Intelligent control method and system for refrigeration house

Abstract

The invention provides an intelligent control method and system for a refrigeration house, and relates to the field of control. A refrigeration house space thermodynamic diagram model is constructed based on temperature measuring points, and a historical temperature sequence, a house door opening and closing state sequence, cargo warehouse-in and warehouse-out records and external environment parameters are obtained; predicting a temperature change track of each node in a future control time domain by using a space-time diagram convolutional network; establishing an equipment health state evaluation model based on a hidden Markov model, evaluating the health state of each refrigeration compressor unit and quantifying the health state into an equipment health index; and solving the rolling optimization objective function in the control time domain to obtain an optimal start-stop time sequence and operation power combination instruction of each refrigeration compressor unit and auxiliary equipment, and issuing the combination instruction to a corresponding field controller for execution. The temperature change of each area in the refrigeration house can be accurately predicted, the operation load can be reasonably distributed according to the health condition of the unit, the service life of equipment is effectively prolonged, and the sudden failure risk and the maintenance cost of the whole life cycle are reduced.

Description

Technical Field

[0001] The present application relates to the field of control, and in particular to a cold storage intelligent control method and system. Background Art

[0002] Cold storage maintains a low temperature environment to ensure the quality of stored goods. Stable, efficient, and economical operation is crucial for cold storage temperature control. Currently, most cold storage temperature control systems still use a traditional threshold control strategy. This involves installing one or a few temperature sensors within the cold storage. The refrigeration compressor starts when the measured temperature exceeds a preset upper threshold, and stops when it falls below a lower threshold. While this control approach is logically simple and easy to implement, it exhibits numerous drawbacks in practice. The control strategy is lagging and reactive, unable to predictably respond to load fluctuations. This leads to large temperature fluctuations within the storage, making it difficult to meet high-quality storage requirements. Furthermore, it completely ignores the impact of dynamic factors such as time-of-use electricity prices, external environmental changes, door openings, and the movement of goods in and out of the storage on energy consumption, often causing refrigeration units to operate at full capacity during peak electricity price periods. Furthermore, frequent, unoptimized starts and stops not only impact the power grid but also increase mechanical wear on core equipment like compressors, shortening their service life and increasing maintenance costs and the risk of failure.

[0003] To overcome the shortcomings of traditional control methods, some improved solutions, such as introducing simple prediction models for load forecasting and combining time-of-use electricity prices to perform pre-cooling operations during periods of low electricity prices, are proposed to reduce operating costs. However, existing prediction models are often oversimplified, typically treating the entire cold storage as a lumped parameter system. This is unable to accurately characterize the complex thermodynamic coupling relationships between different areas within the warehouse, as well as the local dynamic thermal disturbances caused by the activities of goods and personnel. This leads to insufficient temperature prediction accuracy and poor control effects. Moreover, existing technologies are single or fragmented in their optimization objectives, usually focusing only on energy conservation or temperature stability, and failing to incorporate the health status and loss costs of the equipment itself into the closed loop of operational decision-making. The control system is unable to perceive the health status of each refrigeration unit in real time and dynamically adjust its operating load accordingly. This may cause the equipment to operate in a sub-healthy state for a long time, which not only reduces operating efficiency but also creates safety risks such as sudden failures. Summary of the Invention

[0004] In view of the above problems, the present invention proposes a cold storage intelligent control method, comprising the following steps: The system obtains the historical temperature sequence of each temperature measurement point in the cold storage, the sequence of door opening and closing status, the cargo entry and exit records, external environmental parameters, and time-of-use electricity price data; and collects the vibration spectrum, current, power consumption, and cumulative operating time data of each refrigeration compressor unit in real time; constructs a thermodynamic graph model of the cold storage space based on the temperature measurement points, in which the temperature measurement points are nodes of the graph model and the thermodynamic coupling relationships between nodes are edges. Using a spatiotemporal graph convolutional network, based on the historical temperature sequence, the sequence of door opening and closing status, the cargo entry and exit records, and external environmental parameters, it predicts the temperature change trajectory of each node in the future control time domain; For each refrigeration compressor unit, an equipment health status assessment model based on a hidden Markov model is established. The real-time collected vibration spectrum, current, power consumption, and cumulative operating time data are used to evaluate the health status of each unit and quantify it into an equipment health index. A rolling optimization objective function for model predictive control is established. The rolling optimization objective function is a weighted function and includes: a storage temperature deviation penalty term calculated based on the predicted temperature change trajectory, an operating energy consumption cost term calculated based on the time-of-use electricity price data, and an equipment loss cost term that is negatively correlated with the equipment health index. The rolling optimization objective function is solved in the control time domain to obtain the optimal start and stop timing and operating power combination instructions for each refrigeration compressor unit and auxiliary equipment, and the combination instructions are sent to the corresponding field controller for execution.

