Intelligent modeling method for fire storage collaborative peak regulation process based on Mmba-2 large model

Abstract

The invention discloses an intelligent modeling method for a fire storage collaborative peak regulation process based on a Mmba-2 large model. The invention discloses the intelligent modeling method for the fire storage collaborative peak regulation process based on the Mmba-2 large model. Firstly, a fused salt heat storage and supercritical coal-fired unit collaborative peak regulation process coupling model is simplified into a five-input five-output coupling system; secondly, combining a sparse attention mechanism and a novel Lion optimizer to obtain an improved Mamba-2, and improving the modeling performance of the model on a multivariable long-time sequence; and finally, verifying the effectiveness of the modeling method for the coupled fused salt heat storage and supercritical coal-fired unit collaborative peak shaving process based on the Mmba-2 large model by relying on a simulation platform. The coupling system collaborative peak regulation process model structure accurately represents the hydrodynamic characteristic and the heat storage-release characteristic of the system, and provides reliable guarantee for construction of a high-proportion new energy power system.

Description

Technical Field

[0001] The present invention relates to the technical field of power generation process modeling of a thermal-storage coordinated deep peak-shaving system, and more specifically, to an intelligent modeling method for a coupled molten salt thermal storage and supercritical coal-fired unit coordinated deep peak-shaving process based on a Mamba-2 large model. Background Art

[0002] The high penetration of random wind power and photovoltaic power generation has increasingly posed a challenge to the stable and safe operation of the power system. As the main power source, thermal power generating units need to improve their flexible operation capabilities to suppress the load and frequency fluctuations caused by wind and photovoltaic power generation systems. In particular, higher requirements are placed on the peak-shaving depth and peak-shaving speed of large coal-fired units. The mature molten salt heat storage technology has been successfully applied to the energy supply field with its advantages of fast response speed and good economic benefits. Therefore, the thermal storage coordinated peak-shaving system integrating large coal-fired units with molten salt heat storage has become an important choice in the construction of new power systems.

[0003] The fire-storage coordinated peak-shaving system is a new type of power generation technology that effectively integrates large coal-fired units and molten salt heat storage. It can further reduce the minimum output power of large coal-fired units to meet the peak-shaving needs of the power grid. When the power grid issues a peak-shaving instruction, the large coal-fired unit quickly adjusts the output power to the set value, and the remaining electric energy is supplied to the molten salt heat storage system. The molten salt heater converts electric energy into thermal energy for heating deoxygenated water. Subsequently, the deoxygenated water flows through the preheater, evaporator and superheater in the molten salt heat storage system, and is finally converted into high-temperature steam to become industrial steam supply. The efficient coordination of large coal-fired units and molten salt heat storage systems can improve the peak-shaving depth, speed and economic benefits of the units. However, during the peak-shaving process, due to the unstable combustion of the direct-flow boiler in the large coal-fired unit, the intermediate point temperature and main steam pressure often change, and the contradiction between the heating rate and the temperature overshoot must be handled. Secondly, the numerical span of the variables such as temperature and pressure of the working fluid in the fire-storage coordinated peak-shaving system is large, which puts high demands on the generalization ability of the model and the robustness of the controller. In addition, the molten salt heat storage system needs to go through three different stages of heating, energy storage and energy release during the peak-shaving process. The working fluid undergoes phase change in the evaporator, and its dynamic characteristics are complex and changeable. Not only is there a strong nonlinear coupling relationship between the output variables, but there is also a nonlinear mapping relationship between the input and output variables. Therefore, this paper takes the fire-storage collaborative peak-shaving system as the research object, and designs an intelligent modeling method for it in combination with the improved Mamba-2 large model to serve the design of the peak-shaving process controller.

[0004] The Mamba-2 large model is considered to be one of the most effective algorithms for processing multivariate long time series. Its stable selective state space model structure provides effective support for its rapid capture of the nonlinear mapping relationship between input variables and output variables, and has therefore been successfully applied in multi-series prediction, image classification and recognition. The core of Mamba-2 is to obtain the mapping relationship between input variables and output variables through the state transfer equation, thereby achieving accurate prediction of the next step. Therefore, the intelligent modeling method for the fire-storage coordinated peak-shaving process is designed in combination with the improved Mamba-2 with fast and accurate prediction capabilities, laying the foundation for designing joint control strategies to improve the system peak-shaving capability. Summary of the invention

[0005] The present invention aims to provide an intelligent modeling method for the fire-storage coordinated peak-shaving process based on the Mamba-2 large model, and deeply explore the deep peak-shaving capability of large coal-fired units to promote the construction of new power systems. The method fully considers the dynamic characteristics of large coal-fired units and molten salt heat storage systems, and uses the improved Mamba-2 large model that is good at processing multivariate long-term series prediction to design an accurate intelligent modeling method for them. The improved Mamba-2 fully considers the differences and mapping relationships of the numerical values ​​of each variable in the fire-storage coordinated peak-shaving process, combines normalization and sparse attention mechanisms to enhance the learning ability of the context, and can obtain a mathematical model of the wide-load peak-shaving process of the fire-storage coordinated peak-shaving system, which is conducive to improving the deep peak-shaving capability of the system.

