Mechanism data hybrid driven dynamic modeling method and device for cooling equipment

Abstract

This invention discloses a method and apparatus for dynamic modeling of cooling equipment driven by mechanistic data hybridization, belonging to the field of centralized cooling system modeling technology. The method establishes a differential-algebraic equation mechanistic model of the centralized cooling system equipment, setting the physical parameter vector as the parameters to be calibrated; it updates the posterior distribution of the physical parameter vector based on sliding window operating data to obtain the calibrated mechanistic model; it outputs residual correction through a neural network error compensator, performs posterior distribution update and residual correction in parallel, and superimposes the two outputs to obtain the equipment dynamic prediction result; further, it establishes a three-level model library at the second, minute, and hour levels, and publishes updated parameters through a unified parameter database. This invention can characterize physical parameter drift and structural residuals online and improve the consistency between models at different time scales.

Description

Technical Field

[0001] This invention relates to the field of centralized cooling system modeling and operation control technology, specifically to a dynamic modeling method and apparatus for cooling equipment driven by hybrid mechanism data, which can be used for dynamic prediction, control simulation and scheduling analysis of chillers, water pumps, cold storage devices and heat pumps in centralized cooling systems. Background Technology

[0002] Centralized cooling systems typically consist of chillers, chilled water pumps, cooling water pumps, cooling towers, cold storage devices, heat exchange equipment, distribution networks, and terminal equipment, providing centralized cooling to building complexes or industrial parks. To perform equipment control, system regulation, and scheduling optimization, it is necessary to establish a model that describes the dynamic relationships between equipment inputs, states, and outputs, and to use this model to predict operating parameters such as equipment cooling capacity, power consumption, flow rate, and supply and return water temperatures.

[0003] Existing centralized cooling system equipment models mainly include mechanistic models and data-driven models. Mechanistic models typically establish differential equations or differential-algebraic equations based on thermodynamics, heat transfer, and fluid mechanics, calculating equipment response through pre-set heat transfer coefficients, efficiency parameters, and hydraulic parameters. Data-driven models utilize historical operating data to train the mapping relationship between inputs and outputs. At different control levels, equipment models with different simulation step sizes and degrees of simplification can also be established.

[0004] For mechanistic models using fixed physical parameters, changes in heat transfer performance, mechanical efficiency, and hydraulic characteristics during equipment operation can cause actual physical parameters to drift relative to the initial parameters, which fixed parameters cannot reflect in a timely manner. Using parameter identification alone for correction is insufficient to fully characterize structural residuals not explicitly described by the mechanistic equations. While purely data-driven models can fit nonlinear residuals, their outputs lack the physical constraints provided by the mechanistic equations, making them prone to predictions that deviate from actual energy transfer patterns under operating conditions with insufficient sample coverage.

[0005] Therefore, the existing technology has at least the following problems: First, the slow drift of the equipment's physical parameters and the structural residuals of the mechanism model coexist, and the single-path model calibration or error fitting method is difficult to maintain the dynamic prediction accuracy of the equipment while maintaining physical interpretability; Second, the second-level, minute-level and hour-level models adopt different degrees of dynamic simplification, and it is difficult to eliminate the structural deviation caused by model simplification when only sharing parameter values, which can easily lead to inconsistent predictions of the same equipment's operating capabilities at different control levels. Summary of the Invention

[0006] The technical problem to be solved by this invention is: to provide a dynamic modeling method that can update the posterior distribution of physical parameters and the output residual correction amount separately, for situations where the physical parameters of centralized cooling system equipment drift during operation and the mechanism model has structural residuals; at the same time, to uniformly manage the parameters and structural deviations of models at different time scales, so as to improve the consistency of the representation of equipment operating capacity by second-level, minute-level and hour-level models.

[0007] To address the aforementioned technical problems, this invention provides a dynamic modeling method for cooling equipment driven by hybrid mechanism data, comprising:

[0008] A differential algebraic equation mechanism model of the centralized cooling system equipment is established, and the physical parameter vector of the mechanism model is set as the parameter to be calibrated.

[0009] Based on the sliding window running data and the residuals output by the mechanism model, a likelihood function is constructed, the posterior distribution of the physical parameter vector is updated, and the posterior mean parameter is taken to obtain the calibrated mechanism model.

[0010] The observables of unmodeled factors are input into the neural network error compensator, which outputs the residual correction.

[0011] The update of the posterior distribution and the output of the residual correction are run in parallel, and the output of the calibrated mechanism model is superimposed with the residual correction to obtain the dynamic prediction result of the equipment.

[0012] Based on the calibrated mechanism model, a three-level model library of seconds, minutes, and hours is established. The updated parameter values ​​of any level model are then published to the other two levels of models through a unified parameter database.

[0013] Further, the likelihood function is constructed based on the residuals of the sliding window running data and the output of the mechanism model, and the posterior distribution of the physical parameter vector is updated. The posterior mean parameter is then taken to obtain the calibrated mechanism model, including:

[0014] The prior distributions of the heat transfer coefficient component and the efficiency component in the physical parameter vector are fitted using sample data from equipment manufacturers. The prior distribution of the heat transfer coefficient component is a truncated normal distribution, and the prior distribution of the efficiency component is a Beta distribution.

[0015] Take minute-level sampling data within the sliding window, and construct the likelihood function based on the deviations of the supply and return water temperatures, equipment power, and chilled water flow rates output by the mechanism model relative to the measured values ​​in the minute-level sampling data;

[0016] The posterior distribution is obtained by using Markov chain Monte Carlo sampling, based on the prior distribution and the likelihood function.

[0017] The mean of the posterior distribution is taken as the value of the parameter to be calibrated and input into the mechanism model to obtain the calibrated mechanism model, and the covariance of the posterior distribution is retained for uncertainty propagation analysis.

[0018] The update of the posterior distribution is repeated by sliding in new sampled data at the right end and out old sampled data at the left end of the sliding window.

[0019] Furthermore, the step of employing Markov chain Monte Carlo sampling to obtain the posterior distribution based on the prior distribution and the likelihood function includes:

[0020] The physical parameter vector and the hidden initial values ​​of the mechanism model are augmented into a joint sampling object;

[0021] Based on the physical grouping of heat transfer parameters, mechanical efficiency parameters, and hydraulic parameters, block-based Metropolis-Hastings updates are performed on the joint sampling objects;

[0022] The empirical covariance of the accepted samples is used to propose the covariance within the block by scaling the Robbins-Monro attenuation coefficient online, so that the sampling acceptance rate is maintained within the preset acceptance rate range.

[0023] When the likelihood function can be reorganized into a linear Gaussian state-space form, Kalman filtering recursion is used instead of forward dynamic simulation to calculate the likelihood value.

[0024] Save the tail sample of the previous round of Markov chain as the starting point for sampling within the sliding window in the next round, and cancel repeated burning.

[0025] Furthermore, the method also includes:

[0026] After each round of updating the posterior distribution, a hold-out test is performed on the time segment of the running data that did not participate in the current round of updating.

[0027] When the posterior prediction coverage of the leaveout test deviates from the preset coverage range, the system reverts to the previous valid posterior distribution and re-executes the update of the posterior distribution after extending the length of the sliding window.

[0028] When the residual of the test is continuously horizontally displaced, the update of the posterior distribution is paused and a device mutation maintenance alarm is output.

[0029] Furthermore, the step of inputting the observables of the unmodeled factors into the neural network error compensator and outputting the residual correction includes:

[0030] The cumulative trend of the temperature difference between the inlet and outlet of the heat exchanger, the drift of the superheat of the compressor suction gas, the difference between the outdoor dry-bulb temperature and the dew point temperature, and the coordinates of the current operating point are used as observable measurements of the unmodeled factors and input into the neural network error compensator.

