A public heating, ventilation and air conditioning device cluster collaborative optimization control system based on multi-source load prediction

Abstract

The application relates to the general control or adjustment technical field and discloses a public heating and ventilation device cluster cooperative optimization control system based on multi-source load prediction, which comprises a power distribution network side multi-dimensional state sensing module, a random power load trend prediction module, a variable stiffness power boundary locking module and a cooperative power distribution module. Through real-time collection of voltage transient, active power and external power demand disturbance data, an electric energy demand evolution curve is generated, a covariance matrix of prediction confidence is calculated, a climbing rate constraint boundary of power throughput change rate is dynamically locked by using the covariance matrix, and a quadratic programming problem of total power loss minimization is solved under the boundary limitation. The application inversely constructs a physical level power fluctuation defense envelope by using prediction uncertainty, effectively suppresses invalid electric stress impact, and realizes accurate phase synchronization and physical protection of state evolution of a controlled object.

Description

Technical Field

[0001] This invention relates to a collaborative optimization control system for a cluster of public HVAC equipment based on multi-source load prediction, belonging to the general field of control or regulation technology. Background Technology

[0002] Currently, in the power distribution networks of large public buildings or industrial parks, HVAC equipment clusters typically constitute a large proportion of the power load. In order to optimize the efficiency of power utilization and respond to the peak shaving and valley filling needs of the power grid, the existing mainstream technical solutions generally adopt a distributed control architecture based on variable frequency speed regulation technology. Such systems typically collect multi-source feedforward data such as meteorological parameters and passenger flow density to calculate the future load demand curve, and send frequency or power adjustment commands to the chiller compressors, high-power circulating pumps and terminal fans in the cluster accordingly. This control mode based on feedforward prediction matches the time delay characteristics of the controlled object through pre-adjustment, thereby improving the overall control accuracy and energy efficiency ratio of the system under steady-state conditions.

[0003] Existing technologies suffer from physical bottlenecks in underlying hardware-driven responses, and software algorithms lack logical coherence in handling uncertainties from multi-source data. Chinese invention patent CN101021914A discloses a method and system for predicting HVAC loads, which employs a dual architecture of a mechanistic model and a compensator model. It uses historical data to identify deviations and correct simulated loads in order to approximate the actual demand. However, this method still relies on deterministic control thinking, logically assuming that the predicted value after compensation is a reliable true value, and ignoring the random noise and time-varying confidence level of the input data itself. Due to the lack of a mechanism for quantitatively evaluating the reliability of predicted signals, the control system cannot distinguish between load and random noise when facing non-stationary weather disturbances. It mechanically adjusts the output based on the prediction results, and the rigid control strategy is prone to causing the actuator to overreact to high-frequency noise, resulting in energy consumption oscillations and mechanical wear. However, in actual engineering applications, when the above control logic is applied to complex working conditions with high dynamic characteristics, the existing hard-connected control architecture exposes its fundamental defects at the power electronic drive level. Since multi-source predicted data naturally has random fluctuation characteristics and time-varying uncertainty, and traditional control systems lack a dynamic matching mechanism between data confidence and circuit physical response capability when converting these predicted data into control commands for the actuator, the system tends to treat all predicted inputs as deterministic true values, causing the controller to frequently drive high-power frequency converters to perform violent acceleration and deceleration actions in an attempt to track high-frequency noise signals that have no actual physical meaning.

[0004] Therefore, the technical problem to be solved by this invention is how to construct a cluster collaborative control mechanism that can sense the confidence level of prediction data in real time and establish a dynamic mapping relationship between the physical constraints of the underlying drive circuit and the optimization of the upper logic, so as to ensure accurate response to real load while suppressing invalid electrical stress impact. Summary of the Invention

[0005] To address the problems mentioned in the background art, the technical solution of the present invention is as follows: A collaborative optimization control system for public HVAC equipment clusters based on multi-source load prediction, the system comprising:

[0006] The distribution network side multi-dimensional state perception module is used to collect voltage transient data of the distribution network bus, real-time active power data of the public HVAC equipment cluster as a controlled flexible electrical load, and external power demand disturbance data in real time.

[0007] The random power load trend prediction module is used to generate the future time domain power demand evolution curve based on real-time active power data and external power demand disturbance data, and simultaneously calculate the covariance matrix that characterizes the prediction confidence of the power demand evolution curve at each moment; thereby providing the controller with statistical characteristic parameters that characterize the input signal quality, so as to achieve decoupling between control parameters and signal confidence.

[0008] The variable stiffness power boundary locking module stores the nonlinear inverse mapping relationship between the covariance matrix and the system's power response rate. Based on the norm of the covariance matrix, it dynamically calculates the ramp rate constraint boundary that limits the power throughput change rate of the public HVAC equipment cluster. When the uncertainty of the prediction indicated by the covariance matrix increases, it forcibly shrinks the ramp rate constraint boundary to construct a physical defense envelope to suppress power surges on the grid side. It achieves real-time adaptive decoupling between the control law gain and the reliability of the input data. The variable stiffness power boundary locking module constitutes a general signal stress protection logic.

[0009] The collaborative power allocation module is used to solve a quadratic programming problem with the goal of minimizing the total active power loss of the public HVAC equipment cluster under the hard constraints of the ramp rate constraint boundary, and to generate and issue frequency adjustment commands for each variable frequency drive execution unit in the public HVAC equipment cluster.

[0010] Preferably, when constructing a physical defense envelope to suppress power surges on the grid side, the variable stiffness power boundary locking module executes hard locking logic that includes the following inequality constraints: Where u(k) is the frequency adjustment command value issued to the variable frequency drive execution unit in the current control cycle, and u(k-1) is the frequency adjustment command value in the previous control cycle. The power change rate limit threshold is obtained by real-time mapping based on the norm of the covariance matrix Σ, and Δt is the time step of the control cycle.

