AI-driven commercial refrigerator dynamic energy saving and load prediction system

Abstract

This invention relates to the field of artificial intelligence and equipment control technology, and discloses an AI-driven dynamic energy-saving and load prediction system for commercial freezers. The system includes: reusing the compressor's operation and shutdown cycles as an active detection process of the freezer's thermodynamic state; extracting multi-scale dynamic features that simultaneously characterize total heat load and heat exchange efficiency from the temperature time-series data generated during this process; inputting these features into a machine learning model for operating condition classification; and then achieving adaptive adjustment of the control strategy. This invention utilizes the data stream generated by the system's own operation to achieve a shift from passive response to a single temperature to insight into the core latent variable of dynamic heat load. This enables the control system to identify real operating conditions and make optimal energy-saving decisions, improving the energy efficiency and operational reliability of the equipment in complex dynamic environments.

Description

Technical Field

[0001] This invention relates to an AI-driven dynamic energy-saving and load prediction system for commercial freezers, belonging to the field of artificial intelligence and equipment control technology. Background Technology

[0002] Currently, the common technical approach in the control of commercial refrigeration equipment is to use temperature sensors to monitor the temperature inside the cabinet and control the start and stop of the compressor based on one or a set of fixed temperature thresholds. Even with the introduction of machine learning algorithms, the technical approach mainly focuses on predicting future temperature changes in order to more precisely adjust the start and stop thresholds. This method has a clear effect on maintaining stable temperature inside the cabinet and is widely deployed due to its simple structure and high cost-effectiveness. However, when this control method is applied to the actual operating environment of commercial freezers with high dynamics and strong interference, an inherent technical limitation becomes the main factor restricting its performance. That is, the decision-making basis of this method is only the resultant physical quantity of temperature, and it lacks direct measurement of the process core variable of dynamic heat load that drives temperature changes.

[0003] This limitation in decision-making mechanisms can directly lead to potential operational risks and energy efficiency losses in complex commercial scenarios. For example, a freezer that has just undergone a large-scale restocking of warm goods and a freezer that has been running stably after closing at night, even if they display the exact same temperature reading at a certain moment, have completely different internal thermodynamic states. The former has a huge latent heat load and its temperature is about to rise rapidly, while the latter is in a low-load stable state. Existing control systems, unable to perceive this core difference, will adopt the same control strategy for both, which inevitably leads to a delayed response to the former and over-control of the latter. Specifically, existing technologies mainly have the following shortcomings: 1. Single dimension of decision information: The system cannot physically distinguish between multiple states such as high external load, internal latent heat release, or decreased heat exchange efficiency. All its energy-saving strategies are imprecise responses based on the superficial phenomenon of temperature; 2. Static and fixed control logic: The system does not have the ability to dynamically adjust its control objectives and strategies according to the real-time changing load state, thus generating unnecessary energy consumption under low-load conditions and potential safety risks under high-load conditions.

[0004] To address this issue, a direct approach is to add components such as door switch sensors, infrared sensors, or product scanning devices to the system to obtain richer status information. However, this approach significantly increases hardware costs, installation complexity, and system failure rates. For cost-sensitive commercial freezers requiring high reliability, this approach is generally not technically and economically viable for large-scale application. In addition to hardware limitations, existing technologies have also failed to address the core issue in terms of control logic innovation. For example, Chinese invention patent CN211147016U discloses a horizontal commercial energy-saving and environmentally friendly freezer that uses internal partitions and independent sealing covers to achieve physical partitioning, aiming to reduce cold air leakage during single door openings to achieve energy savings. However, this solution based on physical structure optimization does not change the underlying refrigeration control logic, still relying on traditional fixed threshold temperature control. It cannot detect real-time changes in dynamic heat load caused by factors such as restocking and customer flow. Although this solution improves insulation performance to some extent, it fails to resolve the mismatch between the control strategy and actual operating conditions, limiting its energy-saving potential. The essence of the problem then transforms into a more universal technical issue in the field of machine learning applications. Therefore, the technical problem to be solved by this invention is how to use the existing single temperature sensor in the system to extract the structured features that can effectively characterize the core latent variable of dynamic heat load from the temperature time series data stream generated by it, and on this basis, to build a machine learning control model that can perform accurate state recognition and adaptive decision-making. Summary of the Invention

[0005] This invention provides an AI-driven dynamic energy-saving and load prediction system for commercial refrigerators. Its main purpose is to solve the problem that under the technical conditions of using only existing temperature sensor data, it is impossible to effectively extract features that can characterize dynamic heat load, which leads to the control system being unable to accurately identify the real operating conditions and make adaptive adjustments.

[0006] To achieve the above objectives, the present invention provides an AI-driven dynamic energy-saving and load prediction system for commercial refrigerators, the system comprising:

[0007] A thermodynamic detection cycle definition module is configured to define a complete work and stop cycle of the compressor of a commercial freezer as a detection cycle for actively detecting the thermodynamic state inside the freezer.

[0008] A time-series data acquisition module is configured to acquire temperature time-series data covering the cooling process during compressor operation and the temperature recovery process after compressor stops operating within the cabinet during the detection period.

[0009] A multi-scale feature extraction module is configured to generate a feature vector characterizing the thermodynamic state based on temperature time-series data. This generation operation includes: within the initial time window of compressor startup, determining the time required from startup to the first time the temperature time-series data shows a continuous and stable decrease exceeding a preset noise threshold, as an initial thermal response delay feature characterizing the evaporator heat exchange efficiency; within the time interval of the cooling process, extracting the rate of change of the temperature time-series data as a cooling rate feature characterizing the health of the refrigeration system; and within the time interval of the warming process, extracting the rate of change of the temperature time-series data as a warming rate feature directly related to the total heat load of the freezer.