[0031] The residual correction amount for the cooling capacity, power consumption and chilled water outlet temperature of the calibrated mechanism model is output by the neural network error compensator.

[0032] The residual sequence formed by the difference between the measured value and the output of the calibrated mechanism model is used as the training target to perform supervised training on the neural network error compensator, and online incremental learning is performed as the running data accumulates.

[0033] Further, the residual correction amount output by the neural network error compensator for the cooling capacity, power consumption, and chilled water outlet temperature of the calibrated mechanism model includes:

[0034] The neural network error compensator outputs the correction rate for the thermal energy state and the cold energy state, and converts the correction rate into the residual correction amount by heat capacity.

[0035] A latent variable layer is set between the encoder and the output layer of the neural network error compensator. The dimension of the latent variable layer is smaller than the dimension of the encoder output, so that the residual correction amount is constrained within a preset set of disturbance modes.

[0036] The weight matrix of the output layer is maintained by Cayley transformation and synthesized into a semi-negative definite quadratic form to ensure that the power term corresponding to the residual correction satisfies the energy dissipation inequality.

[0037] An orthogonal projection penalty term for the posterior distribution mean response and a latent variable dimension sparse regularization term are added to the loss function of the neural network error compensator.

[0038] The residual correction amount is sequentially subjected to amplitude limiting and component filtering in the dissipation direction and then superimposed onto the output of the calibrated mechanism model. When the sample coverage at the operating point is lower than the preset coverage threshold, the residual correction amount is attenuated according to the proportion of the sample coverage.

[0039] Furthermore, the establishment of a differential algebraic equation mechanism model for the centralized cooling system equipment, and the setting of the physical parameter vector of the mechanism model as parameters to be calibrated, includes:

[0040] The centralized cooling system equipment includes chiller units, chilled water pumps, cooling water pumps, cold storage devices, and heat pump units;

[0041] The mechanism model of the chiller unit includes the compressor energy efficiency equation, the evaporator and condenser efficiency-heat transfer unit number method heat transfer equation, and the refrigerant mass flow balance equation.

[0042] The mechanism models of the chilled water pump and the cooling water pump include the head-flow characteristic curve equation, the shaft power equation, and the similarity law conversion relationship as a function of rotational speed.

[0043] The mechanism model of the cold storage device includes the lumped parameter energy balance differential equation of stratified water temperature and the cold charging and releasing characteristic equation.

[0044] The mechanism model of the heat pump unit includes a vapor compression cycle equation and is coupled with the heating and cooling switching logic of the four-way reversing valve.

[0045] For each type of equipment in the centralized cooling system, alternative mechanism structures with different complexities are constructed, and the alternative mechanism structure with the largest marginal likelihood is used as the structure of the mechanism model of the corresponding equipment.

[0046] Furthermore, the establishment of a three-level model library (second-level, minute-level, and hour-level) based on the calibrated mechanism model, and the dissemination of updated parameter values ​​from any level model to the other two levels models via a unified parameter database, includes:

[0047] The simulation step size of the second-level model is 0.1 seconds to 1 second, and the dynamic characteristics of the complete differential-algebraic equations of the calibrated mechanism model are preserved.

[0048] The simulation step size of the minute-level model is 1 to 5 minutes. Based on the second-level model, rigid fast variables are eliminated, while pipeline transmission and distribution delay and unit regulation inertia are retained.

[0049] The simulation step size of the hourly model is 1 hour. Steady-state simplification is performed based on the minute-level model, while retaining the capacity constraints of the cold storage device and the unit's energy efficiency characteristic curves.

[0050] The rated cooling capacity and efficiency parameters of the same device are stored as unique values ​​in the unified parameter database and are shared by the second-level model, the minute-level model and the hour-level model.

[0051] Furthermore, the method also includes:

[0052] The hourly model is used as the first fidelity model, the minute-level model as the second fidelity model, and the second-level model as the third fidelity model. The response outputs of the first fidelity model, the second fidelity model, and the third fidelity model are collected along the operating condition grid composed of chilled water inlet temperature, load rate, and outdoor dry-bulb temperature.

[0053] Using the response output of the first fidelity model as the prior mean, train a Gaussian process to differentiate between the second fidelity model and the first fidelity model;

[0054] Using the response output of the second fidelity model after scaling as the prior mean, a Gaussian process of difference between the third fidelity model and the second fidelity model is trained, and a noise term is registered for the observation variance.

[0055] Write the hyperparameters and training timestamps of the differential Gaussian process into the unified parameter database;

[0056] Based on the measured values, check whether the deviation between the response output of the first fidelity model after correction by the differential Gaussian process and the response output of the third fidelity model exceeds the preset quantile envelope. If it exceeds the preset quantile envelope, perform incremental retraining on the differential Gaussian process and output a cross-scale mismatch flag.

[0057] The present invention also provides a dynamic modeling device for cooling equipment driven by hybrid mechanism data, comprising:

[0058] The mechanism modeling module is used to establish the differential algebraic equation mechanism model of the centralized cooling system equipment, and to set the physical parameter vector of the mechanism model as the parameters to be calibrated.

[0059] The Bayesian calibration module is used to construct a likelihood function based on the sliding window running data and the residuals output by the mechanism model, update the posterior distribution of the physical parameter vector, take the posterior mean parameter, and obtain the calibrated mechanism model.

[0060] The residual compensation module is used to input the observables of unmodeled factors into the neural network error compensator and output the residual correction amount.

[0061] The dual-path overlay module is used to run the update of the posterior distribution and the output of the residual correction in parallel, and to overlay the output of the calibrated mechanism model with the residual correction to obtain the dynamic prediction result of the equipment.

[0062] The multi-scale management module is used to establish a three-level model library of seconds, minutes and hours based on the calibrated mechanism model, and to publish the updated parameter values ​​of any level model to the other two levels models through a unified parameter database.

[0063] By using a posterior distribution update path to track changes in the physical parameter vector and employing a neural network error compensator to characterize the structural residuals formed by unmodeled factors, the slow drift of physical parameters and residuals that cannot be directly expressed by the mechanistic model can be processed through different paths. After the two paths run in parallel, the output of the calibrated mechanistic model is superimposed with the constrained residual correction amount. This can reduce the deviation of the dynamic prediction results of the equipment as the equipment operating state changes while retaining the physical meaning of the mechanistic model, and can retain the constraint effect of the mechanistic model by attenuating the residual correction amount when the sample coverage is insufficient.

[0064] By using a unified parameter database to publish parameters across second-level, minute-level, and hour-level models, and by using the Gaussian process to register and correct structural deviations caused by model simplifications between models of different fidelity, models at all levels not only share the parameter values ​​of the same device, but also quantify, audit, and update the response differences caused by different degrees of simplification. This reduces decision-making biases caused by inconsistent model predictions between scheduling, regulation, and control levels. Attached Figure Description

[0065] Figure 1 A schematic diagram of the overall process of the mechanism data hybrid-driven dynamic modeling method for centralized cooling system equipment provided in an embodiment of the present invention;

[0066] Figure 2 This is a schematic diagram of the differential algebraic equation mechanism modeling process for centralized cooling system equipment provided in an embodiment of the present invention;

[0067] Figure 3 A schematic diagram of the online update and reliability verification process for the posterior distribution of physical parameters provided in this embodiment of the invention;

[0068] Figure 4 This is a schematic diagram of the residual compensation process for an energy dissipation-constrained neural network provided in an embodiment of the present invention;

[0069] Figure 5 This is a schematic diagram of the multi-timescale model establishment and cross-scale consistency alignment process provided in an embodiment of the present invention;

[0070] Figure 6 A schematic diagram of the module structure of the dynamic modeling device for centralized cooling system equipment driven by hybrid mechanism data provided in an embodiment of the present invention. Detailed Implementation

[0071] The present invention will be further described below with reference to the accompanying drawings and embodiments. It should be understood that the following embodiments are used to illustrate the technical solutions of the present invention, and are not intended to limit the scope of protection of the present invention. Where there is no conflict, the technical features of the various embodiments can be combined with each other.

