A deep learning-based furnace temperature prediction and control system

Abstract

The application discloses a kind of furnace temperature prediction and control system based on deep learning, it is related to thermal power generation technology field.The system includes: data preprocessing module, for obtaining and processing DCS operation data, constructs historical time series data;Key parameter prediction module, using the deep learning time series prediction model of fusion quality balance, energy balance and NOx generation dynamics constraint, output the future change trend of multiple parameters;Multi-objective rolling optimization module, generate optimal control set value sequence;Instruction safety fusion module, after the safe fusion of optimization instruction and DCS original protection logic, issue and execute;And man-machine interaction and performance monitoring module.The application is through the integration framework of physical guided deep learning prediction and rolling time domain optimization, effectively solve the problem of large delay and strong coupling in thermal power unit combustion control, while improving load response speed, guarantee main steam parameter stability, improve boiler efficiency and reduce pollutant emission.

Description

Technical Field

[0001] This invention relates to the field of thermal power generation technology, specifically to a furnace temperature prediction and control system based on deep learning. Background Technology

[0002] Power plant boilers are the core thermal equipment in thermal power plants, and the stability, economy, and environmental friendliness of their combustion process directly affect the safe operation and overall efficiency of the unit. As a key parameter reflecting the combustion state, precise control of furnace temperature is crucial for ensuring stable main steam parameters, improving boiler efficiency, and reducing pollutant emissions. Currently, large-scale thermal power generating units generally employ coordinated control systems (CCS) and burner management systems (BMS) based on classical physical models, and achieve closed-loop control through distributed control systems (DCS).

[0003] However, existing technologies still face the following pressing problems in practical applications: there is a significant delay in the response from changes in fuel quantity to furnace temperature and main steam parameters. Furthermore, the strong coupling between the generator, furnace, and power systems creates an inherent contradiction between load tracking and parameter stability when responding to Automatic Generation Control (AGC) commands. Traditional proportional-integral-derivative (PID) control and conventional feedforward control struggle to achieve rapid and error-free regulation, resulting in large furnace temperature fluctuations that negatively impact combustion efficiency and equipment lifespan.

[0004] The combustion process in the furnace involves the coordinated operation of multiple subsystems, including coal feeding, forced draft, induced draft, and water supply. The variables exhibit complex nonlinear relationships and strong coupling characteristics. In actual operation, frequent fluctuations in coal quality and changes in operating conditions such as the start-up and shutdown of the coal mill further exacerbate the model mismatch problem, making it difficult for traditional control strategies based on fixed mechanism models to maintain good performance under all operating conditions.

[0005] In recent years, although some power plants have attempted to introduce artificial intelligence technology, most applications have been limited to local parameter tuning, fault diagnosis, or soft measurement, lacking an integrated intelligent control framework that deeply integrates full-condition prediction, multi-objective dynamic optimization, and low-level control execution. The design of predictive models and controllers is often decoupled, making it difficult to form closed-loop optimization and limiting further improvements in intelligence levels. Summary of the Invention

[0006] Therefore, this invention provides a furnace temperature prediction and control system based on deep learning to solve the problem of how to improve the load response capability and operational stability of thermal power generating units in the prior art.

[0007] To achieve the above objectives, the present invention provides the following technical solution:

[0008] A furnace temperature prediction and control system based on deep learning includes a data preprocessing module, a key parameter prediction module, a multi-objective rolling optimization module, a command safety fusion module, and a human-machine interaction and performance monitoring module.

[0009] The data preprocessing module is used to acquire panoramic data of boiler operation in the distributed control system in real time through a standard communication interface, and to perform validity verification, filtering, time-series alignment and standardization on the panoramic data in order to construct and output historical time-series data characterizing the dynamic state of the boiler.

[0010] The key parameter prediction module has a built-in deep learning time series prediction model that integrates physical guidance, which is used to take the historical time series data as input and output future working condition information within a preset time period.

[0011] The multi-objective rolling optimization module is used to receive the future operating condition information and the current automatic power generation control load command of the power grid, and based on the rolling time domain optimization strategy, solve a multi-objective optimization problem in a finite time domain under the condition of satisfying the physical limits of the equipment and the safety operation constraints, so as to output the optimal control setpoint sequence in the future period of time.