[0010] A machine learning classification module is configured to input feature vectors into a preset machine learning classification model to output a classification result that characterizes the current dynamic load state of the freezer.

[0011] An adaptive control module is configured to automatically adjust the range of temperature control thresholds used to control the next start-up and shutdown of the compressor based on the classification results.

[0012] The preferred, pre-defined machine learning classification model is generated through supervised learning training using feature vectors collected under various predefined typical operating conditions and their corresponding known dynamic load state labels.

[0013] Preferably, the multi-scale feature extraction module is further configured to: use the temperature recovery rate feature as the core indicator characterizing the instantaneous total heat load of the freezer caused by the heat penetration of the external environment and the release of latent heat from the warm goods inside; the machine learning classification module is configured to use the magnitude of the temperature recovery rate feature as the main criterion for classifying the dynamic load state as a deep energy-saving state, a standard state, or a high-load state.

[0014] Preferably, the dynamic load status output by the machine learning classification module also includes an abnormal warning status; the machine learning classification module is further configured to: when the cooling rate feature calculated in multiple consecutive detection cycles shows a unidirectional decrease, the dynamic load status is determined as an abnormal warning status.

[0015] Preferably, the adaptive control module is further configured to: when the classification result output by the machine learning classification module is a deep energy-saving state, relax the upper and lower limits of the temperature control threshold and extend the minimum sleep time of the compressor; when the classification result output by the machine learning classification module is a high load state, tighten the upper and lower limits of the temperature control threshold and shorten the minimum sleep time of the compressor.

[0016] Preferably, the system further includes: a healthy warming morphology template storage module, configured to extract and store a mathematical template composed of key feature point sequences or polynomial coefficients by averaging or curve fitting multiple temperature time-series data of the warming process acquired under standard load conditions; an online morphology monitoring module, configured to calculate the morphological similarity index between the real-time collected temperature time-series data of the warming process and the mathematical template after the compressor stops working; and an instant intervention module, configured to determine that a continuous abnormal high load event has occurred when the morphological similarity index is continuously lower than a preset similarity threshold, and send an intervention signal to the adaptive control module to force the adaptive control module to switch to high load mode.

[0017] Preferably, the system also includes a system health status self-diagnosis module, which is configured to: establish an online statistical baseline characterizing the normal operating status of the commercial freezer based on multiple feature vectors generated by the multi-scale feature extraction module at the initial stage of system operation; and continuously monitor subsequently extracted feature vectors. The statistical deviation relative to the statistical baseline, expressed as the Mahalanobis distance. The quantification is performed, and the calculation rules are as follows: ,in, This is the currently extracted feature vector. The mean vector of the statistical baseline. It is the inverse covariance matrix of the statistical baseline; when the statistical deviation continues to exceed a preset threshold, an early warning signal indicating that the temperature sensor has drifted is generated.

[0018] Preferably, the system further includes a dynamic baseline compensation module, which is configured to: record the cooling rate features and warming rate features extracted at multiple moments when the machine learning classification module determines that the system is currently in a deep energy-saving state, and establish a dynamic correlation model characterizing the influence of ambient temperature on these two features; before determining the current dynamic load state of the freezer, substitute the real-time extracted cooling rate features into the dynamic correlation model to calculate an expected warming rate baseline containing current ambient temperature information; and subtract the expected warming rate baseline from the real-time extracted warming rate features to generate a net load warming rate feature that has eliminated the influence of ambient temperature changes, and use this net load warming rate feature to replace the original warming rate feature for use by the machine learning classification module.

[0019] Preferably, the feature vector generated by the multi-scale feature extraction module also includes a cycle power-time ratio feature; the multi-scale feature extraction module is further configured to calculate the ratio of the compressor operating time to the total duration of the entire detection cycle to obtain the cycle power-time ratio feature, which is used to characterize the recent average load level.

[0020] Preferably, the online morphology monitoring module is configured to calculate the morphological similarity index by comparing the sign and magnitude of the curvature or second derivative of the temperature time series data of the real-time collected warming process to see if it is within the normal range defined by the mathematical template.

[0021] Compared with the prior art, the beneficial effects of the present invention are:

[0022] 1. By continuously acquiring temperature time-series data during a complete working cycle of the compressor and the subsequent shutdown period, and extracting thermodynamic probe response features containing cooling and warming process characteristics based on the dynamic form of the data, the machine learning classification model is no longer based on isolated temperature point information, but on a structured feature vector that can characterize the current total heat load in the cabinet. This approach changes the basis of control decision-making from a passive response to a single physical quantity to a direct analysis of the system's dynamic process, enabling the system to distinguish between two operating conditions with the same temperature reading but completely different internal heat load states, and thus execute differentiated control strategies.

[0023] 2. While extracting the thermodynamic probe response features characterizing the cycle, further features characterizing the initial thermal response delay are extracted in the initial stage of compressor startup. These two features obtained at different time scales are then merged into the same feature vector and input into a machine learning model for classification. This approach, which combines long-cycle cyclic features with startup transient features, provides the system with two independent evaluation dimensions. This allows the system to not only judge load changes caused by external factors but also simultaneously evaluate heat exchange efficiency changes caused by internal factors. This avoids the erroneous decision of misjudging internal efficiency decay as ineffective defrosting or overload operation due to high external load startup.

