A Smart Closed-Loop Optimization Control System and Method for SCR Denitrification Based on Adaptive Robust Prediction of Key Variables
By constructing an adaptive, robust, predictive, intelligent closed-loop optimization control system for SCR denitrification, the problems of poor adaptability and weak robustness in existing technologies have been solved. This has enabled precise control of ammonia injection and long-term stable operation of the system, reducing maintenance costs and improving the system's robustness and safety.
Patent Information
- Application Number
- CN202611400034.9
- Authority / Receiving Office
- CN · China
- Patent Type
- Applications(China)
- Current Assignee / Owner
- Filing Date
- 2026-09-10
- Publication Date
- 2026-11-17
AI Technical Summary
Existing SCR denitrification ammonia injection closed-loop control systems have poor adaptability, weak robustness, are not globally optimal, and suffer from uncontrollable model degradation, making it difficult to achieve long-term stable and economical automatic operation.
A smart closed-loop optimization control system for SCR denitrification based on adaptive robust prediction of key variables is constructed, including a multi-source sensing and fast response inner loop, an adaptive prediction model middle loop, a robust optimization decision outer loop, and a multi-objective dynamic optimization and command fusion module, to achieve intelligent closed-loop control with self-sensing, self-optimization, and self-evolution.
It achieves long-term, accurate, stable, and economical control of ammonia injection volume, reduces operation and maintenance costs, improves system reliability and safety, and meets stringent environmental protection requirements.
Smart Images

Figure CN123113359A_ABST
Abstract
Description
Technical Field
[0001] This invention belongs to the field of industrial flue gas purification and advanced process control technology, specifically relating to an intelligent closed-loop optimization control system and method for SCR denitrification based on adaptive robust prediction of key variables. Background Technology
[0002] In industrial production scenarios such as coal-fired power plants and industrial boilers, selective catalytic reduction (SCR) technology is the most widely used and technologically mature flue gas denitrification method. Its core control logic is to precisely adjust the flow rate of ammonia injected into the flue gas, and under the catalytic action of a catalyst, reduce nitrogen oxides (NOx) in the flue gas to nitrogen and water, thereby achieving compliant emissions. The accuracy of ammonia injection directly determines the operating effect of the denitrification system. Insufficient ammonia injection will lead to excessive NOx concentration at the outlet, failing to meet environmental emission limits; excessive ammonia injection will cause ammonia escape, resulting in corrosion and blockage of flue gas equipment, secondary environmental pollution, and a significant increase in operating costs. Therefore, achieving precise closed-loop control of ammonia injection in the SCR denitrification system is crucial to ensuring the long-term stable, economical, and efficient operation of the system.
[0003] Existing SCR denitrification ammonia injection closed-loop control systems are mainly divided into four categories. Each type of system has significant technical defects in practical applications and cannot meet the long-term operation requirements of industrial sites. The first type is a conventional closed-loop control system based on single-point feedback. This type of system generally uses the measured value of the outlet NOx concentration monitoring device (CEMS) as the feedback signal. After comparing it with the set value, the total amount of ammonia injection is adjusted by a PID controller. However, its core defect is that there is a large inertia and large pure lag characteristic of several minutes between the ammonia injection point and the outlet CEMS detection point. This causes the PID controller to respond with serious lag. When the boiler load and coal type change, it is easy to have a large overshoot or a long adjustment process, resulting in poor control quality. In order to ensure emission safety, a high NOx set value is often set in actual operation, resulting in long-term excessive ammonia injection and high operating costs. The second category is based on cascade or feedforward-feedback closed-loop control systems using NOx and ammonia slip at the outlet. These systems, based on conventional PID control, introduce ammonia slip concentration as a secondary controlled variable to construct cascade control, or introduce boiler load signals as feedforward compensation. While this improves the system's dynamic response characteristics to some extent, it does not fundamentally solve the core problem of large time lag. Furthermore, the control model uses fixed parameters, making it unable to adapt to the time-varying dynamic characteristics of the SCR system, and its control effect remains significantly limited under rapid disturbance conditions. The third category is model predictive control (MPC) systems based on static models. Some advanced control schemes introduce predictive models based on mechanistic analysis or data identification, forming an optimized closed loop through rolling optimization calculations of ammonia injection quantity. This has certain advantages over conventional PID control, but suffers from two key drawbacks: model mismatch and fragile closed-loop control. The SCR denitrification process is a highly nonlinear and slowly time-varying industrial process. The catalyst activity decays with operating time, the coal quality changes frequently, and the on-site monitoring instruments are prone to drift. After the system has been running for a period of time, the fixed prediction model established offline will inevitably become severely mismatched with the actual operating process, resulting in a significant decrease in the model prediction accuracy, degradation or even divergence in control performance. Frequent manual readjustment of model parameters is required, resulting in extremely high maintenance costs. At the same time, the core prediction model of this type of MPC system is an open-loop static design. When the model fails due to the time-varying nature of the system, the operating basis of the entire optimized closed loop is destroyed, resulting in poor system robustness. The fourth category is intelligent control systems based on data-driven models. In recent years, some researchers have proposed using intelligent algorithms such as neural networks and support vector machines to construct NOx prediction models for ammonia injection rate feedforward compensation or auxiliary decision-making. However, this type of method does not essentially construct a complete and robust closed-loop control system. After the prediction model is trained, the parameters are fixed and it does not have the ability to track the dynamic changes of the system online. Long-term operating performance cannot be guaranteed. Moreover, the prediction model and the controller are loosely coupled, and the model output is only used as a feedforward quantity or reference quantity. The model performance evaluation and online evolution are not included as components of the closed-loop feedback, and the uncertainty of the model itself is not explicitly considered in the control decision. Therefore, it is impossible to achieve long-term autonomous optimization operation of the system.
[0004] In summary, existing SCR denitrification closed-loop control systems generally suffer from core technical defects such as poor adaptability, weak robustness, non-global optimality, and uncontrollable model degradation, making it difficult to achieve long-term, stable, and economical automatic operation of SCR denitrification systems. There is an urgent need to develop a new intelligent closed-loop optimization control scheme to address the many shortcomings of existing technologies. Summary of the Invention
[0005] This invention aims to completely overcome the core technical defects of existing SCR denitrification ammonia injection closed-loop control systems, such as poor adaptability, weak robustness, non-global optimum, and uncontrollable model degradation. It provides an intelligent closed-loop optimization control system and method for SCR denitrification based on adaptive robust prediction of key variables. It constructs an enhanced intelligent closed loop with self-sensing, self-optimization, and self-evolution capabilities, realizing long-term accurate, stable, and economical control of ammonia injection in the SCR denitrification system. It solves the technical problems of model mismatch, insufficient robustness, poor control quality, and high maintenance costs in existing technologies.
