Data-driven modeling and control method for flexible operation of cogeneration unit

Abstract

The invention discloses a data-driven modeling and control method for flexible operation of a combined heat and power generation unit. The method comprises the following steps: firstly, establishing a unit model based on an improved transform neural network framework; secondly, designing a model-free self-adaptive predictive control strategy optimized by adopting an enhanced frost ice optimization algorithm, and using the model-free self-adaptive predictive control strategy to accurately track an automatic power generation control instruction; and finally, verifying the feasibility of the modeling and control strategy based on a simulation platform. According to the method, deep learning modeling and model-free adaptive control are effectively combined, and the provided strategy can meet the flexibility requirement under deep peak regulation of the unit.

Description

Technical Field

[0001] The present invention relates to the technical field of modeling and control of flexible operation of thermal power generating units, and more specifically, to a data-driven modeling and control method for flexible operation of cogeneration units. Background Art

[0002] The traditional power system dominated by thermal power is transitioning to a clean power system dominated by wind power, photovoltaic power generation, etc. However, the large-scale grid connection of intermittent renewable energy poses a threat to the stable operation of the power grid. Therefore, in order to meet the growing stability requirements of the power grid, it is necessary to improve the flexible operation capability of cogeneration units.

[0003] Cogeneration units play an important role in the heat and electricity supply in North China due to their high thermal efficiency and low emissions. Therefore, improving the deep peak-shaving and rapid load response capabilities of cogeneration units has become a research focus. However, the units exhibit strong nonlinearity and severe multivariable coupling characteristics under wide load operating conditions, and the complex dynamic characteristics make it difficult to obtain a high-precision model of the units. In addition, frequent load fluctuations and thermal disturbances exceed the regulation capabilities of traditional controllers. Therefore, there is an urgent need to develop advanced modeling and control technologies to provide feasible solutions for cogeneration units to participate in deep peak-shaving more flexibly and efficiently.

[0004] In recent years, transformer neural networks based on attention mechanisms have been successfully applied to machine translation, natural language processing, sequence modeling and other fields due to their advantages in global feature extraction and long-term dependency capture. Transformer introduces a self-attention mechanism that enables the model to flexibly focus on elements at different positions in the sequence. This mechanism allows the model to associate each position in the sequence with all other positions without being restricted by distance, which is very suitable for the establishment of cogeneration unit models with time-dependent characteristics. However, most existing controllers rely on precise mathematical models, making it difficult for black box models obtained through data-driven modeling to serve controller design. Model-free adaptive control (MFAC), as a data-driven control, does not require precise mathematical models. It only relies on the input / output data of the system to establish a dynamic linearized data model at each operating point of the controlled object, and achieves effective control effects on nonlinear and coupled systems. Therefore, the model-free adaptive control technology is introduced into the load tracking control of multivariable strongly coupled unit units to improve the rapidity and accuracy of the unit's response to automatic power generation control instructions. Summary of the invention

[0005] The present invention aims to provide a data-driven modeling and control method for the flexible operation of cogeneration units, so as to improve the rapidity and accuracy of the load response of the units. The method fully considers the characteristics of the units under wide load operating conditions and large-scale peak regulation, such as multivariable coupling, strong nonlinearity, and large time lag, and combines the improved transformer data-driven modeling technology to accurately describe the complex dynamic characteristics of the system. The model built achieves high modeling accuracy within the wide load range of the unit; combined with the model-free adaptive predictive control strategy optimized by the enhanced frost and ice optimization algorithm (ERIME), the unit effectively tracks the automatic power generation control instructions.

[0006] The data-driven modeling and control method for the flexible operation of a cogeneration unit proposed in the present invention consists of the following five steps:

[0007] S1: Analyze the control difficulties of the boiler-turbine unit of a supercritical cogeneration unit;

[0008] S2: Establish an improved transformer neural network framework for wide load modeling of cogeneration units;

[0009] S3: Describe the principles of model-free adaptive predictive control technology and its design steps;

[0010] S4: Obtain optimal controller parameters based on enhanced frost and ice optimization algorithm;

[0011] S5: Rely on the simulation platform to verify the feasibility and performance of the proposed modeling and control methods.

[0012] S1: The flexibility of the power system requires improving the peak-shaving capability and rapid load response capability of the cogeneration unit. However, under large-scale variable load operation conditions, the difference between the slow dynamic characteristics of the direct current boiler and the fast dynamic characteristics of the steam turbine increases, and the significant multivariable coupling characteristics are not conducive to the design of flexible control strategies. Therefore, it is urgent to develop a data-driven modeling strategy to reasonably describe the dynamic characteristics of the cogeneration unit operating in a wide load range, so as to provide guarantees for the application of flexible control strategies on site. In addition, during the load regulation process, smaller load tracking errors and shorter stabilization times are required. At the same time, the cogeneration unit must maintain good anti-interference performance, stable main steam pressure, intermediate point temperature and heating pressure, and the main steam valve opening, heating valve opening, coal supply and water supply must change smoothly to ensure the economy and safety of the unit operation.