[0024] 3. While utilizing the extracted thermodynamic probe response features for real-time operating condition classification and decision-making, multiple continuously generated features are accumulated over time to establish a statistical baseline characterizing the system's normal operating state. Subsequently, the statistical deviation of newly generated features from this baseline is continuously monitored. This method of reusing instantaneous decision information as long-term state monitoring information utilizes the data stream generated by the system's own operation to establish an internal self-calibrating reference system. When the temperature sensor itself experiences a slow performance drift, leading to systematic distortion of all features, this deviation will show a continuous changing trend. Based on this, the system generates an early warning signal, preventing the entire decision-making system from making judgments based on erroneous sensor data without the system's awareness. The system identifies a low-load operating state and establishes a dynamic correlation model between the cooling rate and the recovery rate to characterize the systematic influence of ambient temperature on the thermodynamic probe response characteristics. Before determining the dynamic load state, the model calculates an expected recovery rate baseline using the real-time measured cooling rate and compensates for the actual measured recovery rate. This approach uses one feature affected by ambient temperature to calibrate another feature also affected by ambient temperature. Without introducing additional environmental sensors, it separates the combined effects of internal load changes and external environmental fluctuations on the temperature curve shape, enabling the feature ultimately used for decision-making to more purely reflect the real load caused by user behavior. Attached Figure Description

[0025] Figure 1 This is a flowchart of the system data flow and adaptive control closed loop of the present invention;

[0026] Figure 2 This is a diagram showing the use case relationships of the system of the present invention in the operation and maintenance scenarios;

[0027] Figure 3 This is a block diagram of the system perception, decision-making and control module of the present invention. Detailed Implementation

[0028] To make the objectives, technical solutions, and advantages of the present invention clearer, the present invention will be further described in detail below with reference to the accompanying drawings. Obviously, the described embodiments are only some embodiments of the present invention, and not all embodiments. Based on the embodiments of the present invention, all other embodiments obtained by those skilled in the art without creative effort are within the scope of protection of the present invention.

[0029] An AI-driven dynamic energy-saving and load prediction system for commercial refrigerators utilizes existing temperature sensors and microcontrollers within the refrigerator. It transforms time-series temperature data streams into a representation of the refrigerator's dynamic heat load through a data processing and machine learning workflow, enabling adaptive adjustment of control strategies. The system includes a thermodynamic detection cycle definition module, a time-series data acquisition module, a multi-scale feature extraction module, a machine learning classification module, and an adaptive control module. Temperature data acquired by the time-series data acquisition module is processed by the multi-scale feature extraction module to generate feature vectors. These feature vectors serve as input to the machine learning classification module, and the output classification results guide the adaptive control module in adjusting its operating strategy. A key technical challenge in the operation of commercial refrigerators is that a single temperature reading cannot distinguish between two different physical states. For example, a refrigerator operating stably at night with an internal air temperature of 4... The freezer and the refrigerator, which had just been replenished with room temperature goods, both had an internal air temperature of 4 degrees Celsius. To address this challenge, the method claimed in this invention is configured to perform the following procedures: First, a thermodynamic detection cycle definition module is configured to define a working and stopping cycle of the compressor of a commercial freezer as a detection cycle for actively detecting the thermodynamic state inside the freezer. This operation establishes a unified observation time reference for subsequent feature extraction. Within this detection cycle, a time-series data acquisition module is configured to use existing temperature sensors in the system to acquire temperature time-series data covering the cooling process during compressor operation and the temperature recovery process after the compressor stops operating, at a sampling frequency of not less than 1Hz. A multi-scale feature extraction module receives the time-series data and performs calculations to transform it into a feature vector that can characterize the current thermodynamic state. The generation of this feature vector is a process of converting discrete temperature data points into a structured input containing dynamic information of the system that can be processed by a machine learning model.

[0030] To transform the raw temperature data stream into structured features with clear physical meaning, the execution logic of the multi-scale feature extraction module was determined to include analysis at both long and short time scales. At the long time scale, the module analyzes both the cooling and warming processes. Given that the health of the refrigeration system is reflected in its cooling efficiency, this module is configured to extract the rate of change of temperature time-series data within the cooling process time interval, serving as a cooling rate feature characterizing the health of the refrigeration system. For example, if the compressor reduces the temperature from 5.0 degrees Celsius to 6.0 degrees Celsius within 300 seconds... Decreased to 2.0 ,but It was calculated as (5.0 - 2.0). / 300s=0.01 / s; Furthermore, since the total heat load inside the cabinet determines the rate of temperature recovery after the compressor stops, this module is accordingly configured to extract the rate of change of temperature time-series data within the time interval of the recovery process, as a recovery rate feature directly related to the total heat load of the freezer. For example, if the temperature drops from 2.0°C within 180 seconds after the compressor stops... It rose back to 3.5 ,but It is calculated as (3.5-2.0) / 180s≈0.0083 / s, the The numerical values ​​of these features become one of the core bases for subsequent machine learning models to determine the system load status. Simultaneously, to capture factors affecting heat exchange efficiency, such as evaporator frosting, analysis needs to be performed on a short timescale. Therefore, this module is further configured to collect temperature data at a frequency of, for example, 10Hz within the initial time window of compressor startup, such as the first 15 seconds after startup. It also determines the time required from startup until the temperature time-series data first shows a continuous and stable decrease exceeding a preset noise threshold, serving as an initial thermal response delay feature characterizing the evaporator's heat exchange efficiency. The preset noise threshold here is determined during the calibration process before the equipment leaves the factory. This is achieved by recording the maximum temperature fluctuation within 1 second after the compressor starts under no-load conditions 100 times, and taking 1.5 times the average value. For example, 0.05. A healthy evaporator corresponds to Typically, it takes less than 1.5 seconds, while an evaporator with thick frost will have a corresponding... It may extend to more than 8 seconds; ultimately, this module will combine the features obtained at different time scales. , and And a cycle power-time ratio characteristic obtained by calculating the ratio of the compressor's operating time to the total duration of the entire detection cycle. These are combined to form a multi-dimensional feature vector, such as [0.01, 0.0083, 1.5, 0.4], and then output.