[0072] In this embodiment, the centralized cooling system equipment can be a chiller unit, chilled water pump, cooling water pump, cold storage device, or heat pump unit. Operating data can be collected by equipment controllers, sensors, energy management systems, or a centralized cooling system monitoring platform, and includes at least supply and return water temperatures, equipment power, chilled water flow rate, valve opening degree, inverter frequency setpoint, and outdoor environmental parameters. A unified parameter database can be deployed on a local server, edge computing device, or cloud server.

[0073] Example 1

[0074] like Figure 1 As shown in the figure, this embodiment provides a dynamic modeling method for cooling equipment driven by hybrid mechanism data, which includes the following steps.

[0075] S1. Establish a differential algebraic equation mechanism model for the centralized cooling system equipment, and set the physical parameter vector of the mechanism model as the parameters to be calibrated;

[0076] The differential-algebraic equation mechanism model, constrained by thermodynamics, heat transfer, and fluid mechanics, expresses the relationship between equipment operating state, instantaneous quantities, control inputs, and physical parameters as follows:

[0077] (Equation 1)

[0078] (Equation 2)

[0079] in, Indicates time; Represents the differential state vector, including temperature, pressure, or energy storage states with dynamic inertia; Represents a vector of algebraic variables, including equipment power, flow rate, or instantaneous heat exchange; This represents the control input vector, including the valve opening degree and the inverter frequency setpoint; Represents a vector of physical parameters; Represents the differential state transition function; Represents algebraic constraint functions; This represents the zero vector with the same dimension as the algebraic constraint function.

[0080] The physical parameter vector may include the heat transfer coefficient of the heat exchanger, the isentropic efficiency of the compressor, the hydraulic efficiency of the water pump, the pipe roughness, and the equivalent heat loss coefficient of the cold storage device. Unlike methods that fix the above parameters to factory design values, this embodiment sets at least some physical parameters as calibration parameters that can be updated based on operating data.

[0081] For parameters that are not suitable for online calibration or are difficult to identify using existing measurement data, they can be kept as fixed parameters. The selection of parameters to be calibrated can be determined based on the results of sensitivity analysis, that is, selecting parameters that have a high degree of influence on cooling capacity, power consumption, chilled water outlet temperature, or chilled water flow rate and can be identified using existing measurement data.

[0082] S2. Construct a likelihood function based on the sliding window running data and the residual output of the mechanism model, update the posterior distribution of the physical parameter vector, take the posterior mean parameter, and obtain the calibrated mechanism model;

[0083] Record the running data within the sliding window as The posterior distribution of the physical parameter vector is calculated according to Bayes' rule:

[0084] (Equation 3)

[0085] in, This represents the posterior distribution of the physical parameter vector given the current sliding window's running data. Represents the likelihood function; Represents the prior distribution of the physical parameter vector; This represents the marginal probability of the data currently being processed by the sliding window.

[0086] The mean of the posterior distribution is used to form the calibration parameters:

[0087] (Equation 4)

[0088] in, This represents the posterior mean parameter corresponding to the current sliding window; This represents the allowed value space of the physical parameter vector.

[0089] Will Substituting the differential-algebraic equations of the mechanistic model shown in Equations 1 and 2, we obtain the calibrated mechanistic model. As the sliding window moves forward, the posterior mean parameters are continuously updated, enabling the mechanistic model to track the slow drift of the equipment's heat transfer performance, mechanical efficiency, and hydraulic characteristics.

[0090] S3. Input the observables of the unmodeled factors into the neural network error compensator and output the residual correction amount;

[0091] The observables of the unmodeled factors are used to characterize the influencing factors that are not directly described by differential algebraic equations but can be observed through running data. The neural network error compensator can employ a multilayer perceptron, recurrent neural network, temporal convolutional network, or a neural network with encoder and latent variable layers.

[0092] When training the neural network error compensator, the difference between the measured output and the output of the calibrated mechanistic model is used as the supervised training target. For input features with temporal correlation, the feature sequences at the current time and within a preset time period before the current time can be input into the neural network error compensator.

[0093] S4. Run the update of the posterior distribution and the output of the residual correction in parallel, and superimpose the output of the calibrated mechanism model with the residual correction to obtain the equipment dynamic prediction result;

[0094] The results of equipment dynamic prediction are expressed as follows:

[0095] (Equation 5)

[0096] in, express The dynamic prediction results of the equipment at any given time; This represents the output of the calibrated mechanistic model using the current posterior mean parameters; This represents the residual correction amount output by the neural network error compensator.

[0097] The Bayesian calibration pathway handles the slow drift of physical parameters such as heat transfer coefficient, compressor efficiency, and hydraulic parameters through posterior distribution updates, while the neural network compensation pathway handles structural residuals not explained by changes in physical parameters. By isolating the explanatory objects of the two pathways, the possibility of the neural network error compensator absorbing physical parameter drift and thus causing distortion in physical parameter identification can be reduced.

[0098] S5. Based on the calibrated mechanism model, establish a three-level model library at the second, minute, and hour levels, and publish the updated parameter values ​​of any level model to the other two levels models through a unified parameter database.

[0099] The second-level model is used for transient simulation of equipment and controller parameter tuning; the minute-level model is used for water supply temperature regulation, hydraulic balance adjustment and equipment combination control; and the hour-level model is used for day-ahead scheduling, cold storage planning and demand response analysis.

[0100] The unified parameter database establishes parameter records according to equipment identifier, parameter name, parameter value, parameter unit, parameter source, update time, applicable operating conditions, and uncertainty. After any level model updates the parameter value, the updated parameter, along with the timestamp and confidence information, is written to the unified parameter database; the other two levels of models read the corresponding parameters and complete the synchronization, so that models at different time scales use the same equipment capability parameters.

[0101] Because the posterior update of physical parameters and the residual compensation of the neural network operate in parallel, physical parameter drift and structural residuals are handled by their respective pathways. This ensures that the dynamic prediction results of the equipment retain the physical meaning of the mechanistic model while being continuously corrected based on operational data. Since the three-level model library shares a unified parameter database, the characterization of the equipment's rated cooling capacity, efficiency, and availability remains consistent across different control levels.

[0102] Example 2

[0103] like Figure 3 As shown, this embodiment further defines the posterior distribution update process of the physical parameter vector based on embodiment 1.

[0104] S2.1. Fit the prior distributions of the heat transfer coefficient component and the efficiency component in the physical parameter vector using sample data from the equipment manufacturer. The prior distribution of the heat transfer coefficient component is a truncated normal distribution, and the prior distribution of the efficiency component is a Beta distribution.

[0105] Equipment manufacturer sample data can include factory performance test data, variable operating condition performance data output by selection software, and baseline operating condition data collected during the commissioning phase. For heat transfer coefficient components with clear upper and lower limits and concentrated around the nominal value, using a truncated normal distribution can avoid sampling negative heat transfer coefficients or values ​​that significantly exceed the allowable range of the equipment structure. For efficiency components with values ​​between zero and one, using a Beta distribution can describe different forms of prior probability density within the parameter domain.

[0106] When setting up prior distributions, similar equipment can be grouped according to equipment model, rated capacity, cumulative operating time, and maintenance status, so that the same equipment category adopts the same prior distribution type, and corresponding distribution hyperparameters can be set for different equipment instances.

