Energy-saving control system and method for parallel connection of single-stage compression refrigerating machine and refrigeration house system

Abstract

The energy-saving control system comprises a hardware layer and an algorithm layer, the hardware layer comprises executing mechanisms such as a compressor unit, an evaporator and a condenser, a sensor module and a PLC, and the algorithm layer comprises an upper computer algorithm module; according to the energy-saving control method, system operation data are collected through the sensor module and input into the upper computer algorithm module after being preprocessed, local agent decision making is achieved through multi-agent reinforcement learning MARL, global collaborative optimization is completed in combination with a collaborative genetic algorithm CGA, decision consistency is guaranteed through conflict detection and a coordination mechanism, and energy-saving control is achieved. And finally, the PLC controller executes the control instruction and realizes closed-loop feedback updating. The temperature stability, the energy efficiency ratio and the equipment reliability are comprehensively considered, global optimal control is achieved, energy consumption is effectively reduced, the overall operation efficiency of the refrigeration house group is improved, and the method is suitable for the fields of food, medicine, industrial cold chain storage and transportation and the like.

Description

Technical Field

[0001] This invention relates to the fields of cold storage and intelligent control technology, specifically to an energy-saving control system and method for a parallel cold storage system with single-stage compression refrigeration units. Background Technology

[0002] Current cold storage refrigeration systems mostly employ a single compressor for refrigeration, manual management, and constant control strategies to achieve automatic control. However, traditional single-compressor refrigeration is not only energy-intensive and slow-responding, but also struggles to meet the actual needs of multi-zone, differentiated temperature control. Especially in multi-cold storage systems, problems such as uneven cooling capacity distribution, frequent compressor start-ups and shutdowns, and insufficient temperature control accuracy often arise. Summary of the Invention

[0003] To overcome the problems of high energy consumption and slow response of traditional single-compressor refrigeration systems, which are difficult to adapt to the actual needs of multi-region and differentiated temperature control, especially the problems of uneven distribution of cooling capacity, frequent start-stop of compressors, and insufficient temperature control accuracy in multi-cold storage systems, this invention provides an energy-saving control system and method for a parallel cold storage system with a single-stage compressor refrigeration unit.

[0004] The technical solution adopted by the present invention to achieve the above objectives includes a hardware layer and an algorithm layer; The hardware layer includes a compressor unit, an evaporator, a condenser, a fan, a liquid supply valve, a sensor module, and a PLC controller; The algorithm layer includes a host computer algorithm module, which includes an environment perception module, a local intelligent agent decision-making module, a global optimization module, a conflict detection module, and a control execution module. The exhaust end of the compressor unit is connected to the inlet of the condenser via an exhaust pipe. The condenser condenses the high-pressure gaseous refrigerant into a high-pressure liquid. The liquid is then distributed to each liquid supply valve via a liquid distribution device connected to the condenser outlet. The outlets of the liquid supply valves are connected to the evaporator inlets of the corresponding cold storage units. The evaporators are arranged in each cold storage unit to absorb the heat load inside the cold storage unit and vaporize the liquid refrigerant. The gas outlet of the evaporator is connected in parallel with the suction main pipe to the suction end of the compressor unit, forming a complete refrigerant cycle. The sensor module collects real-time data on cold storage temperature, compressor unit suction and discharge pressure, and power consumption, and transmits it to the host computer algorithm module. The host computer algorithm module outputs control commands through Multi-Agent Reinforcement Learning (MARL) and Cooperative Genetic Algorithm (CGA), which drive the compressor contactor, fan relay, and liquid supply valve to operate via the PLC controller.