[0013] Based on the analysis of S1, it can be seen that the dynamic characteristics of the cogeneration unit under wide load operation conditions are changeable and multi-variable coupling is serious. It is difficult for traditional modeling methods to reasonably analyze the nonlinear mapping relationship between input variables and output variables. Therefore, it is necessary to design a data-driven modeling method to obtain a high-precision model to lay the foundation for the design of the controller. Step S2 can be concretized as follows:

[0014] S2.1: In order to improve the network's ability to acquire the time dependency of the direct current boiler-turbine coupling system of the cogeneration unit, the bidirectional gated recurrent unit (BiGRU) is applied to the position encoding of the transformer; the basic component unit of BiGRU is GRU, which consists of a reset gate and an update gate; by integrating the forget gate and input gate in the long short-term memory network into the update gate, invalid information can be effectively discarded, long-range dependencies can be captured, the number of parameters can be greatly reduced, and overfitting can be avoided. The description of GRU is as follows:

[0015] z t =σ(W z q′ t +U z h t-1 +b z ) (1)

[0016] r t =σ(W r q′ t +U r h t-1 +b r ) (2)

[0017] c t =tanh(W c q′ t +r t (U c ⊙h t-1 )+b c ) (3)

[0018] h t =c t ⊙(1-z t )+h t-1 ⊙z t (4)

[0019] where q' t is the input at time t, h t-1 is the unit output at time t-1, W z , W r and W c Respectively represent the update gate, reset gate and candidate variable at input q t 'The weight matrix under z , bz , U r , b r , U c , b c Represent the weight matrix and bias of the update gate, reset gate and candidate variable, respectively, z t 、r t and c t They are the outputs of the update gate, reset gate, and candidate variables respectively;

[0020] Information at future time points is conducive to capturing more accurate long-term dependencies. Therefore, a bidirectional GRU is used to fully extract relevant element information of the forward and backward sequences to further improve the modeling performance of the transformer network, as described below:

[0021]

[0022] S2.2: The transformer neural network based on the self-attention mechanism is good at capturing global features, while the convolutional neural network (CNN) can effectively utilize local features. Therefore, the complementary advantages of transformer and CNN are expected to solve the problem of dynamic characteristics analysis of cogeneration units with strong coupling of multiple variables. The multi-head self-attention mechanism is directly applied to the input Z, which is described as follows:

[0023]

[0024]

[0025] Where MSA(Q,K,V) represents a multi-head self-attention layer, Q, K, and V represent query, key, and value, respectively, MLP and LayerNorm represent a multi-layer feedforward network and layer normalization, respectively. and Z time Represents the output of MSA and MLP;

[0026] In order to further enhance the transformer network's ability to extract local features, a multi-receptive field convolution (MRFC) module is designed and integrated into the transformer's multi-head self-attention layer, as described below:

[0027]

[0028]

[0029] Among them, MRFC represents the multi-receptive field convolution layer, Represents the output of the convolution mechanism;

[0030] The MRFC module consists of token-wise convolution, point-wise convolution, Swish activation function and normalization. The token-wise convolution consists of convolutions of different receptive fields in parallel. Convolution kernels of different sizes are used to process the input in parallel, obtain features from multiple receptive fields, and use zero padding to ensure that the size of the feature map remains unchanged. The feature maps of multiple receptive fields are connected along the channel dimension to obtain the output of the point-wise convolution layer. Next, the point-wise convolution layer uses a convolution of size 1 to reduce the number of channels in the feature map of the token-wise convolution layer. The Swish activation function and normalization are conducive to further enhancing the convergence and generalization ability of the network.

[0031] The decoder receives long-term dependency and local feature information from the input variables provided by the encoder as well as the historical trend relationship of the output variables; this information is then integrated and fed into the multi-head self-attention layer to capture the nonlinear mapping relationship between the input variables and the output variables.

[0032] After determining the improved transformer neural network framework for wide load modeling of cogeneration units, the principle of model-free adaptive predictive control technology and its design steps are concretized in step S3:

[0033] S3.1: Represent the 350MW CHP unit as a multi-input multi-output nonlinear discrete-time system as follows:

[0034] Y(k+1)=f(Y(k),…Y(kn y ),U(k),…,U(kn u )) (10)

[0035] where U(k)∈R m×1 , Y(k)∈R l×1 They represent the m-dimensional input and l-dimensional output at time k, f is an unknown function, n y and n u is an unknown positive integer, representing the order of the system;

[0036] The once-through boiler-turbine coupled system of the cogeneration unit described in equation (10) can be described as follows after compact dynamic linearization:

[0037] ΔY(k+1)=Φ c (k)ΔU(k)(11)

[0038] where Φ c (k)∈R l×m , represents the pseudo-Jacobi matrix;

[0039] S3.2: The design goal of MFAPC is to find the best future control action to make the process output as close as possible to the reference trajectory of the CHP unit. The control performance consideration is to minimize the following function:

[0040]

[0041] Among them, N p and N u are the prediction time domain and the control time domain respectively, λ>0 is the weight coefficient, Y * (k+i) is the expected output of the system at time k+i;

[0042] According to the equivalent dynamic linear model (11), Y(k+i) in (12) can be expressed as:

[0043]

[0044]

[0045] Then, formula (13) can be reformulated as:

[0046]

[0047] in represents the Kronecker product;

[0048] definition Substituting formula (15) into formula (12), we can get ΔU Nu Taking the derivative and setting it equal to zero, we can get the following control algorithm:

[0049]

[0050] Therefore, the actual control input value at the current moment is:

[0051]

[0052] in

[0053] S3.3: In formula (14), Φ c (k+i),i=0,1,…,N u -1 is an unknown number and must be replaced by an estimated and predicted value. Consider the following estimation standard function:

[0054]

[0055] where μ>0 is a weight coefficient used to penalize excessive changes in the PJM estimate;

[0056] For formula (18) with respect to Φ c(k) By finding the extreme value, we can get the PJM estimation algorithm, that is,

[0057]

[0058] Where η∈(0,2] is the step size factor;