[0011] Preferably, the system further includes a time-domain phase orthogonal correction module, used to calculate the first derivative characteristics of real-time active power data and the slope characteristics of the power demand evolution curve; the time-domain phase orthogonal correction module is used to trigger a time-domain translation operator to dynamically align the control window of the cooperative power allocation module in response to the phase difference between the first derivative characteristics and the slope characteristics exceeding a preset synchronization tolerance, and to establish a phase-locked loop between the control command output and the state evolution of the controlled object through time-domain feature compensation, so as to ensure that the effective time of the frequency adjustment command and the starting point of the actual power load change are phase synchronized on the time axis.

[0012] Preferably, the collaborative power distribution module includes a low-level register overwrite unit, which is used to directly address and access the hardware control register of the variable frequency drive execution unit through the fieldbus protocol; the low-level register overwrite unit is used to directly write the ramp rate constraint boundary generated by the variable stiffness power boundary locking module into the acceleration and deceleration time parameter register of the variable frequency drive execution unit, so as to establish a physical-level power change rate limiting mechanism at the driver hardware level that is independent of the upper-level communication state.

[0013] Preferably, the variable stiffness power boundary locking module includes a distribution network voltage stability enhancement unit for monitoring voltage transient data of the distribution network bus. When the voltage drop in the voltage transient data exceeds a preset voltage safety threshold, the distribution network voltage stability enhancement unit triggers an emergency response mode aimed at supporting the voltage of the distribution network bus and forcibly locks the ramp rate constraint boundary to the creep rate required to maintain the minimum steady-state operation of the system until the voltage transient data returns to the voltage safety threshold range and remains within a preset time window.

[0014] Preferably, the stochastic power load trend prediction module includes a multi-source heterogeneous data cleaning unit, which performs spatiotemporal alignment processing on external power demand disturbance data; the external power demand disturbance data includes meteorological parameter data characterizing environmental heat load and building passenger flow density data characterizing basic power load; the multi-source heterogeneous data cleaning unit is used to remove outlier noise from the external power demand disturbance data and map the processed data into a normalized power load influence factor, which serves as the input basis for generating the power demand evolution curve.

[0015] Preferably, the variable frequency drive execution unit includes a chiller frequency converter as a large inertial base load power unit and a circulating pump frequency converter as a high frequency regulating load power unit; the cooperative power distribution module is used to adjust the operating frequency of the large inertial base load power unit in order to respond to low-frequency large amplitude power load changes, and adjust the operating frequency of the high frequency regulating load power unit to compensate for high-frequency small amplitude power load fluctuations, based on the power demand evolution curve and under the premise of meeting the ramp rate constraint boundary.

[0016] Preferably, the variable stiffness power boundary locking module includes a stability gain scheduling unit, which is used to automatically switch the control strategy of the system when the prediction confidence is lower than a preset confidence threshold; the stability gain scheduling unit is used to switch from the prediction-based feedforward control mode to the conservative feedback control mode based on real-time deviation, and reduce the proportional gain coefficient of the frequency regulation command to prevent power oscillation on the distribution network side caused by prediction deviation.

[0017] Preferably, the system also includes an energy efficiency optimization feedback module, which is used to collect actual power loss data of public HVAC equipment clusters in real time; the energy efficiency optimization feedback module is used to calculate the residual between the actual power loss data and the theoretical energy consumption model, and use the residual to correct the weight of the objective function of the quadratic programming problem in the collaborative power allocation module online.

[0018] Preferably, the collaborative power distribution module includes a dead-zone avoidance unit for identifying the inefficient operating frequency band of the variable frequency drive execution unit; the dead-zone avoidance unit is used to introduce frequency skipping constraints when generating frequency adjustment commands to ensure that the issued frequency adjustment commands avoid inefficient operating frequency bands, so that the public HVAC equipment cluster always operates in a frequency range with high energy efficiency ratio.

[0019] Compared with the prior art, the beneficial effects of the present invention are:

[0020] 1. In the collaborative optimization control of public HVAC equipment clusters, an inverse physical mapping mechanism is established between prediction confidence and the limit of the rate of change of actuator action to achieve adaptive stress protection for power drive circuits. This provides a general control logic method that can effectively handle non-stationary random disturbances. This method is applicable to various complex controlled systems with random input noise and large inertial physical constraints. Existing technologies, when facing high-frequency random disturbances from multi-source prediction data, often directly transmit fluctuation signals to the actuator, causing the inverter or servo mechanism to frequently execute ineffective acceleration and deceleration actions, thereby triggering current surges in the drive circuit and alternating stress fatigue of mechanical components. This invention introduces the covariance matrix as a quantitative indicator of uncertainty. When the confidence of the prediction data decreases, the maximum allowable ramp rate of the actuator is dynamically reduced in the underlying constraint logic of the controller. This mechanism transforms high-frequency logic instruction noise into low-frequency smooth creep actions at the physical execution level, cutting off the transmission path of data noise to physical electrical stress. While ensuring the steady-state response of the system, it reduces the heat loss and mechanical wear rate of power devices and extends the mean time between failures (MTBF) of heterogeneous load clusters.

[0021] 2. By using a closed-loop fingerprint recognition mechanism based on drive power characteristics, the invalid power injection and logic mismatch caused by the physical dead zone of the actuator are eliminated. Addressing the common mechanical lag or dead zone problem in HVAC cluster equipment after long-term operation, conventional control often continuously increases command gain due to a lack of physical feedback, resulting in electrical energy being consumed for motor coil heating rather than generating actual work. This invention utilizes a collaborative allocation unit to monitor the transient power waveform of the drive loop in real time. By extracting energy fingerprint features associated with mechanical actions, specific equipment in the dead zone state is identified. Once a mismatch between the energy characteristics of the command and response is detected, the system stops injecting invalid power into that equipment and re-encodes the corresponding adjustment requirements before allocating them to other equipment in the cluster that are in the active range. This collaborative compensation logic based on real energy flow feedback ensures that every unit of energy injected into the grid is converted into an effective physical adjustment quantity, avoiding low-frequency energy consumption oscillations caused by the cumulative effect of the dead zone.