[0031] Subsequently, a machine learning classification module, whose core is a pre-trained machine learning classification model embedded in the microcontroller, such as a support vector machine or decision tree model, is configured to take feature vectors generated by the multi-scale feature extraction module as input. This machine learning classification model is generated through supervised learning training before the equipment leaves the factory, using hundreds of feature vectors and their corresponding known dynamic load state labels collected under various predefined typical operating conditions, such as no-load, full-load, frequent door opening, and restocking. For example, state 0 is deep energy saving, state 1 is standard, and state 2 is high load. After receiving the feature vectors, this module performs one inference and finally outputs a classification result representing the current dynamic load state of the freezer. To address the gradual degradation of the refrigeration system, the dynamic load state also includes an abnormal warning state. The machine learning classification module is further configured to calculate the cooling rate feature over multiple consecutive detection cycles, such as 10 consecutive cycles. If the load shows a unidirectional decrease and the cumulative decrease exceeds a threshold calibrated based on the initial health state, such as exceeding 20% ​​of the initial value, the dynamic load state will be determined as an abnormal warning state, and a maintenance alarm will be triggered.

[0032] An adaptive control module, which is a state machine, automatically adjusts the range of the temperature control threshold used to control the next start and stop of the compressor based on the classification results output by the machine learning classification module. Specifically, the adaptive control module is further configured to execute differentiated control logic. When the classification result is a deep energy-saving state, the system determines that the current heat load is low, and then relaxes the upper and lower limits of the temperature control threshold, for example, from [2]. 5 Relaxed to [1] 7 ], and extend the minimum sleep time of the compressor, for example, from 10 minutes to 25 minutes; conversely, when the classification result is a high load state, the system judges that there is a large latent heat load, and tightens the upper and lower limits of the temperature control threshold, for example, tightening it to [2 4 Furthermore, the minimum sleep time of the compressor is shortened, for example, from 10 minutes to 3 minutes; in this way, the system's control strategy changes from static and fixed to dynamic and adaptive, enabling energy consumption and operating strategies to match real-time changing operating conditions.

[0033] To address persistent high-load events such as cabinet doors not being fully closed, the system also includes a parallel online morphological monitoring mechanism. At its core is a healthy recovery morphological template storage module. This module is configured to, during the initial stage of system operation, extract and store a mathematical template composed of key feature point sequences or polynomial coefficients by averaging or curve fitting multiple temperature time-series data points acquired under standard load conditions during the recovery process. An online morphological monitoring module is then configured to, after the compressor stops operating, calculate the morphological similarity index between the real-time collected temperature time-series data of the recovery process and the mathematical template. The metric is calculated by comparing the curvature of the real-time temperature curve or the sign and magnitude of the second derivative to see if they fall within the normal range defined by the mathematical template. An instant intervention module is configured to determine a persistent high-load event when the morphological similarity index remains below a preset similarity threshold, such as below 0.85 for 60 consecutive seconds. This module then sends an intervention signal to the adaptive control module to force it to switch to high-load mode and sends an alarm message to the background. The introduction of this mechanism adds a layer of continuous status monitoring to the system in addition to periodic detection.

[0034] Furthermore, to ensure the reliability of the decision-making system during long-term operation, the system also integrates a self-diagnostic and adaptive compensation capability. Considering that potential performance drift of the temperature sensor could cause biases in feature extraction and decision-making, the system also includes a system health status self-diagnosis module. This self-diagnosis module is configured to, in the initial stage of system operation, such as the first 7 days, establish an online statistical baseline characterizing the normal operating status of the commercial freezer based on multiple feature vectors generated by the multi-scale feature extraction module. This baseline is defined by a mean vector. and an inverse covariance matrix Defined; in subsequent operations, this module continuously monitors the subsequently extracted feature vectors. The statistical deviation relative to the statistical baseline, expressed as the Mahalanobis distance. The quantification is performed, and the calculation rules are as follows: ,in, This is the currently extracted feature vector. The mean vector of the statistical baseline. The inverse covariance matrix of the statistical baseline is used. When the moving average of the statistical deviation exceeds a preset threshold for an extended period of time, such as exceeding three times the initial value of the baseline for 24 consecutive hours, the system generates an early warning signal indicating that the temperature sensor may be drifting.

[0035] Meanwhile, to avoid interference from ambient temperature fluctuations in the feature extraction process, the system further includes a dynamic baseline compensation module. The operating mechanism of this module is to learn the influence of ambient temperature on the system's thermodynamic properties when the system enters a deep energy-saving state, i.e., a quasi-no-load state. Specifically, this dynamic baseline compensation module is configured to record the extracted cooling rate features at multiple moments when the machine learning classification module determines that the system is currently in a deep energy-saving state. Characteristics of the reheat rate Based on this, a dynamic correlation model characterizing the influence of ambient temperature on these two characteristics is established, such as a linear regression model. Before determining the current dynamic load status of the freezer, the system will input the real-time extracted cooling rate characteristics into the dynamic correlation model to calculate a baseline of expected reheat rate that includes the current ambient temperature information. Ultimately, the system extracts the reheat rate characteristics in real time. In the middle, subtract the expected recovery rate baseline to generate a net load recovery rate characteristic that has eliminated the effects of ambient temperature variations. Furthermore, the net load recovery rate feature is used to replace the original recovery rate feature for use by the machine learning classification module; this compensation mechanism, without introducing additional environmental sensors, utilizes the physical correlation between features to suppress environmental interference.

