An online self-adaptive high-efficiency intelligent control method for micro gas turbine

Abstract

The application discloses a kind of micro gas turbine online self-adapting high-efficiency intelligent control method, principle as follows: the fuel flow and rotating speed of gas turbine control quantity such as input airborne deep prediction model, the state parameter prediction value of gas turbine is calculated in combination with the environmental condition of current time.The performance deviation determination threshold is set, when the state parameter prediction value is less than or equal to performance deviation determination threshold, the optimal rotating speed is calculated according to high-efficiency optimal rotating speed adjustment control method;When the state parameter prediction value is greater than performance deviation determination threshold, optimal rotating speed control law is obtained based on online deep extreme limit learning machine training, and update.The application decouples strong time-varying factor (environmental temperature) and weak time-varying factor (performance degradation).Strong time-varying factor is considered in high-efficiency control inner loop, and weak time-varying factors such as performance degradation are considered in adaptive update control outer loop, which maximizes the frequency of update control law, and is more convenient for engineering implementation of controller.

Description

Technical Field

[0001] This invention belongs to the field of gas turbine control technology, and particularly relates to an online adaptive high-efficiency intelligent control method for micro gas turbines. Background Technology

[0002] Wang et al. conducted the first comprehensive analytical study on the variable-speed operation performance of a single-shaft MGT. They derived an analytical solution for the optimal speed and analyzed the impact of different pressure ratios and temperature ratios on the performance of the variable-speed unit. The study showed that the temperature ratio has a much greater impact on efficiency and optimal speed than the pressure ratio, and the optimal speed can be summarized as a curve related to power. Through numerical simulations of a 3MW single-shaft gas turbine at different shaft speeds, they derived the optimal operating speed within a certain load range. The results showed that at 40% load, the gas turbine can achieve a 3.0% absolute efficiency improvement when operating at 83% of its rated speed. Compared with the efficiency of constant-speed operation, this is equivalent to a 12.8% relative efficiency improvement. Furthermore, the lower the ambient temperature, the more significant the efficiency improvement. Duan et al., based on a nonlinear simulation model of the MGT, compared the impact of constant-speed and variable-speed operation modes on the thermal efficiency of the gas turbine under regenerative and simple cycles. The results showed that the variable-speed operation mode of the regenerative cycle can achieve the highest overall thermal efficiency. Other studies have established dynamic simulation models of MGT power generation systems based on the Rowen model. By adopting a variable speed operation mode, the operating efficiency of the gas turbine is improved by establishing a polynomial relationship between the optimal speed and power.

[0003] The polynomial-based highest efficiency control described above only calculates the optimal speed value at the current moment by power calculation, ignoring the changes in optimal speed caused by ambient temperature and engine performance deviations. Summary of the Invention

[0004] To address the aforementioned technical problems, this invention proposes an online adaptive high-efficiency intelligent control method for micro gas turbines, thereby resolving the issues present in the prior art.

[0005] To achieve the above objectives, this invention provides an online adaptive high-efficiency intelligent control method for micro gas turbines, comprising the following steps:

[0006] A high-efficiency optimal speed regulation control method based on BP neural network;

[0007] A deep extreme learning machine prediction model is constructed based on deep learning networks and extreme learning machines.

[0008] An online deep extreme learning machine update model is constructed based on deep learning networks and online sequential extreme learning machines;

[0009] The fuel flow rate and rotational speed of the gas turbine are input into the deep extreme learning machine prediction model to calculate the predicted values ​​of the gas turbine's state parameters;

[0010] Set a performance deviation judgment threshold. When the predicted value of the state parameter is less than or equal to the performance deviation judgment threshold, calculate the optimal speed according to the high-efficiency optimal speed regulation and control method.

[0011] When the predicted value of the state parameter is greater than the performance deviation judgment threshold, the optimal speed control law is obtained by training based on the online deep extreme learning machine. The BP speed regulation law in the deep extreme learning machine prediction model and the high-efficiency optimal speed regulation control method is updated based on the optimal speed control law. The gas turbine is controlled based on the adjusted BP speed regulation law.

[0012] Preferably, the high-efficiency optimal speed regulation control method includes:

[0013] Using ambient temperature and load power as inputs, the speed at which the gas turbine achieves maximum efficiency under the input conditions is calculated using the BP speed regulation law.

[0014] Based on the optimal speed value, the gas turbine speed controller adjusts the fuel flow rate to make the actual shaft speed of the gas turbine track the change of the optimal speed.

[0015] Preferably, the method for regulating the BP speed includes:

[0016] N speed values ​​are uniformly selected within the operating speed range of the gas turbine. These N speed values ​​are input into the nonlinear model of the engine. Speed ​​values ​​that cause overheating and surge of the gas turbine are removed to obtain a candidate dataset. The speed corresponding to the highest efficiency is selected from the candidate dataset. Then, a BP neural network is trained using the selected speed and the corresponding ambient temperature and power. The trained BP neural network is used to obtain the speed value of the gas turbine with the highest efficiency under the same operating conditions.

[0017] Preferably, the deep learning network is constructed using an autoencoder, which includes an encoder and a decoder; the encoding process of the encoder is as follows:

[0018] h = g(W) 1 x+b 1 );

[0019] In the formula, g(·) represents the encoder activation function, and W 1 and b 1 Let represent the encoder's weight and bias matrices, respectively, and let x represent the input data;

[0020] The decoding process of the decoder is as follows:

[0021] z = f(W)2 x+b 2 );

[0022] In the formula, f(·) represents the decoder activation function, and W 2 and b 2 These represent the decoder weights and bias matrix, respectively.