[0059] S3.4: Using Φ c The estimated value of (k) is used to establish an automatic regression model for the valuation sequence, and the following prediction formula is obtained:

[0060]

[0061] Among them, θ i ,i=1,…,n p is the coefficient, n p is an appropriate order, usually between 2 and 7;

[0062] According to formula (20), the prediction algorithm can be obtained as follows:

[0063]

[0064] where j = 1, 2, ..., N u –1;

[0065] definition It is described as follows:

[0066]

[0067] in

[0068] Based on the principle and design steps of the model-free adaptive predictive control technology described in step S3, it is proposed in step S4 to obtain the optimal controller parameters based on the enhanced frost and ice optimization algorithm. The specific process is as follows:

[0069] S4.1: After determining the structure of the model-free adaptive predictive controller, the optimal parameters λ, μ, η and δ of the controller are obtained by the enhanced frost optimization algorithm; in order to improve the quality and diversity of the population, the new initial population is generated by the Logistic chaotic map, and its formula is:

[0070]

[0071] R i,j (0) = Lb i,j +G(i)·(Ub i,j +Lb i,j -R i,j )(twenty four)

[0072] Among them, G(i)∈rand(-1,0), N is the number of rime individuals, Ub i,j and Lb i,j is the upper and lower limits of the jth particle of the ith rime individual, R i,j (0) represents the initial position of the jth particle of the i-th rime individual, It is a logical parameter, and its value is 3;

[0073] S4.2: Utilize the soft frost search strategy to quickly explore the entire space in early iterations. Its mathematical model is expressed as follows:

[0074]

[0075] in is the updated particle position, R best,j represents the jth particle of the best rime individual, r1 is a random value in the range of (-1,1), and together with cosψ controls the movement direction of the rime particle, β represents the environmental factor, h is the adhesion force between two rime particles, which is a random number in the range of (0,1), and r2 is a random number between 0 and 1;

[0076]

[0077] Where t represents the current number of iterations, and T represents the maximum number of iterations;

[0078]

[0079] Where ω is the number of segments used to control the step function β, the default value is 5, and [·] indicates rounding; In formula (25), E represents the adhesion coefficient that affects the probability of particle condensation, which is given by the following formula:

[0080]

[0081] In order to fully cover the search space, the shrinkage and spiral update strategies of the whale optimization algorithm are introduced in the soft frost search mechanism, which helps the algorithm to explore the solution space extensively; the shrinkage and spiral update are selected with a probability of 50% to update the position of the frost particles, which is given by the following formula:

[0082]

[0083] where r3 is a random vector in [0,1], a decreases linearly from 2 to 0 during the iteration, and R i,j represents the position of the ith rime particle before the update, b is a constant used to define the logarithmic spiral shape, and l is a random number in [-1,1]; if |B| ≥ 1, then R * is a random position vector, if |B|<1, R* represents the current best candidate, || represents the absolute value;

[0084] S4.3: Inspired by the puncture phenomenon, the hard frost puncture mechanism is used to update the position to improve the convergence of the algorithm and the ability to jump out of the local optimal solution. The particle position update formula is as follows:

[0085]

[0086] Where r4 represents a random number between 0 and 1, F normr (S i ) represents the normalized value of the current rime particle fitness value;

[0087] S4.4: In order to improve the efficiency of global exploration, a positive greedy selection mechanism is used to participate in population update. The fitness value of the updated particle is compared with that before the update. If the updated one is better than the one before the update, the particle solution is replaced. This method encourages outstanding particles to appear more frequently, thereby improving the quality of the solution. The mathematical model is as follows:

[0088]

[0089] S4.5: In order to expand the search range of the optimal solution, the Levy flight strategy is introduced to add disturbance to the optimal solution. The position iteration formula of the Levy flight strategy is given by the following formula:

[0090]

[0091] where ν is the step size control parameter, is the dot product, Levy(γ) is the random search path, which is distributed as follows:

[0092]

[0093] Where γ=1.5, ζ and τ satisfy the normal distribution, σ u and σ v The definition is as follows:

[0094]

[0095] Based on the controller parameters obtained by the enhanced frost and ice optimization algorithm in step S4, a data-driven modeling and control strategy for the cogeneration unit is established. In step S5, the feasibility of the proposed modeling and control strategy is verified and analyzed based on the simulation platform. The specific process is as follows:

[0096] S5.1: Select 10,000 sets of 40%-100% rated load operation data of a 350MW cogeneration unit, and obtain the model of the unit under wide load operation conditions through the improved transformer neural network;

[0097] S5.2: Based on the model obtained in step S5.1, a model-free adaptive predictive control strategy is designed, and the optimal parameters of the controller are obtained through the enhanced frost and ice optimization algorithm proposed in step S4;

[0098] S5.3: Taking the actual automatic power generation control command as the target, based on the optimal controller parameters obtained in step S5.2, a command tracking experiment with a large range of variable loads is carried out;

[0099] S5.4: Determine the external disturbance u d1 Set the value and send it to the controlled object to simulate the disturbance caused by coal quality and variable load during the actual operation of the unit and observe the changes in the controlled quantity of the unit;

[0100] S5.5: Compare the control strategy proposed in the present invention with PID and MFAC, and statistically analyze the control index adjustment time, overshoot, integral absolute error (IAE) and integral square error (ISE).

[0101]

[0102] Beneficial effects of the present invention:

[0103] The present invention combines the development trend of gradually improving the level of renewable energy power consumption in my country's power production field, and from the perspective of modeling and control, improves the rapidity and accuracy of the unit's tracking of automatic power generation control instructions, providing a feasible solution for achieving deep and rapid peak regulation of the unit.

