Constant temperature adjusting method and system for air conditioner refrigeration house

Abstract

The invention discloses a constant-temperature adjusting method and system for an air conditioner refrigeration house, and belongs to the technical field of refrigeration and automatic control. The method comprises the steps that a multi-source sensing array is deployed in a cold storage door slot area, and pressure difference and heat flow changes are monitored in real time; identifying a door opening event based on the monitored data and determining a thermal intrusion primary path; calling a pre-trained thermal intrusion dynamic prediction model, and outputting a temperature disturbance prediction field at a future moment in the library; calculating a space directional cold air flow regulation and control instruction according to a prediction result, and driving a high-speed pulse cold air nozzle to form a hedging barrier at the initial stage of door opening; and after the door is closed, comparing predicted and actually measured temperature field residuals, and triggering online fine adjustment of the model to maintain long-term prediction precision. The corresponding system is composed of a door crack multi-source sensing module, an event recognition module, a dynamic prediction module, a directional regulation and control module, an online fine adjustment module and a temperature field reconstruction module. Millisecond-level sensing, prediction and source interception of door opening thermal disturbance are achieved, and the stability and control energy efficiency of the temperature of the refrigeration house are remarkably improved.

Description

Technical Field

[0001] This invention belongs to the field of refrigeration and automatic control technology, specifically relating to a method and system for constant temperature regulation of air-conditioned cold storage. Background Technology

[0002] With the widespread application of cold chain logistics in industries such as food and medicine, the constant control of the internal temperature of cold storage has become the key to ensuring the quality of stored goods. At present, most cold storage facilities adopt the traditional PID control strategy based on fixed temperature measurement points, and make lag adjustments based on the feedback of the internal temperature sensor. However, in actual operation, the frequent opening and closing of cold storage doors will introduce external hot air, causing rapid fluctuations in the internal temperature. Such instantaneous thermal disturbances are extremely detrimental to the preservation of temperature-sensitive goods.

[0003] Existing technologies suffer from several shortcomings: First, at the sensing level, systems typically rely solely on sparsely distributed temperature sensors within the storage facility, failing to capture the speed, direction, and spatial distribution characteristics of hot air intrusion in real time, resulting in significant sensing delays and information gaps. Second, at the execution level, conventional cooling compensation methods (such as starting and stopping compressors and activating fans) have slow response times and fixed airflow organization patterns, making it difficult to effectively intercept thermal disturbances before they spread to the main areas of the storage facility. Finally, at the control level, most solutions employ static or rule-based adjustment strategies, unable to dynamically and accurately match and distribute cooling capacity according to variables such as door opening status and external environment, easily leading to insufficient compensation, localized overcooling, or energy waste. Although some research has attempted to introduce intelligent control algorithms, they still suffer from limited prediction accuracy and have not fundamentally solved the problem of physical execution delays when dealing with complex and variable airflow disturbances.

[0004] To address the above problems, this invention proposes a solution. Summary of the Invention

[0005] This invention provides a method and system for constant temperature regulation in air-conditioned cold storage, aiming to solve the problem of drastic temperature fluctuations caused by delayed response, lack of heat intrusion prediction, and insufficient dynamic adaptability of actuators in existing cold storage constant temperature control systems during door opening and closing events.

[0006] In a first aspect, the present invention provides a method for constant temperature regulation in an air-conditioned cold storage, comprising:

[0007] S1, Deploy a multi-source sensing array in the cold storage door seam area to collect pressure difference change signals and heat flow vector distribution data in the door seam area in real time;

[0008] S2, based on the pressure difference change signal and heat flow vector distribution data, identifies the door opening event and determines the main path information of hot air intrusion;

[0009] S3, based on the door opening event and hot air intrusion information, calls the hot intrusion dynamic prediction model and outputs the predicted value of the temperature disturbance increment in the three-dimensional space of the warehouse within a set time in the future.

[0010] S4, based on the predicted value of temperature disturbance increment, generates a spatial directional cold airflow control command, and activates the high-speed pulse cold air nozzles at the corresponding positions to form a directional cold air barrier;

[0011] S5. After the door is closed, the dynamic prediction model for heat intrusion is fine-tuned online based on the residual between the actual temperature field and the predicted temperature field inside the warehouse.

[0012] Preferably, in step S1, the multi-source sensing array includes a differential pressure sensor array and an infrared heat flux density sensor arranged along the inner side of the door frame.

[0013] The differential pressure sensor array consists of multiple differential pressure sensing points arranged at preset intervals, used to collect differential pressure change signals;

[0014] Infrared heat flux density sensors are used to collect heat flux vector distribution data in the door gap area.

[0015] Preferably, step S2 specifically includes:

[0016] The door gap area is divided into multiple sub-regions, and the differential pressure signal within each sub-region is differentiated to obtain the differential pressure gradient.

[0017] When the differential pressure gradient continuously exceeds the first time threshold and is greater than the first preset threshold, a door opening event is determined to occur.

[0018] The heat flux vector distribution data is processed to identify the heat intrusion core area, and the main path direction and rate of heat intrusion are calculated by the displacement of the heat intrusion core area between consecutive frames.

[0019] Preferably, in step S3, the dynamic prediction model for thermal intrusion is a spatiotemporal coupled model constructed based on the Convolutional Long Short-Term Memory (ConvLSTM) network, and its inputs include the pressure difference time series, the heat flux density time series, external environmental parameters, and the initial temperature field inside the reservoir.

[0020] The model outputs the predicted temperature perturbation increment within the three-dimensional grid of the library over a future set time period.

[0021] Preferably, step S4 specifically includes:

[0022] Based on the predicted temperature disturbance increment and the main path of thermal intrusion, a weighted disturbance intensity distribution field is generated through spatial weighting calculation.

[0023] Based on the weighted perturbation intensity distribution field and the second preset threshold, the target area of ​​the cold airflow is determined;

[0024] Based on the geometric information of the target area and the direction of thermal intrusion, the combination of high-speed pulsed cooling nozzles to be activated, the deflection angle of each nozzle, and the jet flow rate are calculated.

[0025] Preferably, the high-speed pulsed cooling gas nozzle includes a piezoelectric ceramic drive valve, a Venturi ejector cavity, and an adjustable deflector vane connected in sequence.

[0026] Piezoelectric ceramic driven valves are used to achieve rapid opening and closing of air circuits;

[0027] The Venturi ejector cavity is used to eject air from the mixing chamber to form a mixed cold gas jet;

[0028] Adjustable deflector vanes are used to adjust the direction of the cold air jet.

[0029] Preferably, step S5 specifically includes:

[0030] After the door is closed, the actual three-dimensional temperature field inside the warehouse is reconstructed based on data from multiple temperature sensors arranged inside the warehouse.

[0031] The actual temperature field is compared with the predicted temperature field at the corresponding time output in step S3, and the spatial residual is calculated.