[0023] Preferably, the method for obtaining the predicted values ​​of the state parameters includes:

[0024] The deep extreme learning machine prediction model includes a deep extreme learning machine-autoencoder deep network and a transformation model based on extreme learning machine;

[0025] First, the output matrix is ​​obtained based on the Deep Extreme Learning Machine-Autoencoder Deep Network. Then, a nonlinear relationship between the output matrix and the output value of the Deep Extreme Learning Machine Prediction Model is established through an ELM-based transformation model. Based on this nonlinear relationship, the predicted value is obtained.

[0026] Preferably, when the deep extreme learning machine prediction model updates the data, it balances the old sample dataset and the new sample dataset based on a dynamic forgetting factor;

[0027] The balancing method includes: calculating the training error of the deep extreme learning machine prediction model, not forgetting the old sample dataset when the training error is less than a training error threshold, and forgetting the old sample dataset when the training error is greater than or equal to the training error threshold.

[0028] Preferably, the update function expression for the dynamic forgetting factor is:

[0029]

[0030] In the formula, e represents the current training error of the model, and λ min μ represents the minimum value of the forgetting factor, and μ represents the control λ towards λ. min The step size of the approximation rate, e t This indicates the set training error threshold.

[0031] Preferably, the control method further includes evaluating the performance of the deep extreme learning machine prediction model, and the evaluation indicators include root mean square error, mean absolute error, mean absolute percentage error, and correlation coefficient.

[0032] Compared with the prior art, the present invention has the following advantages and technical effects:

[0033] This invention discloses an online adaptive high-efficiency intelligent control method for micro gas turbines. The principle of the adaptive high-efficiency intelligent control is as follows: Control variables such as fuel flow rate and speed of the gas turbine are input into an onboard deep prediction model. The predicted values ​​of the gas turbine's state parameters are calculated based on the current environmental conditions. A performance deviation judgment threshold is set. When the predicted value of the state parameters is less than or equal to the performance deviation judgment threshold, the optimal speed is calculated according to the high-efficiency optimal speed regulation control method. When the predicted value of the state parameters is greater than the performance deviation judgment threshold, training is performed based on an online deep extreme learning machine to obtain the optimal speed control law, which is then updated. This invention decouples strongly time-varying factors (ambient temperature) and weakly time-varying factors (performance degradation). Strongly time-varying factors are considered in the high-efficiency control inner loop, while weakly time-varying factors such as performance degradation are considered in the adaptive update control outer loop, minimizing the frequency of control law updates and facilitating the engineering implementation of the controller. The online adaptive high-efficiency control strategy proposed in this invention overcomes the shortcomings of traditional polynomial-based control strategies, which cannot adapt to the time-varying nature of gas turbines. This invention designs a BP optimal speed control law, comprehensively considering the influence of ambient temperature and power on the optimal speed. By designing an online update mechanism, it solves the problem of optimal speed adaptability under engine performance deviations. It can update the optimal speed under different ambient temperatures and engine performance deviations, ensuring the effectiveness of high-efficiency control of the gas turbine in various operating scenarios. Attached Figure Description

[0034] The accompanying drawings, which form part of this application, are used to provide a further understanding of this application. The illustrative embodiments and descriptions of this application are used to explain this application and do not constitute an undue limitation of this application. In the drawings:

[0035] Figure 1 This is a schematic diagram of the online adaptive high-efficiency intelligent control method according to an embodiment of the present invention;

[0036] Figure 2 This is a block diagram of the speed optimization control law in an embodiment of the present invention;

[0037] Figure 3 This is a graph showing the effect of ambient temperature on the optimal speed of the MGT in an embodiment of the present invention.

[0038] Figure 4 This is a block diagram of the optimal speed regulation control under environmental factor coupling in an embodiment of the present invention;

[0039] Figure 5 This is a diagram of the DL structure constructed by an automatic encoder (AE) according to an embodiment of the present invention;

[0040] Figure 6 This is a block diagram of the DL_ELM structure according to an embodiment of the present invention;

[0041] Figure 7 This is a schematic diagram illustrating the dynamic forgetting factor adjustment mechanism according to an embodiment of the present invention.

[0042] Figure 8 This is a prediction result diagram of the DL_ELM algorithm in an embodiment of the present invention, where (a) is T t,in Fitted plot, (b) is T t,in Relative error plot, (c) is T t,out Fitted plot, (d) is T t,out Relative error plot, (f) is N out Fitted plot, (h) is N out Relative error plot;

[0043] Figure 9 This is a schematic diagram showing the changes in the performance parameters of the gas turbine under different control schemes when the load power increases from 0.4 to 0.6 according to an embodiment of the present invention. (a) is the output power curve, (b) is the speed plan curve, (c) is the efficiency curve, (d) is the fuel flow curve, (e) is the turbine inlet temperature curve, and (f) is the turbine exhaust temperature curve.

[0044] Figure 10 This is a schematic diagram of the error values ​​of the DL_ELM airborne prediction model under different load powers according to an embodiment of the present invention;

[0045] Figure 11 The following are high-efficiency operation curves of the gas turbine under different ambient temperatures according to embodiments of the present invention, wherein (a) is an ambient temperature curve, (b) is a speed planning curve, (c) is an efficiency curve, (d) is an output power curve, (e) is a turbine inlet temperature curve, and (f) is a turbine exhaust temperature curve.

[0046] Figure 12 This is a schematic diagram illustrating the error values ​​of the DL_ELM airborne prediction model under different ambient temperatures according to an embodiment of the present invention.

