Method for fine regulation of the temperature field inside a building

Abstract

The application discloses a kind of building internal temperature field fine regulation and control methods, comprising: the microzone sensing data of access are aligned and grid mapping;Equivalent mixed thickness, front shape index and other observable substitution quantity are estimated by using aligned fusion data, and the state of boundary layer is estimated, and combined with physical envelope prediction and nominal data prediction, a robust prediction state set is established;Combining the dew point margin of prediction state set and the corrected near-wall humidity, the control hard constraint set and dew point gray cost are calibrated;According to the prediction state set, under the limitation of constraint and cost, the event-driven control flow is synthesized by adopting tubular model predictive control and combining control barrier function safety filtering, according to the prediction state set, under the limitation of constraint and cost, the event-driven control flow is synthesized by adopting tubular model predictive control and combining control barrier function safety filtering.The present application can effectively suppress transient disturbance, and prospectively and robustly avoid near-wall condensation.

Description

Technical Field

[0001] This invention belongs to the field of building environment and automatic control, specifically a method for precise control of the internal temperature field of buildings. Background Technology

[0002] In fields such as semiconductor manufacturing, biopharmaceuticals, data centers, precision optical experiments, and high-end medical applications, extremely stringent requirements are placed on the uniformity and stability of the temperature field within the building's internal microenvironment. However, in actual operation, unavoidable opening events such as personnel entry and exit, and material transfer (e.g., the opening and closing of cleanroom doors) can instantly disrupt the original thermal and moisture balance, introducing strong convection and heat-moisture intrusion. These short-duration, high-gradient transient disturbances, occurring within seconds, not only impact delicate process windows but can also easily cause the near-wall surface temperature of critical equipment or products (such as silicon wafers and cold server racks) to plummet below the local dew point, triggering catastrophic condensation risks. Therefore, researching how to achieve precise temperature field control and proactively avoid condensation under such strong disturbances from opening events is of significant engineering value for ensuring high-end manufacturing yields and the safe operation of precision equipment.

[0003] Currently, temperature control in such building environments primarily relies on traditional heating, ventilation, and air conditioning (HVAC) systems, such as adjusting the heating and cooling coils of air handling units (AHUs) or fan coil units (FCUs). These systems are typically based on temperature and humidity measurements within the building's interior (e.g., the main return air vent) and employ proportional-integral-derivative (PID) control logic, with response times often on the order of minutes. To suppress direct infiltration through openings, air curtains (air curtain machines) are often used for physical isolation, typically operating at constant airflow speeds and lacking adaptability. During the design phase, researchers frequently utilize computational fluid dynamics (CFD) techniques to perform offline simulation analysis of airflow organization in opening areas to optimize airflow layout and design parameters.

[0004] Existing technologies for addressing such transient, near-wall control challenges primarily suffer from the unobservability of critical states, the lag in control response, and the miscalculation of safety risks. Specifically: First, the physical core of condensation—the near-wall boundary layer—has dynamic states (such as millimeter-level thickness and leading edge morphology) that are virtually impossible to measure directly and reliably in strong transient turbulence due to limitations in sensor cost, size, and interference. Furthermore, the bulk parameters relied upon by traditional control are physically decoupled from the near-wall condensation risk, leading to misaligned control. Second, PID control based on bulk measurements is inherently reactive. By the time the near-wall sensor finally detects a temperature exceedance, a phase transition may have already occurred, resulting in control lag. Conventional predictive models struggle to capture this highly nonlinear, convection-driven physical process, failing to provide a reliable, forward-looking safety envelope for control. Finally, and most critically, existing methods severely underestimate the true risk. They generally use readily available body humidity to calculate dew point, completely ignoring near-wall water vapor buildup caused by the invading front (i.e., near-wall wet film effect). This miscalculation leads to a false safety margin, causing the control system to unintentionally push the system toward the actual condensation boundary when optimizing energy consumption or comfort. Summary of the Invention

[0005] The purpose of this invention is to provide a method for finely controlling the internal temperature field of a building, addressing the problems of transient temperature disturbances and near-wall condensation risks caused by opening events.

[0006] Technical solution: A method for fine control of the internal temperature field of a building, comprising:

[0007] Align and mesh map the accessed micro-area sensing data to obtain aligned and fused data;

[0008] By using aligned and fused data, the boundary layer state is estimated and robust prediction is performed to establish a predicted state set;

[0009] By combining the predicted state set with the aligned fusion data, the dew point margin is evaluated, thereby calibrating the control hard constraint set and the dew point grayscale cost.

[0010] Based on the predicted state set, and under the constraints of the control hard constraint set and the dew point grayscale cost, event-driven control is synthesized to generate the execution command flow.

[0011] Preferably, dimensionless acceptance and online learning are implemented for the execution command stream and aligned fused data.

[0012] Optionally, estimating the boundary layer state and performing robust prediction includes:

[0013] The boundary layer temperature difference field was reconstructed from the aligned and fused data, and the frontal zone mask was identified.

[0014] Observable substitutions are calculated on the normal measurement point series using the boundary layer temperature difference field and the frontal zone mask;

[0015] The observable substitutes include the equivalent mixing thickness, the front shape index, and the invading front velocity, and are used to build the predictive state set.

[0016] Optionally, the observable substitutes are calculated, including:

[0017] Calculate the dimensionless temperature profile of the normal measuring point column and obtain the slope at two points;

[0018] The equivalent mixing thickness is determined as the reciprocal of the slopes at two points;

[0019] Calculate the temperature step response energy ratio and arrival time difference;

[0020] The forward shape index is defined as the ratio of the temperature difference step response energy ratio to the arrival time difference.

[0021] Optionally, assess dew point margin, including:

[0022] Get the forward with a mask;

[0023] Within the near-wall region defined by the forward band mask, the humidity obtained from the aligned fusion data is corrected by using the worst-case humidity gradient assumption or mass transfer correlation to obtain the corrected local humidity.

[0024] The local dew point temperature is calculated using the corrected local humidity and boundary layer temperature difference field.

[0025] Local dew point temperature is used to calibrate the control hard constraint set and dew point grayscale cost.

[0026] Optionally, identifying the forward with a mask includes:

[0027] A velocity-adaptive front matching kernel based on a convection-diffusion-initiation hysteresis physical structure is constructed, and spatiotemporal convolution is performed on the boundary layer temperature difference field to obtain the front response map;

[0028] The structural tensor of the boundary layer temperature difference field is calculated to extract the normal estimate, and a normal derivative map is generated to suppress oblique noise.

[0029] By combining the forward response map and the normal derivative map, a forward band mask is generated.

[0030] Optionally, synthesizing event-driven control generates a command flow for execution, including:

[0031] Obtain the lower bound of the predicted dew point margin from the predicted state set;

[0032] Construct a control barrier function based on the lower bound of the predicted dew point margin;

[0033] The execution command stream is subjected to safety filtering. The safety filtering is fine-tuned by solving a quadratic programming problem to ensure that the control barrier function always meets the safety conditions.

[0034] Optionally, the calibration of the control hard constraint set and the dew point grayscale cost includes:

[0035] Obtain the lower bound of the predicted dew point margin;

[0036] The set of hard control constraints is determined to require that the lower bound of the predicted dew point margin is not lower than the minimum dew point margin threshold.

[0037] Extract the nominal forecast dew point margin from the forecast state set;

[0038] The cost of dew point grayscale is defined as a function of the difference between the nominal predicted dew point margin and the minimum dew point margin threshold, in order to quantify critical safety risks.

[0039] Optionally, a set of predicted states is established, including:

[0040] By using observable substitutes to drive a reduced-order physical observer, a short-term envelope prediction is generated as a physical upper bound.

[0041] By aligning and fusing data with historical execution command flows, a data-driven prediction model is driven to generate short-term nominal forecasts;

[0042] Combining short-time envelope prediction and short-time nominal prediction forms a prediction state set, which is used to derive the lower bound of the prediction dew point margin.

[0043] Optionally, driving a reduced-order physics observer includes:

[0044] The equivalent diffusion coefficient in the reduced-order physical observer is explicitly mapped to a function of the equivalent mixing thickness;

[0045] The front source terms in the reduced-order physics observer are explicitly mapped to functions of the front shape exponent;

[0046] Solve for a reduced-order physical observer with parameter mapping to generate short-time envelope predictions.

[0047] Optionally, synthetic event-driven control includes:

[0048] Rolling optimization is performed using tubular model predictive control.

[0049] Using short-time nominal forecasts, the nominal trajectory of the tubular model predictive control is solved;

[0050] Using short-time envelope prediction, a safety disturbance tube for tubular model predictive control is defined.

[0051] Dew point grayscale cost is introduced into the objective function of rolling optimization, and a set of control hard constraints is included in the constraints of rolling optimization.

[0052] The solution optimized by rolling is used to generate the execution command stream.

[0053] Optionally, obtaining the lower bound of the predicted dew point margin includes:

[0054] Short-time envelope prediction and short-time nominal prediction are used, and extrapolation is performed through a surface heat transfer closed model to obtain the envelope prediction and nominal prediction of surface temperature, respectively.

[0055] Calculate the envelope prediction of the surface temperature and the difference between the nominal prediction and the local dew point temperature to obtain the predicted dew point margin set;

[0056] Select a lower bound from the set of predicted dew point margins as the lower bound of the predicted dew point margin.