[0005] Preferably, the acquisition of the historical temperature sequence of each temperature measurement point in the cold storage, the door opening and closing state sequence, the goods entry and exit records, the external environmental parameters and the time-of-use electricity price data specifically includes: Temperature data is periodically collected through temperature sensors deployed at multiple locations within the cold storage. Door opening and closing events are recorded through door status sensors. The cargo identification system records cargo attributes including type and volume, as well as inbound and outbound information. Furthermore, the time-of-use electricity price list for the coming day is obtained from the grid operator through a network interface.

[0006] Preferably, predicting the temperature change trajectory of each node in the future control time domain specifically includes: The temperature sequence of each measuring point in a historical period, the sequence of warehouse door opening and closing status, the cargo entry and exit records, and the external environmental parameters are used as the input of the spatiotemporal graph convolutional network to predict the temperature value of each node in a future control time domain at multiple discrete time steps.

[0007] Preferably, the evaluating the health status of each unit specifically includes: Feature quantities that can reflect the operating status of the equipment are extracted from the vibration spectrum, current, and power consumption data; using the hidden Markov model, the probability of each unit being in a preset multiple health level status is calculated based on the feature quantities and the cumulative operating time; and based on the distribution of the probabilities, the equipment health index is calculated by weighted summation, wherein the higher the health level, the larger the corresponding weight coefficient.

[0008] Preferably, the value of the equipment loss cost item in the rolling optimization objective function is proportional to the predicted start and stop times of each unit in the control time domain, and inversely proportional to the equipment health index of each unit.

[0009] Preferably, the step of solving the rolling optimization objective function adopts a heuristic optimization algorithm; and the step of sending the combined instruction to the field controller for execution includes: The optimal instructions for one or more future time steps are issued for execution, and after each control cycle, the latest system status is obtained and the next round of optimization and instruction issuance is carried out in a rolling manner.

[0010] Preferably, for each refrigeration compressor unit, an equipment health status assessment model based on a hidden Markov model is established, and the real-time collected vibration spectrum, current, power consumption and cumulative operating time data are used to assess the health status of each unit and quantify it into an equipment health index, specifically: a) using a non-parametric Bayesian method to perform cluster analysis on the historical vibration spectrum, current, power consumption, and cumulative operating time data to obtain implicit health states representing different degradation stages of the equipment; b) building a deep generative observation model based on physical information constraints, which uses prior knowledge of compressor thermodynamics, electromagnetics, and / or mechanical dynamics as structural constraints to calculate the observation probability of real-time data under various states; c) combining the health state and the deep generative observation model of the physical information constraints, and using a sequential Monte Carlo method to estimate the probability distribution of each unit in each implicit health state; d) performing expectation calculation on the obtained probability distribution to obtain the equipment health index.

[0011] The present invention also proposes a cold storage intelligent control system, comprising: The prediction unit is used to obtain the historical temperature sequence of each temperature measurement point in the cold storage, the sequence of the door opening and closing status, the cargo entry and exit records, the external environmental parameters, and the time-sharing electricity price data; and to collect the vibration spectrum, current, power consumption, and cumulative operating time data of each refrigeration compressor unit in real time; based on the temperature measurement points, a thermodynamic graph model of the cold storage space is constructed, wherein the temperature measurement points are nodes of the graph model, and the thermodynamic coupling relationship between the nodes is an edge. The spatiotemporal graph convolutional network is used to predict the temperature change trajectory of each node in the future control time domain based on the historical temperature sequence, the sequence of the door opening and closing status, the cargo entry and exit records, and the external environmental parameters; An evaluation unit is configured to establish, for each refrigeration compressor unit, an equipment health status evaluation model based on a hidden Markov model, and utilize real-time collected vibration spectrum, current, power consumption, and accumulated operating time data to evaluate the health status of each unit and quantify it into an equipment health index; establish a rolling optimization objective function for model predictive control, wherein the rolling optimization objective function is a weighted function and includes: a storage temperature deviation penalty term calculated based on the predicted temperature change trajectory, an operating energy consumption cost term calculated based on the time-of-use electricity price data, and an equipment loss cost term that is negatively correlated with the equipment health index; A control unit is used to solve the rolling optimization objective function in the control time domain, obtain the optimal start and stop timing and operating power combination instructions for each refrigeration compressor unit and auxiliary equipment, and send the combination instructions to the corresponding field controller for execution.