[0032] If the absolute value of the residual at any grid point exceeds the third preset threshold, the online model fine-tuning mechanism will be triggered.

[0033] Preferably, the online model fine-tuning mechanism employs a double buffer for data management:

[0034] The main buffer stores complete data packets for historical gate operation cycles in a circular fashion;

[0035] The auxiliary buffer is specifically used to store event packets marked as predicted failures;

[0036] During fine-tuning, the model parameters are iteratively updated using data in the auxiliary buffer.

[0037] Secondly, the present invention provides an air-conditioned cold storage constant temperature control system, comprising:

[0038] The multi-source sensing module for door gaps is used to perform millisecond-level detection in the door gap area of ​​cold storage, and to collect pressure difference change signals and heat flow vector distribution data;

[0039] The door opening event recognition module is used to identify door opening events based on sensing data and determine the main path information of hot air intrusion.

[0040] The thermal intrusion dynamic prediction module stores a thermal intrusion dynamic prediction model, which is used to predict the temperature disturbance increment within the library based on event and path information.

[0041] The spatial directional cold airflow control module is used to generate control commands based on predicted values ​​and drive high-speed pulse cold air nozzles to perform directional cold air compensation.

[0042] The temperature field reconstruction module inside the warehouse is used to reconstruct the actual temperature field based on the temperature sensor data inside the warehouse.

[0043] The online model fine-tuning module is used to fine-tune the dynamic prediction model of thermal intrusion based on the residual between the predicted temperature field and the actual temperature field.

[0044] Preferably, the spatial directional cold airflow control module includes multiple high-speed pulsed cold air nozzles, each nozzle comprising a piezoelectric ceramic drive valve, a Venturi ejector cavity, and an adjustable deflector.

[0045] In summary, this application includes at least one of the following beneficial technical effects:

[0046] 1. This invention achieves millisecond-level thermal intrusion event identification and path tracking by deploying a high-density multi-source sensing array at the door seam, and combined with feedforward-feedback control, drives high-speed cold air nozzles to perform directional interception in the early stage of hot air diffusion, effectively blocking thermal disturbances from the source and significantly improving the temperature stability of the core area inside the warehouse.

[0047] 2. This invention uses a spatiotemporal coupling prediction model to make high-precision dynamic predictions of the development of thermal intrusion, and calculates the optimal space cooling compensation scheme accordingly, so as to realize the on-demand, fixed-point and directional delivery of cooling capacity, avoiding the response lag and coarse compensation of traditional methods, and reducing energy waste while controlling temperature precisely.

[0048] 3. This invention automatically triggers the online fine-tuning mechanism of the prediction model through closed-loop monitoring and residual analysis. It uses recent prediction failure case data to quickly correct model parameters, enabling the system to continuously adapt to changes in the external environment and internal layout, and ensuring control accuracy and stability during long-term operation. Attached Figure Description

[0049] Figure 1 This is a flowchart illustrating a method for constant temperature control in an air-conditioned cold storage facility according to the present invention.

[0050] Figure 2 This is a schematic diagram of the process of a constant temperature control system for an air-conditioned cold storage according to the present invention. Detailed Implementation

[0051] To further illustrate the technical means and effects adopted by the present invention to achieve the intended purpose, the following detailed description of specific embodiments based on the present invention is provided in conjunction with the accompanying drawings and preferred embodiments.

[0052] This invention provides a method and system for constant temperature regulation of air-conditioned cold storage. By deploying a high-density multi-source sensing array in the door gap area of ​​the cold storage, millisecond-level identification and spatial distribution modeling of hot air intrusion events are achieved. Based on this, a feedforward-feedback fusion control mechanism is constructed to drive high-speed pulsed cold air nozzles to implement directional cold airflow compensation at the beginning stage of door opening, thereby completing the source blocking before the thermal disturbance propagates to the main space inside the cold storage.

[0053] Example 1

[0054] A method for constant temperature control in air-conditioned cold storage, referring to Figure 1 As shown, it includes the following steps:

[0055] S1 integrates a multi-point array micro differential pressure sensor and an infrared heat flux density sensor around the cold storage door frame.

[0056] S101, Select and arrange sensors.

[0057] A silicon-based MEMS differential pressure sensor was selected as the micro differential pressure sensing unit. Its range is -5Pa to +5Pa and its resolution is not less than 0.01Pa. During installation, 12 sensing points were evenly arranged at 20cm intervals along the door height direction on the inner side of the cold storage door frame, about 5mm away from the edge of the door seam, to form a high-density differential pressure sensing array in the vertical direction.

[0058] Meanwhile, the infrared heat flux density sensor uses a thermopile-type infrared array with a field of view of 60° and a response wavelength range covering 8μm to 14μm. The infrared heat flux density sensor is installed on the inner side of the top of the door frame, aligned with the center line of the door gap, to synchronously monitor the heat flux distribution in the door gap area.

[0059] S102, Configure data acquisition and transmission links.

[0060] The differential pressure sensor operates at a sampling frequency of 200Hz to capture instantaneous pressure difference changes at various points in the door gap in real time. The infrared heat flux density sensor is set to a sampling frequency of 100Hz to simultaneously acquire heat flux vector distribution data in the door gap area, including heat flux direction, intensity, and spatial gradient.

[0061] All sensor data is aggregated to the central processing unit via an industrial Ethernet bus, ensuring that the data transmission delay from the sensing end to the processing end does not exceed 5 milliseconds, guaranteeing the real-time performance and timing consistency of the sensed data, and providing a foundation for subsequent millisecond-level control.

[0062] S103, Overview of System Initialization and Sensing Functions.

[0063] The differential pressure sensor array is responsible for capturing the slight pressure difference change caused by the influx of external hot air when the door is opened. At the same time, the infrared heat flux density sensor monitors the heat flow status in the door gap area in real time.

[0064] After the system is powered on, all sensors automatically perform self-test and calibration processes to ensure that each sensing point works normally and the data output is stable. At this point, the system enters the real-time monitoring ready state.

[0065] Through the above steps, the arrangement and configuration of the multi-source sensing array for cold storage door gaps are completed. This high-density, high-response sensor combination can monitor the pressure difference and heat flow dynamics of the airflow in the door gaps in real time at the millisecond level. Furthermore, this sensing front end provides accurate and timely input data for the precise identification of subsequent door opening events, the determination of heat intrusion paths, and dynamic prediction models.

[0066] S2, based on the collected differential pressure change signal and heat flow vector distribution data, determines the current movement state of the door through a door opening status recognition algorithm.

[0067] S201, divide the monitoring area and process the differential pressure signal.

[0068] The door gap area is divided into three sub-regions along the height direction: upper, middle, and lower. Each sub-region corresponds to four consecutive differential pressure sensing points. Then, the differential pressure signal in each sub-region is processed by sliding window differentiation with a window length of 10 milliseconds to calculate the real-time absolute value of the differential pressure gradient.