[0006] The present invention proposes an intelligent modeling method for the fire-storage coordinated peak-shaving process based on the Mamba-2 large model, which consists of the following three steps:

[0007] S1: Analyze the difficulties in modeling the deep peak-shaving process of supercritical coal-fired units coupled with molten salt heat storage;

[0008] S2: Establish an intelligent modeling method for the coordinated peak-shaving process of coupled molten salt heat storage and supercritical coal-fired units based on the Mamba-2 large model;

[0009] S3: Rely on the simulation platform to verify and analyze the effectiveness of the intelligent modeling method designed based on the Mamba-2 large model.

[0010] S1: The purpose of the fire-storage coordinated peak-shaving system is to reduce the minimum output power of thermal power generating units and provide space for the grid connection of large-scale wind power generation and photovoltaic power generation systems. In order to achieve safer, more stable, faster and full-condition peak-shaving of the fire-storage coordinated peak-shaving system, the mathematical model of the deep peak-shaving process of the supercritical coal-fired unit coupled with molten salt heat storage can be established in combination with the outstanding Mamba-2 large model to provide model guarantee for the design of control strategies. However, the supercritical coal-fired unit coupled with molten salt heat storage is a large and complex thermal system with many variables that need to be monitored, and the value of each variable will change greatly during the peak-shaving process. Therefore, in order to facilitate the design of the control strategy of the fire-storage coordinated peak-shaving system, it is urgent to establish an accurate state space model of the large-scale peak-shaving process. The supercritical coal-fired unit coupled with molten salt heat storage is mainly composed of a direct-current boiler, a steam turbine and a double-tank molten salt heat storage system. According to the dynamic characteristics of the steam-water process, it can be simplified into a coupled system with five input and five output variables. The five input variables are the coal feed rate u B (t / h), water supply u W (t / h), main steam valve opening μ T (%), low temperature molten salt flow rate D MS (℃) and deoxygenated water flow u DW (t / h). The five output variables are output power N E (MW), separator outlet temperature T I (℃), main steam pressure P st (MPa), high temperature tank inlet molten salt temperature T HM (℃) and supplementary industrial steam flow D SW (m 3 / h). During the peak regulation process, the variables of the thermal storage coordinated peak regulation system have strong coupling and nonlinear characteristics and the variables change rapidly and over a large span. It is difficult to quickly establish an accurate peak regulation process model using traditional mathematical modeling methods.

[0011] Based on the analysis of S1, it can be seen that the deep peak-shaving process of supercritical coal-fired units coupled with molten salt heat storage involves large-span changes in the values ​​of multiple variables. Therefore, the improved Mamba-2 large model with a stable selective state space model is used to learn multivariate long time series to obtain the nonlinear mapping relationship between the variables, and an accurate state space model of the peak-shaving system can be established. The five input variables and five output variable data of the deep peak-shaving process of supercritical coal-fired units coupled with molten salt heat storage are imported into the improved Mamba-2 large model for training, and the trained model is tested and output with different multiple groups of data, so as to complete the identification of the state space model of the peak-shaving process of the fire-storage coordinated peak-shaving system. Then step S2 can be concretized as follows:

[0012] S2.1: For a supercritical coal-fired unit coupled with molten salt heat storage, the state space model is as follows:

[0013]

[0014] Input variable matrix u(t) = [u B (t),u w (t),μ T (t),D MS (t),u DW (t)], output variable matrix y(t) = [N E (t),T I (t),P st (t),T HM (t),D SW (t)]. Then, the state space model is discretized, and the supercritical coal-fired unit coupled with molten salt heat storage is rewritten as:

[0015]

[0016] Then, rewrite the model as a differential equation and substitute the derivative definition to obtain the solution of the state space model:

[0017]

[0018] Parameters A, B, C, It changes with the on-site peak-shaving process data of the supercritical coal-fired unit coupled with molten salt heat storage. Δ is the time scale parameter. To speed up the calculation, the above parameters are calculated by matrix multiplication in the form of convolution:

[0019]

[0020] Finally, calculate Mamba's predicted output sequence:

[0021]

[0022] However, Mamba's ability to handle selective copying and contextual learning tasks still needs to be improved. Therefore, the Mamba-2 model with structured state space duality (SSD) as the core is considered to model the peak-shaving process of supercritical coal-fired units coupled with molten salt heat storage:

[0023]

[0024] In the SSD layer of Mamba-2, the matrix A is further restricted to be a scalar. The diagonal elements of the matrix A have the same value. Since Mamba-2 can use tensors to accelerate the calculation of SSM, it is designed to be modular to facilitate efficient modeling of multivariate long time series of the peak-shaving process of supercritical coal-fired units coupled with molten salt heat storage. Parallel parameter design and additional normalization are considered in the modular implementation of Mamba-2. In the parallel parameter design, the SSD layer acts as a mapping relationship between the variable parameters A, B, x and the predicted output value y:

[0025] (A (T) ,B (T,N) ,C (T,N) ,x (T,P) )→y (T,P) (7)

[0026] The SSD layer allows the accelerator to use matrix multiplication units, which significantly improves the training efficiency of Mamba-2. In addition, in order to obtain an improved Mamba-2 with higher computing performance, a sparse attention mechanism is added after the input data module, and an additional minimum-maximum normalization module is added to the data block before the final model output to improve modeling accuracy.

