pressure sequences. Furthermore, it introduces knowledge graph reasoning and wolf pack algorithm optimization to achieve deep integration of historical experience and real-time status, as well as adaptive adjustment of model parameters. Finally, it outputs accurate power generation prediction results through a dynamic correction mechanism. This method effectively solves the problems of prediction lag and bias accumulation in fuel switching scenarios, significantly improving the real-time performance, adaptability, and accuracy of the prediction model.

[0021] Next, in order to make the content of the present invention easier to understand, the technical solutions in the embodiments of the present invention will be clearly and completely described below with reference to the accompanying drawings.

[0022] Example 1

[0023] A deep learning-based time series forecasting method for gas turbine power plants includes the following steps:

[0024] Step S10: Receive a fuel switching command signal from the gas turbine control system, the command signal indicating a switch from the first fuel to the second fuel; simultaneously, acquire real-time gas turbine operating time-series data from the sensor network, the time-series data including combustion chamber temperature sequence, exhaust pressure sequence, fuel calorific value sequence, and fuel flow rate sequence, wherein the fuel calorific value sequence is synchronously measured by a fuel composition analyzer, the sensor network collects temperature sequence via thermocouples, pressure sequence via pressure transmitters, and fuel flow rate sequence via flow meters, and all data are stored in an industrial database with a unified timestamp;

[0025] Step S10 serves as the data input for the entire prediction method, its core function being the real-time capture of the trigger signal for the fuel switching event and related operating parameters. The fuel switching command signal is a clear instruction issued by the gas turbine control system, indicating a switch from conventional fuels such as natural gas to alternative fuels such as biomethane. The sensor network consists of various sensors distributed throughout key parts of the gas turbine. Thermocouples are used to accurately measure temperature changes within the high-temperature environment of the combustion chamber, pressure transmitters monitor pressure fluctuations in the exhaust system, and flow meters record the amount of fuel flowing through the pipeline per unit time. The fuel calorific value sequence is a crucial parameter, representing the heat released by the complete combustion of a unit mass of fuel; differences in this sequence lead to variations in combustion efficiency.

[0026] Step S20: The time series data is fused with the corresponding parameters in the fuel chemical characteristic parameter database. Fuel chemical characteristic parameters are the inherent properties of the raw materials, including calorific value, density, and chemical composition ratio. The fusion process inputs the temperature sequence into the combustion efficiency calculation module to generate efficiency mutation characteristics. This involves calculating the real-time combustion efficiency of the gas turbine based on the fuel chemical characteristics and combustion chamber temperature sequence, and extracting the combustion efficiency mutation characteristics caused by fuel switching (i.e., the magnitude and rate of change in combustion efficiency before and after switching). The pressure sequence is input into the preliminary combustion stability assessment module to generate pressure fluctuation characteristics. The efficiency mutation characteristics are then... The pressure fluctuation characteristics, along with the calorific value sequence and fuel flow rate sequence, constitute the input tensor. This input tensor is fed into a deep learning model enhanced with an attention mechanism. Based on the Transformer architecture, the model calculates an adaptive correlation weight matrix between time steps to locate the abrupt change time points in the fuel calorific value sequence. A feature mask is applied to extract the abrupt change correlation weights of the temperature sequence and the calorific value sequence from this adaptive correlation weight matrix. The abrupt change correlation weights are multiplied by the pressure fluctuation characteristics and input into a Sigmoid activation function for normalization to obtain a fuel switching impact factor, representing the instantaneous impact of fuel type abrupt changes on combustion efficiency.

[0027] Specifically, the combustion efficiency calculation module uses a thermal balance model, and the real-time combustion efficiency calculation formula is as follows: in, The temperature series was calculated using a thermodynamic model. The efficiency mutation characteristics are calculated based on the fuel calorific value sequence and fuel flow rate sequence; the efficiency mutation characteristics are obtained by calculating the magnitude and rate of change of combustion efficiency before and after the switch, i.e., the magnitude of change. in, and These represent the combustion efficiency before and after the switch, respectively. For time intervals.

[0028] The preliminary combustion stability assessment module generates pressure fluctuation characteristics using statistical methods. Specifically, the pressure fluctuation characteristics are generated by calculating the sliding window variance of the pressure sequence. And the peak factor is obtained; in, For pressure sequence values, This represents the average pressure within the sliding window. The window size is used; the peak factor is calculated using the following formula: Pressure fluctuation characteristics are represented by a weighted combination of variance and peak factor, such as: pressure fluctuation characteristics are w1 * N(σ²) + w2 * N(PF), where N(·) represents the normalization function;

[0029] Specifically, for the fuel switching impact factor: the calculated mutation correlation weight, which characterizes the contribution of the historical temperature sequence to the current fuel calorific value mutation; the pressure fluctuation characteristics calculated from the pressure sequence; the mutation correlation weight and the pressure fluctuation characteristics are multiplied to obtain an intermediate product; the intermediate product is input into a nonlinear activation function for mapping and range constraint, and the output value of the function is the fuel switching impact factor.

[0030] Step S30: If the fuel switching impact factor is greater than a preset threshold, then,

[0031] Combustion stability index is calculated in real time by combining the temperature and pressure sequences in the time series data. The combustion stability index is dynamically generated based on the weighted sum of the standard deviation of the temperature sequence fluctuation and the rate of change of the pressure sequence. The standard deviation of the fluctuation is calculated from the sliding window variance of the temperature sequence, and the rate of change is obtained from the differential operation of the pressure sequence.

[0032] Otherwise, maintain the default stability index.

[0033] Step S30 quantifies the stability of the combustion process, providing crucial information for subsequent power generation adjustments. The combustion stability index is a dynamic indicator used to assess the smoothness of the gas turbine combustion process after fuel switching. The standard deviation of the temperature series reflects the stability of the combustion temperature; a larger standard deviation indicates more severe temperature fluctuations. The rate of change of the pressure series is obtained through differential calculation (i.e., subtracting the previous pressure value from the current pressure value), indicating the speed of pressure change. These two indicators are combined into the combustion stability index using a weighted sum, with the weighting coefficients determined empirically based on combustion instability events in historical data. The sliding window variance calculation uses a fixed-time window, continuously updating the variance value as new data is added to ensure that the latest changes in combustion status are captured. This step transforms the traditional static combustion stability assessment into a dynamic real-time assessment, particularly suitable for transient processes like fuel switching. When switching fuel from natural gas to biomethane, the slower combustion rate of biomethane may cause a decrease in the combustion stability index, indicating a need to adjust the power generation forecast.

[0034] Step S40: Generate a dynamic correction value for power generation after fuel switching based on the combustion stability index. The dynamic correction value is used to adjust the slope of the prediction curve near the time point triggered by the fuel switching command signal through a linear interpolation method. The dynamic correction value is then fused with the output value of the basic power generation prediction model to output the real-time power generation prediction result.

[0035] Step S40 is the final output stage of this invention, realizing dynamic correction of power generation prediction. The dynamic correction value is an adjustment calculated based on the combustion stability index, used to compensate for prediction deviations caused by fuel switching. The linear interpolation method creates a smooth transition region near the fuel switching point, avoiding abrupt changes in the prediction results. Specifically, within a preset time window before and after the switching point, the system adjusts the slope of the prediction curve according to the changing trend of the combustion stability index, enabling the prediction value to smoothly transition to a new stable state. The basic power generation prediction model is a traditional model that does not consider changes in fuel parameters, and its prediction results will have a large deviation during fuel switching. By fusing the dynamic correction value with the basic prediction value, the system can output a more accurate real-time power generation prediction. This step solves the prediction lag problem caused by fuel switching in the prior art, realizes precise energy dispatch, and avoids energy waste and excessive carbon emissions caused by inaccurate predictions.

[0036] Example: When the system detects that the combustion stability index drops to 0.65, the dynamic correction value is calculated to be -15MW; within a 10-second window before and after the switching point (14:00), the slope of the prediction curve is adjusted from 0 to -1.5MW / s; the correction value of -15MW is merged with the 300MW predicted by the basic model to output a real-time power generation prediction of 285MW, which accurately reflects the actual power generation capacity after the biomethane fuel switch.