[0022] 3. By utilizing an environmental response inversion mechanism under active perturbation excitation, this invention addresses the control model distortion caused by physical environment boundary drift. The spatial physical characteristics of public buildings often change dynamically with factors such as pedestrian flow and door opening and closing, leading to a disconnect between the preset thermodynamic model parameters and actual operating conditions. This results in calculation deviations in control stiffness. During the quasi-steady-state operation of the system, this invention injects energy perturbation signals with amplitudes below the sensing threshold into the cluster and uses sensors to collect the environmental transient response envelope generated by this signal. The logic reconstruction unit compares the measured envelope with the theoretical model, calculates the physical environment alignment coefficient, and directly corrects the gain matrix of the control law. This enables the control system to perceive and self-calibrate changes in the physical environment's heat capacity, ensuring that the controller's output strength always maintains impedance matching with the current actual physical load characteristics under varying operating conditions, avoiding overshoot or response hysteresis caused by model mismatch. Attached Figure Description

[0023] Figure 1 This is a flowchart of the collaborative optimization control logic for public HVAC equipment clusters in this invention.

[0024] Figure 2 This is a schematic diagram of the data interaction architecture and functional closed-loop principle of the system of the present invention.

[0025] The objectives, features, and advantages of this invention will be further explained in conjunction with the embodiments and with reference to the accompanying drawings. Detailed Implementation

[0026] The technical solutions of the embodiments of this application will be clearly described below with reference to the accompanying drawings. Obviously, the described embodiments are only some embodiments of this application, not all embodiments. All other embodiments obtained by those skilled in the art based on the embodiments of this application are within the scope of protection of this application.

[0027] A collaborative optimization control system for public HVAC equipment clusters based on multi-source load forecasting, characterized in that the system includes:

[0028] The distribution network side multi-dimensional state perception module is used to collect voltage transient data of the distribution network bus, real-time active power data of the public HVAC equipment cluster as a controlled flexible electrical load, and external power demand disturbance data in real time.

[0029] The random power load trend prediction module is used to generate the future time domain power demand evolution curve based on real-time active power data and external power demand disturbance data, and simultaneously calculate the covariance matrix that characterizes the prediction confidence of the power demand evolution curve at each moment; thereby providing the controller with statistical characteristic parameters that characterize the input signal quality, so as to achieve decoupling between control parameters and signal confidence.

[0030] The variable stiffness power boundary locking module stores the nonlinear inverse mapping relationship between the covariance matrix and the system's power response rate. Based on the norm of the covariance matrix, it dynamically calculates the ramp rate constraint boundary that limits the power throughput change rate of the public HVAC equipment cluster. When the uncertainty of the prediction indicated by the covariance matrix increases, it forcibly shrinks the ramp rate constraint boundary to construct a physical defense envelope to suppress power surges on the grid side. It achieves real-time adaptive decoupling between the control law gain and the reliability of the input data. The variable stiffness power boundary locking module constitutes a general signal stress protection logic.

[0031] The collaborative power allocation module is used to solve a quadratic programming problem with the goal of minimizing the total active power loss of the public HVAC equipment cluster under the hard constraints of the ramp rate constraint boundary, and to generate and issue frequency adjustment commands for each variable frequency drive execution unit in the public HVAC equipment cluster.

[0032] Preferably, when constructing a physical defense envelope to suppress power surges on the grid side, the variable stiffness power boundary locking module executes hard locking logic that includes the following inequality constraints: Where u(k) is the frequency adjustment command value issued to the variable frequency drive execution unit in the current control cycle, and u(k-1) is the frequency adjustment command value in the previous control cycle. The power change rate limit threshold is obtained by real-time mapping based on the norm of the covariance matrix Σ, and Δt is the time step of the control cycle.

[0033] Preferably, the system further includes a time-domain phase orthogonal correction module, used to calculate the first derivative characteristics of real-time active power data and the slope characteristics of the power demand evolution curve; the time-domain phase orthogonal correction module is used to trigger a time-domain translation operator to dynamically align the control window of the cooperative power allocation module in response to the phase difference between the first derivative characteristics and the slope characteristics exceeding a preset synchronization tolerance, and to establish a phase-locked loop between the control command output and the state evolution of the controlled object through time-domain feature compensation, so as to ensure that the effective time of the frequency adjustment command and the starting point of the actual power load change are phase synchronized on the time axis.

[0034] Preferably, the collaborative power distribution module includes a low-level register overwrite unit, which is used to directly address and access the hardware control register of the variable frequency drive execution unit through the fieldbus protocol; the low-level register overwrite unit is used to directly write the ramp rate constraint boundary generated by the variable stiffness power boundary locking module into the acceleration and deceleration time parameter register of the variable frequency drive execution unit, so as to establish a physical-level power change rate limiting mechanism at the driver hardware level that is independent of the upper-level communication state.

[0035] Preferably, the variable stiffness power boundary locking module includes a distribution network voltage stability enhancement unit for monitoring voltage transient data of the distribution network bus. When the voltage drop in the voltage transient data exceeds a preset voltage safety threshold, the distribution network voltage stability enhancement unit triggers an emergency response mode aimed at supporting the voltage of the distribution network bus and forcibly locks the ramp rate constraint boundary to the creep rate required to maintain the minimum steady-state operation of the system until the voltage transient data returns to the voltage safety threshold range and remains within a preset time window.

[0036] Preferably, the stochastic power load trend prediction module includes a multi-source heterogeneous data cleaning unit, which performs spatiotemporal alignment processing on external power demand disturbance data; the external power demand disturbance data includes meteorological parameter data characterizing environmental heat load and building passenger flow density data characterizing basic power load; the multi-source heterogeneous data cleaning unit is used to remove outlier noise from the external power demand disturbance data and map the processed data into a normalized power load influence factor, which serves as the input basis for generating the power demand evolution curve.