[0047] Figure 13 This is a schematic diagram of the error values ​​of the DL_ELM prediction model under different turbine degradation levels according to an embodiment of the present invention;

[0048] Figure 14 This is a block diagram of the optimal speed regulation control according to an embodiment of the present invention. Detailed Implementation

[0049] It should be noted that, unless otherwise specified, the embodiments and features described in this application can be combined with each other. This application will now be described in detail with reference to the accompanying drawings and embodiments.

[0050] It should be noted that the steps shown in the flowchart in the accompanying drawings can be executed in a computer system such as a set of computer-executable instructions, and although a logical order is shown in the flowchart, in some cases the steps shown or described may be executed in a different order than that shown here.

[0051] Example 1

[0052] like Figure 1 As shown, this embodiment provides an online adaptive high-efficiency intelligent control method for a micro gas turbine, including the following steps:

[0053] A high-efficiency optimal speed regulation control method based on BP neural network;

[0054] A deep extreme learning machine prediction model is constructed based on deep learning networks and extreme learning machines.

[0055] Construct an online deep extreme learning machine based on deep learning networks and online sequential extreme learning machines;

[0056] The fuel flow rate and rotational speed of the gas turbine are input into the deep extreme learning machine prediction model to calculate the predicted values ​​of the gas turbine's state parameters;

[0057] Set a performance deviation judgment threshold. When the predicted value of the state parameter is less than or equal to the performance deviation judgment threshold, calculate the optimal speed according to the high-efficiency optimal speed regulation and control method.

[0058] When the predicted value of the state parameter is greater than the performance deviation judgment threshold, the optimal speed control law is obtained by training based on the online deep extreme learning machine. The BP speed regulation law in the deep extreme learning machine prediction model and the high-efficiency optimal speed regulation control method is updated based on the optimal speed control law. The gas turbine is controlled based on the adjusted BP speed regulation law.

[0059] Furthermore, the principle of the adaptive update control law is as follows: 1. Input the control quantities such as fuel flow rate and speed of the gas turbine into the onboard depth ELM (DL_ELM) prediction model, and calculate the predicted values ​​of the gas turbine's state parameters based on the current environmental conditions. 2. Set the performance deviation judgment threshold e. max When the deviation between the predicted model output value and the sensor measurement value |ΔT| < e max At that time, the optimal speed is calculated according to the original control law. |ΔT|>e max 3. Train the real-time optimal speed control law using an online deep extreme learning machine (DL_OSELM), and update the high-efficiency optimal speed control law and the airborne DL_ELM prediction model.

[0060] Furthermore, a high-efficiency optimal speed regulation control method based on a BP neural network is established as follows:

[0061] In reality, ambient temperature significantly affects the thermal efficiency of gas turbines. To analyze the impact of ambient temperature on the optimal control law for high-efficiency gas turbines, this study investigates the high-efficiency control of gas turbines under the influence of ambient temperature.

[0062] The study of optimal high-efficiency control of a gas turbine (MGT) can be understood as finding the speed value that maximizes the efficiency of the MGT under different operating conditions. The steady-state operating point of the gas turbine is determined by both environmental conditions and control variables. Considering the highest efficiency control under ambient temperature, it can be simplified to solving for the optimal speed values ​​under different operating conditions (T0, G). f The optimal speed value under the combination. Figure 2 The optimal speed calculation block diagram is shown. Eight parameters are uniformly selected within the range of ambient temperature T0 from 273.15K to 308.15K, and 90 points are uniformly selected within the normal operating flow range for fuel flow. Therefore, the optimal speed value of the gas turbine needs to be calculated under 720 operating conditions. 300 speed values ​​are uniformly selected within the range of 36000 r / min to 51000 r / min and input into the engine's nonlinear model. Speed ​​values ​​that cause MGT overheating and surge are removed, and the speed value that maximizes the gas turbine's efficiency under the same operating conditions is found, constituting the optimal speed regulation and control law for the gas turbine. From... Figure 3 It can be seen that the ambient temperature T0 significantly affects the optimal speed control law of the gas turbine. Under the same load power, the optimal speed of the gas turbine gradually increases with the increase of ambient temperature.

[0063] Neural networks, characterized by their simplicity, fast convergence, and strong generalization ability, are widely used in nonlinear multidimensional variable fitting problems. This invention uses ambient temperature and load power as input conditions and optimal speed as the output condition, employing a BP neural network to train the optimal speed control law. It is noteworthy that fuel flow rate is not used as an input parameter here. This is because actual control systems adjust the gas turbine's operating conditions by providing load commands; using load power instead of fuel flow rate as the input parameter simplifies controller design. The block diagram for the high-efficiency optimal speed regulation control considering the influence of ambient temperature is shown below. Figure 4 As shown. First, using the current ambient temperature and load power as input conditions, the optimal speed for the gas turbine to achieve maximum efficiency under the current conditions is calculated based on the trained BP speed regulation law. Second, the gas turbine speed controller adjusts the fuel flow to make the actual shaft speed of the gas turbine track the changes in the optimal speed.

[0064] Furthermore, the online adaptive high-efficiency control method consists of an optimal high-efficiency control element and an adaptive update control law element. An optimal high-efficiency control element was developed based on the BP optimal speed control strategy to calculate the optimal speed under different loads and ambient temperatures, thereby achieving high-efficiency operation of the gas turbine.