[0037] Furthermore, step S10 includes the following specific steps:

[0038] Step S11: Measure the chemical composition ratio of the fuel in real time using an online gas chromatograph, and match the measurement results with a pre-stored standard fuel calorific value table to calculate the calorific value sequence;

[0039] Step S11 addresses the issue of accurately measuring the calorific value of fuel. An online gas chromatograph is an instrument capable of analyzing gaseous components in real time. By separating various components in the fuel (such as methane, ethane, propane, etc.) and measuring their concentration ratios, it obtains the precise chemical composition of the fuel. A standard fuel calorific value table stores calorific value data corresponding to different component ratios. For example, the calorific value of pure methane is 35.8 MJ / m³, while the calorific value of methane containing 10% ethane is approximately 37.2 MJ / m³. By matching the measured component ratios with the standard table, the system can dynamically calculate the accurate calorific value of the current fuel. This is crucial for subsequent predictions, as the differences in calorific values ​​between different fuels directly affect combustion efficiency and power generation. Traditional methods often use fixed calorific values ​​for prediction, which cannot be updated in a timely manner when switching fuels, leading to prediction errors.

[0040] Step S12: The mass flow meter collects the fuel flow sequence at a preset sampling interval and aligns it with the calorific value sequence through a time synchronization protocol to ensure that the two are stored with the same timestamp in the industrial database;

[0041] Step S12 ensures precise time alignment between fuel flow rate data and calorific value data. Mass flow meters directly measure the mass of fuel flowing through the pipe per unit time (in kg / s), providing a more accurate reflection of actual fuel energy input than volumetric flow meters. Preset sampling intervals (e.g., 1 second) guarantee high temporal resolution, capturing rapidly changing fuel flow rates. Time synchronization protocols (e.g., IEEE 1588 Precision Time Protocol) ensure precise time alignment of data from different sensors, avoiding data misalignment due to clock discrepancies. This precise alignment is crucial for subsequent calculations of total energy input, as total energy = calorific value × flow rate. If the two are not synchronized, energy calculation errors will occur, affecting the accuracy of power generation predictions.

[0042] Example: The mass flow meter records the natural gas flow rate as 280 kg / s at 1-second intervals; the time synchronization protocol ensures that this flow rate data corresponds precisely to the calorific value of 38.5 MJ / m³ (converted to mass calorific value) at the same time; when switching to biomethane, the flow rate is adjusted to 295 kg / s to compensate for the lower calorific value, and the system confirms that this change occurs at the same time point through time alignment.

[0043] Step S13: After performing anti-interference filtering on the temperature sequence output by the thermocouple, input it into the industrial database. Perform normalization on the pressure sequence output by the pressure transmitter to a preset value range. The normalization process uses the maximum and minimum values ​​of historical pressure data as boundaries.

[0044] Step S13 preprocesses the raw sensor data to improve data quality. Anti-interference filtering (such as Kalman filtering or wavelet denoising) is used to eliminate random noise and interference signals in thermocouple measurements. Normalization transforms the pressure sequence to a uniform numerical range (e.g., 0-1). Using the maximum / minimum values ​​of historical data as boundaries preserves the relative trend of data change while avoiding the impact of data of different magnitudes on model training. The normalization formula is:

[0045] ;

[0046] Where P is the original pressure value. and These are the historical minimum and maximum stress values. This preprocessing ensures the quality and consistency of the input data, which is the foundation for accurate subsequent predictions.

[0047] Step S14: Before storing the time series data, perform missing value imputation. The missing value imputation is based on the linear extrapolation algorithm of adjacent time points to ensure data continuity and dynamically verify the rationality of the imputation result. If the imputation deviation exceeds the preset tolerance, the imputation algorithm is re-executed.

[0048] Step S14 addresses the common problem of missing data in industrial environments. Factors such as sensor malfunctions and communication interruptions can lead to missing points in time series data, which, if left unaddressed, will affect subsequent analysis. The linear extrapolation algorithm estimates the missing value using linear interpolation based on known data points before and after the missing point. For example, if the value at time t-1 is X1 and the value at time t+1 is X2, then the missing value at time t is interpolated as (X1+X2) / 2. A dynamic verification mechanism checks the reasonableness of the interpolated value with surrounding data. If the interpolated value deviates too much from the expected value based on trend prediction (exceeding the preset tolerance), a more complex interpolation algorithm (such as spline interpolation) is used to recalculate. This process ensures the integrity and continuity of the time series and avoids prediction errors caused by missing data.

[0049] Step S15: The fuel switching command signal is transmitted via the Modbus TCP protocol. The transmission process includes a signal verification mechanism. Data fusion operation is triggered only when the verification passes. The signal verification mechanism verifies the integrity of the command signal through a cyclic redundancy check code.

[0050] Step S15 ensures reliable transmission and accurate identification of the fuel switching command signal. Modbus TCP is an industry-standard communication protocol suitable for device communication in factory automation environments. The signal verification mechanism uses a Cyclic Redundancy Check (CRC) code to check the integrity of transmitted data. CRC is an error-detecting code that can detect bit errors that may occur during data transmission. Only when the CRC check passes, confirming that the command signal is complete and error-free, will the system trigger subsequent operations.

[0051] Example: The control system sends a "Fuel switch to biomethane" command, which includes the target fuel code and timestamp; after receiving the command, it calculates the CRC checksum and compares it with the checksum attached to the command; if they match, the command is confirmed to be valid and data fusion is triggered; if they do not match (e.g., due to data corruption caused by electromagnetic interference), it requests a resend of the command to avoid error handling.

[0052] Furthermore, step S20 includes the following specific steps:

[0053] Step S21: Divide the input tensor into time windows of fixed length. Within each window, calculate the abrupt correlation weight between the temperature sequence and the calorific value sequence through a multi-head self-attention layer. The abrupt correlation weight reflects the contribution of the calorific value change to the temperature fluctuation at the moment of fuel switching.

[0054] Step S21 is the core application of the attention mechanism, used to capture key features at the moment of fuel switching. A fixed-length time window (e.g., 10 seconds) divides the continuous time series into manageable segments, facilitating the model's focused analysis of changes before and after the switching event. The multi-head self-attention layer is a key component of the Transformer model, capable of simultaneously analyzing the correlation between data from different perspectives. The abrupt correlation weight quantifies the impact of calorific value changes on temperature fluctuations; for example, a higher correlation weight is assigned when a 1 MJ / m³ decrease in calorific value leads to a 25°C temperature drop. This weight is not pre-set but automatically learned by the model from historical data, adapting to different gas turbine models and operating conditions. Compared to traditional fixed-weight methods, this technique more accurately identifies the impact of fuel switching, especially when dealing with fuel switching with significant calorific value differences, such as between natural gas and biomethane.

[0055] Step S22: The product of the mutation correlation weight and the pressure fluctuation characteristics is used to output the fuel switching influence factor through a nonlinear activation function. The nonlinear activation function adopts the Sigmoid function to constrain the influence factor range within a preset interval.

[0056] Step S22 transforms intermediate features into an interpretable fuel switching impact factor. The fuel switching impact factor is a normalized value between 0 and 1, with 1 representing the maximum impact. The product of the mutation correlation weight and the pressure fluctuation feature reflects the combined effect of fuel characteristic changes and combustion dynamic response. (Sigmoid function) The results are mapped to the (0,1) interval to ensure that the influence factors have clear physical meaning and comparability. The S-shaped curve characteristic of the Sigmoid function allows moderate influences to be linearly represented, while extreme cases (close to 0 or 1) are appropriately compressed, which conforms to the actual characteristics of the combustion process—the influence usually does not increase indefinitely. The innovation of this step lies in fusing complex multidimensional features into a concise and interpretable index, facilitating subsequent processing and decision-making.

[0057] Step S23: During the model training phase, labeled data from historical fuel switching scenarios are used, wherein the labeled data contains known true values ​​of fuel switching influencing factors, and the model parameters are optimized using the mean squared error loss function.