[0027] In order to improve the contextual learning ability of Mamba-2 for multivariate long time series in the peak-shaving process of supercritical coal-fired units coupled with molten salt heat storage, a sparse attention mechanism was added after the input data module. The sparse attention mechanism is one of the attention mechanisms used to optimize the training process of deep learning models. Its main goal is to only calculate the similarity between the query vector and some key vectors to reduce time and space complexity and save storage space. The nonlinear mapping relationship between variables in the peak-shaving process of supercritical coal-fired units coupled with molten salt heat storage has an important influence on the modeling performance of the unit, and usually requires precise calculation. Another advantage of the sparse attention mechanism is its adjustability. By adjusting the width of the sliding window, the number of computing elements can be controlled to achieve a balance between computing efficiency and modeling performance. Assume that the input matrix of the network is Then the local self-attention operation of the first window G1 on the target grid feature can be expressed as:

[0028]

[0029]

[0030] where h and ω are the height and width of the sliding window, respectively. is the mesh feature in the first window, d k is the scaling factor, D represents the dimension of the grid feature; the query, key and value are respectively in and is the learning parameter. The output of the local self-attention mechanism is

[0031] In order to improve the modeling efficiency of Mamba-2, the grid sparsity λ∈[0,1) is introduced to remove the grid features of the window with probability λ. For the grid features in the sliding window, the features are randomly retained in the proportion of (1-λ) and participate in each batch calculation. Then, the grid features in the first window are removed from Reduced to After sliding (h×ω) times, the target grid features of all windows are connected and the network output is expressed as:

[0032] Z=concatenate(Z1,Z2,…,Z (h×ω) ) (10)

[0033] In addition, to speed up the calculation, the Lion optimizer is added to improve the Mamba-2 model. Compared with the Adam optimizer, the Lion optimizer has fewer parameters, so the cache required for calculation is smaller. In addition, Lion removes the most computationally intensive division and root operations in Adam during the update process, making the calculation faster. Specifically, the parameter update rules in Lion are also different, and the momentum update comes after the variable update. The update process of the Lion optimizer is described as:

[0034]

[0035] where sign (·) is a sign function whose value is equal to 1 or -1. represents the gradient of the loss function, μ t is a sign function used to calculate the sign of the gradient. β1 and β2 are the exponential decay rates of the first-order momentum and the second-order momentum when the parameters are updated, respectively. The learning rate is denoted by η t The gradient momentum accumulation is denoted as m t , using m t Accelerate the gradient descent process, λ t is a constant.

[0036] So far, an intelligent modeling method for the fire-storage coordinated peak-shaving process based on the Mamba-2 large model has been designed.

[0037] Based on the intelligent modeling method of the coordinated peak-shaving process of coupled molten salt heat storage and supercritical coal-fired units based on the Mamba-2 large model designed in step S2.1, the effectiveness of the designed modeling method is verified and analyzed on the simulation platform in step S3. The specific process is as follows:

[0038] S3.1: 20,000 and 5,000 sets of actual historical operation data of the deep peak-shaving process of supercritical coal-fired units coupled with molten salt heat storage are selected respectively to train and verify the designed Mamba-2 model, and finally output the state space model.

[0039] S3.2: When wind and photovoltaic power generation resources are sufficient, the supercritical coal-fired unit coupled with molten salt heat storage quickly reduces its output power to the set value according to the grid peak load instruction, and the values ​​of the five input variables and the five output variables change greatly. In order to obtain the unit state space model of the peak load process, 20,000 groups of units containing coal supply u B (t / h), water supply u W (t / h), main steam valve opening μ T (%), low temperature molten salt flow rate D MS (℃), deoxygenated water flow u DW (t / h), output power N E (MW), separator outlet temperature T I (℃), main steam pressure P st (MPa), high temperature tank inlet molten salt temperature T HM (℃) and supplementary industrial steam flow D SW (m 3 / h) of the actual peak load data of the units were imported into the data module to train the Mamba-2 model. Subsequently, another 5000 sets of actual peak load data of the units were imported to verify the Mamba-2 model.

[0040] S3.3: Observe the training and validation process Five output variables of the Mamba-2 model: Output power N E (MW), separator outlet temperature T I (℃), main steam pressure P st (MPa), high temperature tank inlet molten salt temperature T HM (℃) and supplementary industrial steam flow D SW (m 3 / h).