[0069] S202, recognizes door opening event.

[0070] The system continuously monitors the absolute value of the pressure gradient in the upper sub-region. When the absolute value of the pressure gradient is detected to be greater than the threshold of 0.8 Pa / millisecond for more than 20 milliseconds, it is determined that the door opening event has been initiated. This judgment is mainly based on the physical characteristic that hot air will preferentially enter from the upper part of the door gap due to the buoyancy effect, thereby improving the accuracy of event recognition and anti-interference ability.

[0071] S203, analyze heat flow data to determine the main path of heat intrusion.

[0072] While recognizing the door opening event, the system performs parallel analysis based on the heat flow vector data output in real time by the infrared heat flow density sensor.

[0073] The specific method is as follows: the heat flux density matrix acquired by the infrared array is scanned pixel by pixel. For each pixel, its value is compared with other pixels in a preset neighborhood (e.g., a 3×3 window). If the pixel value is the maximum value in this window and its value exceeds a preset heat flux intensity threshold, then the pixel is marked as a candidate high-temperature point. The threshold can be dynamically set according to the external ambient temperature, for example, set to a certain proportion (e.g., 1.2 times) of the theoretical heat flux intensity value corresponding to the current ambient temperature.

[0074] After identifying all candidate high-temperature points, these points are used as initial seeds to execute a region growing algorithm. The growing criterion is to merge any pixel adjacent to the current region whose heat flux intensity also exceeds the aforementioned threshold into that region. Alternatively, a connected component analysis method can be used to cluster all pixels with heat flux intensities exceeding the threshold and adjacent to each other (using four-connectivity or eight-connectivity criteria) into different connected regions.

[0075] Finally, in all generated candidate regions, calculate the total pixel area (i.e., physical area) and the average heat flux intensity within each region. Select the connected region with the largest area or the highest average heat flux intensity and mark it as the heat intrusion core region at the current moment.

[0076] S204, calculate the intrusion path and rate.

[0077] First, the spatial coordinates of the thermal intrusion core area of ​​the current frame are calculated using the centroid algorithm;

[0078] Next, the centroid coordinates of the core region of the previous frame are obtained, and the known time interval Δt between the two frames is determined (e.g., corresponding to the sensor sampling frequency).

[0079] Then, the displacement vector of the centroid between the two frames is calculated, and the direction of this displacement vector is output as the direction vector of the main path of thermal intrusion.

[0080] Meanwhile, the magnitude of the displacement vector is divided by the time interval Δt, and the result is output as a scalar value of the thermal intrusion rate.

[0081] Thus, step S2 completes the transformation from raw signal to decision information. Its output "gate opening event" signal, thermal intrusion "direction vector" and "velocity scalar" together constitute the key input parameters necessary for subsequent steps to perform high-precision prediction of thermal disturbance and characterize the dynamics of the thermal intrusion source, realizing the transformation from passive temperature feedback to active airflow feedforward.

[0082] S3, based on the door opening event start signal and the main path information of hot air intrusion, calls the pre-stored dynamic prediction model of hot intrusion and outputs the predicted value of temperature disturbance increment of each grid point in the three-dimensional space of the library within the next five seconds.

[0083] S301, Prepare model input data.

[0084] Before calling the pre-stored dynamic prediction model for thermal intrusion, four types of time-series data need to be prepared as input:

[0085] Differential pressure time sequence: Differential pressure data collected from 12 differential pressure sensing points at the door gap in the past 2 seconds were obtained. The differential pressure data was sampled at a frequency of 200Hz to form a sequence with a length of 400 time steps.

[0086] Heat flux density time series: The heat flux density data recorded by the infrared heat flux density sensor in the past 2 seconds are acquired synchronously, with a sampling rate of 100Hz, forming a sequence with a length of 200 time steps.

[0087] External environmental parameters: Real-time measurements from external environmental temperature and humidity sensors, updated once per second.

[0088] Initial state within the reservoir: The current three-dimensional temperature field distribution within the reservoir is obtained in real time from the temperature field reconstruction module within the reservoir. This is used as the initial boundary condition for prediction. The spatial resolution of this temperature field data is 0.5 meters, and it needs to cover the entire reservoir space.

[0089] S302, Perform model inference and obtain prediction results.

[0090] The four types of data prepared in step S301 are input into the pre-stored dynamic prediction model for thermal intrusion for inference. This model is a spatiotemporal coupled model built on a convolutional long short-term memory network (ConvLSTM). Its core function is to learn the complex nonlinear mapping relationship between the airflow disturbance at the door gap and the dynamic evolution of the temperature field inside the warehouse.

[0091] After the model inference is completed, it directly outputs the predicted field of temperature perturbation increment in the three-dimensional space within the library for the next 5 seconds. This predicted field has a clear spatiotemporal resolution.

[0092] On the timeline, prediction results are output at 100-millisecond intervals, for a total of 50 consecutive time steps;

[0093] Spatially, the database is divided into cubic grids with sides of 0.5 meters, and the predicted value corresponds to the center point of each grid.

[0094] S303, Description of model training and deployment.

[0095] To ensure prediction accuracy, the dynamic prediction model for thermal intrusion needs to undergo sufficient offline training:

[0096] S303a, Training Data Acquisition and Preparation.

[0097] The training data was collected in a physical model experimental environment scaled to a real cold storage facility. The dataset was built by performing more than 1,000 perturbation experiments, which systematically covered key operating variables, including different combinations of door opening speed (0.5 to 2.5 m / s), ambient temperature (0 to 40 degrees Celsius), and relative humidity (30% to 90%).

[0098] In each experiment, the system needs to record two types of data synchronously and frequently:

[0099] First, the raw data from the door gap sensor, including differential pressure and heat flux density sequences;

[0100] Second, the true value of the temperature field inside the warehouse, which serves as a monitoring signal, is acquired by a high-precision three-dimensional laser scanning thermometer with a time resolution of 100 milliseconds.

[0101] As a result, a large number of matching data pairs were constructed, from multi-source sensor time series to three-dimensional temperature field change series.

[0102] S303b, Model Training Process.

[0103] Using the prepared dataset, supervised training is performed on the spatiotemporal coupling model based on ConvLSTM. During training, mean squared error (MSE) is used as the loss function, and the Adam optimizer is used for iterative optimization.

[0104] Typically, the dataset is divided into a training set and a validation set in a certain ratio (e.g., 8:2). The training process continues until the model's prediction accuracy on the independent validation set tends to stabilize (i.e., the preset early stopping condition is met), thereby obtaining a final model with good generalization ability.

[0105] S303c, ensuring model deployment and real-time performance.

[0106] Once the trained and validated model is completed, it is deployed on the system's central processing unit, which uses a high-performance embedded multi-core processor and is equipped with a dedicated neural network acceleration coprocessor that supports INT8 quantization inference technology to greatly improve computational efficiency.