[0036] Example 1: On a summer afternoon, a commercial freezer equipped with the system of this invention was in standard operating condition, with its internal temperature maintained at 2°C. Up to 5 In between, at this time, the convenience store clerk had more than one hundred bottles with a temperature of 25 degrees Celsius. Bottled beverages were added to the freezer, creating a concentrated latent heat load. Subsequently, the convenience store entered its peak customer traffic period, and over the next hour, the freezer door was frequently opened, introducing a continuous convective heat load. During the first compressor start-stop cycle after the replenishment operation, the system entered a detection cycle, and the time-series data acquisition module recorded the temperature time-series data of the cooling and warming processes. The multi-scale feature extraction module calculated based on this data the rate of change of the temperature time-series data during the warming process, i.e., the warming rate feature, due to the presence of a large amount of warm beverages. This value was calculated as high, and it was associated with an increase in the total heat load inside the cabinet; the machine learning classification module received the value containing this high value. After analyzing the feature vectors, the current dynamic load state of the freezer is classified as a high load state; accordingly, the adaptive control module adjusts the control strategy based on this classification result, adjusting the range of the temperature control threshold for the next start-up and shutdown from the standard [2]. 5 Tighten to [2] 4 This reduces the minimum sleep time of the compressor, allowing it to operate at a higher frequency to cope with increased latent heat load.

[0037] As peak passenger flow continues, the system repeats the aforementioned feature extraction and classification decision-making process in each subsequent detection cycle. Due to frequent door opening actions, the temperature recovery rate feature... By maintaining a consistently high temperature range, the machine learning classification module continuously classifies the operating condition as either standard or high-load, thus the adaptive control module maintains a relatively aggressive temperature control strategy. This operating mode utilizes the temperature recovery rate characteristic of the total heat load. The cooling rate characteristic of system health In the collaborative analysis process, the machine learning classification module's decision-making does not rely solely on a single feature, but rather on a comprehensive judgment of the entire feature vector. This resolves the technical problem in traditional control methods that cannot distinguish between high external loads and internal efficiency degradation, potentially leading to erroneous decisions. The system changes its decision-making basis from a single temperature result to a structured analysis of the system's thermodynamic dynamics. After the convenience store closes in the evening, the freezer remains unopened for several hours, entering a stable low-load operation phase. During the first detection cycle after closing, due to the lack of external heat intrusion and internal latent heat release, the multi-scale feature extraction module calculates the rewarming rate feature... The value is extremely low; based on this, the machine learning classification module determines the dynamic load state as a deep energy-saving state; after receiving this result, the adaptive control module adjusts the strategy again, widening the upper and lower limits of the temperature control threshold to [1]. 7 This strategy extends the minimum sleep time of the compressor, reducing the number of compressor starts and allowing the system to enter a low-energy operation mode while ensuring food safety at night. The entire process requires no manual intervention, and the system autonomously switches between different operating modes based on the identification of heat load at different time scales.

[0038] Example 2: To objectively verify the energy-saving effect and temperature control capability of the method of the present invention under different heat load conditions, a comparative test platform was built; the platform consisted of two commercial refrigerators of the same model and specifications, placed in an environment where the temperature was constantly controlled at 30°C. ±0.5 Inside the test chamber, one of the refrigerators served as a control group, and its operation was controlled by a standard temperature control threshold that was fixed at [2]. 5 The first freezer is controlled by an electronic temperature controller; the second freezer, used as a test unit employing the method of this invention, is controlled by a microcontroller integrating all the functional modules described in the aforementioned specific embodiments; both freezers are equipped with independent energy meters with an accuracy of 0.01 kWh, and a temperature sensor independent of the control system for data recording, with a measurement accuracy of ±0.1 kWh. The data sampling frequency is set to 1Hz. This sampling frequency setting is based on the Nyquist sampling theorem and is a value greater than twice that of the main frequency component of the system temperature change, after balancing the two factors of capturing the details of dynamic temperature changes and avoiding data redundancy.

[0039] The experiment was designed to simulate a typical 24-hour business operation cycle, divided into three consecutive 8-hour phases. The first phase, a high-load operation phase, simulated peak daytime business hours. A 50W resistance heat source, controlled by a timer, was switched on for 2 minutes every 15 minutes inside the freezer to simulate the convective heat load introduced by frequent customer door openings. The second phase, a standard operation phase, simulated off-peak hours, with the heat source switching frequency reduced to 2 minutes every 30 minutes. The third phase, a low-load operation phase, simulated nighttime closing. No external heat load was applied during the test. Throughout the 24-hour test cycle, the cumulative power consumption and average temperature inside the two freezers were continuously recorded. During the test, the system of the test group using the method of this invention continuously executed the complete process from temperature time series data acquisition to feature vector generation and dynamic load state classification in each detection cycle. A key data processing process is shown in Table 1. This table selects data from representative detection cycles in the three stages to show how the original input data is processed into key intermediate features and finally generates decision output.

[0040] Table 1: Examples of data processing for the sample group of the present invention at different operating stages.