[0058] Step S23 describes the model training process and is fundamental to ensuring prediction accuracy. Annotated data for historical fuel switching scenarios are collected from actual power plant operations, containing complete records of the switching process and the true extent of the fuel switching impact. The mean squared error loss function measures the difference between the model's predicted impact factors and the true values, and the model parameters are optimized by minimizing this difference using a backpropagation algorithm. This supervised learning method allows the model to learn from historical experience and continuously improve prediction accuracy. Compared to unsupervised learning, supervised learning utilizes valuable expert knowledge and is particularly suitable for prediction tasks with clear physical meaning, such as fuel switching.

[0059] Step S24: When the standard deviation of the temperature sequence within the time window exceeds its preset threshold, an attention head is added to enhance the ability to capture abrupt changes. The fluctuation intensity is dynamically evaluated through sliding window statistics.

[0060] Step S24 enables dynamic adaptation of the model structure, a significant innovation of this invention. Attention heads are units in the Transformer model that process different feature relationships in parallel. More attention heads can capture more complex feature relationships, but also increase computational burden. This step innovatively adjusts the number of attention heads dynamically based on the intensity of temperature series fluctuations (evaluated via the standard deviation of a sliding window). When fuel switching causes drastic temperature fluctuations (standard deviation exceeding a threshold), the system automatically increases the number of attention heads to enhance the ability to capture abrupt changes; when operation is stable, the number of attention heads is reduced to save computational resources. This dynamic adjustment avoids the limitations of traditional fixed-structure models, enabling the model to adaptively optimize according to real-time operating conditions, making it particularly suitable for handling transient processes like fuel switching.

[0061] Step S25: Perform feature scaling before inputting the tensor into the model. The scaling factor is dynamically calculated from the density parameter in the fuel chemical property parameter database to ensure the consistency of physical dimensions for different fuel types. The output range of the nonlinear activation function is dynamically constrained by preset boundary parameters.

[0062] Step S25 addresses the standardization issue of data across different fuel types. Feature scaling transforms features with different dimensions and ranges to a uniform scale, preventing certain features from dominating model learning due to their large numerical ranges. The scaling factor is dynamically calculated based on fuel density, as density directly affects the conversion between mass flow rate and volumetric flow rate. For example, natural gas density is approximately 0.7 kg / m³, while biomethane density may be 0.65 kg / m³; this difference needs to be considered in feature processing. Dynamic boundary parameters adjust the output range of activation functions such as the sigmoid function according to the current fuel type, ensuring the comparability of influencing factors across different fuel switching scenarios. This step guarantees the model's generalization ability across different fuel types and is crucial for achieving "universal fuel switching prediction."

[0063] Example: For natural gas, the characteristic scaling factor is 1.0; when switching to lower-density biomethane, the system calculates a scaling factor of 1.07 based on the density parameter of 0.65 kg / m³ in the database; at the same time, the boundary parameters of the Sigmoid function are adjusted from the usual [-5,5] to [-4.5,5.5] to adapt to the typical impact range of biomethane switching.

[0064] Furthermore, step S30 includes the following specific steps:

[0065] Step S31: Calculate the standard deviation of all temperature data within the current window by sliding a window of a preset time length on the temperature series. The calculated standard deviation values ​​form a temperature fluctuation standard deviation series in chronological order.

[0066] Step S31 quantifies the temperature stability of the combustion process, a key indicator for evaluating combustion quality. The sliding window standard deviation calculation measures the standard deviation of temperature data within a fixed time window (e.g., 5 seconds). A larger standard deviation indicates more severe temperature fluctuations and less stable combustion. The sliding window moves forward continuously with the arrival of new data, generating a continuous sequence of fluctuating standard deviations that reflects real-time changes in combustion stability. Compared to temperature values ​​at a single time point, fluctuating standard deviations better reflect the essential characteristics of the combustion process and are less affected by short-term noise. This step provides the core input for subsequent calculations of the combustion stability index.

[0067] Example: When natural gas is burning stably, the temperature standard deviation is 15°C within a 5-second window; when switching to biomethane, the temperature fluctuation intensifies, and the standard deviation increases to 30°C; after 10 seconds, it gradually stabilizes at 20°C; the system can identify the change in combustion state in a timely manner by monitoring this change.

[0068] Step S32: Perform differential operation on the pressure sequence to calculate the pressure change and pressure change rate at adjacent time points, and form a pressure change rate sequence by arranging the obtained pressure change rates in chronological order.

[0069] Step S32 captures the dynamic pressure changes during combustion, providing another important dimension for assessing combustion stability. Differential calculations are used to calculate the pressure change between adjacent time points, using the formula: ΔP = P(t) - P(t-Δt), where P(t) is the current pressure, P(t-Δt) is the previous pressure, Δt is the sampling interval, and the pressure change rate refers to the pressure change per unit time (Δt). The pressure change rate sequence reflects the speed of pressure fluctuations within the combustion chamber; drastic pressure fluctuations often indicate combustion instability or impending combustion oscillations. Similar to temperature fluctuations, the pressure change rate is also an important indicator of the combustion process, but it provides a different dimension of information. Combining temperature and pressure information allows for a more comprehensive assessment of the combustion state.

[0070] Step S33: The combustion stability index is generated by the weighted sum of the temperature fluctuation standard deviation sequence and the pressure change rate sequence. The weight coefficients are determined by a linear regression model trained on historical combustion instability event data and continuously optimized by an online learning algorithm to adapt to new data distribution.

[0071] Step S33 fuses multi-source information into a unified combustion stability index. The weighted sum formula is: SSI stands for Combustion Stability Index. For the standard deviation of temperature fluctuation, Here, w1 and w2 represent the pressure change rate, and w1 and w2 are weighting coefficients. These weighting coefficients are not fixed values ​​but are determined using a linear regression model trained on historical combustion instability event data. This model learns the relationship between temperature fluctuations and pressure changes and actual combustion stability. An online learning algorithm allows the weights to adjust over time. The combustion stability index is typically normalized to a range of 0-1, where 1 represents complete stability and 0 represents extreme instability. The innovation of this step lies in the dynamic weighting mechanism, which enables the system to automatically adjust the evaluation criteria based on the characteristics of different gas engines.

[0072] Step S34: After receiving the fuel switching impact factor, if the impact factor is greater than the preset threshold, real-time evaluation is initiated; otherwise, the default stability index is maintained. The preset threshold is dynamically set based on the fuel chemical characteristic parameters.

[0073] Step S34 enables intelligent switching of the evaluation mechanism and optimizes the allocation of computing resources. A preset threshold is the critical point for determining whether a detailed evaluation needs to be initiated; its value is dynamically adjusted based on fuel chemical characteristics. When the fuel switching impact factor exceeds the threshold, indicating a significant switching impact, the system initiates a real-time evaluation mode, increasing computing resource investment. If the impact factor is small, the system maintains the default evaluation, saving computing resources. This dynamic threshold mechanism avoids evaluating all switching events with equal intensity, improving system efficiency. In actual power plant operation, frequent small parameter adjustments are common, but only significant fuel switching requires detailed evaluation.

[0074] Step S35: Introduce a time decay factor. The time decay factor is dynamically adjusted based on the duration of the fuel switching command signal. The longer the time after switching, the smaller the decay factor becomes, so as to reduce the influence of historical data on the current index. The window length of the sliding window standard deviation calculation module is dynamically configured according to the fuel type.

[0075] Step S35 addresses the issue of historical data interfering with the current assessment. The time decay factor is a weight that decreases over time (formula: The formula (where λ is the decay rate and t is the time elapsed after the switch) is used to reduce the impact of historical data on the current stability assessment during the initial stage of the switch. Immediately after the fuel switch, historical data (before the switch) differs significantly from the current state and should be assigned a lower weight. As time progresses, the system gradually enters a stable combustion state with the new fuel, increasing the reference value of historical data and consequently raising its weight. The sliding window length is also dynamically adjusted according to the fuel type; for example, for slower-burning biomethane, a longer window is used to capture smoother trends. This dynamic adjustment mechanism allows the system to adapt to the combustion characteristics of different fuels, providing a more accurate stability assessment.