[0107] Based on this hardware, the entire process from data input to completing model inference and outputting prediction results is controlled within 80 milliseconds, thereby ensuring that the system can meet the strict real-time response requirements after the door opening event is triggered.

[0108] S4. Based on the predicted temperature disturbance increment, a spatially oriented cold airflow control command is generated to activate the high-speed pulse cold air nozzle at the corresponding position.

[0109] S401 performs spatial weighted focusing on the predicted temperature field.

[0110] Based on the predicted increment of the three-dimensional spatial temperature disturbance within the library in the next 5 seconds obtained in step S3, a spatial weighted calculation is performed to focus on the region most significantly affected by thermal intrusion. The specific operation is as follows:

[0111] First, the reference benchmark for spatial weighted calculation is determined to be the main path axis of thermal intrusion. This axis is defined as a spatial straight line that passes through the centroid coordinate point of the thermal intrusion core area determined by step S203, and its direction is consistent with the thermal intrusion direction vector output by step S203.

[0112] Secondly, a weighted calculation is performed point-by-point, and the library space has been divided into a cubic grid with a side length of 0.5 meters. For each grid point (x, y, z):

[0113] Obtain the predicted value of the temperature perturbation increment ΔT(x,y,z) at the current prediction time for this point (in °C);

[0114] Calculate the vertical distance from the grid point to the main path axis of the aforementioned heat intrusion, i.e., the Euclidean distance d(x,y,z) (in meters).

[0115] The weight value w(x,y,z) corresponding to this distance is calculated using the Gaussian decay function. The specific formula is as follows:

[0116] ;

[0117] The attenuation coefficient of 0.3 meters is based on the fitting results of experimental data on typical cold storage dimensions and thermal diffusion characteristics. It is used to control the rate at which the weight decays with distance and can be calibrated and adjusted according to the actual cold storage dimensions and airflow organization characteristics.

[0118] Calculate the weighted perturbation intensity at this grid point:

[0119] .

[0120] After traversing all grid points, a weighted perturbation intensity distribution field with the same spatial resolution as the predicted temperature field is generated.

[0121] The significance of the weighting process is that by applying spatial attenuation weighting to the original predicted values, the influence of grid points that are far from the main path of thermal intrusion (low weight) but have large predicted temperature changes is reduced, while the disturbance signal of grid points located near the main path (high weight) is highlighted.

[0122] This allows for more precise identification of the core area that is about to heat up due to the direct intrusion of hot air, laying the foundation for accurately delineating the target area requiring cooling intervention in the next step, and also avoiding over-cooling response in non-critical areas.

[0123] S402, determine the target area for the cold airflow.

[0124] The goal is to precisely delineate the core area requiring immediate cooling compensation within the entire storage space. The operation is based on the weighted perturbation intensity distribution field S(x,y,z) generated in step S401.

[0125] First, a threshold of 0.5℃ was set as a key control parameter, determined based on calibration results from extensive prior simulations and field experiments. The experiments primarily examined the relationship between the system's effectiveness in suppressing temperature peaks caused by door opening and the additional cooling energy consumption under different thresholds. This threshold roughly corresponds to the inflection point where the suppression effect plateaus while energy consumption begins to increase significantly, ensuring temperature control accuracy in the core area while avoiding unnecessary cooling waste. In the actual system, this threshold serves as a configurable parameter, allowing operators to make minor adjustments (e.g., ±0.1℃) based on the temperature-sensitive characteristics of the stored items to achieve the optimal balance between temperature control accuracy and energy consumption.

[0126] Next, regional screening and spatial clustering are performed. The system traverses all grid points and marks the points with S(x,y,z)>0.5℃ as hotspots. These hotspots may be spatially discrete and need to be aggregated to form a continuous intervention area.

[0127] The rule for determining spatial continuity during clustering is that two grid points are considered connected only when they are directly adjacent in the positive and negative directions of the three coordinate axes X, Y, and Z (i.e., the two grid cells share a complete surface). All hot spots connected by this rule are aggregated into an independent cold airflow target area.

[0128] Through the above process, the system can extract one or more spatially coherent regions with thermal disturbance intensity exceeding a set threshold from the complex predicted temperature field. These targeted regions provide direct and quantitative spatial basis for the precise spatial matching of nozzles and the distribution of cold air flow in subsequent steps.

[0129] S403, calculates nozzle combination and injection parameters.

[0130] Based on the geometric center coordinates of the target area and its corresponding thermal intrusion direction vector, combined with nozzle layout information, the optimal nozzle activation combination, injection angle, and flow parameters are calculated. Specifically, this includes:

[0131] S403a, based on the known fixed installation coordinates of the 16 nozzles and the nominal spray cone angle of each nozzle (e.g., 60°), calculate the potential coverage area of ​​each nozzle. For a given target area, the system evaluates all nozzles and includes nozzles that simultaneously meet the following two conditions into the candidate activation set:

[0132] The spatial coverage condition is determined by whether the direction of the line connecting the nozzle position to the geometric center of the target area is within the deflection angle range (-45° to +45°) of the adjustable deflection guide vane of the nozzle. If so, the nozzle is considered to have the potential to cover the geometric center of the area in that direction.

[0133] For directional countermeasures, based on the spatial coverage condition, the angle θ between the theoretically optimal jet direction of the nozzle (i.e., the direction vector from the nozzle to the geometric center of the target area) and the thermal intrusion direction vector is calculated. Nozzles with an absolute value of θ close to 180° are preferred to achieve frontal interception of the thermal intrusion airflow. In practice, a threshold can be set (e.g., |θ|>150°); meeting this threshold is considered to satisfy the "countermeasure" requirement.

[0134] S403b calculates a unique deflection angle for each activated nozzle.

[0135] The deflection angle is set to allow the cold air jet to be more precisely aimed at the front of the intruding hot airflow, thus achieving effective interception. The system calculates the deflection angle individually for each activated nozzle, primarily based on the hot intrusion direction vector.

[0136] Specifically, the thermal intrusion direction vector is decomposed, and its vertical component (usually corresponding to the Z-axis direction in space, i.e., the height direction) determines the basic deflection direction and angle of the nozzle guide vane in the vertical plane. For example, if the vertical component of the thermal intrusion direction vector is positive (indicating that the hot airflow has an upward tendency), then the guide vane of the corresponding nozzle is controlled to deflect upward by a corresponding angle to block the rising hot airflow. The influence of the horizontal component is mainly addressed by the coordinated response of the nozzle combinations at different horizontal positions selected in step S403a.

[0137] S403c determines the injection flow rate for each enabled nozzle.

[0138] First, the required total cooling air compensation flow rate Q_total is calculated. This total flow rate is obtained by multiplying the average weighted perturbation intensity within the target area (i.e., the average value of S(x,y,z) obtained in step S401 over all grid points within the target area) by a preset scaling factor (e.g., 0.1 m³ / (s·℃)). The calculation formula is as follows:

[0139] ;

[0140] in, This represents the average value of the weighted perturbation intensity S(x,y,z) of all grid points within the target region Ω; k is a preset scaling factor (e.g., 0.1 m³ / (s·℃)).