[0041]

[0042] Referring to Table 1, during the high-load phase, due to frequent external heat injection, the system's calculated temperature recovery rate characteristics... The numerical value is relatively high, and the machine learning classification module classifies the operating condition as a high-load state, driving the adaptive control module to execute a more proactive cooling strategy; while in the low-load stage, the temperature recovery rate characteristic... The power consumption was significantly reduced, and the system classified it as a deep energy-saving state, accordingly relaxing the control threshold and reducing the number of compressor starts. This process confirms that the method can effectively identify changes in external operating conditions using calculated intermediate characteristic values ​​and translate the identification results into differentiated control actions. After completing a continuous 24-hour test, the recorded overall data was analyzed. The total power consumption of the control group was 2.15 kWh, and its average cabinet temperature during the entire test was 3.4 kWh. The total power consumption of the test group using the method of this invention was 1.78 kWh, and its average cabinet temperature was 3.6 degrees Celsius. Data shows that, compared with the control group, the test group using the method of this invention achieved a 17.2% reduction in power consumption while maintaining the average temperature inside the cabinet within a reasonable range. Most of the energy-saving benefits occurred during the low-load operation phase. The experimental results indicate that the method of this invention, by introducing dynamic load identification and adaptive control based on machine learning, can distinguish different operating conditions and execute matching energy-saving strategies under changing operating environments.

[0043] Example 3: This example combines Figures 1 to 3 This section describes an AI-driven dynamic energy-saving and load prediction system for commercial refrigerated display cases, such as... Figure 1 As shown, the process begins with sensing the physical state of the commercial refrigerator, including information such as temperature, door opening / closing, and product load. Then, the time-series data acquisition module collects temperature time-series data from the compressor's complete operating and shutdown cycles. This data stream is first sent to the multi-scale feature extraction module to generate a structured feature vector representing the thermodynamic state. Simultaneously, this real-time feature vector is also sent to the system health self-diagnosis module to monitor the statistical deviation of the feature vector over a long period, thereby providing early warnings of sensor drift and other faults and outputting system warning signals. The feature vector generated by the multi-scale feature extraction module is processed by the dynamic baseline compensation module using a dynamic correlation model established from historical deep energy-saving state data to generate a compensated net load feature that eliminates environmental temperature fluctuation interference. This feature vector is then input into the machine learning classification module, which outputs a classification result representing the current dynamic load, such as deep energy saving, standard, or high load state. This classification result is used to update the model of the dynamic baseline compensation module online and serves as the decision basis for the adaptive control module. Based on the classification result, the adaptive control module dynamically adjusts the temperature control threshold for compressor start-up and shutdown, ultimately instructing the compressor to execute the optimal energy-saving decision.

[0044] like Figure 2As shown, convenience store clerks, as the primary users, interact with the system to achieve the basic goal of maintaining daily food safety. They also achieve the goal of efficiently cooling and replenishing goods by dynamically adjusting energy-saving strategies. Maintenance technicians, as another type of user, receive system health alerts to diagnose and maintain the health status of sensors. There are also connections between the functional modules within the system. For example, the function of dynamically adjusting energy-saving strategies utilizes data provided by the function of compensating for the influence of ambient temperature, while the function of receiving system health alerts relies on the output generated by the function of diagnosing the health status of sensors.

[0045] like Figure 3 As shown, this system is based on commercial freezers, compressors, and temperature sensors. Its core functional modules include a thermodynamic detection period definition module, which sets the time reference for data acquisition by the time-series data acquisition module. The data acquired by the time-series data acquisition module is input to a multi-scale feature extraction module, which is responsible for calculating the initial thermal response delay. Cooling rate Reheating rate and cycle work time ratio Multiple features are then fed into the machine learning classification module to output classification results such as deep energy-saving state, standard state, or high load state. Finally, the adaptive control module executes control strategies including temperature threshold adjustment, sleep time adjustment, and compressor start-stop based on the results. In addition, the system also includes multiple parallel or auxiliary functional modules. Among them, the dynamic baseline compensation module is used for ambient temperature compensation, the system health status self-diagnosis module is used for sensor drift detection, and the branch consisting of the health recovery mode template storage module, the online mode monitoring module, and the real-time intervention module is responsible for real-time monitoring and intervention of abnormal high load events.

[0046] Example 4: In this example, to ensure the stability of the calibration process, the calibration is conducted in a standardized test environment. The core of this environment is a test chamber with controllable ambient temperature and a commercial refrigerator to be calibrated, which already has the system of this invention installed inside. The refrigerator is initially set to be empty, dry, and clean. The calibration procedure first establishes a dynamic correlation model to compensate for changes in ambient temperature, given that ambient temperature is a characteristic affecting the cooling rate. Characteristics of the reheat rate Determining the variables and their relationship is a prerequisite for improving the robustness of the model; therefore, the ambient temperature of the test chamber was set to 15°C. 25 and 35 The system operates stably at three temperature points. At each ambient temperature point, the freezer system is set to operate autonomously in a deep energy-saving state without load, continuously recording at least 10 complete detection cycles. After each cycle, the system automatically extracts the cooling rate characteristics corresponding to that cycle. Characteristics of the reheat rate Thus, at each ambient temperature setpoint, a set of 10 groups was obtained. A sample set consisting of data points; after collecting data from all three temperature points, these 30 data points are used as a sample, and the least squares method is applied to perform linear regression analysis to calculate the dynamic correlation model. slope parameter in With intercept parameter The value was calculated in a specific implementation of this calibration procedure. , These parameters were then embedded in the freezer's microcontroller, serving as the basis for environmental temperature compensation during online operation.

