A method, system and storage medium for energy-saving intelligent control of cabinet air conditioners
By collecting and analyzing data such as supply air, return air, and hot spot temperature, and using an airflow response inversion network to separate the degree of return air short circuit from the effective cooling capacity of hot spots, the problem of hot spot overheating caused by return air short circuit in rack air conditioners is solved, and intelligent energy-saving control of rack air conditioners is realized.
Patent Information
- Application Number
- CN202611547663.4
- Authority / Receiving Office
- CN · China
- Patent Type
- Applications(China)
- Current Assignee / Owner
- Filing Date
- 2026-10-08
- Publication Date
- 2026-11-10
AI Technical Summary
Existing cabinet air conditioners cannot accurately determine the cooling effect when the return air is short-circuited, leading to hot spots overheating and a lack of effective energy-saving control measures.
By collecting data such as supply air, return air, hot spot temperature, humidity, and cooling input power, and using an airflow response inversion network to perform dual-measurement point joint inversion, the degree of return air short-circuiting and the effective cooling capacity of hot spots are separated, and target control commands are generated to achieve intelligent management and control.
Without relying on return air temperature, it accurately judges hot spot temperature changes, avoids hot spot overheating, achieves energy-saving control of rack air conditioning, and reduces energy consumption.
Smart Images

Figure CN123065102A_ABST
Abstract
Description
Technical Field
[0001] This invention relates to the field of intelligent control technology for cabinet air conditioners, and more specifically, to an energy-saving intelligent management and control method, system, and storage medium for cabinet air conditioners. Background Technology
[0002] Unattended outdoor communication sites often use enclosed cabinets to centrally install communication power supplies, transmission equipment and batteries. The heat density inside the cabinets is relatively high and the heat dissipation conditions are limited, requiring door-mounted cabinet air conditioners for continuous cooling. The internal circulation of door-mounted cabinet air conditioners is isolated from the ambient circulation by partitions, and energy-saving control can only be carried out on the closed internal circulation air supply organization. Existing rack air conditioners mostly use return air temperature as the control quantity to adjust the compressor and fan. There are also solutions that reconstruct the temperature field or predict the temperature after collecting the temperature and then adjust the cooling output. Due to the influence of cable bundles, empty space sealing plates and equipment resistance in the rack, the cold air delivered from the evaporator side is prone to return to the return air vent in advance along the low resistance channel near the door, forming a return air short circuit. These solutions are mostly based on passively observed temperature sequences or require dense measurement points and a large amount of labeled data, making it difficult to determine whether the delivered cold air has actually reached the equipment hot spot under the edge conditions of a single rack. The severity of return air short circuits cannot be directly measured on-site. Furthermore, return air short circuits can cause the return air temperature to drop prematurely, masking the fact that the hot spots in the equipment are still heating up. Control systems based on return air temperature will then determine that the cooling is sufficient and reduce the output, causing the hot spots to overheat. At the same time, the failure of the hot spot temperature to drop may be caused by either a return air short circuit or insufficient cooling capacity at the current operating setting. These two issues are difficult to distinguish in terms of temperature, and there is a lack of technical means to identify the degree of return air short circuits online on-site and to implement energy-saving control accordingly. Summary of the Invention
[0003] This invention provides a method, system, and storage medium for energy-saving intelligent management of cabinet air conditioners, solving the technical problems in related technologies where the degree of return air short circuit cannot be directly measured and using return air temperature as the control basis can easily lead to misjudgment of sufficient cooling and cause hot spots to overheat.
[0004] This invention provides a method for intelligent energy-saving management and control of cabinet air conditioners, comprising the following steps: S1 collects supply air, return air, hot spot temperature, return air humidity, cooling input power, load proxy quantity, actuator and protection status according to the update context of the control cycle. After timing alignment, physical consistency verification and security permission determination, a status dataset is generated. S2, when the state dataset indicates that the compressor is stable and the hot spot has a safe reserve space, implement recoverable restricted air volume changes, extract the early return air response, the hot spot delayed response and the concurrent energy consumption to obtain the airflow response feature set; S3, use the airflow response inversion network to perform dual-point simultaneous inversion on the airflow response feature set, obtain the degree of return air short-circuit and the effective cooling capacity of hot spots, form a capability record of candidate operating combinations and output a set of thermal state parameters; S4. Based on the natural temperature rise rate of the thermal state parameter set and the effective cooling capacity of hot spots in the candidate operating combinations, extrapolate the future hot spot temperature from the current hot spot temperature, and combine the dew point boundary to screen the operating combinations that meet the thermal safety constraints, and generate the target control command. S5, after executing the target control command, drives the compressor and fan through the pre-interlock of power supply, pressure, water leakage and start-stop interval, compares the actual response of the hot spot with the expected response, and generates execution feedback and parameter update permission; S6 will execute feedback and parameter update permission return and verify remote settings, write control context and model parameters into non-volatile memory, and output the next control cycle update context according to the safe recovery sequence after power is restored.
[0005] Furthermore, the safety permit determination is as follows: when the following conditions are met simultaneously: there is a reserved space between the hot spot temperature and the control boundary, the power supply is stable, the pressure protection does not activate, the water leakage input is normal, the cabinet door is closed, the compressor is stable, the evaporator fan can feedback the actual speed, and the key status frames are continuously valid, the safety permit flag is set to valid; otherwise, the safety permit flag is set to invalid and a primary reason is attached. The state dataset also includes the natural temperature rise rate, derived from the slope of hot spot temperature versus time in a stable segment with no effective cooling, and a load continuity flag, derived from whether the load agent experiences a step within the observation window.
[0006] Furthermore, the restricted airflow change is a closed loop consisting of the current speed, the target speed, and the recovery speed. The compressor maintains its original output state within the identification window, and the recovery segment is used to verify whether the response direction corresponds to the airflow change. When extracting the early return air response and the delayed hot spot response, the return air residual sequence and the hot spot residual sequence are obtained by using the pre-excitation baseline trend. These are then used as inputs to the airflow response inversion network by resampling the evaporator fan speed, cooling input power, and phase markers into a multi-channel sequence of equal length. At the same time, the periods with load steps are marked as external disturbances by load proxy and load continuity flags.
[0007] Furthermore, the airflow response inversion network consists of a temporal convolutional network backbone, a single-head temporal attention unit, and dual output heads; The dual-output head outputs the degree of return air short-circuit and its uncertainty, as well as the unit effective air volume hot spot cooling coefficient and its uncertainty. The envelope parameter head then regresses the start time and shape parameters of the hot spot delay response, the early return air response, and the effective component return response, respectively.
[0008] Furthermore, the dual-measurement point joint inversion is as follows: the air volume increment of the limited air volume change is divided into short-circuit component and effective component according to the degree of return air short-circuit. The early response of return air is reconstructed by the short-circuit component and the delayed response of hot spot is reconstructed by the effective component. The two reconstruction relationships are combined to separate the degree of return air short-circuit and the effective cooling capacity of hot spot. The return air cooling coefficient of the short-circuit component is anchored by the measured temperature difference between the supply air temperature and the return air temperature. The airflow response inversion network is trained without labels using hotspot reconstruction loss, return air reconstruction loss, causal symmetry loss, and load gating loss. The effective cooling capacity of the hot spot is obtained by multiplying the hot spot cooling coefficient per unit effective air volume by the effective air volume. The capacity is recorded according to the candidate operating combination, combined with the compressor output status, evaporator fan speed, supply air temperature range, ambient temperature range and hot spot location.
[0009] Furthermore, the future hotspot temperature is obtained by adding the product of the natural temperature rise rate and the control time domain to the current hotspot temperature, and then subtracting the product of the effective cooling capacity of the hotspot and the control time domain, and superimposing the prediction reserved space. The prediction reserved space is converted from the uncertainty output by the airflow response inversion network. The greater the uncertainty, the larger the prediction reserved space. During the screening process, operating combinations with expected supply air temperatures below the sum of the dew point boundary and the anti-condensation reserve space are excluded, while operating combinations that pass the hot spot temperature, dew point boundary, start-stop interval, and protection check are considered as safe candidates. When comparable input power records are available, the operating combination with the lowest input power and valid parameter confidence status is selected from the safety candidates. Operating combinations without current rack measured capacity records are not included in the energy reduction selection.
[0010] Furthermore, the pre-execution interlock drives the compressor and fan when the power supply is stable, the pressure protection does not activate, the water leakage input is normal, and the compressor start-stop interval is met; otherwise, it maintains the current permitted state. When the status frame is continuously valid, the command and feedback are consistent, the compressor does not activate protection, the cabinet door is not opened, the load agent quantity does not change, the hot spot address does not change, the power data is complete, and the actual response direction is consistent with the expectation, the parameter update permission is set to valid and the capacity record is refreshed accordingly; when the parameter update permission is invalid, the capacity update of the corresponding operating combination is frozen, and the original capacity record is not used to reduce the cooling output.