[0012] Preferably, the acquisition of the historical temperature sequence of each temperature measurement point in the cold storage, the door opening and closing state sequence, the goods entry and exit records, the external environmental parameters and the time-of-use electricity price data specifically includes: Temperature data is periodically collected through temperature sensors deployed at multiple locations within the cold storage. Door opening and closing events are recorded through door status sensors. The cargo identification system records cargo attributes including type and volume, as well as inbound and outbound information. Furthermore, the time-of-use electricity price list for the coming day is obtained from the grid operator through a network interface.

[0013] Preferably, predicting the temperature change trajectory of each node in the future control time domain specifically includes: The temperature sequence of each measuring point in a historical period, the sequence of warehouse door opening and closing status, the cargo entry and exit records, and the external environmental parameters are used as the input of the spatiotemporal graph convolutional network to predict the temperature value of each node in a future control time domain at multiple discrete time steps.

[0014] Preferably, the evaluating the health status of each unit specifically includes: Feature quantities that can reflect the operating status of the equipment are extracted from the vibration spectrum, current, and power consumption data; using the hidden Markov model, the probability of each unit being in a preset multiple health level status is calculated based on the feature quantities and the cumulative operating time; and based on the distribution of the probabilities, the equipment health index is calculated by weighted summation, wherein the higher the health level, the larger the corresponding weight coefficient.

[0015] Preferably, the value of the equipment loss cost item in the rolling optimization objective function is proportional to the predicted start and stop times of each unit in the control time domain, and inversely proportional to the equipment health index of each unit.

[0016] Preferably, the solution of the rolling optimization objective function adopts a heuristic optimization algorithm; the sending of the combined instructions to the field controller for execution includes: sending and executing the optimal instructions for one or more future time steps, and after each control cycle, obtaining the latest system status, and rolling for the next round of optimization and instruction issuance.

[0017] Preferably, for each refrigeration compressor unit, an equipment health status assessment model based on a hidden Markov model is established, and the real-time collected vibration spectrum, current, power consumption and cumulative operating time data are used to assess the health status of each unit and quantify it into an equipment health index, specifically: a) using a non-parametric Bayesian method to perform cluster analysis on the historical vibration spectrum, current, power consumption, and cumulative operating time data to obtain implicit health states representing different degradation stages of the equipment; b) building a deep generative observation model based on physical information constraints, which uses prior knowledge of compressor thermodynamics, electromagnetics, and / or mechanical dynamics as structural constraints to calculate the observation probability of real-time data under various states; c) combining the health state and the deep generative observation model of the physical information constraints, and using a sequential Monte Carlo method to estimate the probability distribution of each unit in each implicit health state; d) performing expectation calculation on the obtained probability distribution to obtain the equipment health index.

[0018] Compared with existing technologies, this invention, by constructing a thermodynamic graph model of cold storage space and utilizing a spatiotemporal graph convolutional network, can accurately predict future temperature changes in various areas within the storage, improving the accuracy and uniformity of storage temperature control, thereby better ensuring the quality of stored items. This invention quantifies the real-time assessment of equipment health status as loss costs, and incorporates this into the unified optimization goal of model predictive control along with storage temperature control accuracy and operating energy consumption costs. This allows for the management of equipment losses while seeking the lowest operating electricity bill, rationally allocating operating loads based on the health of the units, and avoiding overuse of critical equipment in a sub-healthy state, thereby effectively extending equipment service life, reducing the risk of sudden failures, and reducing maintenance costs throughout the entire life cycle. BRIEF DESCRIPTION OF THE DRAWINGS

[0019] Figure 1 is a flow chart of a specific embodiment; Figure 2 It is a visual diagram of the thermodynamic diagram model of the cold storage space; Figure 3 Schematic diagram of the overall optimization objective function; Figure 4 Schematic diagram of iteration. DETAILED DESCRIPTION

[0020] The following will be combined with the drawings in the embodiments of the present application to clearly and completely describe the technical solutions in the embodiments of the present application. Obviously, the embodiments described are only part of the embodiments of the present application, not all of the embodiments. Based on the embodiments in the present application, all other embodiments obtained by those skilled in the art without making creative efforts are within the scope of protection of this application.