[0076] Example: In the first second after switching to biomethane, the time decay factor is 0.9, and historical data has little impact; after the 10th second, the decay factor drops to 0.3, and historical data has almost no impact on the current assessment; at the same time, the system extends the sliding window from 5 seconds to 7 seconds to adapt to the slower combustion dynamics of biomethane.

[0077] Furthermore, step S40 includes the following specific steps:

[0078] Step S41: Determine the switching time point according to the fuel switching command signal, and select a window of preset time length before and after the switching point on the time axis. The range of the preset time length window is dynamically expanded according to the magnitude of the fuel switching influencing factor.

[0079] Step S41 defines the time range for dynamic correction, which is fundamental to achieving precise correction. The switching time point is precisely marked by the fuel switching command signal and serves as the reference origin for the entire correction process. A preset time window (e.g., 5 seconds before and after the switching point) defines the time area where dynamic correction is required. This window range is not fixed but dynamically adjusted according to the magnitude of the fuel switching impact factor—the larger the impact factor, the wider the window range, because a large-impact switch requires more time to reach a new equilibrium. For example, when the impact factor is 0.3, the window may be ±3 seconds; when the impact factor is 0.8, the window may expand to ±10 seconds. This dynamic window mechanism ensures that the correction process covers the entire transition period, neither over-correcting stable period data nor omitting key transition data.

[0080] Step S42: Within the window, calculate the correction slope based on the combustion stability index. The correction slope is determined by the derivative of the linear interpolation function, where the input of the interpolation function is the combustion stability index and the output is the power generation adjustment range. The parameters of the linear interpolation function are determined by the real-time change rate of the combustion stability index.

[0081] Step S42 calculates the correction rate for the power generation forecast, which is the core of the dynamic correction. A linear interpolation function establishes the mapping relationship between the combustion stability index (SSI) and the power generation adjustment magnitude (ΔP1): ΔP1 = a·SSI + b, where a and b are parameters. The correction slope is the derivative of this function, d(ΔP1) / d(SSI) = a, representing the magnitude of power generation adjustment required for each unit change in SSI. Parameter a is determined by the real-time rate of change of SSI—when SSI decreases rapidly (e.g., during fuel switching), a takes a larger value, indicating a need for rapid forecast adjustment; when SSI changes slowly, a takes a smaller value, achieving a smooth transition. This parameter adjustment based on the real-time rate of change allows the correction process to adapt to different switching speeds and intensities, avoiding under-correction or over-correction caused by a fixed slope.

[0082] Step S43: Obtain the dynamic correction value by multiplying the correction slope by the base power generation prediction value, which is output by the original deep learning model without integrated fuel parameters.

[0083] Step S43 converts the correction slope into a specific power generation adjustment. The baseline power generation forecast comes from a traditional forecasting model that does not account for changes in fuel parameters, resulting in a fixed bias during fuel switching. By correlating the correction slope with the baseline forecast (actually with SSI changes), the system can calculate an accurate correction. For example, if the baseline model predicts that power generation will remain at 300MW after the switch, but the actual value should be 285MW, the dynamic correction is -15MW. The key innovation of this step lies in directly correlating the correction with the combustion state, rather than a simple time-delay adjustment, thus more accurately reflecting the physical processes.

[0084] Step S44: The dynamic correction value and the basic power generation prediction value are integrated using a weighted average strategy. The weight is dynamically allocated by the fuel switching influence factor. The larger the influence factor, the higher the weight of the dynamic correction value. The weight allocation process of the weighted average strategy introduces the output result of the fuel type identification module as an auxiliary input.

[0085] Step S44 achieves intelligent fusion of prediction results, balancing the contributions of new and old information. The weighted average formula is: ,in This is the final predicted value. The base forecast value is ΔP1, the dynamic correction value is ΔP1, and w is the weight of the dynamic correction value. The weight w is determined by the fuel switching impact factor (f), and is usually set to... (k is an adjustment parameter) to ensure that the larger the influence factor, the higher the correction weight. The output of the fuel type identification module (e.g., "biomethane") serves as an auxiliary input, allowing for further fine-tuning of the weights—for fuel types known to have a significant impact, the system may increase the k value to enhance the correction weight. This dynamic weighting mechanism ensures that corrections are fully applied during significant switching events and maintains predictive stability during minor fluctuations.

[0086] Step S45: Before outputting the real-time power generation prediction result, a smoothing process is performed. The smoothing process uses an exponential moving average algorithm, and the attenuation coefficient is adaptively set according to the combustion stability index. The lower the stability index, the smaller the attenuation coefficient to retain the abrupt change characteristics. The attenuation coefficient adjustment mechanism of the smoothing process ensures that the prediction curve retains key features near the abrupt change point.

[0087] Step S45 performs post-processing on the prediction results to improve output quality. The formula for the exponential moving average algorithm is: Where P(t) is the current predicted value, The smoothed value, This refers to the attenuation coefficient. Dynamically adjusted based on the combustion stability index—when the SSI is low (combustion unstable), Increasing (close to 1) reduces smoothness and preserves abrupt changes; when SSI is high (combustion stability), By reducing and increasing smoothness, random fluctuations are eliminated. This adaptive smoothing mechanism avoids excessive fluctuations in forecast results (which affect scheduling decisions) while preserving real dynamic changes (such as the decrease in power generation due to fuel switching), achieving an optimal balance between smoothness and responsiveness.

[0088] Example 2

[0089] Preferably, in Embodiment 2, a stability verification is performed between step S20 and step S30, wherein the stability verification includes the following steps:

[0090] Step Z10: Compare the fuel switching impact factor with the preset threshold. If the impact factor is greater than the threshold, perform a double check of the combustion stability index.

[0091] Step Z10 serves as the trigger mechanism for stability verification, enabling intelligent allocation of computing resources. Preset thresholds are dynamically set based on fuel type and turbine model; for example, the threshold for switching from natural gas to biomethane might be 0.6, while the threshold for switching between two similar fuels might be 0.8. When the fuel switching impact factor exceeds the threshold, it indicates a significant switching impact, requiring additional verification to ensure the accuracy of the stability assessment. If the impact factor is small, the system skips the double verification, saving computing resources. This conditional verification mechanism avoids performing equally intensive verification on all switching events, improving system efficiency. In actual power plant operation, frequent small parameter adjustments are common, but only significant fuel switches require detailed verification.

[0092] Step Z11: Extract the high-frequency fluctuation components of the temperature sequence through the fast Fourier transform module. The frequency band division is dynamically determined based on the historical spectral characteristics of the fuel type.

[0093] Step Z11 analyzes temperature fluctuations from a frequency domain perspective, providing supplementary verification information. The Fast Fourier Transform (FFT) converts the time-domain temperature signal into a frequency-domain representation, revealing the fluctuation intensity of different frequency components. Frequency band division is based on the typical combustion oscillation frequencies corresponding to fuel types in historical data—natural gas combustion may have a characteristic frequency band of 5-10Hz, while biomethane may have a band of 3-8Hz. The system dynamically determines these frequency bands and extracts the fluctuation components within the corresponding frequency range as high-frequency fluctuation characteristics. Combustion instability often manifests as energy concentration in specific frequency bands; this step can identify potential combustion problems earlier and is more sensitive than time-domain analysis. During fuel switching, an abnormal increase in high-frequency fluctuations is often a precursor to combustion instability.

[0094] Step Z12: Detect instantaneous abrupt change points in the pressure sequence using a wavelet analysis module. The sensitivity of the abrupt change point detection is adaptively adjusted by the fluctuation amplitude of the fuel calorific value sequence.

[0095] Step Z12 focuses on identifying transient anomalies in the pressure signal, a crucial indicator of combustion stability. Wavelet analysis, a time-frequency analysis method, can precisely locate abrupt changes in the signal (such as sudden pressure increases or decreases), offering greater accuracy than traditional threshold detection. The sensitivity of abrupt change detection is dynamically adjusted based on fuel calorific value fluctuations—increasing sensitivity to capture more potential problems when calorific value fluctuations are large (e.g., during initial fuel switching); decreasing sensitivity to avoid false alarms when calorific value is stable. This adaptive mechanism allows the system to maintain optimal detection performance under different operating conditions. Pressure abrupt changes are often a direct manifestation of combustion instability or impending combustion oscillations; early detection can prevent serious problems.