[0011] Furthermore, the non-volatile memory uses a current area and a candidate area for storage. The candidate area is first written with data content, version, length and check value, and then marked as the current valid version after readback verification is completed. After power is restored, the compressor is kept stopped and the power supply, sensors and fans are self-tested. When the equipment layout version, sensor calibration version and model version are consistent, the model parameters of the capacity record and airflow response inversion network are used for recovery. When the versions are inconsistent or the validity period expires, it is marked as pending verification and conservative temperature control is adopted.
[0012] Furthermore, a cabinet air conditioning energy-saving intelligent control system includes: The data acquisition module collects supply air, return air, hot spot temperature, return air humidity, cooling input power, load proxy quantity, actuator and protection status based on the update context of the previous control cycle. After timing alignment, physical consistency verification and security permission determination, it generates a status dataset. The response extraction module, when the state dataset indicates that the compressor is stable and the hot spot has a safe reserved space, implements recoverable restricted air volume changes, extracts the early return air response, the hot spot delayed response, and the concurrent energy consumption, and obtains the airflow response feature set. The capability assessment module uses an airflow response inversion network to perform dual-measurement point simultaneous inversion on the airflow response feature set, obtains the degree of return air short-circuiting and the effective cooling capacity of hot spots, forms a capability record of candidate operating combinations and outputs a set of thermal state parameters. The instruction generation module extrapolates the future hot spot temperature from the current hot spot temperature based on the natural temperature rise rate of the thermal state parameter set and the hot spot effective cooling capacity of the candidate operation combination. It then combines the dew point boundary to screen the operation combination that meets the thermal safety constraints and generates the target control instruction. The feedback update module, after executing the target control command via power supply, pressure, water leakage and start-stop intervals, drives the compressor and fan, compares the actual response of the hot spot with the expected response, and generates execution feedback and parameter update permission. The power failure recovery module will execute feedback and parameter update permission return and verify remote settings, write the control context and model parameters into non-volatile memory, and output the next control cycle update context according to the safe recovery sequence after power is restored.
[0013] Furthermore, a computer-readable storage medium stores a program thereon, which, when executed by a processor, implements the steps in the above-described intelligent energy-saving control method for cabinet air conditioning.
[0014] The beneficial effects of this invention are as follows: When the compressor output is stable, a recoverable limited air volume change consisting of the current speed, target speed, and recovery speed is applied, and the airflow response inversion network performs a dual-measurement joint inversion of the early return air response and the hot spot delayed response. Since a return air short circuit will enhance the early return air response and weaken the hot spot delayed response, the network anchors the short circuit component with the early return air response and the effective cooling with the hot spot delayed response. The return air cooling coefficient of the short circuit component is anchored by the measured temperature difference between the supply air and the return air. Thus, without the short circuit true value label, the degree of return air short circuit and the effective cooling capacity of the hot spot are separated. This allows energy-saving control to judge based on the actual cooling amount obtained by the hot spot rather than the surface temperature drop of the return air, avoiding hot spot overheating caused by return air short circuit. The system selects operating combinations based on the effective cooling capacity of hotspots and input power from among the safety candidates checked through hotspot temperature, dew point boundary, and protection checks. Operating combinations without actual measured capacity records for the current rack are not included in the energy consumption reduction selection to avoid optimistic extrapolation. This reduces rack air conditioning energy consumption while maintaining the hotspot safety boundary. At the same time, through versioned storage of the current zone and candidate zones, parameter update permission, and power-on self-test recovery, abnormal segments are not written into the effective capacity record. Local pressure, leakage, power supply, and temperature protection do not rely on learned parameters, ensuring the continuity and reliability of control data before and after power failure. Attached Figure Description
[0015] Figure 1 This is a structural diagram of an energy-saving intelligent control system for cabinet air conditioning according to the present invention; Figure 2 This is a flowchart of an energy-saving intelligent control method for cabinet air conditioners according to the present invention; Figure 3 This is a flowchart of an energy-saving intelligent control method for cabinet air conditioners according to the present invention. Detailed Implementation
[0016] The subject matter described herein will now be discussed with reference to exemplary embodiments. It should be understood that these embodiments are discussed only to enable those skilled in the art to better understand and implement the subject matter described herein, and changes may be made to the function and arrangement of the elements discussed without departing from the scope of this specification. Various processes or components may be omitted, substituted, or added as needed in the examples. Furthermore, some features described in the examples may be combined in other examples.
[0017] At least one embodiment of the present invention discloses an energy-saving intelligent control method for cabinet air conditioners, such as... Figure 1The diagram shown is an overall structural diagram of an energy-saving intelligent control system for cabinet air conditioning according to the present invention. The method consists of four parts: sensing and protection acquisition, controller, refrigeration actuator, and storage and remote control. The sensing and protection acquisition includes a supply air temperature sensor, a return air temperature and humidity sensor, a hot spot temperature sensor, a power acquisition unit, and pressure leakage cabinet door protection, which provides sensing signals to the controller. The controller operates in a six-step cycle. The first step involves acquiring status data and granting safety permissions based on the update context of the previous control cycle, and outputting a status dataset. The second step, when the compressor is stable and the hotspot has sufficient safety margin, implements recoverable restricted airflow changes, extracting the early return air response, the delayed hotspot response, and concurrent energy consumption to obtain an airflow response feature set. The third step uses an airflow response inversion network to perform a dual-measurement point simultaneous inversion on the airflow response feature set, obtaining the unmeasurable return air short-circuit degree and the effective cooling capacity of the hotspot, and outputting a thermal state parameter set. The fourth step, based on the natural temperature rise rate and the effective cooling capacity of the hotspot… The system extrapolates future hotspot temperatures and, after screening operating combinations that meet thermal safety constraints based on dew point boundaries, generates target control commands. In the fifth step, after implementing pre-interlocking for power supply, pressure, water leakage, and start-stop intervals, control commands are sent to the compressor, evaporator, and condenser. After comparing the actual hotspot response with the expected response, execution feedback and parameter update permission are output. In the sixth step, the control context and model parameters are written to a non-volatile storage medium, and data is exchanged with a remote platform via RS485. After power is restored, the updated context for the next control cycle is output according to the safety recovery sequence. This updated context is sent back to the first step to form a closed loop. Each step is connected using versioned data; data with inconsistent versions will not be processed subsequently. Airflow response identification is not forcibly activated in every cycle; even when identification is not enabled or the segment is invalid, cooling is maintained based on hotspot temperature, return air temperature, dew point boundary, and independent hardware protection. Specifically, as follows... Figures 2 to 3 As shown, it includes: S1 collects supply air, return air, hot spot temperature, return air humidity, cooling input power, load proxy quantity, actuator and protection status according to the update context of the control cycle. After timing alignment, physical consistency verification and security permission determination, a status dataset is generated. Specifically, in each control cycle, the controller first reads the update context of the previous control cycle to obtain the valid sensor set and recovery status, and then adds the acquisition time and acquisition sequence number to each sampled value using the local monotonic clock; sensors directly connected to the controller's analog-to-digital conversion port write to the same circular buffer after sampling, while sensors connected via the communication bus carry both the node address and the receiving time; the controller does not use the data arrival order to replace the data occurrence order, but aligns the data from each channel into the same state frame according to the allowed time difference, which is a prerequisite for subsequent timing identification to be successful; The status frame includes at least the supply air temperature, return air temperature, hot spot temperature, return air humidity, compressor command status, compressor feedback status, evaporator fan command speed, evaporator fan feedback speed, cooling input power, load proxy, power supply voltage, pressure protection status, water leakage status, cabinet door status, and communication quality; when multiple hot spot sensors are present, the status frame also includes the temperature of each hot spot and the address corresponding to the highest hot spot. The load proxy is used to characterize the heat intensity of the equipment in the cabinet and is a necessary input for distinguishing between airflow problems and changes in operating conditions. When the cabinet is configured with equipment-side current acquisition or total cabinet power acquisition, the controller directly uses the current or power as the load proxy. When neither is available, the controller constructs an equivalent load proxy by combining the same-direction change amplitude of multiple hot spot temperatures with the change in the unit's own cooling input power. The load proxy is aligned with the temperature in the same frame, so that changes in hot spot temperature can be distinguished as being caused by airflow or equipment heat. The power acquisition unit collects the refrigeration input power composed of the compressor, evaporator fan and condenser fan, or collects the power of each actuator separately and sums them in the controller for the same time period; the power value is only associated with the temperature response segment covering the same start and end time, and the power value of the previous control cycle cannot be combined with the current response segment to avoid obtaining the wrong hot spot effective cooling capacity. The compressor output status is determined by both commands and feedback; the feedback for variable frequency compressors includes the actual operating frequency or drive confirmation status, while the feedback for fixed frequency compressors includes the contactor status and operating current; if the control command has changed but the feedback has not yet stabilized, the status frame is marked as an execution transition and is not used for airflow response identification or for updating