[0021] In a specific embodiment, the present invention proposes a cold storage intelligent control method, which is characterized by comprising the following steps: S1, obtain the historical temperature sequence of each temperature measurement point in the cold storage, the sequence of door opening and closing status, the cargo entry and exit records, external environmental parameters, and time-of-use electricity price data; and collect the vibration spectrum, current, power consumption, and cumulative operating time data of each refrigeration compressor unit in real time; construct a cold storage space thermodynamic graph model based on the temperature measurement points, where the temperature measurement points are nodes of the graph model and the thermodynamic coupling relationship between nodes is the edge. Using the spatiotemporal graph convolutional network, based on the historical temperature sequence, the sequence of door opening and closing status, the cargo entry and exit records, and external environmental parameters, predict the temperature change trajectory of each node in the future control time domain; Specifically, historical temperature sequences are collected by deploying multiple Pt100 platinum resistance temperature sensors at key locations such as the internal walls of the cold storage and near cargo stacks. Hall sensors or travel switches are installed on the warehouse doors to obtain the door opening and closing status sequences. Goods entry and exit records are obtained by scanning the RFID tags or barcodes on the goods and combining them with the warehouse management system (WMS). External environmental parameters such as outdoor temperature and humidity are obtained by calling the public meteorological service API interface. Time-of-use electricity price lists for the next 24 hours are obtained from the power supply department or smart meters. Accelerometers, current transformers, and smart electricity meters are installed on each refrigeration compressor unit to collect vibration, current, and power consumption data, respectively. All data are stored in a time series database with a unified timestamp.

[0022] Abstract N temperature measurement points into N nodes of the graph model; determine the heat transfer influence intensity between nodes by calculating the Pearson correlation coefficient between the historical temperature series of each measurement point or using the Granger causality test, connect the node pairs with an intensity higher than the preset threshold into edges, and the weight of the edge is the influence intensity value, thereby constructing a weighted adjacency matrix, such as Figure 2 As shown in the figure, a spatiotemporal graph convolutional network model consisting of graph convolutional layers and gated recurrent units (GRUs) is constructed. The model uses the temperature of each node, door status, estimated heat load for inflow and outflow, and external ambient temperature over a historical period of time, such as the past 24 hours, as multidimensional input features. The graph convolutional layer captures the spatial thermodynamic coupling relationship between different measurement points at the same time. The GRU layer then learns the dynamic evolution of the temperature of each measurement point over time. Finally, it outputs temperature predictions for all nodes every 15 minutes over a future control time domain, such as the next 6 hours, forming a complete temperature change trajectory.

[0023] S2. For each refrigeration compressor unit, establish an equipment health status assessment model based on a hidden Markov model. Utilize the real-time collected vibration spectrum, current, power consumption, and accumulated operating time data to assess the health status of each unit and quantify it into an equipment health index. Establish a rolling optimization objective function for model predictive control. The rolling optimization objective function is a weighted function that includes: a storage temperature deviation penalty term calculated based on the predicted temperature change trajectory, an operating energy consumption cost term calculated based on the time-of-use electricity price data, and an equipment loss cost term that is negatively correlated with the equipment health index. A fast Fourier transform (FFT) is performed on the real-time vibration signal to extract its amplitude, kurtosis, and margin at characteristic frequencies as vibration features. The root mean square value and crest factor of the current signal are extracted as current features. A hidden Markov model (HMM) with four hidden states—normal, light wear, severe wear, and impending failure—is established. The HMM is trained using the Baum-Welch algorithm using historical data from the equipment's entire lifecycle, from normal to failure, which includes the extracted feature vector sequence. A state transition matrix is ​​obtained. During actual operation, the real-time feature vector sequence is input into the trained HMM model and decoded using the Viterbi algorithm to obtain the most likely sequence of health states. The probabilities of each health state are weighted and summed. For example, the normal state has a weight of 1, light wear has a weight of 0.7, severe wear has a weight of 0.3, and impending failure has a weight of 0. This results in a continuous value between 0 and 1 that serves as the equipment health index.

[0024] The temperature deviation penalty term is the sum of the squares of the differences between all predicted temperature points and the set temperature upper and lower limits within the control time domain, multiplied by a penalty coefficient. The operating energy consumption cost term is the sum of the start and stop state variables of each unit at each control time step, 0 or 1, multiplied by its rated power and the time-of-use electricity price during that period, and then summed over the entire time domain. The equipment loss cost term is designed as an exponential function of the equipment health index, for example , where k is the adjustment coefficient. The lower the health index, the exponential growth of the cost item. Multiplying these three items by their respective weight coefficients w1, w2, and w3 and adding them together constitutes the rolling optimization objective function, as shown in Figure 3 shown.

[0025] S3, solving the rolling optimization objective function in the control time domain, obtaining the optimal start and stop timing and operating power combination instructions for each refrigeration compressor unit and auxiliary equipment, and sending the combination instructions to the corresponding field controller for execution.