[0057] Optionally, the synthetic event-driven control also includes:

[0058] Based on the aligned and fused data, the isodense reference temperature that makes the supply gas density equal to the cavity density is solved;

[0059] Based on the invading forward velocity and forward shape index, the capture coefficient is estimated, and the coaxial annular suction flow rate is set;

[0060] The synthesis was controlled by an isodense reference temperature and a coaxial annular suction flow rate.

[0061] Optionally, dimensionless acceptance testing and online learning can be implemented, including:

[0062] Injecting the execution command stream into the control model yields a predicted opening surface temperature difference after control injection.

[0063] Construct the residual flow of temperature difference between the predicted temperature difference at the opening surface after controlled injection and the temperature difference field of the boundary layer;

[0064] Using the residual flow of temperature difference, a set of updated model and control parameters is generated;

[0065] The model and control parameter update sets are written back to the estimated boundary layer state and robust prediction and synthetic event-driven control are performed to close the entire process.

[0066] Optionally, alignment and grid mapping include:

[0067] Generate adaptive opening surface meshes with near-wall priority;

[0068] The adaptive opening mesh is inspected for observability. The observability inspection is used to ensure that the mesh has at least a preset number of nodes that span the near-wall layer in the normal direction.

[0069] Adaptive open-face meshes are used to estimate boundary layer states and perform robust predictions.

[0070] Optionally, implementing dimensionless acceptance includes:

[0071] Calculate the dimensionless thermal recovery index using the boundary layer temperature difference field;

[0072] The dew point safety index is calculated using local dew point temperature and surface temperature.

[0073] Dimensionless thermal recovery index and dew point safety index are used for online learning.

[0074] Beneficial effects: This invention can effectively suppress transient disturbances and proactively and robustly avoid near-wall condensation. Attached Figure Description

[0075] Figure 1 A flowchart provided for an embodiment of this application.

[0076] Figure 2 This is a flowchart illustrating the estimation of boundary layer states and the execution of robust predictions for embodiments of this application.

[0077] Figure 3 A flowchart for calculating observable substitutes for embodiments of this application.

[0078] Figure 4 This is a flowchart for evaluating dew point margin in an embodiment of this application.

[0079] Figure 5 A flowchart for identifying the forward band mask in an embodiment of this application. Detailed Implementation

[0080] To enable those skilled in the art to better understand the present invention, the technical solutions of the present invention will be clearly and completely described below with reference to the accompanying drawings of the embodiments of the present invention. Obviously, the described embodiments are only some embodiments of the present invention, and not all embodiments. Based on the embodiments of the present invention, all other embodiments obtained by those skilled in the art without creative effort should fall within the scope of protection of the present invention.

[0081] Example 1: A framework for achieving precise temperature field control inside buildings (e.g., clean rooms, data centers, operating rooms, or precision manufacturing workshops).

[0082] This method aims to address, for example, the rapid and intense temperature transients caused by air exchange when an opening event occurs between a controlled environment inside a building and the outside (such as a corridor), and the potential risk of condensation on near-wall surfaces.

[0083] In this embodiment, a method for finely controlling the temperature field inside a building includes:

[0084] Step 1: Align and mesh-map the received micro-area sensing data to obtain aligned and fused data.

[0085] Specifically, micro-area sensing data can include, but is not limited to: readings from miniature temperature sensor arrays (e.g., thermocouples, RTDs, or fiber Bragg gratings) deployed near opening areas (e.g., door frames, walls), humidity sensor readings, micro differential pressure sensor readings, and signals used to identify opening events (e.g., door magnetic switch signals) and actuator (e.g., air conditioners, heaters, fans) status. Alignment and mesh mapping represent the unification of these data from different sources with different sampling rates and timestamps in both time and space. Time alignment can include clock synchronization and channel delay compensation; spatial alignment (mesh mapping) means mapping the physical location of the sensors onto a unified two-dimensional or three-dimensional mesh coordinate system for subsequent calculations, thereby obtaining aligned and fused data that characterizes the spatiotemporal state of the opening area.

[0086] Step 2: Using aligned and fused data, estimate the boundary layer state and perform robust prediction to establish a predicted state set.

[0087] The boundary layer refers to the thin layer of air that closely adheres to the near-wall surface of an opening region (such as a door frame or wall). The temperature and humidity gradients within this layer change most dramatically during the opening transient, making it the region with the highest risk of condensation. Estimating the boundary layer state involves using aligned and fused data, employing specific algorithms (as described in subsequent examples) to reconstruct the temperature difference field of the region, identify the disturbance front, and calculate key indicators (i.e., observable substitution quantities) to characterize its dynamic properties. Robust prediction refers to predicting not only the nominal evolution trend of the temperature field within a short time window (e.g., 0-10 seconds) but also its worst-case boundary (i.e., the physical envelope) under the influence of uncertainties (such as sensor noise, model errors, and external disturbances). These two together constitute the predicted state set.

[0088] Step 3: Combine the predicted state set with the aligned fusion data to evaluate the dew point margin, thereby calibrating the control hard constraint set and the dew point grayscale cost.

[0089] Dew point margin refers to the difference between the surface temperature and the local air dew point temperature, and is a direct indicator of condensation risk. Assessing dew point margin involves using aligned fused data (obtaining current humidity) and a predicted state set (obtaining predicted surface temperature) to calculate the real-time dew point margin and the predicted future dew point margin (especially its lower bound). Calibrating the control hard constraint set involves setting a safety baseline that the control system must not violate under any circumstances, based on the predicted lower bound of the dew point margin (e.g., requiring the predicted minimum dew point margin to always be greater than a safety threshold, such as 2°C). Calibrating the dew point grayscale cost involves defining a continuous cost function that gradually increases as the predicted nominal dew point margin approaches (but has not yet violated) the hard constraint, thereby guiding the control system (such as subsequent MPC) to proactively avoid risks.

[0090] Step 4: Based on the predicted state set and under the constraints of the control hard constraint set and the dew point grayscale cost, synthesize event-driven control and generate the execution command flow.

[0091] Synthetic event-driven control refers to control logic triggered by an opening event (such as a door opening signal) and performing rolling optimization based on the predicted state set established in step two. Under the constraints of the hard control constraint set and the dew point grayscale cost, control synthesis (e.g., Model Predictive Control, MPC) must treat the dew point hard constraint as an inviolable constraint and include the dew point grayscale cost as a penalty term in the objective function when solving for the optimal control sequence. The execution command flow is the final output of control synthesis, such as a series of specific instructions (e.g., set temperature, fan speed, valve opening, etc.) sent to the air conditioning system (AHU), fan coil unit (FCU), heater, or coaxial exhaust fan.

[0092] Step 5: Perform dimensionless acceptance and online learning on the executed command stream and aligned fused data.

[0093] Dimensionless acceptance refers to the objective evaluation of the control effect after a control event, using dimensionless indicators (such as the thermal recovery index and dew point safety index) that are independent of the control model and based on physical meaning. Online learning refers to using the residuals between the measured data (aligned and fused data) of the current event and the predicted data after control injection to update and write back the parameters of the prediction model in step two and the control model in step four online, thereby enabling the system to perform better in the next event and achieving adaptive closed-loop control.

[0094] Example 2 provides a more detailed preferred implementation of step one (alignment and mesh mapping) in Example 1.

[0095] In this embodiment, alignment and mesh mapping include:

[0096] Step 2.1: Generate a near-wall-priority adaptive opening mesh. Near-wall priority means that in the mesh division of the opening surface, the area closer to the solid wall (such as a door frame or wall) has a higher mesh resolution (i.e., a denser mesh), while the central area farther from the wall can have a relatively lower mesh resolution (i.e., a sparser mesh). For example, the mesh size in the near-wall area (e.g., within 0-10 mm from the wall) can be set to 2-5 mm, while the mesh size in the far-field area (e.g., beyond 10 mm from the wall) can be 10-20 mm. The basis for generating this adaptive mesh can be determined during the system installation and commissioning phase by evaluating the baseline (i.e., the steady state when the door is closed) temperature gradient. Areas with large near-wall gradients require a finer mesh to capture them.

[0097] Step 2.2: Perform observability check on the adaptive opening surface mesh. The observability check is used to ensure that the mesh has at least a preset number of nodes that cross the near-wall layer in the normal direction.

[0098] The observability check involves inspecting the mesh generated in step 2.1 to ensure it meets the minimum requirements for subsequent boundary layer state estimation (as in Example 3). The normal direction refers to the direction perpendicular to the near-wall surface. The preset number of nodes spanning the near-wall layer means, for example, that at least three (or preferably five) sensor nodes or mesh interpolation points must fall within the boundary layer thickness (e.g., 0-12 mm) in the direction perpendicular to the wall. If the sensor placement and mesh generation fail to meet this requirement, subsequent state estimation will fail due to insufficient information. This check ensures effective input for subsequent algorithms.

[0099] Step 2.3: Adaptive open-face meshes are used to estimate the boundary layer state and perform robust predictions.

[0100] In other words, the Ω generated in steps 2.1 and 2.2 open_adapt (Adaptive open surface mesh) will serve as a unified digital basis for all subsequent calculations involving spatial fields (such as temperature difference field and dew point margin field).

[0101] Furthermore, in step one of Embodiment 1, in order to improve data quality and robustness, the following processing is preferably included:

[0102] Sensor health scoring and drift compensation: Before alignment and fusion, the system can assess the health status of each sensor in real time. For example, based on its statistical characteristics such as noise variance, packet loss rate, and zero-drift stability, a health score H is generated for each sensor x. score (x). For sensors with excessively low scores, their data can be removed or their weight in subsequent fusion calculations can be reduced. Simultaneously, the system can robustly estimate the slow drift parameters of the sensors using baseline data and compensate for them before data fusion.