[0096] Example: During the biomethane switching process, the calorific value fluctuates significantly. The system lowered the threshold of wavelet analysis from 0.1 MPa to 0.05 MPa; it successfully detected a pressure change of 0.07 MPa at 14:00:02, 2 seconds earlier than the temperature change, providing an early warning.

[0097] Step Z13: Cross-validate the high-frequency fluctuation component and the instantaneous mutation point result with the combustion stability index calculated in step S30. If the cross-validation deviation exceeds the preset deviation threshold, the model retraining mechanism is triggered.

[0098] Step Z13 performs cross-validation of multi-source information, improving the reliability of stability assessment. Cross-validation compares the results of frequency domain analysis (step Z11) and time-frequency analysis (step Z12) with the time-domain stability index from step S30. If the differences between the three assessment results are too large (exceeding a preset deviation threshold), it indicates potential assessment bias or abnormal operating conditions, triggering a model retraining mechanism. For example, if the time-domain SSI shows stability (0.8), but the frequency domain analysis shows abnormally high high-frequency energy (25%), the assessment may be inaccurate. This cross-validation mechanism significantly reduces the risk of misjudgment, ensuring the accuracy of combustion stability assessment, which is crucial for safe operation.

[0099] Example: The system detected a time-domain SSI of 0.75, but the energy in the 4-7Hz frequency band reached 22%, and detected 3 pressure mutation points; the cross-validation deviation exceeded the threshold, triggering model fine-tuning and avoiding misjudgment of unstable combustion states.

[0100] Step Z14: Retrieve the most recent time series data of a preset duration from the industrial database, and use an online learning algorithm to fine-tune the parameters of the last two layers of the deep learning model enhanced by the attention mechanism. The fine-tuning process of the model retraining mechanism adopts the mini-batch gradient descent algorithm, and the batch size is dynamically optimized according to the amount of data.

[0101] Step Z14 enables rapid model adaptation and is crucial for continuous system optimization. When cross-validation identifies issues, the system retrieves recent (e.g., 5-minute) high-resolution data from an industrial database and uses an online learning algorithm to fine-tune only the parameters of the last two layers of the model, rather than completely retraining, significantly reducing adjustment time. The mini-batch gradient descent algorithm divides the data into small batches (batch size dynamically adjusted based on data volume) and updates model parameters batch by batch, balancing training speed and stability. This lightweight retraining mechanism allows the model to quickly adapt to new operating conditions without causing system instability due to frequent complete retraining. For common but impactful events like fuel switching, this mechanism ensures the prediction model maintains high accuracy.

[0102] Example: After detecting an anomaly in the biomethane combustion stability assessment, the system acquires data from the last 300 seconds; using mini-batch gradient descent with a batch size of 32, it fine-tunes only the last two layers of the model; the fine-tuning is completed within 5 seconds, improving the model's prediction accuracy for biomethane combustion by 15%.

[0103] Example 3

[0104] Preferably, in Embodiment 3, data feedback is performed after step S40, wherein the data feedback includes the following steps:

[0105] Step Z20: Calculate the deviation between the output real-time power generation prediction result and the actual power generation measurement value. The actual power generation measurement value is collected by the power plant's energy meter at a preset time interval. The collection frequency of the preset time interval is dynamically configured by the real-time demand of the energy dispatch system.

[0106] Step Z20 establishes a closed-loop feedback between prediction and reality, which is the foundation for continuous system optimization. The deviation calculation formula is: ,in To predict power generation, The data represents actual measured values. The electricity meter collects actual power generation at configurable time intervals (e.g., 5 seconds or 10 seconds), with the interval length dynamically adjusted according to energy dispatch needs—high-frequency data collection (5 seconds) is used during critical processes such as fuel switching, while low-frequency data collection (10 seconds) is used during stable operation. This dynamic data collection strategy balances data accuracy and system load, providing sufficient feedback data at critical moments requiring high-precision predictions. The deviation value is a direct indicator of prediction quality and an important input for subsequent optimization.

[0107] Step Z21: Input the deviation calculation result into the adaptive learning rate adjustment module. The module dynamically modifies the optimizer learning rate of the deep learning model with attention mechanism enhancement according to the deviation size. The larger the absolute value of the deviation, the higher the learning rate. The dynamic modification mechanism of the adaptive learning rate adjustment module introduces the historical deviation trend as a reference.

[0108] Step Z21 enables intelligent adjustment of the learning rate, optimizing model training efficiency. The learning rate is a key hyperparameter in deep learning for controlling the step size of parameter updates; traditional fixed learning rates are difficult to adapt to different operating conditions. This step innovatively links the learning rate to prediction bias: when the bias is large (e.g., in the initial stage of fuel switching), the learning rate is increased (e.g., from 0.001 to 0.01) to accelerate model adjustment; when the bias is small, the learning rate is decreased for fine-tuning. Historical bias trends are used as a reference to avoid drastic fluctuations in the learning rate caused by a single large bias—only if there are multiple consecutive large biases does the learning rate increase significantly. This mechanism allows the model to respond quickly to significant changes while maintaining accuracy under stable operating conditions.

[0109] Step Z22: Generate a feature vector for the fuel switching scenario. The feature vector includes the fuel switching influencing factor, combustion stability index and dynamic correction value, and is stored in the feature database for subsequent model iteration.

[0110] Step Z22 systematically records the key features of fuel switching events, providing a data foundation for continuous model optimization. Feature vectors are compact representations of multidimensional data, including, in this example, core parameters such as fuel switching impact factors (0-1, representing the degree of impact), combustion stability index (0-1, representing combustion quality), and dynamic correction values ​​(MW, representing power generation adjustments). These features are standardized and stored in a feature database, forming a structured fuel switching knowledge base. Compared to the original time series, feature vectors significantly compress the data volume while retaining key information, facilitating rapid subsequent retrieval and analysis. This step realizes the transformation from raw data to knowledge, forming the basis of the system's "learning ability."

[0111] Step Z23: Perform principal component analysis to reduce the dimensionality of the feature vectors, retaining a preset proportion of variance information. The variance retention ratio in the principal component analysis dimensionality reduction process is adaptively adjusted according to the storage capacity of the feature database.

[0112] Step Z23 optimizes feature storage efficiency while preserving key information. Principal Component Analysis (PCA) is a linear dimensionality reduction technique that projects high-dimensional feature vectors into a low-dimensional space by identifying the main directions of change (principal components) in the data. The variance retention ratio (e.g., 95%) represents the degree of variation in the original data retained after dimensionality reduction; a higher ratio results in less information loss but a worse dimensionality reduction effect. The system dynamically adjusts this ratio based on the storage pressure of the feature database—retaining 95% variance when storage is sufficient and reducing it to 90% when storage is tight. This adaptive mechanism strikes a balance between information preservation and storage efficiency, ensuring the long-term sustainability of the system. The dimensionality-reduced features can still effectively characterize the essential properties of fuel switching events and are suitable for subsequent model iterations.

[0113] Step Z24: The output of the adaptive learning rate adjustment module is applied in real time to the model quantization process in step S20 to form a closed-loop optimization mechanism. The execution frequency of the closed-loop optimization mechanism is determined by the occurrence frequency of the fuel switching event.

[0114] Step Z24 completes the entire feedback loop, enabling the system to self-optimize. The output of the adaptive learning rate adjustment module (the updated learning rate) is applied in real time to the model quantization process in step S20, allowing the prediction model to continuously improve based on the latest feedback. The execution frequency of the closed-loop optimization mechanism is related to the frequency of fuel switching events—during frequent switching (such as the testing phase), a complete optimization is performed every 5 minutes; during stable operation (such as normal production), it is performed once per hour. This adaptive execution strategy ensures that the system optimizes quickly when needed, while reducing unnecessary computational overhead during stable periods. This closed-loop mechanism gives the system the characteristic of "becoming more accurate with use," with prediction accuracy continuously improving over long-term operation.