the capability record of candidate operating combinations. The controller saves the calibration version, installation location identifier, and the most recent valid time for each sensor. After maintenance personnel replace the sensor or change the installation location, a new calibration version must be generated. The new calibration version can continue to participate in temperature protection, but the airflow response parameters accumulated in the original version are marked as pending verification until a new restricted airflow change obtains a valid response segment. In one embodiment of the present invention, the controller first performs data integrity verification; if a required field in the status frame is missing, the sampling time difference exceeds the allowable range, the check code is incorrect, or the sensor address is inconsistent with the configuration, the status frame shall not be used as identification input; if the continuous missing fields reach the maintenance criterion, the controller records the sensor or communication failure and switches the control basis to the still valid highest safety level temperature information. Then, a physical consistency check is performed; the controller determines whether the temperature value falls within the sensor's range, whether the change per unit time can be explained by the current status of the fan and compressor, and compares the relative relationship between the supply air temperature, return air temperature, and hot spot temperature; if the supply air temperature is consistently higher than the return air temperature during continuous cooling operation, or if a certain temperature jumps between adjacent samples while other states remain unchanged, the sensor is marked as suspicious. Suspicious markers do not immediately equate to sensor failure; the controller checks adjacent sensors, compressor current, pressure status, and cabinet door status to distinguish between sudden changes in equipment load, cabinet door opening, compressor protection, and sensor abnormalities; when multiple hotspots exist, suspicious hotspots do not participate in the highest value comparison, but their fault status is continuously uploaded until continuous effective sampling or maintenance reset meets the recovery conditions. Return air humidity and return air temperature are used together to determine the dew point boundary; when humidity data is missing, the controller does not assume that the air is dry, but uses a preset conservative supply air lower limit; after the humidity sensor recovers, it needs to be continuously and stably sampled before the real-time dew point boundary is reactivated to avoid a single recovered value immediately causing a change in compressor output; Safety authorization requires the following conditions to be met simultaneously: there is a reserved space between the hot spot temperature and the control boundary; the power supply is stable; the pressure protection is not activated; the water leakage input is normal; the cabinet door is closed; the compressor output is stable; the evaporator fan can provide feedback on the actual speed; and the critical status frames are continuously valid. If any condition is not met, the safety authorization will be invalid, with a primary reason and related status attached. The controller generates a load continuity flag based on whether a step occurs in the load agent quantity within the observation window; when the load agent quantity changes smoothly within the allowable fluctuation range, the load continuity flag is set to continuous; when a step exceeds the allowable amplitude or a continuous rise occurs, it is set to discontinuous, and the time of occurrence is recorded; this flag is used for triggering judgment and subsequent response attribution of limited air volume changes, avoiding activation of identification or updating of capacity records during periods of drastic changes in equipment load. The natural temperature rise rate is uniformly obtained in this step and used throughout the process. It represents the upward trend of the hot spot temperature in the short time domain when there is no additional effective cooling capacity under the current load. It is obtained from the slope of the hot spot temperature with respect to time in a stable segment where there is no effective cooling capacity, such as when the compressor is at low output or shut down. The hot spot safety margin is determined based on the observation duration that the limited air volume change may continue and the natural temperature rise rate. The faster the temperature rise, the more margin is required. If the natural temperature rise rate cannot be obtained from the effective segment, the controller adopts the conservative temperature rise boundary verified by engineering and does not start identification by lowering the judgment requirements. The final state dataset with version number is generated; this dataset includes aligned state frames, valid sensor set, highest hot spot location, current dew point boundary, natural temperature rise rate, compressor stability flag, load continuity flag, safety clearance flag, and clearance reason; subsequent steps must verify that the dataset version is consistent with the current cycle, and no air volume change is implemented if the versions are inconsistent.
[0018] S2, when the state dataset indicates that the compressor is stable and the hot spot has a safe reserve space, implement recoverable restricted air volume changes, extract the early return air response, the hot spot delayed response and the concurrent energy consumption to obtain the airflow response feature set; Specifically, this step receives the status dataset output from the previous step, first verifying the safety permission, compressor stability flag, load continuity flag, and evaporator fan controllable status; the controller selects a target speed adjacent to the current fan speed and verified by the equipment airflow range, so that this action is sufficient to form an identifiable change at the return air measurement point, while not exceeding the fan's allowable range and the equipment's air supply requirements; The restricted airflow change is represented by a closed loop consisting of the current speed, the target speed, and the recovery speed. The controller first maintains the current speed and obtains the pre-excitation trend, then adjusts the fan to the target speed, completes the early return air observation and hot spot delay observation, and then restores the original speed. The recovery phase is used to verify whether the response direction corresponds to the airflow change, rather than relying on a single unidirectional action to draw conclusions. This closed loop is also the source of the causal symmetry constraint in the network training later. The compressor maintains its original output state throughout the identification window; the target frequency of the variable frequency compressor remains unchanged, and the fixed frequency compressor must not be started or stopped within the window; if pressure regulation, drive protection, or upper-level command causes the compressor state to change, the controller immediately terminates the identification, restores the fan to the safe speed before entering the identification, and marks the segment as compressor state interference. The condenser fan operates according to the pressure control of the refrigeration system itself, and its state changes are recorded as auxiliary information. If the condenser fan changes cause the compressor power or refrigeration pressure to exceed the stability criterion, this segment will not be used for calculation. Therefore, the main controllable changes used for identification are limited to the evaporator-side air volume. During periods of restricted airflow variation, the controller continuously monitors hotspot temperature, temperature rise trend, power supply, pressure, water leakage, and cabinet door status. If any safety condition becomes invalid, the controller will immediately terminate the action and restore the safe airflow without waiting for the observation window to close. Data obtained before termination is only used for fault tracing and not for updating the degree of return air short circuit or candidate capabilities. Set recovery conditions between repeated identifications; only when the hot spot temperature, return air temperature, compressor output and power return to a stable state, and the thermal response generated by the previous action has decayed, is it allowed to change the air volume again; this condition avoids the superposition of continuous actions and frequent changes in fan status. In one embodiment of the present invention, the controller uses the pre-excitation state frame to determine the short-term reference trends of return air temperature and hot spot temperature; the reference trend only extends to the current observation window and is not used as a long-term thermal model; the temperature value after the fan reaches the target speed is compared with the reference value at the same time to form the return air residual sequence and the hot spot residual sequence, and the sampling quality flag corresponding to each residual is retained; the residual is the difference between the measured temperature and the reference trend, which strips away the slow changes that already exist before excitation, so that the part caused by the change in air volume can be observed separately; Early return air observation begins after the fan feedback reaches the target speed, covering the response phase required for the airflow to reach the return air inlet from the evaporator; delayed hotspot observation covers the phase after the airflow passes through the equipment inlet, internal heat exchange path, and equipment outlet, affecting the hotspot sensor; the two observation phases can be partially connected, but the hotspot conclusions cannot be replaced solely by the data from the return air observation phase. The controller resamples the return air residual sequence, hot spot residual sequence, supply air residual sequence, evaporator fan command and feedback speed, compressor output status flag, cooling input power, and phase marker at the relative excitation time into a multi-channel sequence of equal length at a uniform sampling interval, which is then used as the input to the inversion network in the next step. The sensor noise range is obtained from the static operation segment of the equipment. Changes that do not exceed the noise range are marked as low-confidence samples in the sequence, which are used by the network to adjust the weights. Simultaneous energy consumption is accumulated from the time the fan reaches the target speed until the hot spot delay observation ends. The state frame used for power integration must have the same time boundary as the temperature sequence. If power sampling is interrupted, the airflow timing response can still be retained for maintenance diagnosis, but the record of the effective cooling capacity of the hot spot corresponding to the unit energy consumption cannot be generated, nor can it participate in the minimum power selection in subsequent steps. After returning to the original speed, the controller continues to observe the reverse changes of the return air and the hot spot; if the return air cools down when the speed is increased and rises again when the speed is restored, and both directions are consistent with the fan feedback, the return air response obtains a repeatability mark; the hot spot response is also checked according to the same principle, but considering the delay caused by the equipment's heat capacity, it is not required that the two measuring points change at the same time; If a load jump occurs in the equipment inside the cabinet during the observation period, the hot spot temperature may change rapidly while the return air change lags behind. The controller identifies this situation based on the load proxy quantity and load continuity flag generated in the first step, and marks the period when the load continuity is discontinuous as an external disturbance. This flag, along with the load proxy quantity, is written into the feature set and used as the basis for load gating in the next step. Output airflow response feature set; this feature set includes fan speed before and after action, compressor steady state, early return air residual response, hot spot delay residual response, supply air residual sequence, two response times, synchronous energy consumption, recovery phase response, load continuity, load proxy, resampled multi-channel input sequence, segment validity and invalid reasons; the next step only receives feature sets where segment validity is valid.