[0012] The instruction security fusion module is used to securely fuse the control setpoint sequence with the original logic protection, rate limiting and priority management in the distributed control system, generate the final executable security optimization setpoint, and send it to each sub-loop of the basic control layer for execution.

[0013] The human-computer interaction and performance monitoring module is used to provide a graphical human-computer interface to display the prediction curve in real time, set the optimization target weight and statistical system performance indicators, and monitor the performance of the deep learning time series prediction model. When the prediction error continues to exceed the preset threshold, it issues a prompt signal or initiates the model parameter fine-tuning process.

[0014] Furthermore, the panoramic data acquired by the data preprocessing module includes unit load command, main steam pressure, main steam temperature, reheat steam temperature, total coal feed, output of each coal mill, total air volume, primary and secondary air ratio, flue gas oxygen content, SCR inlet nitrogen oxide concentration, furnace negative pressure, wall temperature of each heating surface, and multi-point temperature of the furnace.

[0015] Furthermore, the deep learning time-series prediction model in the key parameter prediction module incorporates physical constraints based on mass balance, energy balance, and simplified nitrogen oxide generation kinetics into its loss function during training to ensure that the model's prediction trend conforms to basic physicochemical laws; the future operating condition information includes the changing trends of furnace temperature, main steam pressure, main steam temperature, and SCR inlet nitrogen oxide concentration.

[0016] Furthermore: the loss function is specifically expressed as follows:

[0017] ;

[0018] in, This is the total loss function; For data fitting terms; The mean of the sum of squares of the residuals for the quality balance constraint; The mean of the sum of squares of the residuals under energy balance constraints; The mean of the sum of squares of the residuals under NOx kinetic constraints; These are the weighting coefficients for the mass balance constraints; These are the weighting coefficients for energy balance constraints; These are the weighting coefficients for NOx kinetic constraints; To input history window, For the corresponding real future output, These are model parameters; This is for predicting the output.

[0019] Furthermore: the mean of the sum of squares of the residuals of the mass balance constraint The calculation formula is:

[0020] ;

[0021] ;

[0022] in, The total weight of the working fluid in the furnace at step k; Let K be the mass flow rate entering the furnace at step k. Let K be the mass flow rate leaving the furnace at step k. The mass conservation discretization deviation at step k; is the discretization time step; K is the total number of time steps involved in the constraint calculation.

[0023] Furthermore: the mean of the sum of squares of the residuals of the energy balance constraint The calculation formula is:

[0024] ;

[0025] ;

[0026] in, The discretization deviation for energy conservation at the k-th step; For storing thermal energy in the furnace in step k; Input the thermal power of the furnace at step k; The heat power carried away from the furnace in step k; is the discretization time step; K is the total number of time steps involved in the constraint calculation.

[0027] Furthermore: the mean of the sum of squares of the residuals of the NOx kinetic constraints The calculation formula is:

[0028] ;

[0029] ;

[0030] in, Discretization bias of the NOx generation kinetic equation for step k; Input the thermal power of the furnace at step k; Let K be the flame temperature at step k. Let be the oxygen concentration at step k; Let be the residence time of the flue gas at step k; is the discretization time step; K is the total number of time steps involved in the constraint calculation.

[0031] Furthermore, when solving the multi-objective rolling optimization problem, the objective function J of the multi-objective rolling optimization module comprehensively considers the main steam parameter tracking error, pollutant emissions, boiler efficiency, and control action change rate.

[0032] Furthermore, the objective function is specifically expressed as follows:

[0033]

[0034] ;

[0035] ;

[0036] ;

[0037] ;

[0038] ;

[0039] Where k is the discrete-time index, k=0 corresponds to the current time; N is the prediction time domain length; M is the control time domain length; α, β, γ, δ, ε are the weight coefficients in the objective function; The expected value of the main steam pressure; P(k) is the expected value of temperature; P(k) is the predicted value of main steam pressure; T(k) is the predicted value of temperature; N(k) is the predicted value of NOx concentration at the SCR inlet. This is a predicted value for boiler efficiency; ; and These are the lower and upper bounds of the state variable, respectively; and These are the lower and upper bounds of the control variable, respectively. To control the maximum allowable value of the rate of change; It is a state vector; The control input vector; x(0) is the current state measurement value; For reference trajectory; To control the sequence.