[0037] Preferably, the variable frequency drive execution unit includes a chiller frequency converter as a large inertial base load power unit and a circulating pump frequency converter as a high frequency regulating load power unit; the cooperative power distribution module is used to adjust the operating frequency of the large inertial base load power unit in order to respond to low-frequency large amplitude power load changes, and adjust the operating frequency of the high frequency regulating load power unit to compensate for high-frequency small amplitude power load fluctuations, based on the power demand evolution curve and under the premise of meeting the ramp rate constraint boundary.

[0038] Preferably, the variable stiffness power boundary locking module includes a stability gain scheduling unit, which is used to automatically switch the control strategy of the system when the prediction confidence is lower than a preset confidence threshold; the stability gain scheduling unit is used to switch from the prediction-based feedforward control mode to the conservative feedback control mode based on real-time deviation, and reduce the proportional gain coefficient of the frequency regulation command to prevent power oscillation on the distribution network side caused by prediction deviation.

[0039] Preferably, the system also includes an energy efficiency optimization feedback module, which is used to collect actual power loss data of public HVAC equipment clusters in real time; the energy efficiency optimization feedback module is used to calculate the residual between the actual power loss data and the theoretical energy consumption model, and use the residual to correct the weight of the objective function of the quadratic programming problem in the collaborative power allocation module online.

[0040] Preferably, the collaborative power distribution module includes a dead-zone avoidance unit for identifying the inefficient operating frequency band of the variable frequency drive execution unit; the dead-zone avoidance unit is used to introduce frequency skipping constraints when generating frequency adjustment commands to ensure that the issued frequency adjustment commands avoid inefficient operating frequency bands, so that the public HVAC equipment cluster always operates in a frequency range with high energy efficiency ratio.

[0041] Example 1: A collaborative optimization control system for a public HVAC equipment cluster based on multi-source load forecasting is deployed in the energy center of a large transportation hub with high-density passenger flow. The controlled object in this scenario is a public HVAC equipment cluster containing multiple centrifugal chillers and variable frequency circulating pumps. During the system's continuous steady-state operation, the multi-dimensional state perception module on the distribution network side collects real-time voltage transient data of the distribution network bus, real-time active power data of the public HVAC equipment cluster, and external power demand disturbance data. The external power demand disturbance data includes passenger flow density change rate and environmental heat load parameters. When the system faces sudden and drastic fluctuations in meteorological data or discontinuous passenger flow guidance signals, the stochastic power load trend prediction module... Based on the above input data, the module generates the future time-domain power demand evolution curve and simultaneously calculates the covariance matrix representing the prediction confidence of the power demand evolution curve at each moment. At this time, the system detects that the norm of the covariance matrix increases, indicating that although the current load forecast data shows a leap trend in numerical value, its statistical certainty has been greatly reduced, and the input signal contains high-frequency random noise. Under this high uncertainty input condition, the variable stiffness power boundary locking module calculates the power throughput change rate limit threshold of the current control cycle in real time based on the preset nonlinear inverse mapping relationship, that is, the negative correlation function between the covariance matrix norm and the allowable ramp rate. This module constructs a physical defense envelope and shrinks the ramp rate constraint boundary of the public HVAC equipment cluster.

[0042] In practical implementation, the system does not directly run the abstract function image, but instead executes microcontroller instructions based on table lookup and reciprocal operations: The system reads the baseline maximum ramp rate preset in non-volatile memory, which is calibrated to 2.50 Hz / s; it reads the norm value of the covariance matrix and multiplies it by the preset sensitivity coefficient of 0.80 to obtain the safety reduction factor; the system divides the baseline maximum ramp rate by 1.0 and adds the safety reduction factor to calculate the current dynamic allowable rate of change. For example, when the covariance norm is 0.50, the dynamic allowable rate of change is... The rate of change is limited to 1.78 Hz / s. Next, the underlying drive logic performs a crucial physical dimension transformation: it reads the inverter's maximum output frequency parameter (usually set to 50.0 Hz), divides it by the dynamic allowable rate of change, and obtains the corresponding acceleration / deceleration time value. If the dynamic allowable rate of change is 1.78 Hz / s, the calculated acceleration / deceleration time is 28.1 s. This value is formatted as a 16-bit unsigned integer. Specifically, this module dynamically embeds inequality constraints with the following hard-locking logic into the quadratic programming solver of the model predictive control: Where u(k) is the frequency adjustment command value issued to the variable frequency drive execution unit in the current control cycle, and u(k-1) is the frequency adjustment command value in the previous control cycle. The power change rate limit threshold is obtained by real-time mapping based on the norm of the covariance matrix Σ, where Δt is the time step of the control cycle. Under this constraint, although the predicted power demand evolution curve indicates a significant power ramp-up demand, due to... When calculated to an extremely small value, such as 0.5 Hz / s, the frequency adjustment command output by the cooperative power allocation module is forcibly clamped within the physical defense envelope when solving a quadratic programming problem with the goal of minimizing total active power loss.

[0043] The variable stiffness power boundary locking module performs inverse damping mapping according to the formula. The power change rate limit threshold for the current control cycle is calculated in real time. This characterizes the slope of the maximum allowable change in the frequency adjustment command at the current moment. This is the baseline value for the maximum response rate under rated physical conditions for a public HVAC equipment cluster. The Frobenius norm of the predicted covariance matrix is ​​calculated in real time. α is the sensitivity attenuation coefficient determined by the variance of the historical load prediction error of the offline identification system. The nonlinear inverse proportional function relationship establishes the system's dynamic constraint boundary, which is negatively correlated with the confidence level of the input signal. When the covariance norm increases, indicating that the predicted data contains high-frequency random noise, the system is forced to shrink according to a hyperbolic law. Numerical construction of the high-damping physical envelope; the time-domain phase orthogonal correction module employs a discrete-time sliding window cross-correlation algorithm, and buffers in parallel the first-order difference sequence of real-time active power and the slope sequence of the power demand evolution curve for N control cycles, and calculates... The sliding window length N is set to a range of 50 to 150 control cycles, preferably 100. This length N must meet the requirement of covering the minimum thermal inertia time constant of the controlled HVAC equipment cluster by 3 to 5 times to ensure that complete dynamic features can be extracted and the global maximum value of the cross-correlation function corresponds to the number of lag steps. As input parameters to the time-domain translation operator, the control command sequence generated by the cooperative power allocation module is reversed and shifted in the reverse time sequence to eliminate the phase deviation between the control command and the actual load response caused by the transmission hysteresis of the thermodynamic system.