[0141] Then, the total flow rate Q_total is allocated to each nozzle in the candidate activation set, following a weighted allocation method to achieve spatial optimization of cooling resources. Each candidate nozzle i is assigned a weight w_i, which comprehensively considers its distance d_i from the geometric center of the target area and the degree of antagonistic matching between its injection direction and the heat intrusion direction (reflected by the cosine of the included angle θ, |cosθ|; a larger value indicates a higher degree of matching). A feasible weight calculation method is as follows:

[0142] .

[0143] Finally, the flow rate Q_i allocated to each nozzle i is calculated according to the following formula:

[0144] ;

[0145] Here, Σw_i is the sum of the weights of all candidate nozzles. This allocation method ensures that nozzles closer to the core of the target area and whose spray direction is closer to the direction of thermal intrusion receive a larger flow of cold air, thereby achieving precise and efficient cold air delivery.

[0146] S404 is equipped with a high-speed pulse cooling nozzle actuator.

[0147] To achieve sub-second response and high-precision spatial directional cold gas injection, this system uses a specially designed high-speed pulse cold gas nozzle. This nozzle is an integrated actuator, mainly composed of three parts connected in sequence: a piezoelectric ceramic drive valve, a Venturi ejector chamber, and an adjustable deflector.

[0148] As the core of the gas circuit for rapid switching, the piezoelectric ceramic driven valve is directly connected to the refrigerant supply line. Based on the inverse piezoelectric effect of piezoelectric ceramics, the valve achieves millisecond-level opening and closing, with an opening response time of no more than 30 milliseconds, ensuring that the system can start airflow in a very short time after the door is opened. The maximum design flow rate of the valve body is 0.5 cubic meters per second, providing sufficient instantaneous cooling compensation capability for a single point.

[0149] The Venturi ejector chamber is connected downstream of the drive valve. Its structure includes a tapering inlet section, a narrow throat, and a widening outlet section. The throat, with a diameter of 8 mm, is a key design parameter. When the high-pressure refrigerant flows at high speed through the throat via the drive valve, a negative pressure zone is formed around the throat. This negative pressure draws in and entrains the low-temperature air in the chamber surrounding the ejector nozzle, causing it to mix thoroughly with the mainstream refrigerant from the drive valve within the diffuser section of the chamber. This not only increases the overall flow rate of the final ejected airflow but also optimizes the spatial distribution efficiency of the cooling capacity by mixing in the air within the chamber, forming a mixed cold air jet with a larger flow rate and a more suitable temperature distribution.

[0150] The adjustable deflector is installed at the outlet end of the Venturi ejector cavity and is a key mechanism for achieving spatial orientation of the cold gas jet. The deflector is driven by a micro stepper motor and can deflect precisely within the range of -45° to +45° with an angle adjustment resolution of 1°. The system controls the stepper motor in real time according to the command calculated in step S403 and dynamically adjusts the deflection angle of the deflector, thereby precisely changing the final ejection direction of the cold gas jet and ensuring that it can accurately aim at and intercept the predicted thermal intrusion front.

[0151] In summary, this high-speed pulsed cold air nozzle achieves millisecond-level rapid on / off of the gas path (pulsation) through a piezoelectric ceramic driven valve, enhances the jet flow rate and optimizes the mixing effect (high speed and mixing) through a Venturi ejector cavity, and completes fine spatial adjustment of the ejection direction (directivity) through an adjustable deflector.

[0152] S405 executes injection commands and ensures timing synchronization.

[0153] The system captures the door opening event start signal in real time through a hardware interrupt mechanism, and completes the following actions within 50ms thereafter:

[0154] When the door opening event start signal in step S202 is sent to the system central processing unit through the hardware interrupt pin, the unit responds immediately. First, it encapsulates the injection parameters (including nozzle ID, deflection angle, injection flow rate, etc.) that have been calculated in step S403 into a drive command package. Within a very short time after the door opens (e.g., within 10 milliseconds), these commands are quickly sent to the local drive units of each selected high-speed pulse cooling nozzle through the control bus.

[0155] Upon receiving the command, the drive unit immediately takes action. For each nozzle, its piezoelectric ceramic drive valve opens to the specified opening degree within 30 milliseconds according to the flow command. The refrigerant from the evaporator outlet then enters the nozzle. The refrigerant is accelerated as it flows through the Venturi ejector chamber and mixes with the ejected air in the chamber to form a high-speed cold air jet. At the same time, the adjustable deflector at the end of the nozzle is also rotated to the specified angle calculated in step S403b within a few milliseconds under the drive of the micro stepper motor.

[0156] Within a 50-millisecond time window triggered by the door opening signal, a dynamic cold air barrier consisting of multiple nozzles working together is initially formed at a designated spatial location inside the door. The spatial layout (which nozzles spray where), direction (guide vane deflection angle), and intensity (flow rate of each nozzle) of this barrier are strictly set according to the instructions of S403, thereby ensuring that it spatially counteracts the predicted main path of heat intrusion and achieves on-demand compensation in terms of cooling capacity, ultimately achieving the goal of suppressing thermal disturbance from the source.

[0157] S5: After the door is closed, continuously monitor the feedback data from multiple temperature sensors inside the warehouse and compare it with the prediction results. If the absolute value of the residual exceeds the preset threshold of 0.3℃, the online fine-tuning mechanism of the model is triggered.

[0158] S501 collects real-time temperature data and reconstructs the temperature field inside the warehouse.

[0159] Upon detecting a door closure signal, the system immediately initiates a temperature field reconstruction process. First, it synchronously reads the measurements from 40 high-precision platinum resistance temperature sensors located at different positions within the warehouse at a frequency of once per second. These sensors are spatially non-uniformly distributed, primarily covering shelf gaps, mainstream airflow areas, and dead zones.

[0160] Subsequently, the temperature field reconstruction module within the reservoir, based on real-time data from these sparse measuring points, generates a three-dimensional temperature field distribution with a spatial resolution of 0.5 meters, covering the entire reservoir volume, through a data interpolation algorithm that incorporates simplified physical constraints. Specifically:

[0161] Interpolation core: The inverse distance weighted interpolation method is used as the basis. The temperature T_j at any unknown grid point j is calculated from the temperatures T_i of all sensor measurement points i within a certain radius around it. The calculation formula is:

[0162] ;

[0163] Among them, weight d_ij is the spatial distance from grid point j to sensor i.