[0104] Aiming at the serious multivariable coupling and multivariable dynamic characteristics problems existing in cogeneration units operating in a wide load range, this paper proposes an improved transformer structure combining BiGRU and MRFC modules. The new neural network framework aims to comprehensively analyze the global and local characteristics of the DC boiler-turbine coupling system, and the obtained high-precision model provides a reference for the design of the controller.

[0105] In order to solve the problems of slow response speed and low precision when the unit is running under fast load change conditions, the present invention designs an MFAPC strategy optimized by ERIME based on the established high-precision black box model, and applies it to the cogeneration unit to improve the flexible operation capability of the system. BRIEF DESCRIPTION OF THE DRAWINGS

[0106] Figure 1The figure is a schematic diagram of the cogeneration unit involved in the present invention.

[0107] Figure 2 The simplified structural diagram of the cogeneration unit involved in the present invention.

[0108] Figure 3 The present invention provides a flow chart of the data-driven modeling and control scheme of the cogeneration unit involved in the present invention. DETAILED DESCRIPTION

[0109] The specific implementation modes of the present invention will be further described below with reference to the accompanying drawings.

[0110] Please refer to the attached figure in the instruction manual. Figure 1 , Figure 1 FIG. 1 is a schematic diagram of the cogeneration unit involved in the present invention. Figure 1 As shown in the figure, the combined heat and power unit is mainly composed of a direct-current boiler, a steam turbine and a heating system. First, the pulverized coal and air are fully mixed and burned in the furnace. The feed water exchanges heat with the heat generated by combustion through the water-cooled wall to form a steam-water mixture. Then, the steam separated by the steam-water separator is heated by the first and second reheaters to form high-temperature and high-pressure main steam. The main steam enters the high-pressure cylinder, drives the turbine blades to rotate, and then drives the generator to generate electricity. The steam discharged from the high-pressure cylinder is heated by the reheater, and then drives the medium-pressure cylinder and the low-pressure cylinder to do work. Part of the steam discharge of the medium-pressure cylinder provides a stable heat source for users. Finally, the exhaust steam of the low-pressure cylinder and the heating system passes through the condenser to form condensed water, and then passes through the low-pressure and high-pressure heaters and economizers to form boiler feed water.

[0111] Please refer to the attached figure in the instruction manual. Figure 2 , Figure 2 This is a simplified structural diagram of the combined heat and power unit involved in the present invention. Considering the balance between fidelity and simplicity, the modeling object is simplified to a four-input and four-output system. The four input variables of the modeling object are: main steam valve opening μ T , Heating valve opening μ H , coal supply B and water supply W, the four output variables are: active power N, main steam pressure P T , heating pressure P H and the midpoint temperature T.

[0112] Please refer to the attached figure in the instruction manual. Figure 3 , Figure 3The flow chart of the data-driven modeling and control scheme of the cogeneration unit involved in the present invention. First, the designed MFAPC strategy is used to obtain the optimal control law and act on the improved transformer neural network structure integrating BiGRU and MRFC modules. Then, the developed ERIME algorithm receives the output of the neural network, obtains the optimal parameters of MFAPC by minimizing the objective function, and further improves the tracking performance of the unit for automatic power generation control instructions. This example is based on the actual operating data of a 350MW cogeneration unit in North China, and the method steps include:

[0113] S1: Analyze the control difficulties of the boiler-turbine unit of a supercritical cogeneration unit;

[0114] S2: Establish an improved transformer neural network framework for wide load modeling of cogeneration units;

[0115] S3: Describe the principles of model-free adaptive predictive control technology and its design steps;

[0116] S4: Obtain optimal controller parameters based on enhanced frost and ice optimization algorithm;

[0117] S5: Rely on the simulation platform to verify the feasibility and performance of the proposed modeling and control methods.

[0118] S1: The flexibility of the power system requires improving the peak-shaving capability and rapid load response capability of the cogeneration unit. However, under large-scale variable load operation conditions, the difference between the slow dynamic characteristics of the direct current boiler and the fast dynamic characteristics of the steam turbine increases, and the significant multivariable coupling characteristics are not conducive to the design of flexible control strategies. Therefore, it is urgent to develop a data-driven modeling strategy to reasonably describe the dynamic characteristics of the cogeneration unit operating in a wide load range, so as to provide guarantees for the application of flexible control strategies on site. In addition, during the load regulation process, smaller load tracking errors and shorter stabilization times are required. At the same time, the cogeneration unit must maintain good anti-interference performance, stable main steam pressure, intermediate point temperature and heating pressure, and the main steam valve opening, heating valve opening, coal supply and water supply must change smoothly to ensure the economy and safety of the unit operation.

[0119] Based on the analysis of S1, it can be seen that the dynamic characteristics of the cogeneration unit under wide load operation conditions are changeable and multi-variable coupling is serious. The traditional modeling method is difficult to reasonably analyze the nonlinear mapping relationship between input variables and output variables. Therefore, it is necessary to design a data-driven modeling method to obtain a high-precision model to lay the foundation for the design of the controller. Step S2 can be concretized as follows:

[0120] S2.1: In order to improve the network's ability to acquire the time dependency of the direct current boiler-turbine coupling system of the cogeneration unit, the bidirectional gated recurrent unit (BiGRU) is applied to the position encoding of the transformer; the basic component unit of BiGRU is GRU, which consists of a reset gate and an update gate; by integrating the forget gate and input gate in the long short-term memory network into the update gate, invalid information can be effectively discarded, long-range dependencies can be captured, the number of parameters can be greatly reduced, and overfitting can be avoided. The description of GRU is as follows:

[0121] z t =σ(W z q' t +U z h t-1 +b z ) (37)

[0122] r t =σ(W r q' t +U r h t-1 +b r ) (38)

[0123] c t =tanh(W c q' t +r t (U c ⊙h t-1 )+b c ) (39)

[0124] h t =c t ⊙(1-z t )+h t-1 ⊙z t (40)

[0125] where q' t is the input at time t, h t-1 is the unit output at time t-1, W z , W r and W c Respectively represent the update gate, reset gate and candidate variable at input q' t The weight matrix under z 、b z , U r 、b r , U c 、b c Represent the weight matrix and bias of the update gate, reset gate and candidate variable, respectively, z t 、r t and c tThey are the outputs of the update gate, reset gate, and candidate variables respectively;

[0126] Information at future time points is conducive to capturing more accurate long-term dependencies. Therefore, a bidirectional GRU is used to fully extract relevant element information of the forward and backward sequences to further improve the modeling performance of the transformer network, as described below:

[0127]

[0128] S2.2: The transformer neural network based on the self-attention mechanism is good at capturing global features, while the convolutional neural network (CNN) can effectively utilize local features. Therefore, the complementary advantages of transformer and CNN are expected to solve the problem of dynamic characteristics analysis of cogeneration units with strong coupling of multiple variables. The multi-head self-attention mechanism is directly applied to the input Z, which is described as follows:

[0129]

[0130]

[0131] Where MSA(Q,K,V) represents a multi-head self-attention layer, Q, K, and V represent query, key, and value, respectively, MLP and LayerNorm represent a multi-layer feedforward network and layer normalization, respectively. and Z time Represents the output of MSA and MLP;

[0132] In order to further enhance the transformer network's ability to extract local features, a multi-receptive field convolution (MRFC) module is designed and integrated into the transformer's multi-head self-attention layer, as described below:

[0133]

[0134]

[0135] Among them, MRFC represents the multi-receptive field convolution layer, Represents the output of the convolution mechanism;

[0136] The MRFC module consists of token-wise convolution, point-wise convolution, Swish activation function and normalization. The token-wise convolution consists of convolutions of different receptive fields in parallel. Convolution kernels of different sizes are used to process the input in parallel, obtain features from multiple receptive fields, and use zero padding to ensure that the size of the feature map remains unchanged. The feature maps of multiple receptive fields are connected along the channel dimension to obtain the output of the point-wise convolution layer. Next, the point-wise convolution layer uses a convolution of size 1 to reduce the number of channels in the feature map of the token-wise convolution layer. The Swish activation function and normalization are conducive to further enhancing the convergence and generalization ability of the network.

[0137] The decoder receives long-term dependency and local feature information from the input variables provided by the encoder as well as the historical trend relationship of the output variables; this information is then integrated and fed into the multi-head self-attention layer to capture the nonlinear mapping relationship between the input variables and the output variables.

[0138] After determining the improved transformer neural network framework for wide load modeling of cogeneration units, the principle of model-free adaptive predictive control technology and its design steps are concretized in step S3:

[0139] S3.1: Represent the 350MW CHP unit as a multi-input multi-output nonlinear discrete-time system as follows:

[0140] Y(k+1)=f(Y(k),…Y(kn y ),U(k),…,U(kn u )) (46)

[0141] where U(k)∈R m×1 , Y(k)∈R l×1 They represent the m-dimensional input and l-dimensional output at time k, f is an unknown function, n y and n u is an unknown positive integer, representing the order of the system;

[0142] The combined heat and power unit once-through boiler-turbine coupled system described in equation (46) can be described by compact dynamic linearization:

[0143] ΔY(k+1)=Φ c (k)ΔU(k)(47)

[0144] where Φ c (k)∈R l×m , represents the pseudo-Jacobi matrix;

[0145] S3.2: The design goal of MFAPC is to find the best future control action to make the process output as close as possible to the reference trajectory of the CHP unit. The control performance consideration is to minimize the following function:

[0146]

[0147] Among them, N p and N u are the prediction time domain and the control time domain respectively, λ>0 is the weight coefficient, Y * (k+i) is the expected output of the system at time k+i;

[0148] According to the equivalent dynamic linear model (11), Y(k+i) in (12) can be expressed as:

[0149]

[0150]

[0151] Then, formula (49) can be reformulated as:

[0152]

[0153] in represents the Kronecker product;

[0154] definition Substituting formula (51) into formula (48), we can get ΔU Nu Taking the derivative and setting it equal to zero, we can get the following control algorithm:

[0155]

[0156] Therefore, the actual control input value at the current moment is:

[0157]

[0158] in

[0159] S3.3: In formula (50), Φ c (k+i),i=0,1,…,N u -1 is an unknown number and must be replaced by an estimated and predicted value. Consider the following estimation standard function:

[0160]

[0161] where μ>0 is a weight coefficient used to penalize excessive changes in the PJM estimate;

[0162] For formula (18) with respect to Φ c(k) By finding the extreme value, we can get the PJM estimation algorithm, that is,

[0163]

[0164] Where η∈(0,2] is the step size factor;

[0165] S3.4: Using Φ c The estimated value of (k) is used to establish an automatic regression model for the valuation sequence, and the following prediction formula is obtained:

[0166]

[0167] Among them, θ i ,i=1,…,n p is the coefficient, n p is an appropriate order, usually between 2 and 7;

[0168] According to formula (56), the prediction algorithm can be obtained as follows:

[0169]