[0044] The collaborative power distribution module sends the constrained frequency adjustment command to each variable frequency drive execution unit. Simultaneously, the underlying register overwrite unit directly addresses and accesses the hardware control registers of each variable frequency drive via the fieldbus protocol. It directly writes the contracted ramp rate constraint boundary generated by the variable stiffness power boundary locking module into the acceleration / deceleration time parameter register of the variable frequency drive execution unit. This action establishes a physical-level power change rate limitation defense line at the hardware level, independent of the upper-level communication state. During the system's operating cycle, even if the predicted data continues to fluctuate significantly, the actual power output of the public HVAC equipment cluster only exhibits a smooth creep response. No voltage surges or reactive power oscillations induced by high-frequency commands are detected on the distribution network bus. When the covariance matrix norm falls back to the preset confidence interval in subsequent periods, the system automatically releases the ramp rate clamp, restoring its ability to quickly track load changes. This process demonstrates… The system achieves physical layer isolation of signal noise within a single control architecture by transforming the uncertainty of prediction into physical constraints of the actuator in real time. This solves the impedance mismatch problem between highly dynamic prediction data and a large inertial physical system, ensuring the mechanical safety and grid stability of the equipment cluster under non-stationary disturbances. The underlying register overwrite unit of the collaborative power distribution module utilizes the periodic process data object (PDO) communication channel of the industrial fieldbus protocol to directly address the volatile control word address of the random access memory (RAM) of the microcontroller of the frequency converter drive actuator, without accessing the parameter area of ​​the electrically erasable programmable read-only memory (EEPROM). Specific addressing writes avoid the physical limit of the number of erase / write cycles of the non-volatile storage medium, supporting continuous high-frequency dynamic refresh of the ramp rate constraint boundary within the millisecond-level control cycle of the system, preventing damage to the storage chip. When the distribution network voltage stabilization enhancement unit monitors the instantaneous voltage drop of the bus exceeding the calibrated threshold... When the system triggers the highest priority hardware interrupt service routine, it directly bypasses the flexible adjustment logic and forcibly locks the value written to the inverter torque limit register to maintain the minimum circulating head required for the fluid pipeline network. It then uses the thermal inertia characteristics of the hydraulic system to temporarily disconnect the grid-side power load during the voltage sag until the voltage monitoring data returns to the steady-state range of ±5% of the rated range.

[0045] Example 2: To verify the actual control performance and engineering feasibility of the collaborative optimization control system for public HVAC equipment clusters based on multi-source load prediction in a non-ideal industrial environment, a hardware-in-the-loop (HIL) simulation verification platform was built. The RT-BOX real-time simulator was selected as the controlled object model carrier. Internally, a nonlinear thermodynamic model of a centrifugal chiller unit conforming to the ASHRAE standard and a large-inertia hydraulic network model were run. The power distribution network data used dynamic power flow data from the IEEE 33-node standard test system. External disturbance sources were constructed based on 720 hours of continuous historical measured load data from a large-scale transportation hub. To reproduce the signal pollution characteristics in real engineering scenarios, Gaussian white noise with a signal-to-noise ratio of 20 dB was actively injected into the original active power acquisition signal during the experiment, and a frequency of 0.1 was superimposed. Random narrowband interference from Hz to 1.0Hz was used to simulate the effects of sensor aging and drift and the degradation of measurement accuracy by the electromagnetic environment. The experimental design included three parallel control groups to verify the gradient: the experimental group of this invention adopted the control strategy of dynamically adjusting the ramp rate constraint based on the predictive covariance matrix; the control group A adopted the traditional fixed constraint model predictive control, with its ramp rate limit fixed at 5% of the rated power of the equipment per second; the control group B was a partially missing control group. Although a load prediction module was introduced, the variable stiffness power boundary locking module was removed, and the predicted mean was directly used as the control target. The experiment focused on typical sudden disturbance conditions: in the interval from 1200 to 1500 seconds of the simulation time axis, a drastic fluctuation in environmental heat load caused by simulated severe convective weather was introduced, while intermittent packet loss failure of the passenger flow density sensor was superimposed.

[0046] In the initial steady-state phase after the experiment started, the outputs of each system group were able to track the baseline load. When entering the disturbance interval at 1200 seconds, the raw input data showed that the external power demand disturbance data exhibited a significant upward trend with fluctuations. Furthermore, due to sensor packet loss, some data became outliers. At this point, the covariance matrix norm ||Σ|| calculated by the random power load trend prediction module of this invention jumped rapidly from 0.15 in the steady state to over 0.85, indicating the high uncertainty of the current prediction data. Subsequently, the key intermediate variable, namely the power change rate limit threshold, became increasingly important. Under the inverse mapping effect of the variable stiffness power boundary locking module, the steady-state power output rapidly shrinks from 2.5 kW / s to 0.4 kW / s. This process forces the divergent information entropy at the logic layer to be transformed into a hard constraint boundary at the physical layer. Data records show differences in control response. Comparison group A, due to the use of fixed constraints, had its controller attempting to track the noisy predicted command at full speed, causing the output frequency of the variable frequency drive execution unit to fluctuate wildly between 35 Hz and 48 Hz. The instantaneous value of the total harmonic distortion of the distribution network bus voltage once exceeded the standard limit of 5%, and the mechanical action frequency of the chiller unit's guide vane opening reached 12 times per minute, posing a risk of mechanical wear. Although comparison group B utilized the predicted mean, it lacked... Despite boundary locking for uncertainty, the output power still experienced three invalid overshoots exceeding 15% following the spurious peak of the predicted curve. In contrast, the actual power output curve of the present invention's sample group presented a smooth, progressively rising curve, filtering out high-frequency jitter at the input end. The inverter's operating frequency remained within twice per minute. Although at the 1250-second mark, the instantaneous cooling capacity of the present invention's sample group lagged behind the predicted average demand by approximately 8%, the temperature fluctuation of the chilled water tank was only 0.5℃, not exceeding the comfort threshold, proving that the system utilized thermal inertia to mitigate the power surge. Further boundary exploration experiments examined the parameter sensitivity of the mapping function between the covariance matrix norm and the ramp rate limit threshold, setting the mapping function as follows: Where K is the adjustment coefficient. As the preset minimum power response deviation constant, experimental data shows that when the K value is set below 0.2, the system's power response is too sluggish under low confidence, causing the return water temperature to deviate from the set value by more than 2 degrees Celsius in the early stage of disturbance, which fails to meet the basic temperature control requirements. When the K value is set above 1.5, the ability to suppress covariance fluctuations is weakened, and the smoothness index of the output power waveform, i.e., the second-order difference variance, deteriorates by 60%. When the K value is in the preferred range of 0.5 to 1.0, the system can achieve the best balance between maintaining the stability of the bus voltage, i.e., the fluctuation rate is less than 2%, and ensuring the comfort of the terminal temperature, which confirms the engineering rationality and necessity of the parameter selection in this invention.