[0006] To achieve the above objectives, this invention provides an intelligent closed-loop optimization control system for SCR denitrification based on adaptive robust prediction of key variables, comprising:
[0007] The multi-source sensing and rapid response inner loop is used to quickly collect global operating condition process data in real time and execute second-level rapid ammonia injection adjustment;
[0008] In the adaptive prediction model, an evolvable digital twin is configured to output the predicted mean and uncertainty variance of NOx and ammonia escape concentrations at the outlet for multiple future time windows, and the model achieves online self-evolution based on the prediction performance.
[0009] The robust optimization decision outer loop is configured with an optimizer of stochastic model predictive control or robust model predictive control to integrate predictive uncertainties to construct a chance-constrained optimization problem and solve the globally optimal ammonia injection command.
[0010] The multi-objective dynamic optimization and instruction fusion module is used to dynamically fuse the inner loop fast control quantity and the outer loop optimization instruction based on the model confidence and measurement confidence, and output the final partitioned ammonia injection instruction and execute it stably.
[0011] Preferably, the multi-source sensing and rapid response inner loop includes a sensing layer and a rapid control execution layer; the sensing layer includes a global operating condition sensing unit, a rapid process sensing unit, and a key variable precise monitoring unit; the rapid control execution layer includes multiple zone intelligent controllers, with the number of zones corresponding to the number of zones in the ammonia injection grid.
[0012] Preferably, the rapid process sensing unit employs a NOx concentration field measurement array using ceramic semiconductor or ultraviolet differential absorption spectroscopy technology, arranged in a grid pattern on the cross-section of the flue before the inlet of the first catalyst layer in the SCR reactor.
[0013] Preferably, the partition intelligent controller adopts a feedforward-feedback composite control algorithm, is configured with signal anomaly detection and redundancy switching logic, and the inner loop operation cycle is shorter than the outer loop robust optimization decision cycle.
[0014] Preferably, the evolvable digital twin of the loop in the adaptive prediction model adopts a deep temporal probability model, which outputs the conditional probability distribution of key states in multiple future time windows, including the prediction mean and prediction variance.
[0015] Preferably, the adaptive prediction model includes a loop configuration performance evaluation module that triggers online model learning based on the statistical test results of the prediction error sequence, and updates the model using a catastrophic forgetting prevention strategy and SCR mechanism constraints.
[0016] Preferably, the cost function of the robust optimization decision outer loop includes a prediction uncertainty penalty term, the NOx and ammonia escape constraints are constructed as chance constraints, and the optimization problem includes actuator physical constraints and rate constraints.
[0017] Preferably, the multi-objective dynamic optimization and instruction fusion module calculates dynamic fusion weights based on model confidence and measurement confidence, and the final partitioning instruction is a weighted sum of the inner loop fast control quantity and the outer loop optimization instruction, and configures instruction safety limiting and degradation switching strategies.
[0018] Another aspect of the present invention provides an intelligent closed-loop optimization control method for SCR denitrification based on adaptive robust prediction of key variables, used to control the aforementioned intelligent closed-loop optimization control system for SCR denitrification based on adaptive robust prediction of key variables, comprising the following steps:
[0019] S1. System initialization and verification: Deploy hardware and verify sensors and actuators; train the initial model of the digital twin using historical data; set the inner loop action threshold, model evolution trigger conditions, optimization parameters, and initial values of fusion weights.
[0020] S2. Real-time synchronous acquisition of multi-source data for global operating conditions, rapid processes, and precise monitoring of key variables;
[0021] S3, Inner loop fast adjustment, each zone intelligent controller judges the deviation based on the fast process sensing signal, and outputs the zone fast control quantity in parallel when the threshold is exceeded, and the inner loop runs independently according to the set cycle;
[0022] S4. Model Prediction and Evolution: When the outer loop cycle is reached, input global data into the digital twin to obtain the predicted mean and variance. The performance evaluation module calculates the error index and starts online learning to update the model when the triggering conditions are met.
[0023] S5. Robust optimization decision-making: The robust optimization decision-maker uses the current system state as the initial value and the evolved model as the prediction model to solve the opportunity-constrained optimization problem with fused uncertainties and outputs the global optimization control quantity.
[0024] S6. Command Fusion and Secure Execution: The command fusion module calculates the model confidence and measurement confidence, determines the dynamic fusion weights to generate the final partition command, and issues it for execution after security limit verification.
[0025] S7. Wait for the next outer loop cycle, return to S2, continuously monitor system performance and adaptively adjust the control cycle, and execute the degradation switching process when a fault occurs.
[0026] Preferably, in step S3, the inner loop control cycle is shorter than the outer loop robust optimization decision cycle in step S5, and the inner loop performs ultra-short-time disturbance compensation fine-tuning on the outer loop command.
[0027] Preferably, in step S4, the online learning of the model adopts a catastrophic forgetting prevention strategy, introducing SCR mechanism soft constraints as regularization terms to prevent the model from learning associations that violate physical laws.
[0028] Preferably, the cost function of the optimization problem in step S5 includes an uncertainty penalty term, and the NOx and ammonia escape opportunity constraints are transformed into a deterministic equivalent form for solution.
[0029] Preferably, in step S6, the dynamic fusion weight is calculated based on the model confidence and the measurement confidence, and a downgrade switching strategy is executed when a fault occurs, whether it is a measurement fault or a model fault.
[0030] Compared with the prior art, the present invention has the following beneficial technical effects:
[0031] (1) By constructing a multi-objective dynamic optimization and instruction fusion module, the present invention uses the model prediction error as a closed-loop feedback signal to continuously fine-tune and update the prediction model online. It can automatically track slow time-varying factors of the system such as catalyst activity decay, coal type change, and instrument drift, so that the controller has self-healing and self-optimization capabilities. It fundamentally solves the technical pain points of traditional model predictive control caused by model mismatch, such as performance degradation and the need for frequent manual maintenance, greatly reduces system operation and maintenance costs, and improves the long-term operation stability of the system.
