A heating ventilation system hybrid model predictive control method and device

Abstract

The application discloses a heating system hybrid model predictive control method and device, constructs a device coupling relationship matrix based on heating system operation data, identifies a load response controllable interval, extracts a soft and hard thermodynamic constraint parameter to generate a model embedded constraint coefficient, constructs a thermodynamic constraint prediction model and generates a model failure identifier through working condition coverage analysis; according to the model failure identifier, a rolling time domain load prediction is performed to generate a predicted load sequence, a prediction confidence identifier is extracted and mapped to a device response delay, a response priority is evaluated to generate a layered device activation parameter group; the multi-device coordination constraint is analyzed to identify a priority regulation device, the adjustable margin of the device after the time window overlap reduction of adjacent devices is quantified, and the device regulation weight is adjusted in combination with the model embedded constraint coefficient to output a multi-device coordination control instruction, so that the thermodynamic constraint perception and delay adaptive multi-device coordination predictive control are realized.

Description

Technical Field

[0001] This invention relates to the field of intelligent control technology for heating, ventilation, and air conditioning systems, and in particular to a hybrid model predictive control method and device for heating, ventilation, and air conditioning systems. Background Technology

[0002] HVAC systems consist of various equipment, including chiller units, circulating water pumps, and terminal air conditioning units. These units are interconnected through chilled water piping and heat exchange processes, and changes in the operating status of one unit often directly reduce the adjustability of other units. Existing predictive control schemes simplify the modeling of this physical coupling constraint. The predictive model suffers from decreased confidence under atypical operating conditions due to the scarcity of historical samples, and lacks the ability to sense weak areas in the model and proactively adjust the control strategy. At the level of multi-unit collaborative scheduling, the inherent response characteristics of the equipment and the additional delays introduced by disturbances lead to biases in activation timing estimation. When the forced holding periods of multiple units overlap, the available time for group control is compressed. Existing schemes lack quantitative assessment of this compression effect, and the executability of coordination commands under complex constraint scenarios is difficult to guarantee.

[0003] Therefore, a method is urgently needed to solve at least one of the above problems. Summary of the Invention

[0004] This invention discloses a hybrid model predictive control method and device for HVAC systems. It constructs a device coupling relationship matrix by collecting operational data and performing thermodynamic coupling analysis. A constraint-aware predictive model is built based on the embedding of soft and hard constraint weights. A model failure identifier is generated by combining operating condition coverage analysis. Based on the predicted confidence level, device response delays are mapped to generate a hierarchical device activation parameter group. The overlap reduction margin of the group control time window is quantified and constraint coefficients are fused to output multi-device coordinated control commands, achieving constraint-aware and delay-adaptive multi-device coordinated predictive control.

[0005] The first aspect of this invention proposes a hybrid model predictive control method for HVAC systems, comprising the following steps:

[0006] Collect operating data of the HVAC system, and construct a device coupling relationship matrix based on the operating data through thermodynamic constraint coupling analysis.

[0007] Based on the equipment coupling relationship matrix, the controllable range of load response is identified. Soft and hard thermodynamic constraint parameters are extracted from the operating data to generate model embedded constraint coefficients. The model embedded constraint coefficients are used to perform deep neural network modeling on the controllable range of load response to form a thermodynamic constraint prediction model. Based on the thermodynamic constraint prediction model, the range of insufficient operating condition coverage is identified to generate a model failure identifier.

[0008] Based on the model failure identifier, the thermodynamic constraint prediction model is subjected to rolling time-domain load prediction to generate a predicted load sequence. The prediction confidence identifier is extracted from the predicted load sequence. The response characteristic is mapped through the prediction confidence identifier to obtain the equipment response delay. The response priority is evaluated based on the equipment response delay to generate a hierarchical equipment activation parameter group.

[0009] The multi-device coordination constraint analysis is performed on the hierarchical device activation parameter group to determine the priority adjustment device. The shortest operating constraint time window of the priority adjustment device is detected. The device adjustability margin is determined according to the shortest operating constraint time window. The device adjustment weight is adjusted using the device adjustability margin and the model embedded constraint coefficient, and the multi-device coordination control command is output.

[0010] A second aspect of the present invention provides a hybrid model predictive control device for HVAC systems, comprising:

[0011] The data acquisition module is used to collect the operating data of the HVAC system and to construct the equipment coupling relationship matrix based on the operating data through thermodynamic constraint coupling analysis.

[0012] The model building module is used to identify the controllable load response range based on the equipment coupling relationship matrix, extract soft and hard thermodynamic constraint parameters from the operating data to generate model embedded constraint coefficients, use the model embedded constraint coefficients to perform deep neural network modeling on the controllable load response range to form a thermodynamic constraint prediction model, and identify the insufficient operating condition coverage range based on the thermodynamic constraint prediction model to generate model failure identifiers.

[0013] The prediction optimization module is used to perform rolling time-domain load prediction on the thermodynamic constraint prediction model based on the model failure identifier to generate a predicted load sequence, extract prediction confidence identifiers from the predicted load sequence, perform response characteristic mapping through the prediction confidence identifiers to obtain equipment response delay, and evaluate response priority based on the equipment response delay to generate a hierarchical equipment activation parameter group.

[0014] The coordination control module is used to perform multi-device coordination constraint analysis on the hierarchical device activation parameter group to determine the priority adjustment device, detect the shortest operating constraint time window of the priority adjustment device, determine the device adjustability margin according to the shortest operating constraint time window, adjust the device adjustment weight using the device adjustability margin and the model embedded constraint coefficient, and output multi-device coordination control commands.

[0015] The beneficial effects of this invention are reflected in the following points: 1. Based on the coupling strength of strongly coupled equipment pairs, the initial controllable range is dimensionally shrunk and the intersection is obtained, establishing a controllable range of load response under multi-equipment linkage constraints; soft and hard thermodynamic constraint parameters are transformed into differentiated weights and embedded in the loss function, and high-risk constraint boundary regions obtain stronger penalty strength during training; model failure indicators are generated through operating condition coordinate density analysis and differentiated conservative prediction strategies are triggered in weak coverage areas, realizing a prediction model construction mechanism that combines thermodynamic constraint perception and operating condition coverage adaptation. 2. The inherent delay of equipment and disturbance-induced delay are processed separately. The delay boundary margin is jointly divided by inherent delay clustering and disturbance delay compressible boundary, establishing a quantitative mapping relationship between the prediction confidence decay magnitude and the additional amount of response delay uncertainty. Insufficient margin identification and activation level stability analysis jointly drive dynamic priority adjustment, getting rid of the mismatch problem of fixed priority during prediction accuracy fluctuation periods. 3. By using the constraint dependency graph, multi-device conflicts are classified into two categories: capacity-related and time-related, and eliminated by sorting them separately. The overlapping compression effect of the forced holding period of adjacent devices is quantified into a group control reduction coefficient and incorporated into the margin correction. The adjusted margin of the corrected devices and the embedded constraint coefficients of the model are jointly used to determine the adjustment weight of each device. The adjustment sequence and margin allocation of multiple devices in the shortest running constraint superposition period have clear basis. Attached Figure Description

[0016] Figure 1 This is a flowchart illustrating a hybrid model predictive control method for HVAC systems according to the present invention.

[0017] Figure 2 This is a structural block diagram of a hybrid model predictive control device for HVAC systems according to the present invention. Detailed Implementation

[0018] In the following description, specific details such as particular system architectures and techniques are set forth for illustrative purposes and not for limitation, in order to provide a thorough understanding of the embodiments of this application. However, those skilled in the art will understand that this application may also be implemented in other embodiments without these specific details. In other instances, detailed descriptions of well-known systems, apparatuses, circuits, and methods have been omitted so as not to obscure the description of this application with unnecessary detail.

[0019] References to "one embodiment" or "some embodiments" as described in this specification mean that one or more embodiments of this application include a specific feature, structure, or characteristic described in connection with that embodiment. Therefore, the phrases "in one embodiment," "in some embodiments," "in other embodiments," "in still other embodiments," etc., appearing in different parts of this specification do not necessarily refer to the same embodiment, but rather mean "one or more, but not all, embodiments," unless otherwise specifically emphasized. The terms "comprising," "including," "having," and variations thereof mean "including but not limited to," unless otherwise specifically emphasized.

[0020] The technical solutions of the embodiments of this application will be described below.

[0021] like Figure 1 As shown, this embodiment of the invention provides a hybrid model predictive control method for HVAC systems, including the following steps S110-S140:

[0022] Step S110: Collect the operating data of the HVAC system, and construct the equipment coupling relationship matrix based on the thermodynamic constraint coupling analysis of the operating data.

[0023] Specifically, operational data of the HVAC system is collected. This data is acquired in real-time from sensors and controllers of various sub-equipment within the HVAC system. It covers a wide range of physical quantities, including chilled water supply and return temperatures, cooling water supply and return temperatures, compressor power and operating status, chilled water pump and cooling water pump frequencies, expansion valve opening, supply and return air temperatures and damper openings of terminal air handling units, and indoor and outdoor ambient temperature and humidity. Each physical quantity is grouped according to its sub-equipment identifier, and the acquisition cycle is uniformly set to 1 minute. The acquisition times of various sensor types are uniformly calibrated to eliminate data misalignment caused by clock deviations. The HVAC system operational data is organized hierarchically by subsystem, with data from the chiller unit and terminal units grouped separately. The chiller unit data focuses on collecting compressor operating status, evaporator temperature, and condenser temperature, while the terminal unit data focuses on collecting supply air temperature and damper openings. These two sets of data play different constraint description roles in subsequent thermodynamic constraint coupling analysis. After the data acquisition is completed, the measured values ​​of each measuring point in the HVAC system operation data are checked for validity. The check includes two items: range exceeding the limit and adjacent cycle jump. When the measured value exceeds the upper limit of the sensor range or the change amplitude between two adjacent cycles exceeds the physically achievable rate, it is marked as an abnormal point. Abnormal points are replaced by linear interpolation and an abnormal label is added.