[0047] Example 3: This example combines Figures 1 to 2 This document describes a collaborative optimization control system for a cluster of public HVAC equipment based on multi-source load forecasting. Figure 1As shown, this module is responsible for collecting voltage transients, real-time active power, and external disturbance data. The above information is aggregated into a real-time data stream and input to the stochastic power load trend prediction module. The stochastic power load trend prediction module generates an energy demand evolution curve and calculates the prediction confidence covariance matrix. Then, the data containing the energy demand curve and the covariance matrix Σ is transmitted to the variable stiffness power boundary locking module. The variable stiffness power boundary locking module dynamically calculates the ramp rate constraint boundary based on the covariance norm to construct a physical defense envelope, and then outputs the ramp rate constraint boundary hard limit. After receiving the above limit, the collaborative power allocation module solves a quadratic programming problem under the hard limit to generate a frequency adjustment command. Finally, the frequency adjustment command u(k) is sent to the controlled objects of the public HVAC equipment cluster containing the variable frequency drive execution unit.

[0048] like Figure 2 As shown, the process branch consists of two steps: injecting a pseudo-random binary sequence excitation signal and locking the physical limit of the maximum frequency change rate. The calibration results are then associated with the public HVAC equipment cluster. On the other hand, based on the input status of the distribution network and external disturbance sources, the system performs parallel operations of collecting voltage transient data, collecting real-time active power data, and generating power demand evolution curves. The data flow passes through the calculation of prediction confidence and variable stiffness power boundary locking stages to construct a physical defense envelope. Finally, the system generates frequency regulation commands by solving a quadratic programming problem and applies them to the public HVAC equipment cluster.

[0049] Example 4: In scenarios involving long-term continuous operation, the technical solution of this invention faces the challenge of gradually deteriorating signal-to-noise ratios due to component aging in field sensors. Under such long-term timescales, the stochastic power load trend prediction module does not maintain static calculation logic but instead executes an online adaptive covariance estimation procedure based on recursive updates. This ensures that the covariance matrix, representing the prediction confidence, can track the statistical characteristic drift of the input signal in real time. When the system is in online operation, the multi-dimensional state perception module on the distribution network side continuously collects real-time active power data at a preset high-frequency sampling rate and compares it with the predicted value from the previous moment, generating a real-time prediction residual sequence. This sequence is then fed into the covariance matrix iteration unit. This unit does not use a fixed statistical window but instead employs a recursive algorithm with a dynamic forgetting factor to update the current covariance matrix estimate. The specific iterative update logic follows the following mathematical relationship with weighted memory characteristics: ,in, Let be the covariance matrix updated at time k. Let be the covariance matrix of the previous time step. The predicted residual vector is calculated at the current moment, and λ is a dynamic forgetting factor with a value between 0 and 1. In standard operation mode, the system adjusts the value of λ in real time according to the stationarity index of the residual sequence. When the variance of the residual sequence is detected to remain stable in the most recent N periods, the system automatically reduces the value of λ and assigns higher weight to historical statistical data to filter out occasional measurement noise. When a trend divergence of the residual variance is detected, the system rapidly increases the value of λ to force the covariance matrix to quickly forget outdated statistical features, thereby keenly capturing newly emerging fluctuation patterns in the current signal.

[0050] Based on the above real-time iteratively updated covariance matrix The variable stiffness power boundary locking module executes an online calibration process for the mapping coefficient K. The system runs a long-term performance monitoring process in the background, continuously calculating the moving average of the control action frequency and the bus voltage deviation rate. When the system detects an unnecessary increase in the control action frequency without an increase in the covariance norm—indicating signs of overly sensitive regulation—the process fine-tunes the mapping coefficient K by a preset step size, tightening the limit on the ramp rate. Conversely, if the bus voltage deviation rate approaches the safety threshold, it fine-tunes the K value in the opposite direction, releasing greater regulatory dynamics. In the accelerated aging test simulating the linear decay of sensor accuracy, as the input... The variance of white noise in the input signal increases linearly from the initial 0.01 to 0.05. The conventional fixed-parameter control system gradually exhibits high-frequency oscillations, resulting in a 300% increase in the number of actuator actions. However, the system using the adaptive procedure in this embodiment, through the adaptive adjustment of the dynamic forgetting factor λ, enables the calculated covariance matrix norm to accurately reflect the increase in noise level, thereby driving the mapping coefficient K to automatically evolve to a new equilibrium point. The final test data shows that although the input signal quality decreases, the smoothness of the power command output by the system only decreases by 5%, and the fluctuation rate of the distribution network bus voltage is always locked within the safe range of 2.5%.