[0107] S2.2. Take the minute-level sampling data within the sliding window, and construct the likelihood function based on the deviations of the supply and return water temperature, equipment power, and chilled water flow rate output by the mechanism model relative to the measured values ​​in the minute-level sampling data;

[0108] In one implementation, the likelihood function takes the form of a multivariable Gaussian residual:

[0109] (Equation 6)

[0110] in, This represents the likelihood function value corresponding to the current sliding window; This indicates the number of valid sampling moments within the current sliding window; Indicates the first The residual vector at each valid sampling time; The covariance matrix representing the measurement residuals; The dimension of the residual vector; The determinant of the measurement residual covariance matrix; Indicates the sequence number of the valid sampling time.

[0111] The residual vector is constructed by subtracting the corresponding measured values ​​from the supply water temperature, return water temperature, equipment power, and chilled water flow rate output by the mechanistic model. Before forming the likelihood function, time alignment, outlier removal, missing value marking, and dimensional normalization can be performed on the minute-level sampled data. Data determined to be from invalid sensors or equipment shutdowns are not included in the parameter updates for the current window.

[0112] S2.3. Markov chain Monte Carlo sampling is used to obtain the posterior distribution based on the prior distribution and the likelihood function;

[0113] Metropolis-Hastings sampling, block Metropolis-Hastings sampling, or other Markov chain Monte Carlo methods that can generate posterior samples based on prior distributions and likelihood functions can be used. During the sampling process, constraints are imposed on the physical parameters, determined by the equipment structure and operating boundaries, to prevent the sampling results from deviating from the physically feasible region.

[0114] S2.4. Take the mean of the posterior distribution as the value of the parameter to be calibrated and input it into the mechanism model to obtain the calibrated mechanism model, and retain the covariance of the posterior distribution for uncertainty propagation analysis;

[0115] The posterior mean parameter is used to perform dynamic predictions of the equipment, while the posterior covariance is used to describe the degree of uncertainty in the current parameter perception. Multiple sets of physical parameters can be extracted from the posterior distribution, and mechanistic models can be run on each set separately. The prediction confidence interval is formed by the sample distribution of the equipment output.

[0116] When the posterior covariance increases, it indicates that the current operating data has a reduced ability to constrain the relevant parameters. Upper-level control or scheduling programs can reduce the weighting of the high-uncertainty model output based on the predicted confidence interval width.

[0117] S2.5. The update of the posterior distribution is repeated by sliding in new sampled data at the right end and out old sampled data at the left end of the sliding window.

[0118] The sliding window length can be determined based on the device's dynamic time constant, parameter drift rate, and sampling frequency. For example, a sliding window with minute-level sampling data can cover multiple consecutive operating cycles, ensuring the window contains effective data for the device under different load rates. If the window is too short, the data's ability to identify the parameters to be calibrated is insufficient; if the window is too long, earlier data may weaken the model's ability to track recent parameter changes. Therefore, the window length can be adaptively adjusted based on the posterior variance and hold-out test results.

[0119] Example 3

[0120] This embodiment further defines the Markov chain Monte Carlo sampling process based on Embodiment 2.

[0121] S2.3.1. Augment the physical parameter vector and the hidden state initial values ​​of the mechanism model into a joint sampling object;

[0122] For the internal state of the equipment that cannot be directly measured at the starting point of the sliding window, it is used as the initial value of the hidden state and jointly sampled with the physical parameter vector. This can reduce the systematic impact of the initial value setting error of the hidden state on the posterior distribution of the physical parameters.

[0123] S2.3.2. Perform block-based Metropolis-Hastings updates on the joint sampling objects according to the physical grouping of heat transfer parameters, mechanical efficiency parameters and hydraulic parameters;

[0124] Heat transfer parameters can include evaporator heat transfer coefficient, condenser heat transfer coefficient, and equivalent heat loss coefficient of cold storage device; mechanical efficiency parameters can include compressor isentropic efficiency and motor efficiency; hydraulic parameters can include pump characteristic curve coefficient and pipeline equivalent resistance coefficient. Block updates enable highly correlated parameters to be jointly proposed within the same parameter block, reducing the low acceptance rate problem caused by overall proposals of high-dimensional parameters.

[0125] S2.3.3. Based on the empirical covariance of the accepted samples, the proposed covariance is scaled online according to the Robbins-Monro attenuation coefficient to keep the sampling acceptance rate within the preset acceptance rate range;

[0126] The intra-block proposal covariance can be updated according to the following formula:

[0127] (Equation 7)

[0128] in, Indicates the first The parameter block in the first... The proposed covariance used in the next iteration; Indicates the first The parameter block in the first... Online scaling factor during the next iteration; Indicates as of the date The empirical covariance of the accepted samples in the next iteration; This represents a small positive number used to guarantee the positive definiteness of the proposed covariance; Indicates the relationship with the first The identity matrix corresponding to the dimension of each parameter block; Indicates the parameter block number; This indicates the sampling iteration number.

[0129] The online scaling factor can be updated according to the difference between the current acceptance rate and the target acceptance rate using a Robbins-Monro decay step size. The preset acceptance rate range can be set based on the parameter block dimension, for example, from 0.234 to 0.44. When the acceptance rate is below the preset acceptance rate range, the proposal step size is decreased; when the acceptance rate is above the preset acceptance rate range, the proposal step size is increased.

[0130] S2.3.4. When the likelihood function can be rearranged into a linear Gaussian state-space form, Kalman filtering recursion is used instead of forward dynamic simulation to calculate the likelihood value;

[0131] For lumped parameter models of cold storage devices or models that satisfy the linear Gaussian condition after local linearization, Kalman filtering is used to sequentially perform state prediction, measurement update, and prediction error covariance update. The innovation quantity and innovation covariance at each sampling time are accumulated to form a likelihood value, avoiding the need to perform a complete nonlinear dynamic simulation for each parameter proposal.

[0132] When the linear Gaussian condition is not met, the likelihood value can be calculated by numerical integration of differential-algebraic equations, or by approximation methods such as extended Kalman filtering or unscented Kalman filtering.

[0133] S2.3.5. Save the tail sample of the previous round of Markov chain as the starting point for sampling within the sliding window in the next round, and cancel repeated burning.

[0134] Since most of the running data in adjacent sliding windows overlaps, the posterior distribution of physical parameters usually changes continuously. Using the tail sample of the previous round of the Markov chain as the sampling starting point for the next round allows the next round of the Markov chain to start running from a position close to the target posterior distribution, thereby reducing the computational overhead caused by repeated burning.

[0135] Example 4

[0136] This embodiment, based on embodiment 2, sets up a reliability test for the posterior distribution update result and a mechanism for interlocking mutation events, including the following steps.

[0137] S2.6. After each round of updating the posterior distribution, perform a hold-out test on the time segment of the running data that did not participate in the current round of updating;

[0138] The operational data time segment can be continuous sampling data formed after the right end of the current sliding window, or it can be data reserved from the current calibration period that has not participated in the posterior distribution update. The predicted distribution of each operational data time segment is calculated based on the posterior samples, and the proportion of measured values ​​falling into the posterior prediction interval is statistically analyzed.

[0139] S2.7. When the posterior prediction coverage of the leaveout test deviates from the preset coverage range, revert to the previous valid posterior distribution, and re-execute the update of the posterior distribution after extending the length of the sliding window;

[0140] If the posterior prediction coverage is lower than the preset coverage range, it may indicate that the current window data volume is insufficient, the operating condition coverage is insufficient, or the sampling has not yet converged; if the posterior prediction coverage is abnormally higher than the preset coverage range, it may indicate that the posterior distribution is too wide and cannot provide effective parameter constraints.