[0041] S3.4: Calculate the modeling errors of the five output variables of the Mamba-2 model, including the mean absolute error (MAE), mean square error (MSE), mean absolute percentage error (MAPE), symmetric mean absolute percentage error (SMAPE) and modeling accuracy (R), comprehensively evaluate the performance of the peak-shaving process model of the supercritical coal-fired unit coupled with molten salt heat storage, and output the unit state space model. Beneficial effects of the present invention:

[0042] Starting from building a control-oriented process model, the present invention designs a Mamba-2 modeling method for a supercritical coal-fired unit coupled with molten salt heat storage, which is used to quickly obtain an accurate state-space model of the unit's peak-shaving process. This responds to the actual development needs of today's large-scale thermal power generating units and can promote the construction of a high-proportion new energy power system.

[0043] In the design of the Mamba-2 modeling method for a supercritical coal-fired unit coupled with molten salt heat storage, the present invention fully considers the dynamic characteristics of the unit's peak-shaving process and the key control variables of the molten salt heat storage system, and integrates the sparse attention mechanism and the new Lion optimizer to improve the learning ability of the Mamba-2 model for multivariate long time series, so that the designed intelligent modeling method based on the Mamba-2 large model can meet the needs of fast and accurate acquisition of the peak-shaving process model.

[0044] The present invention simplifies the supercritical coal-fired unit coupled with molten salt heat storage into a coupled system with five input variables and five output variables, accurately characterizing the hydrodynamic characteristics of the unit's peak-shaving process. The improved Mamba-2 model is used to accurately capture the nonlinear mapping relationship between the input and output variables, and the state space model of the peak-shaving process of the supercritical coal-fired unit coupled with molten salt heat storage is successfully obtained, which lays a good foundation for the design of control strategies to improve the deep peak-shaving capacity of the service unit. BRIEF DESCRIPTION OF THE DRAWINGS

[0045] Figure 1 It is a simplified structure of the peak-shaving process model of a supercritical coal-fired unit coupled with molten salt heat storage according to the present invention;

[0046] To facilitate model establishment, the supercritical coal-fired unit coupled with molten salt heat storage participating in power grid peak regulation is simplified into a coupled system with five input variables and five output variables, and the variables have strong coupling dynamic characteristics. Among them, the five input variables of the system are the coal feed rate u B (t / h), water supply u W (t / h), main steam valve opening μ T (%), low temperature molten salt flow rate D MS (℃) and deoxygenated water flow u DW (t / h). The five output variables of the system are output power N E (MW), separator outlet temperature T I (℃), main steam pressure P st (MPa), high temperature tank inlet molten salt temperature T HM (℃) and supplementary industrial steam flow D SW (m 3 / h).

[0047] Figure 2The present invention is an intelligent modeling method for the coordinated peak-shaving process of coupled molten salt heat storage and supercritical coal-fired units based on the Mamba-2 large model;

[0048] The intelligent modeling method based on the Mamba-2 large model consists of multiple improved Mamba-2 modules with the same structure. Each improved Mamba-2 contains a data input module, a mapping module, a convolution module, an activation module, a selective state space model, a gating unit, a minimum-maximum normalization module and a data output model. The improved Mamba-2 module efficiently processes multivariate long time series and solves the modeling problem of the peak-shaving process of supercritical coal-fired units coupled with molten salt heat storage. DETAILED DESCRIPTION

[0049] The specific implementation modes of the present invention will be further described below with reference to the accompanying drawings.

[0050] Please refer to the attached figure in the instruction manual. Figure 1 , Figure 1 The simplified structure of the peak-shaving process model of the supercritical coal-fired unit coupled with molten salt heat storage in the present invention includes the coal feeding amount u B (t / h), water supply u W (t / h), main steam valve opening μ T (%), low temperature molten salt flow rate D MS (℃) and deoxygenated water flow u DW (t / h) Five input variables, output power N E (MW), separator outlet temperature T I (℃), main steam pressure P st (MPa), high temperature tank inlet molten salt temperature T HM (℃) and supplementary industrial steam flow D SW (m 3 / h) five output variables.

[0051] Please refer to the attached figure in the instruction manual. Figure 2 , Figure 2 The intelligent modeling method for the peak-shaving process of coupled molten salt heat storage and supercritical coal-fired units based on the Mamba-2 large model designed by the present invention is introduced into the Mamba-2 module for training and verification, and finally an accurate state space model of the peak-shaving process of supercritical coal-fired units coupled with molten salt heat storage is output.

[0052] S1: Analyze the difficulties in modeling the deep peak-shaving process of supercritical coal-fired units coupled with molten salt heat storage;

[0053] S2: Establish an intelligent modeling method for the coordinated peak-shaving process of coupled molten salt heat storage and supercritical coal-fired units based on the Mamba-2 large model;

[0054] S3: Rely on the simulation platform to verify and analyze the effectiveness of the intelligent modeling method designed based on the Mamba-2 large model.