[0115] Example 3

[0116] Preferably, in Embodiment 3, knowledge graph reasoning is performed between step S10 and step S20, wherein the knowledge graph reasoning includes the following steps:

[0117] Step Z30: Construct a knowledge graph for gas turbine fuel switching. The knowledge graph stores the relationship between fuel type, chemical characteristic parameters and historical switching events in the form of triples. Each triple includes entity nodes and relation edges. The entity nodes are generated from historical fuel switching event data in the industrial database and cover fuel type entities, gas turbine model entities and operating condition entities.

[0118] Step Z30 establishes a knowledge representation framework for fuel switching, which forms the basis for the system's "experience learning." The knowledge graph organizes knowledge in the form of triples (head entity, relation, tail entity), such as ("natural gas", "calorific value", "38.5 MJ / m³") or ("natural gas → biomethane", "typical impact", "5% decrease in power generation"). Entity nodes include fuel type (natural gas, biomethane, etc.), turbine model (e.g., a certain type of heavy-duty gas turbine), and operating conditions (load rate, ambient temperature, etc.). Relational edges represent the associations between entities, such as "calorific value", "degree of impact", and "recommended adjustment amount". This knowledge is extracted from historical fuel switching events, transforming unstructured time-series data into structured domain knowledge. Compared to traditional databases, knowledge graphs can better represent complex relationships, support semantic reasoning, and enable the system to "understand" rather than merely "memorize" historical experience.

[0119] Step Z31: Dynamically construct relation edges through correlation analysis of parameter changes in historical event data. The relation edges represent the degree of quantitative impact of fuel type conversion on combustion efficiency. The correlation analysis generates relation edge weights by calculating the Pearson correlation coefficient of parameter changes.

[0120] Step Z31 quantifies the strength of relationships in the knowledge graph, making inference more precise. The Pearson correlation coefficient (r) measures the degree of linear correlation between two variables, and is calculated using the following formula: Where cov is the covariance and σ is the standard deviation. For example, calculating the correlation coefficient between changes in fuel calorific value and changes in power generation, if r = -0.85, it indicates a strong negative correlation between a decrease in calorific value and a decrease in power generation. This correlation coefficient serves as the weight of the relation edge, representing the strength of the "fuel calorific value → power generation" relationship. Dynamic construction means that as new data is added, the correlation coefficient is updated, keeping the knowledge graph up-to-date. This statistically based correlation analysis is more objective and accurate than fixed weights set by expert experience.

[0121] Step Z32: When a fuel switching command signal is received, the command signal is input to the graph neural network inference engine. The inference engine queries the knowledge graph based on the target fuel type in the command signal and outputs a reference vector of the switching influence between the target fuel type and the current fuel type.

[0122] Step Z32 enables rapid reasoning based on a knowledge graph, providing historical references for the current switchover. A Graph Neural Network (GNN) is a deep learning model specifically designed for graph-structured data, capable of learning complex relationships between entities. When a new switchover instruction arrives, the GNN inference engine searches the knowledge graph for similar historical events (same or similar fuel type, turbine model, operating conditions) and aggregates relevant information to generate a switchover impact reference vector. This vector contains key indicators such as expected changes in power generation and combustion stability, for example, [rate of change in power generation, rate of change in temperature, pressure fluctuation amplitude] = [-0.045, -0.03, 0.15]. Compared to directly using historical averages, GNNs can consider more contextual factors, providing more personalized predictive references.

[0123] Step Z33: Concatenate the switching influence reference vector with the fuel calorific value sequence obtained in step S10 to generate an enhanced fuel feature vector;

[0124] Step Z33 integrates historical experience with real-time data to enhance feature representation capabilities. Vector concatenation connects the switching impact reference vector with the real-time fuel calorific value sequence along the feature dimension. For example, the reference vector is 3-dimensional, and the calorific value sequence is 10-dimensional (10 seconds of history), resulting in a 13-dimensional enhanced feature vector. This fusion allows the model to simultaneously possess both "historical experience" and "current state" perspectives, making it more accurate in predicting switching impacts than using only real-time data. The enhanced feature vector preserves the dynamic characteristics of the time series while incorporating domain knowledge, making it particularly suitable for handling fuel switching processes, which have clear physical laws but are influenced by multiple factors.

[0125] Step Z34: Input the enhanced fuel feature vector into the input tensor fusion process of step S20 to replace the original calorific value sequence in the efficiency mutation feature calculation. The query process of the graph neural network inference engine adopts a graph attention mechanism to dynamically adjust the relation edge weights according to the current gas turbine model entity.

[0126] Step Z34 completes the knowledge-enhanced prediction process and is a key innovation of this invention. Enhanced feature vectors replace the original calorific value sequence in subsequent calculations, enabling the model to make predictions based on both historical experience and real-time data. Graph attention, an advanced variant of GNN, dynamically adjusts the weights of relational edges in the knowledge graph based on the current query—giving higher weights to experiences more relevant to the current gas turbine model. For example, if a gas turbine is more sensitive to biomethane, the system will place greater emphasis on historical data for that model. This personalized reasoning makes predictions more accurate, solving the "one-size-fits-all" problem of traditional models. The entire process achieves a dual-drive of "data + knowledge," significantly improving prediction accuracy in fuel switching scenarios.

[0127] Example 4

[0128] Preferably, in Embodiment 4, wolf pack algorithm optimization is performed after step S40, wherein the wolf pack algorithm optimization includes the following steps:

[0129] Step Z40: Obtain the fuel switching scenario feature vector from the feature database. The feature vector is generated by step Z22 and includes the fuel switching influencing factor, combustion stability index and dynamic correction value.

[0130] Step Z40 provides the optimization objective and initial data for the wolf pack algorithm. The fuel switching scenario feature vector is a structured data representation generated in step Z22, containing key parameters such as the fuel switching impact factor (0-1, representing the degree of impact), combustion stability index (0-1, representing combustion quality), and dynamic correction value (MW, representing the adjustment amount of power generation). These features are standardized to form a multi-dimensional optimization space in which the wolf pack algorithm searches for the optimal solution. Compared with directly optimizing the original time series, using feature vectors significantly reduces the problem complexity while retaining key information, making the optimization process more efficient. The feature vector also contains metadata such as timestamps and gas turbine models, allowing the optimization results to be customized for specific conditions.

[0131] Step Z41: Input the feature vector into the wolf pack algorithm optimization module. The optimization module initializes the wolf pack individual set, which includes alpha wolf individuals, scout wolves individuals, and predator wolves individuals. The alpha wolf individual represents the global optimal solution, the scout wolf individual represents the local search direction, and the predator wolves individuals represent the parameter update step size.

[0132] Step Z41 establishes the basic framework of the wolf pack algorithm, simulating wolf pack hunting behavior for optimization. Individual wolves represent candidate solutions in the algorithm, with each individual representing a set of hyperparameter configurations for a deep learning model (such as learning rate and regularization coefficients). The alpha wolf represents the currently known best solution, guiding the entire pack's search direction; scout wolves explore new areas to avoid getting trapped in local optima; and aggressive wolves perform fine-tuning searches, determining the specific step size for parameter updates. This division of roles simulates the cooperative hunting strategy of real wolf packs, balancing global exploration with local exploitation. Compared to traditional optimization algorithms, the wolf pack algorithm has stronger global search capabilities and faster convergence speed, making it particularly suitable for high-dimensional non-convex optimization problems.

[0133] Step Z42: Calculate the fitness value of each wolf pack individual based on the feature vector. The fitness value is determined by the weighted sum of the dynamic correction value and the historical prediction deviation in the feature vector, and generate the parameter update direction vector through time series correlation analysis of the feature vector.

[0134] Step Z42 defines the optimization objective and evaluation criteria. Fitness is an indicator of the quality of individual wolves, calculated as: fitness = w1·|ΔP| + w2·error, where ΔP is the dynamic correction value, error is the historical prediction bias, and w1 and w2 are weights. A lower fitness value indicates a better individual, as an ideal prediction should have small correction values ​​and small biases. Time-series correlation analysis calculates the correlation between the dimensions of the feature vectors, generating a parameter update direction vector to guide the wolves to move towards more promising areas. For example, if a negative correlation is found between the learning rate and the prediction bias, the direction vector will suggest reducing the learning rate. This data-driven correlation analysis makes the optimization process more intelligent and targeted.