[0005] Preferably, the specific functions of each module in the host computer algorithm module are as follows: Environmental sensing module: Collects temperature, load demand and compressor operating status (suction and discharge pressure, running time, number of start and stop times) for each cold storage area, and performs noise reduction, anomaly detection and normalization on the data; Local agent decision-making module: Each cold storage is configured with an independent agent, which generates local control commands (start and stop of fans and liquid supply valves) based on MARL, and the optimization objective is determined by the reward function; Global optimization module: Based on CGA, it performs global optimization on the operating status of the compressor group and the instructions of local intelligent agents. The optimization target is determined by the fitness function, with additional constraints for temperature overrun and equipment life protection. Conflict detection module: Determines conflicts between local agent instructions and global optimization results, and feeds back the corrected instructions to the reward function and fitness function for parameter updates; Control execution module: Outputs optimized control commands to the PLC controller in digital form to achieve stable temperature and energy-saving operation of the cold storage group.

[0006] A control method for an energy-saving control system of a parallel cold storage system with single-stage compressor refrigeration units includes the following steps: S1: Environmental Sensing: Collects the temperature of each cold storage unit through sensor modules. Load requirements, compressor suction and discharge pressure \ The operating status and energy consumption data are preprocessed to form a state vector; S2: Local agent decision-making: Based on MARL, an independent agent is configured for each cold storage. The agent makes decisions on the start and stop actions of the fan and liquid supply valve according to the state vector, with the reward function as the optimization objective. S3: Global Cooperative Optimization: CGA is used to globally optimize the compressor's operating status (start-stop combination, load distribution) and local agent instructions, and global control instructions are generated with the fitness function as the target. S4: Conflict Detection and Coordination: Coordinate local decisions and global optimization results through conflict and feedback mechanisms to ensure decision consistency; S5: Control Execution and Feedback Update: The PLC controller executes global control commands to drive the compressor, fan, and liquid supply valve. After the cycle ends, the control effect is fed back to the reward function and fitness function to update the parameter weights and achieve closed-loop optimization.

[0007] Preferably, the reward function in S2 is defined as:

[0008] in, As a penalty for temperature error, , The current temperature of the cold storage. To set the target temperature; This is an estimate of the compressor power allocated proportionally to the warehouse; This is a compressor start indication function. It is set to 1 when the current action causes the compressor to switch from a stopped state to a started state, and 0 otherwise. This is a function indicating frequent device switching. It takes the value 1 when the number of switching times per unit time exceeds the threshold, and 0 otherwise. This is a temperature stability indication function; it takes the value 1 when the temperature is within the set error range, and 0 otherwise. These are the weighting coefficients.

[0009] Preferably, the The calculation methods include the following two: When reflecting the total power allocation: ,in This is the current total power of the compressor. This is the system's rated power; When reflecting the load balance among cold storage facilities; ,in This represents the average power.

[0010] Preferably, the fitness function in S3 is:

[0011] in, The normalized total energy consumption predicted for the candidate scheme; Normalize the number of candidate start / stop cycles; Accumulated for temperature violations; This is an indicator of load imbalance. For safety constraints; These are the weighting coefficients.

[0012] Preferably, the conflict and feedback mechanism described in S4 follows these rules: (1) If a local action causes the cold storage temperature to exceed the set safety boundary, a global optimization scheme will be forcibly adopted; (2) If the global solution causes the compressor to start and stop at high frequency (exceeding the maximum number of start-ups constraint) or violates the minimum switching cycle constraint of the equipment, the local action is executed first, and a penalty is imposed on the global fitness function; (3) The action results after conflict resolution are fed back to the reward function and fitness function to update the parameter weights and ensure that the system gradually converges.

[0013] Preferably, the temperature safety boundary constraints include the target temperature and the allowable fluctuation range, and the reward function applies a strong penalty when the temperature is exceeded; the equipment life protection constraints include the minimum downtime of the compressor, the maximum number of starts, and the minimum switching cycle of the fan and the liquid supply valve, and the safety protection logic is triggered when the temperature is violated and the Penalty term is included in the fitness function.

[0014] The beneficial effect of this invention is that by acquiring state data and constructing features, the operating parameters of the cold storage, such as temperature, pressure, and energy consumption, are input to the local intelligent agent. The local intelligent agent realizes independent decision-making and local temperature control of the fan and liquid supply valve based on multi-agent reinforcement learning. Then, combined with the collaborative genetic algorithm, the start-stop combination and load distribution of multiple single-stage compressors are globally optimized to achieve collaborative optimization between the local intelligent agent and the global compressor group, thereby reducing system energy consumption while ensuring accurate temperature control of multiple cold storages.