[0055] S1: The purpose of the fire-storage coordinated peak-shaving system is to reduce the minimum output power of thermal power generating units and provide space for the grid connection of large-scale wind power generation and photovoltaic power generation systems. In order to achieve safer, more stable, faster and full-condition peak-shaving of the fire-storage coordinated peak-shaving system, the mathematical model of the deep peak-shaving process of the supercritical coal-fired unit coupled with molten salt heat storage can be established in combination with the outstanding Mamba-2 large model to provide model guarantee for the design of control strategies. However, the supercritical coal-fired unit coupled with molten salt heat storage is a large and complex thermal system with many variables that need to be monitored, and the value of each variable will change greatly during the peak-shaving process. Therefore, in order to facilitate the design of the control strategy of the fire-storage coordinated peak-shaving system, it is urgent to establish an accurate state space model of the large-scale peak-shaving process. The supercritical coal-fired unit coupled with molten salt heat storage is mainly composed of a direct-current boiler, a steam turbine and a double-tank molten salt heat storage system. According to the dynamic characteristics of the steam-water process, it can be simplified into a coupled system with five input and five output variables. The five input variables are the coal feed rate u B (t / h), water supply u W (t / h), main steam valve opening μ T (%), low temperature molten salt flow rate D MS (℃) and deoxygenated water flow u DW (t / h). The five output variables are output power N E (MW), separator outlet temperature T I (℃), main steam pressure P st (MPa), high temperature tank inlet molten salt temperature T HM (℃), supplementary industrial steam flow D SW (m 3 / h). During the peak regulation process, the variables of the thermal storage coordinated peak regulation system have strong coupling and nonlinear characteristics and the variables change rapidly and over a large span. It is difficult to quickly establish an accurate peak regulation process model using traditional mathematical modeling methods.

[0056] Based on the analysis of S1, it can be seen that the deep peak-shaving process of supercritical coal-fired units coupled with molten salt heat storage involves large-span changes in the values ​​of multiple variables. Therefore, the improved Mamba-2 large model with a stable selective state space model is used to learn multivariate long time series to obtain the nonlinear mapping relationship between the variables, and an accurate state space model of the peak-shaving system can be established. The five input variables and five output variable data of the deep peak-shaving process of supercritical coal-fired units coupled with molten salt heat storage are imported into the improved Mamba-2 large model for training, and the trained model is tested and output with different multiple groups of data, so as to complete the identification of the state space model of the peak-shaving process of the fire-storage coordinated peak-shaving system. Then step S2 can be concretized as follows:

[0057] S2.1: For a supercritical coal-fired unit coupled with molten salt heat storage, the state space model is as follows:

[0058]

[0059] Input variable matrix u(t) = [u B (t),u w (t),μ T (t),D MS (t),u DW (t)], output variable matrix y(t) = [N E (t),T I (t),P st (t),T HM (t),D SW (t)]. Then, the state space model is discretized, and the supercritical coal-fired unit coupled with molten salt heat storage is rewritten as:

[0060]

[0061] Then, rewrite the model as a differential equation and substitute the derivative definition to obtain the solution of the state space model:

[0062]

[0063] Parameters A, B, C, It changes with the on-site peak-shaving process data of the supercritical coal-fired unit coupled with molten salt heat storage. Δ is the time scale parameter. To speed up the calculation, the above parameters are calculated by matrix multiplication in the form of convolution:

[0064]

[0065] Finally, calculate Mamba's predicted output sequence:

[0066]

[0067] However, Mamba's ability to handle selective copying and contextual learning tasks still needs to be improved. Therefore, the Mamba-2 model with structured state space duality (SSD) as the core is considered to model the peak-shaving process of supercritical coal-fired units coupled with molten salt heat storage:

[0068]

[0069] In the SSD layer of Mamba-2, the matrix A is further restricted to be a scalar. The diagonal elements of the matrix A have the same value. Since Mamba-2 can use tensors to accelerate the calculation of SSM, it is designed to be modular to facilitate efficient modeling of multivariate long time series of the peak-shaving process of supercritical coal-fired units coupled with molten salt heat storage. Parallel parameter design and additional normalization are considered in the modular implementation of Mamba-2. In the parallel parameter design, the SSD layer acts as a mapping relationship between the variable parameters A, B, x and the predicted output value y:

[0070] (A (T) ,B (T,N) ,C (T,N) ,x (T,P) )→y (T,P) (18)

[0071] The SSD layer allows the accelerator to use matrix multiplication units, which significantly improves the training efficiency of Mamba-2. In addition, in order to obtain an improved Mamba-2 with higher computing performance, a sparse attention mechanism is added after the input data module, and an additional minimum-maximum normalization layer is added to the data block before the final model output to improve modeling accuracy.