[0032] (2) Through robust decision outer loop design, this invention transforms the hard environmental constraints of NOx and ammonia escape into probabilistic constraints by considering the core intrinsics of the control algorithm and quantitatively taking into account the uncertainty of model prediction. This enables the control system to strictly comply with emission limits with a high probability when facing unknown disturbances, measurement noise, and model errors, greatly improving the system's operational reliability and automatic commissioning rate, effectively avoiding the risk of exceeding emission standards, and meeting stringent environmental regulatory requirements.
[0033] (3) This invention combines a fast-response inner loop (second-level response, suppressing uneven flue gas distribution and short-term concentration fluctuations) with a robust optimization outer loop (minute-level decision-making, pursuing global optimal operation), and adopts dynamic command fusion technology based on real-time reliability to achieve deep coupling between the two, forming a collaborative control mode of global optimization and local fine-tuning, achieving the best balance between rapid anti-disturbance capability and long-term economic optimization, effectively reducing the fluctuation range of NOx concentration at the outlet, reducing ammonia escape emissions, and improving the economic efficiency of system operation.
[0034] (4) The system of the present invention not only realizes the closed loop of ammonia injection control action, but also constructs a complete intelligent closed loop from operation data acquisition, high-fidelity probability model construction, robust optimization decision execution, decision effect feedback verification, and model knowledge iterative update. The system can autonomously accumulate operation experience and evolve control knowledge. It is the core technical support for realizing the long-term, autonomous and optimized operation of SCR denitrification system, and provides a replicable and scalable technical paradigm for intelligent control of industrial processes.
[0035] (5) This invention constructs a full-process safety protection system through multiple fault-tolerant designs such as signal anomaly discrimination and redundancy switching, model fault degradation, measurement fault degradation, and command safety limiting. Under abnormal conditions such as sensor failure, model failure, and optimization solution failure, it can smoothly switch to the backup control strategy to avoid control interruption or dangerous command output, greatly improve the reliability and safety of system operation, and reduce the risk of industrial site operation.
[0036] The following will further explain the concept, specific structure, and technical effects of the present invention in conjunction with the accompanying drawings, so as to fully understand the purpose, features, and effects of the present invention. Attached Figure Description
[0037] Figure 1 This is a schematic diagram of the four-ring nested enhanced intelligent closed-loop architecture of the system of the present invention;
[0038] Figure 2 This is a system flowchart of the control method of the present invention. Detailed Implementation
[0039] The technical solution of the present invention will be described in detail below with reference to the accompanying drawings and specific embodiments. The embodiments are only used to explain the present invention and are not intended to limit the scope of protection of the present invention.
[0040] This invention provides an intelligent closed-loop optimization control system for SCR denitrification based on adaptive robust prediction of key variables. It constructs an enhanced intelligent closed-loop architecture with four nested loops, including an inner loop of multi-source sensing and rapid response, a middle loop of adaptive prediction model, an outer loop of robust optimization decision, and a multi-objective dynamic optimization and command fusion module. The four core modules work together to form a multi-level, strongly coupled, and self-evolving intelligent closed-loop control system.
[0041] The multi-source sensing and rapid response inner loop is the core of the system's underlying execution and rapid feedback. It is responsible for real-time acquisition of multi-source data, signal anomaly processing, and second-level rapid ammonia injection adjustment. It effectively suppresses short-term concentration fluctuations caused by flue gas turbulence and makes up for the shortcomings of insufficient response speed of the outer loop optimization control. It consists of two parts: the sensing layer and the rapid control execution layer.
[0042] The sensing layer consists of a global operating condition sensing unit deployed on the boiler side, a rapid process sensing unit deployed downstream of the ammonia injection grid of the SCR reactor, and a key variable precision monitoring unit (advanced flue gas analyzer) at the reactor outlet. These three types of units work together to achieve comprehensive and multi-level sensing of the system's operating status. The global operating condition sensing unit includes sensors for boiler load signals, total air volume, coal consumption, economizer outlet flue gas temperature, and oxygen content, providing comprehensive environmental background data for system operation and providing basic operating condition information for model prediction and optimization decisions. The rapid process sensing unit, as the core rapid feedback device of this closed loop, employs a NOx concentration field measurement array using ceramic semiconductor or differential ultraviolet absorption spectroscopy (DOAS) technology. This array is installed at the flue section 1-2 meters in front of the ammonia injection grid of the SCR reactor, arranged in a grid pattern. The number of measuring points (M) is determined according to the flue size, typically 8-24, ensuring accurate capture of the NOx concentration distribution characteristics of the flue section. The rapid process sensing unit has a response time (T90) of less than 5 seconds and a measurement accuracy better than ±2% of full scale, significantly reducing the control feedback delay from minutes in traditional CEMS to seconds, fundamentally solving the problem of large lag in traditional control. The precision monitoring unit, or advanced flue gas analysis and measurement device, provides accurate and compliant measurements of outlet NOx concentration and ammonia slip concentration. It is mainly used for model validation, performance evaluation, and environmental emission compliance monitoring, ensuring the authority and accuracy of system operation data.
[0043] The rapid control execution layer consists of N independent zone intelligent controllers (typically N=4-12, strictly corresponding to the number of zones in the ammonia injection grid). Each controller corresponds to one zone of the ammonia injection grid and one or more rapid process sensing unit measurement points downstream of that zone. There is a clear mapping relationship between the number of measurement points M and the number of zones N. For example, when M=16 and N=8, the average of data from every two adjacent measurement points can be used as the feedback value for the corresponding zone, achieving precise matching between the concentration field distribution and zone control. Each zone controller receives the rapid process sensing signal corresponding to its zone and compares it with the set target NOx concentration value for that zone (usually a uniform distribution value of the total outlet target value or a distribution value determined based on historical flow field optimization), constructing a rapid inner loop of feedforward-feedback composite control; the inner loop control cycle ( The time interval is set to 1-5 seconds. An improved PI or predictive function control (PFC) algorithm is used, and the bandwidth is designed to effectively suppress concentration fluctuations with a period of less than 5 minutes caused by flue gas turbulence, thereby achieving rapid suppression of short-term disturbances.