[0026] The optimization objective function, along with a series of constraints, such as the storage temperature must be within an acceptable range, the minimum compressor start / stop interval, and a total power limit, forms a mixed-integer nonlinear programming problem. This problem is solved using a particle swarm optimization algorithm (PSO) or a genetic algorithm (GA). Within a set number of iterations, the space of decision variables is searched. The decision variables are the start / stop states of each compressor (0 or 1) at each time step within the control domain. The algorithm then outputs a sequence of start / stop times that minimizes the objective function. This sequence is the optimal control instruction. For example, the instruction is: Compressor 1 turns on from 0:00 AM to 2:00 AM, and Compressor 2 turns on from 1:30 AM to 4:00 AM. This instruction sequence is sent to the cold storage's programmable logic controller (PLC) via industrial Ethernet. The PLC then switches the contactors of the corresponding compressors on and off, completing the control execution. The entire optimization process is repeated over a rolling cycle, for example, every 30 minutes.

[0027] In an optional embodiment, the acquisition of historical temperature sequences of temperature measurement points in the cold storage, sequences of door opening and closing status, cargo entry and exit records, external environmental parameters, and time-of-use electricity price data specifically includes: periodically collecting temperature data through temperature sensors deployed at multiple locations in the cold storage; recording door opening and closing events through door status sensors; recording cargo attributes including cargo category and volume, and entry and exit information through a cargo identification system; and obtaining a time-of-use electricity price list for the next day from a power grid operator through a network interface.

[0028] For example, in a typical 3,000-cubic-meter cold storage facility, 12 high-precision PT100 temperature sensors might be evenly distributed along its length, width, and height. These sensors are set to collect temperature data every minute, generating a real-time temperature field data stream, such as T1 equals -18.5 degrees Celsius and T2 equals -18.3 degrees Celsius. Furthermore, a Hall effect sensor is installed on each cold storage door. When the door is open, the sensor outputs a high-level signal, and when it is closed, a low-level signal. This accurately records the start time and duration of each opening. For example, a door opening at 10:02 a.m. is recorded, lasting three minutes.

[0029] In terms of cargo management, ultra-high frequency RFID reader gates or visual recognition cameras installed at the warehouse entrance can automatically identify pallet label information, recording, for example, the arrival of a batch of frozen dumplings, totaling five pallets with a total volume of approximately 6 cubic meters, at 2:30 PM on October 26, 2023, at an initial temperature of -12°C. Electricity price information is automatically retrieved from the local power company's server at midnight each day through a dedicated API. This table details the specific electricity prices during peak, off-peak, and valley periods. For example, the valley price is 0.35 yuan per kilowatt-hour from 12:00 AM to 8:00 AM, while the peak price is as high as 1.2 yuan per kilowatt-hour from 6:00 PM to 9:00 PM.

[0030] In an optional embodiment, the prediction of the temperature change trajectory of each node in the future control time domain specifically includes: using the temperature sequence of each measuring point in a historical period, the opening and closing state sequence of the warehouse door, the goods in and out records and the external environmental parameters as the input of the spatiotemporal graph convolutional network, and predicting the temperature value of each node in a future control time domain at multiple discrete time steps.

[0031] More specifically, the 12 temperature measurement points within the cold storage are defined as 12 nodes of the graph. Based on the physical spatial location of these temperature measurement points within the warehouse, a weighted adjacency matrix is ​​established to represent the intensity of heat transfer between nodes. The closer the measurement points, the greater the weight. The model input data is very rich in dimensionality. For example, the temperature value sequence of all 12 measurement points recorded every 5 minutes over the past 24 hours is extracted to form a 12 times 288 matrix. The input also includes a binary sequence representing the door open / close status, such as 0011100, and cargo entry events, such as the entry of 5 cubic meters of cargo at the 120th time step.

[0032] After receiving multidimensional input data, the spatiotemporal graph convolutional network model uses its unique graph convolutional layer to capture the mutual influence of temperatures at different locations in the spatial dimension. It then uses temporal convolutional layers or recurrent units to learn the evolution of factors such as temperature at warehouse doors and cargo in the temporal dimension. The model outputs a future control domain, such as the predicted temperature of all 12 nodes at discrete time steps of 5 minutes over the next 60 minutes. For example, for node 5, the model will produce a series of predicted values, such as -18.1 degrees Celsius at the 5th minute, -18.0 degrees Celsius at the 10th minute, and -17.5 degrees Celsius at the 60th minute, thus forming a complete temperature trajectory.

[0033] In an optional embodiment, the evaluation of the health status of each unit specifically includes: extracting characteristic quantities that can reflect the operating status of the equipment from the vibration spectrum, current, and power consumption data; using the hidden Markov model, according to the characteristic quantities and the cumulative operating time, calculating the probability of each unit being in a preset multiple health level status; and based on the distribution of the probability, calculating the equipment health index by weighted summation, wherein the higher the health level, the larger the corresponding weight coefficient.