[0103] Baseline set robust estimation and stability quantification: Before the opening event occurs, the system estimates the baseline set B0={T0(x),H0(x),ΔP0} (i.e., baseline temperature field, humidity field, and differential pressure) and the baseline variance Σ from the stable segment. base This estimation preferably employs anomaly-resistant estimation methods (such as RANSAC or Huber regression) to eliminate transient noise. Simultaneously, the baseline stability index S is calculated. base (For example, through change point detection). B0 and Σ base This forms the basis for subsequent calculations of the temperature difference field and the setting of the noise threshold, while S base It can be used to adaptively adjust the parameters of subsequent prediction algorithms (such as Examples 4 and 6).

[0104] Measurement of observation uncertainty: Integrated sensor health score H score (x), baseline variance Σ base In addition to channel delay jitter (from clock synchronization mapping), the system can synthesize a spatiotemporally distributed observation variance field Σ. obs (x). This uncertainty quantification result will preferably be used for the compact design of the disturbance tube radius in subsequent (such as Example 6) tubular model predictive control, making the control more robust to noise.

[0105] Generation of a low-latency, slimmed-down data stream: For control loops requiring extremely high real-time performance (e.g., milliseconds) (such as the CBF security filtering in Example 7), the full aligned and fused data D0(t) may be too large. Therefore, the system can generate a lightweight, slimmed-down data stream D0 in parallel. lite (t). This data stream retains only the features that are critical to control, such as the peak value of the temperature difference field, the arrival time of key points (such as the normal measurement point series), and the slope of the leading edge, thereby greatly reducing the data bandwidth usage of the real-time loop while ensuring that key information is not lost.

[0106] Example 3: A preferred implementation of step 2 in Example 1, which involves estimating the boundary layer state, is described in detail.

[0107] In this embodiment, estimating the boundary layer state and performing robust prediction includes:

[0108] Step 3.1: Reconstruct the boundary layer temperature difference field and identify the frontal zone mask.

[0109] Specifically, the air temperature T in the near field of the opening can be reconstructed from the aligned and fused data D0(t) and the baseline set B0 of Example 2. open (x,t). Boundary layer temperature difference field ΔT bl (x,t) is calculated as the difference between the near-field air temperature at the opening and the baseline temperature T0(x), i.e., ΔT bl (x,t)=Topen (x,t)-T0(x).

[0110] Based on this, identify the forward with mask F mask (x,t). This identification process preferably does not employ traditional static edge detection operators (such as Difference of Gaussians (DoG)) because they struggle to distinguish genuine convective fronts from irrelevant temperature fluctuations in strong turbulence and noise environments. This embodiment employs a velocity-adaptive detection method for opening transients.

[0111] Specifically, identifying the front band mask involves constructing a velocity-adaptive front matching kernel based on a convection-diffusion-initiation hysteresis physics structure. This matching kernel h... vf (n,τ) is constructed exemplarily as: h vf (n,τ)=u(τ)•exp(-n² / (4•α _pre •τ))•(1-exp(-τ / τ c Where n is the normal coordinate; τ is time; u(τ) is the step function, representing the start-up hysteresis; α _pre τ is the estimated near-wall equivalent thermal diffusivity; c Establish a time constant for the forward. This kernel function is designed to physically match the dynamic setup process of the invading forward (i.e., (1 - exp(-τ / τ)). c The term )) represents the development of the frontier from scratch and the diffusion and broadening process over time (i.e., the term exp(-n² / (...)).

[0112] By analyzing the boundary layer temperature difference field ΔT bl Spatiotemporal convolution is performed on (x,t), i.e., R front (x,t)=ΔT bl (x,t)*h vf Receive forward response diagram R front (x,t). Meanwhile, to suppress potential oblique noise in the opening region (e.g., artifacts caused by lateral airflow), a normal estimate needs to be extracted. This can be achieved by calculating the boundary layer temperature difference field ΔT. bl The structure tensor J(x,t) is used to implement the structure, for example, J(x,t)=Gσ*[(▽ΔT) bl (▽ΔT) bl ) T ], where Gσ is the Gaussian smoothing kernel, ▽ΔT bl This represents the gradient of the temperature difference field. Extract the principal axis (i.e., the eigenvector corresponding to the largest eigenvalue) of this tensor J. n This serves as the estimate of the normal at that point. Subsequently, at e n One-dimensional enhancement filtering is performed in the direction to generate the normal derivative map d. nΔT(x,t) enhances the true gradient signal perpendicular to the wall and suppresses oblique noise.

[0113] Finally, combining the forward response diagram R front (x,t) (for example, performing it based on baseline noise statistics Σ) base Adaptive thresholding θ(t) = μ noise +κ•σ noise ) and normal derivative diagram d n ΔT(x,t) (for example, as a consistency criterion, the normal derivative must also be significant), together generate the forward band mask F. mask (x,t). d represents the partial derivative operator.

[0114] This combination of matching kernel and structural tensor orientation, unlike static edge detection, offers stronger physical interpretability and front selectivity. It can accurately capture the truly threatening convective front (i.e., the main body of cold / hot air intrusion) in transient disturbances at openings, providing an accurate spatial range F for subsequent dew point correction (as in Example 5) and substitution calculations. mask (x,t).

[0115] Step 3.2: Calculate the observable substitute. Utilize the boundary layer temperature difference field ΔT obtained in Step 3.1. bl (x,t) and the forward band mask F mask (x,t) is used to calculate observable substitutions along a normal measurement point series. This allows for direct measurement of the boundary layer thickness δ at the millimeter level. bl This is extremely difficult and unreliable in engineering, especially in the intense turbulent environment at the moment of opening. Therefore, this embodiment uses observable substitutes to indirectly characterize the boundary layer state. These substitutes (such as equivalent mixing thickness and front shape index) are indicators that are more easily and robustly obtained through sparse measurement points in a turbulent background.

[0116] Specifically, in F mask Within the region covered by (x,t), along the normal direction (e.g., e in step 3.1) n (Direction) Select at least two measurement point columns, for example, a near-wall measurement point column n1 (about 2-3 mm from the wall) and a far-field measurement point column n2 (about 8-12 mm from the wall).

[0117] The calculation of observable substitutes includes: calculating the dimensionless temperature profile of the normal measurement point series. For example, the dimensionless temperature θ(n,t) = (T... open (n,t)-T0) / max(ΔT bl ), where T open(n,t) represents the temperature at measurement point n, and T0 is the baseline temperature. The slope s(t) at the two points is obtained, i.e., s(t) = [θ(n2,t) - θ(n1,t)] / (n2 - n1). The equivalent mixing thickness δ is then calculated. mix (t) is determined to be the reciprocal of the slopes at the two points, i.e., δ mix (t)=1 / |s(t)|. This δ mix Physically, it reflects the mixing scale of heat in the normal direction.

[0118] Furthermore, the temperature difference step response energy ratio ρ is calculated. E And the arrival time difference Δτ. Specifically, Δτ refers to the time difference between the arrival of the disturbance front (e.g., a temperature step) at n1 and n2, which can be calculated through cross-correlation. ρ E This refers to the energy ratio of the temperature difference time series measured at n1 and n2 under a step event. The front shape index I... front (t) is determined as the ratio of the temperature difference step response energy ratio to the arrival time difference, i.e., I front (t)=ρ E / Δτ. This index I front The larger the value, the sharper the striker and the stronger the pull.

[0119] The observable substitute also includes the invading front velocity v_f(t). This velocity v_f(t) can be estimated by calculating the peak position of the cross-correlation of time series of adjacent rings or adjacent normal measurement point sequences (e.g., n1 and n2).

[0120] Finally, the output equivalent mixing thickness δ mix (t), Forward Shape Index I front (t) and the invading forward velocity v_f(t) will be used together to establish the predicted state set, especially as the real-time driving force for the physical envelope prediction in the subsequent embodiment four.

[0121] In some preferred embodiments, to improve the accuracy of state estimation, the following steps may also be introduced:

[0122] Measuring point observability reweighting: Reconstructing the boundary layer temperature difference field ΔT in step 3.1 bl When (x,t), not all sensors contribute equally. This can be combined with the observability verification report (R) generated in Example 2. obsmesh ,) and sensor health score (H score (x), generate a measurement point weight field W. sens (x). Observability score O score (x) Low (e.g., not meeting the normal node count requirement) or health score H score (x) Lower measurement points will be assigned lower weights W sens (x). The weight field Wsens (x) will be used to reconstruct the temperature difference field in a weighted manner and will serve as the spatial weight in the objective function of subsequent control optimization (such as in Example 6).

[0123] Near-mouth flow direction and shear estimation: The normal vector upon which the structural tensor depends in step 3.1 is preferably not entirely dependent on a fixed geometric normal vector. This can be achieved by inverting differential pressure data (ΔP). raw (t), and surface temperature (T) surf The coupling relationship between (x,t) is used to estimate the actual near-mouth flow direction e. n (x,t) and shear index S shear (t). Using e n Replacing the geometric normal with (x,t) allows for more accurate extraction of the normal derivative and provides shear perturbation information for the reference quantity calculation in Example 6.

[0124] Example 4: A preferred implementation of step 2 in Example 1, which involves performing robust prediction and establishing a prediction state set.

[0125] Because the opening transient (e.g., 0-2 seconds) is strongly dominated by convection and has high uncertainty, while the subsequent recovery phase (e.g., 2-10 seconds) is dominated by diffusion and control behavior, it is difficult to simultaneously achieve both accuracy and robustness in both phases using a single model. Therefore, this embodiment preferably adopts a dual-window prediction architecture, which combines short-term physical envelope prediction with long-term nominal data-driven prediction.