[0015] By constructing a reward function that comprehensively considers temperature stability, energy efficiency ratio, and equipment reliability, global optimal control is achieved, effectively reducing energy consumption and improving the overall operating efficiency of cold storage groups. This approach is applicable to fields such as food, pharmaceuticals, and industrial cold chain storage and transportation. Attached Figure Description

[0016] Figure 1 This is a schematic diagram of the system structure; Figure 2 For the control strategy flowchart; Figure 3 This is a diagram of the algorithm coordination framework. Detailed Implementation

[0017] The embodiments of the present invention will now be described in detail with reference to the accompanying drawings.

[0018] like Figure 1 As shown, the energy-saving control system of a single-stage compressor parallel cold storage system of the present invention includes a compressor unit, an evaporator, a condenser, a fan, a liquid supply valve, a sensor module, a PLC controller, and a host computer algorithm module. The exhaust end of the compressor unit is connected to the inlet of the condenser through an exhaust pipe. After the condenser condenses the high-pressure gaseous refrigerant into a high-pressure liquid, it supplies liquid to each liquid supply valve through a liquid distribution device connected to the condenser outlet. The outlet of each liquid supply valve is connected to the evaporator inlet of the corresponding cold storage. The corresponding evaporators are arranged in each cold storage to absorb the heat load in the cold storage and vaporize the liquid refrigerant. The gas outlet of the evaporator is connected in parallel with the suction main pipe to the suction end of the compressor unit, forming a complete refrigerant cycle of condenser-distribution-liquid supply valve-evaporator-return gas confluence-compressor.

[0019] The sensor module includes a temperature sensor, an intake and exhaust pressure sensor, and an energy meter.

[0020] Temperature sensors are installed at representative locations such as the evaporator outlet and the center of each cold storage unit, with a calibration error of <= ±0.2℃, for real-time monitoring of the temperature inside the cold storage unit.

[0021] The suction and discharge pressure sensors are installed at the suction and discharge ports of the compressor to collect evaporation pressure and condensation pressure, reflecting the refrigerant cycle conditions. The electricity meter is installed in the compressor power supply circuit and the fan branch to monitor operating energy consumption.

[0022] All devices (such as PLCs, industrial computers, etc.) are synchronized to the same clock. The sampling period for the reference temperature is set to 10s, the sampling period for the pressure cycle is set to 10s, the MARL control cycle is set to 10-30s, and the CGA scheduling cycle is set to 60-300s.

[0023] The sensor module collects real-time data on cold storage temperature, compressor unit suction and discharge pressure, and power consumption, and transmits this data to the host computer algorithm module. The host computer algorithm module, using multi-agent reinforcement learning and coordination algorithms, outputs control commands to the PLC controller. The PLC controller then drives the compressor contactor, fan relay, and liquid supply valve via digital outputs to control temperature changes. This system architecture ensures a closed-loop operation of "data acquisition—algorithm optimization—execution control—feedback correction."

[0024] like Figure 2 As shown, the specific execution process of the control strategy of the present invention is as follows: Data Acquisition: Real-time collection of temperature data from each warehouse. Door magnetic status, fan / liquid supply valve on / off status, compressor suction / discharge pressure \ Denoising and anomaly detection are performed on the original signal to remove significant outliers and label them; feature parameters are constructed, including temperature difference. Calculation, Recommendation =1℃, and triggers a penalty when the temperature deviation exceeds the stability dead zone; temperature decrease rate Compressor load index: average power Differential pressure Cumulative runtime Number of start / stop operations For each feature class, perform min-max or z-score standardization to ensure that different dimensions can be compared in a weighted manner.