[0141] During rollback, the unified parameter database retains the posterior mean and posterior covariance parameters that passed the hold-out test in the previous round, preventing unverified parameters from entering the equipment control, system adjustment, or scheduling optimization process. After extending the sliding window, the likelihood function is reconstructed and the posterior distribution is updated.

[0142] S2.8. When the residual of the test is continuously horizontally shifted, pause the update of the posterior distribution and output a device mutation maintenance alarm.

[0143] Continuous horizontal displacement refers to the situation where the residual values ​​remain in the same direction of change across multiple consecutive sampling times, and the mean residual value deviates from the historical baseline by more than a preset threshold. This phenomenon differs from an increase in random noise and may correspond to a sudden change in the equipment's structural state or operating conditions.

[0144] Pausing posterior distribution updates can prevent abrupt changes in data from being misinterpreted as slow-varying parameter drift. Equipment abrupt change maintenance alarms can include equipment identifier, abrupt change time, affected output, residual offset direction, and offset magnitude, allowing operators to check equipment and sensor status. Posterior distribution updates should only resume after the cause of the abrupt change has been eliminated or the mechanistic model structure has been redefined.

[0145] Example 5

[0146] This embodiment, based on embodiment 1, further defines the input, output, and training methods of the neural network error compensator.

[0147] S3.1. The cumulative trend of the temperature difference between the inlet and outlet of the heat exchanger, the drift of the superheat of the compressor suction gas, the difference between the outdoor dry-bulb temperature and the dew point temperature, and the coordinates of the current operating point are taken as observable measurements of the unmodeled factors and input into the neural network error compensator.

[0148] The cumulative trend of the temperature difference between the inlet and outlet of the heat exchanger is used to characterize the change in heat exchange performance over time; the drift of the compressor suction superheat is used to characterize the change in the refrigerant circulation state; the difference between the outdoor dry-bulb temperature and the dew point temperature is used to characterize the influence of ambient humidity; the coordinates of the current operating point can be composed of at least two of the following: load rate, chilled water inlet temperature, cooling water inlet temperature, flow rate, and equipment frequency.

[0149] Each observable can be standardized using the mean and standard deviation of the training dataset before being input into the neural network error compensator. The cumulative trend can be obtained using moving averages, exponential smoothing, or trend decomposition methods.

[0150] S3.2. The residual correction amount of the cooling capacity, power consumption and chilled water outlet temperature of the calibrated mechanism model is output by the neural network error compensator.

[0151] The neural network error compensator is configured with output nodes corresponding to cooling capacity, power consumption, and chilled water outlet temperature, respectively. Each output node generates a residual correction amount with the same unit as the output quantity of the calibrated mechanistic model, or with the same unit after inverse standardization.

[0152] S3.3. Using the residual sequence formed by the difference between the measured value and the output of the calibrated mechanism model as the training target, supervised training is performed on the neural network error compensator, and online incremental learning is performed as the running data accumulates.

[0153] The target residuals of the training samples are generated using the posterior mean parameter, which has already passed the hold-out test, to avoid unreliable parameter updates contaminating the neural network training objective. Online incremental learning can be performed at fixed time intervals or triggered when the statistical characteristics of the residuals change significantly.

[0154] To avoid frequent changes in network parameters due to short-term noise, the data used for online incremental learning can consist of both recent samples and historical representative samples, with upper limits set for the learning rate and the magnitude of a single parameter update. In other implementations, the neural network error compensator can be updated offline periodically, and the updated network parameters can be put into operation after being verified with independent validation data.

[0155] Example 6

[0156] like Figure 4As shown, this embodiment applies an energy dissipation constraint to the residual correction amount based on embodiment 5.

[0157] S3.2.1. The neural network error compensator outputs the correction rate of the thermal energy storage state and the cold energy storage state, and converts the correction rate into the residual correction amount by heat capacity;

[0158] The neural network error compensator first outputs the time rate of change of the energy state, rather than directly outputting an unconstrained temperature or power correction. For an equivalent heat capacity of... For thermal storage objects, the temperature residual correction rate can be expressed as:

[0159] (Equation 8)

[0160] in, Indicates the temperature residual correction rate; This represents the internal energy state correction rate output by the neural network error compensator; This represents the equivalent heat capacity of the object being modified.

[0161] The residual corrections for cooling capacity and power consumption can be obtained based on the power balance relationship of the corresponding energy channels. By linking the residual corrections to the energy state change process, the energy direction of the neural network error compensator can be constrained.

[0162] S3.2.2. A latent variable layer is set between the encoder and the output layer of the neural network error compensator. The dimension of the latent variable layer is smaller than the dimension of the encoder output, so that the residual correction amount is constrained within a preset set of disturbance modes.

[0163] The encoder maps observables of unmodeled factors to high-dimensional features, and the latent variable layer compresses these high-dimensional features into low-dimensional perturbation representations. The set of perturbation modes may include heat transfer resistance variation modes, refrigerant cycle state variation modes, and environmental load modes.

[0164] The dimension of the latent variable layer can be determined by the residual fitting error on the validation data and the model complexity. If the dimension of the latent variable layer is too small, effective perturbation information will be lost, while if the dimension is too large, the neural network may express high-frequency non-physical residuals outside the training samples.

[0165] S3.2.3. The weight matrix of the output layer is maintained orthogonally by Cayley transformation and synthesized into a semi-negative definite quadratic form, so that the power term corresponding to the residual correction satisfies the energy dissipation inequality;

[0166] Perform a Cayley transformation on the antisymmetric parameter matrix to obtain an orthogonal weight matrix. The energy dissipation constraint can be expressed as:

[0167] (Equation 9)

[0168] in, This represents the rate of energy change corresponding to the residual compensation path; This represents the perturbation mode vector output by the latent variable layer; Represents a positive semidefinite dissipation matrix; This represents the energy input rate resulting from heat transfer at a known boundary. Because... Since it is a non-negative quantity, the residual compensation path cannot generate positive net energy in the absence of known boundary energy input, thereby reducing the possibility of energy constraint in the residual correction failure mechanism model.

[0169] S3.2.4. Add an orthogonal projection penalty term for the posterior distribution mean response and a latent variable dimension sparse regularization term to the loss function of the neural network error compensator;

[0170] The training loss can be expressed as:

[0171] (Equation 10)

[0172] in, This represents the total training loss of the neural network error compensator; This represents the residual fitting loss; This indicates that the energy dissipation inequality violates the penalty term; This represents the orthogonal projection penalty term for the response to the mean of the posterior distribution; This indicates a sparse regularization term for the latent variable dimensions; Indicates the weight of the residual fitting loss; Indicates the weight of the dissipation constraint; Indicates the orthogonal projection penalty weight; This represents the sparse regularization weight.

[0173] The orthogonal projection penalty term is used to reduce the correlation between the neural network output and the model response caused by changes in the posterior mean of the physical parameters, preventing the slowly varying parameter effects that can be explained by the Bayesian calibration pathway from being repeatedly absorbed by the neural network. During the online incremental learning phase, the encoder can be frozen, and only the dissipative output layer after the latent variable layer can be adjusted to reduce the drift of perturbation modes with short-term running data.

[0174] S3.2.5. The residual correction amount is sequentially subjected to amplitude limiting and component filtering in the dissipation direction and then superimposed on the output of the calibrated mechanism model. When the sample coverage at the operating point is lower than the preset coverage threshold, the residual correction amount is attenuated according to the proportion of the sample coverage.

[0175] When applying amplitude limiting to the residual correction, an upper limit can be set based on the rated value of the corresponding output, historical residual quantiles, or the allowable deviation range of the mechanistic model. Direction filtering is used to remove residual components that cause non-dissipative energy changes.