[0164] Physical constraints are introduced: To better conform to the physical laws of airflow in cold storage, a simplified computational fluid dynamics reference field based on steady-state flow field is introduced before the pure mathematical interpolation mentioned above. This reference field is a baseline model of airflow velocity and temperature distribution in the cold storage under typical refrigeration unit operating conditions, obtained in advance through CFD simulation. During real-time reconstruction, the current sparse measurement point data is assimilated with this reference field. The spatial gradient prior information provided by the reference field is used to correct the results of inverse distance weighted interpolation, especially in areas with sparse sensor coverage, so that the reconstructed temperature field is more consistent with the airflow trend and the distortion that pure interpolation may cause is reduced.

[0165] Output and Alignment: The reconstruction process is completed within 2 seconds after the door closes. The output three-dimensional temperature field data has a timestamp aligned with the door closing time. The spatial grid coordinate system is completely consistent with the grid used to predict the temperature field in step S3, ensuring that the two can be directly compared point by point.

[0166] After the gate operation disturbance has basically subsided, this step quickly acquires a snapshot that reflects the actual spatial distribution of temperature inside the storage chamber, which serves as a reliable benchmark for evaluating the accuracy of the previous predictions.

[0167] S502 calculates the residual between the predicted temperature field and the actual temperature field.

[0168] The purpose of this step is to quantitatively assess the accuracy of the previous predictions. In practice, the system needs to solve two key issues: time correspondence and spatial alignment.

[0169] Time Correspondence: After the door opening event is triggered, the prediction model will continuously output the temperature field for multiple future time steps. For effective comparison, the system defaults to selecting the predicted temperature field on the time axis that is closest to the door closing time as the comparison benchmark. For example, if the model outputs time steps t1, t2, ... (interval of 100 milliseconds), and the door closing time is T_close, the system will automatically calculate and select the prediction field corresponding to the time step t_k that satisfies the minimum value of |T_close - t_k|, ensuring that the comparison is between theoretical predictions and actual observations under the same spatiotemporal conditions.

[0170] Spatial Alignment: Thanks to step S501, which ensured that the reconstructed actual temperature field and the predicted temperature field used the exact same spatial grid division scheme (0.5-meter cube grid) and a unified global coordinate system origin, the grid points of the two are strictly one-to-one corresponding in spatial location. When performing residual calculation, the system directly reads the predicted temperature value T_pred(x,y,z) of the grid point in the predicted field and the reconstructed temperature value T_real(x,y,z) at the same location in the actual field based on the same three-dimensional index (x,y,z).

[0171] Residual calculation: For each grid point in the database, the difference is calculated using the formula residual(x,y,z) = T_real(x,y,z) - T_pred(x,y,z). After traversing all grid points, a residual field of the same size as the temperature field is obtained, which intuitively reflects the spatial distribution of the prediction error.

[0172] S503 determines whether model fine-tuning should be triggered.

[0173] After the system completes the residual calculation, it will scan the entire residual field, check the absolute value of the residuals of all grid points, and set a key criterion: as long as the absolute value of the residual of any grid point exceeds the preset threshold of 0.3℃, the system will determine that there is a significant deviation in the prediction, and record it as a prediction failure event.

[0174] Once a prediction is determined to be ineffective, the system will immediately send a trigger signal to the online model fine-tuning module to initiate the subsequent fine-tuning process (S504-S506).

[0175] The 0.3℃ threshold is a core parameter. Its value is mainly determined by the stringent requirements of high-standard scenarios such as GSP pharmaceutical warehousing for temperature fluctuations (usually requiring better than ±0.5℃), while also taking into account the model's ability to predict common disturbance patterns. The aim is to capture those critical prediction errors that may affect storage quality.

[0176] Of course, this threshold is not absolutely fixed. Operators can fine-tune it within ±0.1℃ based on the actual temperature sensitivity of the stored items. For example, when storing vaccines, which are extremely sensitive to temperature, the threshold can be tightened to 0.2℃; while for food ingredients with slightly higher tolerance, it can be relaxed to 0.4℃, thus achieving the best balance between control precision and system sensitivity.

[0177] S504, prepare the data needed for fine-tuning (double buffer mechanism).

[0178] To efficiently organize the data used for online model fine-tuning, the system establishes a dual-buffer data management structure, with each buffer having its own specific function:

[0179] The main buffer acts as a circular queue, continuously storing relevant data from the last 100 complete door opening and closing operation cycles. After each door operation event, the system packages the multi-source sensing raw sequence (differential pressure, heat flux density) associated with the event, the synchronously recorded external environmental parameters (temperature and humidity), the true value of the actual temperature field reconstructed in step S501, and the door opening event and heat intrusion path information output in step S2 into a single data packet and stores it in the main buffer. When the buffer is full (reaching 100 data packets), the newest data packet overwrites the oldest one, achieving rolling data updates.

[0180] Auxiliary buffer: This buffer is specifically used to store event data packets marked as prediction failures, filtered from the main buffer. Whenever a prediction failure event is determined in step S503, the system not only triggers fine-tuning but also immediately stores the complete data packet corresponding to the current event (which can be obtained from the latest record in the main buffer) along with the residual field information obtained in this calculation into the auxiliary buffer. The auxiliary buffer has a capacity limit (e.g., 20 data packets). When it is full, new failure event data will replace the oldest stored data, ensuring that it always focuses on recent typical operating conditions where the model prediction is inaccurate.

[0181] Once the online model fine-tuning process is triggered, the fine-tuning module extracts all stored data packets from the auxiliary buffer as the training dataset for this fine-tuning. The advantage of this design is that:

[0182] The main buffer retains comprehensive historical data for possible long-term analysis or model retraining; while the auxiliary buffer automatically filters and aggregates problematic data, enabling online fine-tuning to quickly and specifically correct parameters for particular conditions where the model has recently performed poorly (such as a specific door opening speed combined with high external humidity), avoiding repeated screening from massive amounts of normal data and improving the efficiency and targeting of fine-tuning.

[0183] S505 performs online fine-tuning of model parameters.

[0184] Once the fine-tuning process is triggered, the system invokes the online model fine-tuning module to optimize the parameters of the dynamic prediction model for heat intrusion using the dataset extracted from the auxiliary buffer. The specific execution process is as follows:

[0185] The goal of each fine-tuning is to narrow the gap between model predictions and actual observations. Therefore, the mean squared error (MSE) is continued to be used as the loss function to calculate the difference between the predicted temperature field and the actual temperature field.

[0186] A stochastic gradient descent (SGD) algorithm with a batch size of 8 and a fixed learning rate of 0.001 was used to iteratively update all weight parameters of the dynamic prediction model for thermal intrusion. This local update refers to making small adjustments to the parameters while keeping the overall model architecture unchanged, in order to adapt to newly observed operating conditions.

[0187] To ensure the real-time nature of online fine-tuning, the maximum number of iterations for the algorithm is set to 10 after each fine-tuning trigger. Typically, when the amount of data in the auxiliary buffer is limited (e.g., containing only a few recent failure events), this number of iterations is sufficient for the loss function to converge to a better local solution.