[0025] The collected temperature and pressure data signals are processed in a unified manner to extract key variables representing the cooling load level, compressor operating efficiency, and condenser heat exchange status. For example, the deviation between the cold storage temperature and the set temperature reflects the refrigeration demand of the cold storage; the suction and discharge pressure difference reflects the compressor load level; and unit energy consumption reflects equipment operating efficiency. These characteristics serve as inputs to the intelligent control algorithm, providing a data foundation for subsequent optimization.

[0026] Local agent decision-making module: Each cold storage unit is equipped with an independent agent, adopting a centralized training and distributed execution (CTDE) architecture. During training, information is shared, and during runtime, decisions are made independently. Multi-agent reinforcement learning algorithms are used to dynamically control the fans and liquid supply valves. The agents select switching actions based on temperature deviation, energy consumption levels, and historical control behavior, thereby adjusting the heat exchange of the evaporator and achieving precise control of the cold storage temperature.

[0027] The state of an agent is defined as follows: , The actions are defined as follows: A0: State remains unchanged (maintain the status quo); A1: Open the liquid supply valve; A2: Turn on the fan; A3: Start the fan after a delay after the liquid supply valve is opened (valve → delay → fan); A4: Turn off the liquid supply in advance after the fan is turned on (to prevent overcooling).

[0028] The reward function for multi-agent reinforcement learning is:

[0029] in, The temperature error penalty (squared term to avoid averaging error masking excessive deviation) is mainly used to constrain the temperature control accuracy of cold storage and encourage the system to make the temperature close to the set value. This is an estimate of the compressor power allocated proportionally to the library, primarily used to encourage the system to reduce total power consumption or to allocate load appropriately. This is the compressor start-up indicator function. The function takes a value of 1 when an action causes the compressor to start, and 0 otherwise. If switching occurs too frequently in a short period of time, a penalty will be imposed. The purpose is to punish equipment such as fans and liquid supply valves that frequently start and stop. The purpose of providing a small reward for temperature within the error range to encourage steady-state operation is to encourage the system to operate stably within the "allowable fluctuation range" for a long time, rather than frequently approaching the set point. The weighting coefficients are tuned in HIL.

[0030] This reward function, through a combination of temperature accuracy (squared penalty), local power reduction penalty, start / stop / switching penalty, and dead zone stability reward, can both ensure the quality of cold storage temperature control and effectively suppress unnecessary compressor start / stop and system energy consumption under the constraints of digital actuators, thereby improving equipment lifespan.

[0031] in The calculation method is as follows:

[0032] The current temperature of the cold storage. To set the target temperature. (Among them) The larger the value >0, the more stringent the requirements for temperature accuracy.

[0033] The calculation method is as follows: When the calculation result is the total power

[0034] in This is the current total power of the compressor. This is the system's rated power.

[0035] Calculation method when the calculation result is the balance between cold storage facilities:

[0036] This represents the average power. When... The larger the system, the more it tends to operate in an energy-efficient manner.

[0037] The calculation method is as follows: When the value is 1, the compressor starts at time t; when the value is 0, it stops. The larger the value, the more sensitive it is to compressor startup.

[0038] The calculation method is as follows: The value is 1 if the number of switching times per unit time > the threshold, otherwise it is 0; The larger the value, the more stable the system.

[0039] The calculation method is as follows: The value is 1 when the temperature is within the range, and 0 otherwise; A larger value indicates that the system encourages stable operation.

[0040] Global fitness function design: The design goal is to provide start-stop combinations and operating periods for the compressor group in the future scheduling, so as to minimize system energy consumption, control start-stop, reduce temperature violations, and achieve good load balancing.

[0041]

[0042] Define the normalization of each term: The normalized total energy consumption (including start-up losses) predicted for candidate schemes. The candidate start-stop count is normalized to reflect the start-stop frequency of system equipment (especially compressors). The cumulative temperature violation data indicates the degree of temperature fluctuation or deviation from the target value in the cold storage, and is an important indicator of the system's temperature control quality. This is a load imbalance indicator, representing whether the temperature remains within the operating fluctuation range. For safety constraints; These are the weighting coefficients.

[0043] in The calculation method is as follows:

[0044] The total power of the system (including compressors, fans, etc.) is T, and the scheduling period is T.