[0051] Example 5: Before the system is formally connected to the power grid to perform closed-loop control tasks, an offline physical boundary calibration process is initiated to construct a static reference manifold that matches the specific site hardware characteristics for the variable stiffness power boundary locking module. This process injects a set of pseudo-random binary sequence excitation signals with gradient amplitudes into a cluster of public HVAC equipment in an unloaded or low-load state. Simultaneously, high-frequency vibration sensors and power quality analyzers are used to record the transient response spectrum of the mechanical shaft system and the voltage distortion rate of the distribution network bus, respectively. Thus, the physical limit of the maximum frequency change rate that each actuator can withstand without triggering mechanical resonance and exceeding the harmonic limit of the power grid is locked through measured data, and it is set as the lower bound asymptote of the mapping function. At the same time, the upper bound asymptote of the mapping function is determined by combining the thermal inertia time constant obtained from the inversion of the thermal parameters of the building envelope. This generates an initial lookup table that can cover the safe operation envelope of the equipment.

[0052] Following the physical boundary calibration, the system executes a baseline initialization procedure for the prediction model based on historical data backtracking. This addresses the covariance matrix divergence problem caused by the lack of prior statistical features in the cold start phase of the stochastic power load trend prediction module. The procedure calls upon historical load data and meteorological records from the building's past complete operating cycles stored in the local database. An open-loop rolling deduction is performed in the shadow prediction model running in the background. Statistical methods are used to calculate the probability density distribution of the prediction residuals under different typical meteorological conditions, and the expected value of the residual covariance under each typical operating condition is extracted as the initial state matrix for the online recursive algorithm. This step ensures that when the system is first put into online operation, its dynamic forgetting factor λ and mapping adjustment coefficient K evolve based on localized statistical laws rather than general default values, thereby maintaining the convergence and stability of the control strategy in the early cold start stage when the algorithm has not accumulated enough real-time data.

[0053] Example 6: In the pre-deployment phase of this system for a specific building scenario, in order to address potential empirical biases during control parameter initialization, a standardized engineering procedure for environmental response feature identification and model parameter calibration is implemented. During non-business hours, a 24-hour active excitation test process is initiated. The system control module applies a set of sweep sine wave speed commands with a frequency range of 0.01Hz to 0.5Hz to the variable frequency water pumps and cooling tower fan groups. The amplitude is limited to ±10% of the rated speed to ensure operational safety. During this period, a high-precision temperature sensor array distributed at the return air vents of the terminal air conditioning units synchronously records the transient response data of the building's internal thermal environment to the dynamic disturbances of the hydraulic distribution system with a sampling period of 1 second.

[0054] In this process, the system explicitly defines a three-dimensional eigenvector structure for calculating the prediction confidence covariance matrix: the first dimension is the real-time active power residual, with a normalized benchmark of 500.0 kW (system rated power); the second dimension is the chilled water return temperature residual, with a normalized benchmark of 2.0℃ (maximum allowable temperature drift); and the third dimension is the passenger flow density change rate residual, with a normalized benchmark of 50 people per minute. The system uses a sliding window with 60 sampling points to calculate the statistical covariance of the above three normalized dimensions over time in real time. This multi-dimensional construction method ensures that the matrix not only reflects power fluctuations but also captures the coupling uncertainty between the thermal system and external disturbances. Using the above measured input-output data pairs, the system constructs a building-specific thermal inertia transfer function model offline based on a subspace identification algorithm, and extracts the key time constants characterizing the system's response lag to control commands. and the steady-state gain coefficient of the system Based on the identification The control period Δt within the variable stiffness power boundary locking module is automatically set to [value]. One-tenth of that, to meet the closed-loop control bandwidth requirement of Shannon's sampling theorem; at the same time, based on The penalty weight matrices Q and R in the collaborative power allocation module are initialized to achieve a normalized value that balances the magnitude of the energy consumption term and the control increment term in the objective function. Through this physical parameter calibration procedure based on measured data, the system achieves precise matching between the control parameters and the dynamic characteristics of the specific physical object, eliminating the risk of control divergence or response hysteresis caused by improper parameter settings, and ensuring the universality and stability of the system under different building thermal environments.

[0055] It will be apparent to those skilled in the art that the present invention is not limited to the details of the exemplary embodiments described above, and that the present invention can be implemented in other specific forms without departing from the spirit or essential characteristics of the present invention.

[0056] Finally, it should be noted that the above embodiments are only used to illustrate the technical solutions of the present invention and are not intended to limit it. Although the present invention has been described in detail with reference to preferred embodiments, those skilled in the art should understand that modifications or equivalent substitutions can be made to the technical solutions of the present invention without departing from the spirit and scope of the technical solutions of the present invention.