[0024] A device coupling matrix is ​​constructed based on thermodynamic constraint coupling analysis using operational data. Thermodynamic constraint coupling analysis is based on the physical connections formed between sub-devices through fluid pipelines and heat exchange processes. A direct hydraulic constraint exists between the chilled water pump frequency and chilled water flow rate, while a thermodynamic constraint is formed between the chilled water flow rate and the chiller evaporation temperature through heat transfer equations. The superposition of these two levels of constraints results in an indirect coupling relationship between the chilled water pump frequency and the chiller evaporation temperature. Coupling analysis extracts time-series data of the state variables of each sub-device from the operational data, calculates mutual information for each variable pair, and quantifies the statistical dependence between two variables. Higher mutual information indicates a stronger coupling between the operating states of the two sub-devices. After mapping the operating points of each data acquisition cycle to the thermodynamic state space, operating points with close proximity in the state space form operating condition neighborhoods. The mutual information of the sample time-series data for each variable pair within a neighborhood is calculated, and the average value is used as an estimate of the coupling strength within that operating condition interval. The granularity of the operating condition interval division is set to one-tenth of the total state space range to ensure a relatively sufficient number of samples within the neighborhood under typical data scales. The coupling strength between all sub-device pairs is filled into a square matrix to form a device coupling relationship matrix. Diagonal elements set to 1 indicate the self-constraint of the sub-device, while off-diagonal elements reflect the coupling strength between the corresponding sub-device pairs. The device coupling relationship matrix is ​​a symmetric matrix. Sub-device pairs with a coupling strength exceeding 0.6 are marked as strongly coupled device pairs. Strongly coupled device pairs require coordinated handling in subsequent identification of controllable load response ranges and multi-device coordinated control. Adjusting only one device while ignoring the strong coupling relationship may cause the system to deviate from the thermodynamic constraint boundary.

[0025] Step S120: Identify the controllable range of load response based on the equipment coupling relationship matrix, extract soft and hard thermodynamic constraint parameters from the operating data to generate model embedded constraint coefficients, use the model embedded constraint coefficients to perform deep neural network modeling on the controllable range of load response to form a thermodynamic constraint prediction model, and identify the insufficient operating condition coverage range based on the thermodynamic constraint prediction model to generate model failure identifiers.

[0026] Specifically, the controllable load response range is identified based on the equipment coupling relationship matrix. The controllable load response range is the operating range within which the HVAC system can achieve target load tracking through adjustment under current equipment constraints. Within this range, each sub-equipment has sufficient adjustment margin to respond to load commands. Outside this range, there are limitations as equipment has reached its operating boundary and cannot be further adjusted. The coupling strength of strongly coupled equipment pairs in the equipment coupling relationship matrix reflects the narrowing effect of controllable boundaries during coordinated adjustment. When the coupling strength between two devices is high, the adjustment space of a single device is constrained by the current state of the other device. The strong coupling relationship between the chilled water pump and the chiller constrains the adjustable range of the pump frequency to the current evaporation temperature of the chiller. Defining the controllable range solely based on the pump's rated range will lead to an overly wide boundary estimate. The controllable load response range identification uses the rated operating boundary of each sub-equipment as the initial controllable range. Based on the coupling strength of strongly coupled equipment pairs in the equipment coupling relationship matrix, the initial range is shrunk dimension by dimension. The shrinkage ratio of each dimension is determined by a linear mapping of the coupling strength of the corresponding strongly coupled equipment pair. The mapping range is from a shrinkage ratio of 0.05 corresponding to a coupling strength of 0.6 to a shrinkage ratio of 0.5 corresponding to a coupling strength of 1.0. The higher the coupling strength, the larger the shrinkage ratio. After shrinkage, the intersection of the controllable ranges of each sub-equipment in the operating space is the controllable load response range, which is represented by a multi-dimensional boundary polygon. Each dimension corresponds to a type of operating condition parameter, and the coordinates of the boundary vertices record the precise range. After the equipment coupling relationship matrix is ​​updated, the controllable load response range is synchronously re-identified.

[0027] In some embodiments, the step of extracting soft and hard thermodynamic constraint parameters from the operational data to generate model embedded constraint coefficients includes: identifying constraint types based on the operational data to generate a soft constraint parameter set and a hard constraint parameter set; calculating constraint boundary spacing based on the soft constraint parameter set and the hard constraint parameter set to generate a soft and hard constraint critical distance distribution; identifying critical narrowing intervals using the soft and hard constraint critical distance distribution to generate a constraint compression risk set; and applying differentiated weight calibration to the soft constraint parameter set and the hard constraint parameter set using the constraint compression risk set to generate model embedded constraint coefficients.

[0028] Based on operational data, constraint types are identified to generate soft constraint parameter sets and hard constraint parameter sets. Constraint type identification focuses on the physical operating boundaries and control strategy boundaries of each sub-equipment of the HVAC system. Hard constraints originate from equipment physical limits and safety protection settings. Exceeding hard constraints will lead to equipment protection activation or damage. The maximum discharge temperature limit of the chiller compressor and the minimum flow protection value of the chilled water pump are both hard constraints. These constraints are fixed in the controller parameter register as protection trip settings and do not change with operating conditions. Soft constraints are set based on engineering experience in optimizing operating efficiency and ensuring comfort. Soft constraints allow for temporary exceedances under special operating conditions, but will lead to a decrease in operating efficiency or comfort. The recommended operating range for chilled water supply temperature and the comfort-priority range for terminal air valve openings are both soft constraints. The boundaries of soft constraints have a flexible allowable range in actual operation. The historical extreme values ​​of each measuring point in the HVAC system operation data are combined with the protection setting values ​​in the controller parameter register to determine the hard constraint boundary. Operating points in the historical measured values ​​of the operation data that exceed the recommended operating range but do not trigger protection actions are identified as soft constraint tolerable deviation ranges. In the identification results, each constraint parameter is grouped by sub-equipment to form a hard constraint parameter set and a soft constraint parameter set. Each entry in the soft constraint parameter set and the hard constraint parameter set is accompanied by the corresponding sub-equipment identifier and constraint direction label. The constraint direction distinguishes between upper limit constraint and lower limit constraint. The same sub-equipment may have soft constraint and hard constraint entries in both upper limit and lower limit directions at the same time.

[0029] The constraint boundary spacing is calculated based on the soft constraint parameter set and the hard constraint parameter set to generate the critical distance distribution of soft and hard constraints. The constraint boundary spacing is obtained by subtracting the soft constraint boundary value of each entry in the soft constraint parameter set from the hard constraint boundary value of the corresponding sub-device in the hard constraint parameter set in the same direction. The smaller the difference, the closer the soft constraint boundary is to the hard constraint boundary, the smaller the margin between the device and the hard constraint trigger when approaching the soft constraint boundary, and the higher the operational safety risk. The difference is normalized based on the rated operating range of the corresponding sub-device. The normalized spacing of the soft and hard constraint boundaries is D_norm = |B_hard - B_soft| / R_rated, where B_hard is the hard constraint boundary value, B_soft is the soft constraint boundary value, and R_rated is the rated operating range of the sub-device in the corresponding constraint direction. B_hard in the upper constraint direction should be greater than B_soft, and B_hard in the lower constraint direction should be less than B_soft. If these conditions are not met, it is marked as an abnormal constraint configuration. The closer D_norm is to zero, the more closely the soft and hard constraint boundaries are approaching each other. Under extreme high-temperature conditions, the gap between the upper limit of the soft constraint and the protection value of the hard constraint for the chiller exhaust temperature of the HVAC system narrows significantly. The distribution of the critical distance between soft and hard constraints shows low values ​​under this condition, indicating a low adjustment safety margin. The D_norm of all sub-equipment in all constraint directions is arranged in the order of sub-equipment identification to form the distribution of the critical distance between soft and hard constraints. The low-value segments in the critical distance distribution correspond to the sub-equipment and constraint direction combinations that are close to the soft and hard constraint boundaries. These are the key detection targets for subsequent identification of the critical narrowing interval. When the low-value segments in the critical distance distribution are concentrated in a specific subsystem, it indicates that the overall operating margin of that subsystem is low under the current condition.

[0030] The critical narrowing intervals are identified using the critical distance distribution of soft and hard constraints to generate a constraint compression risk set. The critical narrowing intervals are determined by filtering combinations of sub-devices and constraint directions whose normalized spacing is lower than the narrowing threshold in the critical distance distribution of soft and hard constraints. The narrowing threshold is set as the lower quartile of the entire critical distance distribution of soft and hard constraints. Combinations below the lower quartile indicate that their soft and hard constraint spacing is within the 25% range of their historical narrowest point, indicating a high risk of constraint compression. The critical narrowing interval identification process involves filtering entries from the critical distance distribution of soft and hard constraints whose normalized spacing is lower than the narrowing threshold. When multiple constraint directions of the same sub-device simultaneously exhibit critical narrowing in the filtering results, the sub-device is marked as a multi-directional critical device. Multi-directional critical devices face constraint compression risks in multiple directions simultaneously during adjustment. For example, in a chiller operating under high temperature and high humidity conditions, if the upper limit of the exhaust temperature and the lower limit of the chilled water outlet temperature simultaneously approach their respective hard constraint boundaries, and the D_norm in both directions is lower than the narrowing threshold, the chiller is marked as a multi-directional critical device. The constraint compression risk set consists of all entries in the critical narrowing interval. Each entry in the constraint compression risk set is accompanied by the corresponding sub-device identifier, constraint direction and D_norm value. The smaller the D_norm value, the higher the constraint compression risk. In the subsequent weight calibration, the risk-added weight of the corresponding entry in the constraint compression risk set is greater. When there are no entries below the narrowing threshold in the distribution of soft and hard constraint critical distances, the constraint compression risk set is an empty set.

[0031] By applying differentiated weight calibration to the soft and hard constraint parameter sets using a constraint compression risk set, the embedded constraint coefficients of the model are generated. The differentiated weight calibration is based on the D_norm value of each item in the constraint compression risk set. Items with smaller D_norm correspond to higher weights in the model. High-weight constraints exert a stronger penalty on the model output during the deep neural network modeling stage, ensuring that the model's predictions near the high-risk constraint boundary region are sufficiently conservative. The base weight of each item in the hard constraint parameter set is uniformly set to 1.0 to ensure that hard constraints always have absolute priority. The base weight of each item in the soft constraint parameter set is set to 0.5. Items appearing in the constraint compression risk set have an additional risk weight W_risk = 1 / (D_norm + ε) added on top of the base weight, where D_norm is the normalized constraint boundary spacing of the item in the constraint compression risk set, and ε is a zero-prevention constant set to 0.01. The smaller D_norm is, the larger W_risk is. The final weight W_final = W_base + W_risk, where W_base is the base weight. When the soft constraint W_final exceeds the upper limit of 5.0, it is truncated to 5.0; when the hard constraint W_final exceeds the upper limit of 8.0, it is truncated to 8.0, ensuring that the maximum achievable weight of hard constraints is higher than that of soft constraints. Items in the soft and hard constraint parameter sets that do not appear in the constraint compression risk set retain their base weights. When the constraint compression risk set is empty, all items in the model's embedded constraint coefficients use the base weights without applying risk additional weights. The W_final of all constraint entries is organized by sub-device and constraint direction to form the model embedded constraint coefficients. The model embedded constraint coefficients guide the network to pay more attention to high-risk constraint regions during training.