[0135] Step Z43: Execute the iterative optimization process: the scout wolf performs a bounding search on the fitness value to generate a parameter update direction vector, the ferocious wolf calculates the step size adjustment based on the parameter update direction vector and updates the hyperparameters of the deep learning model, and the alpha wolf evaluates whether the updated hyperparameters meet the convergence condition.

[0136] Step Z43 describes the core optimization process of the wolf pack algorithm. Encirclement search involves individual wolves exploring around the current optimal solution (the alpha wolf), generating a parameter update direction vector to determine the search direction. The step size adjustment is calculated by individual wolves, determining the magnitude of the parameter update—large steps are used in the early stages of the search to quickly approach the optimal region, while small steps are used for fine-tuning as the optimal solution approaches. The alpha wolf continuously evaluates the updated parameter configuration, checking if convergence conditions are met (e.g., fitness value change is less than a threshold, or the maximum number of iterations is reached). This iterative process simulates the wolf pack's strategy of gradually encircling its prey, efficiently searching the solution space. Compared to single-point optimization, swarm intelligence algorithms are better able to avoid local optima and find the global optimum.

[0137] Step Z44: Input the optimal set of hyperparameters that meets the convergence condition into the parameter configuration unit of the deep learning model with enhanced attention mechanism, so as to dynamically adjust the learning rate, regularization coefficient and number of attention heads of the model. The convergence condition is dynamically set based on the changing trend of the combustion stability index in the feature vector. The output of the optimal set of hyperparameters triggers the instantaneous reloading mechanism of the model parameters.

[0138] Step Z44 applies the optimization results, enabling adaptive model adjustment. The optimal hyperparameter set includes key parameters such as the learning rate, regularization coefficient, and number of attention heads, which directly affect the model's predictive performance. The convergence condition is not a fixed value but is dynamically set based on the changing trend of the combustion stability index—when the stability index changes rapidly (in the early stages of fuel switching), the convergence condition is relaxed, allowing for a wider range of parameter adjustments; when the stability index stabilizes, the convergence condition is tightened for fine-tuning. The instant reload mechanism ensures that new parameters take effect immediately without restarting the system, guaranteeing real-time optimization. This dynamic hyperparameter adjustment allows the model to always operate in its optimal state, adapting to constantly changing operating conditions.

[0139] Example 5

[0140] Based on the same inventive concept as the deep learning-based time series prediction method for gas turbine power plants in the foregoing embodiments, this application provides a deep learning-based time series prediction system for gas turbine power plants, including:

[0141] The data acquisition module is used to receive fuel switching command signals from the gas turbine control system; and to acquire gas turbine operating time series data in real time. The time series data includes combustion chamber temperature sequence, exhaust pressure sequence, fuel calorific value sequence and fuel flow sequence, and all data are stored in an industrial database with a unified timestamp.

[0142] The data fusion module is used to fuse the time series data with the corresponding parameters in the fuel chemical property parameter database to generate an input tensor; the input tensor is input into an attention mechanism-enhanced deep learning model, which is based on the Transformer architecture and calculates adaptive correlation weights between time steps to quantify the fuel switching impact factor, which is calculated based on the abrupt deviation between the temperature sequence and the calorific value sequence in the input tensor.

[0143] The data processing module is used to receive the fuel switching influencing factor through the fuel type identification module, and calculate the combustion stability index in real time by combining the temperature series and pressure series in the time series data. The combustion stability index is dynamically generated based on the weighted sum of the standard deviation of the temperature series fluctuation and the rate of change of the pressure series, wherein the standard deviation of the fluctuation is calculated from the sliding window variance of the temperature series, and the rate of change is obtained by the differential operation of the pressure series.

[0144] The power generation prediction module is used to generate a dynamic correction value for power generation after fuel switching based on the combustion stability index. The dynamic correction value is used to adjust the slope of the prediction curve near the time point triggered by the fuel switching command signal through a linear interpolation method. The dynamic correction value is then fused with the output value of the basic power generation prediction model to output the real-time power generation prediction result.

[0145] The above description is merely a preferred embodiment of the present invention and is not intended to limit the present invention. Any modifications, equivalent substitutions, and improvements made within the spirit and principles of the present invention should be included within the scope of protection of the present invention.

Claims

1. A deep learning-based time series prediction method for gas turbine power plants, characterized in that, Includes the following steps: Step S10: Receive fuel switching command signal from the gas turbine control system; acquire gas turbine operation time series data in real time, the time series data including combustion chamber temperature sequence, exhaust pressure sequence, fuel calorific value sequence and fuel flow sequence, and all data are stored in an industrial database with a unified timestamp; Step S20: Fuse the time series data with the corresponding parameters in the fuel chemical property parameter database to generate an input tensor; The input tensor is fed into a deep learning model with an attention mechanism. The model, based on the Transformer architecture, calculates an adaptive correlation weight matrix between time steps to locate the abrupt change time point of the fuel calorific value sequence. A feature mask is applied to extract the abrupt change correlation weights of the temperature sequence and the calorific value sequence from the adaptive correlation weight matrix. The abrupt change correlation weights are multiplied by the pressure fluctuation features and input into the Sigmoid activation function for normalization into a fuel switching influencing factor. Step S30: Obtain the fuel switching influence factor. If the fuel switching influence factor is greater than a preset threshold, then, in combination with the temperature series and pressure series in the time series data, calculate the combustion stability index in real time. The combustion stability index is dynamically generated based on the weighted sum of the standard deviation of the temperature series fluctuation and the rate of change of the pressure series. The standard deviation of the fluctuation is calculated from the sliding window variance of the temperature series, and the rate of change is obtained from the differential operation of the pressure series. Otherwise, maintain the default stability index; Step S40: Generate a dynamic correction value for power generation after fuel switching based on the combustion stability index. The dynamic correction value is used to adjust the slope of the prediction curve near the time point triggered by the fuel switching command signal through a linear interpolation method. The dynamic correction value is then fused with the output value of the basic power generation prediction model to output the real-time power generation prediction result.

2. The deep learning-based time series prediction method for gas turbine power plants as described in claim 1, characterized in that, For step S10, the following specific steps are included: Step S11: Measure the chemical composition ratio of the fuel in real time using an online gas chromatograph, and match the measurement results with a pre-stored standard fuel calorific value table to calculate the calorific value sequence; Step S12: The mass flow meter collects the fuel flow sequence at a preset sampling interval and aligns it with the calorific value sequence through a time synchronization protocol to ensure that the two are stored with the same timestamp in the industrial database; Step S13: After performing anti-interference filtering on the temperature sequence output by the thermocouple, input it into the industrial database. Perform normalization on the pressure sequence output by the pressure transmitter to a preset value range. The normalization process uses the maximum and minimum values ​​of historical pressure data as boundaries. Step S14: Before storing the time series data, perform missing value imputation. The missing value imputation is based on the linear extrapolation algorithm of adjacent time points to ensure data continuity and dynamically verify the rationality of the imputation result. If the imputation deviation exceeds the preset tolerance, the imputation algorithm is re-executed. Step S15: Transmit the fuel switching command signal to the data processing unit via the Modbus TCP protocol. The transmission process includes a signal verification mechanism. The data fusion operation in step S20 is triggered only when the verification passes. The signal verification mechanism verifies the integrity of the command signal through a cyclic redundancy check code.

3. The deep learning-based time series prediction method for gas turbine power plants as described in claim 1, characterized in that, Step S20 includes the following specific steps: Step S21: Divide the input tensor into time windows of fixed length. Within each window, calculate the abrupt correlation weight between the temperature sequence and the calorific value sequence through a multi-head self-attention layer. The abrupt correlation weight reflects the contribution of the calorific value change to the temperature fluctuation at the moment of fuel switching. Step S22: The product of the mutation correlation weight and the pressure fluctuation characteristics is used to output the fuel switching influence factor through a nonlinear activation function. The nonlinear activation function adopts the Sigmoid function to constrain the influence factor range within a preset interval. Step S23: During the model training phase, labeled data from historical fuel switching scenarios are used, wherein the labeled data contains known true values ​​of fuel switching influencing factors, and the model parameters are optimized using the mean squared error loss function. Step S24: Dynamically adjust the number of attention heads. The number is adaptively determined based on the standard deviation of the temperature sequence within the time window. When the standard deviation exceeds a preset threshold, the number of attention heads is increased to enhance the ability to capture abrupt changes. The fluctuation intensity is dynamically evaluated through sliding window statistics. Step S25: Perform feature scaling before inputting the tensor into the model. The scaling factor is dynamically calculated from the density parameter in the fuel chemical property parameter database to ensure the consistency of physical dimensions for different fuel types. The output range of the nonlinear activation function is dynamically constrained by preset boundary parameters.