[0019] S3, use the airflow response inversion network to perform dual-point simultaneous inversion on the airflow response feature set, obtain the degree of return air short-circuit and the effective cooling capacity of hot spots, form a capability record of candidate operating combinations and output a set of thermal state parameters; Specifically, the degree of return air short-circuiting is a quantity that cannot be directly measured, and there is no true value available for labeling on-site. At the same time, the temperature response caused by limited airflow changes is often weak and is also coupled with equipment load changes and sensor noise. Using only a fixed threshold to compare the order of return air and hot spot responses can only provide a rough binary judgment, which will fail when encountering weak responses, uncertain delays, or slow load increases. Therefore, this step uses a physically guided airflow response inversion network to map the multi-channel input sequence from the previous step into continuous estimates of two latent variables: the degree of return air short-circuiting and the effective cooling capacity of hot spots, with their respective uncertainties. The backbone of the network adopts a temporal convolutional network (TCN). The input is a multi-channel sequence of equal length, with each channel corresponding to a residual or running state. The network consists of several causal dilated convolutional residual blocks stacked sequentially. Each residual block first performs a one-dimensional causal convolution, then nonlinear activation, and is superimposed with a shortcut connection that directly adds the input to the output. Causal convolution only allows the output to depend on the current time and samples before it, which conforms to the physical fact that the response occurs unidirectionally along time and also allows the network to perform incremental computation as data arrives point by point. The convolution dilation rate of adjacent residual blocks is doubled block by block, so that the later residual blocks can cover a longer time span without increasing the parameters within the layer, thereby adapting to the different hotspot delay lengths caused by different wind channel structures. After passing through all residual blocks, the sequence is encoded into a set of temporal feature vectors of equal length. A single-headed temporal attention unit is set on top of the temporal features to automatically focus on the time period when hotspots actually begin to respond, instead of relying on a manually set fixed observation window. The query vector, key vector, and value vector of the attention are obtained from the temporal features through three sets of linear mappings. The correlation between the query vector and the key vector at each time point is converted into weights by a normalized exponential function. Then, the value vectors at each time point are weighted and summed to obtain a compact response representation vector. The normalized exponential function is a conventional operation, and its expression is not elaborated here. The compact representation vector is fed into two output heads. The return air short-circuit degree head outputs a point estimate of the return air short-circuit degree and a logarithmic variance. The point estimate is normalized to a closed interval of zero to one. Zero indicates that the airflow is smooth and the increased air volume basically acts on the heat source. One indicates that almost all of the increased air volume returns directly to the return air vent through the low-resistance channel. The hot spot effective cooling capacity head outputs the hot spot cooling coefficient per unit effective air volume and its logarithmic variance. In addition to these two latent variables for subsequent use, the same representation vector also passes through a lightweight envelope parameter head to regress the start time and shape parameters of the hot spot delayed response, the early return air response, and the effective component return response, respectively, thus giving three independent weight functions of the duration of action, instead of sharing a single attention weight. The two logarithmic variances describe the uncertainty of this inversion. Segments with poor data quality or weak responses will have greater uncertainty, reserving space for the next step of amplifying the prediction. The reason the network can inversely determine the degree of return air short-circuit without a short-circuit truth label is that it has a built-in dual-measurement point reconstruction relationship that conforms to the actual airflow distribution, and requires the reconstruction result to simultaneously approximate the measured return air residual and hot spot residual. The airflow increment caused by the constrained airflow change is divided into two paths: the short-circuit component returns directly to the return air inlet through the low-resistance channel, mainly causing an early rapid decrease in return airflow without significantly affecting the hot spot; the effective component passes through the heat source path, first causing a delayed decrease in the hot spot, and then returning to the return air inlet. Based on this, the hot spot delay response reconstruction formula is:
[0020] The early response reconstruction formula for return air is:
[0021] in, and These represent the inversion network at the sampling time. The reconstructed hotspot temperature and return air temperature; and These are the extrapolation of the pre-incentive hotspot baseline trend and the return wind baseline trend to the time point. Temperature; The value range for the degree of return air short circuit is a closed interval from zero to one. This refers to the increase in evaporator fan airflow caused by changes in limited airflow. The hot spot cooling coefficient per unit effective air volume; The return air cooling coefficient for the short-circuit component is the unit airflow cooling coefficient of the short-circuit component at the return air inlet. , , These are the weighting functions of the duration of action from their respective start times for the hotspot delayed response, early return air response, and effective component return response. , , This represents the start time of the corresponding response; It should be noted that the dimensional relationship between the coefficients and the duration weighting function is as follows: The unit effective air volume hotspot cooling coefficient is defined according to unit time. Multiplying it by the effective air volume yields the hotspot cooling rate, which is then multiplied by the duration weighting function to obtain the cumulative cooling amount at that sampling moment. Subtracting this from the baseline temperature gives the temperature dimension. The two reconstructed coefficients have the same dimensions. The duration weighting function describes the process of the response gradually accumulating from the initial moment and then falling back after the air volume is withdrawn. Its dimension is time. The short-circuit component return air cooling coefficient and the hotspot cooling coefficient take the same unit air volume and unit time caliber. The above two equations incorporate the physical process of splitting the airflow increment according to the short-circuit degree into the network. The airflow increment is obtained by converting the command speed and feedback speed of the evaporator fan, and is a directly measurable quantity. The short-circuit component of the airflow passes almost no heat source, and the unit airflow temperature drop it causes at the return air inlet is close to the measured temperature difference between the supply air temperature and the return air temperature. Therefore, the return air cooling coefficient of the short-circuit component is directly anchored by the supply air residual and the return air residual, and is not treated as a free unknown quantity. In this way, the amplitude of the early return air response corresponds to the product of the short-circuit degree and this anchoring coefficient. Given a known coefficient, the degree of short-circuiting of the return air can be determined separately. The magnitude of the hot spot delayed response corresponds to the product of the effective component and the hot spot cooling coefficient per unit effective air volume. After the degree of short-circuiting is known, this cooling coefficient can be determined, thus separating the two originally intertwined quantities: the degree of short-circuiting and the cooling capacity. If only the hot spot is used as a single relationship, the failure of the hot spot to decrease can be interpreted as either a large short-circuit or a weak cooling capacity, and the two cannot be distinguished. This is precisely why it is necessary to introduce the early response of the return air and anchor the cooling coefficient of the short-circuit component. The three applied duration weight functions each characterize the process of the corresponding response accumulating and then falling back as the application duration increases. They are obtained by regressing the starting time and shape parameters of the envelope parameter head, respectively. The start of the hot spot delayed response is significantly later than the early return air response. The two do not share parameters, so there is no need to preset a fixed delay, nor is it an untraceable black function. These weight parameters remain stable for the same duct structure after training convergence. This reconstruction relationship describes a phenomenon unique to return air short circuits that is contrary to conventional intuition: the faster the return air cools down after the air is increased while the hot spot remains still, the more severe the short circuit is, rather than the more effective the cooling. Based on this, the network takes the strength of the early response of the return air as positive evidence of a short circuit, rather than taking the drop in return air temperature directly as effective cooling, thus fundamentally correcting the misjudgment when return air temperature is the only control variable. In one embodiment of the present invention, the network training is carried out in a self-supervised manner during prototype debugging and field operation, without the need for manual labeling of the return air short-circuit degree; each effective restricted air volume change segment constitutes a training sample, which carries not the artificial short-circuit value, but the measured return air residual sequence and hot spot residual sequence of that segment. The supervision signal comes from whether the two measured sequences can be reproduced by dual-measurement point reconstruction; the total loss is composed of four weighted sums, namely hot spot reconstruction loss, return air reconstruction loss, causal symmetry loss and load gating loss. The training objective is to enable the network output to explain the short-circuit degree and cooling capacity of the measured response under the common constraints of these four factors. Hotspot reconstruction loss is the mean square error between the reconstructed hotspot temperature sequence and the measured hotspot residual sequence; return air reconstruction loss is the mean square error between the reconstructed return air temperature sequence and the measured return air residual sequence. These two factors work together to ensure that the return air short-circuit degree and the hotspot cooling coefficient per unit effective airflow must simultaneously explain both the early return air response and the delayed hotspot response. It is the introduction of return air reconstruction loss that allows the short-circuit degree and cooling capacity to be separated, rather than relying solely on a single hotspot curve to make them inseparable. Causal symmetry loss utilizes the closed-loop action formed by the acceleration and recovery segments of the limited airflow change; when airflow is smooth, the response directions of the two segments are opposite. Conversely, the magnitudes are comparable, and the more severe the short circuit, the weaker and more asymmetrical the responses of the two segments. This penalty is applied when the network inversion results do not conform to this symmetric relationship. The load gating loss identifies load step segments and reduces their weights based on the load continuity flag and load proxy given in the previous step, to prevent the network from mistaking the hotspot temperature rise caused by increased equipment heat generation as insufficient cooling caused by a short circuit. The weights of the four losses are determined based on the verification segments during the prototype debugging phase. For example, in this embodiment, the default weights of the return air reconfiguration loss, causal symmetry loss, and load gating loss relative to the hotspot reconfiguration loss are 1.0, 0.5, and 0.3, respectively. Mean squared error is a routine calculation. Network parameters are updated using conventional error backpropagation and mini-batch stochastic gradient descent, and their iterative formulas will not be elaborated here. The uncertainty corresponding to the logarithmic variance is used in training in a heteroscedastic manner, so that the network actively outputs greater uncertainty for segments with weak response and low sampling quality, rather than forcibly giving a definite degree of short circuit. During the prototype debugging stage, the initial weights are obtained by offline pre-training using accumulated effective segments. A certain air duct can also be temporarily blocked or opened to create a small number of segments that are slightly unobstructed or slightly short-circuited as weak calibration samples to fine-tune the absolute scale of the degree of short circuit. This is consistent with the previous pre-survey and prototype setup practices