[0032] A thermodynamic constraint prediction model is formed by using deep neural network modeling of the controllable load response range using embedded constraint coefficients. The thermodynamic constraint prediction model uses historical operating data within the controllable load response range as training samples. The input of the thermodynamic constraint prediction model is the current operating state vector of each sub-equipment and the target load command. The output of the thermodynamic constraint prediction model is the predicted adjustment value of each sub-equipment that meets the thermodynamic constraints. The thermodynamic constraint prediction model adopts a multi-layer fully connected architecture, and the number of hidden layers is determined by the dimension of the controllable load response range and the size of the training samples. The model embeds constraint coefficients into the training loss function, which consists of a prediction error term and a constraint penalty term. The total loss is L = L_pred + Σ(W_final_i × L_constraint_i), where L_pred is the mean square of the prediction error, W_final_i is the weight of the i-th constraint in the embedded constraint coefficients, and L_constraint_i is the one-sided violation of the i-th constraint boundary when the predicted value exceeds the boundary. L_constraint_i is zero when the predicted value is within the constraint boundary, and the square of the excess value is taken when it exceeds the boundary. Higher weight constraint entries correspond to stronger penalties. During training, operating condition samples outside the controllable load response range are not included in normal training to avoid abnormal operating condition data outside the controllable load response range interfering with the thermodynamic constraint prediction model's learning of the thermodynamic laws within the controllable range. Samples near the boundary of the controllable load response range are repeatedly included in training with higher sampling weights to strengthen the thermodynamic constraint prediction model's learning of the constraint characteristics of the boundary region. After the thermodynamic constraint prediction model is trained, the prediction error and constraint violation rate are evaluated on the validation set. The thermodynamic constraint prediction model is retrained when the prediction error exceeds the validation threshold or the constraint violation rate exceeds 1%.

[0033] In some embodiments, the step of identifying insufficient coverage intervals and generating model failure identifiers based on the thermodynamic constraint prediction model includes: extracting sample densities for each operating condition interval from the thermodynamic constraint prediction model to generate an operating condition density sequence; identifying density drop inflection points based on the operating condition density sequence to generate a set of coverage density drop intervals; jointly evaluating the drop magnitude and drop location using the coverage density drop interval set to generate a boundary failure risk degree; and using the boundary failure risk degree to calibrate the insufficient coverage intervals and generate model failure identifiers.

[0034] For example, the step of extracting sample densities for each operating condition interval from the thermodynamic constraint prediction model to generate an operating condition density sequence includes: extracting operating condition coordinates of training samples from the thermodynamic constraint prediction model to generate an operating condition coordinate set; calculating the distance between adjacent samples using the operating condition coordinate set to generate a sample distance distribution; identifying areas of sudden expansion of distance based on the sample distance distribution to generate a set of covered blank intervals; and calibrating the degree of coverage weakness of the covered blank interval set to generate an operating condition density sequence.

[0035] A set of operating condition coordinates is generated by extracting the operating condition coordinates of training samples based on the thermodynamic constraint prediction model. Each training sample in the thermodynamic constraint prediction model corresponds to an operating condition coordinate point. The coordinate dimension is consistent with the number of operating condition parameters. The values ​​of each dimension are linearly mapped to the range of 0 to 1 with historical maximum and minimum values ​​as boundaries. After normalization, the dimensions of each dimension are unified, which facilitates the equal contribution of each dimension in subsequent spacing calculations. The set of operating condition coordinates consists of the coordinate points of all training samples of the thermodynamic constraint prediction model arranged in chronological order of acquisition. Dense areas of the set of operating condition coordinates correspond to historically high-frequency operating conditions, while sparse areas correspond to historically low-frequency or non-occurring operating conditions. During seasonal transitions, the operating condition parameters of the HVAC system change rapidly, and the corresponding coordinate points are concentrated along a specific direction in the operating condition space, while the areas on both sides are insufficiently covered because they have not been traversed in historical operations. When the coordinate dimension of the operating condition coordinate set exceeds 3 dimensions, the spacing is calculated directly in the high-dimensional space without dimensionality reduction, so as to retain the complete constraint information of each operating condition parameter dimension. Dimensionality reduction would distort the relative positional relationships between some operating condition parameters. When the number of training samples in the working condition coordinate set is less than 100, it is marked as sparse sample. In the case of sparse sample, the confidence of subsequent spacing calculation is reduced accordingly, and the scope of the insufficient coverage interval in the model failure identifier is expanded accordingly to conservatively estimate the failure risk.

[0036] The sample distance distribution is generated by calculating the distance between adjacent samples using the operating condition coordinate set. The K-nearest neighbor method is used to calculate the distance between adjacent samples. For each coordinate point in the operating condition coordinate set, its K_nn nearest neighbors in the operating condition space are found. The value of K_nn is set to the square root of the total number of training samples. This setting adaptively adjusts the neighborhood size when the number of samples is different; when there are few samples, a small K_nn value indicates a compact neighborhood, while when there are many samples, a large K_nn value indicates a loose neighborhood to smooth out local noise. The mean of the K-nearest neighbor distances is used as the local distance estimate for that coordinate point. The local distance reflects the density of samples near that coordinate point; a small local distance indicates a dense sample area around the coordinate point, while a large local distance indicates a sparse sample area around the coordinate point. The local spacing of all coordinate points in the operating condition coordinate set is arranged in ascending order of the first dimension value of the operating condition parameter to form the sample spacing distribution. Historical samples are scarce near extreme winter and summer operating conditions in HVAC systems, resulting in significantly larger local spacing of the corresponding coordinate points. The sample spacing distribution in this region exhibits high-value peaks. The width of these high-value peaks reflects the span of the operating condition parameters in the sparse region; a larger width indicates a wider coverage gap. Sharp high-value peaks indicate a more concentrated coverage gap, while wide, gentle high-value segments indicate a wider extension of the sparse coverage. A large overall mean of the sample spacing distribution indicates insufficient overall sparse coverage of the training data, while a large standard deviation indicates uneven distribution of the training data, with both local dense and local sparse areas. When the operating condition coordinate set is labeled as sparse, the sample spacing distribution is labeled as low reliability. Under low reliability, the threshold for identifying sudden increases in spacing is appropriately relaxed to improve the detection rate of coverage gaps and avoid missing true coverage faults due to low-confidence spacing estimation caused by insufficient samples.

[0037] Based on the sample spacing distribution, regions with abrupt spacing expansion are identified to generate a set of coverage gaps. These regions correspond to work condition parameter intervals where the local spacing in the sample spacing distribution suddenly increases from a normal level to a significantly high value. This abrupt expansion indicates a sharp decrease in training sample density within this interval, suggesting a coverage discontinuity in the work condition space. The abrupt expansion regions are identified by calculating the mean of the sample spacing distribution using a sliding window, identifying continuous intervals where the mean exceeds a sudden expansion threshold. This threshold is set as the total mean of the sample spacing distribution plus two standard deviations. Exceeding this threshold indicates that the local spacing of the corresponding interval is significantly higher than the overall average. Under the assumption of a normal distribution, this threshold of two standard deviations corresponds to a false positive rate of approximately 5%, balancing the detection rate and the false positive rate. The starting point of a sudden expansion interval is the coordinate point where the local spacing first exceeds the sudden expansion threshold, and the ending point is the coordinate point where the local spacing falls back below the threshold. The range of operating parameters between the starting and ending points constitutes an entry in the coverage blank interval set. Historical operating data for HVAC system cooling capacity is scarce within the partial load rate range, and the sample spacing distribution in this range shows significantly larger local spacing, forming a typical sudden expansion area. The corresponding coverage blank interval indicates that the thermodynamic constraint prediction model needs to add a conservative margin for predictions within this load rate range. If there is a single interval coordinate point below the threshold between two adjacent sudden expansion intervals, they are merged into one interval. The maximum local spacing of the merged interval is the larger of the two values ​​to avoid artificially dividing a continuous coverage blank area into two segments due to a brief recovery. The coverage blank interval set is composed of all sudden expansion intervals. Each interval in the coverage blank interval set includes three attributes: the coordinates of the starting and ending operating parameters, the maximum local spacing within the interval, and the span of the operating parameters within the interval. When there are no continuous intervals in the sample spacing distribution that exceed the sudden expansion threshold, the coverage blank interval set is an empty set.

[0038] The coverage weakness of the set of coverage gaps is calibrated to generate a load condition density sequence. The coverage weakness of each interval in the set of coverage gaps is determined by the maximum local spacing within the interval and the span of the load condition parameters within the interval. The coverage weakness score S_weak = D_max × W_span, where D_max is the normalized value of the maximum local spacing within the interval and W_span is the normalized value of the span of the load condition parameters. Both normalizations are based on the maximum value of the total sample spacing distribution and the total span of the load response controllable interval. The higher the S_weak, the wider the coverage gap occupies in the load condition space and the deeper the fault, and the higher the prediction uncertainty of the thermodynamic constraint prediction model in the corresponding load condition interval. The S_weak of each interval in the covered blank interval set is attached to the corresponding operating condition parameter interval. The operating condition parameter intervals not covered by the covered blank interval set correspond to the normal density segments in the operating condition density sequence. The density value of the normal density segment is determined by taking the reciprocal of the mean of the sample interval distribution in the corresponding interval. The smaller the mean of the interval, the higher the density value. The two types of segments are connected after linear transition smoothing on the operating condition parameter axis. The width of the transition area is one grid step at the boundary of each adjacent segment. After inverting the density values ​​of the normal density segments and the S_weak values ​​of each segment in the blank coverage set, the sequence is smoothly connected sequentially on the operating parameter axis through a linear transition, forming a continuous sequence covering the entire range of operating parameters, which is the operating condition density sequence. In the operating condition density sequence, high-value segments correspond to typical operating conditions with sufficient training data, while low-value segments correspond to operating condition ranges with weak coverage by the thermodynamic constraint prediction model. Typical daytime full-load operating conditions of HVAC systems correspond to high-value segments in the operating condition density sequence, while extreme low-temperature or extreme-high-temperature operating conditions correspond to low-value segments. The location and depth of low-value segments jointly guide the risk level classification of subsequent model failure identification. When the blank coverage set is empty, the operating condition density sequence is directly constructed by taking the reciprocal of the entire sample interval distribution.

[0039] The density drop inflection point is identified based on the load condition density sequence to generate a set of density drop intervals. The density drop inflection point is marked by the location in the load condition density sequence where the sample density suddenly drops from a high value to a low value. A sudden drop indicates a significant discontinuity in the training data coverage near the corresponding load condition interval. The drop inflection point is determined by identifying the location where the amplitude of the negative difference portion exceeds the drop threshold after performing first-order differencing on the load condition density sequence. The drop threshold is set as the upper quartile of the amplitude of the negative difference portion; exceeding the upper quartile indicates that the density drop at that location is within the historical steepest 25% range. After identifying the drop inflection point, the interval is extended until the density recovers to more than 80% of the pre-drop level. If recovery is not achieved by the end of the sequence, the end of the sequence is used as the termination point, and an end-truncation label is added to the drop interval. The load condition sub-intervals within the extended range together constitute a single drop interval. The set of coverage density drop intervals consists of all drop intervals. Each drop interval has three attributes: the starting position of the drop, the drop magnitude, and the lowest sample density within the interval. The drop magnitude is determined by the difference between the mean density before the drop and the mean density after the drop. If two adjacent drop intervals are separated by only a single normal density point, they are merged into one interval. When there are no differential points in the working condition density sequence that exceed the drop threshold, the set of coverage density drop intervals is an empty set.