[0126] Establishing the predicted state set includes:

[0127] Step 4.1: Short-time envelope prediction (e.g., 0–2 seconds). The goal of this prediction window is to generate a physical upper bound, i.e., the worst-case scenario under the influence of uncertainty (e.g., the worst dew point margin corresponding to the lowest temperature or highest humidity). This is achieved by using observable substitutes to drive a reduced-order physical observer.

[0128] Specifically, the observer can be a one-dimensional convection-diffusion-source term reduced-order observer constructed on the normal n. An example of the governing equation (PDE) for this observer is: dΔT / dt+v f •dΔT / dn-α Eff •d²ΔT / dn²=S front (t) where ΔT is the boundary layer temperature difference field; t is time; n is the normal coordinate; v f The velocity of the invading front calculated in Example 3; α Eff S is the equivalent diffusion coefficient; front This is the source term for the forward. Note that d is the partial derivative operator.

[0129] It should be noted that the key parameter α in the observer...Eff and S front It is not fixed, but is explicitly mapped in real time to the observable substitutes calculated in Example 3 to anchor the physical model: (1) Driving the reduced-order physical observer includes: adjusting the equivalent diffusion coefficient α in the reduced-order physical observer. Eff Explicitly mapped to the equivalent mixing thickness δ mix A function of α(t). For example, α Eff =f(δ mix (t)). Its physical meaning is that when δ is observed in real time... mix When α increases (indicating more intense boundary layer mixing), Eff The corresponding increase makes the predicted envelope more conservative. (2) The forward source term S in the reduced-order physics observer is increased. front Explicitly mapped to the forward shape index I front A function of (t). For example, S front =g(I front (t)). Its physical meaning is that when I is observed in real time... front When it increases (indicating that the invading vane is sharper and the suction is stronger), the source term S front The pulse width or amplitude is also adjusted accordingly to match the observed leading edge pattern.

[0130] By solving for the reduced-order physical observer after parameter mapping, for example, by integrating forward from 0 to 2 seconds using the current temperature difference field as the initial value, the short-time envelope prediction ΔT, which serves as the physical upper bound, can be obtained. blEnv (x,t+τ). This explicit mapping from measured observables to physical parameters allows the physical envelope to predict ΔT. blEnv Being constantly anchored to real-time observations greatly enhances the reliability of the worst-case upper bound.

[0131] Step 4.2: Short-term nominal prediction (e.g., 2-10 seconds). The goal of this prediction window is to generate the nominal trajectory within a longer future time window, i.e., the most likely evolutionary trend. This is achieved by using a fusion of the data D0(t) and the historical execution command stream U. Exec (t), driving the data-driven prediction model.

[0132] As an example, data-driven prediction models can employ either the controlled Koopman operator (EDMDc) or a neural operator. If EDMDC is used, the system first processes the high-dimensional temperature difference field ΔT... bl The process is elevated to a nonlinear space of observations (basis functions), and then a linear controlled dynamics matrix (Koopman operator) is learned in this space, which describes how the state evolves over time and how it responds to the control input U. Exec(t). If a neural operator (such as the Fourier neural operator FNO or DeepONet) is used, the system will directly learn a state field from the current state field and the control flow U. Exec The operator (function-to-function) mapping from (t) to the future state field. The advantage of this method lies in its robustness to changes in mesh resolution and geometry.

[0133] In some preferred embodiments, to ensure the physical consistency of the data-driven model, a residual regularization term of the convection-diffusion equation (PDE) can be introduced during model training or online updates. For example, an L term can be added to the loss function. _PDE =||dtΔT+u•▽ΔT-α•▽²ΔT-B•U Exec ||², where u is the flow field, α is the diffusion coefficient, and B is the control matrix. This regularization term penalizes predictions that violate known physical laws, enabling data-driven models to make reasonable physical inferences even in regions with sparse data.

[0134] Furthermore, to adapt the model to different stages of the opening event (such as preset, mezzanine, and pullback), an event descriptor E can be introduced. open More preferably, a stage probability vector P can be calculated in real time. _stage (t)=[p _pre ,p mid ,p _tail This vector represents the probability of being in the preset, mezzanine (strong transient), or pullback phase at the current moment. Data-driven models (such as neural operators) will conditionalize based on this probability vector, that is, automatically switch or smoothly integrate different local dynamic models at different phases.

[0135] Finally, this step generates a short-time nominal prediction ΔT. bl_hat (x,t+τ).

[0136] Step 4.3: Combine the predicted state set. Combine the short-time envelope prediction ΔT blEnv (From step 4.1) and short-term nominal forecast ΔT bl_hat (From step 4.2), forming the predicted state set.

[0137] In some preferred embodiments, multi-predictor ensembles can be employed to improve the robustness of the predictions. For example, the physical envelope (step 4.1), the neural operator (step 4.2), and a simplified linear convection-diffusion model can be run simultaneously, and the ensemble prediction result ΔT can be obtained through error weighting or a bagging strategy. bl Ens (x,t+τ) and its confidence bound C Ens (x,t+τ). The confidence bound C Ens(x,t+τ) can be preferably used for the adaptive adjustment of the radius of the tubular MPC disturbance tube in Example 6.

[0138] Furthermore, the online learning closed loop in Example 8 (i.e., temperature difference residual flow r) T (x,t) can be used to update the model parameters online in this embodiment. For example, it can be used to update α in step 4.1 through Bayesian recursion. Eff and S front The parameters enable it to continuously adapt to environmental changes.

[0139] Predicted state set (including ΔT) bl_Env and ΔT bl_hat This will be used to derive the lower bound of the predicted dew point margin and serve as the nominal trajectory and disturbance tube for the tubular MPC in Example 6.

[0140] Example 5: A preferred implementation of step 3 in Example 1, which assesses the dew point margin to calibrate the control hard constraint set and the dew point grayscale cost. This step is crucial for ensuring system operational safety (i.e., preventing condensation).

[0141] In this embodiment, evaluating the dew point margin includes:

[0142] Step 5.1: Obtain the leading edge mask; within the near-wall region defined by the leading edge mask, the humidity obtained from the aligned fusion data is corrected using the worst-case humidity gradient assumption or mass transfer correlation to obtain the corrected local humidity.

[0143] Specifically, readings from volumetric humidity sensors (e.g., those installed in the center of a room or at an air conditioning return vent) are far from representing the true humidity within the near-wall boundary layer during the transient period of an opening. Due to the sudden drop in near-wall temperature caused by cold air intrusion, air stagnates within this thin layer, and water vapor easily accumulates, forming a near-wall wet film effect. This results in local relative humidity far exceeding the volumetric value, which is the primary source of condensation risk. Therefore, the F... (The sentence is incomplete and requires further context to be fully translated.) mask Humidity is corrected within the region (x,t). As one implementation method, the worst-case humidity gradient assumption refers to the local relative humidity (RH) near the wall. _local The estimate is the volumetric humidity RH_bulk plus a safety margin ΔRH given by safety specifications or experience. _worst (For example, ΔRH) _worst =5%RH or 10%RH).

[0144] As a preferred, physics-based approach, mass transfer correlation refers to using thermo-mass analogies (e.g., Chilton-Colburn analogies) or direct calculation. The Reynolds number Re is obtained through flow field estimation (e.g., from the invading front velocity v_f), and combined with the air Schmidt number Sc, the Sherwood number Sh (e.g., Sh = f(Re,Sc)) can be estimated. Sh describes the thickness of the normal humidity boundary layer, from which the bulk humidity RH_bulk can be converted to the equivalent RH near the wall. _local This method is more accurate.

[0145] Step 5.2: Calculate the local dew point temperature using the corrected local humidity and boundary layer temperature difference field. Specifically, RH _local (From step 5.1) and local air temperature T _local (The temperature difference field ΔT reconstructed from Example 3) bl The dew point is substituted into a precise dew point calculation formula, such as the Magnus approximation formula.

[0146] For example, T _dew_local =(b*L) / (aL), where L=log(RH) _local / 100)+(a*T _local / (b+T _local Here, T _dew_local The calculated local dew point temperature is denoted by s; a and b are Magnus empirical coefficients (e.g., when the temperature unit is Celsius, a is approximately equal to 17.62 and b is approximately equal to 243.12).

[0147] Step 5.3: Local dew point temperature is used to calibrate the control hard constraint set and dew point grayscale cost. This step is based on prediction rather than real-time values ​​to achieve forward-looking safety. Specifically, calibrating the control hard constraint set and dew point grayscale cost includes: obtaining the lower bound of the predicted dew point margin (the specific calculation process of this lower bound is detailed in Example 7). The control hard constraint set U... _safe (t) is determined as the lower bound LM required to predict the dew point margin. _hat_min (x,t+τ) is not lower than the minimum dew point margin threshold LM _min In other words, this hard constraint requires that LM be satisfied throughout the entire prediction time domain τ and at all spatial points x. _hat_min (x,t+τ)≥LM _min .

[0148] Preferably, the minimum dew point margin threshold LM _min It is not a fixed empirical value (e.g., 2°C), but can be adaptively adjusted. In a preferred embodiment, LM _min Based on surface material properties M surfCleanroom class (e.g., ISO 1 or ISO 5) is dynamically generated. For example, the system can have a built-in threshold mapping table (Map). _LM For highly sensitive, nucleation-prone substrate materials (such as silicon wafers), LM _min It can be dynamically set to 3.0°C; while for inert stainless steel walls, LM _min It can be set to 1.5°C. This adaptive threshold avoids unnecessarily overly conservative or insufficiently safe conditions.