[0034] Taking a compressor unit as an example, the vibration signals collected by the accelerometer during operation were analyzed. The vibration spectrum was obtained through a fast Fourier transform, and the amplitudes at specific frequencies, such as 50 Hz and 100 Hz, were extracted as features. Simultaneously, the RMS current value and total harmonic distortion (THD) were extracted from the current data. These features, such as a vibration amplitude of 0.05g and a current harmonic distortion of 3%, along with the cumulative operating hours of the unit since its last overhaul, served as observations for the hidden Markov model.

[0035] Based on the hidden Markov model trained with historical data, the probability of the unit being in different states at the current moment can be calculated according to the currently input feature quantity observation sequence.

[0036] In an optional embodiment, the value of the equipment loss cost item in the rolling optimization objective function is proportional to the predicted number of starts and stops of each unit in the control time domain, and is inversely proportional to the calculated equipment health index of each unit.

[0037] Specifically, the cost item aims to reduce equipment wear and tear and extend its service life through economic leverage. In one embodiment, the calculation formula can be expressed as follows: the equipment wear and tear cost is equal to a basic loss coefficient multiplied by the predicted number of starts and stops, divided by the health index of the unit. For example, the basic loss coefficient is set to 5. This value represents the equivalent cost of the average electrical and mechanical impact caused by each start and stop on a brand new device. Within the control time domain of the next hour, if an optimization plan predicts that a unit needs to be started twice, then its basic loss is 10.

[0038] Costs are adjusted based on the real-time health of the equipment. For example, if Unit A has a health index of 0.95, indicating excellent health, the predicted loss cost for two starts and stops is approximately 10.53 yuan. Meanwhile, Unit B, due to long-term operation, has a health index of 0.6 and is also predicted to have two starts and stops, with a loss cost of 16.67 yuan. When seeking the solution with the lowest total cost, the optimization algorithm tends to minimize or stabilize the use of equipment in poor health, thereby protecting vulnerable equipment.

[0039] In an optional embodiment, the solution of the rolling optimization objective function adopts a heuristic optimization algorithm; the sending of the combined instructions to the field controller for execution includes: sending and executing the optimal instructions for one or more future time steps, and after the end of each control cycle, obtaining the latest system status, and rolling for the next round of optimization and instruction issuance.

[0040] Specifically, since the objective function includes multiple nonlinear and conflicting terms such as electricity costs and equipment losses, and the decision variables are the on / off states of multiple units at multiple time steps, this is a complex combinatorial optimization problem, and is preferably solved using a particle swarm optimization algorithm. At the beginning of each control cycle, for example at 10:00 a.m., the optimization time domain of the next 60 minutes will be used, with a time step of 5 minutes, to find the unit start and stop instruction sequence that can minimize the total cost in the next 60 minutes. For example, the optimal sequence is that unit 1 is turned on in the first 5 minutes and turned off in the next 55 minutes, unit 2 is completely shut down, and unit 3 is turned on from the 25th to the 35th minute.

[0041] Not all instructions for these 60 minutes will be executed at once. For example, after the above optimal sequence is calculated at 10:00 a.m., only the instruction for the first time step, that is, to turn on unit 1 between 10:00 and 10:05, will be sent to the on-site programmable logic controller PLC for execution. When the time comes to 10:05, the control cycle ends, and the latest temperature of the cold storage, unit status and other information will be collected again. Based on the system status, the particle swarm algorithm is started again, and the optimization calculation for the new time domain of 10:05 to 11:05 in the future is re-performed. The cycle of optimization, execution, feedback, and re-optimization will continue to roll uninterruptedly, such as Figure 4 shown.

[0042] In an optional embodiment, for each refrigeration compressor unit, an equipment health status assessment model based on a hidden Markov model is established. The real-time collected vibration spectrum, current, power consumption and cumulative operating time data are used to assess the health status of each unit and quantify it into an equipment health index, specifically: a) using a non-parametric Bayesian method to perform cluster analysis on the historical vibration spectrum, current, power consumption, and cumulative operating time data to obtain implicit health states representing different degradation stages of the equipment; b) building a deep generative observation model based on physical information constraints, which uses prior knowledge of compressor thermodynamics, electromagnetics, and / or mechanical dynamics as structural constraints to calculate the observation probability of real-time data under various states; c) combining the health state and the deep generative observation model of the physical information constraints, and using a sequential Monte Carlo method to estimate the probability distribution of each unit in each implicit health state; d) performing expectation calculation on the obtained probability distribution to obtain the equipment health index.