[0170] where j = 1, 2, ..., N u –1;

[0171] definition It is described as follows:

[0172]

[0173] in

[0174] Based on the principle and design steps of the model-free adaptive predictive control technology described in step S3, it is proposed in step S4 to obtain the optimal controller parameters based on the enhanced frost and ice optimization algorithm. The specific process is as follows:

[0175] S4.1: After determining the structure of the model-free adaptive predictive controller, the optimal parameters λ, μ, η and δ of the controller are obtained by the enhanced frost optimization algorithm; in order to improve the quality and diversity of the population, the new initial population is generated by the Logistic chaotic map, and its formula is:

[0176]

[0177] R i,j (0) = Lb i,j +G(i)·(Ub i,j +Lb i,j -R i,j )(60)

[0178] Among them, G(i)∈rand(-1,0), N is the number of rime individuals, Ub i,j and Lb i,j is the upper and lower limits of the jth particle of the ith rime individual, R i,j (0) represents the initial position of the jth particle of the i-th rime individual, It is a logical parameter, and its value is 3;

[0179] S4.2: Utilize the soft frost search strategy to quickly explore the entire space in early iterations. Its mathematical model is expressed as follows:

[0180]

[0181] in is the updated particle position, R best,j represents the jth particle of the best rime individual, r1 is a random value in the range of (-1,1), and together with cosψ controls the movement direction of the rime particle, β represents the environmental factor, h is the adhesion force between two rime particles, which is a random number in the range of (0,1), and r2 is a random number between 0 and 1;

[0182]

[0183] Where t represents the current number of iterations, and T represents the maximum number of iterations;

[0184]

[0185] Where ω is the number of segments used to control the step function β, the default value is 5, and [·] indicates rounding; In formula (25), E represents the adhesion coefficient that affects the probability of particle condensation, which is given by the following formula:

[0186]

[0187] In order to fully cover the search space, the shrinkage and spiral update strategies of the whale optimization algorithm are introduced in the soft frost search mechanism, which helps the algorithm to explore the solution space extensively; the shrinkage and spiral update are selected with a probability of 50% to update the position of the frost particles, which is given by the following formula:

[0188]

[0189] where r3 is a random vector in [0,1], a decreases linearly from 2 to 0 during the iteration, and R i,j represents the position of the ith rime particle before the update, b is a constant used to define the logarithmic spiral shape, and l is a random number in [-1,1]; if |B| ≥ 1, then R * is a random position vector. If |B|<1, R* represents the current best candidate, || represents the absolute value;

[0190] S4.3: Inspired by the puncture phenomenon, the hard frost puncture mechanism is used to update the position to improve the convergence of the algorithm and the ability to jump out of the local optimal solution. The particle position update formula is as follows:

[0191]

[0192] Where r4 represents a random number between 0 and 1, F normr (S i ) represents the normalized value of the current rime particle fitness value;

[0193] S4.4: In order to improve the efficiency of global exploration, a positive greedy selection mechanism is used to participate in population update. The fitness value of the updated particle is compared with that before the update. If the updated one is better than the one before the update, the particle solution is replaced. This method encourages outstanding particles to appear more frequently, thereby improving the quality of the solution. The mathematical model is as follows:

[0194]

[0195] S4.5: In order to expand the search range of the optimal solution, the Levy flight strategy is introduced to add disturbance to the optimal solution. The position iteration formula of the Levy flight strategy is given by the following formula:

[0196]

[0197] where ν is the step size control parameter, is the dot product, Levy(γ) is the random search path, which is distributed as follows:

[0198]

[0199] Where γ=1.5, ζ and τ satisfy the normal distribution, σ u and σ v The definition is as follows:

[0200]

[0201] Based on the controller parameters obtained by the enhanced frost and ice optimization algorithm in step S4, a data-driven modeling and control strategy for the cogeneration unit is established. In step S5, the feasibility of the proposed modeling and control strategy is verified and analyzed based on the simulation platform. The specific process is as follows:

[0202] S5.1: Select 10,000 sets of 40%-100% rated load operation data of a 350MW cogeneration unit, and obtain the model of the unit under wide load operation conditions through the improved transformer neural network;

[0203] S5.2: Based on the model obtained in step S5.1, a model-free adaptive predictive control strategy is designed, and the optimal parameters of the controller are obtained through the enhanced frost and ice optimization algorithm proposed in step S4;

[0204] S5.3: Taking the actual automatic power generation control command as the target, based on the optimal controller parameters obtained in step S5.2, a command tracking experiment with a large range of variable loads is carried out;

[0205] S5.4: Determine the external disturbance u d1 Set the value and send it to the controlled object to simulate the disturbance caused by coal quality and variable load during the actual operation of the unit and observe the changes in the controlled quantity of the unit;

[0206] S5.5: Compare the control strategy proposed in the present invention with PID and MFAC, and statistically analyze the control index adjustment time, overshoot, integral absolute error (IAE) and integral square error (ISE).

[0207]

[0208]

Claims

1. A data-driven modeling and control method for the flexible operation of a cogeneration unit is characterized by: The following steps are involved: S1: Analyze the control difficulties of the boiler-turbine unit of a supercritical cogeneration unit; S2: Establish an improved transformer neural network framework for wide load modeling of cogeneration units; S3: Describe the principles of model-free adaptive predictive control technology and its design steps; S4: Obtain optimal controller parameters based on enhanced frost and ice optimization algorithm; S5: Rely on the simulation platform to verify the feasibility and performance of the proposed modeling and control methods.