[0149] Step 5.4: Extract the nominal predicted dew point margin from the predicted state set; define the dew point grayscale cost as a function of the difference between the nominal predicted dew point margin and the minimum dew point margin threshold to quantify the critical safety risk. Nominal predicted dew point margin LM _hat_nom Based on the nominal prediction ΔT in Example 4 bl_hat Derived. Dew point grayscale cost W _LM It is a continuous function used to provide a warning signal to the control optimizer (such as MPC in Example 6) when the system approaches but has not yet violated the hard constraints.

[0150] For example, the grayscale cost W _LM (x,t) can be calculated as: W _LM (x,t)=max{0,LM _min -LM _hat_nom (x,t+τ)}. It can be seen that when the nominal prediction LM... _hat_nom Much larger than LM _min (When it is very safe) W _LM =0; when LM _hat_nom Approaching LM _min At that time, W _LM It begins to increase linearly. This allows the controller to distribute control energy (e.g., heating) proportionally to smoothly mitigate risks, rather than waiting until hard constraints are met before applying emergency braking.

[0151] Finally, the calibrated U _safe (t) and W _LM (x,t) will be output to the control synthesis module in Example 6 as a constraint and cost for its optimization.

[0152] In some preferred embodiments, to further improve the accuracy and robustness of dew point assessment, the method further includes:

[0153] Online identification of surface heat transfer parameters: The calculation of dew point margin relies on accurate prediction of surface temperature. This prediction depends on a closed-loop surface heat transfer model (including convection, radiation, and conduction). To make the model parameters Ψ... surf ={h conv ,ε _rad ,k_sub The system maintains accurate convective heat transfer coefficient, radiation coefficient, and substrate thermal conductivity. During non-critical periods (e.g., after a door is closed or when the baseline is stable), it can actively inject a small-amplitude thermal excitation signal (e.g., a low-amplitude pseudo-random binary sequence) into the actuator (e.g., a heater) and read the response R from the surface temperature sensor. surf (t). Using the stimulus and response data, the system can identify Ψ online using parameter inversion algorithms (such as the least squares method). surf This ensures that the model used for surface temperature extrapolation in Example 7 is always accurate.

[0154] Condensation precursor detection: In addition to passive assessment based on dew point margin (LM), the system can also actively detect condensation precursors. In one implementation, the system can utilize a high-frequency temperature sensor or infrared thermal imager to extract the high-frequency spatial component of the surface temperature, i.e., the thermal texture feature F_tex(x,t). When tiny, un-dropletized water vapor begins to condense on the surface, it causes local latent heat release, resulting in specific changes in the thermal texture. By monitoring F_tex(x,t), the condensation precursor S can be determined. n uc(x,t). This precursor signal S n The uc can be detected before the LM exceeds the limit and sent as an emergency signal to the controller (as in Examples 6 and 7) to tighten the safety barrier.

[0155] Hysteresis and Residence Time Shaping: To prevent the control system from experiencing hysteresis when the true value of LM is exactly at LM _min Frequent control switching (i.e., jitter) during nearby fluctuations can be mitigated by introducing a risk hysteresis parameter Hys. c ondense. For example, setting an entry risk threshold LM. _min _on=2.0°C, but set an exit risk threshold LM. _min _off=2.5°C, and the system is required to have LM>LM. _min The _off state must remain in place for at least 5 seconds before the risk flag is deactivated. This enhances the stability of the control.

[0156] Example 6 is a preferred implementation of the synthetic event-driven control in step four of Example 1, specifically involving the calculation of reference quantities (feedforward) and the rolling optimization (feedback) of tubular MPC.

[0157] In this embodiment, the synthetic event-driven control includes:

[0158] Step 6.1: Calculate the feedforward reference quantity and settings. Synthetic event-driven control also includes: based on aligned and fused data, solving for the isodense reference temperature T_isopyc(t) that makes the supply gas density equal to the cavity density.

[0159] When traditional air conditioning (e.g., 18°C) enters a 22°C cleanroom, the higher density of the cold air causes a buoyancy effect (i.e., the cold air sinks), resulting in strong and difficult-to-control mixing and turbulent infiltration. This embodiment calculates the isodense reference temperature T. _isopyc (For example, heating the supply air to a specific temperature), such that the supply gas (e.g., a heat curtain) mixes with the invading cold air at the opening, resulting in a mixing density ρ. mix As close as possible to the air density ρ inside the cavity _room This minimizes buoyancy-driven disturbances, keeping the intruding airflow suspended at the opening height, making it easily captured by subsequent suction. Optionally, where the actuator (e.g., mixing chamber, heater) bandwidth allows, the near-isenthalpic thermal curtain reference temperature T can be further calculated. curtain_ref (t), the goal of which is to match the supply enthalpy with the cavity enthalpy in order to achieve more precise thermal neutralization.

[0160] Step 6.2: Coaxial suction flow rate setting. Based on the invading front velocity v_f(t) and the front shape index I. front (t), estimate the trapping coefficient, and set the coaxial annular suction flow rate Q. _suction (t). This is a push-pull combined strategy. The heat curtain in step 6.1 is a push, while this step is a pull. The invading forward velocity v_f and the forward shape exponent I... front (All from Example 3) collectively characterize the momentum and entrainment strength of the intrusion. When v _f Very high or I front A high value (sharp front) indicates strong intrusion kinetic energy. The system estimates the required capture factor based on these indicators and accordingly increases the flow rate Q of the coaxial annular suction fan installed around the opening (e.g., door frame). _suction (t) is used to actively capture and expel the intruding airflow (which has been suspended by the isodense strategy of step 6.1) before it diffuses into the cleanroom core area. The isodense reference temperature and the coaxial annular suction flow rate are used as a combination for control, serving as feedforward and initial conditions.

[0161] Step 6.3: Tube Model Predictive Control (Tube-MPC) Rolling Optimization. This step is the core of feedback control. Synthetic event-driven control includes rolling optimization using tube model predictive control. Tube MPC is a robust control method. Standard MPC optimizes only a nominal trajectory; if the model is inaccurate or there are disturbances, the actual trajectory may deviate and lead to danger. The core idea of ​​tube MPC is that while optimizing the nominal trajectory, it also considers the existence of uncertainties and ensures that the real (unknown) system state is always constrained within a tube with a certain radius centered on the nominal trajectory. The specific implementation is as follows:

[0162] Nominal trajectory solution: Predicting ΔT using short-time nominal data bl_hat (From Example 4), solve for the nominal trajectory of the tubular model predictive control.

[0163] Disturbance tube definition: Predicting ΔT using short-time envelope. blEnv (From Example 4), define a safety disturbance tube for tubular model predictive control. The radius ρ of this tube is... _tube It can be determined by ΔT blEnv and ΔT bl_hat The difference between them is used to define the boundary. The controller introduces a compaction constraint, which requires that the nominal trajectory must be greater than the safety boundary (such as LM). _min There is an extra ρ _tube The safe distance ensures that the safety boundary is met even in the worst case (when the actual state deviates to the edge of the pipe).

[0164] Objective function and constraints: Introducing dew point grayscale cost W into the objective function of rolling optimization. _LM (From Example 5), and includes a set of control hard constraints U in the constraints of rolling optimization. _safe (t) (from Example 5). An exemplary objective function J can be constructed as:

[0165] J=Σ[w1•||ΔT bl (t+τ)|| rms 2 +w2•||d n ΔT bl (t+τ)||2 2 +w3•(W LM (x,t+τ),mask)+w4•||ΔU(t+τ)||2 2 ]. Where Σ represents the prediction time domain H MPC The constraints include the integral or summation over the given area; w1 penalizes the norm of the temperature difference field (control objective); w2 penalizes the normal gradient (suppressing sharp fronts); w3 penalizes the dew point grayscale cost (safety objective); w4 penalizes the control increment (energy saving objective). The constraints also include physical amplitude limits and slope (rate of change) limits for actuators (such as heater power and fan speed), and most importantly, U from Example 5. _safe (t) (dew point hard constraint). The solution of the rolling optimization (i.e., minimizing J and satisfying all constraints, the control sequence ΔU) is used to generate the execution command flow U. Exec (t).

[0166] In some preferred embodiments, to improve the performance and robustness of MPC:

[0167] Online Adaptive MPC Parameters: The performance of the MPC is highly dependent on its parameters. Preferably, these parameters are adaptive online. (a) Tube Radius Self-Adjustment: The radius ρ_tube(t) of the perturbation tube can be dynamically adjusted. It is preferably the quantized observation variance field Σ in Example 2. obs (x) and the prediction confidence bound C integrated in Example 4 E A function of ns(x,t+τ). When the sensor noise is high (Σ obs High) or high uncertainty in the prediction model (C) E When the width is less than ns, ρ_tube(t) automatically increases (more conservative control); conversely, it decreases (more aggressive control). (b) Prediction window self-adjustment: The prediction window H_MPC(t) can also be adaptive. It is preferably the baseline stability index S estimated in Example 2. base The function. When the baseline is stable (S base When the baseline is low, use a longer window H_MPC (e.g., 10 seconds) to achieve optimal performance; when the baseline is unstable (S base When H_MPC is high, it is automatically shortened (e.g., 3 seconds) to reduce computational burden and avoid relying on long-term unreliable predictions.