[0188] The entire fine-tuning process, including data loading, forward propagation, loss calculation, backpropagation, and parameter updates, needs to be completed within 500 milliseconds. This time constraint ensures that the fine-tuning operation will not consume too many system resources, thereby ensuring that the system still has the ability to predict and respond to the next sudden door opening event in milliseconds.

[0189] Through the above steps, the system can quickly and specifically calibrate the neural network model in a very short time using recently collected typical case data of inaccurate predictions, thereby improving its adaptability to the current operating environment and its prediction accuracy.

[0190] S506, Update model weights and apply the changes.

[0191] After fine-tuning, the new model weights immediately override the original parameters and are directly used in subsequent door opening event predictions. The system operates without interruption, achieving a seamless transition.

[0192] After the fine-tuning process (S505) is completed, the system safely applies the new parameters and ensures the continuity and consistency of the forecasting service by following these steps:

[0193] The new weight parameters obtained from the fine-tuning are first written to a temporary, independent memory region or a copy of the model file, which is physically isolated from the storage location used by the currently running prediction model.

[0194] After confirming that the new weight data has been completely written, the system performs an atomic pointer update operation. Specifically, the system maintains a pointer (or handle) to the weights of the currently effective model. During the update, an uninterruptible instruction instantly switches this pointer from pointing to the address of the old weights to pointing to the address of the new weights. This operation ensures that at any given time, all concurrent prediction requests read a complete and consistent set of weights, avoiding inconsistencies that may arise from step-by-step overwriting.

[0195] After the pointer switch is completed, all subsequent door opening events will automatically and seamlessly use the new weight parameters for inference in their corresponding prediction tasks (S3). At this point, the model update takes effect at the system level.

[0196] Once the system confirms that the old weight model is no longer referenced by any prediction task (e.g., through reference counting or a delay mechanism), it safely releases the memory resources it occupies.

[0197] Through the above mechanism, the system realizes hot updating of model weights. The entire update action is completed within milliseconds. For the continuously running control system, the prediction service is uninterrupted, realizing a smooth and seamless transition from the old model to the new model, and ensuring the continuity and reliability of the adaptive control process.

[0198] Example 2

[0199] This embodiment provides an air-conditioned cold storage constant temperature control system for implementing the constant temperature control method described in Embodiment 1, referring to... Figure 2 As shown, the system includes a multi-source sensing module for door gaps, a door opening event recognition module, a dynamic prediction module for thermal intrusion, a spatially directional cold airflow control module, an online model fine-tuning module, and a temperature field reconstruction module inside the warehouse.

[0200] The multi-source sensing module for door gaps is responsible for millisecond-level detection in the initial stages of heat intrusion. Its hardware includes:

[0201] Micro differential pressure sensor array: Twelve silicon-based MEMS differential pressure sensors are vertically and evenly arranged at 20cm intervals along the inner side of the cold storage door frame, approximately 5mm from the door gap, forming a high-density differential pressure sensing array. The sensor range is -5Pa to +5Pa, with a resolution of no less than 0.01Pa, and sampling at a frequency of 200Hz.

[0202] Infrared heat flux density sensor: A thermopile type infrared array sensor with a field of view of 60° is installed on the inner side of the top of the door frame, aligned with the center line of the door gap, and monitors the direction, intensity and spatial distribution of heat flux at a frequency of 100Hz.

[0203] All sensor data is transmitted to the central processing unit via an industrial Ethernet bus, with end-to-end latency controlled within 5 milliseconds.

[0204] The door opening event recognition module runs in the central processing unit and processes the raw data collected by the sensing module in real time. Its core function is:

[0205] The door gap is divided into three sub-regions: upper, middle, and lower. The differential pressure signal is differentiated by a sliding window (10ms) to calculate the differential pressure gradient.

[0206] When the pressure gradient in the upper sub-region continuously exceeds the threshold of 0.8 Pa / ms for more than 20 ms, it is determined to be a gate opening event. At the same time, the infrared heat flow data is processed (such as region growth or connected component analysis) to identify the core area of ​​heat intrusion, and the main path direction and rate of heat intrusion are calculated by the centroid displacement between consecutive frames.

[0207] The thermal intrusion dynamic prediction module is the core of the system's intelligent decision-making, and it deploys a spatiotemporal coupled prediction model based on a convolutional long short-term memory network (ConvLSTM).

[0208] The model receives the door gap pressure difference sequence (past 2 seconds, 200Hz), heat flux density sequence (past 2 seconds, 100Hz), real-time external temperature and humidity, and the current three-dimensional temperature field inside the warehouse (as the initial state) as inputs.

[0209] After model inference, the model outputs a predicted field of temperature perturbation increments within the library over the next 5 seconds, with 100ms intervals and a spatial resolution of 0.5m grid. The entire inference process is completed within 80 milliseconds, thanks to the central processing unit deployed on a high-performance embedded multi-core processor equipped with a neural network acceleration coprocessor.

[0210] The space-oriented cold airflow control module is a rapid response mechanism that performs precise compensation based on forecast results. It mainly includes:

[0211] Based on the predicted temperature field, the target area is delineated by spatial weighted focusing and threshold determination (such as 0.5℃), and the optimal nozzle activation combination, injection angle and flow distribution are calculated for this area.

[0212] The nozzle is a customized component, consisting of three parts in sequence: a piezoelectric ceramic drive valve (response time ≤30ms), a venturi ejector chamber (throat diameter 8mm, used to eject air from the mixing chamber), and an adjustable deflector (-45° to +45° adjustable, resolution 1°).

[0213] The refrigerant flow rate into the nozzle is regulated by an electromagnetic proportional valve (response time 10ms, control accuracy 2%). The system refrigerant supply pressure is maintained within a stable range of 300 kPa to 500 kPa by a variable frequency screw compressor and an energy storage and pressure stabilizing tank.

[0214] The system captures the door opening signal through hardware interrupts and completes the entire process from issuing the command to multiple nozzles working together to form a directional cold air barrier within 50 milliseconds.

[0215] The temperature field reconstruction module inside the warehouse provides the system with a temperature benchmark for the entire warehouse space, which is used to verify predictions and trigger self-learning.

[0216] It relies on 40 high-precision platinum resistance temperature sensors placed in key locations within the warehouse (shelf gaps, main flow areas, and dead corners).

[0217] At a frequency of once per second, an inverse distance weighted interpolation algorithm that incorporates simplified CFD physical constraints is used to reconstruct sparse measurement point data into a three-dimensional temperature field covering the entire database with a resolution of 0.5 meters, and this process is completed within 2 seconds after the door closes.

[0218] The online model fine-tuning module enables the system to have long-term adaptive capabilities.

[0219] A dual-buffer mechanism is adopted. The main buffer stores the complete data of the most recent 100 gate operation cycles in a rolling manner; the auxiliary buffer is specifically used to store event data marked as prediction failure (such as residuals exceeding 0.3℃).