and does not introduce additional on-site burden. On-site retraining of network weights is carried out in a low-frequency batch manner, rather than online training segment by segment. New weights are first written into the candidate area for back-read verification and are only upgraded to the current model version after being confirmed by repeatability. Network inference is only performed when an open segment is identified and valid. The output return air short-circuit degree is updated based on multiple valid segments, but the new segment must be consistent with the current sensor calibration version, equipment layout version, and hotspot address. If the equipment layout version changes after cabinet maintenance, equipment plugging / unplugging, or cable rearrangement, the old parameters are retained for traceability but do not directly participate in the control of the new version. If adjacent valid segments give opposite conclusions, the controller does not simply average to mask the difference, but checks the load status, hotspot address, and fan operation direction, and sets the return air short-circuit degree to pending verification. In the pending verification state, the hotspot temperature boundary can still be used in the next step, but the return air short-circuit degree is not used to reduce the fan output. The effective cooling capacity of a hot spot is defined as the rate at which a candidate operating combination can achieve a hot spot cooling rate under the current rack layout, i.e., the amount of hot spot cooling per unit time. It is obtained by multiplying the hot spot cooling coefficient per unit effective air volume output by the effective air volume of the combination. The effective air volume is the part of the air volume increment after deducting the short-circuit component. Therefore, the cooling capacity already includes the reduction of return air short circuit. This capacity, together with the compressor output status, evaporator fan speed, supply air temperature range, ambient temperature range, and hot spot address, serves as a record index to avoid mixing data from different condensing conditions or different hot spot locations into a single capacity value. The controller also generates an energy record corresponding to this capacity based on the energy consumption during the same period. This energy record is only used for comparing combinations of adjacent operating conditions within the same rack and is not expressed as a cross-product general energy efficiency indicator. A single restricted airflow change only excites the current combination to reach the target speed, thus only updating the capability record of the excited combination. The capabilities of other candidate operating combinations come from the records accumulated during their respective executions. For normal stable operating segments, the controller can also record the rate of change of hot spot temperature under the current candidate operating combination, but it needs to exclude start-up transitions, compressor frequency adjustment transitions, cabinet door openings, load steps, and protection actions. Stable operating records and restricted airflow change records have different source markers, and the latter step prioritizes using records verified by airflow closure actions and network inversion. The variable frequency implementation saves capability records according to verified discrete output states and does not arbitrarily interpolate between two output states. The fixed frequency implementation saves the effective cooling capacity of hot spots during the compressor operation phase and the natural temperature rise rate during the compressor shutdown phase, and constrains the control cycle through the shortest operating time and the shortest shutdown time, without using short-cycle start-stop as an energy-saving means. The parameter confidence state is jointly determined by fragment repeatability, data integrity, record freshness, environmental condition coverage, and network output uncertainty. It is a traceable rule result rather than a score given by the network alone. Each state can be traced back to a specific fragment, such as whether the response direction is reproduced, whether the power data is complete, whether the calibration version is consistent, whether the current environment falls within the applicable range of the record, and whether the uncertainty of this inversion is within the allowable range. The capability record has an expiration period. Even if the equipment layout remains unchanged, filter dust accumulation, changes in fan performance, and refrigerant status will change the actual response. After the expiration period, the capability record can only be used as a conservative reference. The system needs to obtain a valid fragment again when the safety conditions are met next time. Output a set of thermal state parameters; this set of parameters includes the current return air short-circuit degree and its uncertainty, short-circuit determination status, effective cooling capacity of hotspots for candidate operating combinations that have been recorded, synchronous power records, natural temperature rise rate and its availability status, applicable environmental range, hotspot address, sensor calibration version, equipment layout version, parameter confidence status and validity period; the natural temperature rise rate generated in the previous step is transmitted along with the same version of this parameter set, and can be directly used for hotspot temperature estimation in subsequent steps.
[0022] S4. Based on the natural temperature rise rate of the thermal state parameter set and the effective cooling capacity of hot spots in the candidate operating combinations, extrapolate the future hot spot temperature from the current hot spot temperature, and combine the dew point boundary to screen the operating combinations that meet the thermal safety constraints, and generate the target control command. Specifically, this step receives the thermal state parameter set output from the previous step and reads the state dataset of the same version as the parameter set. The controller first determines the current temperature and corresponding address of the hot spot, and then uses the natural temperature rise rate transmitted with the thermal state parameter set without repeating the calculation. This rate is not updated during cabinet door opening, equipment startup, load step, and sensor suspicion. When the natural temperature rise is unavailable, the conservative temperature rise boundary obtained from engineering verification is adopted, and the prediction reserve space for candidate screening is increased. The controller establishes a candidate operating combination table; each combination in the variable frequency embodiment includes a discrete compressor output state and a permissible evaporator fan speed, and each combination in the fixed frequency embodiment includes a compressor operating state and a permissible fan speed; each combination is associated with the hot spot effective cooling capacity, recent input power, applicable environmental range and parameter confidence state obtained from the previous inversion step; For each candidate operating combination, the controller estimates the hot spot temperature at the end of the next control time domain, starting from the current hot spot temperature. The estimation method is as follows: first, the natural temperature rise rate is multiplied by the control time domain length to obtain the temperature rise when there is no cooling effect; then, the effective cooling capacity of the hot spot in this combination is multiplied by the same time domain length to obtain the effective cooling amount. The two are superimposed and canceled out by the current hot spot temperature, that is, the current hot spot temperature plus the temperature rise and then minus the effective cooling amount, to obtain the baseline estimate of the future hot spot temperature. Since the cooling capacity comes from network inversion, the controller also needs to convert the uncertainty of the network output into a prediction reserve space and add it to the baseline estimate. The greater the uncertainty, the larger the reserve space, so that combinations with insufficient data can bear more stringent temperature screening conditions. Effective cooling capacity is directly related to the degree of return air short-circuiting. The greater the degree of return air short-circuiting in the previous step, the less effective airflow can reach the hot spot at the same fan speed, and the lower the corresponding effective cooling capacity of the hot spot. Therefore, even if the fan speed is increased under high short-circuit conditions, the estimated future hot spot temperature will not decrease significantly. Based on this, the controller no longer automatically regards higher fan speeds as stronger cooling capacity, but maintains the verified fan combination and generates airflow organization maintenance prompts. When the hot spot temperature approaches the boundary, the hot spot safety requirements still determine whether to enter protective cooling. Combinations without current rack measured capacity records are not included in the energy consumption reduction selection; if the combination is a high-cooling combination activated for protection needs, it can be used as the protection action according to the equipment's factory safety capacity, but the energy-saving capacity record will not be updated accordingly; if adjacent combinations have valid records, the controller can only use the lower bound of the more conservative capacity for hot spot cooling, and cannot presume that the unknown combination has a linear response. The controller retrieves the current dew point boundary from the status dataset and checks the supply air temperature records of candidate operating combinations. Combinations where the expected supply air temperature is lower than the sum of the dew point boundary and the anti-condensation reserve are excluded. When humidity data is unavailable, a conservative lower limit for supply air is adopted, and low supply air temperature combinations cannot be allowed due to lack of humidity. Combinations that pass hotspot temperature estimation, dew point boundary, compressor start-stop time, environmental applicable range, and all protection permissions are valid constitute a safe candidate set. If any condition fails, the reason for exclusion is retained. In one embodiment of the present invention, the controller compares the recent measured input power of each combination in the safety candidate set; the power record is from a stable segment consistent with the candidate operating combination index and selected according to the current ambient temperature range; when there is no corresponding power record in the current environment, the combination can continue to operate, but it cannot be determined as the lowest power consumption combination based solely on nominal power; when there is a comparable power record, the controller selects the combination with the lowest input power and valid parameter confidence state; when the power difference between two combinations falls within the metering resolution range, the combination that maintains the current compressor state is selected first, and then the combination with fewer fan speed changes is selected to reduce start-stop, frequency adjustment and air volume fluctuations; The target control command includes the compressor target output, evaporator fan target speed, condenser fan control mode, command generation time, command validity period, expected hot spot temperature, expected air supply boundary, used capacity record version, and selection reason; if any necessary field is missing from the command, the subsequent step must not split and execute part of the action. When the safe candidate set is empty, the controller does not lower the hot spot temperature boundary, nor does it delete the prediction reserved space; the system selects the safe cooling combination that has been verified by the equipment protection and outputs an alarm according to the reason for failure; if all combinations are excluded due to pressure, power or water leakage protection, the system enters the corresponding protection state instead of continuing to force the compressor to cool; when the return air short circuit degree is high and the safe cooling combination still cannot maintain the hot spot boundary in the prediction time domain, the controller outputs a combination alarm for abnormal airflow organization and insufficient cooling capacity. The maintenance information also provides the hot spot address, the most recent effective airflow action, the return air response and the hot spot response, which is convenient for checking the sealing plate, cable bundle, air duct seal and the air inlet and outlet direction of the equipment; The remote platform can modify hotspot control boundaries, operating modes, and allowed output ranges, but new settings require verification of address, permissions, version, and range. Remote settings cannot mask local pressure, leakage, power supply, and compressor start-stop interlocks. In the event of a communication interruption, this step continues to use the most recently confirmed local setting version and marks the remote link as interrupted in the status. Output target control commands and retain candidate screening records during the generation process for traceability; the candidate screening records include temperature estimates, dew point checks, start-up and shutdown checks, protection checks, power comparability status and final selection results for each combination, enabling energy-saving control to explain why a certain combination was selected, and enabling maintenance personnel to distinguish between insufficient parameters and insufficient equipment capacity.