[0044] The zone intelligent controller has built-in signal anomaly detection and redundancy switching logic. It continuously monitors the status of received rapid sensing signals. If the signal exceeds the range, changes abruptly beyond the set threshold (e.g., changes exceeding 20% of the full scale per second), or is seriously inconsistent with the signals of adjacent zones and system operating conditions, it is determined to be a signal anomaly. When a signal is abnormal, the system automatically switches to the backup control strategy, prioritizing the use of the average value of other normal measurement points in the same zone as the feedback signal. If there are no normal measurement points, it switches to the soft measurement value constructed based on upstream operating conditions (e.g., the flow and pressure of the ammonia injection branch pipe corresponding to this zone), or switches to the allocation value of the global command given by the robust optimization outer loop in this zone, ensuring that the inner loop control is not interrupted and improving the reliability of system operation.
[0045] The inner and outer loops adopt a time-sequential coordinated operation rule, with the inner loop (cycle) ) and outer-loop robust optimization decision loop (cycle) (Typically 30-60 seconds) It runs asynchronously and independently. At the beginning of each outer loop control cycle, the inner loop outputs its latest partition control quantity. The command is uploaded to the command fusion module. The outer loop optimization calculation is carried out based on the current global state of the system (including the effect of the inner loop fast adjustment). After the fused command is issued, the inner loop continues to run independently within its own fast cycle, and fine-tunes the outer loop command to compensate for ultra-short time disturbances, forming a collaborative control mode of "outer loop sets the tone, inner loop refines", which takes into account the advantages of global optimization and local fast adjustment.
[0046] The adaptive prediction model loop is the core of the system's modeling and self-evolution. It is responsible for modeling the system's dynamic characteristics, predicting the probability of future states, evaluating model performance, and updating the model online. It constructs an evolvable digital twin with the ability to track the system's time-varying characteristics in real time, fundamentally solving the model mismatch problem of traditional model predictive control. The core component is an evolvable digital twin based on a deep time-series probabilistic model, and it is also equipped with a performance evaluation module and an online learning module.
[0047] Evolvable digital twins, taking historical system operation data and real-time multi-source data (global operating condition data, rapid process data, and precise outlet monitoring data) as input, can not only predict the outlet NOx concentration and ammonia escape concentration in multiple future time windows, but more importantly, can output the probability distribution characteristics of the prediction results, providing a quantitative basis for uncertainty in robust optimization decisions. The preferred deep temporal probabilistic model employs a deep Gaussian process (Deep GP) or a probabilistic recurrent neural network (such as Bayesian LSTM / GRU) that combines attention mechanisms (e.g., Transformer) or temporal convolutional networks (TCN). The model output is the conditional probability distribution of the key states in the next H steps. In practical applications, it is usually parameterized into a Gaussian distribution, that is, the output is the predicted mean of the next H steps. , …, The prediction variance (uncertainty) Σ{t+1}, …, Σ{t+H} is used to achieve synchronous output of prediction results and uncertainty.
[0048] Model performance is comprehensively evaluated using three core metrics: root mean square error of prediction (RMSE), mean absolute percentage error (MAPE), and prediction interval coverage (PICP). PICP is defined as the proportion of actual measurements falling within a specific confidence interval (e.g., 95%) of the model output. It is used to accurately quantify the degree of calibration of uncertainty estimation and provide a quantitative basis for triggering model evolution.
[0049] The self-evolution mechanism is one of the core innovations of this invention. Through three key designs—online learning trigger threshold and statistical testing methods, catastrophic forgetting prevention strategies, and constraints integrating the SCR mechanism—the model is safely, effectively, and continuously updated online. The system has a built-in continuously running performance evaluation module that maintains the model prediction error sequence within a sliding time window (e.g., the past 24 hours). When the error sequence meets any of the following conditions, the online learning process is triggered: First, the mean or standard deviation of the error sequence exceeds a preset threshold (e.g., RMSE > 5 mg / Nm³) for multiple consecutive periods (e.g., 5 outer loop periods); Second, the error sequence is tested using a CUSUM (cumulative sum) control chart or EWMA (exponentially weighted moving average) control chart, revealing statistically significant drift (exceeding control limits); Third, the prediction interval coverage (PICP) remains consistently below the target level (e.g., 90%), ensuring that the model evolution trigger conditions are scientific, reasonable, and quantifiable.
[0050] The online learning process employs a catastrophic forgetting prevention strategy, prioritizing Elastic Weight Consolidation (EWC) or online Bayesian update strategies for model fine-tuning. The EWC strategy calculates the Fisher information matrix of model parameters on historical data and adds a penalty term for important parameters to the loss function. This protects parameters that are already learned and are important for the process dynamics from being significantly modified when adjusting model parameters using new data, effectively preventing the model from forgetting historical operating condition knowledge when learning new operating condition features and ensuring the predictive stability of the model across the entire operating condition range.
[0051] During the online learning process, constraints incorporating the SCR reaction mechanism are introduced as regularization terms to prevent the model from learning false correlations that violate physical laws, thereby improving the model's physical rationality and generalization ability. For example, constraints are introduced to define the linear relationship between the predicted ammonia slip concentration and the ammonia injection rate and NOx removal rate, or to constrain the trend of slow decay of catalyst activity with operating time under the same operating conditions. This ensures that the model's prediction results always conform to the physical mechanism characteristics of the SCR denitrification process, avoiding unreasonable prediction results.
[0052] The robust optimization decision outer loop is the core of the system's global optimization and robust control. It is responsible for calculating the globally optimal ammonia injection strategy, quantifying uncertainty, constructing robust constraints, and solving optimization problems. The core of the control algorithm intrinsically considers the uncertainty of model prediction, ensuring that the system can still meet environmental emission constraints with a high probability even under the presence of unknown disturbances, measurement noise, and model errors. The core component is an optimization controller built based on the theory of stochastic model predictive control (Stochastic MPC) or robust model predictive control (RMPC).
[0053] Optimize the controller in each control cycle Internally, starting from the current global state of the system, and using an evolvable digital twin as an internal probabilistic prediction model, the system performs rolling optimization calculations on the ammonia injection strategy for a period of time in the future, and outputs the globally optimal ammonia injection control command, taking into account the three major optimization objectives of denitrification efficiency, ammonia slip control, and operating cost.
[0054] The design of costs and constraints that integrate uncertainty is achieved through three key designs: a cost function with uncertainty penalty terms, robust constraint expressions for NOx and ammonia escape at the outlet, and actuator constraints. This allows for the explicit and quantitative consideration of predictive uncertainty in optimization decision-making, thereby fundamentally improving the robustness of the system.