[0220] When fine-tuning is triggered, the weights of the prediction model are rapidly iterated using data from the auxiliary buffer with a small batch size (e.g., 8) and a learning rate (e.g., 0.001) using stochastic gradient descent (typically within 10 iterations), and the entire process is completed within 500 milliseconds.

[0221] The fine-tuned new model weights are seamlessly updated via atomic pointer switching, ensuring uninterrupted prediction services.

[0222] In summary, this system achieves source suppression and precise compensation for thermal disturbances caused by door opening through closed-loop collaboration of four stages: multi-source sensing of door gaps, dynamic prediction of thermal intrusion, spatial directional airflow control, and online model fine-tuning. This fundamentally overcomes the problems of slow response, coarse compensation, and slow execution in existing technologies.

[0223] It will be apparent to those skilled in the art that the present invention is not limited to the details of the exemplary embodiments described above, and that the present invention can be implemented in other specific forms without departing from the spirit or essential characteristics of the present invention. Therefore, the embodiments should be regarded as exemplary and non-limiting in all respects.

[0224] Furthermore, it should be understood that although this specification describes embodiments, not every embodiment includes only one independent technical solution. This narrative style is merely for clarity. Those skilled in the art should consider the specification as a whole, and the technical solutions in each embodiment can also be appropriately combined to form other embodiments that can be understood by those skilled in the art.

Claims

1. A method for constant temperature control in an air-conditioned cold storage facility, characterized in that, include: S1, Deploy a multi-source sensing array in the cold storage door seam area to collect pressure difference change signals and heat flow vector distribution data in the door seam area in real time; S2, based on the pressure difference change signal and heat flow vector distribution data, identify the door opening event and determine the main path information of hot air intrusion; S3, based on the door opening event and hot air intrusion information, call the hot intrusion dynamic prediction model and output the predicted value of temperature disturbance increment in the three-dimensional space of the warehouse within a set time in the future; S4. Based on the predicted temperature disturbance increment, a spatial directional cold airflow control command is generated to activate the high-speed pulse cold air nozzle at the corresponding position to form a directional cold air barrier. S5. After the door is closed, the dynamic prediction model for heat intrusion is fine-tuned online based on the residual between the actual temperature field and the predicted temperature field inside the warehouse.

2. The air-conditioned cold storage constant temperature control method according to claim 1, characterized in that, In step S1, the multi-source sensing array includes a differential pressure sensor array and an infrared heat flux density sensor arranged along the inner side of the door frame. The differential pressure sensor array consists of multiple differential pressure sensing points arranged at preset intervals, used to collect differential pressure change signals; The infrared heat flux density sensor is used to collect heat flux vector distribution data in the door gap area.

3. The air-conditioned cold storage constant temperature control method according to claim 2, characterized in that, Step S2 specifically includes: The door gap area is divided into multiple sub-regions, and the differential pressure signal within each sub-region is differentiated to obtain the differential pressure gradient. When the differential pressure gradient continuously exceeds the first time threshold and is greater than the first preset threshold, it is determined that a door opening event has occurred; The heat flow vector distribution data is processed to identify the heat intrusion core area, and the main path direction and rate of heat intrusion are calculated by the displacement of the heat intrusion core area between consecutive frames.

4. The method for constant temperature regulation of an air-conditioned cold storage according to claim 1, characterized in that, In step S3, the dynamic prediction model for thermal intrusion is a spatiotemporal coupled model constructed based on the Convolutional Long Short-Term Memory (ConvLSTM) network. Its inputs include pressure difference time series, heat flux density time series, external environmental parameters, and the initial temperature field inside the reservoir. The model outputs the predicted temperature perturbation increment within the three-dimensional grid of the library over a future set time period.

5. The method for constant temperature control of an air-conditioned cold storage according to claim 1, characterized in that, Step S4 specifically includes: Based on the predicted temperature disturbance increment and the main path of thermal intrusion, a weighted disturbance intensity distribution field is generated through spatial weighting calculation. Based on the weighted disturbance intensity distribution field and the second preset threshold, the cold airflow target area is determined; Based on the geometric information and thermal intrusion direction of the target area, the combination of high-speed pulse cooling nozzles to be activated, the deflection angle of each nozzle, and the jet flow rate are calculated.

6. The method for constant temperature regulation of an air-conditioned cold storage according to claim 5, characterized in that, The high-speed pulsed cold air nozzle includes a piezoelectric ceramic drive valve, a venturi ejector cavity, and an adjustable deflector vane connected in sequence. The piezoelectric ceramic drive valve is used to achieve rapid opening and closing of the air circuit; The Venturi ejector cavity is used to eject air from the mixing chamber to form a mixed cold air jet; The adjustable deflector is used to adjust the direction of the cold air jet.

7. The method for constant temperature regulation of an air-conditioned cold storage according to claim 1, characterized in that, Step S5 specifically includes: After the door is closed, the actual three-dimensional temperature field inside the warehouse is reconstructed based on data from multiple temperature sensors arranged inside the warehouse. The actual temperature field is compared with the predicted temperature field at the corresponding time output in step S3, and the spatial residual is calculated. If the absolute value of the residual at any grid point exceeds the third preset threshold, the online model fine-tuning mechanism will be triggered.

8. The method for constant temperature control of an air-conditioned cold storage according to claim 7, characterized in that, The online fine-tuning mechanism of the model employs a dual-buffer management system for data. The main buffer stores complete data packets for historical gate operation cycles in a circular fashion; The auxiliary buffer is specifically used to store event packets marked as predicted failures; During fine-tuning, the model parameters are iteratively updated using the data in the auxiliary buffer.

9. An air-conditioned cold storage constant temperature control system, used to implement the air-conditioned cold storage constant temperature control method as described in any one of claims 1 to 8, characterized in that, include: The multi-source sensing module for door gaps is used to perform millisecond-level detection in the door gap area of ​​cold storage, and to collect pressure difference change signals and heat flow vector distribution data; The door opening event recognition module is used to recognize door opening events based on the sensing data and determine the main path information of hot air intrusion; The thermal intrusion dynamic prediction module stores a thermal intrusion dynamic prediction model, which is used to predict the temperature disturbance increment in the library based on the event and path information. A spatial directional cold airflow control module is used to generate control commands based on the predicted values ​​and drive high-speed pulse cold air nozzles to perform directional cold air compensation. The temperature field reconstruction module inside the warehouse is used to reconstruct the actual temperature field based on the temperature sensor data inside the warehouse. The online model fine-tuning module is used to fine-tune the dynamic prediction model of thermal intrusion online based on the residual between the predicted temperature field and the actual temperature field.

10. The air-conditioned cold storage constant temperature control system according to claim 9, characterized in that, The spatial directional cold airflow control module includes multiple high-speed pulsed cold air nozzles, each nozzle comprising a piezoelectric ceramic drive valve, a Venturi ejector cavity, and an adjustable deflector.

Images