[0065] When gas turbine performance degrades, relying solely on the optimal speed control loop is insufficient to achieve stable, high-efficiency operation. To adapt the optimal speed control law to changes in gas turbine characteristics, an adaptive update control outer loop is added to the optimal high-efficiency control loop, forming an online adaptive high-efficiency control law. The principle of the adaptive update control law is as follows: 1. Input the gas turbine's fuel flow rate and speed, among other control parameters, into the onboard depth ELM (DL_ELM) prediction model, and calculate the predicted values ​​of the gas turbine's state parameters based on the current environmental conditions. 2. Set a performance deviation judgment threshold e. max When the deviation between the predicted model output value and the sensor measurement value |ΔT| < e max At that time, the optimal speed is calculated according to the original control law. |ΔT|>e max 3. Train the real-time optimal speed control law using an online deep extreme learning machine (DL_OSELM), and update the high-efficiency optimal speed control law and the airborne DL_ELM prediction model.

[0066] The adaptive high-efficiency control system designed in this invention decouples strong time-varying factors (ambient temperature) from weak time-varying factors (performance degradation). Strong time-varying factors are considered in the inner loop of the high-efficiency control, while weak time-varying factors such as performance degradation are considered in the outer loop of the adaptive update control. This minimizes the frequency of control law updates and facilitates the engineering implementation of the controller. The airborne predictive model and the online identification model are the core of the online adaptive high-efficiency control strategy, respectively influencing the implementation of the adaptive update control law and the optimal speed regulation strategy. Therefore, the airborne predictive model and the online identification model must possess high simulation accuracy to ensure the effectiveness of the online adaptive high-efficiency control system.

[0067] Furthermore, the deep extreme learning machine prediction model is structured as follows:

[0068] 1. Extreme Learning Machine (ELM)

[0069] Given a training sample S = {(x i ,Y i )|x i ∈R s ,Y i ∈R t An ELM with i = 1, 2, 3…M, having s input nodes, L hidden nodes, and t output nodes can be described as follows:

[0070]

[0071] Where x is the input parameter, β = [β1, β2, ..., β] L [] represents the output weight coefficient matrix connecting the hidden layer and the output layer, g(·) is the non-linear activation function, and w i ,b i The parameters for the hidden layer nodes are randomly generated. The hidden layer output matrix Ω can be represented as:

[0072]

[0073] The ELM algorithm transforms the network between the input and hidden layers into a fixed nonlinear transformation by randomly initializing hidden nodes, thus reducing the number of training parameters. After this fixed nonlinear transformation in the hidden layers, the input parameters are mapped to a new linear space, called the ELM feature space. The unknown parameter β can be obtained by minimizing ||Ωβ-Y||.

[0074]

[0075] in, It is the Moore-Penrose generalized inverse of the matrix. To avoid overfitting and ensure generalization ability, l is added to the objective function. 2,1 Norm regularization ||β|| 2 C / 2. The loss function is shown in Equation 19.

[0076]

[0077] make We can obtain:

[0078] β=(Ω T Ω+CI) -1 Ω T Y (5)

[0079] 2. Deep Learning Autoencoder

[0080] Deep learning is a branch of neural networks that simulates the multi-layered cognition of the human brain, obtaining high-level features and distributed data structures. Deep learning overcomes the limitation of traditional shallow neural networks in accurately representing complex problems and has been widely applied in many fields. This section primarily introduces deep learning built using autoencoders (AEs). Figure 5 As shown, suppose a deep learning algorithm (DL) consists of Q hidden layers. Where L... k This represents the number of nodes in the k-th hidden layer.

[0081] Each layer of the array (AE) consists of an encoder and a decoder. Assuming x is the input data, W... 1and b 1 These are the encoder weights and bias matrix, W, respectively. 2 and b 2 These are the decoder weights and bias matrix, respectively. The encoding process is shown in equation (6):

[0082] h = g(W) 1 x+b 1 (6)

[0083] Where g(·) is the activation function, and h is the output of the hidden layer.

[0084] The decoding process is shown in equation (7):

[0085] z = f(W) 2 x+b 2 (7)

[0086] Here, the output z is the reconstruction matrix of the input x, containing the important features of x. The goal of AE is to minimize the reconstruction error. Its loss function is as follows:

[0087]

[0088] 3. The proposed DL_ELM prediction model

[0089] like Figure 6 As shown, DL_ELM consists of the DL_ELM_AE deep network and an ELM-based transformation model.

[0090] 1) DL_ELM_AE combines the ability of deep learning to extract complex nonlinear features with the fast training speed of the ELM algorithm, exhibiting excellent nonlinear feature extraction capabilities. DL_ELM_AE achieves feature extraction of input parameters by stacking Q ELM_AEs. The biggest difference between ELM_AE and ELM is that the ELM_AE algorithm is an unsupervised learning method where input and output are equivalent. Taking the i-th ELM_AE as an example, assuming the output matrix of the (i-1)-th hidden layer is H... (i-1) The output weights of ELM_AE can be calculated according to equation (5):

[0091]

[0092] like Figure 6 As shown, DL_ELM_AE uses the output of each hidden layer as the input to the next ELM_AE. Define H... (0) If x = , then the output matrix of the DL_ELM_AE hidden layer can be calculated as follows:

[0093] H (i) =g(H (i-1) (β (i) )T i = 1, 2, ..., Q (10)

[0094] Finally, the output matrix a(t) of the DL_ELM_AE algorithm can be calculated:

[0095] a(t)=f5(x)=g((…g(x(β (1) ) T )…)(β (Q) ) T (11)

[0096] 2) The purpose of constructing the ELM-based transformation model is to establish a nonlinear relationship between a(t) and the model output value Y. Here, a(t) acts as the input, and Y acts as the output. The predicted value of the model can be calculated according to formula (1):