Claims

1. A collaborative optimization control system for public HVAC equipment clusters based on multi-source load forecasting, characterized in that the system... include: The distribution network side multi-dimensional state perception module is used to collect voltage transient data of the distribution network bus, real-time active power data of the public HVAC equipment cluster as a controlled flexible electrical load, and external power demand disturbance data in real time. The random power load trend prediction module is used to generate the future time domain power demand evolution curve based on real-time active power data and external power demand disturbance data, and simultaneously calculate the covariance matrix that characterizes the prediction confidence of the power demand evolution curve at each moment; thereby providing the controller with statistical characteristic parameters characterizing the quality of the input signal. The variable stiffness power boundary locking module stores the nonlinear inverse mapping relationship between the covariance matrix and the system's power response rate. Based on the norm of the covariance matrix, it dynamically calculates the ramp rate constraint boundary that limits the power throughput change rate of the public HVAC equipment cluster. When the uncertainty of the prediction indicated by the covariance matrix increases, it forcibly shrinks the ramp rate constraint boundary to construct a physical defense envelope to suppress power surges on the grid side. It achieves real-time adaptive decoupling between the control law gain and the reliability of the input data. The variable stiffness power boundary locking module constitutes a general signal stress protection logic. The collaborative power allocation module is used to solve a quadratic programming problem with the goal of minimizing the total active power loss of the public HVAC equipment cluster under the hard constraints of the ramp rate constraint boundary, and to generate and issue frequency adjustment commands for each variable frequency drive execution unit in the public HVAC equipment cluster. And the variable stiffness power boundary locking module performs hard locking logic containing the following inequality constraints when constructing a physical defense envelope to suppress power impact on the grid side: wherein, is the frequency adjustment instruction value issued to the variable frequency drive execution unit in the current control period, is the frequency adjustment instruction value of the last control period, is the power change rate limit threshold obtained in real time according to the norm of the covariance matrix , is the time step of the control period; the variable stiffness power boundary locking module performs reverse damping mapping to calculate the power change rate limit threshold in real time according to the formula , wherein characterizes the maximum change slope of the frequency adjustment instruction allowed at the current time, is the maximum response rate reference value of the public heating equipment cluster under the rated physical working condition, is the Frobenius norm of the predicted covariance matrix calculated in real time, is the sensitivity attenuation coefficient determined by offline identification of historical load prediction error variance of the system; a nonlinear inverse function relationship establishes a negative correlation between the system dynamic constraint boundary and the confidence of the input signal.

2. The public heating, ventilation, and air conditioning (HVAC) cluster coordinated optimization control system based on multi-source load forecasting of claim 1, wherein, The system also includes a time-domain phase orthogonal correction module, which is used to calculate the first derivative characteristics of real-time active power data and the slope characteristics of the power demand evolution curve. The time-domain phase orthogonal correction module is used to trigger the time-domain translation operator to dynamically align the control window of the cooperative power allocation module when the phase difference between the first derivative characteristics and the slope characteristics exceeds the preset synchronization tolerance. Through time-domain feature compensation, a phase-locked loop is established between the control command output and the state evolution of the controlled object to ensure that the effective time of the frequency adjustment command and the starting point of the actual power load change are phase synchronized on the time axis.

3. The public heating, ventilation, and air conditioning (HVAC) cluster coordinated optimization control system based on multi-source load forecasting of claim 1, wherein, The collaborative power allocation module includes a low-level register overwrite unit, which is used to directly address and access the hardware control register of the variable frequency drive execution unit via the fieldbus protocol. The low-level register overwrite unit is used to directly write the ramp rate constraint boundary generated by the variable stiffness power boundary locking module into the acceleration and deceleration time parameter register of the variable frequency drive execution unit, so as to establish a physical-level power change rate limiting mechanism at the driver hardware level that is independent of the upper-level communication state.

4. The public heating, ventilation, and air conditioning (HVAC) cluster coordinated optimization control system based on multi-source load forecasting of claim 1, wherein, The variable stiffness power boundary locking module includes a distribution network voltage stability enhancement unit, which is used to monitor voltage transient data of the distribution network bus. When the voltage drop in the voltage transient data exceeds the preset voltage safety threshold, the distribution network voltage stability enhancement unit is used to trigger an emergency response mode aimed at supporting the voltage of the distribution network bus, and forcibly lock the ramp rate constraint boundary to the creep rate required to maintain the minimum steady-state operation of the system, until the voltage transient data returns to the voltage safety threshold range and remains within the preset time window.

5. A collaborative optimization control system for public HVAC equipment clusters based on multi-source load forecasting as described in claim 1, characterized in that, The stochastic power load trend prediction module includes a multi-source heterogeneous data cleaning unit, which performs spatiotemporal alignment processing on external power demand disturbance data. The external power demand disturbance data includes meteorological parameter data characterizing environmental heat load and building passenger flow density data characterizing basic power load. The multi-source heterogeneous data cleaning unit is used to remove outlier noise from the external power demand disturbance data and map the processed data into a normalized power load influence factor, which serves as the input basis for generating the power demand evolution curve.

6. A collaborative optimization control system for public HVAC equipment clusters based on multi-source load forecasting as described in claim 3, characterized in that, The variable frequency drive execution unit includes a chiller inverter as a large inertial base load power unit and a circulating pump inverter as a high-frequency regulating load power unit. The collaborative power distribution module is used to adjust the operating frequency of the large inertial base load power unit in response to low-frequency, large-amplitude power load changes, and adjust the operating frequency of the high-frequency regulating load power unit to compensate for high-frequency, small-amplitude power load fluctuations, based on the power demand evolution curve and under the premise of meeting the ramp rate constraint boundary.

7. A collaborative optimization control system for public HVAC equipment clusters based on multi-source load forecasting as described in claim 1, characterized in that, The variable stiffness power boundary locking module includes a stability gain scheduling unit, which is used to automatically switch the system's control strategy when the prediction confidence is lower than a preset confidence threshold. The stability gain scheduling unit is used to switch from a prediction-based feedforward control mode to a conservative feedback control mode based on real-time deviation, and reduce the proportional gain coefficient of the frequency regulation command to prevent power oscillations on the distribution network side caused by prediction deviation.

8. A collaborative optimization control system for public HVAC equipment clusters based on multi-source load forecasting as described in claim 1, characterized in that, The system also includes an energy efficiency optimization feedback module, which is used to collect real-time data on the actual power loss of public HVAC equipment clusters. The energy efficiency optimization feedback module is used to calculate the residual between the actual power loss data and the theoretical energy consumption model, and to use the residual to correct the weight of the objective function of the quadratic programming problem in the collaborative power allocation module online.

9. A collaborative optimization control system for public HVAC equipment clusters based on multi-source load forecasting as described in claim 1, characterized in that, The collaborative power distribution module includes a dead-zone avoidance unit, which is used to identify the inefficient operating frequency band of the variable frequency drive execution unit. The dead-zone avoidance unit is used to introduce frequency skipping constraints when generating frequency adjustment commands to ensure that the issued frequency adjustment commands avoid inefficient operating frequency bands, so that the public HVAC equipment cluster always operates in a frequency range with high energy efficiency ratio.

Images