2. A data-driven modeling and control method for flexible operation of a cogeneration unit according to claim 1, characterized in that: The control difficulties of analyzing the boiler-turbine unit of the supercritical cogeneration unit described in step S1 are as follows: S1: The flexibility of the power system requires improving the peak-shaving capability and rapid load response capability of the cogeneration unit. However, under large-scale variable load operation conditions, the difference between the slow dynamic characteristics of the once-through boiler and the fast dynamic characteristics of the steam turbine increases, and the significant multivariable coupling characteristics are not conducive to the design of flexible control strategies; Therefore, it is urgent to develop a data-driven modeling strategy to reasonably describe the dynamic characteristics of CHP units operating in a wide load range and provide a guarantee for the application of flexible control strategies in the field; In addition, during load regulation, smaller load tracking error and shorter stabilization time are required; At the same time, the cogeneration unit must maintain good anti-interference performance, stable main steam pressure, intermediate point temperature and heating pressure. The main steam valve opening, heating valve opening, coal supply and water supply must change smoothly to ensure the economy and safety of the unit operation.

3. A data-driven modeling and control method for flexible operation of a cogeneration unit according to claim 1, characterized in that: Establish an improved transformer neural network framework for wide load modeling of cogeneration units, including: S2.1: In order to improve the network's ability to acquire the time dependency of the direct current boiler-turbine coupling system of the cogeneration unit, the bidirectional gated recurrent unit (BiGRU) is applied to the position encoding of the transformer; the basic component unit of BiGRU is GRU, which consists of a reset gate and an update gate; by integrating the forget gate and input gate in the long short-term memory network into the update gate, invalid information can be effectively discarded, long-range dependencies can be captured, the number of parameters can be greatly reduced, and overfitting can be avoided. The description of GRU is as follows: z t =σ(W z q′ t +U z h t-1 +b z ) (1) r t =σ(W r q′ t +U r h t-1 +b r ) (2) c t =tanh(W c q′ t +r t (U c ⊙h t-1 )+b c ) (3) h t =c t ⊙(1-z t )+h t-1 ⊙z t (4) where q′ t is the input at time t, h t-1 is the unit output at time t-1, W z , W r and W c Respectively represent the update gate, reset gate and candidate variable at input q′ t The weight matrix under z 、b z , U r 、b r , U c 、b c Represent the weight matrix and bias of the update gate, reset gate and candidate variable, respectively, z t 、r t and c t They are the outputs of the update gate, reset gate, and candidate variables respectively; Information at future time points is conducive to capturing more accurate long-term dependencies. Therefore, a bidirectional GRU is used to fully extract relevant element information of the forward and backward sequences to further improve the modeling performance of the transformer network, as described below: S2.2: The transformer neural network based on the self-attention mechanism is good at capturing global features, while the convolutional neural network (CNN) can effectively utilize local features. Therefore, the complementary advantages of transformer and CNN are expected to solve the problem of dynamic characteristics analysis of cogeneration units with strong coupling of multiple variables. The multi-head self-attention mechanism is directly applied to the input Z, which is described as follows: Where MSA(Q,K,V) represents a multi-head self-attention layer, Q, K, and V represent query, key, and value, respectively, MLP and LayerNorm represent a multi-layer feedforward network and layer normalization, respectively. and Z time Represents the output of MSA and MLP; In order to further enhance the transformer network's ability to extract local features, a multi-receptive field convolution (MRFC) module is designed and integrated into the transformer's multi-head self-attention layer, as described below: Among them, MRFC represents the multi-receptive field convolution layer, Represents the output of the convolution mechanism; The MRFC module consists of token-wise convolution, point-wise convolution, Swish activation function and normalization. The token-wise convolution consists of convolutions of different receptive fields in parallel. Convolution kernels of different sizes are used to process the input in parallel, obtain features from multiple receptive fields, and use zero padding to ensure that the size of the feature map remains unchanged. The feature maps of multiple receptive fields are connected along the channel dimension to obtain the output of the point-wise convolution layer. Next, the point-wise convolution layer uses a convolution of size 1 to reduce the number of channels in the feature map of the token-wise convolution layer. The Swish activation function and normalization are conducive to further enhancing the convergence and generalization ability of the network. The decoder receives long-term dependency and local feature information from the input variables provided by the encoder as well as the historical trend relationship of the output variables; this information is then integrated and fed into the multi-head self-attention layer to capture the nonlinear mapping relationship between the input variables and the output variables.

4. The data-driven modeling and control method for flexible operation of a cogeneration unit according to claim 1, characterized in that: Describe the principles of model-free adaptive predictive control technology and its design steps: S3.1: Represent the 350MW CHP unit as a multi-input multi-output nonlinear discrete-time system as follows: Y(k+1)=f(Y(k),…Y(k-n y ),U(k),…,U(k-n u )) (10) where U(k)∈R m×1 , Y(k)∈R l×1 They represent the m-dimensional input and l-dimensional output at time k, f is an unknown function, and n y and n u is an unknown positive integer, representing the order of the system; The once-through boiler-turbine coupled system of the cogeneration unit described in equation (10) can be described as follows after compact dynamic linearization: ΔY(k+1)=Φ c (k)ΔU(k) ​​(11) where Φ c (k)∈R l×m , represents the pseudo-Jacobi matrix; S3.2: The design goal of MFAPC is to find the best future control action to make the process output as close as possible to the reference trajectory of the cogeneration unit. The control performance consideration is to minimize the following function: Among them, N p and N u are the prediction time domain and the control time domain respectively, λ>0 is the weight coefficient, Y * (k+i) is the expected output of the system at time k+i; According to the equivalent dynamic linear model (11), Y(k+i) in (12) can be expressed as: definition Then, formula (13) can be reformulated as: in represents the Kronecker product; definition Substituting formula (15) into formula (12), we can get ΔU Nu Taking the derivative and setting it equal to zero, we can get the following control algorithm: Therefore, the actual control input value at the current moment is: U(k)=U(k-1)+g T ΔU Nu (k) (17) in S3.3: In formula (14), Φ c (k+i),i=0,1,…,N u -1 is an unknown number and must be replaced by an estimated and predicted value. Consider the following estimation standard function: where μ>0 is a weight coefficient used to penalize excessive changes in the PJM estimate; For formula (18) with respect to Φ c (k) By finding the extreme value, we can get the PJM estimation algorithm, that is, Where η∈(0,2] is the step size factor; S3.4: Using Φ c The estimated value of (k) is used to establish an automatic regression model for the valuation sequence, and the following prediction formula is obtained: Among them, θ i ,i=1,…,n p is the coefficient, n p is an appropriate order, usually between 2 and 7; According to formula (20), the prediction algorithm can be obtained as follows: where j = 1, 2, ..., N u –1; definition It is described as follows: in δ∈(0,1].