[0043] Using nonparametric Bayesian methods, such as HDP-HMM, historical data is analyzed to identify the stages that equipment progresses through from new to damaged. For example, it may be discovered that, in addition to typical bearing wear, there is an earlier stage of lubricant degradation. For each discovered state, a deep learning observation model is trained that is specific to that state and constrained by physical laws such as energy efficiency and mechanical dynamics. During the real-time evaluation phase, this model uses Sequential Monte Carlo or particle filtering to estimate the probability of the equipment being in each of the aforementioned health states. This model combines the latest sensor data, such as vibration and / or current, the degree to which this data matches the physical models for each state, and the impact of current operating instructions, such as high power operation, which accelerates state deterioration and transitions. For example, at a given moment, the system calculates that the equipment has a 70% probability of being in a lubricant deterioration state, with a score of 80, and a 30% probability of being in a stable and healthy state, with a score of 95. By weighted summing these probabilities, a device health index of 84.5 is obtained, enabling a precise quantitative assessment of the equipment's condition. In a specific embodiment, the deep learning observation model is implemented using a variational autoencoder (VAE). Specifically, for the VAE model corresponding to each implicit state, that is, the deep generative observation model, when a new set of observation data is input, the encoder of the deep generative observation model will compress it into a low-dimensional latent feature vector, and the decoder will reconstruct the original data from this feature vector. The observation probability is judged by calculating the similarity between the reconstructed data and the original data. If the VAE trained for a healthy state can reconstruct the current data, it means that the probability of the data being observed in a healthy state is very high.

[0044] The above embodiments are only used to illustrate the technical solutions of the present invention, rather than to limit them. Although the present invention has been described in detail with reference to the above embodiments, it should be understood by those skilled in the art that the technical solutions described in the above embodiments can still be modified, or some of the technical features thereof can be replaced by equivalents. However, these modifications or replacements do not deviate from the spirit and scope of the technical solutions of the embodiments of the present invention. In addition, the various different implementations of the embodiments of the present invention can also be arbitrarily combined, as long as they do not violate the ideas of the embodiments of the present invention, and they should also be regarded as the contents disclosed in the embodiments of the present invention.

Claims

1. A cold storage intelligent control method, characterized in that: The following steps are involved: Obtain historical temperature sequences of each temperature measurement point in the cold storage, door opening and closing status sequences, cargo entry and exit records, external environmental parameters, and time-of-use electricity price data; The vibration spectrum, current, power consumption, and cumulative operating time data of each refrigeration compressor unit are collected in real time. A thermodynamic graph model of the cold storage space is constructed based on the temperature measurement points, wherein the temperature measurement points are nodes of the graph model and the thermodynamic coupling relationships between the nodes are edges. The temperature change trajectory of each node in the future control time domain is predicted using a spatiotemporal graph convolutional network based on the historical temperature sequence, the door opening and closing state sequence, the cargo entry and exit records, and the external environmental parameters. For each refrigeration compressor unit, an equipment health status assessment model based on a hidden Markov model is established. The real-time collected vibration spectrum, current, power consumption, and cumulative operating time data are used to evaluate the health status of each refrigeration compressor unit and quantify it into an equipment health index. A rolling optimization objective function for predictive control is established. The rolling optimization objective function is a weighted function and includes: a storage temperature deviation penalty term calculated based on the predicted temperature change trajectory, an operating energy consumption cost term calculated based on the time-of-use electricity price data, and an equipment loss cost term that is negatively correlated with the equipment health index. The rolling optimization objective function is solved in the control time domain to obtain the optimal start and stop timing and operating power combination instructions for each refrigeration compressor unit and auxiliary equipment, and the combination instructions are sent to the corresponding field controller for execution.

2. The method according to claim 1, characterized in that The acquisition of the historical temperature sequence of each temperature measurement point in the cold storage, the door opening and closing state sequence, the cargo entry and exit records, the external environmental parameters and the time-of-use electricity price data specifically includes: Temperature data is periodically collected through temperature sensors deployed at multiple locations within the cold storage. Door opening and closing events are recorded through door status sensors. The cargo identification system records cargo attributes including type and volume, as well as inbound and outbound information. Furthermore, the time-of-use electricity price list for the coming day is obtained from the grid operator through a network interface.

3. The method according to claim 1, characterized in that The prediction of the temperature change trajectory of each node in the future control time domain specifically includes: The temperature sequence of each measuring point in a historical period, the sequence of warehouse door opening and closing status, the cargo entry and exit records, and the external environmental parameters are used as the input of the spatiotemporal graph convolutional network to predict the temperature value of each node in a future control time domain at multiple discrete time steps.