[0097]

[0098] 4. Online update strategy based on dynamic forgetting factor

[0099] Both traditional ELM algorithms and gradient-based identification algorithms are offline identification methods. When sufficient sample stimulation is obtained and system characteristics do not change over time, offline identification methods can achieve satisfactory modeling results. However, traditional offline models cannot track changes in gas turbine characteristics caused by component performance degradation. This invention proposes an online version of the DL_ELM prediction model (Online Time Series Deep Extreme Learning Machine). It utilizes a dynamic forgetting factor to balance the old and new sample datasets, adaptively forgetting old sample data when engine characteristics change, and updating model parameters in real time. Assume the initial training sample set S0 = {(x... i ,Y i )|x i ∈R s ,Y i ∈R t ,i=1,2,3…M0}. Calculate the output weight matrix using equation (5). in When new samples Upon arrival, the output weight β1 can be calculated as:

[0100]

[0101] Ω L,1 Y1 and Y2 are the hidden layer matrix and output weights associated with S1, respectively. However, for time-varying systems like gas turbines, the timeliness of old sample data is limited. Therefore, a forgetting factor λ∈(0,1) is introduced to gradually forget the old sample data. Equation (13) can be rewritten as:

[0102]

[0103] in

[0104] Similarly, when the (m+1)th block of the dataset Upon arrival, the output weight β m+1 The update formula is as follows:

[0105]

[0106] As can be seen from equation (15), the forgetting factor λ∈(0,1] determines the proportion of old and new samples in the model. When λ<1, the old sample data will be gradually forgotten as the number of iterations increases. In actual engineering processes, the characteristics of the system do not change according to the same pattern, and a constant forgetting factor cannot guarantee that the model will obtain satisfactory performance under different changes in the system. In order to accurately track the characteristic changes of the gas turbine, this invention improves the forgetting factor by using a new dynamic forgetting factor (DFF) to gradually discard the old collected data and pay more attention to the newly collected data according to the estimation error to enhance the tracking ability. In addition, an update selection strategy (USS) is proposed to determine whether the model needs to be updated according to the new data, so that the proposed method is more in line with the actual modeling situation. The update function of the forgetting factor is shown in equation (16):

[0107]

[0108] Where e is the current training error of the model, and λ min It is the minimum value of the forgetting factor, μ is the control λ to λ min The step size of the approximation rate, e t This is the set training error threshold. When e < e t When e ≥ e, old sample data is not forgotten (λ = 1). t At this point, it is proven that the model identified using old sample data can no longer represent the true characteristics of the gas turbine. At this time, λ should gradually approach the set minimum value of the forgetting factor λ. min The model dynamically forgets old sample data during the parameter update process. Figure 7 A schematic diagram of the forgetting factor update function.

[0109] In summary, the proposed DFF_DL_OSELM algorithm is shown in Table 1:

[0110] Table 1

[0111]

[0112] This embodiment is verified through the following simulation experiment:

[0113] 1. Simulation parameter settings

[0114] This invention selects fuel signal G f Ambient temperature T0 and shaft speed n are used as input signals. Exhaust temperature T t,out Turbine inlet temperature T t,in Output power N out The DL_ELM prediction model is built using the output signal of the model. All state parameters are normalized before simulation. The model can be described as follows:

[0115]

[0116] Eight values ​​were uniformly selected from the ambient temperature T0 within the range of 273.15K to 308.15K, and the fuel flow rate G... f The input dataset consists of 26 uniformly selected values ​​within the range [0.0050, 0.0075] and 301 uniformly selected values ​​for n within the range [36000, 51000]. This dataset is then used as input to simulate the mechanism model of the gas turbine, yielding the corresponding Ti values. t,in T t,out and N out The parameter output values ​​were filtered out, removing data points that caused MGT to exceed limits, resulting in 9351 sets of data. Additionally, the ambient temperature was changed to 279.65K, 389.65K, and 299.65K respectively to simulate the mechanistic model, obtaining 3254 sets of data. The first dataset was used as the training dataset, and the second dataset as the validation dataset.

[0117] The root mean square error (RMSE), mean absolute error (MAE), mean absolute percentage error (MAPE), and correlation coefficient (R²) were used. 2 The performance of the model is evaluated using equations (18) to (21).

[0118]

[0119] Where Y(i) is the actual target value, M represents the target value to be predicted, and M is the number of samples.

[0120] 2. Accuracy Verification of Airborne Prediction Model

[0121] To verify the superiority of the DL_ELM algorithm proposed in this invention, the algorithm is compared with the traditional ELM algorithm and the Multilevel Online Extreme Learning Machine (ML-ELM) algorithm. For the traditional ELM, the number of hidden layer nodes L is set to 100, 300, and 3000 respectively. The number of hidden layers in the ML_ELM and DL_ELM algorithms is set to 3, and the regularization parameter C of the DL_ELM algorithm is set to 10.-6 The sigmoid function was chosen as the activation function: g(x) = 1 / (1 + exp(-x)). A genetic algorithm was used to find the optimal combination of remaining hyperparameters within a relatively large space. The number of hidden layer nodes after optimization of the ML_ELM algorithm was 170, 60, and 180, respectively, while the number of hidden layer nodes for the DL_ELM algorithm was set to 248, 177, and 393, respectively. The recognition accuracy of the ELM, ML_ELM, and DL_ELM algorithms is recorded in Table 2. Table 2 shows that increasing the number of hidden layer nodes in ELM does not improve the model accuracy; instead, it significantly increases the model training time. ML_ELM, by expanding the hidden layers of the ELM algorithm, achieved lower RMSE, MAE, MAPE, and a higher model correlation coefficient R0. 2 Therefore, increasing the number of hidden layers in the ELM algorithm is an effective way to improve model identification accuracy. Compared with the ELM and ML_ELM algorithms, the proposed DL_ELM algorithm achieves the lowest RMSE, MAE, MAPE, and the highest model correlation coefficient R. 2 .