[0055] The cost function J for the optimization problem is designed as follows:
[0056]
[0057] in, For prediction step size (prediction time domain length); For time k to time k+i Concentration prediction; for Setting value; The predicted value of ammonia escape at time k+i is given by time k. for Predict the covariance matrix; The control quantity (ammonia injection quantity) at time k+i-1 is planned for time k; Q, R, and S are weight matrices; The term represents a penalty term for the uncertainty of the prediction, where γ is the uncertainty penalty coefficient. This represents the total variance of the predicted state. Minimizing this value drives the controller to prioritize models that predict more certain future operating states, effectively enhancing system robustness.
[0058] The NOx and ammonia escape constraints are constructed as opportunity constraints, transforming hard environmental constraints into probabilistic constraints, ensuring that the system can still meet emission requirements with a high probability even when uncertainties exist. The concentration constraint expression is:
[0059]
[0060] in The confidence level is typically set to 95% or 99%; based on the Gaussian distribution assumption of the twin output, this constraint can be transformed into a deterministic equivalent form:
[0061]
[0062] in The inverse cumulative distribution function of the standard normal distribution is in The value at; for Emission limits.
[0063] The expression for the ammonia slip concentration constraint is:
[0064]
[0065] The conversion method is consistent with the NOx concentration constraint. The ammonia slip limit is set to ensure that the ammonia slip concentration is strictly controlled within the compliant range.
[0066] The optimization problem also includes the physical constraints of the ammonia injection control valve, ensuring that control commands are within the reasonable operating range of the actuator and avoiding invalid command output. Actuator constraints include:
[0067] Valve opening limit constraint:
[0068]
[0069] in, Let k+i be the opening degree of the ammonia injection valve; This is the lower limit of the valve opening range; This is the upper limit of the valve opening.
[0070] Valve speed constraint:
[0071]
[0072] in, The valve opening increment at time k+i; Upper limit of valve opening increment;
[0073] Dead zone constraint:
[0074] To avoid frequent small-amplitude valve movements, a setting is provided. ( The valve opening is only allowed to change when a lower limit is set; otherwise, ΔU=0, which reduces actuator wear and extends equipment service life.
[0075] For optimization problems involving quadratic cost functions and linearization constraints, the effective set method, interior point method, or scenario tree method specifically designed for stochastic MPC can be used for solving; control period The setting of 30-60 seconds should ensure that the calculation time of the optimization solution is less than 50% of the cycle time (e.g., <30 seconds) to ensure the real-time nature of the optimization decision. In industrial DCS / PLC systems, efficient QP solvers such as qpOASES and OSQP can be used to complete the optimization solution online, ensuring computational efficiency and stability.
[0076] The system is configured with online adjustment rules for multi-objective weights. The weight coefficients Q, R, S, and γ are not fixed and can be dynamically fine-tuned according to the system's operating mode to improve control adaptability under different operating conditions. For example, under the condition of rapid rise and fall of boiler load, γ (emphasizing system robustness) and R (strictly suppressing ammonia slip) can be appropriately increased. Under the condition of stable load and high model confidence, Q can be appropriately increased (emphasizing NOx concentration tracking accuracy) to pursue system operating economy. The weight adjustment rules can be implemented based on fuzzy logic or simple operating condition lookup table method to flexibly adapt to the needs of different operating scenarios.
[0077] The multi-objective dynamic optimization and instruction fusion module is the core of the system's instruction coordination. It undertakes the dynamic fusion of the inner loop rapid control quantity and the outer loop optimization instruction, instruction safety verification, fault degradation switching, and timing synchronization functions. It achieves the complementary advantages of rapid response and global optimization, ensuring that the system operates efficiently and collaboratively under normal operating conditions and transitions safely and stably under fault conditions.
[0078] The module employs a dynamic weighting formula and confidence calculation logic, and the instruction fusion module receives the fast control input from the inner loop. (Dimension N×1, where N is the number of partitions) and the globally optimized total ammonia injection output from the outer ring. First of all The instruction vectors for partition optimization are allocated based on the current flue gas flow rate or historical contribution ratio of each partition. Final ammonia injection command for each zone The calculation formula is:
[0079]
[0080] Among them, the fusion weight ∈[0,1], determined through dynamic calculation.
[0081] Fusion weights Based on model confidence With measurement confidence level Calculate the model confidence level. , The mean variance of recent model predictions reflects the reliability of model predictions; the confidence level is measured. M represents the total number of measurement points in the fast process sensing unit, reflecting the reliability of the fast measurement signal; fusion weights The calculation formula is:
[0082]
[0083] parameter Used to adjust the relative importance of model confidence and measurement confidence; when model uncertainty is high ( Hour, As the number of rapid measurement signal anomalies increases, the system becomes more reliant on fast inner-loop control; when the number of rapid measurement signal anomalies is high ( Hour, The system relies more on global optimization of the outer loop control to achieve complementary advantages based on real-time reliability, thereby significantly improving the reliability of system operation.
[0084] Final partitioning instructions Before being sent to the actuator, the command must undergo strict safety limit verification to avoid outputting dangerous or invalid commands: First, the commands for each zone must not exceed the opening range of the corresponding control valve. Second, the rate of change of instructions in each partition must not exceed the rate constraint. Third, the sum of instructions for each zone (total ammonia injection) must be within the safe range estimated based on the current flue gas volume and inlet NOx concentration to prevent excessive ammonia injection from causing ammonia escape and exceeding the standard, and to ensure the safe operation of the system.
[0085] The module is configured with a comprehensive degradation and switching strategy to handle abnormal operating conditions such as measurement failures, model failures, and optimization solution failures, ensuring the system's safe and stable operation under fault conditions and preventing control failures. Measurement fault switching strategy: If all fast sensing units fail ( =0), system switches to =0 mode: The robust optimization outer loop operates independently, relying on global operating condition perception data and outlet CEMS data for prediction and optimization decisions. Model failure / optimization failure switching strategy: If the digital twin's prediction error continues to exceed the limit and online learning fails to converge, or if the robust optimization problem fails to be solved several times in a row, the system automatically switches to a rule-based backup controller (such as a gain-scheduled PID controller with feedback from the outlet flue gas analyzer) and issues an equipment maintenance alarm to prompt maintenance personnel to troubleshoot the fault in a timely manner.