[0045] The calculation method is as follows:

[0046] nk represents the total number of compressors, and nk represents the number of times compressor k starts and stops within a scheduling cycle.

[0047] The calculation method is as follows:

[0048] M represents the total number of cold storage facilities; The calculation method is as follows: When the statistical result is the percentage of time spent outside the dead zone:

[0049] Let i be the total time that the cold storage temperature exceeds the dead temperature, and T be the total running time.

[0050] When the statistical result is the cumulative range exceeding the dead zone:

[0051] Within the permitted range; The calculation method is as follows

[0052] The value is 1 if the constraint is violated, otherwise it is 0. The penalty weight is used to constrain j.

[0053] predict For each candidate scheme, a simplified thermodynamic model (RC model or linear approximation) is used to perform rapid simulation within the scheduling window to obtain the temperature trajectory and compressor load. Then, the power is calculated and accumulated according to the COP curve or empirical model, while also incorporating the energy consumption of each compressor start-up. .

[0054] Predicted temperature violation Simulation or empirical estimation of the cumulative time / range of each cold storage unit exceeding ±1℃ under the scheme.

[0055] The Cooperative Genetic Algorithm (CGA) uses candidate solutions from MARL as the initial population, with a population size of 40-120 and 20-80 generations (to ensure completion within the CGA cycle). It employs a tournament selection mechanism (to increase the survival probability of fitness individuals and ensure the preservation of superior genes), two-point crossover (to facilitate the exploration of the global optimum), and a mutation rate of 0.01-0.05 (to avoid random perturbations leading to the optimum). To ensure that actions meet equipment operational safety requirements, a crossover and mutation constraint repair mechanism is introduced. If the minimum running time and minimum downtime are violated, the mechanism is repaired to continuously meet the minimum running / downtime requirements.

[0056] The local cooling demand (e.g., cold storage A needs cooling while cold storage B temporarily suspends operation) is input as a constraint into the cooperative genetic algorithm module through multi-agent reinforcement learning. The cooperative genetic algorithm performs global evolutionary calculations and outputs compressor start-up / shutdown and load distribution schemes. Subsequently, the global optimization results are fed back to each local agent to guide the next control decision, thus forming a cooperative control mechanism of local and global closed-loop control.

[0057] Conflict Detection and Fusion: Used to coordinate local and global decisions. When the local control action is inconsistent with the global scheduling plan, the system makes a ruling based on the principle of safety priority and adjusts the conflicting actions to ensure that the compressor operates normally and the storage temperature is maintained within the target range. The adjustment results are fed back to the local agent and the global algorithm for the correction of the reward function and fitness function, respectively. Through iterative iteration, the system gradually reduces conflicting decisions, achieves consistency between local and global control, and improves energy efficiency and operational stability.

[0058] The optimization results are calculated by the code and sent to the PLC. The PLC's digital port is connected to the compressor contactor, fan relay and liquid supply valve to realize specific start and stop operations. At the same time, the condenser is turned on in stages according to the condensing pressure to ensure condensation effect and system stability.

[0059] Control Execution and Feedback Updates: At the end of each mechanism cycle, the system feeds back the temperature control effect, compressor energy consumption level, and equipment operating status to the reward function and fitness function, which are used to update the agent and optimize the population. Through continuous iterative training and optimization, the system can gradually achieve a balance between cooling demand and optimal energy consumption.

[0060] like Figure 3 As shown, this invention proposes a coordinated control framework for a cold storage cluster system based on Multi-Agent Reinforcement Learning (MARL) and Coordinating Genetic Algorithm (CGA). This framework comprises the following five levels.