[0040] Furthermore, the human-computer interaction and performance monitoring module also provides an interface with the power plant monitoring information system to obtain historical operating data to support model training and upload the optimization results to the power plant monitoring information system.

[0041] This invention has the following advantages: It employs a physics-guided deep learning time-series prediction model, using mass balance, energy balance, and nitrogen oxide generation kinetics as constraints in the loss function. This significantly improves the accuracy and robustness of the prediction model under all operating conditions, ensuring that the predicted trend conforms to physical laws. Furthermore, it introduces a multi-objective rolling time-domain optimization strategy, dynamically tracking AGC load commands while comprehensively balancing main steam parameter stability, denitrification efficiency, and boiler energy efficiency, achieving dual optimization of rapid unit response and stable operation. Finally, through a command safety fusion module, it seamlessly integrates optimized commands with the underlying DCS protection logic, ensuring that intelligent control is implemented within safe boundaries, thereby comprehensively improving the economy, safety, and environmental friendliness of thermal power unit operation.

[0042] Other features and advantages of the present invention will be set forth in the following description. Attached Figure Description

[0043] To more intuitively illustrate the prior art and this application, exemplary drawings are provided below. It should be understood that the specific shapes and structures shown in the drawings should not generally be regarded as limiting conditions for implementing this application; for example, based on the technical concept disclosed in this application and the exemplary drawings, those skilled in the art are able to easily make conventional adjustments or further optimizations to the addition / reduction / classification, specific shapes, positional relationships, connection methods, size ratios, etc. of certain units (components).

[0044] Figure 1 This is a flowchart illustrating the implementation of a deep learning-based furnace temperature prediction and control system, as provided in an embodiment of this application. Detailed Implementation

[0045] The following specific embodiments illustrate the implementation of the present invention. Those skilled in the art can easily understand other advantages and effects of the present invention from the content disclosed in this specification. Obviously, the described embodiments are only some embodiments of the present invention, not all embodiments. It should be understood that these embodiments are merely for further explanation of the present invention and should not be construed as limiting the scope of protection of the present invention. Those skilled in the art can make some non-essential improvements and adjustments to the present invention based on the above-described content.

[0046] Please see Figure 1 A furnace temperature prediction and control system based on deep learning includes a data preprocessing module, a key parameter prediction module, a multi-objective rolling optimization module, an instruction safety fusion module, and a human-machine interaction and performance monitoring module.

[0047] The data preprocessing module acquires real-time panoramic data of boiler operation from the DCS via a standard communication interface. This panoramic data includes unit load commands, main steam pressure, main steam temperature, reheat steam temperature, total coal feed, output of each coal mill, total air volume, primary and secondary air ratio, flue gas oxygen content, selective catalytic reduction (SCR) inlet nitrogen oxide concentration, furnace negative pressure, wall temperature of each heating surface, and multi-point temperature of the furnace. The data preprocessing module performs validity verification, filtering, time-series alignment, and standardization on the acquired panoramic data, constructing a historical time-series sequence characterizing the dynamic state of the boiler, and outputting the cleaned historical time-series data.

[0048] The key parameter prediction module adopts a deep learning time series prediction model that integrates physics guidance. It takes the cleaned historical time series data mentioned above as input and outputs future operating condition information for the next 10-30 minutes. The future operating condition information includes the changing trends of multiple key controlled variables and intermediate states. Among them, the key controlled variables include furnace temperature, main steam pressure, main steam temperature, and SCR inlet nitrogen oxide concentration.

[0049] Furthermore, during the training process, the loss function of the deep learning time series prediction model incorporates physical constraints based on mass balance, energy balance, and simplified nitrogen oxide generation kinetics to ensure that the predicted trend conforms to basic physicochemical laws and enhance the extrapolation robustness of the deep learning time series prediction model under unexperienced operating conditions.