[0072] In order to improve the contextual learning ability of Mamba-2 for multivariate long time series in the peak-shaving process of supercritical coal-fired units coupled with molten salt heat storage, a sparse attention mechanism was added after the input data module. The sparse attention mechanism is one of the attention mechanisms used to optimize the training process of deep learning models. Its main goal is to only calculate the similarity between the query vector and some key vectors to reduce time and space complexity and save storage space. The nonlinear mapping relationship between variables in the peak-shaving process of supercritical coal-fired units coupled with molten salt heat storage has an important influence on the modeling performance of the unit, and usually requires precise calculation. Another advantage of the sparse attention mechanism is its adjustability. By adjusting the width of the sliding window, the number of computing elements can be controlled to achieve a balance between computing efficiency and modeling performance. Assume that the input matrix of the network is Then the local self-attention operation of the first window G1 on the target grid feature can be expressed as:

[0073]

[0074]

[0075] where h and ω are the height and width of the sliding window, respectively. is the mesh feature in the first window, d k is the scaling factor, D represents the dimension of the grid feature; the query, key and value are respectively in and is the learning parameter. The output of the local self-attention mechanism is

[0076] In order to improve the modeling efficiency of Mamba-2, the grid sparsity λ∈[0,1) is introduced to remove the grid features of the window with probability λ. For the grid features in the sliding window, the features are randomly retained in the proportion of (1-λ) and participate in each batch calculation. Then, the grid features in the first window are removed from Reduced to After sliding (h×ω) times, the target grid features of all windows are connected and the network output is expressed as:

[0077] Z=concatenate(Z1,Z2,…,Z (h×ω) ) (twenty one)

[0078] In addition, to speed up the calculation, the Lion optimizer is added to improve the Mamba-2 model. Compared with the Adam optimizer, the Lion optimizer has fewer parameters, so the cache required for calculation is smaller. In addition, Lion removes the most computationally intensive division and root operations in Adam during the update process, making the calculation faster. Specifically, the parameter update rules in Lion are also different, and the momentum update is after the variable update. The update process of the Lion optimizer is described as:

[0079]

[0080] where sign (·) is a sign function whose value is equal to 1 or -1. represents the gradient of the loss function, μ t is a sign function used to calculate the sign of the gradient. β1 and β2 are the exponential decay rates of the first-order momentum and the second-order momentum when the parameters are updated, respectively. The learning rate is denoted by η t The gradient momentum accumulation is denoted as m t , using m t Accelerate the gradient descent process, λ t is a constant.

[0081] So far, an intelligent modeling method for the coordinated peak-shaving process of coupled molten salt heat storage and supercritical coal-fired units based on the Mamba-2 large model has been completed.

[0082] Based on the intelligent modeling method of the coupled molten salt heat storage and supercritical coal-fired unit coordinated peak-shaving process based on the Mamba-2 large model designed in step S2.1, the effectiveness of the designed modeling method is verified and analyzed on the simulation platform in step S3. The specific process is as follows:

[0083] This example is based on the actual peak load data of the 660MW ultra-supercritical coal-fired unit + 40MW / 80MWh molten salt thermal storage system of Jiangsu Jingjiang Power Plant. The method steps include:

[0084] In this example, the operating conditions of the peak load regulation process of the supercritical coal-fired unit coupled with molten salt heat storage are as follows:

[0085] Table 1 Operating conditions of 40MW / 80MWh+660MW supercritical coal-fired unit coupled with molten salt thermal storage

[0086] S3.1: 20,000 and 5,000 sets of actual historical operation data of the deep peak-shaving process of supercritical coal-fired units coupled with molten salt heat storage are selected respectively to train and verify the designed Mamba-2 model, and finally output the state space model.

[0087] S3.2: When wind and photovoltaic power generation resources are sufficient, the supercritical coal-fired unit coupled with molten salt heat storage quickly reduces its output power to the set value according to the grid peak load instruction, and the values ​​of the five input variables and the five output variables change greatly. In order to obtain the unit state space model of the peak load process, 20,000 groups of units containing coal supply u B (t / h), water supply u W (t / h), main steam valve opening μ T (%), low temperature molten salt flow rate D MS (℃), deoxygenated water flow u DW (t / h), output power N E (MW), separator outlet temperature T I (℃), main steam pressure P st (MPa), high temperature tank inlet molten salt temperature T HM (℃) and supplementary industrial steam flow D SW (m 3 / h) of the actual peak load data of the units were imported into the data module to train the Mamba-2 model. Subsequently, another 5000 sets of actual peak load data of the units were imported to verify the Mamba-2 model.

[0088] S3.3: Observe the training and validation process Five output variables of the Mamba-2 model: Output power N E (MW), separator outlet temperature T I (℃), main steam pressure P st(MPa), high temperature tank inlet molten salt temperature T HM (℃) and supplementary industrial steam flow D SW (m 3 / h).

[0089] S3.4: Calculate the modeling errors of the five output variables of the Mamba-2 model, including the mean absolute error (MAE), mean square error (MSE), mean absolute percentage error (MAPE), symmetric mean absolute percentage error (SMAPE) and modeling accuracy (R), comprehensively evaluate the performance of the peak-shaving process model of the supercritical coal-fired unit coupled with molten salt heat storage, and output the unit state space model.

[0090]

[0091]

[0092]

[0093]

[0094]

[0095] Statistical results show that, combined with the intelligent modeling method of the coupled molten salt heat storage and supercritical coal-fired unit coordinated peak-shaving process based on the Mamba-2 large model designed by the present invention, the modeling accuracy of the five output variables of the system is higher than 99%, and the obtained state space model can be used for control strategy design, which effectively promotes the transformation of thermal power generating units from "main power source" to "regulatory resource".