[0061] In this embodiment of the invention, the first layer, the environment and state perception layer, monitors the operating status of the cold storage system in real time. This layer receives the operating status of actuators such as temperature and compressor load parameters for each cold storage unit, processes the data, and forms a state vector as algorithm input. This state vector is sent in real time to the second layer, the local agent decision layer, as input for each cold storage agent. The local agent decision layer calculates local actions based on the aforementioned state vector and generates real-time rewards according to the reward function. This layer maintains continuous communication between the upper and lower layers, sending local features such as local actions, immediate rewards, and loads to the third layer, the global optimization layer. Simultaneously, it receives constraints and global evaluations returned from the global optimization layer and makes necessary adjustments based on the policy output (if a local action meets safety and control requirements, it is passed as a candidate action to subsequent layers; otherwise, the policy is incorrectly corrected). The global optimization layer receives the "action set" from the local agent decision layer and decodes it into an initial population for a cooperative genetic algorithm, obtaining globally optimized actions through an evolutionary process. The global optimization layer sends optimized actions and corrections to local actions to the fourth-layer conflict detection and fusion layer. Simultaneously, it feeds back the global fitness score to the local agent decision layer, enabling it to converge more effectively in the next policy update. The global optimization layer also receives reward trends from the local agent decision layer, controlling stability indicators and device start-up / shutdown behavior to dynamically adjust the weights of the fitness function, ensuring consistency between local learning and global optimization. The conflict detection and fusion layer receives local actions from the local agent decision layer and global optimized actions from the global optimization layer, comparing them to determine if conflicts exist, including mutually exclusive device actions, local actions violating global temperature safety constraints, and global actions causing excessive start-up / shutdown of local devices. If a conflict exists, actions are fused and corrected according to a preset strategy to obtain the final executable action. The conflict detection and fusion layer transmits the action rejection reason and conflict information to the local agent decision layer to correct the reward function and feeds back the executed actions to the global optimization layer to update the fitness function. In this invention, a closed-loop collaborative control mechanism between multi-agent control and global optimization is constructed by setting up a conflict detection and fusion layer and a feedback learning layer. The conflict detection and fusion layer is used to perform consistency and feasibility analysis on the control commands generated by multiple local agents. When a conflict is detected between local control strategies in terms of compressor load distribution, cooling capacity utilization, or system operation constraints, the control commands are coordinated and fused according to the global system operation status to form a control decision that satisfies the overall system constraints. The feedback learning layer monitors the execution result of the decision in real time and sends the feedback information back to the environment and state perception layer for correcting the control parameters.Through information transmission between the two layers, this invention can ensure the temperature and control accuracy of multiple cold storage units while avoiding control conflicts caused by independent decision-making by multiple agents, achieving continuous adaptive optimization of the control strategy, thereby improving the overall operating efficiency and reliability of the system. The final fifth-layer feedback learning layer returns the actual action results, temperature trajectory, compressor start / stop status, and control effects to the upper layer for updating the local agents and dynamically correcting the global optimization, thus forming an adaptive learning mechanism.

[0062] The Multi-Agent Reinforcement Learning (MARL) algorithm operates in a cold storage area, treating each cold storage unit as an agent. Each agent independently generates local decisions based on the temperature deviation and energy consumption cost within its region. Its reward function is:

[0063] Each cold storage area corresponds to an independent agent, whose reward function is composed of multiple constraint indicators, including a local temperature deviation squared term, a power allocation penalty term, an equipment start-up and shutdown penalty term, and a stable operation reward term. The temperature deviation squared term constrains the tracking accuracy of the cold storage temperature to the set value; the power allocation term reflects the degree of reliance of a single area on the system's cooling capacity; the start-up and shutdown penalty term suppresses frequent start-up and shutdown behavior of the compressor and actuators; and the stability reward term provides positive incentives when the cold storage temperature enters and remains within the set dead zone, thereby guiding the agent to learn to prioritize stable and low-energy-consumption control strategies while meeting temperature control accuracy requirements. Through this reward structure, a single agent can achieve a trade-off optimization between temperature stability, energy consumption control, and equipment lifespan within its region.