[0050] Assume the training sample set is ;in To input history window, The model parameters are the corresponding true future output. The predicted output is The loss function is then defined as:

[0051] ;

[0052] in, This is the total loss function; For data fitting terms; The mean of the sum of squares of the residuals for the quality balance constraint; The mean of the sum of squares of the residuals under energy balance constraints; The mean of the sum of squares of the residuals under NOx kinetic constraints; These are the weighting coefficients for the mass balance constraints; These are the weighting coefficients for energy balance constraints; These are the weighting coefficients for NOx kinetic constraints;

[0053] Under mass balance constraints, the mass conservation of the boiler system can be simplified to:

[0054] ;

[0055] Where M represents the mass of the working fluid inside the furnace, This includes the amount of coal fed and the amount of air supplied. This includes steam flow rate and flue gas quality;

[0056] Discretize the above differential equation to obtain the residual at the Kth step;

[0057] ;

[0058] The mass balance constraint term is taken as the mean of the sum of squares of the residuals at all time steps; therefore, the formula for calculating the mass balance constraint is:

[0059] ;

[0060] in, The total weight of the working fluid in the furnace at step k; Let K be the mass flow rate entering the furnace at step k. Let K be the mass flow rate leaving the furnace at step k. The mass conservation discretization deviation at step k is given.

[0061] Under energy balance constraints, the energy conservation of a boiler can be expressed as:

[0062] ;

[0063] Where E represents the heat storage energy inside the furnace. The heat released by fuel combustion The residuals after discretization are: (This includes heat carried away by the steam, heat dissipation losses, etc.)

[0064] ;

[0065] The energy balance constraint term is:

[0066] ;

[0067] in, The discretization deviation for energy conservation at the k-th step; For storing thermal energy in the furnace in step k; Input the thermal power of the furnace at step k; The heat power carried away from the furnace in step k;

[0068] In the constraints of NOx generation kinetics, simplified empirical formulas are used, such as the Zeldovich mechanism or a multinomial regression model, and the formula can be:

[0069] ;

[0070] Where T is the flame temperature. Oxygen concentration, Let be the dwell time; then the discretized residual is:

[0071] ;

[0072] The NOx constraint term is then:

[0073] ;

[0074] in, Discretization bias of the NOx generation kinetic equation for step k; Input the thermal power of the furnace at step k; Let K be the flame temperature at step k. Let be the oxygen concentration at step k; Let be the residence time of the flue gas at step k; is the discretization time step; K is the total number of time steps involved in the constraint calculation.

[0075] The multi-objective rolling optimization module receives future operating condition information from the key parameter prediction module and the current grid load command (AGC load command). The core of the multi-objective rolling optimization module is a controller based on rolling time-domain optimization. This controller solves a finite-time-domain optimal control problem in each control cycle, assuming the discrete-time system is:

[0076] ;

[0077] in, The state vector contains the main steam pressure P(k), main steam temperature T(k), and furnace temperature T. f NOx concentration N(k), boiler efficiency η(k), etc.

[0078] The control input vector includes total coal feed uc(k), total air volume ua(k), air-coal ratio correction ur(k), and burnout damper opening uo(k); f is a deep learning prediction model with fixed parameters, which is considered a known nonlinear function.

[0079] Given the current state measurement x(0) and the reference trajectory The prediction time domain length is N, and the control time domain length is M (usually M≤N); let the control sequence be... The optimization problem is as follows:

[0080]

[0081] in, ;

[0082] ;

[0083] ;

[0084] ;

[0085] ;

[0086] Where k is the discrete-time index, k=0 corresponds to the current time; N is the prediction time domain length (number of steps); M is the control time domain length, when k≥M, u(k)=u(M−1); α, β, γ, δ, ε are the weight coefficients in the objective function, which can be adjusted online according to operational requirements; The expected value of the main steam pressure; The expected temperature value is usually set by AGC load instructions and operators; P(k) is the predicted main steam pressure; T(k) is the predicted temperature; N(k) is the predicted NOx concentration at the SCR inlet. For the predicted value of boiler efficiency, 1 / This represents the reciprocal of efficiency; the smaller the value, the higher the efficiency. For k=0, u(−1)=u(0) can be defined or handled separately; and These are the lower and upper bounds of the state variable, respectively; and These are the lower and upper bounds of the control variable, respectively. This is the maximum allowable value for controlling the rate of change.