[0023] S5, after executing the target control command, drives the compressor and fan through the pre-interlock of power supply, pressure, water leakage and start-stop interval, compares the actual response of the hot spot with the expected response, and generates execution feedback and parameter update permission; Specifically, after receiving the target control command, this step first checks the command validity period, status dataset version, capability record version, and current control mode. If the command has expired, the current status version has changed, or the remote mode has been legally switched, the controller abandons the command and requests re-filtering to avoid using combinations formed under old hot states for new hot states. The power interlock checks whether the input voltage is within the equipment's allowable range, whether there are continuous fluctuations in the power supply, and whether the compressor driver reports undervoltage, overvoltage, or phase loss. When the voltage is abnormal, the controller stops the compressor, limits the output, or maintains fan cooling according to the hardware capabilities, and records the trigger source. After the power is restored, it still needs to be stabilized and the compressor is not started immediately during a single sampling recovery. The refrigeration interlock checks for high-pressure protection, low-pressure protection, compressor discharge or casing temperature protection, and driver faults; hard contact protection has the ability to disconnect independently of software commands, and the software simultaneously reads the protection status and stops the relevant outputs; after the protection is released, the controller retains the fault lockout or delayed recovery requirements, depending on the equipment protection level and maintenance settings. The water leakage interlock is determined based on the water leakage input from the water tray, drain pipe, or bottom of the cabinet. When a water leakage occurs, the controller stops the cooling operation that may continue to generate condensate, maintains necessary ventilation and monitoring, and uploads the location of the water leakage. If the cabinet is equipped with an electric heater, the controller also checks the interlock between the heater and the water leakage status to prevent continued heating under unknown water accumulation conditions. The compressor start-stop interlock checks the current continuous running time, the current continuous shutdown time, the most recent protection action, and the refrigeration pressure balance status. Even if the target combination requires a change in the compressor status, if the minimum running time or the minimum shutdown time is not met, this step maintains the current allowed status and feeds back the start-stop restriction to the previous step for selection in the next cycle. When executing compressor and fan commands, the order of execution is as follows: when it is necessary to increase the cooling capacity, first confirm that the evaporator fan has formed the allowable air volume, then start or increase the compressor output, and control the condenser fan according to the pressure; when it is necessary to reduce the cooling capacity, first reduce the compressor output or complete the shutdown, and then put the evaporator fan into the corresponding ventilation state to reduce the risk of condensation caused by local temperature of the heat exchanger and residual cold. The variable frequency actuator gradually reaches the target state according to the allowed rate of change, while the fixed frequency actuator starts or stops according to the contactor feedback confirmation. When the command state and the feedback state are inconsistent, the controller does not treat the command value as an executed value, nor does it start the effect evaluation. If the consistency is still not achieved after the execution confirmation condition is exceeded, the controller outputs an actuator fault and enters the corresponding conservative state. Electric heating is used in embodiments where the ambient temperature is low and the cabinet interior needs to be protected against low-temperature operation; heating control is interlocked with compressor cooling and is simultaneously constrained by fan feedback, over-temperature protection and power capacity; the return air short-circuit identification in this embodiment is mainly used for cooling operation, and the heating state does not enter the cooling candidate capacity record to avoid data contamination between the two thermal processes; In one embodiment of the present invention, after the actuator reaches the target combination, the controller establishes an effect observation segment; the segment input includes the hot spot temperature before execution, the expected hot spot temperature, the actual compressor feedback, the actual fan feedback, the supply air temperature, the return air temperature, the hot spot temperature, the return air humidity, the input power, and the protection status; the observation time domain is consistent with the prediction time domain of the previous step or is its verifiable part; The controller compares the actual hot spot temperature change direction with the expected change direction and checks whether the actual hot spot temperature is within the range formed by the expected temperature and the predicted reserved space. If the actual hot spot temperature decreases as expected or remains within a safe range, and other states are continuously effective, this command is marked as a consistent response. This result proves that the current combination can produce the expected heat effect under this operating condition. If the actual hot spot temperature is higher than the expected range, the controller checks the load load, supply air temperature, return air short circuit degree, fan feedback, cooling pressure, and hot spot address in sequence. An increase in load corresponds to a change in operating condition, an increase in supply air temperature corresponds to a change in cooling capacity, a decrease in return air temperature but not in hot spot temperature corresponds to a change in airflow path, and insufficient fan feedback corresponds to an actuator problem. Different fault symbols are used for each cause. If the actual temperature of the hot spot is close to the control boundary, the controller does not need to wait for the complete effect observation segment to end, and immediately requests the previous step to reselect a stronger safety candidate operation combination; if no safety candidate exists, the safety cooling combination is executed and an alarm is triggered; the effect evaluation serves parameter confirmation and cannot delay temperature protection action. Parameter update permission requires simultaneous fulfillment of the following conditions: continuous validity of status frames, consistency between command and feedback, no compressor protection activation, cabinet door not opened, no step change in load proxy quantity, no change in hotspot address, complete power data, and the actual response direction consistent with the expected direction. Each condition forms a Boolean state and reason code to facilitate tracing why a certain segment did not enter the capability record. When the parameter update permission is valid, the controller only updates the data layer, that is, it refreshes the hotspot effective cooling capability record and power record of the corresponding candidate running combination with the current segment. During the update, the original record is retained, and the time, environmental range, and repeatability status of the current sample are added first. Only after the repeatability requirement is met can the new record become a usable effective capability record. The current segment is also included in the training sample set for low-frequency retraining of network weights, but does not trigger online weight updates on a segment-by-segment basis, thereby separating the frequently changing data records from the low-frequency stable model parameters for maintenance. When the parameter update permission is invalid, the controller freezes the capability update of the candidate operating combination; freezing is not the same as deleting the original record. The controller decides whether to continue using it, reduce the confidence level, or mark it for review based on the reason for invalidity; when the sensor calibration is changed, the equipment layout is changed, or the response directions are inconsistent multiple times, the original record shall not be used to reduce the cooling output; Output execution feedback and parameter update permission; execution feedback includes target combination, actual combination, interlocking results, hotspot actual response, supply and return air response, power segment, deviation type, protection action and recovery requirements; this output serves as both the storage and communication input for the next step and as the historical basis for judging the state continuity of the next cycle.