[0168] Energy efficiency adaptive weights: The weights w1 to w4 in the objective function J can also be adaptive. Preferably, there is an upper-level learning loop (see Example 8). Example 8 calculates an Energy Efficiency-Comfort Pareto Index (ECI). This ECI value is written back to adjust the weight vector W_MPC(t)={w1,...,w4} in this example. For example, if the ECI from the previous run indicates excessive energy consumption, w4 (energy-saving weight) will be automatically increased in this round.

[0169] Gate-induced pressure disturbance feedforward: Gate opening events are not only temperature disturbances but also pressure disturbances. The system can independently observe the pressure flow dP caused by this disturbance. obs (t). The dP obs (t) can be used to generate a separate pressure feedforward control signal U. _ffP (t), this signal is directly superimposed onto the execution command stream U. Exec (t) (for example, for controlling micro-overpressure systems) to cancel out pressure waves as they arrive, thereby decoupling pressure disturbances from temperature disturbances.

[0170] Process interlocking and equipment state coordination: In practical applications, the control system must comply with the plant's process constraints. Therefore, the constraint set of the MPC preferably also includes a safety interlock set L. _interlock The interlock set is based on the equipment status e of the factory. _state (t) is read (e.g., suction is prohibited when a process device is running).

[0171] Example 7 elaborates on a key security layer following Example 6 (MPC). This example serves as the hub connecting prediction (Example 4), assessment (Example 5), and control (Example 6), providing ultimate security assurance.

[0172] First, this embodiment details the lower bound of the predicted dew point margin LM upon which embodiments five and six are based. _hat_min The specific calculation process.

[0173] Step 7.1: Obtaining the lower bound of the predicted dew point margin includes: using short-time envelope to predict ΔT blEnv Compared with short-term nominal forecast ΔT bl_hat (All from Example 4), and the surface temperature envelope prediction T was obtained by extrapolation using a surface heat transfer closed model. surf_hat_Env With nominal forecast T surf_hat_nom As in Example 5, this surface heat transfer closed-loop model is a physical model that considers convection, radiation, and conduction, and its parameter Ψ surf Preferably, it is obtained through online identification (micro-thermal excitation).

[0174] Step 7.2: Calculate the envelope prediction T of the surface temperature surf_hat_Env and nominal forecast T surf_hat_nom With local dew point temperature T _dew_local The difference between (from Example 5) is used to obtain the predicted dew point margin set {LM}. _hat_Env ,LM _hat_nom}

[0175] Step 7.3: Select a lower bound from the predicted dew point margin set as the lower bound LM for the predicted dew point margin. _hat_min . Specifically, L.M. _hat_min (x,t+τ)=min{LM _hat_nom (x,t+τ),LM _hat_Env (x,t+τ)}. This LM _hat_min It is the most conservative and robust estimate of the future security state by the system, and it will be used simultaneously for the hard constraint calibration in Example 5 and the security filtering in this example.

[0176] Secondly, this embodiment describes the execution process of safety filtering. Synthetic event-driven control, generating the execution command flow, includes: obtaining the lower bound LM of the predicted dew point margin from the predicted state set. _hat_min (As calculated in steps 7.1-7.3).

[0177] Step 7.4: Based on the predicted dew point margin lower bound LM _hat_minConstruct the control barrier function h(t). The control barrier function (CBF) is a mathematical tool used to provide forward-looking safety guarantees. h(t) is defined as the distance between the system and the unsafe boundary (i.e., LM = LM) over the entire space-time domain. _min The minimum distance of h(t). An exemplary construction of h(t) is: h(t) = min _x,τ [LM _hat_min (x,t+τ)-LM _min ], where min _x,τ This represents the minimum value at all spatial points x and prediction time τ. The system is defined as safe when h(t) ≥ 0; the system is unsafe when h(t) < 0. The goal of CBF is to never allow h(t) to cross 0.

[0178] Step 7.5: Perform security filtering on the command execution flow. Security filtering involves fine-tuning the command execution flow by solving a quadratic programming problem (QP) to ensure that the control barrier function always meets the security conditions. This step occurs during the MPC calculation of the initial U in Example 6. Exec After (t), the goal of MPC is performance optimization (minimizing J), while the goal of CBF is safety first (ensuring h(t) ≥ 0). CBF filtering is implemented through a small-scale quadratic programming (QP) with extremely high computational speed (far faster than MPC). The goal of this QP is to find a modified safety command U. Exec_safe This aligns it with the initial command U of the MPC. Exec As close as possible (i.e., min||U) Exec_safe -U Exec || 2 Simultaneously, a CBF safety constraint must be satisfied. This safety constraint (i.e., the safety condition) is typically expressed as dh / dt ≥ -γ•h(t) (where γ is a positive coefficient, and dh / dt is the time derivative of h(t), which can be calculated through a model). This inequality guarantees that even as h(t) decreases, its rate of decrease is limited by its distance from 0, thus ensuring that h(t) will never reach 0 in a finite amount of time.

[0179] In other words, if the output of Example 6 (MPC) is U Exec If (t) already satisfies the CBF security condition, then the QP solution of the CBF filter is U. Exec_safe =U Exec (Zero changes). However, if the MPC outputs a dangerous U in pursuit of performance (e.g., energy saving), Exec (t) (which will cause h(t) to rapidly approach 0), CBF's QP will immediately intervene to calculate a minimally invasive correction command U. Exec_safe(For example, forcibly increasing heater power), this command only corrects U Exec The dangerous part in (t) is used to ensure that dh / dt≥-γ•h(t) always holds.

[0180] This architecture, which combines MPC optimization with CBF filtering, integrates the foresight performance of MPC with the formal safety guarantee of CBF.

[0181] Step 7.6: Withdrawal and Control Injection. After CBF filtering, the system preferably plans a dew point-priority withdrawal-decompression trajectory. That is, when the event ends and control exits, the control quantity U is not simply... _wave (t) linearly decreases to zero, but according to LM _hat_min The spatial distribution prioritizes protecting areas with the lowest dew point margin, and an exponential retracement curve U is planned. _wave_tail (t), to avoid localized condensation during the exit phase. The final synthesized, safety-filtered execution command stream U Exec (t) will be injected back into the control model to generate the predicted opening surface temperature difference ΔT after control injection. open_hat (t+τ). This ΔT open_hat (t+τ) is the best estimate of the actual response of the system, which will be passed to Example 8 for calculating the residuals and performing online learning.

[0182] Example 8 is a preferred implementation method for elaborating on step five of Example 1, which involves dimensionless acceptance and online learning.

[0183] In this embodiment, implementing dimensionless acceptance and online learning includes:

[0184] Step 8.1: Implement dimensionless acceptance. The purpose of acceptance is to objectively quantify the overall effect of the transient control (including thermal recovery performance and condensation safety performance) using indicators that are independent of the control model and have clear physical meaning after the event. Specifically, implementing dimensionless acceptance includes:

[0185] Using the reconstructed boundary layer temperature difference field (i.e., ΔT in Example 3) bl (x,t)) calculates the dimensionless thermal recovery index. This index, exemplarily referred to as HRI(t), is used to quantify the process by which the temperature difference field at the opening surface recovers to the baseline level.

[0186] An exemplary calculation formula is: HRI(t) = ΔT_rms(t) / ΔT _ref +λ•(τ _ref •|dΔT _rms / dt| / ΔT _ref ), where ΔT _rms (t) refers to the temperature difference field ΔT at the opening surface.bl The root mean square of (x,t) in space; ΔT _ref It is a peak reference, for example, the maximum value of ΔT_rms within 0.5 seconds after the door opening event (t0) can be taken, representing the peak intensity of the disturbance; λ is an empirical coefficient, τ _ref It is a reference time constant, |dΔT _rms / dt| is ΔT _rms The rate of change over time. The first term ΔT of HRI(t) _rms The term (t) / ΔT_ref represents the relative magnitude of the current temperature difference deviation, and the second term λ•(...) represents the relative rate of temperature field recovery. When HRI(t) remains below a qualified threshold ε (e.g., ε=0.1) for a certain period of time (e.g., 5 seconds), the difference between the current time t and the event start time t0 can be defined as the thermal recovery time T. _hot_rec .

[0187] Using the calculated local dew point temperature and surface temperature (i.e., T in Example 5) _dew_local and real-time surface temperature T surf The dew point safety index is calculated. This index, exemplarily called DSI(t), is used to quantify the minimum distance from the condensation safety boundary throughout the entire event.

[0188] An exemplary calculation formula is: DSI(t) = min_x[(T surf (x,t)-T _dew_local (x,t)) / LM _min ]. Where, min_x represents the minimum value at all spatial points x on the open surface; LM _min This is the minimum dew point margin threshold defined in Example 5. If DSI(t) remains greater than or equal to 1 throughout the entire event (i.e., t>t0), it indicates that the entire adjustment process meets the hard dew point safety constraint and passes safely; otherwise, the system will output the out-of-limit segment where DSI(t)<1 and the corresponding execution command stream U. Exec (t), for subsequent auditing and model correction. Dimensionless thermal recovery index and dew point safety index are used for online learning.

[0189] Step 8.2: Implement online learning. The purpose of online learning is to utilize the data obtained during the acceptance phase to form a feedback loop, enabling the predictive and control models to continuously self-optimize. Specifically, implementing dimensionless acceptance and online learning includes:

[0190] Inject the execution command stream (i.e., the final output U in Example 7) into the control model. Exec (t)) to obtain the predicted temperature difference at the opening surface after controlled injection (i.e., ΔT in Example 7). open_hat (t+τ)).