[0040] The risk of boundary failure is assessed by jointly evaluating the magnitude and location of the sudden drop in density across a set of covered intervals. The magnitude of the sudden drop reflects the severity of the density decrease in that interval, while the location reflects the interval's position in the operating space. Sudden drop intervals closer to the boundary of the controllable load response interval are more likely to be encountered in actual control, resulting in a greater safety impact from model failure. Sudden drop intervals farther from the controllable boundary, even with severe faults, are less likely to be entered in actual control, thus posing a relatively lower risk. The proximity of the sudden drop location to the boundary of the controllable load response interval is measured by the Euclidean distance between their nearest points in the operating space; a smaller distance indicates that the sudden drop interval is closer to the controllable boundary, and the higher the failure risk. The boundary failure risk R_fail = A_norm × Prox_norm, where A_norm is the normalized value of the drop amplitude, equal to the drop amplitude divided by the maximum density value of the load condition density sequence, and Prox_norm is the normalized value of the location proximity, equal to 1 minus the ratio of the Euclidean distance to the maximum distance. The value of R_fail ranges from 0 to 1, with a higher value indicating a higher model failure risk corresponding to the drop interval. R_fail is calculated separately for each drop interval in the coverage density drop interval set, and the risk values ​​of each interval are output in order of the starting position of the drop interval. Training samples for HVAC systems under extreme high temperature and humidity outdoor conditions combined with high load rates are usually scarce. This load condition combination is close to the boundary of the load response controllable interval. The system needs to operate near this interval under extreme summer weather conditions, and the corresponding R_fail is high and requires special attention. When the coverage density drop interval set is empty, the R_fail of all intervals is set to zero, and no subsequent model failure indicators are generated.

[0041] The model generates failure indicators by calibrating the inadequate coverage intervals using boundary failure risk levels. In the boundary failure risk level, the intervals where the R_fail value drops sharply beyond the risk threshold are identified as inadequate coverage intervals. The risk threshold is set to the mean of the total boundary failure risk level plus one standard deviation. This threshold ensures that only intervals with significantly higher-than-average risk are included in the inadequate coverage assessment, avoiding frequent triggering of conservative handling in too many low-risk intervals and reducing control efficiency. Inadequate coverage intervals are described in the form of a combination of operating parameters coordinates. When the HVAC system operates under extreme summer temperatures and the building is simultaneously at full load, the combined operating parameters of outdoor dry-bulb temperature and chiller load rate fall into the inadequate coverage interval. The thermodynamic constraint prediction model lacks sufficient sample support for the joint prediction of chilled water supply temperature and compressor power under this combination, and directly using the model prediction results to formulate control commands carries a high risk of constraint violation. Differentiated conservative strategies are adopted for different levels of failure risk. For high-failure intervals, a larger safety margin is added to the control commands based on the predicted values, and the target setpoint for chilled water supply temperature is shifted towards a safer direction to avoid the evaporation temperature approaching the hard constraint boundary. For moderate-failure intervals, only a smaller margin is added to balance safety and control accuracy. The model failure identifier consists of three fields: the coordinate range of the under-covered interval, the corresponding R_fail value, and the risk level. The risk level is divided into two levels, moderate failure and high failure, based on the R_fail value. The high failure interval triggers a conservative prediction strategy and compresses the rolling prediction step size to half of the default value to control the propagation range of prediction bias. The coordinate range of each under-covered interval in the model failure identifier is recalibrated weekly after the thermodynamic constraint prediction model is updated. As the HVAC system accumulates more extreme operating data, the sample density of some under-covered intervals increases, and the corresponding entries are removed from the model failure identifier.

[0042] Step S130: Based on the model failure identifier, perform rolling time-domain load prediction on the thermodynamic constraint prediction model to generate a predicted load sequence, extract the prediction confidence identifier from the predicted load sequence, obtain the equipment response delay by mapping the response characteristics through the prediction confidence identifier, and generate a hierarchical equipment activation parameter group based on the evaluation of response priority according to the equipment response delay.

[0043] Specifically, based on model failure indicators, a rolling time-domain load prediction is performed on the thermodynamically constrained prediction model to generate a predicted load sequence. The rolling time-domain load prediction uses the current HVAC system operating state vector and the outdoor weather forecast for the future period as inputs, driving the thermodynamically constrained prediction model to predict the load response trajectory of each sub-equipment in the future prediction time domain. The prediction time domain is set to 1 hour by default, with a time step of 1 minute. At the beginning of each prediction step, the current operating condition coordinates are compared line by line with the coordinate range field in the model failure indicator. The corresponding R_fail value field and risk level field in the model failure indicator are read. The R_fail value field determines the additional proportion of the safety margin for this prediction step, and the risk level field determines the activation level of the conservative strategy. When the risk level field is high failure, the thermodynamically constrained prediction model narrows the upper limit of each prediction for this prediction step downwards, with the narrowing magnitude determined by a linear mapping of the R_fail value, and the rolling prediction step size is compressed to 30 seconds. When the risk level field is moderate failure, only a small conservative margin is added to the prediction output without adjusting the step size. The predicted load sequence consists of the predicted load values ​​of each sub-equipment in all prediction steps arranged in chronological order. Each time step in the predicted load sequence corresponds to a record containing the predicted load of each sub-equipment and a confidence interval. The width of the confidence interval is determined by the standard deviation of the historical prediction error of the thermodynamic constraint prediction model in the corresponding operating condition range. The confidence interval of prediction steps that fall into the range that is not covered by the model failure indicator is widened accordingly. The predicted load sequence is updated on a rolling basis with each control cycle. Expired records of the predicted load sequence are retained in a sliding window for the most recent 2 hours to support error backtracking analysis.

[0044] In some embodiments, extracting the prediction confidence identifier from the predicted load sequence includes: calculating the prediction error using a sliding window based on the predicted load sequence to generate an error time series distribution; identifying the error surge inflection point from the error time series distribution to generate a confidence decay time set; sorting the decay magnitudes according to the confidence decay time set to generate a confidence decay magnitude sequence; and identifying high-risk periods based on the confidence decay magnitude sequence to generate a prediction confidence identifier.

[0045] The prediction error is calculated using a sliding window based on the predicted load sequence to generate a temporal distribution of the error. The prediction error is determined by the absolute value of the difference between the predicted value and the measured load value at each time step. A continuously large absolute error indicates a decrease in the prediction accuracy of the thermodynamically constrained prediction model under the current operating conditions. The sliding window width is set to 10 time steps. The mean of the absolute error within the window is used as the local prediction error at the center of the window. The local prediction error is better able to smooth transient fluctuations and reflect changes in accuracy trends than the single-step error. The load of the HVAC system fluctuates drastically during periods of rapid entry and exit of building personnel. After smoothing the sliding mean, this brief increase will not be misjudged as a continuous decline in accuracy. When the predicted load sequence covers multiple sub-devices, the local prediction errors of each sub-device are calculated independently and then weighted averaged. The weighted average formula is E_w(t)=Σ(W_dev_i×e_i(t)) / ΣW_dev_i, where W_dev_i is the importance weight of the i-th sub-device, determined by the sum of the outgoing edge coupling strengths of that device in the device coupling relationship matrix, and e_i(t) is the local prediction error of that sub-device at time t. The error weights of core sub-devices such as chiller power are higher than those of auxiliary devices such as terminal air valves. The E_w(t) values ​​at all times are arranged in chronological order to form the error time series distribution. A continuous increase in the overall level of the error time series distribution indicates that the thermodynamic constraint prediction model is entering a stage of declining prediction accuracy, and it is necessary to pay attention to whether new operating conditions are entering the insufficient coverage range of the model failure indicator.

[0046] This study identifies inflection points of sudden error increases in the error time series distribution to generate a set of confidence decay time points. These inflection points correspond to moments in the error time series distribution where the local prediction error suddenly rises from a low level to a significantly high value. This sudden increase indicates a sharp decline in the prediction accuracy of the thermodynamic constraint prediction model near that moment, leading to a sharp drop in the reliability of the prediction results. The control strategies before and after the sudden increase need to be adjusted accordingly to address the change in prediction accuracy. The inflection points are determined by performing first-order differencing on the error time series distribution and identifying moments where the difference value exceeds a sudden increase threshold. This threshold is set as the upper quartile of the positive value portion of the difference sequence. Only positive differencing is used in the threshold calculation to distinguish between two different modes: a sudden increase in error and a slow decrease in error. Negative differencing indicates that the prediction error is decreasing and does not belong to a confidence decay event. The sudden increase inflection points identified in the error time series distribution are extracted as elements of the confidence decay time set. Each element in the confidence decay time set records the timestamp of the sudden increase and the difference between the local prediction error before and after the sudden increase. During the start-up and shutdown of the HVAC system, the thermodynamic state changes rapidly, and the evaporation and condensation temperatures fluctuate significantly within several time steps after the switch. The predicted load sequence error typically increases sharply during this period. The inclusion of the corresponding moment in the confidence decay time set indicates a decrease in prediction confidence during this period, requiring a reduction in the aggressiveness of control commands. When there are no differential points in the error time series distribution exceeding the sudden increase threshold, the confidence decay time set is empty, indicating that the overall error of the predicted load sequence is stable within the current prediction period, and the prediction confidence remains at a normal level.

[0047] A confidence decay magnitude sequence is generated by sorting the decay magnitudes based on the confidence decay time set. The sudden increase difference recorded in each element of the confidence decay time set reflects the confidence decay magnitude at the corresponding time. The larger the decay magnitude, the more drastic the decrease in prediction accuracy at that time, and the lower the reliability of the prediction results for the corresponding period, requiring a more conservative response margin in the activation scheduling. The decay magnitude sorting arranges all elements in the confidence decay time set in descending order of sudden increase difference. The sorting result intuitively presents the severity distribution of each confidence decay event. The times ranked higher correspond to the periods with the most drastic decrease in the accuracy of the predicted load sequence, and the high-risk period calibration should be triggered first. When multiple chillers in an HVAC system simultaneously enter the adjustment boundary under extreme conditions, the error time series distribution may show multiple sudden increase points in a short period of time. The decay magnitudes of each sudden increase point are different. The descending order sorts the most severe decay events first, ensuring that the highest-risk period receives the highest priority conservative strategy response. Adjacent elements with a time interval of less than 180 seconds in the confidence decay time set are merged into a single record. The larger of the two decay amplitudes is then used after merging. This merging operation avoids identifying consecutive sudden increases within a short period as multiple independent decay events, thus overestimating the number of high-risk periods. Multiple consecutive sudden increases during the cold start-up and shutdown transition phase are merged into a single high-amplitude record, which more accurately reflects the overall decay level of that transition period. The confidence decay amplitude sequence consists of the timestamps and decay amplitude values ​​of all sorted elements arranged in descending order. The length of the confidence decay amplitude sequence is equal to the number of elements in the merged confidence decay time set. If the confidence decay time set is empty, the confidence decay amplitude sequence is empty, and subsequent high-risk period labeling is not performed.