Claims

1. An intelligent modeling method for the fire-storage coordinated peak-shaving process based on the Mamba-2 large model is characterized by: The following steps are involved: S1: Analyze the difficulties in modeling the deep peak-shaving process of supercritical coal-fired units coupled with molten salt heat storage; S2: Establish an intelligent modeling method for the coordinated peak-shaving process of coupled molten salt heat storage and supercritical coal-fired units based on the Mamba-2 large model; the details are as follows: S2.1: For a supercritical coal-fired unit coupled with molten salt heat storage, the state space model is as follows: Input variable matrix u(t) = [u B (t),u w (t),μ T (t),D MS (t),u DW (t)], output variable matrix y(t) = [N E (t),T I (t),P st (t),T HM (t),D SW (t)]; Then, the state space model is discretized, and the supercritical coal-fired unit coupled with molten salt heat storage is rewritten as: Then, rewrite the model as a differential equation and substitute the derivative definition to obtain the solution of the state space model: Parameters A, B, C, It changes with the on-site peak-shaving process data of the supercritical coal-fired unit coupled with molten salt heat storage; Δ is the time scale parameter; to speed up the calculation, the above parameters are calculated by matrix multiplication in the form of convolution: Finally, calculate Mamba's predicted output sequence: However, Mamba's ability to handle selective copying and contextual learning tasks still needs to be improved; therefore, the Mamba-2 model with structured state space duality (SSD) as the core is considered to model the peak-shaving process of supercritical coal-fired units coupled with molten salt heat storage: In the SSD layer of Mamba-2, the matrix A is further restricted to a scalar; the diagonal elements of the matrix A have the same value; since Mamba-2 can use tensors to accelerate the calculation of SSM, it is designed to be modular to facilitate efficient modeling of multivariate long time series of the peak-shaving process of supercritical coal-fired units coupled with molten salt heat storage; Parallel parameter design and additional normalization are considered in the Mamba-2 modular implementation; in the parallel parameter design, the SSD layer is used as the mapping relationship between the variable parameters A, B, x and the predicted output value y: (A (T) ,B (T,N) ,C (T,N) ,x (T,P) )→y (T,P) (7) The SSD layer allows the accelerator to use matrix multiplication units, which significantly improves the training efficiency of Mamba-2. In addition, to obtain an improved Mamba-2 with higher computing performance, a sparse attention mechanism is added after the input data module, and an additional min-max normalization module is added to the data block before the final model output to improve modeling accuracy. To improve the contextual learning capability of Mamba-2 for multivariate long time series of peak-shaving process of supercritical coal-fired units coupled with molten salt heat storage, a sparse attention mechanism is added after the input data module; The sparse attention mechanism is one of the attention mechanisms used to optimize the training process of deep learning models. Its main goal is to only calculate the similarity between the query vector and some key vectors to reduce time and space complexity and save storage space. The nonlinear mapping relationship between variables in the peak-shaving process of supercritical coal-fired units coupled with molten salt heat storage has an important impact on the modeling performance of the unit and usually requires precise calculation. Another advantage of the sparse attention mechanism is its adjustability. By adjusting the width of the sliding window, the number of computing elements can be controlled to achieve a balance between computing efficiency and modeling performance. Assume that the input matrix of the network is Then the local self-attention operation of the first window G1 on the target grid feature can be expressed as: Where h and ω are the height and width of the sliding window respectively; is the mesh feature in the first window, d k is the scaling factor, D represents the dimension of the grid feature; the query, key and value are respectively in and To learn the parameters, the output of the local self-attention mechanism is In order to improve the modeling efficiency of Mamba-2, the grid sparsity λ∈[0,1) is introduced to remove the grid features of the window with probability λ; For the grid features in the sliding window, randomly retain the features at a ratio of (1-λ) to participate in each batch calculation; then, the grid features in the first window are Reduced to After sliding (h×ω) times, the target grid features of all windows are connected and the network output is expressed as: Z=concatenate(Z1,Z2,…,Z (h×ω) ) (10) In addition, to speed up the calculation, the Lion optimizer is added to improve the Mamba-2 model. Compared with the Adam optimizer, the Lion optimizer has fewer parameters, so the cache required for calculation is smaller. In addition, Lion removes the most computationally intensive division and root operations in Adam during the update process, making the calculation faster. Specifically, the parameter update rules in Lion are also different, and the momentum update is after the variable update. The update process of the Lion optimizer can be described as follows: where sign (·) is a sign function whose value is equal to 1 or -1; represents the gradient of the loss function, μ t is a sign function used to calculate the sign of the gradient; β1 and β2 are the exponential decay rates of the first-order momentum and the second-order momentum when the parameters are updated, respectively, and the learning rate is recorded as η t ; Gradient momentum accumulation is recorded as m t , using m t Accelerate the gradient descent process, λ t is a constant; So far, an intelligent modeling method design for the coordinated peak-shaving process of coupled molten salt heat storage and supercritical coal-fired units based on the Mamba-2 large model has been completed; Based on the intelligent modeling method of the coupled molten salt heat storage and supercritical coal-fired unit coordinated peak-shaving process based on the Mamba-2 large model designed in step S2.1, the effectiveness of the designed modeling method is verified and analyzed in step S3 by relying on the simulation platform: S3: Rely on the simulation platform to verify and analyze the effectiveness of the intelligent modeling method designed based on the Mamba-2 large model.