[0122] Table 2

[0123]

[0124]

[0125] Figure 8 The prediction performance of the DL_ELM algorithm is presented. Figure 8 The left side shows the fitting plot of the predicted and observed values ​​of the output parameters, while the right side shows the relative error of each predicted parameter. Figure 8 It can be seen that the DL_ELM algorithm is effective for T. t,in T t,out N out The fitted points between the predicted and observed values ​​are concentrated near the ideal curve. The DL_ELM algorithm applies this to T... t,in T t,out The relative error of the prediction is no greater than 1% for N out The prediction relative errors are concentrated below 1%, with a small number of outliers between 1% and 4%, and the maximum relative error is 3.9%. This indicates that the method has a good predictive effect on the output parameters of gas turbines.

[0126] 3. Online identification model accuracy verification

[0127] To verify the ability of the proposed DL_ELM airborne prediction model to track changes in gas turbine characteristics, the performance of the DL_ELM algorithm was tested using data from compressor and turbine component performance degradation at 1%, 3%, and 5%, respectively. Table 3 shows that compared to the prediction results in Table 2 when the gas turbine is in a healthy state, the RMSE, MAE, and MAPE of the DL_ELM algorithm significantly increase after component performance degradation, and the correlation coefficient R0 also increases. 2 The performance of the DL_ELM algorithm decreased significantly as the degree of engine component degradation increased. This indicates that DL_ELM, trained based on healthy state data, is not suitable for predicting the state parameters of gas turbines after performance degradation. This invention develops an online version of DL_ELM (the DFF-based online deep extreme learning machine (DFF_DL_OSELM) algorithm) to adapt to the time-varying characteristics of gas turbines.

[0128] Table 3

[0129]

[0130] This invention selects the MAPE index as the update index for the dynamic forgetting factor in Equation 32. To illustrate the superiority of the proposed DFF_DL_OSELM algorithm, online update models for Deep Online Extreme Learning Machine (DL_FOSELM) and Deep Online Time Series Extreme Learning Machine (DL_OSELM) based on a fixed forgetting factor are established respectively. DL_FOSELM forgets old sample data in chronological order at a fixed ratio. DL_OSELM is a special case of DL_FOSELM (forgetting factor λ = 1), which does not discard original training data during training. The hyperparameter values ​​of DL_OSELM, DL_FOSELM, and DFF_DL_OSELM are determined through multiple simulations and recorded in Table 4.

[0131] Table 4

[0132]

[0133] A.MGT's recognition performance when the model is updated online in a healthy state

[0134] The acquired 9351 sets of operational data under gas turbine health conditions were used as training data for the online update model. The first 6000 sets were used for initial training, and the remainder were used as the update dataset. RMSE and R... 2 RMSE, MAE, and MAPE are used as metrics to evaluate the training accuracy of the model, and the results are recorded in Table 5. It can be seen that the RMSE and RPE of the DFF_DL_OSELM algorithm and the DL_OSELM algorithm... 2The RMSE, MAE, and MAPE indices of the DL_FOSELM algorithm are similar to those of the offline DL_ELM algorithm, proving that it can meet the performance tracking requirements under healthy gas turbine conditions. However, the RMSE, MAE, and MAPE indices of DL_FOSELM all show significant increases, indicating poor identification performance under healthy gas turbine conditions. This phenomenon is because after the arrival of new sample data, DL_FOSELM forgets old sample data at a fixed ratio, resulting in the loss of some useful information and ultimately causing a decline in identification performance. The DL_OSELM algorithm retains all data information during training, therefore, it does not adversely affect model training when the gas turbine characteristics remain unchanged. The DFF_DL_OSELM algorithm can adaptively adjust the size of the forgetting factor; when the model training error is small, increasing the value of the forgetting factor maximizes the retention of useful information from model training.

[0135] Table 5

[0136]

[0137]

[0138] B.MGT performance degradation and online model update recognition effect

[0139] A component degradation factor for the gas turbine was set, and simulation data were obtained when the performance of the gas turbine compressor and turbine degraded by 3%. 986 sets of compressor degradation data and 910 sets of turbine degradation data were obtained. The first 800 sets of data from the degradation dataset were used as supplementary data for the identification model, and the remaining data were used to verify the model's tracking effect on the time-varying characteristics of the gas turbine. Unlike its performance in a healthy gas turbine state, the DL_FOSELM algorithm achieved excellent identification results after the gas turbine characteristics changed. This is mainly because the DL_FOSELM algorithm gradually discards invalid old sample data according to the time series using a forgetting factor, giving new data greater weight, and can quickly track changes in system characteristics. However, the DL_OSELM algorithm lacks a forgetting mechanism, causing erroneous information in the old sample data to interfere with the model update process, ultimately leading to a decline in model training performance. When system characteristics change significantly, the DFF_DL_OSELM algorithm can quickly update the model by reducing the value of the forgetting factor and rapidly discarding erroneous information from the old model. The performance comparison of the online model update under gas turbine performance degradation is shown in Table 6.