[0024] S6 will execute feedback and parameter update permission to return and verify remote settings, write control context and model parameters to non-volatile memory, and output the next control cycle update context according to the safe recovery sequence after the power is restored. Specifically, after receiving execution feedback and parameter update permission, this step constructs RS485 uplink data. The uplink data includes device address, protocol version, status dataset version, supply air temperature, return air temperature, hotspot temperature and address, operating combination, power, short circuit judgment status, parameter validity status, alarm code, and the reason for the most recent control selection. The communication frame uses a start identifier, length, device address, function identifier, data area, and integrity check field. The receiving end checks the length, address, function permission, and integrity before parsing the data. Duplicate frames are identified according to their sequence number, out-of-order frames do not overwrite new statuses, and frames that fail verification only increase the communication error count without changing local control settings. Remote downlink settings include setting version, target device address, operating mode, hotspot control boundary, allowable output range, and activation conditions. The controller first checks if the setting version is newer than the current version, then checks the value range and local permissions. The verified settings are written to the activation area. When local security conditions are met, the controller switches to the current setting and sends back the activation result. When remote commands conflict with local protection, local protection has execution priority. The remote platform can request to exit energy-saving selection, re-identify, or enter maintenance mode, but it cannot disable pressure protection, water leakage protection, power protection, and compressor start / stop constraints. During communication interruption, the local controller continues to execute the most recently confirmed setting and periodically attempts to restore the link. The non-volatile memory uses a current area and a candidate area to store the control context. The candidate area is first written with the data content, version, length, and check value. After the readback and verification are completed, it is marked as the current valid version. When the write is interrupted, the previous complete version can still be found during the startup process, avoiding the use of partially written data as the basis for control. The control context includes sensor calibration version, equipment layout version, hotspot address, valid candidate operation combination index, hotspot effective cooling capacity, power record, return air short circuit judgment status, parameter validity period, remote setting version, most recent protection event, and most recent safe operation combination. The model parameters of the inversion network, along with the equipment layout version and sensor calibration version on which it was trained, are stored in the current area, along with the model version number. Original high-frequency segments are stored in the circular record area, not sharing a valid flag with the control parameter area. Response segments with invalid parameter update permissions can be stored in the traceability area, but cannot be written into valid capability records or included in the network retraining sample set. When a parameter update permission is valid, the new segment is first written into a candidate record and verified, then its repetition status determines whether it is promoted to a valid record. Similarly, the new weights obtained from low-frequency retraining of network weights are first written into the candidate area and verified before being promoted to the current model version. This storage relationship prevents communication anomalies, power outages, or single, occasional responses from altering the energy-saving control basis. In one embodiment of the present invention, after power is restored, the controller first keeps the compressor stopped, reads the power supply voltage, pressure protection, water leakage input, cabinet door input, and various temperature sensors, and verifies the check values, version continuity, and sensor calibration versions of the current area and candidate areas. If a candidate area is incomplete, it is discarded, and the previous complete current area is used. The power supply needs to be continuously and stably confirmed, and the compressor needs to meet the shutdown time and refrigeration pressure recovery conditions. The evaporator fan first runs at the self-test speed. The controller compares the command with the feedback and observes whether the supply air and return air sensors produce reasonable changes. If the fan self-test fails, the compressor remains in a prohibited start state and outputs a fan fault. Sensor self-tests include checks on range, rate of change, address, and physical relationships. When a hotspot sensor is unavailable, if a verified backup hotspot location exists, the system switches to the backup location and increases safety margin. If no backup location exists, the system enters a conservative cooling and alarm state, reducing output without using the hotspot capability record from before the power outage. When the stored equipment layout version, sensor calibration version, model version, and current detection version are consistent, the valid capability record and inversion network model parameters can participate in candidate screening and inference after recovery. When the versions are inconsistent, the validity period expires, or the cabinet door status cannot be confirmed during the power outage, the capability record is marked as pending verification, the inversion network does not participate in reducing output temporarily, and conservative temperature control is used for the first control. When cooling is restored, the controller first determines whether it is necessary to immediately enter the safe cooling combination based on the current hot spot temperature and protection status. If emergency cooling is not required, the safe operation combination before the power outage is restored according to the compressor start and stop conditions. When multiple cabinets share the site power supply, the start requests can be staggered according to the equipment address or the upper-level scheduling to reduce the power impact of simultaneous start-up. The first complete control cycle after restoration re-executes the status acquisition, without using the limited air volume changes that did not end before the power outage. Only when the new status dataset is continuous and valid, the hot spot safety reserve space is met, and the compressor output stabilizes again can the new identification be started, so as to avoid splicing two discontinuous thermal processes before and after the power outage into a single response segment. Output the update context for the next control cycle; the update context includes the current valid settings, valid sensor set, recovery status, capability record availability status, model availability status, most recent safe operation combination, pending alarms, and communication link status; the status acquisition step reads this context and then starts a new acquisition.
[0025] In one embodiment of the present invention, an application example is provided. The following example is used to illustrate the flow of each field between steps. The example cabinet is a door-mounted sealed communication cabinet, configured with a variable frequency compressor, an adjustable evaporator fan, one supply air temperature point, one return air temperature and humidity point, and two hot spot temperature points. The controller selects the hot spot point with the higher temperature and valid status. The inversion network is deployed in the cabinet monitoring unit. The sampling data before and after the limited air volume change is shown in Table 1. In the table, compressor output status two indicates the same discrete output level and remains unchanged throughout the segment. The power is the total input power of the compressor, evaporator fan, and condenser fan at the same sampling time. Table 1 shows examples of sampling data before and after changes in restricted air volume;
[0026] When the status acquisition step receives the data from Table 1, it first confirms the address, sampling time and calibration version of the two hot spots, and then uses the point with the higher effective temperature to form the control hot spot temperature; the compressor output status in the table is continuous, the fan feedback can reach the command speed, and the power supply, pressure, water leakage and cabinet door status are all assumed to be normal. Therefore, this step outputs a safe and valid status dataset. If any protection input is abnormal, the subsequent example action will not occur. The limited airflow excitation step adjusts the evaporator fan from its original speed to the target speed at a relative time of 20 seconds and restores it at a relative time of 50 seconds, while the compressor output state remains unchanged. The controller resamples the supply air, return air, hot spot residual sequence and the fan speed, power, and phase marker into a multi-channel sequence of equal length. The return air temperature drops significantly in the early stage after the action, while the hot spot temperature only shows a limited change. After the airflow is restored, the return air temperature rises, forming a closed response segment that is fed into the inversion network. The inversion network first subtracts the baseline trend of the hotspot temperature still rising before the action, obtaining the return air residual sequence and the hotspot residual sequence. Then, two reconstruction relations are used to fit these two residuals simultaneously. The return air decreases significantly in the early stage after the action, indicating that the boosted air has indeed been delivered and quickly returned to the return air inlet, and the air volume itself is not lost. The hotspot, however, hardly moves, indicating that the delivered air does not pass through the heat source. The network sets the degree of return air short-circuiting to a high value based on the early response of the return air, and determines that the effective component is very small based on the hotspot delayed response being close to zero. This distinguishes between severe short-circuiting and insufficient cooling capacity at this level. This segment gives a low value and a large uncertainty for the hotspot cooling coefficient per unit effective air volume. The return air velocity decreases during the recovery segment and increases during the recovery segment, satisfying the causal symmetry constraint and supporting the above inversion results. The candidate screening step reads the current temperature of the hotspot, the natural temperature rise rate, and the capability records of each candidate operating combination. The high fan speed combination has a low hotspot cooling coefficient per unit effective air volume and moderate uncertainty. The estimated future hotspot temperature is not significantly lower than the current combination, but it corresponds to higher input power. Therefore, it does not receive higher priority due to the decrease in return air temperature. Among the combinations that meet the hotspot future temperature boundary, dew point boundary, and start-stop constraints, the controller selects the target combination with lower input power and valid confidence state. Before execution, the power supply, pressure, water leakage, and compressor start-stop duration are checked again. After the actuator feedback is consistent, the actual response of the hot spot is compared with the expected response. If the hot spot response meets the expectation and the load is continuous, the parameter update permission is valid. Otherwise, the airflow response is marked as insufficient and the current capability update is frozen. In the communication and storage step, the target combination, return air short circuit judgment status, hot spot response, update permission, and cause code are uploaded to the site platform. Valid samples refresh the capability record and are written into the candidate area for readback verification. At the same time, they are included in the network retraining sample set for subsequent low-frequency retraining. This example illustrates that the return air cooling does not automatically equal the hot spot receiving cooling. The controller determines whether the fan speed change forms effective hot spot cooling based on the inverse short circuit degree and hot spot response, and uses the actual hot spot response rather than the return air appearance for the candidate operating combination capability record. In one embodiment of the present invention, this implementation transforms the airflow path that cannot be directly observed within the door-mounted cabinet into an early return air response and a delayed hot spot response by using three spatial temperature locations and a single recoverable restricted airflow change. Then, a physically guided airflow response inversion network maps these two responses to the degree of return air short-circuiting and the effective cooling capacity of the hot spot. The compressor output remains stable within the identification window, so the network can process the changes in evaporator-side airflow and the changes in cooling capacity separately, reducing erroneous attribution. This processing method is different from existing schemes that rely on passive temperature reconstruction or temperature prediction for recooling. It actively excites and inverts the unmeasurable airflow short-circuiting, which is designed to address the difficulty of the lack of true values for return air short-circuiting. The key to truly solving the problem lies in the dual-measurement point flow splitting and inversion. A return air short circuit strengthens the return air response while weakening the hot spot response; since these two factors are inversely related, the network divides the increased airflow into short-circuit and effective components based on the degree of short circuit. It anchors the short-circuit component using the early return air response and the effective component using the delayed hot spot response, thus separating the two causes—severe short circuit and insufficient cooling capacity—that were previously indistinguishable on a single hot spot curve. Without artificial short-circuit labels, the inversion network completes self-supervised training based on the causal symmetry of hot spot reconstruction, return air reconstruction, acceleration and recovery, and load gating. This avoids misjudgments by a single threshold in cases of weak response, uncertain delay, and load coupling. Simultaneously, it explicitly transmits insufficient data to the candidate selection process through uncertainty output. The effective cooling capacity of the hot spot is formed by effective segments of the current cabinet, duct version, and sensor version. Candidate operating combinations must first pass future hot spot temperature, dew point, compressor start / stop, and hardware protection checks, and then be compared with recent measured power. This order ensures that energy-saving selection conforms to thermal safety boundaries. The execution results are returned as parameter records through response verification. Abnormal segments will not be included in the effective capability data or the network retraining sample set. Sensor anomalies, communication interruptions, storage verification failures, and power outage recovery all have clear degradation paths. Local pressure, water leakage, power supply, and temperature protection do not depend on learning parameters and network output. This forms a deployable, verifiable, and maintainable closed-loop control process. The method steps, system modules, and storage media programs use the same data fields and version relationships. The state dataset, airflow response feature set, thermal state parameter set, target control command, execution feedback, parameter update permission, and update context maintain the same meaning throughout the text. The degree of return air short-circuit, effective cooling capacity of hot spots, and cooling input power are not substituted for each other.