[0176] The residual correction amount after safety processing can be expressed as:

[0177] (Equation 11)

[0178] in, This indicates the residual correction amount after safety processing; This represents the attenuation coefficient determined by the sample coverage, and its value ranges from zero to one. This represents the projection operator that performs component filtering in the direction of dissipation; This represents the amplitude limiting function; This represents the raw residual output of the neural network error compensator; This represents the upper limit vector of the magnitude of each residual output component.

[0179] When the sample coverage is not lower than the preset coverage threshold, One can be chosen; when the sample coverage is lower than the preset coverage threshold, It decreases as coverage decreases; when the operating point is significantly outside the coverage range of the training samples. It can be set to zero, causing the dynamic prediction results of the equipment to degenerate into the output of the calibrated mechanism model.

[0180] Example 7

[0181] like Figure 2 As shown, this embodiment further defines the centralized cooling system equipment and its differential algebraic equation mechanism model based on embodiment 1.

[0182] S1.1. The centralized cooling system equipment includes chiller units, chilled water pumps, cooling water pumps, cold storage devices, and heat pump units;

[0183] S1.2. The mechanism model of the chiller unit includes the compressor energy efficiency equation, the evaporator and condenser efficiency-heat transfer unit number method heat transfer equation, and the refrigerant mass flow balance equation.

[0184] In one implementation, the compressor efficiency equation calculates the compressor power based on the evaporation temperature, condensation temperature, compressor frequency, and isentropic efficiency. The evaporator and condenser efficiency-number-of-units heat transfer equation calculates the heat transfer based on the heat exchanger heat transfer coefficient, heat transfer area, capacity factor on both sides, and inlet temperature. The refrigerant mass flow balance equation is used to constrain the refrigerant circulation mass flow rate between the compressor, evaporator, and condenser.

[0185] S1.3. The mechanism model of the chilled water pump and the cooling water pump includes the head-flow characteristic curve equation, the shaft power equation, and the similarity law conversion relationship as the speed changes;

[0186] The pump model can obtain the head, flow rate, and shaft power at the current speed based on the head-flow characteristic curve at the reference speed. The pump efficiency can be used as a parameter to be calibrated, and its posterior distribution can be updated based on real-time power, flow rate, and differential pressure data.

[0187] S1.4. The mechanism model of the cold storage device includes the lumped parameter energy balance differential equation of stratified water temperature and the cold charging and releasing characteristic equation;

[0188] For water-based cold storage devices, the cold storage tank can be divided into multiple temperature layers along its height, and energy balance differential equations can be established for each layer. For ice-based cold storage devices, charging and releasing characteristic equations can be established based on the ice quantity, heat exchange flow rate, and inlet and outlet temperatures.

[0189] S1.5. The mechanism model of the heat pump unit includes a vapor compression cycle equation and is coupled with the heating and cooling switching logic of the four-way reversing valve;

[0190] The state of the four-way reversing valve is used to determine the switching of heat exchange functions between the evaporator and condenser. During the switching process, the continuity of pressure, temperature, and heat exchanger thermal inertia states can be maintained to describe the transient response when switching operating modes.

[0191] S1.6. For each type of equipment in the centralized cooling system, construct alternative mechanism structures with different complexities, and use the alternative mechanism structure with the largest marginal likelihood as the structure of the mechanism model of the corresponding equipment.

[0192] For similar devices, low-order lumped parameter structures, mid-order gray-box structures, and complete differential-algebraic equation structures can be constructed. The marginal likelihood of each candidate mechanism structure is calculated using benchmark data, and both model fitting ability and model complexity are considered in the comparison. The candidate mechanism structure with the highest marginal likelihood is selected as the default mechanism model structure for the corresponding device.

[0193] Performing structural selection and physical parameter calibration separately can reduce the possibility of systematic shifts in the posterior distribution of physical parameters due to model-structure mismatch. In other implementations, alternative mechanism structures can be selected by comprehensively considering marginal likelihood, computational delay, and the real-time requirements of the target control level. Example 8

[0194] This embodiment, based on embodiment 1, further defines the three-level model library of second-level, minute-level, and hour-level models.

[0195] S5.1. The simulation step size of the second-level model is 0.1 seconds to 1 second, and the dynamic characteristics of the complete differential-algebraic equations of the calibrated mechanism model are preserved;

[0196] The second-level model preserves the dynamic changes in equipment temperature, pressure, flow rate, and energy storage status, and is used for equipment start-up and shutdown sequence simulation, valve action response analysis, pump frequency conversion control, and controller parameter tuning. The simulation step size can be selected from 0.1 seconds to 1 second based on the equipment's minimum dynamic time constant and computational resources.

[0197] S5.2. The simulation step size of the minute-level model is 1 minute to 5 minutes. Based on the second-level model, rigid fast variables are eliminated, while pipeline transmission and distribution delay and unit regulation inertia are retained.

[0198] For state variables whose rate of change is significantly faster than the minute-level adjustment cycle, quasi-steady-state approximation or algebraic elimination methods can be used to eliminate them. The network transmission and distribution delay and the unit's adjustment inertia affect the minute-level supply water temperature and flow response, and therefore are retained. The minute-level model can be used for system loop hydraulic balance, supply water temperature resetting, and unit load distribution.

[0199] S5.3. The simulation step size of the hour-level model is 1 hour. Steady-state simplification is performed based on the minute-level model, while retaining the capacity constraints of the cold storage device and the unit energy efficiency characteristic curves.

[0200] The hourly model sets the rapid dynamic state of the equipment to a steady-state relationship, while preserving the accumulated energy storage status of the cold storage device across hours, the power constraints of charging and discharging, and the characteristics of unit energy efficiency varying with load rate. This model can be used for day-ahead scheduling, peak-valley load transfer, and demand response simulation.

[0201] S5.4. The rated cooling capacity and efficiency parameters of the same device in the unified parameter database are stored as unique values ​​and are shared by the second-level model, the minute-level model and the hour-level model.

[0202] The unified parameter database links the three-level models through unique device identifiers. Only one currently valid record is maintained for the rated cooling capacity, efficiency parameters, and other shared physical parameters of the same device. When parameters are updated, historical versions, update times, and update sources are retained, enabling the three-level models to read the same currently valid parameters and trace the parameter change process.

[0203] Example 9

[0204] like Figure 5 As shown, this embodiment, based on embodiment 8, utilizes a differential Gaussian process to perform consistent alignment of structural deviations in the three-level model.

[0205] S5.5. The hourly model is used as the first fidelity model, the minute-level model as the second fidelity model, and the second-level model as the third fidelity model. The response outputs of the first fidelity model, the second fidelity model, and the third fidelity model are collected along the operating condition grid composed of chilled water inlet temperature, load rate, and outdoor dry-bulb temperature.

[0206] The operating condition grid can be a sparse Cartesian grid or a non-uniform grid formed based on the historical operating condition distribution. For operating condition areas with high operating frequency or large output change gradients, denser sampling points can be set; for less frequently occurring edge operating condition areas, sampling points can be reserved for model extrapolation verification.

[0207] The response output of the Level 3 model should include at least the cooling capacity, coefficient of performance, and chilled water outlet temperature. When collecting the response output, the Level 3 model should use uniform equipment parameters and consistent boundary conditions so that the response differences mainly reflect the structural deviations caused by model simplification.

[0208] S5.6. Using the response output of the first fidelity model as the prior mean, train a Gaussian process to differentiate between the second fidelity model and the first fidelity model;

[0209] The relationship between the first fidelity model and the second fidelity model can be expressed as:

[0210] (Equation 12)

[0211] in, This represents the operating condition vector consisting of chilled water inlet temperature, load rate, and outdoor dry-bulb temperature. This indicates that the first-fidelity model is in the working condition vector. The response output at the location; This indicates that the second fidelity model is in the working condition vector. The response output at the location; This represents the scaling factor used to upgrade the first fidelity model to the second fidelity model; This represents a Gaussian process that represents the difference between the second fidelity model and the scaled first fidelity model.