[0191] Predicting the temperature difference ΔT at the opening surface after controlled injection open_hat (t) and the boundary layer temperature difference field ΔT bl The residual flow of temperature difference between (x,t) (i.e., the actual measured values). For example, r_T(x,t) = ΔT bl (x,t)-ΔT open _hat(t). The residual flow r_T(x,t) characterizes the true physical dynamics that the current models (including prediction and control models) fail to accurately describe.

[0192] Using the residual flow due to temperature difference, an update set for the model and control parameters is generated. This update set is exemplarily referred to as Θ. _update It analyzes r through algorithms (such as Bayesian recursion, gradient descent, or Kalman filtering). _T It is obtained from (x,t). For example, r _T (x,t) can be used to update the internal parameters of the data-driven prediction model (such as EDMDC or neural operators) in Example 4, and can also be used to update the parameters of the physical envelope observer (such as α) in Example 4. Eff and S front It can also be used to correct the assumption of near-wall humidity correction (worst humidity gradient) in Example 5.

[0193] Update the model and control parameter set Θ _update The process is written back to estimate the boundary layer state and perform robust prediction and synthetic event-driven control to close the entire loop. This is a key closed loop: Θ _update The data is written back to Embodiment 4 (prediction module) and Embodiment 6 (control module), enabling the model to make more accurate predictions and more optimized control based on the learned experience when the next opening event occurs.

[0194] In some preferred embodiments, this embodiment further includes calculating comparable indices across operating conditions: for example, calculating the recovery Strouhal number θ. _rec =T _hot_rec •v _f ,pk / L open Among them, T _hot_rec It is the heat recovery time; v _f ,pk is the peak speed of the invading forward (from Example 3); L open It is the characteristic scale of the opening. θ _rec It is a dimensionless number used to fairly compare the recovery efficiency of control schemes under different intrusion velocities and opening sizes. Another example is the calculation of the Energy Efficiency-Comfort Pareto Index (ECI=HRI). End / E ctrl HRI EndIt is the residual temperature difference index at the recovery end (representing comfort level); E ctrl The command flow U executed in this event Exec (t) is the total energy consumed. The ECI index (or its Pareto front) is preferably used to construct another learning loop: the ECI is written back to Example 6 (control module) for online adjustment, namely, energy efficiency adaptive weights, that is, adaptively adjusting the ratio of w1 (comfort weight) and w4 (energy saving weight) in the MPC objective function J to achieve automatic trade-off between energy efficiency and comfort under different operating conditions.

[0195] Example 9: In a certain scenario, the implementation process of this method is roughly as follows:

[0196] Assuming the scenario is an ISO 5 level semiconductor cleanroom, its controlled baseline state B0 is: internal temperature T0 = 22.0°C, internal relative humidity H0 = 40.0%, and internal relative corridor slight overpressure ΔP0 = 10 Pa. Minimum safe dew point margin threshold LM. _min The temperature was set to 2.0°C. At t=0, a door leading to the corridor (24°C, 50%RH) was opened. The system (Example 2) started, received micro-area sensing data, and loaded a near-wall-priority adaptive opening surface grid Ω. open _adapt (e.g., 2mm near-wall resolution, 15mm far-field resolution).

[0197] At t=0.5s, the intrusion of cold air (Note: This example assumes the corridor is at 20°C for the purpose of illustrating the temperature difference) is detected. Example 3 Startup: System Reconstruction of Boundary Layer Temperature Difference Field ΔT bl (x,t). Step 3.1 (Forward Mask Recognition): By using the velocity adaptive matching kernel and structure tensor, F was accurately located. mask (x,t). Step 3.2 (Calculation of Substitution Quantities) in F mask Calculations were made on the normal measurement point array (n1=2mm, n2=10mm) within the region, yielding: intrusion front velocity v_f=0.5m / s; slope at two points s=-0.15 / mm, resulting in the equivalent mixed thickness δ. mix =1 / |s|=6.67mm; Forward shape index I front =1.2.

[0198] Establish the predicted state set: Step 4.1 (Physical envelope prediction, 0-2s): Reduced-order observer starts. α E ff is mapped in real time to f(δ) mix =6.67mm); S front Mapped to g(I) front =1.2). Solving this PDE yields the short-time envelope prediction ΔT. blEnvThis shows that at t=1.8s, the worst-case near-wall air temperature will drop to 16.5°C. Step 4.2 (nominal data prediction, 2-10s): The neural operator model is started, combined with the control input U Exec (From the previous moment), the nominal trajectory ΔT is predicted. bl_hat It shows that at t=2.0s, the nominal air temperature will drop to 18.0°C.

[0199] Collaborative startup procedures for Examples 5 and 7: Step 5.1 (Humidity Correction): The system at F mask Within the region, the worst-case humidity gradient assumption is used, and the humidity of the body area H0=40.0% is corrected to RH. _local =40.0% + 10.0% = 50.0%. Step 5.2 (Dew Point Calculation): Based on T _local (Using the predicted worst-case air temperature of 16.5°C) and RH _local =50.0%, the local dew point temperature T is calculated using the Magnus formula. _dew_local ≈5.9°C. Steps 7.1-7.3 (lower margin): The system utilizes a closed-loop surface heat transfer model (assuming its parameter Ψ). surf Extrapolation was performed using the preferred method identified in Example 5.

[0200] ΔT blEnv (Worst air temperature 16.5°C) → Extrapolate to obtain T surf_hat_Env (Worst surface temperature) = 17.8°C.

[0201] ΔT bl_hat (Nominal air temperature 18.0°C) → Extrapolation yields T surf_hat_nom (Nominal surface temperature) = 18.8°C.

[0202] Predicted lower bound of dew point margin:

[0203] LM _hat_min =min(T surf_hat_Env ,T surf_hat_nom )-T _dew_local =17.8°C - 5.9°C = 11.9°C.

[0204] Note: In this case, LM _hat_min Much larger than LM _min =2.0°C, system safe. To demonstrate constraint activation, assume a more stringent operating condition, resulting in T surf_hat_Env =7.5°C, T surf_hat_nom =8.2°C).

[0205] Lower bound of margin for harsh operating conditions (LM) _hat_min =7.5°C - 5.9°C = 1.6°C.

[0206] Nominal margin LM under harsh operating conditions _hat_nom =8.2°C - 5.9°C = 2.3°C.

[0207] Step 5.3 (Hard Constraint): Due to LM _hat_min =1.6°C <LM _min =2.0°C, control hard constraint set U _safe (t) is activated. Step 5.4 (Grayscale Cost): W _LM =max(0,LM _min -LM _hat_nom =max(0, 2.0°C - 2.3°C) = 0. (Note: In this example, the grayscale cost is 0 because the nominal prediction is still outside the threshold. If the nominal prediction LM...) _hat_nom It also drops to 1.8°C, then W _LM =0.2).

[0208] Steps for starting Example 6: Steps 6.1-6.2 (feedforward): The system is based on v_f=0.5m / s and I front =1.2, calculate the required coaxial suction flow rate Q_suction=150CMH; and calculate the isodensity reference temperature T_isopyc. Step 6.3 (MPC optimization): Tubular MPC is started, its objective function J includes w3•W _LM (In this example, it is 0), and its constraint set contains U. _safe (t) (activated in this example). The MPC solver attempts to satisfy U _safe While minimizing J (mainly temperature difference w1 and energy consumption w4), we also minimize J (mainly temperature difference w1 and energy consumption w4).

[0209] Step 7.4 (CBF Construction): Construct the barrier function h(t) = LM _hat_min -LM _min =1.6°C - 2.0°C = -0.4. Step 7.5 (Security Filtering): The system finds h(t) < 0, indicating that the solution of MPC (even considering U) is not satisfactory. _safe The safety conditions are still not met under the current forecast. The CBF safety filter QP is immediately activated to minimize ||U. Exec_safe -U Exec With _mpc||^2 as the objective and dh / dt≥-γ•h(t) as the hard constraint, the solver is forced to fine-tune U. Exec _mpc (e.g., forcibly increasing the heat curtain power from 800W in the MPC solution to 1000W), to obtain the final U Exec (t), to ensure that h(t) is pulled back to a positive value.

[0210] After the event ends (e.g., t=30s), the process of Example 8 is initiated: Step 8.1 (Acceptance): The system calculates the HRI(t) curve of this event and integrates it to obtain T. _hot_rec =12.5s. Simultaneously calculate the DSI(t) curve and find min(DSI(t)) = 1.05, passing safely. Step 8.2 (Learning): The system calculates r_T(x,t) = ΔT bl -ΔT open The system found a systematic underestimation of 0.3°C at the near-wall corner. Θ_update was generated and written back to the neural operator model of Example 4 to correct this bias in the next iteration.

[0211] Example 10 provides a preferred implementation method to ensure the safe operation of the system when the actuator fails, which is a robust extension of the control synthesis in Examples 6 and 7.

[0212] Step 10.1: Fault detection and criterion generation.

[0213] The system continuously transmits the execution command stream U generated in Example 6 or 7. Exec The expected value (t) is compared with the actuator status feedback (e.g., actual heater power, actual valve opening) obtained from the aligned and fused data D0(t) (from Example 2). When the deviation between the expected value and the actual feedback value exceeds a preset threshold (e.g., power deviation > 10%) for a certain period of time (e.g., for 3 seconds), the system generates an actuator fault flag F. _act =1. Alternatively, fault detection can also be achieved by analyzing the residual flow r_T(x,t) (from Example 8). When r_T(x,t) suddenly and continuously deviates significantly from the normal range and cannot be converged by online learning (step 8.2), it can also be determined as F. _act =1.