[0048] High-risk periods are identified based on the confidence decay magnitude sequence, generating prediction confidence markers. High-risk periods cover the prediction periods corresponding to moments in the confidence decay magnitude sequence where the decay magnitude exceeds a calibration threshold. The calibration threshold is set as the mean of the entire confidence decay magnitude sequence plus one standard deviation. Exceeding this threshold indicates a significantly higher decrease in prediction accuracy than the average level of the current prediction period, requiring additional conservative margins in control instructions for this period. The moments corresponding to elements in the confidence decay magnitude sequence exceeding the calibration threshold are extended to the moments in the error time series distribution where the local prediction error falls back below the level before the surge. This extended range constitutes the complete high-risk period. When the frequency conversion switching of chilled water pumps in the HVAC system causes brief flow fluctuations, the predicted load sequence error remains high for several minutes after the switch. The extension operation ensures that the entire high-error period is included in the high-risk calibration range, rather than just marking the surge moment, thus avoiding the middle of the high-error period from falling into a non-high-risk zone and requiring aggressive control strategies. The prediction confidence identifier consists of two fields: the start and end times of the high-risk period and the corresponding attenuation magnitude. The attenuation magnitude field in the prediction confidence identifier is used to quantify the added uncertainty of the device response delay in the subsequent response characteristic mapping. The start and end times field locates the specific prediction period for which the degree of control aggression needs to be reduced. When the confidence attenuation magnitude sequence is an empty sequence, the prediction confidence identifier is marked as high confidence for the entire period. No additional uncertainty is added in the subsequent response characteristic mapping of the prediction confidence identifier.

[0049] Equipment response delay is obtained by mapping response characteristics using prediction confidence indicators. The response characteristic mapping takes the attenuation magnitude and start / end times of each high-risk period in the prediction confidence indicators as input, mapping the attenuation magnitude to the added uncertainty in the response delay of each sub-device. The mapping relationship is determined by the regression relationship between the measured response delay and prediction error of each sub-device in the HVAC system during historical high-error periods. The regression relationship reflects the statistical regularity of the actual response delay of the equipment when prediction uncertainty increases. High-risk periods with higher attenuation magnitudes in the prediction confidence indicators correspond to sub-devices with larger added uncertainty, and the response delay estimation is shifted towards a conservative direction to reserve sufficient time margin for control command execution. In HVAC terminal air conditioning unit valves, flow fluctuations during periods of large prediction errors lead to extended valve arrival times. The mapping relationship encodes this statistical regularity as a function of attenuation magnitude to added delay; the higher the attenuation magnitude, the greater the added delay. The device response delay is determined by the sum of the nominal response delay and the uncertainty addition of each sub-device. The nominal response delay is read from the device file. The uncertainty addition is calculated by a mapping function based on the attenuation of the current effective high-risk period indicated by the predicted confidence level. When the predicted confidence level is high confidence for the entire period, the uncertainty addition is set to zero, and the device response delay equals the nominal response delay. The device response delay is output item by item according to the sub-device identifier. Each item includes three pieces of information: the corresponding nominal delay, the uncertainty addition, and the effective period. After the high-risk period ends, the uncertainty addition of the corresponding item in the device response delay is automatically reset to zero and restored to the nominal value. The restoration time is strictly aligned with the end time of the high-risk period in the predicted confidence level.

[0050] In some embodiments, the step of generating a hierarchical device activation parameter set based on the device response delay evaluation response priority includes: using the device response delay to perform delay component decomposition and boundary joint partitioning to generate a delay boundary margin distribution; identifying devices with insufficient margin based on the delay boundary margin distribution to generate a hierarchical boundary risk device set; performing activation level stability constraint analysis on the hierarchical boundary risk device set to generate hierarchical constraint parameters; and evaluating the response priority of each device through the hierarchical constraint parameters to generate a hierarchical device activation parameter set.

[0051] For example, the step of using the device response delay to perform delay component decomposition and boundary joint partitioning to generate a delay boundary margin distribution includes: performing delay source decomposition on the device response delay to generate inherent delay components and disturbance-induced delay components; performing distribution clustering analysis on the inherent delay components to generate inherent delay cluster centers; identifying compressible delay boundaries based on the disturbance-induced delay components to generate a compressible delay threshold set; and jointly partitioning based on the inherent delay cluster centers and the compressible delay threshold set to generate a delay boundary margin distribution.

[0052] The equipment response delay is decomposed into two sources: inherent delay components and disturbance-induced delay components. Delay source decomposition separates the delay values ​​of each sub-device in the equipment response delay into two categories: inherent delay components, which originate from the equipment's mechanical inertia, controller processing delay, and actuator action time. This component is determined by the equipment model and remains relatively stable under the same operating conditions. The inherent delay of the terminal damper in the HVAC system is determined by the rated speed of the valve drive motor and the valve's full stroke angle. Different models of dampers show significant differences in inherent delay, but the inherent delay of the same damper is relatively stable under normal operating conditions. The disturbance-induced delay component originates from current operating condition disturbances, network transmission fluctuations, and load forecast uncertainties, introducing additional delays. This component fluctuates with changes in operating conditions, and periods with high attenuation in the forecast confidence indicator typically correspond to larger disturbance-induced delay components. The delay decomposition method uses the average delay of stable operating conditions in the historical cumulative operating records of each sub-equipment as the intrinsic delay estimate. The historical stable operating condition is defined as a period in the past 30 days where the local prediction error is lower than the long-term average. During this period, the measured delay is closest to the intrinsic level of the equipment due to the minimum operating condition disturbance. During the nighttime light-load stable operation of the HVAC system, the delay values ​​of each sub-equipment are generally close to the nominal value, and the average value of this period reliably reflects the intrinsic delay level. The disturbance-induced delay component is obtained by subtracting the intrinsic delay estimate from the current value of the equipment response delay. A positive difference indicates that the delay is higher than the intrinsic level under the current operating condition, while a negative difference is truncated to zero. The intrinsic delay component and the disturbance-induced delay component are arranged in the order of sub-equipment identification to form two independent sequences. The intrinsic delay component and the disturbance-induced delay component together carry the complete information of the equipment response delay.

[0053] Distributed clustering analysis is performed using the inherent delay components of the devices to generate inherent delay cluster centers. This analysis clusters the inherent delay estimates of all sub-devices within the inherent delay components. The clustering objective is to group sub-devices with similar inherent delay characteristics into the same cluster. Devices within the same cluster exhibit similar timing characteristics during activation scheduling and can be processed uniformly. Devices in the same cluster share the same activation level boundary threshold to simplify scheduling logic. The clustering method uses the K-means algorithm, where K_cluster represents the number of clusters. K_cluster is determined based on the preset number of levels in the hierarchical device activation parameter group. Each level corresponds to one cluster, and the cluster center represents the typical level of inherent delay within that level. In HVAC systems, the inherent delays of variable frequency drives (VFDs), variable frequency pumps, and terminal air valves are typically on different orders of magnitude. VFDs have inherent delays in the tens of seconds range, VFDs in the several seconds range, and terminal air valves in the sub-second range. These three types of equipment naturally form three separate clusters. The inherent delay cluster centers are chosen precisely to fall near the typical levels of the inherent delays for each type of equipment. The clustering results closely match the physical characteristics of the equipment, and the spacing between cluster centers reflects the resolution between different response speed levels. The inherent delay cluster centers are composed of K_cluster clusters whose center values ​​are arranged in ascending order of delay. Smaller center values ​​correspond to faster-responding equipment clusters. This arrangement corresponds to the priority order of the activation levels. When the minimum spacing between clusters is lower than the maximum standard deviation within a cluster after clustering, the clustering is considered invalid, and the inherent delay cluster center degenerates into a single mean value. At this point, all sub-equipment are grouped into the same activation level for unified scheduling.

[0054] Compressible delay thresholds are generated by identifying compressible delay boundaries based on disturbance-induced delay components. The compressible delay boundary represents the upper limit of delay that can be compressed through optimized control strategies within the disturbance-induced delay component. Distances exceeding this boundary indicate that the current operating condition disturbance exceeds the compensation range of control optimization, requiring adjustments to activation priorities rather than simple delay compression. The historical distribution of each sub-device within the disturbance-induced delay component is obtained through kernel density estimation. The upper quartile of the distribution serves as a candidate value for the compressible delay boundary of that device. Distances below the upper quartile can be partially compressed by accelerating command issuance timing and optimizing prediction step size. Delays exceeding the upper quartile typically originate from unpredictable changes in operating conditions. During the switching of multiple chillers in an HVAC system, flow rates fluctuate drastically, and the corresponding peak delay in the disturbance-induced delay component usually exceeds the upper quartile. Such delays require adjustments to activation priorities rather than delay compression to ensure control response. Each sub-device's compressible delay boundary candidate value is accompanied by a confidence assessment. When the number of historical samples is less than 30, the candidate value is marked as low confidence and increased by 10% as a conservative estimate. This conservative increase ensures that the compressible range estimate is biased towards the conservative side rather than the optimistic side, avoiding the activation of scheduling passively when the actual disturbance delay exceeds the estimated compressible upper limit. The candidate values ​​of all sub-devices are aggregated to form a compressible delay threshold set. When all disturbance-induced delay components are zero, each entry in the compressible delay threshold set is assigned a value of zero, indicating that there is currently no disturbance-induced delay available for compression.

[0055] The delay boundary margin distribution is generated by jointly partitioning based on the inherent delay cluster centers and the compressible delay threshold set. The joint partitioning combines the hierarchical boundaries defined by the inherent delay cluster centers with the compressible upper limits of each device in the compressible delay threshold set, defining a delay boundary margin interval for each sub-device within the current active level. The inherent delay cluster center value of the cluster to which each sub-device belongs determines the hierarchical baseline delay for that device, and the compressible delay boundary of the corresponding device in the compressible delay threshold set determines the upper limit of the device's delay under the current disturbance level. The difference between the upper limit and the hierarchical baseline delay is the device's delay boundary margin. When the difference is negative, it is truncated to zero, and the device is marked as a device exceeding the hierarchical boundary. The margin reflects the delay margin space that the device can still maintain stability at the active level under the current disturbance level. In a HVAC system, when the disturbance-induced delay component of a variable frequency water pump approaches the compressible delay boundary during a high-disturbance period, the pump's delay boundary margin approaches zero, indicating that the pump's response delay is approaching the critical level for hierarchical stability. If not addressed first, the operating condition may have already crossed the hierarchical switching boundary before the next activation command is issued. The delay boundary margins of all sub-devices are arranged in the order of sub-device identification to form the delay boundary margin distribution. The joint partitioning method distinguishes between two different causes: inherent delay and disturbance delay. Compared with the overall estimation of the device response delay, the margin assessment more accurately reflects the true stability of the device under the current operating conditions. The delay boundary margin distribution is updated synchronously with the update cycle of the device response delay to maintain real-time correspondence with the current disturbance state.