[0212] S5.7. Using the response output of the second fidelity model after scaling as the prior mean, train the difference Gaussian process between the third fidelity model and the second fidelity model, and register a noise term for the observation variance;

[0213] The relationship between the second fidelity model and the third fidelity model can be expressed as:

[0214] (Equation 13)

[0215] in, This indicates that the third fidelity model is in the working condition vector. The response output at the location; This represents the scaling factor used to improve the fidelity of the model from the second to the third fidelity. This represents a Gaussian process that indicates the difference between the third-fidelity model and the scaled second-fidelity model. The observation noise of the Gaussian process is represented by the variance of the observation noise, which is written as a noise term into the covariance function of the Gaussian process.

[0216] The scaling factor, kernel length scale, kernel amplitude, and noise variance can be solved using the maximum likelihood method. In addition to outputting the corrected response mean, the difference Gaussian process also outputs the corresponding prediction variance to indicate the reliability of the cross-scale correction results.

[0217] S5.8. Write the hyperparameters and training timestamps of the differential Gaussian process into the unified parameter database;

[0218] In addition to storing the physical parameters of the storage devices, the unified parameter database also includes a structure bias register. The structure bias register can include the output variable name, fidelity level, scaling factor, Gaussian process kernel function type, kernel function hyperparameters, noise variance, training data version, and training timestamp.

[0219] After any substantial update to the structure or parameters of the first-level model, the response output on the standard working condition mesh can be re-acquired, and the corresponding differential Gaussian process and its structural deviation register can be refreshed.

[0220] S5.9. Based on the measured value, check whether the deviation between the response output of the first fidelity model after correction by the difference Gaussian process and the response output of the third fidelity model exceeds the preset quantile envelope. If it exceeds the preset quantile envelope, perform incremental retraining on the difference Gaussian process and output a cross-scale mismatch flag.

[0221] Consistency audits can be performed according to a preset cycle. First, the device response is quickly calculated using the first-fidelity model, and then corrected sequentially using a difference Gaussian process to obtain a response estimate close to the third-fidelity model. Then, this response estimate, the output of the third-fidelity model, and the measured value are compared.

[0222] When the deviation exceeds the preset quantile envelope corresponding to the historical deviation distribution, it indicates that the existing differential Gaussian process cannot adequately describe the structural deviation between the current three-level models. The system collects new operating condition samples to perform incremental retraining and outputs cross-scale mismatch markers to the upper-level control or scheduler. Before the differential Gaussian process passes re-validation, the weight of the response estimate obtained after cross-scale correction in the decision-making process can be reduced.

[0223] Example 10

[0224] like Figure 6 As shown, this embodiment provides a dynamic modeling device for cooling equipment driven by hybrid mechanism data, including a mechanism modeling module, a Bayesian calibration module, a residual compensation module, a dual-path overlay module, and a multi-scale management module.

[0225] The mechanism modeling module is used to establish the differential algebraic equation mechanism model of the centralized cooling system equipment, and to set the physical parameter vector of the mechanism model as the parameters to be calibrated.

[0226] The mechanism modeling module can call the corresponding equipment model template according to the equipment type, configure state variables, algebraic variables, control inputs, physical parameters and boundary conditions, and generate a differential algebraic equation mechanism model that can be executed by a numerical solver. For equipment with multiple alternative mechanism structures, the mechanism modeling module can also select the appropriate model structure based on the marginal likelihood.

[0227] The Bayesian calibration module is used to construct a likelihood function based on the sliding window running data and the residuals output by the mechanism model, update the posterior distribution of the physical parameter vector, take the posterior mean parameter, and obtain the calibrated mechanism model.

[0228] The Bayesian calibration module may include a data window management unit, a prior distribution setting unit, a likelihood calculation unit, a posterior sampling unit, and a reliability testing unit. The data window management unit is used to update the sliding window; the posterior sampling unit is used to perform block Metropolis-Hastings updates; and the reliability testing unit is used to perform hold-out tests and posterior distribution rollback.

[0229] The residual compensation module is used to input the observables of unmodeled factors into the neural network error compensator and output the residual correction amount.

[0230] The residual compensation module may include a feature processing unit, an encoder, a latent variable layer, a dissipation output layer, and a secure projection unit. The secure projection unit is used to perform amplitude limiting, dissipation direction filtering, and sample coverage attenuation on the original residual output.

[0231] The dual-path overlay module is used to run the update of the posterior distribution and the output of the residual correction in parallel, and to overlay the output of the calibrated mechanism model with the residual correction to obtain the dynamic prediction result of the equipment.

[0232] The dual-path overlay module receives the posterior mean parameter output from the Bayesian calibration module and the residual correction amount output from the residual compensation module, respectively. It then inputs the posterior mean parameter into the mechanistic model and overlays the model output with the safety-processed residual correction amount. The dual-path overlay module can also output the posterior covariance, neural network sample coverage, and the confidence level of the device's dynamic prediction results.

[0233] The multi-scale management module is used to establish a three-level model library of seconds, minutes and hours based on the calibrated mechanism model, and to publish the updated parameter values ​​of any level model to the other two levels models through a unified parameter database.

[0234] The multi-scale management module may include a three-level model management unit, a unified parameter database, a structural deviation register, and a cross-scale consistency audit unit. The three-level model management unit stores or calls second-level, minute-level, and hour-level models respectively; the structural deviation register stores the hyperparameters of the differential Gaussian process; and the cross-scale consistency audit unit is used to determine whether the cross-scale response deviation after correction by the differential Gaussian process exceeds the preset quantile envelope.

[0235] The modules described above can be implemented by the processor executing computer programs stored in memory, or they can be implemented using software, hardware, or a combination of both. The division between modules is a functional division. In actual deployment, multiple modules can be implemented by the same processor, or a single module can be implemented collaboratively by multiple processors, edge computing nodes, or servers.

Claims

1. A dynamic modeling method for cooling equipment driven by hybrid mechanism data, characterized in that, include: A differential algebraic equation mechanism model of the centralized cooling system equipment is established, and the physical parameter vector of the mechanism model is set as the parameter to be calibrated. Based on the sliding window running data and the residuals output by the mechanism model, a likelihood function is constructed, the posterior distribution of the physical parameter vector is updated, and the posterior mean parameter is taken to obtain the calibrated mechanism model. The observables of unmodeled factors are input into the neural network error compensator, which outputs the residual correction. The update of the posterior distribution and the output of the residual correction are run in parallel, and the output of the calibrated mechanism model is superimposed with the residual correction to obtain the dynamic prediction result of the equipment. Based on the calibrated mechanism model, a three-level model library of seconds, minutes, and hours is established. The updated parameter values ​​of any level model are then published to the other two levels of models through a unified parameter database.

2. The method according to claim 1, characterized in that, The likelihood function is constructed based on the residuals of the sliding window running data and the output of the mechanism model. This is used to update the posterior distribution of the physical parameter vector. The posterior mean parameter is then taken to obtain the calibrated mechanism model, which includes: The prior distributions of the heat transfer coefficient component and the efficiency component in the physical parameter vector are fitted using sample data from equipment manufacturers. The prior distribution of the heat transfer coefficient component is a truncated normal distribution, and the prior distribution of the efficiency component is a Beta distribution. Take minute-level sampling data within the sliding window, and construct the likelihood function based on the deviations of the supply and return water temperatures, equipment power, and chilled water flow rates output by the mechanism model relative to the measured values ​​in the minute-level sampling data; The posterior distribution is obtained by using Markov chain Monte Carlo sampling, based on the prior distribution and the likelihood function. The mean of the posterior distribution is taken as the value of the parameter to be calibrated and input into the mechanism model to obtain the calibrated mechanism model, and the covariance of the posterior distribution is retained for uncertainty propagation analysis. The update of the posterior distribution is repeated by sliding in new sampled data at the right end and out old sampled data at the left end of the sliding window.