[0140] Table 6

[0141]

[0142] In summary, the DL_ELM algorithm is only suitable for offline training and cannot be used as an online update model for gas turbines. The DL_OSELM algorithm exhibits excellent training performance when gas turbine characteristics remain constant, but its recognition performance significantly decreases after system characteristics change. The DL_FOSELM algorithm can track changes in system characteristics, but its training accuracy is poor when system characteristics remain constant. The dynamic forgetting factor update strategy proposed in this invention adaptively adjusts the forgetting factor based on the training error of the preceding model, effectively achieving model updates after changes in gas turbine characteristics while ensuring model recognition stability under healthy gas turbine operating conditions.

[0143] 4. Comparison of Control Effects

[0144] To verify the superiority of the high-efficiency control system designed in this invention, the control effects of high-efficiency control based on polynomial fitting (Case 1), high-efficiency control based on BP neural network (Case 2), and online adaptive high-efficiency control (Case 3) were compared under constant external environmental conditions, changing external environmental conditions, and changes in engine characteristics. Figure 14 , Figure 4 The control systems for Case 1 and Case 2 were designed based on the principle of [the principle], and both speed controllers used PID controllers. MAPE was selected as [the appropriate controller]. Figure 1 The threshold e is the criterion for determining the adaptive update process in the middle stage. max =0.005.

[0145] (1) Constant external environmental conditions

[0146] Figure 9 This diagram illustrates the performance parameter changes of a gas turbine under different control schemes when the load power increases from 0.4 to 0.6 under ISO conditions (T0 = 273.15 K, P0 = 101.325 kPa). The load power command input for the gas turbine is as follows: Figure 9 As shown in (a). The MAPE value of the airborne prediction model in Case 3 is displayed. Figure 10 It is worth noting that the prediction error based on the DL_ELM prediction model is less than the update threshold e. max The speed control rules in Case 3 and Case 2 are consistent. Figure 9 (b) shows the planned speed curves of the gas turbine. Compared with constant speed control, Case 1, Case 2, and Case 3 reduce the operating flow of the gas turbine by decreasing the shaft speed, effectively increasing the turbine inlet temperature. Figure 9 (e) and overall efficiency ( Figure 9(c)). Under operating conditions of 0.4 and 0.6, the speed regulation control improved thermal efficiency by 4.17% and 3.70% respectively compared to the traditional constant speed control mode. From Figure 9 (e) It can be seen that under constant speed control, the turbine inlet temperature of the gas turbine decreases significantly with the decrease of load power. Case 1, Case 2, and Case 3 can all maintain the turbine inlet temperature at a high level by adjusting the shaft speed. In summary, under constant external conditions, Case 1, Case 2, and Case 3 can all effectively achieve high-efficiency control of the gas turbine.

[0147] (2) Changing external environmental conditions

[0148] To verify the control effect of the proposed high-efficiency control scheme under different ambient temperatures, Figure 11 (a) shows the ambient temperature curve, which simulates the changing trend of ambient temperature throughout the day. The simulation is performed using a model under operating condition 0.5, and the simulation time is 20,000 s. Figure 12 This represents the prediction error MAPE value of the airborne prediction model based on the DL_ELM algorithm in Case 3. During the 20,000-second simulation, the MAPE value of Case 3 did not exceed the update threshold e. max In Case 3, the adaptive update outer loop is not activated, and the speed control law of Case 3 is consistent with that of Case 2.

[0149] from Figure 11 (b) It can be seen that the rotational speed of Case 1 corresponds one-to-one with the load power, and changes in ambient temperature do not alter its rotational speed plan. However, ambient temperature significantly affects the internal temperature field of the gas turbine. During the warming phase of the ambient temperature, the turbine inlet and outlet temperatures of Case 1 gradually increase until they reach the maximum turbine exhaust temperature. During the cooling phase of the ambient temperature, the turbine inlet and outlet temperatures of Case 1 gradually decrease. Figure 11 (e) and Figure 11 (f)). Unlike Case 1, Case 2 and Case 3 adapt to changes in external environmental conditions by adjusting the shaft speed, which can achieve relative stability of turbine inlet temperature, turbine exhaust temperature and thermal efficiency under different ambient temperatures. Figure 11 The decrease in output power of Case 1 in (d) is because the present invention incorporates turbine outlet temperature limit protection control logic based on the actual unit control system. When Case 1 reaches the turbine outlet temperature limit, the limit protection control system reduces the output power of the MGT by decreasing the fuel flow, thus preventing further increase in turbine outlet temperature.

[0150] In summary, Case 1's optimal speed cannot track changes in ambient temperature, leading to efficiency degradation at low temperatures and overheating faults in the gas turbine at high temperatures. Cases 2 and 3, because they consider changes in environmental factors during the optimization of the speed control law based on BP, can adapt their speed plans to changes in ambient temperature, enabling the gas turbine to operate at high efficiency at all times. Therefore, Case 1's high-efficiency control method, which uses polynomial fitting of the functional relationship between optimal speed and load power, cannot achieve high-efficiency operation under varying external environmental factors and has limitations in practical engineering applications. Cases 2 and 3 can effectively achieve high-efficiency operation of the gas turbine under different times and seasonal temperatures.

[0151] (3) Performance degradation of gas turbines

[0152] Under load power of 0.4, 0.5, and 0.6 and ISO conditions, the turbine degradation degree ΔD is set respectively. t =0%, 1%, and 3% to verify the ability of the proposed control scheme to cope with changes in gas turbine characteristics. Figure 13 The MAPE values ​​of the onboard predictive model in Case 3 represent different health states of the gas turbine. It can be seen that after turbine degradation occurs, the onboard predictive model based on DL_ELM accurately identifies the changes in gas turbine characteristics, and the MAPE value increases with the severity of turbine degradation. After turbine degradation, the MAPE value of the predictive model exceeds the update threshold e. max The adaptive update outer loop is activated, and the control system trains the optimal speed control law of the engine based on the DFF_DL_OSELM online identification model.