[0056] Based on the delay boundary margin distribution, devices with insufficient margins are identified, generating a hierarchical boundary risk device set. Sub-devices with delay boundary margins below a margin threshold in the delay boundary margin distribution are extracted as devices with insufficient margins. The margin threshold is set as the lower quartile of the entire delay boundary margin distribution. A margin below the lower quartile indicates that the device's boundary margin is within the historical lowest 25% range. The combination of response delay and operating condition change rate makes this device highly risky of crossing the hierarchical switching boundary during the response process. Devices with insufficient margins need to be given higher priority in activation scheduling to issue control commands as early as possible and shorten the waiting response time. When the response delay of HVAC terminal dampers increases due to increased uncertainty, the delay boundary margin of the dampers narrows accordingly during periods of rapid load change. Devices identified as having insufficient margins need to be prioritized in hierarchical activation to ensure that terminal adjustment actions are completed before the operating condition exceeds the limit. When multiple devices within the same subsystem simultaneously exhibit insufficient margin in the delay boundary margin distribution, it indicates that the overall response capability of the subsystem is approaching saturation under the current operating conditions. A conservative overall scheduling strategy should be applied to the subsystem during hierarchical activation, rather than targeting individual devices. The hierarchical boundary risk device set consists of the identifiers of all devices with insufficient margin and their corresponding delay boundary margin values. Devices with smaller delay boundary margin values ​​are ranked higher in the hierarchical boundary risk device set. Devices ranked higher are given priority in constraint evaluation during subsequent hierarchical stability constraint analysis. When there are no sub-devices below the margin threshold in the delay boundary margin distribution, the hierarchical boundary risk device set is empty, indicating that the delay boundary margins of all current sub-devices are at normal levels.

[0057] Activation level stability constraint analysis is performed on the set of devices at the hierarchical boundary risk to generate hierarchical constraint parameters. The activation level stability constraint analysis takes the current operating status and delay boundary margin of each device in the set as input, and evaluates the constraints that allow each device to stably maintain its current activation level under the current response delay conditions. These constraints describe the timing requirements and upper limits of adjustment that each device must meet during activation scheduling. The device with the smallest delay boundary margin in the set of devices at the hierarchical boundary risk is analyzed first. The analysis includes the minimum adjustment step size and maximum single-step adjustment amplitude required for the device to maintain hierarchical stability under the current operating conditions. The minimum adjustment step size ensures sufficient response time between two adjacent activation commands, while the maximum single-step adjustment amplitude prevents a single large adjustment from causing the operating condition to quickly exceed the hierarchical switching boundary. For HVAC system chilled water pumps, under insufficient margin conditions, the activation level stability constraint requires that the single-step frequency adjustment amplitude not exceed 5% of the rated frequency. Exceeding this amplitude will cause excessively rapid flow changes, leading to hydraulic imbalance and triggering unexpected hierarchical switching. This constraint is determined through statistical analysis of adjustment amplitudes under historical stable operating conditions. The hierarchical constraint parameters are composed of the minimum adjustment step size and the maximum single-step adjustment amplitude of each device in the hierarchical boundary risk device set. Each constraint item in the hierarchical constraint parameters is accompanied by the corresponding device identifier and the current delay boundary margin. When the hierarchical boundary risk device set is empty, the hierarchical constraint parameters adopt the default adjustment step size and amplitude limit of all sub-devices. The default value ensures that the device adjustment efficiency is not over-constrained under normal margin.

[0058] Hierarchical device activation parameter groups are generated by evaluating the response priority of each device through hierarchical constraint parameters. The maximum single-step adjustment range and minimum adjustment step size of each device in the hierarchical constraint parameters jointly determine the maximum effective adjustment amount that the device can achieve per unit time. Devices with larger effective adjustment amounts contribute more efficiently to the load response and have higher ranking potential in activation priority evaluation. The response priority of each device is determined by the ratio of the effective adjustment amount to the device response delay. When the device response delay is zero, the response priority is directly taken as the maximum value of the effective adjustment amounts of all current devices. The higher the ratio, the more effective the adjustment amount that device can achieve within a shorter delay, and the more significant its contribution to rapid load response. In HVAC systems, variable frequency chilled water pumps with low delays and ample adjustment ranges typically have a higher response priority than terminal air valves with large response delays. Prioritizing the activation of variable frequency pumps can more quickly change the total system flow in response to load changes. Terminal air valves are activated sequentially after the pump's adjustment capacity is exhausted to handle fine-grained adjustment needs at the terminal. All sub-devices are sorted from highest to lowest response priority. The sorting result, combined with the adjustment step size and amplitude limit of each device in the hierarchical constraint parameters, constitutes a hierarchical device activation parameter group. The hierarchical device activation parameter group is organized with the activation level as the primary key. Devices within the same level are arranged according to response priority, with higher-level devices responding first to core load changes. Lower-level devices are activated sequentially after the adjustment capacity of higher-level devices is exhausted. The hierarchical device activation parameter group is updated synchronously with the update cycle of device response delay and prediction confidence indicators. When response delay changes or high-risk periods are entered or exited, the activation priority is reassessed, and the hierarchical device activation parameter group is adjusted accordingly to maintain consistency with the current system state.

[0059] Step S140: Perform multi-device coordination constraint analysis on the hierarchical device activation parameter group to determine the priority adjustment device, detect the shortest operating constraint time window of the priority adjustment device, determine the device adjustment margin based on the shortest operating constraint time window, and adjust the device adjustment weight using the device adjustment margin and the model embedded constraint coefficient to output multi-device coordination control command.

[0060] In some embodiments, the step of performing multi-device coordination constraint parsing on the hierarchical device activation parameter group to determine the priority adjustment device includes: extracting the constraint association relationship of each device through the hierarchical device activation parameter group to generate a constraint dependency graph; identifying the conflict type of the constraint dependency graph to generate a capacity conflict set and a temporal conflict set; performing classification conflict elimination priority sorting according to the capacity conflict set and the temporal conflict set to generate a classification conflict elimination priority sequence; and determining the priority adjustment device based on the classification conflict elimination priority sequence.

[0061] Constraint dependency graphs are generated by extracting the constraint relationships of each device through a hierarchical device activation parameter group. The constraint relationship extraction is based on the upper limit of the adjustment range and the minimum adjustment step size of each level of device in the hierarchical device activation parameter group. Combined with the coupling strength between corresponding device pairs in the device coupling relationship matrix, the mutual influence between the adjustment actions of each device is inferred. The higher the coupling strength, the stronger the constraint relationship with the device, and the greater the difficulty of adjustment coordination. Each device in the hierarchical device activation parameter group is considered a node. The directed edges between nodes represent the direction of constraint influence. The starting point of the edge is the device applying the constraint, and the ending point is the constrained device. The weight of the edge, W_edge, is determined by the product of the corresponding coupling strength, C_couple, and the current margin tension of the constrained device: W_edge = C_couple × (1 - R_avail / (C_max + ε_c)), where R_avail is the current remaining adjustment margin of the constrained device, C_max is the upper limit of the rated adjustment margin of the device, and ε_c is a zero-prevention constant set to 0.01. The closer the margin is to exhaustion, the higher the weight, indicating a more significant constraint compression of the ending device by the adjustment action of the starting device. When a circular dependency path appears in the constraint dependency graph, it is marked as a cyclic constraint group. In extreme conditions, chilled water pumps, chillers, and terminal air valves in a HVAC system may form a three-node cyclic constraint group. Adjusting any device within this group will compress the margin of the other two devices. The larger the size of the cyclic constraint group, the higher the difficulty of adjustment and coordination. The number of nodes in the constraint dependency graph equals the total number of devices in the hierarchical device activation parameter group. The edge density of the constraint dependency graph reflects the overall degree of constraint coupling between devices; a high edge density indicates that more refined priority ranking is needed for multi-device coordination and scheduling.

[0062] Conflict type identification is performed on the constraint dependency graph to generate capacity conflict sets and temporal conflict sets. Conflict type identification focuses on the constraint effects of each directed edge in the constraint dependency graph, distinguishing between two types of conflict causes: capacity conflict refers to two devices simultaneously competing for limited adjustment margins, thus compressing each other's available adjustment space; temporal conflict refers to two devices causing one device's adjustment action to lead to a constraint violation state due to improper response timing. Capacity conflict is identified by checking whether the sum of the current remaining margins of the nodes at both ends of the directed edge in the constraint dependency graph is lower than the sum of the demand adjustment amounts of the two devices corresponding to the current time step in the predicted load sequence. If the sum of the remaining margins is insufficient, a capacity conflict is determined. In HVAC systems, during high-load periods, chilled water pumps and chillers simultaneously require significant adjustments, but the overall margin is limited. Neither device can achieve its target adjustment amount, and a margin allocation scheme needs to be determined through priority ranking. Temporal conflicts are identified by examining the temporal relationship between the response delay of the starting device and the adjustment initiation time of the ending device in the constraint dependency graph. A temporal conflict is determined when the response completion time of the starting device is later than the adjustment initiation time of the ending device. This conflict indicates that the ending device started adjustment before the state of the starting device was stable, and there is uncertainty in the constraint boundary estimation. Capacity conflict set and temporal conflict set respectively summarize the conflict entries of the corresponding types. Each entry in the capacity conflict set and temporal conflict set is accompanied by the identification of the involved device and a conflict severity score. The score is determined by the degree of insufficient residual margin or the magnitude of temporal misalignment. When there are no conflicting edges in the constraint dependency graph, both the capacity conflict set and the temporal conflict set are empty sets.