3. The method according to claim 2, characterized in that, The method employing Markov chain Monte Carlo sampling, and obtaining the posterior distribution based on the prior distribution and the likelihood function, includes: The physical parameter vector and the hidden initial values ​​of the mechanism model are augmented into a joint sampling object; Based on the physical grouping of heat transfer parameters, mechanical efficiency parameters, and hydraulic parameters, block-based Metropolis-Hastings updates are performed on the joint sampling objects; The empirical covariance of the accepted samples is used to propose the covariance within the block by scaling the Robbins-Monro attenuation coefficient online, so that the sampling acceptance rate is maintained within the preset acceptance rate range. When the likelihood function can be reorganized into a linear Gaussian state-space form, Kalman filtering recursion is used instead of forward dynamic simulation to calculate the likelihood value. Save the tail sample of the previous round of Markov chain as the starting point for sampling within the sliding window in the next round, and cancel repeated burning.

4. The method according to claim 2, characterized in that, The method further includes: After each round of updating the posterior distribution, a hold-out test is performed on the time segment of the running data that did not participate in the current round of updating. When the posterior prediction coverage of the leaveout test deviates from the preset coverage range, the system reverts to the previous valid posterior distribution and re-executes the update of the posterior distribution after extending the length of the sliding window. When the residual of the test is continuously horizontally displaced, the update of the posterior distribution is paused and a device mutation maintenance alarm is output.

5. The method according to claim 1, characterized in that, The step of inputting observables of unmodeled factors into a neural network error compensator and outputting residual correction amounts includes: The cumulative trend of the temperature difference between the inlet and outlet of the heat exchanger, the drift of the superheat of the compressor suction gas, the difference between the outdoor dry-bulb temperature and the dew point temperature, and the coordinates of the current operating point are used as observable measurements of the unmodeled factors and input into the neural network error compensator. The residual correction amount for the cooling capacity, power consumption and chilled water outlet temperature of the calibrated mechanism model is output by the neural network error compensator. The residual sequence formed by the difference between the measured value and the output of the calibrated mechanism model is used as the training target to perform supervised training on the neural network error compensator, and online incremental learning is performed as the running data accumulates.

6. The method according to claim 5, characterized in that, The residual correction amount output by the neural network error compensator for the cooling capacity, power consumption, and chilled water outlet temperature of the calibrated mechanism model includes: The neural network error compensator outputs the correction rate for the thermal energy state and the cold energy state, and converts the correction rate into the residual correction amount by heat capacity. A latent variable layer is set between the encoder and the output layer of the neural network error compensator. The dimension of the latent variable layer is smaller than the dimension of the encoder output, so that the residual correction amount is constrained within a preset set of disturbance modes. The weight matrix of the output layer is maintained by Cayley transformation and synthesized into a semi-negative definite quadratic form to ensure that the power term corresponding to the residual correction satisfies the energy dissipation inequality. An orthogonal projection penalty term for the posterior distribution mean response and a latent variable dimension sparse regularization term are added to the loss function of the neural network error compensator. The residual correction amount is sequentially subjected to amplitude limiting and component filtering in the dissipation direction and then superimposed onto the output of the calibrated mechanism model. When the sample coverage at the operating point is lower than the preset coverage threshold, the residual correction amount is attenuated according to the proportion of the sample coverage.

7. The method according to claim 1, characterized in that, The establishment of a differential algebraic equation mechanism model for the centralized cooling system equipment, and the setting of the physical parameter vector of the mechanism model as the parameters to be calibrated, includes: The centralized cooling system equipment includes chiller units, chilled water pumps, cooling water pumps, cold storage devices, and heat pump units; The mechanism model of the chiller unit includes the compressor energy efficiency equation, the evaporator and condenser efficiency-heat transfer unit number method heat transfer equation, and the refrigerant mass flow balance equation. The mechanism models of the chilled water pump and the cooling water pump include the head-flow characteristic curve equation, the shaft power equation, and the similarity law conversion relationship as a function of rotational speed. The mechanism model of the cold storage device includes the lumped parameter energy balance differential equation of stratified water temperature and the cold charging and releasing characteristic equation. The mechanism model of the heat pump unit includes a vapor compression cycle equation and is coupled with the heating and cooling switching logic of the four-way reversing valve. For each type of equipment in the centralized cooling system, alternative mechanism structures with different complexities are constructed, and the alternative mechanism structure with the largest marginal likelihood is used as the structure of the mechanism model of the corresponding equipment.

8. The method according to claim 1, characterized in that, The process involves establishing a three-level model library (second-level, minute-level, and hour-level) based on the calibrated mechanism model, and disseminating updated parameter values ​​from any level model to the other two levels models via a unified parameter database, including: The simulation step size of the second-level model is 0.1 seconds to 1 second, and the dynamic characteristics of the complete differential-algebraic equations of the calibrated mechanism model are preserved. The simulation step size of the minute-level model is 1 to 5 minutes. Based on the second-level model, rigid fast variables are eliminated, while pipeline transmission and distribution delay and unit regulation inertia are retained. The simulation step size of the hourly model is 1 hour. Steady-state simplification is performed based on the minute-level model, while retaining the capacity constraints of the cold storage device and the unit's energy efficiency characteristic curves. The rated cooling capacity and efficiency parameters of the same device are stored as unique values ​​in the unified parameter database and are shared by the second-level model, the minute-level model and the hour-level model.

9. The method according to claim 8, characterized in that, The method further includes: The hourly model is used as the first fidelity model, the minute-level model as the second fidelity model, and the second-level model as the third fidelity model. The response outputs of the first fidelity model, the second fidelity model, and the third fidelity model are collected along the operating condition grid composed of chilled water inlet temperature, load rate, and outdoor dry-bulb temperature. Using the response output of the first fidelity model as the prior mean, train a Gaussian process to differentiate between the second fidelity model and the first fidelity model; Using the response output of the second fidelity model after scaling as the prior mean, a Gaussian process of difference between the third fidelity model and the second fidelity model is trained, and a noise term is registered for the observation variance. Write the hyperparameters and training timestamps of the differential Gaussian process into the unified parameter database; Based on the measured values, check whether the deviation between the response output of the first fidelity model after correction by the differential Gaussian process and the response output of the third fidelity model exceeds the preset quantile envelope. If it exceeds the preset quantile envelope, perform incremental retraining on the differential Gaussian process and output a cross-scale mismatch flag.

10. A dynamic modeling device for cooling equipment driven by hybrid mechanism data, characterized in that, include: The mechanism modeling module is used to establish the differential algebraic equation mechanism model of the centralized cooling system equipment, and to set the physical parameter vector of the mechanism model as the parameters to be calibrated. The Bayesian calibration module is used to construct a likelihood function based on the sliding window running data and the residuals output by the mechanism model, update the posterior distribution of the physical parameter vector, take the posterior mean parameter, and obtain the calibrated mechanism model. The residual compensation module is used to input the observables of unmodeled factors into the neural network error compensator and output the residual correction amount. The dual-path overlay module is used to run the update of the posterior distribution and the output of the residual correction in parallel, and to overlay the output of the calibrated mechanism model with the residual correction to obtain the dynamic prediction result of the equipment. The multi-scale management module is used to establish a three-level model library of seconds, minutes and hours based on the calibrated mechanism model, and to publish the updated parameter values ​​of any level model to the other two levels models through a unified parameter database.

Images