[0047] Next, the procedure shifts to calibrating the online morphological monitoring module used for immediate intervention. This step establishes decision boundaries for determining persistent abnormal high-load events such as cabinet doors not being properly closed. Firstly, in 25... Under ambient temperature, the freezer was run for 5 detection cycles under standard load. The system recorded the complete temperature time series data of these 5 recovery processes. By averaging these 5 curves point by point, a baseline curve was generated. This baseline curve was then processed into a mathematical template composed of polynomial coefficients and stored as a healthy recovery pattern template. Next, a 1cm gap was artificially created on the freezer door to simulate an open door condition, and the system was run for 5 more detection cycles under this condition, recording the corresponding recovery curves. The online morphological monitoring module calculated the morphological similarity index between these 5 abnormal curves and the stored healthy recovery pattern template, obtaining a set of similarity values: 0.65, 0.68, 0.66, 0.70, and 0.67. To ensure the reliability of the decision, the preset similarity threshold was determined to be a value between normal and abnormal operating conditions. The setting procedure was to take the maximum similarity value of 0.70 measured under abnormal operating conditions and the baseline similarity of 1.0 under normal operating conditions. An intermediate value was set, with an added safety margin, ultimately set to 0.85. This threshold was also embedded into the microcontroller. Finally, a training dataset was generated for the machine learning classification module. The system sequentially simulated various predefined typical operating conditions in the test chamber, including no load, standard load, full load, and a one-time large heat injection simulating replenishment. Each operating condition was run for a sufficient time to collect data for at least 50 probe cycles. During the data acquisition process, each extracted feature vector was automatically associated with a predefined dynamic load state label corresponding to the current simulated operating condition. Thus, a training dataset containing hundreds of feature vectors with state labels was constructed. This dataset was then used to conduct supervised learning training on a support vector machine or decision tree model on an offline computing platform. The model parameters after training were finally burned into the microcontroller of the freezer. By executing the above complete calibration procedure, the key models, parameters, and thresholds of the system were determined based on objective experimental data before it was put into actual use.

[0048] Example 5: Upon initial deployment of each commercial freezer, a standardized online self-calibration procedure is executed to establish a decision-making benchmark for the system's health status self-diagnosis module. This procedure is set to run automatically for 72 hours after the equipment is initially powered on. During this period, the system controller guides the freezer through three preset operating conditions: no load, standard load, and full load, each lasting 24 hours. Throughout this process, feature vectors generated by the multi-scale feature extraction module are continuously collected. The system uses the set of all feature vectors collected within these 72 hours to calculate an online statistical baseline characterizing the normal operating status of the specific equipment. This baseline includes the mean vector. Covariance Matrix It is stored in a fixed format and used as a reference standard for subsequent sensor drift warning of the device.

[0049] To maintain the compensation accuracy of the dynamic baseline compensation module throughout the equipment's lifecycle and to address the slow changes in equipment performance due to long-term operation, the system is configured to execute a model parameter update procedure based on a sliding time window. This procedure continuously collects cooling rate features extracted from all detection cycles identified as deep energy-saving states by the machine learning classification module within the last 30 calendar days. Characteristics of the reheat rate The data pairs; at the end of each natural day, the system uses all the data within this sliding window to refine the dynamic correlation model. parameters and If the relative deviation between the newly calculated parameter value and the currently used parameter value exceeds a preset update threshold of 5%, the system will update the model parameters to the newly calculated values ​​and use the updated parameters for subsequent environmental temperature compensation calculations.

[0050] Example 6: In another application scenario, the system background data of a sample unit of this invention that has been running stably for six months shows that although the statistical distribution of the daily dynamic load status classification results in the past month has not changed significantly compared with the previous five months, the initial thermal response delay features calculated by the multi-scale feature extraction module have changed. The average cycle time showed a trend of slowly increasing from an initial 1.5s to 6.8s; based on this, the system determined that the evaporator's heat exchange efficiency was gradually decreasing due to frosting, and accordingly, automatically triggered a defrosting cycle during a preset low-electricity-price period at night; after defrosting, the newly collected data... The value recovered to 1.6s, proving that this feature has a direct indicative role in identifying system performance degradation caused by internal factors.

[0051] To further verify the indispensability of the core steps of the present invention for working condition identification based on dynamic feature extraction from the perspective of decision-making mechanism, the following comparative examples are set up.

[0052] Comparative Example 1: This comparative example uses the exact same commercial freezer hardware and test environment as the test group in Example 2. The test process also simulates the same 24-hour commercial operation cycle under high load, standard load, and low load. The only essential difference is that the freezer control system of this comparative example adopts the conventional technical path described in the background art, that is, an electronic thermostat that makes decisions based solely on instantaneous temperature readings. The temperature threshold used to control the start and stop of the compressor is fixedly set at [2]. 5 The control system does not have the ability to process temperature time series data, extract multi-scale dynamic features, or classify dynamic load states using machine learning models. During the 24-hour test cycle, the operating data of the conventional control system was recorded, and its key performance and results analysis at different load stages are shown in Table 2 below.

[0053] Table 2: Performance of conventional control strategies at different operating stages.

[0054]

[0055] After the experiment, the overall data was statistically analyzed. The refrigerator using the conventional control strategy consumed a total of 2.15 kWh of electricity during the 24-hour test period, and the average temperature inside the refrigerator was 3.7°C. The experimental results show that, due to the lack of direct measurement of the core latent variable of dynamic heat load, conventional control strategies cannot effectively distinguish between high-load and low-load conditions. Their fixed control logic exhibits response lag and insufficient temperature control capability under high load, while leading to energy waste under low load. This result, from the opposite perspective, confirms that the step of extracting dynamic feature vectors from temperature time-series data for condition identification in this invention is the key technology for achieving energy saving and precise temperature control.

[0056] 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.

[0057] Finally, it should be noted that the above embodiments are only used to illustrate the technical solutions of the present invention and are not intended to limit it. Although the present invention has been described in detail with reference to preferred embodiments, those skilled in the art should understand that modifications or equivalent substitutions can be made to the technical solutions of the present invention without departing from the spirit and scope of the technical solutions of the present invention.