2. According to claim 1, an intelligent modeling method for fire-storage coordinated peak-shaving process based on the Mamba-2 large model is characterized by: The difficulties in modeling the deep peak regulation process of the supercritical coal-fired unit coupled with molten salt heat storage described in step S1 are as follows: S1: The purpose of the fire-storage coordinated peak-shaving system is to reduce the minimum output power of the thermal power generating unit and provide space for the grid connection of large-scale wind power generation and photovoltaic power generation systems. In order to achieve safer, more stable, faster and full-operating peak-shaving of the fire-storage coordinated peak-shaving system, the mathematical model of the deep peak-shaving process of the supercritical coal-fired unit coupled with molten salt heat storage can be established in combination with the outstanding Mamba-2 large model to provide model guarantee for the design of control strategy. However, the supercritical coal-fired unit coupled with molten salt heat storage is a large and complex thermal system with many variables that need to be monitored, and the value of each variable will change greatly during the peak-shaving process. Therefore, in order to facilitate the design of the control strategy of the fire-storage coordinated peak-shaving system, it is urgent to establish an accurate state space model of the large-scale peak-shaving process. The supercritical coal-fired unit coupled with molten salt heat storage is mainly composed of a direct-current boiler, a steam turbine and a double-tank molten salt heat storage system. According to the dynamic characteristics of the steam-water process, it can be simplified into a coupled system with five input and five output variables. The five input variables are the coal feed rate u B (t / h), water supply u W (t / h), main steam valve opening μ T (%), low temperature molten salt flow rate D MS (℃) and deoxygenated water flow u DW (t / h); the five output variables are output power N E (MW), separator outlet temperature T I (℃), main steam pressure P st (MPa), high temperature tank inlet molten salt temperature T HM (℃) and supplementary industrial steam flow D SW (m 3 / h); During the peak-shaving process, the variables of the thermal-storage coordinated system have strong coupling and nonlinear characteristics and change rapidly over a large span. It is difficult to quickly establish an accurate peak-shaving process model using traditional mathematical modeling methods.

3. According to claim 1, a method for intelligent modeling of a fire-storage coordinated peak-shaving process based on the Mamba-2 large model is characterized by: Based on the Mamba-2 intelligent modeling method designed in step S2, the effectiveness of the proposed intelligent modeling method based on the Mamba-2 large model is verified and analyzed in step S3 by relying on the simulation platform. Specifically, the following are performed: S3.1: 20,000 and 5,000 sets of actual historical operation data of deep peak-shaving process of supercritical coal-fired units coupled with molten salt heat storage were selected respectively, the designed Mamba-2 model was trained and verified, and finally the state space model was output; S3.2: When wind and photovoltaic power generation resources are sufficient, the supercritical coal-fired unit coupled with molten salt heat storage quickly reduces its output power to the set value according to the grid peak load instruction, and the values ​​of the five input variables and the five output variables change greatly. In order to obtain the unit state space model of the peak load process, 20,000 groups of units containing coal supply u B (t / h), water supply u W (t / h), main steam valve opening μ T (%), low temperature molten salt flow rate D MS (℃), deoxygenated water flow u DW (t / h), output power N E (MW), separator outlet temperature T I (℃), main steam pressure P st (MPa), high temperature tank inlet molten salt temperature T HM (℃) and supplementary industrial steam flow D SW (m 3 / h) of the actual peak-shaving data of the units were imported into the data module to train the Mamba-2 model; then, another 5,000 sets of actual peak-shaving data of the units were imported to verify the Mamba-2 model; S3.3: Observe the training and validation process Five output variables of the Mamba-2 model: Output power N E (MW), separator outlet temperature T I (℃), main steam pressure P st (MPa), high temperature tank inlet molten salt temperature T HM (℃) and supplementary industrial steam flow D SW (m 3 / h); S3.4: Calculate the modeling errors of the five output variables of the Mamba-2 model, including mean absolute error (MAE), mean square error (MSE), mean absolute percentage error (MAPE), symmetric mean absolute percentage error (SMAPE) and modeling accuracy (R), comprehensively evaluate the performance of the peak-shaving process model of the supercritical coal-fired unit coupled with molten salt heat storage, and output the unit state space model; Statistical results show that, combined with the intelligent modeling method of the coordinated peak-shaving process of coupled molten salt heat storage and supercritical coal-fired units based on the Mamba-2 large model proposed in the present invention, the modeling accuracy of the five output variables of the system is higher than 99%, and the obtained state space model can be used for control strategy design, which effectively promotes the transformation of thermal power generating units from "main power source" to "regulatory resource".

Images