4. The deep learning-based time series prediction method for gas turbine power plants as described in claim 1, characterized in that, Step S30 includes the following specific steps: Step S31: Input the temperature sequence into the sliding window standard deviation calculation module, wherein the sliding window length is a preset time length, and output the temperature fluctuation standard deviation sequence; Step S32: Input the pressure sequence into the differential calculation module, calculate the pressure change at adjacent time points, and output the pressure change rate sequence; Step S33: The combustion stability index is generated by the weighted sum of the temperature fluctuation standard deviation sequence and the pressure change rate sequence. The weight coefficients are determined by a linear regression model trained on historical combustion instability event data and continuously optimized by an online learning algorithm to adapt to new data distribution. Step S34: After receiving the fuel switching impact factor, if the impact factor is greater than the preset threshold, real-time evaluation is initiated; otherwise, the default stability index is maintained. The preset threshold is dynamically set based on the fuel chemical characteristic parameters. Step S35: Introduce a time decay factor. The time decay factor is dynamically adjusted based on the duration of the fuel switching command signal. The longer the time after switching, the smaller the decay factor becomes, so as to reduce the influence of historical data on the current index. The window length of the sliding window standard deviation calculation module is dynamically configured according to the fuel type.

5. The deep learning-based time series prediction method for gas turbine power plants as described in claim 1, characterized in that, Step S40 includes the following specific steps: Step S41: Determine the switching time point according to the fuel switching command signal, and select a window of preset time length before and after the switching point on the time axis. The range of the preset time length window is dynamically expanded according to the magnitude of the fuel switching influencing factor. Step S42: Within the window, calculate the correction slope based on the combustion stability index. The correction slope is determined by the derivative of the linear interpolation function, where the input of the interpolation function is the combustion stability index and the output is the power generation adjustment range. The parameters of the linear interpolation function are determined by the real-time change rate of the combustion stability index. Step S43: Obtain the dynamic correction value by multiplying the correction slope by the base power generation prediction value, which is output by the original deep learning model without integrated fuel parameters. Step S44: The dynamic correction value and the basic power generation prediction value are integrated using a weighted average strategy. The weights are dynamically allocated by the fuel switching influence factor. The larger the influence factor, the higher the weight of the dynamic correction value. The weight allocation process of the weighted average strategy introduces the output of the fuel type identification module as an auxiliary input. Step S45: Before outputting the real-time power generation prediction result, a smoothing process is performed. The smoothing process uses an exponential moving average algorithm, and the attenuation coefficient is adaptively set according to the combustion stability index. The lower the stability index, the smaller the attenuation coefficient to retain the abrupt change characteristics. The attenuation coefficient adjustment mechanism of the smoothing process ensures that the prediction curve retains key features near the abrupt change point.

6. The deep learning-based time series prediction method for gas turbine power plants as described in claim 1, characterized in that, Between step S10 and step S20, knowledge graph reasoning is performed, wherein the knowledge graph reasoning includes the following steps: Step Z30: Construct a knowledge graph for gas turbine fuel switching. The knowledge graph stores the relationship between fuel type, chemical characteristic parameters and historical switching events in the form of triples. Each triple includes entity nodes and relation edges. The entity nodes are generated from historical fuel switching event data in the industrial database and cover fuel type entities, gas turbine model entities and operating condition entities. Step Z31: Dynamically construct relation edges through correlation analysis of parameter changes in historical event data. The relation edges represent the degree of quantitative impact of fuel type conversion on combustion efficiency. The correlation analysis generates relation edge weights by calculating the Pearson correlation coefficient of parameter changes. Step Z32: When a fuel switching command signal is received, the command signal is input to the graph neural network inference engine. The inference engine queries the knowledge graph based on the target fuel type in the command signal and outputs a reference vector of the switching influence between the target fuel type and the current fuel type. Step Z33: Concatenate the switching influence reference vector with the fuel calorific value sequence obtained in step S10 to generate an enhanced fuel feature vector; Step Z34: Input the enhanced fuel feature vector into the input tensor fusion process of step S20 to replace the original calorific value sequence in the efficiency mutation feature calculation. The query process of the graph neural network inference engine adopts a graph attention mechanism to dynamically adjust the relation edge weights according to the current gas turbine model entity.

7. The deep learning-based time series prediction method for gas turbine power plants as described in claim 1, characterized in that, After step S40, wolf pack algorithm optimization is performed, wherein the wolf pack algorithm optimization includes the following steps: Step Z40: Obtain the fuel switching scenario feature vector from the feature database. The feature vector is generated by step Z22 and includes the fuel switching influencing factor, combustion stability index and dynamic correction value. Step Z41: Input the feature vector into the wolf pack algorithm optimization module. The optimization module initializes the wolf pack individual set, which includes alpha wolf individuals, scout wolves individuals, and predator wolves individuals. The alpha wolf individual represents the global optimal solution, the scout wolf individual represents the local search direction, and the predator wolves individuals represent the parameter update step size. Step Z42: Calculate the fitness value of each wolf pack individual based on the feature vector. The fitness value is determined by the weighted sum of the dynamic correction value and the historical prediction deviation in the feature vector, and generate the parameter update direction vector through time series correlation analysis of the feature vector. Step Z43: Execute the iterative optimization process: the scout wolf performs a bounding search on the fitness value to generate a parameter update direction vector, the ferocious wolf calculates the step size adjustment based on the parameter update direction vector and updates the hyperparameters of the deep learning model, and the alpha wolf evaluates whether the updated hyperparameters meet the convergence condition. Step Z44: Input the optimal set of hyperparameters that meets the convergence condition into the parameter configuration unit of the deep learning model with enhanced attention mechanism, so as to dynamically adjust the learning rate, regularization coefficient and number of attention heads of the model. The convergence condition is dynamically set based on the changing trend of the combustion stability index in the feature vector. The output of the optimal set of hyperparameters triggers the instantaneous reloading mechanism of the model parameters.

8. A deep learning-based time series prediction system for gas turbine power plants, characterized in that, include: The data acquisition module is used to receive fuel switching command signals from the gas turbine control system; and to acquire gas turbine operating time series data in real time. The time series data includes combustion chamber temperature sequence, exhaust pressure sequence, fuel calorific value sequence and fuel flow sequence, and all data are stored in an industrial database with a unified timestamp. The data fusion module is used to fuse the time series data with the corresponding parameters in the fuel chemical property parameter database to generate an input tensor; the input tensor is input into an attention mechanism-enhanced deep learning model, which is based on the Transformer architecture and calculates adaptive correlation weights between time steps to quantify the fuel switching impact factor, which is calculated based on the abrupt deviation between the temperature sequence and the calorific value sequence in the input tensor. The data processing module is used to receive the fuel switching influencing factor through the fuel type identification module, and calculate the combustion stability index in real time by combining the temperature series and pressure series in the time series data. The combustion stability index is dynamically generated based on the weighted sum of the standard deviation of the temperature series fluctuation and the rate of change of the pressure series, wherein the standard deviation of the fluctuation is calculated from the sliding window variance of the temperature series, and the rate of change is obtained by the differential operation of the pressure series. The power generation prediction module is used to generate a dynamic correction value for power generation after fuel switching based on the combustion stability index. The dynamic correction value is used to adjust the slope of the prediction curve near the time point triggered by the fuel switching command signal through a linear interpolation method. The dynamic correction value is then fused with the output value of the basic power generation prediction model to output the real-time power generation prediction result.

Images