[0063] Based on the capacity-related conflict set and the temporal conflict set, a priority sequence for conflict elimination is generated by classifying and prioritizing conflict elimination. The priority sequence establishes elimination priorities for both types of conflict sets. In the capacity-related conflict set, each conflict item is sorted from highest to lowest based on the degree of insufficient margin. Conflicts with higher margin insufficiency correspond to a greater degree of inability for the two devices to simultaneously complete the target adjustment amount; these are eliminated first to release the occupied margin as early as possible, ensuring the core load response requirements are met. In the temporal conflict set, each conflict item is sorted from largest to smallest temporal misalignment. Larger temporal misalignment indicates a more severe misalignment in the adjustment timing of the two devices; prioritizing its elimination can prevent cascading violations caused by constraint estimation errors during the misalignment period. Premature and significant adjustment of terminal air valves before the chiller power response of the HVAC system is completed may cause a rapid drop in evaporation temperature, triggering chiller protection shutdown; temporal misalignment elimination can effectively avoid this cascading risk. After the two conflict sets are sorted, they are merged into a unified classification conflict elimination priority sequence. When the same device appears in both conflict sets, the one with the higher score is retained and duplicate entries are removed. The merging rule is to take the larger value after normalizing the capacity conflict priority score P_cap and the temporal conflict priority score P_seq to determine the sorting position after merging. P_cap is determined by the percentile of the margin deficiency degree in the entire capacity conflict set, and P_seq is determined by the percentile of the temporal misalignment amplitude in the entire temporal conflict set. When both the capacity conflict set and the temporal conflict set are empty sets, the classification conflict elimination priority sequence is an empty sequence. When the classification conflict elimination priority sequence is an empty sequence, the priority adjustment device is directly the device with the highest response priority in the hierarchical device activation parameter group.

[0064] Priority adjustment devices are determined based on the conflict elimination priority sequence. The conflict entry ranked first in the priority sequence corresponds to the device pair that most needs priority processing. From this device pair, the device with the higher constraint influence is selected as the priority adjustment device. The constraint influence is determined by the sum of the outgoing edge weights of the device in the constraint dependency graph. When the constraint influence is equal, the device with the larger response delay is selected as the priority adjustment device to ensure that the slow-response device gets priority adjustment time. The higher the sum of the outgoing edge weights, the more extensive the constraint influence of the device's adjustment state on other devices. When the same device is ranked high in multiple conflict entries in the priority sequence, the device is dependent on multiple conflicts. Once its adjustment state is determined, the uncertainty of multiple conflicts can be eliminated simultaneously. In HVAC systems, chilled water pumps are usually ranked high in multiple entries in the priority sequence. After selecting them as priority adjustment devices, frequency locking can simultaneously eliminate the constraint estimation uncertainty of the chiller and terminal air valves. Once the priority adjustment device is determined, its current available adjustment range and the expected response completion time are recorded as known constraint quantities. These known constraint quantities serve as prerequisites for the adjustment of other devices in the subsequent shortest running constraint time window detection. Subsequent entries in the classification conflict elimination priority sequence are processed sequentially. After each conflict is eliminated by the priority adjustment device, the available margin estimate of the remaining devices is updated before processing the next conflict, until all conflicts are eliminated or the margin is exhausted.

[0065] The shortest operating constraint window for priority adjustment equipment is detected. The shortest operating constraint window refers to the minimum time that the priority adjustment equipment must maintain its current state after completing a single adjustment action. Within this window, no further adjustment commands can be issued to the priority adjustment equipment. This forced hold window ensures that the equipment has sufficient time to reach a new steady state after adjustment, preventing equipment oscillations caused by continuous adjustment. The shortest operating constraint window is detected by reading from the equipment file and current operating status of the priority adjustment equipment. The equipment file records the minimum continuous operating time requirements for each model of equipment. For chilled water pumps in HVAC systems, the shortest operating constraint window is determined by both motor thermal protection requirements and hydraulic stabilization time. Frequent start-stop cycles can lead to excessive motor winding temperature rise and water hammer impact in the pipes. The window is typically between 3 and 10 minutes. The current running time of the priority adjustment equipment is read from the controller's operating log. The difference between the current running time and the shortest operating constraint window yields the remaining constraint time. A positive remaining constraint time indicates that the equipment is still within the forced hold window and further adjustment is not allowed. A zero or negative remaining constraint time indicates that the forced hold window has ended and the equipment can accept new adjustment commands. The shortest operating constraint window is dynamically updated as the thermodynamic state of the priority regulating device changes. When the steady-state time of the device is prolonged due to changes in thermodynamic state during operation, the window is extended accordingly. When the number of priority regulating devices exceeds one, the shortest operating constraint window of each device is detected independently.

[0066] In some embodiments, determining the device adjustability margin based on the shortest operating constraint window includes: generating a remaining window sequence by statistically analyzing the remaining windows of each device based on the shortest operating constraint window; generating a window overlap distribution by identifying overlapping intervals of adjacent handover request windows based on the remaining window sequence; performing a group control adjustment margin reduction evaluation on the window overlap distribution to generate a margin reduction coefficient; and using the margin reduction coefficient to correct the remaining window sequence to generate the device adjustability margin.

[0067] The remaining time windows of each device are statistically analyzed based on the shortest operating constraint time window to generate a sequence of remaining time windows. The remaining time window for each device is measured by the remaining time between the current moment and the end of its shortest operating constraint time window. The shorter the remaining time, the faster the device can accept a new adjustment command; a remaining time of zero indicates that the device can adjust immediately. The start time of the forced hold time window and the length of the shortest operating constraint time window for each device are recorded in the shortest operating constraint time window to calculate the remaining time for each device. The start time is determined by the timestamp of the device's most recent adjustment action, and the length of the shortest operating constraint time window is determined by the device file and the current thermodynamic state. After completing an opening adjustment, the terminal damper of the HVAC system must maintain the current opening for at least 30 seconds to wait for the indoor temperature response to stabilize. This waiting time constitutes the damper's shortest operating constraint time window. If the damper receives a new adjustment command before the end of the shortest operating constraint time window, the command is temporarily stored until the end of the shortest operating constraint time window before execution. The remaining time of all devices in the hierarchical device activation parameter group is arranged according to the device response priority to form a remaining time window sequence. Devices with a value of zero in the remaining time window sequence are adjustable immediately, while devices with a value greater than zero are required to wait for the corresponding time. When all elements in the remaining time window sequence are greater than zero, it means that no device is currently adjustable and the adjustment command must be issued after waiting for the time corresponding to the smallest element.

[0068] The overlap interval of adjacent handover request time windows is identified based on the remaining time window sequence, generating a time window overlap distribution. When the mandatory hold time windows of two adjacent response priority devices overlap on the time axis in the remaining time window sequence, an adjacent handover request time window overlap interval is formed. Time window overlap indicates that both devices are in a mandatory hold state during the overlap period; if a load adjustment command is received during this period, neither device can be called to respond. The identification of the time window overlap interval is achieved by performing an intersection operation on the time axis of the time window ranges of two adjacent devices in the remaining time window sequence. The intersection length is the overlap duration, and the overlap degree O_ij = T_overlap / min(T_i, T_j), where T_overlap is the overlap duration of the time windows of the two devices, and T_i and T_j are the shortest operating constraint time window lengths of the two devices, respectively. When the time window length of either device is zero, O_ij is directly assigned a value of zero, indicating no overlap constraint. The closer O_ij is to 1, the deeper the overlap. After the chilled water pump in the HVAC system completes its adjustment, it is within the shortest time window. If the chiller also happens to be within its shortest time window during the same period, the time windows of the two devices highly overlap, forming a period of adjustment resource vacuum. The time window overlap distribution shows a high value during this period, indicating that necessary adjustments need to be completed before the vacuum period or that instructions should be issued after the vacuum period ends. The O_ij values ​​of all adjacent device pairs in the remaining time window sequence are arranged in order of device pair response priority to form the time window overlap distribution. An overlap of zero indicates that the time windows of the corresponding device pair do not overlap. A time window overlap distribution of zero indicates that there is no overlap of adjacent device time windows in the current layered device activation parameter group.

[0069] A group control adjustment margin reduction assessment is performed on the time window overlap distribution to generate a margin reduction coefficient. The group control adjustment margin reduction assessment converts the overlap of each pair of devices in the time window overlap distribution into a reduction ratio of the corresponding device's adjustment margin. A high overlap indicates a high probability that both devices are simultaneously unadjustable, reducing the overall available adjustment opportunities for this pair in group control scheduling. The margin loss due to overlapping periods needs to be deducted from the margin estimate. The reduction ratio F_i is calculated as F_i = Σ(O_ij × W_joint_ij), where O_ij is the overlap between the i-th device and its adjacent j-th device, and W_joint_ij is the joint weight of the two devices, equal to the harmonic mean of their response priorities. The summation covers all adjacent pairs of devices of the i-th device. A larger F_i indicates a more severe impact of group control time window overlap on the i-th device. During periods of frequent group control switching in the HVAC system, multiple devices simultaneously enter the shortest time window constraint. In the time window overlap distribution, the overlap of multiple devices is simultaneously high, and the reduction ratio of each device is cumulatively superimposed. The comprehensive margin reduction coefficient K_reduce is determined by the weighted average of the reduction ratios of all devices, K_reduce=1-Σ(F_i×P_i) / Σ(P_i), where P_i is the response priority weight of the i-th device, and the value of K_reduce ranges from 0 to 1. The smaller K_reduce is, the more severe the compression of the overall adjustment margin due to the group control time window overlap is. When the total time window overlap distribution is zero, the margin reduction coefficient K_reduce is assigned a value of 1, indicating no reduction.

[0070] The remaining time window sequence is modified using a margin reduction factor to generate the device's adjustability margin. The remaining duration of each device in the remaining time window sequence reflects the device's adjustability opportunity in the time dimension, while the margin reduction factor reflects the degree to which group control constraints compress the adjustability opportunity as a whole. Combining the two yields the actual usable adjustability margin of each device under the current group control state. The correction method multiplies the rated adjustability margin C_rated of each device in the remaining time window sequence by the margin reduction factor K_reduce. The rated adjustability margin C_rated is determined by the ratio of the upper limit of the rated adjustability amplitude to the minimum adjustability step size in the device's file, reflecting the maximum number of adjustability steps that the device can execute under no time window constraints. For devices with a remaining duration greater than zero in the remaining time window sequence, the rated adjustability margin is additionally multiplied by the time window remaining penalty factor P_remain=1-T_remain / T_total, where T_remain is the remaining constraint duration and T_total is the total length of the shortest running constraint time window. The longer the remaining duration, the greater the penalty, and the lower the device's adjustability margin at the current moment. After correction, the adjustable margin of each device is output according to the device identifier. Each entry contains the corrected adjustable margin value and the corresponding remaining time window information. The adjustable margin of the HVAC system variable frequency chilled water pump is the highest when the shortest time window ends and the current operating condition is far from the boundary. Adjustment tasks are assigned first. When the total adjustable margin of the device is zero, the multi-device coordinated control command is delayed until the time window of the first device ends.