4. The method according to claim 1, wherein The health status of each refrigeration compressor unit is evaluated, specifically including: Feature quantities that can reflect the operating status of the equipment are extracted from the vibration spectrum, current, and power consumption data; using the hidden Markov model, the probability of each unit being in a preset multiple health level status is calculated based on the feature quantities and the cumulative operating time; and based on the distribution of the probabilities, the equipment health index is calculated by weighted summation, wherein the higher the health level, the larger the corresponding weight coefficient.

5. The method according to claim 1, characterized in that The value of the equipment loss cost item in the rolling optimization objective function is proportional to the predicted start and stop times of each refrigeration compressor unit in the control time domain, and inversely proportional to the equipment health index of each unit.

6. The method according to claim 1, characterized in that The method for solving the rolling optimization objective function adopts a heuristic optimization algorithm; The sending of the combined instructions to the field controller for execution includes sending the optimal instructions for one or more future time steps for execution, and after each control cycle, obtaining the latest system status and rolling out the next round of optimization and instruction sending.

7. The method according to claim 1, characterized in that For each refrigeration compressor unit, an equipment health status assessment model based on the hidden Markov model is established. The real-time collected vibration spectrum, current, power consumption and cumulative operating time data are used to assess the health status of each unit and quantify it into an equipment health index, specifically: a) Using non-parametric Bayesian methods, cluster analysis is performed on historical vibration spectrum, current, power consumption, and cumulative operating time data to obtain implicit health states that characterize different degradation stages of the equipment; b) constructing a deep generative observation model based on physical information constraints. The equipment health status assessment model uses prior knowledge of compressor thermodynamics, electromagnetics, and / or mechanical dynamics as structural constraints to calculate the observation probability of real-time data under various states; c) combining the health state and the deep generative observation model of the physical information constraints, and using a sequential Monte Carlo method to estimate the probability distribution of each unit in each implicit health state; d) performing expectation calculation on the obtained probability distribution to obtain the equipment health index.

8. A cold storage intelligent control system, characterized in that: include: The prediction unit is used to obtain the historical temperature sequence of each temperature measurement point in the cold storage, the door opening and closing status sequence, the goods in and out records, the external environmental parameters and the time-of-use electricity price data; The vibration spectrum, current, power consumption and cumulative operating time data of each refrigeration compressor unit are collected in real time. A thermodynamic graph model of the cold storage space is constructed based on the temperature measurement points, wherein the temperature measurement points are nodes of the graph model and the thermodynamic coupling relationship between the nodes is an edge. The temperature change trajectory of each node in the future control time domain is predicted by using a spatiotemporal graph convolutional network based on the historical temperature sequence, the door opening and closing state sequence, the goods in and out records and the external environmental parameters. An evaluation unit is configured to establish, for each refrigeration compressor unit, an equipment health status evaluation model based on a hidden Markov model, and utilize real-time collected vibration spectrum, current, power consumption, and accumulated operating time data to evaluate the health status of each unit and quantify it into an equipment health index; establish a rolling optimization objective function for model predictive control, wherein the rolling optimization objective function is a weighted function and includes: a storage temperature deviation penalty term calculated based on the predicted temperature change trajectory, an operating energy consumption cost term calculated based on the time-of-use electricity price data, and an equipment loss cost term that is negatively correlated with the equipment health index; A control unit is used to solve the rolling optimization objective function in the control time domain, obtain the optimal start and stop timing and operating power combination instructions for each refrigeration compressor unit and auxiliary equipment, and send the combination instructions to the corresponding field controller for execution.

9. The system according to claim 8, characterized in that The acquisition of the historical temperature sequence of each temperature measurement point in the cold storage, the door opening and closing state sequence, the cargo entry and exit records, the external environmental parameters and the time-of-use electricity price data specifically includes: Temperature data is periodically collected through temperature sensors deployed at multiple locations within the cold storage. Door opening and closing events are recorded through door status sensors. The cargo identification system records cargo attributes including type and volume, as well as inbound and outbound information. Furthermore, the time-of-use electricity price list for the coming day is obtained from the grid operator through a network interface.

10. The system according to claim 8, wherein: The prediction of the temperature change trajectory of each node in the future control time domain specifically includes: The temperature sequence of each measuring point in a historical period, the sequence of warehouse door opening and closing status, the cargo entry and exit records, and the external environmental parameters are used as the input of the spatiotemporal graph convolutional network to predict the temperature value of each node in a future control time domain at multiple discrete time steps.

Images