[0027] The embodiments of the present invention have been described above. However, the embodiments are not limited to the specific implementation methods described above. The specific implementation methods described above are merely illustrative and not restrictive. Those skilled in the art can make more equivalent embodiments under the guidance of the present embodiments, and all of them are within the protection scope of the present embodiments.
Claims
1. A method for intelligent energy-saving control of cabinet air conditioners, characterized in that, Includes the following steps: S1 collects supply air, return air, hot spot temperature, return air humidity, cooling input power, load proxy quantity, actuator and protection status according to the update context of the control cycle. After timing alignment, physical consistency verification and security permission determination, a status dataset is generated. S2, when the state dataset indicates that the compressor is stable and the hot spot has a safe reserve space, implement recoverable restricted air volume changes, extract the early return air response, the hot spot delayed response and the concurrent energy consumption to obtain the airflow response feature set; S3, use the airflow response inversion network to perform dual-point simultaneous inversion on the airflow response feature set, obtain the degree of return air short-circuit and the effective cooling capacity of hot spots, form a capability record of candidate operating combinations and output a set of thermal state parameters; S4. Based on the natural temperature rise rate of the thermal state parameter set and the effective cooling capacity of hot spots in the candidate operating combinations, extrapolate the future hot spot temperature from the current hot spot temperature, and combine the dew point boundary to screen the operating combinations that meet the thermal safety constraints, and generate the target control command. S5, after executing the target control command, drives the compressor and fan through the pre-interlock of power supply, pressure, water leakage and start-stop interval, compares the actual response of the hot spot with the expected response, and generates execution feedback and parameter update permission; S6 will execute feedback and parameter update permission return and verify remote settings, write control context and model parameters into non-volatile memory, and output the next control cycle update context according to the safe recovery sequence after power is restored.
2. The intelligent energy-saving control method for cabinet air conditioners according to claim 1, characterized in that, The safety permit determination is as follows: when there is a reserved space between the hot spot temperature and the control boundary, the power supply is stable, the pressure protection does not activate, the water leakage input is normal, the cabinet door is closed, the compressor is stable, the evaporator fan can report the actual speed, and the key status frames are continuously valid, the safety permit flag is set to valid; otherwise, the safety permit flag is set to invalid and a primary reason is attached. The state dataset also includes the natural temperature rise rate, derived from the slope of hot spot temperature versus time in a stable segment with no effective cooling, and a load continuity flag, derived from whether the load agent experiences a step within the observation window.
3. The intelligent energy-saving control method for cabinet air conditioners according to claim 1, characterized in that, The restricted airflow change is a closed loop consisting of the current speed, the target speed, and the recovery speed. The compressor maintains its original output state within the identification window, and the recovery section is used to verify whether the response direction corresponds to the airflow change. When extracting the early return air response and the delayed hot spot response, the return air residual sequence and the hot spot residual sequence are obtained by using the pre-excitation baseline trend. These are then used as inputs to the airflow response inversion network by resampling the evaporator fan speed, cooling input power, and phase markers into a multi-channel sequence of equal length. At the same time, the periods with load steps are marked as external disturbances by load proxy and load continuity flags.
4. The intelligent energy-saving control method for cabinet air conditioners according to claim 1, characterized in that, The airflow response inversion network consists of a temporal convolutional network backbone, a single-head temporal attention unit, and dual output heads. The dual-output head outputs the degree of return air short-circuit and its uncertainty, as well as the unit effective air volume hot spot cooling coefficient and its uncertainty. The envelope parameter head then regresses the start time and shape parameters of the hot spot delay response, the early return air response, and the effective component return response, respectively.
5. The intelligent energy-saving control method for cabinet air conditioners according to claim 4, characterized in that, The dual-measurement point joint inversion is as follows: the air volume increment of the limited air volume change is divided into short-circuit component and effective component according to the degree of return air short-circuit. The early response of return air is reconstructed by the short-circuit component and the delayed response of hot spot is reconstructed by the effective component. The two reconstruction relationships are combined to separate the degree of return air short-circuit and the effective cooling capacity of hot spot. The return air cooling coefficient of the short-circuit component is anchored by the measured temperature difference between the supply air temperature and the return air temperature. The airflow response inversion network is trained without labels using hotspot reconstruction loss, return air reconstruction loss, causal symmetry loss, and load gating loss. The effective cooling capacity of the hot spot is obtained by multiplying the hot spot cooling coefficient per unit effective air volume by the effective air volume. The capacity is recorded according to the candidate operating combination, combined with the compressor output status, evaporator fan speed, supply air temperature range, ambient temperature range and hot spot location.
6. The intelligent energy-saving control method for cabinet air conditioners according to claim 1, characterized in that, The future hotspot temperature is obtained by adding the product of the natural temperature rise rate and the control time domain to the current hotspot temperature, and then subtracting the product of the effective cooling capacity of the hotspot and the control time domain, and superimposing the prediction reserved space. The prediction reserved space is converted from the uncertainty output by the airflow response inversion network. The greater the uncertainty, the larger the prediction reserved space. During the screening process, operating combinations with expected supply air temperatures below the sum of the dew point boundary and the anti-condensation reserve space are excluded, while operating combinations that pass the hot spot temperature, dew point boundary, start-stop interval, and protection check are considered as safe candidates. When comparable input power records are available, the operating combination with the lowest input power and valid parameter confidence status is selected from the safety candidates. Operating combinations without current rack measured capacity records are not included in the energy reduction selection.
7. The intelligent energy-saving control method for cabinet air conditioners according to claim 1, characterized in that, The pre-execution interlock drives the compressor and fan when the power supply is stable, the pressure protection does not activate, the water leakage input is normal, and the compressor start-stop interval is met; otherwise, it maintains the current permitted state. When the status frame is continuously valid, the command and feedback are consistent, the compressor does not activate protection, the cabinet door is not opened, the load agent quantity does not change, the hot spot address does not change, the power data is complete, and the actual response direction is consistent with the expectation, the parameter update permission is set to valid and the capacity record is refreshed accordingly; when the parameter update permission is invalid, the capacity update of the corresponding operating combination is frozen, and the original capacity record is not used to reduce the cooling output.
8. The intelligent energy-saving control method for cabinet air conditioners according to claim 1, characterized in that, The non-volatile memory uses a current area and a candidate area for storage. The candidate area is first written with data content, version, length and check value, and then marked as the current valid version after readback and verification. After power is restored, the compressor is kept stopped and the power supply, sensors and fans are self-tested. When the equipment layout version, sensor calibration version and model version are consistent, the model parameters of the capacity record and airflow response inversion network are used for recovery. When the versions are inconsistent or the validity period expires, it is marked as pending verification and conservative temperature control is adopted.
9. A cabinet air conditioner energy-saving intelligent control system, characterized in that, include: The data acquisition module collects supply air, return air, hot spot temperature, return air humidity, cooling input power, load proxy quantity, actuator and protection status based on the update context of the previous control cycle. After timing alignment, physical consistency verification and security permission determination, it generates a status dataset. The response extraction module, when the state dataset indicates that the compressor is stable and the hot spot has a safe reserve space, implements recoverable restricted air volume changes, extracts the early return air response, the hot spot delayed response, and the concurrent energy consumption, and obtains the airflow response feature set. The capability assessment module uses an airflow response inversion network to perform dual-measurement point simultaneous inversion on the airflow response feature set, obtains the degree of return air short-circuiting and the effective cooling capacity of hot spots, forms a capability record of candidate operating combinations and outputs a set of thermal state parameters. The instruction generation module extrapolates the future hot spot temperature from the current hot spot temperature based on the natural temperature rise rate of the thermal state parameter set and the effective cooling capacity of the hot spot of the candidate operation combination. It then combines the dew point boundary to screen the operation combination that meets the thermal safety constraints and generates the target control instruction. The feedback update module, after executing the target control command via power supply, pressure, water leakage and start-stop intervals, drives the compressor and fan, compares the actual response of the hot spot with the expected response, and generates execution feedback and parameter update permission. The power failure recovery module will execute feedback and parameter update permission return and verify remote settings, write the control context and model parameters into non-volatile memory, and output the next control cycle update context according to the safe recovery sequence after power is restored.
10. A computer-readable storage medium having a program stored thereon, characterized in that, When the program is executed by the processor, it implements the steps of the intelligent energy-saving control method for cabinet air conditioning as described in any one of claims 1 to 8.