[0071] The system utilizes equipment adjustability margin and model-embedded constraint coefficients to adjust equipment adjustment weights and output multi-equipment coordinated control commands. Equipment adjustability margin determines the upper limit of each equipment's adjustment capacity at the current moment, while the model-embedded constraint coefficients determine the constraint penalty intensity that each equipment must adhere to during adjustment. Both together constrain the adjustment share undertaken by each equipment in coordinated control. The adjustment weight W_adj_i = C_adj_i / W_constraint_i, where W_adj_i is the adjustment share allocation weight of the i-th equipment, C_adj_i is the equipment adjustability margin of the i-th equipment, and W_constraint_i is the maximum value of W_final of all constraint entries for the corresponding equipment in the model-embedded constraint coefficients. Both are dimensionless values. Equipment with higher margins and lower constraint weights undertakes a larger share of the load response, while equipment with lower margins or higher constraint weights has limited adjustment range. In extreme operating conditions, the chiller exhaust temperature constraint has a very high weight in the embedded constraint coefficients of the HVAC system model. Consequently, the chiller regulation weight is significantly reduced, and the power regulation share is transferred to the chilled water pump and terminal air valves. This forms a constraint-aware multi-device collaborative regulation mode, effectively avoiding the risk of the chiller triggering hard constraint protection under extreme conditions. After the regulation weight of each device is normalized, it is multiplied by the total target regulation amount to obtain the allocated regulation amount. The allocated regulation amount is superimposed with the current state to generate the target control state. Combined with the timing constraints determined by the priority regulation device, a multi-device coordinated control command is formed. The multi-device coordinated control command uses the device identifier as the primary key. Each entry includes three fields: target control state, regulation range, and expected execution time. The expected execution time is calculated based on the response completion time of the priority regulation device and the response delay of other devices to avoid timing conflicts. When the full adjustability margin of the device is zero, the multi-device coordinated control command is delayed until the end of the first device's time window for re-evaluation and issuance.

[0072] To implement the above-described method embodiments, a hybrid model predictive control method for HVAC systems is proposed to achieve the corresponding functions and technical effects. See also... Figure 2 , Figure 2 This application provides a structural block diagram of a hybrid model predictive control device 200 for a heating, ventilation, and air conditioning system, comprising:

[0073] Data acquisition module 201 is used to acquire operating data of HVAC system and construct equipment coupling relationship matrix based on thermodynamic constraint coupling analysis of the operating data;

[0074] Model building module 202 is used to identify the controllable load response range based on the equipment coupling relationship matrix, extract soft and hard thermodynamic constraint parameters from the operating data to generate model embedded constraint coefficients, use the model embedded constraint coefficients to perform deep neural network modeling on the controllable load response range to form a thermodynamic constraint prediction model, and identify the insufficient operating condition coverage range based on the thermodynamic constraint prediction model to generate model failure identifiers.

[0075] Prediction optimization module 203 is used to perform rolling time-domain load prediction on the thermodynamic constraint prediction model based on the model failure identifier to generate a predicted load sequence, extract prediction confidence identifier from the predicted load sequence, perform response characteristic mapping through the prediction confidence identifier to obtain equipment response delay, and evaluate response priority based on the equipment response delay to generate a hierarchical equipment activation parameter group.

[0076] The coordination control module 204 is used to perform multi-device coordination constraint analysis on the hierarchical device activation parameter group to determine the priority adjustment device, detect the shortest operating constraint time window of the priority adjustment device, determine the device adjustability margin according to the shortest operating constraint time window, adjust the device adjustment weight using the device adjustability margin and the model embedded constraint coefficient, and output multi-device coordination control commands.

[0077] The aforementioned hybrid model predictive control device 200 for HVAC systems can implement a hybrid model predictive control method for HVAC systems as described in the above method embodiments. The options in the above method embodiments are also applicable to this embodiment and will not be detailed here. The remaining content of this application's embodiments can be referred to the content of the above method embodiments, and will not be repeated in this embodiment.

[0078] The purpose of the above embodiments is to reproduce and derive the technical solution of the present invention by way of example, and to fully describe the technical solution, purpose and effect of the present invention. The purpose is to enable the public to have a more thorough and comprehensive understanding of the disclosure of the present invention, and not to limit the scope of protection of the present invention.

Claims

1. A hybrid model predictive control method for HVAC systems, characterized in that, include: Collect operating data of the HVAC system, and construct a device coupling relationship matrix based on the operating data through thermodynamic constraint coupling analysis. Based on the equipment coupling relationship matrix, the controllable range of load response is identified. Soft and hard thermodynamic constraint parameters are extracted from the operating data to generate model embedded constraint coefficients. The model embedded constraint coefficients are used to perform deep neural network modeling on the controllable range of load response to form a thermodynamic constraint prediction model. Based on the thermodynamic constraint prediction model, the range of insufficient operating condition coverage is identified to generate a model failure identifier. Based on the model failure identifier, the thermodynamic constraint prediction model is subjected to rolling time-domain load prediction to generate a predicted load sequence. The prediction confidence identifier is extracted from the predicted load sequence. The response characteristic is mapped through the prediction confidence identifier to obtain the equipment response delay. The response priority is evaluated based on the equipment response delay to generate a hierarchical equipment activation parameter group. The multi-device coordination constraint analysis is performed on the hierarchical device activation parameter group to determine the priority adjustment device. The shortest operating constraint time window of the priority adjustment device is detected. The device adjustability margin is determined according to the shortest operating constraint time window. The device adjustment weight is adjusted using the device adjustability margin and the model embedded constraint coefficient, and the multi-device coordination control command is output.

2. The method according to claim 1, characterized in that, The step of extracting soft and hard thermodynamic constraint parameters from the operational data to generate model embedded constraint coefficients includes: Based on the operational data, constraint type identification is performed to generate soft constraint parameter sets and hard constraint parameter sets; Based on the soft constraint parameter set and the hard constraint parameter set, the constraint boundary spacing is calculated to generate the critical distance distribution of soft and hard constraints; The critical narrowing intervals are identified using the aforementioned soft and hard constraint critical distance distribution, generating a constraint compression risk set. By using the constraint compression risk set, differential weight calibration is performed on the soft constraint parameter set and the hard constraint parameter set to generate the model embedded constraint coefficients.

3. The method according to claim 1, characterized in that, The step of generating model failure identifiers based on the insufficient coverage of operating conditions identified by the thermodynamic constraint prediction model includes: The sample density of each operating condition interval is extracted from the thermodynamic constraint prediction model to generate an operating condition density sequence; Based on the operating condition density sequence, identify the density drop inflection point and generate a set of intervals covering the density drop; The risk of boundary failure is generated by jointly assessing the magnitude and location of the sudden drop in coverage density using the set of sudden drop intervals. The model failure identifier is generated by using the boundary failure risk degree calibration coverage of the insufficient interval.

4. The method according to claim 1, characterized in that, The step of extracting the prediction confidence identifier from the predicted load sequence includes: Based on the predicted load sequence, a sliding window prediction error calculation is performed to generate an error time series distribution; From the error time series distribution, identify the inflection points of sudden error increase to generate a set of confidence decay time points; Based on the set of confidence decay time points, the decay magnitudes are sorted to generate a confidence decay magnitude sequence; Based on the confidence decay magnitude sequence, high-risk periods are identified and predicted confidence indicators are generated.

5. The method according to claim 1, characterized in that, The step of generating a hierarchical device activation parameter group based on the device response latency assessment response priority includes: The device response delay is used to perform delay component decomposition and boundary joint partitioning to generate a delay boundary margin distribution. Based on the delay boundary margin distribution, identify devices with insufficient margin and generate a hierarchical boundary risk device set; Activation-level stability constraint analysis is performed on the set of hierarchical boundary risk devices to generate hierarchical constraint parameters; The hierarchical device activation parameter group is generated by evaluating the response priority of each device using the hierarchical constraint parameters.

6. The method according to claim 1, characterized in that, The step of performing multi-device coordination constraint analysis on the hierarchical device activation parameter group to determine the priority adjustment device includes: The constraint dependency graph is generated by extracting the constraint relationships of each device through the hierarchical device activation parameter group. The constraint dependency graph is used to identify conflict types and generate a capacity-based conflict set and a temporal conflict set. Based on the capacity-based conflict set and the temporal conflict set, a classification conflict elimination priority sort is generated to produce a classification conflict elimination priority sequence; The priority adjustment device is determined based on the classification conflict elimination priority sequence.

7. The method according to claim 1, characterized in that, The step of determining the device adjustability margin based on the shortest operating constraint window includes: Based on the shortest running constraint window, the remaining time windows of each device are statistically analyzed to generate a remaining time window sequence; Based on the remaining time window sequence, identify the overlapping intervals of adjacent handover request time windows and generate a time window overlap distribution; The time window overlap distribution is evaluated by group control adjustment margin reduction to generate margin reduction coefficients; The margin can be adjusted by using the margin reduction factor to modify the remaining time window sequence.

8. The method according to claim 3, characterized in that, The step of extracting sample densities for each operating condition interval from the thermodynamic constraint prediction model to generate an operating condition density sequence includes: Based on the aforementioned thermodynamic constraint prediction model, the operating condition coordinates of the training samples are extracted to generate an operating condition coordinate set; The sample spacing distribution is generated by calculating the spacing between adjacent samples using the aforementioned operating condition coordinate set; Based on the sample spacing distribution, identify the spacing abrupt expansion region and generate a set of covered blank intervals; The degree of coverage weakness is calibrated for the set of covered blank intervals to generate a working condition density sequence.

9. The method according to claim 5, characterized in that, The step of using the device response delay to perform delay component decomposition and joint boundary partitioning to generate a delay boundary margin distribution includes: The device response delay is decomposed into delay source components to generate inherent delay components and disturbance-induced delay components; The inherent delay component of the device is used to perform distribution clustering analysis to generate inherent delay cluster centers; Based on the perturbation-induced delay components, compressible delay boundaries are identified, and a set of compressible delay thresholds is generated. The inherent delay cluster centers and the compressible delay threshold set are used to jointly divide and generate a delay boundary margin distribution.

10. A hybrid model predictive control device for HVAC systems, characterized in that, include: The data acquisition module is used to collect the operating data of the HVAC system and to construct the equipment coupling relationship matrix based on the operating data through thermodynamic constraint coupling analysis. The model building module is used to identify the controllable load response range based on the equipment coupling relationship matrix, extract soft and hard thermodynamic constraint parameters from the operating data to generate model embedded constraint coefficients, use the model embedded constraint coefficients to perform deep neural network modeling on the controllable load response range to form a thermodynamic constraint prediction model, and identify the insufficient operating condition coverage range based on the thermodynamic constraint prediction model to generate model failure identifiers. The prediction optimization module is used to perform rolling time-domain load prediction on the thermodynamic constraint prediction model based on the model failure identifier to generate a predicted load sequence, extract prediction confidence identifiers from the predicted load sequence, perform response characteristic mapping through the prediction confidence identifiers to obtain equipment response delay, and evaluate response priority based on the equipment response delay to generate a hierarchical equipment activation parameter group. The coordination control module is used to perform multi-device coordination constraint analysis on the hierarchical device activation parameter group to determine the priority adjustment device, detect the shortest operating constraint time window of the priority adjustment device, determine the device adjustability margin according to the shortest operating constraint time window, adjust the device adjustment weight using the device adjustability margin and the model embedded constraint coefficient, and output multi-device coordination control commands.

Images