[0086] The system uses high-speed industrial Ethernet (such as EtherCAT or Profinet IRT) to communicate between its various units, ensuring the deterministic timing of data transmission. At the same time, a global clock synchronization protocol (such as IEEE1588PTP) is used to align the timestamps of multi-source sensing data, ensuring that the inner loop, model loop, and decision loop perform calculations based on the same system "snapshot" at the same moment. This provides a timing basis for the precise collaborative operation of the multi-layer closed loop and avoids control deviations caused by misaligned data timing.
[0087] This invention provides a control method based on the aforementioned four-ring nested intelligent closed-loop control system to achieve multi-level, full-process collaborative operation, specifically including the following steps:
[0088] S1: System Initialization and Verification. Complete the deployment of hardware equipment, connection of sensors and actuators, and accuracy verification; train the initial digital twin model using no less than 3 months of historical operating data, and verify the basic accuracy of the model (RMSE < 10%FS) and the coverage of the prediction interval (PICP > 85%); set key operating parameters such as inner loop action thresholds (e.g., start adjustment when the deviation of the NOx concentration in the partition exceeds ±5 mg / Nm³ from the target value), model evolution trigger conditions (e.g., CUSUM control limit, PICP threshold), optimization problem parameters, initial values of fusion weights, and actuator constraint parameters, and complete the system initialization configuration.
[0089] S2: Real-time synchronous acquisition of multi-source data. Multi-source data, including boiler load, total air volume, flue gas temperature, oxygen content, zone NOx concentration, outlet NOx concentration, ammonia slip concentration, ammonia injection branch flow rate, and valve opening, are synchronously acquired at a set sampling frequency from the global operating condition sensing unit, rapid process sensing unit, and outlet precise monitoring unit. Data timestamps are aligned via a global clock synchronization protocol to ensure data synchronization and accuracy.
[0090] S3: Inner Loop Rapid Adjustment. Each zone's intelligent controller calculates the deviation between the zone's NOx concentration and the set target value based on the real-time signal output by the corresponding rapid process sensing unit, and determines whether the deviation exceeds the inner loop action threshold. If the deviation exceeds the threshold, each zone controller calculates and outputs the zone's rapid control quantity in parallel. This drives the ammonia injection regulating valve; if the deviation does not exceed the threshold, the output of the previous cycle remains unchanged; the inner loop operates at the set cycle. (1-5 seconds) Independent cycle operation, continuously suppressing short-term concentration fluctuations.
[0091] S4: Model Prediction and Evolution. When the outer loop control cycle... When the time is reached (30-60 seconds), the processed global multi-source data is input into the evolvable digital twin, and the model outputs the predicted average values of the outlet NOx concentration and ammonia slip concentration for the next H steps. With prediction variance Simultaneously, the performance evaluation module retrieves the model's predicted values and the measured values of the exported CEMS within the past sliding time window, calculates error indicators such as RMSE, MAPE, and PICP, and performs CUSUM or EWMA statistical tests. If the model evolution triggering conditions are met, the system starts the online learning process, updates the model parameters under the constraints of the EWC anti-catastrophic forgetting strategy and the SCR mechanism, and completes the model self-evolution. If the triggering conditions are not met, the current model parameters remain unchanged.
[0092] S5: Robust Optimization Decision. The robust optimization decision-maker takes the current global system state (including the effects of rapid adjustment within the inner loop) as the initial state and uses the evolved and updated digital twin as the internal probabilistic prediction model to construct a chance-constrained optimization problem that integrates prediction uncertainties. An efficient QP solver is used to solve the optimization problem, and convergence is checked during the solution process. Termination conditions are set when the decrease in cost function is less than the set tolerance or the maximum number of iterations is reached. If the solution converges, the globally optimal total ammonia injection amount is output. If the solution fails to converge, record the fault information and use the optimal solution from the previous cycle or activate the backup controller to ensure the continuity of system control.
[0093] S6: Instruction Fusion and Secure Execution. The instruction fusion module retrieves the mean variance of the current model prediction and the number of normal measurement points to calculate the model confidence level. With measurement confidence level Substitute into the dynamic weighting formula to calculate the fusion weight. The total global optimized ammonia injection output from the outer ring will be used to optimize the overall ammonia injection volume. Distribute partition optimization command vectors according to partition traffic ratio. Combined with the inner loop rapid control quantity Calculate the final partitioning instructions ;right Perform safety limit verification to ensure that the command meets the safety constraints of opening degree, rate, and total ammonia quantity; after the verification is passed, the final zoning command is issued to the ammonia injection regulating valve of each zone for execution; if the verification fails, the command is corrected and then issued again.
[0094] S7: Cyclic Operation and Performance Monitoring. Waiting for the next outer loop control cycle to arrive, return to step S2 to begin the next data acquisition and control cycle. The system continuously monitors overall operational performance and can adaptively adjust the outer loop control cycle within a preset range based on boiler operating stability and model prediction accuracy. (Shorten the cycle when the load changes drastically); when a serious fault occurs in any part of the system, switch down the control level step by step according to the predefined fault rollback process (e.g., first switch to pure outer loop optimization control, and then switch to backup conventional PID control) to ensure the safe operation of the system.
[0095] The specific content of the present invention is illustrated below through specific embodiments:
[0096] Example 1
[0097] like Figure 1 As shown in the figure, this embodiment provides an intelligent closed-loop optimization control system for SCR denitrification applied to a 660MW coal-fired power generating unit, with the following specific configuration:
[0098] Multi-source sensing and rapid response inner loop: At the inlet flue of the SCR reactor A side, 1.5 meters downstream of the ammonia injection grid, a UV DOAS concentration field measurement array consisting of 16 measuring points (4×4 grid) is installed. The array response time is <5s, and the measurement accuracy is ±1.5%FS. The ammonia injection grid is divided into 8 independent zones. The rapid control execution layer is configured with 8 zone intelligent controllers, with one controller corresponding to one zone and two adjacent array measuring points. The average value of the two measuring points is taken as the zone feedback signal. The inner loop control cycle is... =2 seconds, the zone intelligent controller adopts the PFC algorithm, and the bandwidth is adjusted to suppress flue gas concentration fluctuations with a cycle of <40 seconds; the signal anomaly discrimination threshold is set as follows: single-point signal loss, instantaneous change rate >15%FS / s, deviation from the average value of adjacent zones >20%. When an anomaly occurs, it will first switch to the average value of normal measurement points in the same zone, and switch to the soft measurement value when there are no normal measurement points.