5. The data-driven modeling and control method for flexible operation of a cogeneration unit according to claim 1, characterized in that: The optimal controller parameters are obtained based on the enhanced frost and ice optimization algorithm, and the process includes: S4.1: After determining the structure of the model-free adaptive predictive controller, the optimal parameters λ, μ, η and δ of the controller are obtained by the enhanced frost optimization algorithm; in order to improve the quality and diversity of the population, the new initial population is generated by the Logistic chaotic map, and its formula is: R i,j (0)=Lb i,j +G(i)·(Ub i,j +Lb i,j -R i,j ) (24) Among them, G(i)∈rand(-1,0), N is the number of rime individuals, Ub i,j and Lb i,j is the upper and lower limits of the jth particle of the ith rime individual, R i,j (0) represents the initial position of the jth particle of the i-th rime individual, It is a logical parameter, and its value is 3; S4.2: Utilize the soft frost search strategy to quickly explore the entire space in early iterations. Its mathematical model is expressed as follows: in is the updated particle position, R best,j represents the jth particle of the best rime individual, r1 is a random value in the range of (-1,1), and together with cosψ controls the movement direction of the rime particle, β represents the environmental factor, h is the adhesion force between two rime particles, which is a random number in the range of (0,1), and r2 is a random number between 0 and 1; Where t represents the current number of iterations, and T represents the maximum number of iterations; Where ω is the number of segments used to control the step function β, the default value is 5, and [·] indicates rounding; In formula (25), E represents the adhesion coefficient that affects the probability of particle condensation, which is given by the following formula: In order to fully cover the search space, the shrinkage and spiral update strategies of the whale optimization algorithm are introduced in the soft frost search mechanism, which helps the algorithm to explore the solution space extensively; the shrinkage and spiral update are selected with a probability of 50% to update the position of the frost particles, which is given by the following formula: where r3 is a random vector in [0,1], a decreases linearly from 2 to 0 during the iteration, and R i,j represents the position of the ith rime particle before the update, b is a constant used to define the logarithmic spiral shape, and l is a random number in [-1,1]; if |B| ≥ 1, then R * is a random position vector. If |B|<1, R * represents the current best candidate, || represents the absolute value; S4.3: Inspired by the puncture phenomenon, the hard frost puncture mechanism is used to update the position to improve the convergence of the algorithm and the ability to jump out of the local optimal solution. The particle position update formula is as follows: Where r4 represents a random number between 0 and 1, F normr (S i ) represents the normalized value of the current rime particle fitness value; S4.4: In order to improve the efficiency of global exploration, a positive greedy selection mechanism is used to participate in population update. The fitness value of the updated particle is compared with that before the update. If the updated one is better than the one before the update, the particle solution is replaced. This method encourages outstanding particles to appear more frequently, thereby improving the quality of the solution. The mathematical model is as follows: S4.5: In order to expand the search range of the optimal solution, the Levy flight strategy is introduced to add disturbance to the optimal solution. The position iteration formula of the Levy flight strategy is given by the following formula: where ν is the step size control parameter, is the dot product, Levy(γ) is the random search path, which is distributed as follows: Where γ=1.5, ζ and τ satisfy the normal distribution, σ u and σ v The definitions are as follows: s v =1.

6. A data-driven modeling and control method for flexible operation of a cogeneration unit according to claim 1, characterized in that: In step S5, feasibility verification and performance analysis of the proposed modeling and control method are performed on a simulation platform, specifically including: S5.1: Select 10,000 sets of 40%-100% rated load operation data of a 350MW cogeneration unit, and obtain the model of the unit under wide load operation conditions through the improved transformer neural network; S5.2: Based on the model obtained in step S5.1, a model-free adaptive predictive control strategy is designed, and the optimal parameters of the controller are obtained through the enhanced frost and ice optimization algorithm proposed in step S4; S5.3: Taking the actual automatic power generation control command as the target, based on the optimal controller parameters obtained in step S5.2, a command tracking experiment with a large range of variable loads is carried out; S5.4: Determine the external disturbance u d1 Set the value and send it to the controlled object to simulate the disturbance caused by coal quality and variable load during the actual operation of the unit and observe the changes in the controlled quantity of the unit; S5.5: Compare the control strategy proposed in the present invention with PID and MFAC, and statistically analyze the control index adjustment time, overshoot, integral absolute error (IAE) and integral square error (ISE).

Images