Claims

1. An AI-driven dynamic energy-saving and load prediction system for commercial refrigerators, characterized in that, The system includes: A thermodynamic detection cycle definition module is configured to define a complete work and stop cycle of the compressor of a commercial freezer as a detection cycle for actively detecting the thermodynamic state inside the freezer. A time-series data acquisition module is configured to acquire temperature time-series data covering the cooling process during compressor operation and the temperature recovery process after compressor stops operating within the cabinet during the detection period. A multi-scale feature extraction module is configured to generate a feature vector characterizing the thermodynamic state based on temperature time-series data. This generation operation includes: within the initial time window of compressor startup, determining the time required from startup to the first time the temperature time-series data shows a continuous and stable decrease exceeding a preset noise threshold, as an initial thermal response delay feature characterizing the evaporator heat exchange efficiency; within the time interval of the cooling process, extracting the rate of change of the temperature time-series data as a cooling rate feature characterizing the health of the refrigeration system; and within the time interval of the warming process, extracting the rate of change of the temperature time-series data as a warming rate feature directly related to the total heat load of the freezer. A machine learning classification module is configured to input feature vectors into a preset machine learning classification model to output a classification result that characterizes the current dynamic load state of the freezer. An adaptive control module is configured to automatically adjust the range of temperature control thresholds used to control the next start-up and shutdown of the compressor based on the classification results.

2. The AI-driven dynamic energy-saving and load prediction system for commercial refrigerators according to claim 1, characterized in that, The pre-defined machine learning classification model is generated through supervised learning training using feature vectors collected under various predefined typical operating conditions and their corresponding known dynamic load state labels.

3. The AI-driven dynamic energy-saving and load prediction system for commercial refrigerators according to claim 1, characterized in that, The multi-scale feature extraction module is further configured to use the temperature recovery rate feature as the core indicator characterizing the instantaneous total heat load of the freezer caused by the heat penetration from the external environment and the release of latent heat from the warm goods inside; the machine learning classification module is configured to use the magnitude of the temperature recovery rate feature as the main criterion for classifying the dynamic load state as deep energy-saving state, standard state or high load state.

4. The AI-driven dynamic energy-saving and load prediction system for commercial refrigerators according to claim 1, characterized in that, The dynamic load status output by the machine learning classification module also includes an anomaly warning status; The machine learning classification module is further configured to: when the cooling rate feature calculated over multiple consecutive detection cycles shows a unidirectional decrease, the dynamic load state is identified as an abnormal warning state.

5. The AI-driven dynamic energy-saving and load prediction system for commercial refrigerators according to claim 1, characterized in that, The adaptive control module is further configured to: when the classification result output by the machine learning classification module is a deep energy-saving state, relax the upper and lower limits of the temperature control threshold and extend the minimum sleep time of the compressor; when the classification result output by the machine learning classification module is a high-load state, tighten the upper and lower limits of the temperature control threshold and shorten the minimum sleep time of the compressor.

6. The AI-driven dynamic energy-saving and load prediction system for commercial refrigerators according to claim 1, characterized in that, The system also includes: a healthy recovery morphology template storage module, configured to extract and store a mathematical template composed of key feature point sequences or polynomial coefficients by averaging or curve fitting multiple temperature time-series data of the recovery process acquired under standard load conditions; an online morphology monitoring module, configured to calculate the morphological similarity index between the real-time collected temperature time-series data of the recovery process and the mathematical template after the compressor stops working; and an instant intervention module, configured to determine that a persistent abnormal high load event has occurred when the morphological similarity index is continuously lower than a preset similarity threshold, and send an intervention signal to the adaptive control module.

7. The AI-driven dynamic energy-saving and load prediction system for commercial refrigerators according to claim 1, characterized in that, The system also includes a system health status self-diagnosis module, which is configured to: establish an online statistical baseline characterizing the normal operating status of the commercial freezer based on multiple feature vectors generated by the multi-scale feature extraction module at the initial stage of system operation; and continuously monitor subsequently extracted feature vectors. The statistical deviation relative to the statistical baseline, expressed as the Mahalanobis distance. The quantification is performed, and the calculation rules are as follows: ,in, This is the currently extracted feature vector. The mean vector of the statistical baseline. It is the inverse covariance matrix of the statistical baseline; when the statistical deviation continues to exceed a preset threshold, an early warning signal indicating that the temperature sensor has drifted is generated.

8. The AI-driven dynamic energy-saving and load prediction system for commercial refrigerators according to claim 1, characterized in that, The system also includes a dynamic baseline compensation module, which is configured to: record the extracted cooling rate features and warming rate features at multiple moments when the machine learning classification module determines that the system is currently in a deep energy-saving state, and establish a dynamic correlation model characterizing the influence of ambient temperature on these two features; before determining the current dynamic load state of the freezer, substitute the real-time extracted cooling rate features into the dynamic correlation model to calculate an expected warming rate baseline that includes the current ambient temperature information; and subtract the expected warming rate baseline from the real-time extracted warming rate features to generate a net load warming rate feature that has eliminated the influence of ambient temperature changes, and use this net load warming rate feature to replace the original warming rate feature for use by the machine learning classification module.

9. The AI-driven dynamic energy-saving and load prediction system for commercial refrigerators according to claim 1, characterized in that, The feature vector generated by the multi-scale feature extraction module also includes a cycle power-time ratio feature; the multi-scale feature extraction module is further configured to calculate the ratio of the compressor operating time to the total duration of the entire detection cycle in order to obtain the cycle power-time ratio feature.

10. The AI-driven dynamic energy-saving and load prediction system for commercial refrigerators according to claim 6, characterized in that, The online morphology monitoring module is configured to calculate the morphological similarity index by comparing the sign and magnitude of the curvature or second derivative of the temperature time series data of the real-time temperature recovery process to see if it is within the normal range defined by the mathematical template.

Images