[0099] In the adaptive prediction model, the evolvable digital twin is constructed using a Bayesian neural network (BNN) based on an attention mechanism. The model input features include boiler load, total air volume, flue gas temperature, oxygen content, ammonia injection rates for each zone, estimated inlet NOx, and average NOx concentration for each zone measured by the array over the past 30 minutes. The model output is for the next 10 minutes (prediction time domain H=20, outer loop control cycle). =30 seconds) of predicted mean and variance (log-standard deviation) of outlet NOx concentration and ammonia slip concentration; the online learning trigger condition for the performance evaluation module is set as follows: the EWMA control chart of the prediction error in the past 2 hours exceeds the 3σ control limit, and PICP remains below 88% for 1 hour; the online learning uses the EWC strategy to fine-tune the model, and the importance damping coefficient is... The loss function incorporates a soft-constraint regularization term based on nitrogen molar balance, with the penalty term expressed as follows:
[0100]
[0101] in: These are weighting coefficients to prevent the model from outputting results that violate physical laws.
[0102] Robust optimization decision outer loop: A stochastic MPC framework is adopted, with prediction time domain H=20 (10 minutes) and control time domain Hu=10; the cost function weights are set as Q=diag([1.0]), R=diag([0.5]), S=diag([0.1]), and the uncertainty penalty coefficient γ=0.3; the confidence level of NOx and ammonia escape chance constraints η=95%; the actuator constraints are set as follows: the total ammonia injection rate ranges from 0 to 200 kg / h, the opening degree of the zone valve is 0 to 100%, and the rate is limited to ±5% / s; the optimization problem is transformed into a quadratic programming (QP) problem, which is solved online using the qpOASES solver, with an average computation time of <10 seconds, meeting the real-time requirement of a 30-second outer loop cycle.
[0103] Multi-objective dynamic optimization and instruction fusion module: Model confidence Measure confidence level =Number of effective measurement points / 16; fusion weight calculation parameter β=0.7; command safety limits include: zonal valve opening degree 0-100%, rate ±5% / s, and total ammonia injection quantity dynamically determined based on the current load and inlet NOx estimation value; the system adopts the IEEE1588PTP protocol to achieve sub-microsecond clock synchronization, ensuring the consistency of multi-source data timing.
[0104] Example 2
[0105] like Figure 2 As shown, this embodiment provides a control method based on the system of Embodiment 1, and the specific operation flow is as follows:
[0106] S1: System Initialization and Verification. After power-on, the system completes hardware self-test, sensor and actuator communication test, and accuracy calibration; retrieves historical DCS operation data from the past 3 months, trains the initial BNN model, and verifies that the model's NOx prediction RMSE on the test set is 3.2 mg / Nm³, and the PICP reaches 92%, meeting the accuracy requirements; sets the inner loop action threshold to a NOx concentration deviation of ±4 mg / Nm³ in the partition, and the model evolution trigger threshold to an EWMA error statistic > 2.7 (corresponding to the 95% control limit); optimizes the parameters, initializes the fusion weights, and initializes the actuator constraint parameters according to Example 1.
[0107] S2: Real-time synchronous acquisition of multi-source data. The system synchronously reads multi-source data from 16 array measuring points at a frequency of 1Hz, including NOx concentration, boiler load, flue gas temperature, oxygen content, outlet CEMS data, ammonia injection branch flow rate in each zone, and valve opening. Data timestamps are aligned via the IEEE 1588PTP protocol to ensure data synchronization.
[0108] S3: Inner loop rapid adjustment. Eight zone intelligent controllers operate independently in parallel, with a control cycle of 2 seconds. Taking zone #1 as an example, the controller reads the average value of the corresponding two array measurement points as 45 mg / Nm³, and the zone NOx setpoint is 40 mg / Nm³. The deviation of 5 mg / Nm³ exceeds the 4 mg / Nm³ action threshold. The controller calculates and outputs a rapid control quantity. The corresponding ammonia injection valve opening is finely adjusted by -1.5%; other zones are adjusted independently according to the same logic, and the valve opening remains unchanged when the deviation does not exceed the threshold.
[0109] S4: Model Prediction and Evolution. When the outer loop control cycle reaches 30 seconds, the system retrieves global multi-source data from the past 30 minutes and inputs it into the BNN twin. The model outputs the average predicted NOx concentration at the outlet for the next 10 minutes as 38±2.5 mg / Nm³ (95% confidence interval), and the average predicted ammonia escape concentration as 2.1±0.8 mg / Nm³. The performance evaluation module calculates the prediction error EWMA value for the past 2 hours as 2.8, exceeding the trigger threshold of 2.7, thus satisfying the model evolution condition. The system locks the current BNN parameters, retrieves the latest 1-hour running data, performs model fine-tuning for 5 epochs under EWC constraints, updates the model parameters, and completes self-evolution.
[0110] S5: Robust Optimization Decision. The system uses the current system state (including the effects of inner-loop regulation) as initial values, inputs the updated BNN model for prediction, and constructs a QP optimization problem with chance constraints. The optimization objective is set as follows: minimizing ammonia injection consumption while satisfying Prob(NOx≤45mg / Nm³)≥95% and Prob(ammonia slip≤3ppm)≥95%. The qpOASES solver converges to obtain the globally optimal total ammonia injection amount for the next 30 seconds. =152kg / h.
[0111] S6: Instruction fusion and secure execution. The system calculates the current prediction variance mean as 6.25 (mg / Nm³)², and the model confidence level... =1 / (1+6.25)≈0.14; All 16 array measurement points are normal, and the measurement confidence level C_s=1.0; Substituting into the formula, the fusion weight α=0.7*(1-0.14)+0.3*1.0≈0.9; The flue gas flow rate of each zone is allocated into an 8-dimensional partition optimization instruction vector U_opt; the final partition instructions are calculated. Verify that the opening degree, rate, and total ammonia injection volume of each zone are within the safe range, and then issue instructions to the ammonia injection regulating valves of each zone for execution.
[0112] S7: Cyclic Operation and Performance Monitoring. The system returns to step S2 and enters the next 30-second outer loop cycle. After the model learns online, the monitoring system displays in real time that the EWMA error statistic has dropped to 2.1, indicating stable control performance. When a rapid increase or decrease in boiler load is detected, the system automatically adjusts the outer loop cycle to 20 seconds and adjusts the uncertainty penalty coefficient γ in the cost function from 0.3 to 0.5 to enhance system robustness. In the event of a failure in any component, the system switches the control mode according to the degradation strategy to ensure safe operation.