[0087] In this embodiment, the weighting coefficients α, β, γ, δ, and ε can be adjusted online according to operational requirements;

[0088] The constraints of the optimization problem include boiler dynamic prediction model constraints, equipment physical limit constraints, and safe operation constraints; the boiler dynamic prediction model is a deep learning time series prediction model provided by the key parameter prediction module.

[0089] After solving the optimization problem online, the multi-objective rolling optimization module outputs the optimal control setpoint sequence for a future period of time. The control setpoint sequence includes the total coal feed setpoint, the total air volume setpoint, the air-coal ratio correction value, and the suggested value for the burnout damper opening.

[0090] The command safety fusion module integrates the advanced commands calculated by the multi-objective rolling optimization module with the existing logic protection, rate limiting, and priority management in the DCS to generate the final executable safety optimization settings, which are then distributed to each sub-loop of the basic control layer (fuel master control, air volume master control, denitrification control, etc.). The command safety fusion module ensures that the intelligent optimization commands always operate within the safety allowable range, and all interlocking protections in the original DCS (such as main fuel trip MFT) are independent and have the highest priority.

[0091] The human-computer interaction and performance monitoring module provides a graphical human-computer interface for real-time display of prediction curves, optimization of target weight settings, and system performance metric statistics. At the same time, this human-computer interaction and performance monitoring module can monitor the performance of the deep learning time series prediction model. When the prediction error of the deep learning time series prediction model continues to exceed a preset threshold, it issues a prompt signal or, after authorization, initiates a model parameter fine-tuning process based on recent historical data to achieve online adaptive updates of the deep learning time series prediction model.

[0092] In this embodiment, the system of the present invention also interfaces with a power plant monitoring information system (SIS) to obtain historical operating data to support model training and upload optimization results to the SIS system.

[0093] The above description is only a preferred embodiment of the present invention and is not intended to limit the present invention. Any modifications, equivalent substitutions, and improvements made within the spirit and principles of the present invention should be included within the protection scope of the present invention.

Claims

1. A furnace temperature prediction and control system based on deep learning, characterized in that, It includes a data preprocessing module, a key parameter prediction module, a multi-objective rolling optimization module, an instruction security fusion module, and a human-computer interaction and performance monitoring module; The data preprocessing module is used to acquire panoramic data of boiler operation in the distributed control system in real time through a standard communication interface, and to perform validity verification, filtering, time-series alignment and standardization on the panoramic data in order to construct and output historical time-series data characterizing the dynamic state of the boiler. The key parameter prediction module has a built-in deep learning time series prediction model that integrates physical guidance, which is used to take the historical time series data as input and output future working condition information within a preset time period. The multi-objective rolling optimization module is used to receive the future operating condition information and the current automatic power generation control load command of the power grid, and based on the rolling time domain optimization strategy, solve a multi-objective optimization problem in a finite time domain under the condition of satisfying the physical limits of the equipment and the safety operation constraints, so as to output the optimal control setpoint sequence in the future period of time. The instruction security fusion module is used to securely fuse the control setpoint sequence with the original logic protection, rate limiting and priority management in the distributed control system, generate the final executable security optimization setpoint, and send it to each sub-loop of the basic control layer for execution. The human-computer interaction and performance monitoring module is used to provide a graphical human-computer interface to display the prediction curve in real time, set the optimization target weight and statistical system performance indicators, and monitor the performance of the deep learning time series prediction model. When the prediction error continues to exceed the preset threshold, it issues a prompt signal or initiates the model parameter fine-tuning process.