[0153] Table 7 records the control effects of Case 1, Case 2, and Case 3 under different degrees of degradation. It can be seen that under a 0.5 load condition, as the degree of degradation increases, the turbine inlet and exhaust temperatures in Case 1 and Case 2 gradually rise, triggering the maximum exhaust temperature limit and ultimately causing the gas turbine output power to fail to reach the target value. Case 3, through adaptively updating the speed control law, adjusts the gas turbine speed plan in a timely manner, maintaining the turbine inlet and exhaust temperatures within a safe operating range by increasing the shaft speed after turbine degradation occurs. As shown in Table 7, compared to Case 1 and Case 2, Case 3 ensures that the gas turbine maintains high efficiency operation within a stable operating range.

[0154] Table 7

[0155]

[0156] The above are merely preferred embodiments of this application, but the scope of protection of this application is not limited thereto. Any variations or substitutions that can be easily conceived by those skilled in the art within the scope of the technology disclosed in this application should be included within the scope of protection of this application. Therefore, the scope of protection of this application should be determined by the scope of the claims.

Claims

1. A method for online adaptive high-efficiency intelligent control of a micro gas turbine, characterized in that, Includes the following steps: A high-efficiency optimal speed regulation control method based on BP neural network; A deep extreme learning machine prediction model is constructed based on deep learning networks and extreme learning machines. An online deep extreme learning machine update model is constructed based on deep learning networks and online sequential extreme learning machines; The fuel flow rate and rotational speed of the gas turbine are input into the deep extreme learning machine prediction model to calculate the predicted values ​​of the gas turbine's state parameters; Set a performance deviation judgment threshold. When the predicted value of the state parameter is less than or equal to the performance deviation judgment threshold, calculate the optimal speed according to the high-efficiency optimal speed regulation and control method. The high-efficiency optimal speed regulation control method includes: Using ambient temperature and load power as inputs, the speed at which the gas turbine achieves maximum efficiency under the input conditions is calculated using the BP speed regulation law. Based on the aforementioned highest efficiency speed value, the gas turbine speed controller adjusts the fuel flow rate to make the actual shaft speed of the gas turbine track the change of the optimal speed; The method for regulating the BP speed includes: N speed values ​​are uniformly selected within the operating speed range of the gas turbine. These N speed values ​​are input into the nonlinear model of the engine. Speed ​​values ​​that cause the gas turbine to overheat and surge are removed to obtain a candidate dataset. The speed corresponding to the highest efficiency is selected from the candidate dataset. Then, a BP neural network is trained using the selected speed and the corresponding ambient temperature and power. The speed value with the highest efficiency of the gas turbine under the same operating conditions is obtained through the trained BP neural network. When the predicted value of the state parameter is greater than the performance deviation judgment threshold, the optimal speed control law is obtained by training based on the online deep extreme learning machine. The BP speed regulation law in the deep extreme learning machine prediction model and the high-efficiency optimal speed regulation control method is updated based on the optimal speed control law. The gas turbine is controlled based on the adjusted BP speed regulation law.

2. The online adaptive high-efficiency intelligent control method for micro gas turbines according to claim 1, characterized in that, The deep learning network is constructed using an autoencoder, which includes an encoder and a decoder. The encoder's encoding process is as follows: h=g(W 1 x+b 1 ); In the formula, g(·) represents the encoder activation function, and W 1 and b 1 Let represent the encoder's weight and bias matrices, respectively, and let x represent the input data; The decoding process of the decoder is as follows: z=f(W 2 x+b 2 ); In the formula, f(·) represents the decoder activation function, and W 2 and b 2 These represent the decoder weights and bias matrix, respectively.

3. The online adaptive high-efficiency intelligent control method for micro gas turbines according to claim 1, characterized in that, The method for obtaining the predicted values ​​of the state parameters includes: The deep extreme learning machine prediction model includes a deep extreme learning machine-autoencoder deep network and a transformation model based on extreme learning machine; First, the output matrix is ​​obtained based on the Deep Extreme Learning Machine-Autoencoder Deep Network. Then, a nonlinear relationship between the output matrix and the output value of the Deep Extreme Learning Machine Prediction Model is established through an ELM-based transformation model. Based on this nonlinear relationship, the predicted value is obtained.

4. The online adaptive high-efficiency intelligent control method for micro gas turbines according to claim 1, characterized in that, When updating data, the deep extreme learning machine prediction model balances the old sample dataset and the new sample dataset based on a dynamic forgetting factor. The balancing method includes: calculating the training error of the deep extreme learning machine prediction model, not forgetting the old sample dataset when the training error is less than a training error threshold, and forgetting the old sample dataset when the training error is greater than or equal to the training error threshold.

5. The online adaptive high-efficiency intelligent control method for micro gas turbines according to claim 4, characterized in that, The update function expression for the dynamic forgetting factor is: In the formula, e represents the current training error of the model, and λ min μ represents the minimum value of the forgetting factor, and μ represents the control λ towards λ. min The step size of the approximation rate, e t This indicates the set training error threshold.

6. The online adaptive high-efficiency intelligent control method for micro gas turbines according to claim 1, characterized in that, The control method also includes evaluating the performance of the deep extreme learning machine prediction model, with evaluation metrics including root mean square error, mean absolute error, mean absolute percentage error, and correlation coefficient.

Images