[0113] The preferred embodiments of the present invention have been described in detail above. It should be understood that those skilled in the art can make numerous modifications and variations based on the concept of the present invention without creative effort. Therefore, all technical solutions that can be obtained by those skilled in the art based on the concept of the present invention through logical analysis, reasoning, or limited experimentation on the basis of existing technology should be within the scope of protection defined by the claims.
Claims
1. A smart closed-loop optimization control system for SCR denitrification based on adaptive robust prediction of key variables, characterized in that, include: The multi-source sensing and rapid response inner loop is used to quickly collect global operating condition process data in real time and execute second-level rapid ammonia injection adjustment; In the adaptive prediction model, an evolvable digital twin is configured to output the predicted mean and uncertainty variance of NOx and ammonia escape concentrations at the outlet for multiple future time windows, and the model achieves online self-evolution based on the prediction performance. The robust optimization decision outer loop is configured with an optimizer of stochastic model predictive control or robust model predictive control to integrate predictive uncertainties to construct a chance-constrained optimization problem and solve the globally optimal ammonia injection command. The multi-objective dynamic optimization and instruction fusion module is used to dynamically fuse the inner loop fast control quantity and the outer loop optimization instruction based on the model confidence and measurement confidence, and output the final partitioned ammonia injection instruction and execute it stably.
2. The SCR denitrification intelligent closed-loop optimization control system based on adaptive robust prediction of key variables as described in claim 1, characterized in that, The multi-source sensing and rapid response inner loop includes a sensing layer and a rapid control execution layer; the sensing layer includes a global operating condition sensing unit, a rapid process sensing unit, and a key variable precise monitoring unit; the rapid control execution layer includes multiple zone intelligent controllers, with the number of zones corresponding to the number of zones in the ammonia injection grid.
3. The intelligent closed-loop optimization control system for SCR denitrification based on adaptive robust prediction of key variables as described in claim 2, characterized in that, The rapid process sensing unit employs a NOx concentration field measurement array using ceramic semiconductor or ultraviolet differential absorption spectroscopy technology, arranged in a grid pattern on the cross-section of the flue before the inlet of the first catalyst layer in the SCR reactor.
4. The SCR denitrification intelligent closed-loop optimization control system based on adaptive robust prediction of key variables as described in claim 2, characterized in that, The partition intelligent controller adopts a feedforward-feedback composite control algorithm, and is configured with signal anomaly detection and redundancy switching logic. The inner loop operation cycle is shorter than the outer loop robust optimization decision cycle.
5. The intelligent closed-loop optimization control system for SCR denitrification based on adaptive robust prediction of key variables as described in claim 1, characterized in that, The adaptive prediction model uses a deep temporal probability model to generate an evolutionary digital twin of the loop, which outputs the conditional probability distribution of key states in multiple future time windows, including the prediction mean and prediction variance.
6. The intelligent closed-loop optimization control system for SCR denitrification based on adaptive robust prediction of key variables as described in claim 5, characterized in that, The adaptive prediction model includes a loop configuration performance evaluation module that triggers online model learning based on the statistical test results of the prediction error sequence, and updates the model using a catastrophic forgetting prevention strategy and SCR mechanism constraints.
7. The intelligent closed-loop optimization control system for SCR denitrification based on adaptive robust prediction of key variables as described in claim 1, characterized in that, The cost function of the robust optimization decision outer loop includes a prediction uncertainty penalty term, NOx and ammonia escape constraints are constructed as chance constraints, and the optimization problem includes actuator physical constraints and rate constraints.
8. The intelligent closed-loop optimization control system for SCR denitrification based on adaptive robust prediction of key variables as described in claim 1, characterized in that, The multi-objective dynamic optimization and instruction fusion module calculates dynamic fusion weights based on model confidence and measurement confidence. The final partition instruction is a weighted sum of the inner loop fast control quantity and the outer loop optimization instruction, and the configuration instruction safety limiting and degradation switching strategy are configured.
9. A method for intelligent closed-loop optimization control of SCR denitrification based on adaptive robust prediction of key variables, used to control the intelligent closed-loop optimization control system for SCR denitrification based on adaptive robust prediction of key variables as described in any one of claims 1-8, characterized in that, Includes the following steps: S1. System initialization and verification: Deploy hardware and verify sensors and actuators; train the initial model of the digital twin using historical data; set the inner loop action threshold, model evolution trigger conditions, optimization parameters, and initial values of fusion weights. S2. Real-time synchronous acquisition of multi-source data for global operating conditions, rapid processes, and precise monitoring of key variables; S3, Inner loop fast adjustment, each zone intelligent controller judges the deviation based on the fast process sensing signal, and outputs the zone fast control quantity in parallel when the threshold is exceeded, and the inner loop runs independently according to the set cycle; S4. Model Prediction and Evolution: When the outer loop cycle is reached, input global data into the digital twin to obtain the predicted mean and variance. The performance evaluation module calculates the error index and starts online learning to update the model when the triggering conditions are met. S5. Robust optimization decision-making: The robust optimization decision-maker uses the current system state as the initial value and the evolved model as the prediction model to solve the opportunity-constrained optimization problem with fused uncertainties and outputs the global optimization control quantity. S6. Command Fusion and Secure Execution: The command fusion module calculates the model confidence and measurement confidence, determines the dynamic fusion weights to generate the final partition command, and issues it for execution after security limit verification. S7. Wait for the next outer loop cycle, return to S2, continuously monitor system performance and adaptively adjust the control cycle, and execute the degradation switching process when a fault occurs.
10. The method as described in claim 9, characterized in that, In step S3, the inner loop control cycle is shorter than the outer loop robust optimization decision cycle in step S5. The inner loop performs ultra-short-time disturbance compensation fine-tuning on the outer loop command.
11. The method as described in claim 9, characterized in that, In step S4, the online learning of the model adopts a catastrophic forgetting prevention strategy, introducing SCR mechanism soft constraints as regularization terms to prevent the model from learning associations that violate physical laws.
12. The method as described in claim 9, characterized in that, In step S5, the cost function of the optimization problem includes an uncertainty penalty term, and the NOx and ammonia escape opportunity constraints are transformed into a deterministic equivalent form for solution.
13. The method as described in claim 9, characterized in that, In step S6, the dynamic fusion weights are calculated based on model confidence and measurement confidence. In case of a failure, a degradation switching strategy is executed for either measurement failure or model failure.