2. The furnace temperature prediction and control system based on deep learning according to claim 1, characterized in that, The panoramic data acquired by the data preprocessing module includes unit load command, main steam pressure, main steam temperature, reheat steam temperature, total coal feed, output of each coal mill, total air volume, primary and secondary air ratio, flue gas oxygen content, SCR inlet nitrogen oxide concentration, furnace negative pressure, wall temperature of each heating surface, and multi-point temperature of the furnace.

3. The furnace temperature prediction and control system based on deep learning according to claim 1, characterized in that, The deep learning time-series prediction model in the key parameter prediction module incorporates physical constraints based on mass balance, energy balance, and simplified nitrogen oxide generation kinetics into its loss function during training to ensure that the model's prediction trend conforms to basic physicochemical laws. The future operating condition information includes the changing trends of furnace temperature, main steam pressure, main steam temperature, and SCR inlet nitrogen oxide concentration.

4. The furnace temperature prediction and control system based on deep learning according to claim 3, characterized in that, The loss function is specifically expressed as follows: ； in, This is the total loss function; For data fitting terms; The mean of the sum of squares of the residuals for the quality balance constraint; The mean of the sum of squares of the residuals under energy balance constraints; The mean of the sum of squares of the residuals under NOx kinetic constraints; These are the weighting coefficients for the mass balance constraints; These are the weighting coefficients for energy balance constraints; These are the weighting coefficients for NOx kinetic constraints; To input history window, For the corresponding real future output, These are model parameters; This is for predicting the output.

5. A furnace temperature prediction and control system based on deep learning according to claim 4, characterized in that, The mean of the sum of squares of the residuals of the mass balance constraint The calculation formula is: ； ； in, The total weight of the working fluid in the furnace at step k; Let K be the mass flow rate entering the furnace at step k. Let K be the mass flow rate leaving the furnace at step k. The mass conservation discretization deviation at step k; is the discretization time step; K is the total number of time steps involved in the constraint calculation.

6. The furnace temperature prediction and control system based on deep learning according to claim 4, characterized in that, The mean of the sum of squares of the residuals of the energy balance constraints The calculation formula is: ； ； in, The discretization deviation for energy conservation at the k-th step; For storing thermal energy in the furnace in step k; Input the thermal power of the furnace at step k; The heat power carried away from the furnace in step k; is the discretization time step; K is the total number of time steps involved in the constraint calculation.

7. The furnace temperature prediction and control system based on deep learning according to claim 4, characterized in that, The mean of the sum of squares of the residuals of the NOx kinetic constraints The calculation formula is: ； ； in, Discretization bias of the NOx generation kinetic equation for step k; Input the thermal power of the furnace at step k; Let K be the flame temperature at step k. Let be the oxygen concentration at step k; Let be the residence time of the flue gas at step k; is the discretization time step; K is the total number of time steps involved in the constraint calculation.

8. The furnace temperature prediction and control system based on deep learning according to claim 1, characterized in that, When solving the multi-objective optimization problem, the objective function J of the multi-objective rolling optimization module comprehensively considers the main steam parameter tracking error, pollutant emissions, boiler efficiency, and control action change rate.

9. A furnace temperature prediction and control system based on deep learning according to claim 8, characterized in that, The objective function is specifically expressed as follows: ； ； ； ； ； Where k is the discrete-time index, k=0 corresponds to the current time; N is the prediction time domain length; M is the control time domain length; α, β, γ, δ, ε are the weight coefficients in the objective function; The expected value of the main steam pressure; P(k) is the expected value of temperature; P(k) is the predicted value of main steam pressure; T(k) is the predicted value of temperature; N(k) is the predicted value of NOx concentration at the SCR inlet. This is a predicted value for boiler efficiency; ; and These are the lower and upper bounds of the state variable, respectively; and These are the lower and upper bounds of the control variable, respectively. To control the maximum allowable value of the rate of change; It is a state vector; The control input vector; x(0) is the current state measurement value; For reference trajectory; To control the sequence.

10. A furnace temperature prediction and control system based on deep learning according to claim 1, characterized in that, The human-computer interaction and performance monitoring module also provides an interface with the power plant monitoring information system to obtain historical operating data to support model training and upload optimization results to the power plant monitoring information system.

Images