[0214] Step 10.2: Graceful degradation control synthesis. When F is detected... _act When =1, the system immediately switches its control strategy, abandoning the high-performance optimization goal and instead executing graceful degradation control U. _fallback (t). This degradation control U _fallback The synthesis of (t) will halt the tubular MPC optimization in Example 6 (step 6.3) because MPC relies on an accurate (now defunct) actuator model. Instead, U _fallback (t) will employ a simpler, more robust control logic (e.g., a high-gain PID controller, or a minimized safety QP), whose sole objective is to satisfy the set of hard control constraints U defined in Example 5. _safe (t). In other words, degradation control abandons the control over the thermal recovery time T. _hot_recAlternatively, optimizing Energy Efficiency Index (ECI) might shift the focus of all remaining, usable control resources (e.g., if one of two heaters fails, use the other at full capacity) to prioritize safety, meaning ensuring LM (Low Energy Consumption) at all costs (e.g., regardless of energy consumption). _hat_min ≥LM _min This design ensures that the invention can still fulfill its core safety function (preventing condensation) even in the worst-case scenario of partial hardware failure.

[0215] Specifically, addressing the unobservability problem of boundary layer states, this invention proposes a method for calculating observable alternatives. This method explicitly abandons direct measurement of millimeter-level thicknesses, instead calculating, for example, the equivalent mixed thickness (δ) based on the slope of two points along the normal direction. mix ) and the forward shape index based on energy and time difference (I front These observables are used to indirectly but robustly characterize the boundary layer state. Furthermore, they are used as real-time driving parameters for the physical envelope prediction model (e.g., the equivalent diffusion coefficient α). eff Mapped to δ mix The function ensures that the prediction model is always anchored to real physical observations, thus solving the bottleneck of controlling irrelevance.

[0216] Building upon this foundation, and addressing the hysteresis problem in control response, this invention employs a dual-window robust predictive architecture (e.g., combining short-time physical envelope with long-time nominal data prediction). This architecture provides the control system with a reliable, forward-looking safety envelope that is lacking in the prior art. Subsequent control synthesis (such as tubular model predictive control) and safety filtering (such as control barrier functions) heavily rely on this prediction set, achieving a shift from reactive to predictive-optimization-filtering active control, ensuring action is taken before violations occur.

[0217] Meanwhile, addressing the most critical issue of risk miscalculation, this invention directly confronts and resolves the near-wall wet film effect. By applying a worst-case humidity gradient assumption or mass transfer correlation to force a correction to the volumetric humidity in critical regions (such as within the frontal zone), a more realistic local dew point temperature (T) is calculated. _dew_local This eliminates the false safety margins of the background technology. This ensures that all safety constraints (such as the control hard constraint set and the CBF barrier function h(t)) are calibrated based on real near-wall risks, forming a complete technical closed loop.

Claims

1. A method for fine-tuning the temperature field inside a building, characterized in that, include: Align and mesh map the accessed micro-area sensing data to obtain aligned and fused data; By using aligned and fused data, the boundary layer state is estimated and robust prediction is performed to establish a predicted state set; By combining the predicted state set with the aligned fusion data, the dew point margin is evaluated, thereby calibrating the control hard constraint set and the dew point grayscale cost. Based on the predicted state set, and under the constraints of the control hard constraint set and the dew point grayscale cost, event-driven control is synthesized to generate an execution command flow; Alignment and grid mapping include: Generate adaptive opening surface meshes with near-wall priority; The adaptive opening mesh is inspected for observability. The observability inspection is used to ensure that the mesh has at least a preset number of nodes that span the near-wall layer in the normal direction. Adaptive open-face meshes are used to estimate boundary layer states and perform robust predictions; Estimate the boundary layer state and perform robust prediction, including: The boundary layer temperature difference field was reconstructed from the aligned and fused data, and the frontal zone mask was identified. Observable substitutions are calculated on the normal measurement point series using the boundary layer temperature difference field and the frontal zone mask; The observable substitutes include the equivalent mixing thickness, the front shape index, and the invading front velocity, and are used to build the prediction state set; Calculate observable substitutes, including: Calculate the dimensionless temperature profile of the normal measuring point column and obtain the slope at two points; The equivalent mixing thickness is determined as the reciprocal of the slopes at two points; Calculate the temperature step response energy ratio and arrival time difference; The leading edge shape index is defined as the ratio of the temperature difference step response energy ratio to the arrival time difference. The calculation of observable substitution includes: calculating the dimensionless temperature profile of the normal measurement point series, where the dimensionless temperature θ(n,t) = (T open (n,t)-T0) / max(ΔT bl ), where T open (n,t) is the temperature at measurement point n, and T0 is the baseline temperature. The slope between the two points is obtained as s(t) = [θ(n2,t)-θ(n1,t)] / (n2-n1). The equivalent mixing thickness δ is then calculated. mix (t) is determined to be the reciprocal of the slopes at the two points, δ mix (t)=1 / |s(t)|, this δ mix Physically, it reflects the mixing scale of heat in the normal direction; Calculate the temperature difference step response energy ratio ρ E The arrival time difference Δτ, where Δτ is the time difference between the arrival of the disturbance fronts at n1 and n2, can be calculated through cross-correlation. E This refers to the energy ratio of the temperature difference time series measured at n1 and n2 under a step event, expressed as the front shape index I. front (t) is determined as the ratio of the temperature difference step response energy ratio to the arrival time difference, I front (t)=ρ E / Δτ, the exponent I front The larger the value, the sharper the striker and the stronger the pull. The observable substitution also includes the invading front velocity v_f(t), which can be estimated by calculating the peak position of the cross-correlation of the time series of adjacent annulus or adjacent normal measurement point series; Output equivalent mixing thickness δ mix (t), Forward Shape Index I front (t) and the invading forward velocity v_f(t) will be used together to build the prediction state set; Assessing dew point margin includes: Get the forward with a mask; Within the near-wall region defined by the forward band mask, the humidity obtained from the aligned fusion data is corrected by using the worst-case humidity gradient assumption or mass transfer correlation to obtain the corrected local humidity. The local dew point temperature is calculated using the corrected local humidity and the reconstructed boundary layer temperature difference field. Local dew point temperature is used to calibrate the control hard constraint set and the dew point grayscale cost; The worst-case humidity gradient assumption refers to estimating the local relative humidity near the wall as the volume humidity plus a safety margin given by safety specifications or experience. The calibration control hard constraint set and dew point grayscale cost include: Obtain the lower bound of the predicted dew point margin; The set of hard control constraints is determined to require that the lower bound of the predicted dew point margin is not lower than the minimum dew point margin threshold. Extract the nominal forecast dew point margin from the forecast state set; The cost of dew point grayscale is defined as a function of the difference between the nominal predicted dew point margin and the minimum dew point margin threshold, in order to quantify critical safety risks.

2. The method according to claim 1, characterized in that, Identifying the forward with a mask includes: A velocity-adaptive front matching kernel based on a convection-diffusion-initiation hysteresis physical structure is constructed, and spatiotemporal convolution is performed on the boundary layer temperature difference field to obtain the front response map; The structural tensor of the boundary layer temperature difference field is calculated to extract the normal estimate, and a normal derivative map is generated to suppress oblique noise. By combining the forward response map and the normal derivative map, a forward band mask is generated.

3. The method according to claim 1, characterized in that, Synthetic event-driven control, generating and executing command flows including: Obtain the lower bound of the predicted dew point margin from the predicted state set; Obtaining the lower bound of the predicted dew point margin includes: Short-time envelope prediction and short-time nominal prediction are used, and extrapolation is performed through a surface heat transfer closed model to obtain the envelope prediction and nominal prediction of surface temperature, respectively. Calculate the envelope prediction of the surface temperature and the difference between the nominal prediction and the local dew point temperature to obtain the predicted dew point margin set; Select a lower bound from the set of predicted dew point margins as the lower bound for predicted dew point margins; Construct a control barrier function based on the lower bound of the predicted dew point margin; A control barrier function is a mathematical tool used to provide forward-looking security guarantees; The execution command stream is subjected to safety filtering. The safety filtering is fine-tuned by solving a quadratic programming problem to ensure that the control barrier function always meets the safety conditions.

4. The method according to claim 1, characterized in that, Establishing the predicted state set includes: By using observable substitutes to drive a reduced-order physical observer, a short-term envelope prediction is generated as a physical upper bound. By aligning and fusing data with historical execution command flows, a data-driven prediction model is driven to generate short-term nominal forecasts; The short-time envelope prediction and the short-time nominal prediction are combined to form a prediction state set, which is used to derive the lower bound of the prediction dew point margin. The driving force for the reduced-order physics observer includes: The equivalent diffusion coefficient in the reduced-order physical observer is explicitly mapped to a function of the equivalent mixing thickness; The front source terms in the reduced-order physics observer are explicitly mapped to functions of the front shape exponent; Solve for a reduced-order physical observer with parameter mapping to generate short-time envelope predictions.

5. The method according to claim 4, characterized in that, Synthetic event-driven control includes: Rolling optimization is performed using tubular model predictive control. Using short-time nominal forecasts, the nominal trajectory of the tubular model predictive control is solved; Using short-time envelope prediction, a safety disturbance tube for tubular model predictive control is defined. Dew point grayscale cost is introduced into the objective function of rolling optimization, and a set of control hard constraints is included in the constraints of rolling optimization. The solution optimized by rolling is used to generate the execution command stream.

Images