[0064] However, due to the independent optimization of each agent, overall conflicts may occur, such as multiple cold storage units simultaneously requesting temperature reductions, leading to compressor overload. To address this issue, a cooperative genetic algorithm (CGA) is introduced at the system level. CGA uses MARL's candidate solutions as the initial population and performs global optimization through a fitness function. Its form is as follows:

[0065] Its reward function is composed of multiple constraint indicators, including a comprehensive consideration of total system energy consumption, compressor start-stop frequency, temperature violation rate, and cold load distribution balance. Among them, the total energy consumption term is consistent with the power sharing penalty in the agent's reward function in terms of objective; the temperature violation rate corresponds to the local temperature deviation constraint; the start-stop frequency term forms a cross-layer mapping relationship with the local start-stop penalty term; and the cold load balance indicator is used to characterize the coordination and stability of the operating status of each cold storage area.

[0066] During the optimization process, CGA employs selection, crossover, and mutation operations to generate a new generation of candidate solutions. These solutions are then eliminated through a fitness function, ultimately outputting the globally coordinated optimal control scheme. This scheme is then fed back to the agent layer, forming a closed-loop update mechanism. Through this collaborative framework, the system can achieve a natural transition from "local optimum to global optimum," ensuring a balance between energy consumption, temperature accuracy, and equipment lifespan.

[0067] Compressor Group Operation State Machine and Protection Logic: The operation of the compressor group is described using a finite state machine. The basic states include standby state, single-unit operation state, dual-unit operation state, and protection shutdown state. The transitions between states are determined by load demand, sensor detection results, and algorithm optimization results.

[0068] In standby mode, when the temperature of a cold storage area exceeds the set upper limit, the system assesses whether the current load is within the capacity of a single compressor. If the conditions are met, one compressor is started to enter single-unit operation mode. If multiple areas have high load demands simultaneously, causing the total load to exceed the capacity of a single compressor, the system uses CGA to determine and coordinate the start of a second compressor, entering dual-unit operation mode.

[0069] During operation, the system monitors key parameters such as compressor exhaust temperature, suction pressure, and motor current in real time. When any parameter exceeds the set threshold, the protection logic is immediately triggered, the system switches to protection shutdown state, and ensures equipment cooling and avoids damage through delayed shutdown and forced fan operation.

[0070] To reduce energy consumption and equipment wear, the state machine incorporates minimum start-stop time constraints and cumulative start-stop count constraints to avoid efficiency losses caused by frequent switching. Furthermore, during the transition from dual-unit to single-unit operation, the system prioritizes retaining compressor units with high energy efficiency ratios and stable operation, thereby optimizing energy efficiency while ensuring cooling capacity.

[0071] The combination of this state machine and intelligent algorithm not only ensures the safety and stability of the compressor group's operation, but also effectively reduces power consumption in actual operation. Through coupling with MARL and CGA, the compressor state switching decision no longer depends on a single parameter, but achieves an organic unity of "safety protection - algorithm optimization - energy-saving operation" based on global optimization results.

[0072] The above description is merely a specific embodiment of the present invention. By introducing a two-layer optimization framework combining Multi-Agent Reinforcement Learning (MARL) and Cooperative Genetic Algorithm (CGA), the present invention achieves a combination of local autonomous adjustment and global coordinated scheduling in a cold storage group control system. This overcomes the problems of uneven cold load distribution, insufficient temperature accuracy, high energy consumption, and frequent equipment start-ups and shutdowns inherent in traditional single-compressor control methods. Compared to existing technologies, the present invention significantly improves the energy efficiency, stability, and intelligence level of the cold storage group control system.

[0073] Furthermore, the control method of this invention is not limited to a parallel system of two single-stage compressors, but is also applicable to scenarios involving multiple compressors in parallel and multi-stage compression systems. The parameter weights in the fitness function and reward function can be adjusted according to different cold storage sizes, energy efficiency targets, and equipment lifespan requirements to adapt to different application needs.

[0074] In summary, the energy-saving control method for a parallel cold storage system with a single-stage compressor proposed in this invention breaks through the limitations of traditional constant control and manual management, and realizes the organic combination of "local autonomous control and global collaborative optimization" in a multi-cold storage system. It not only ensures precise temperature control of each cold storage, but also reduces the overall energy consumption of the system and extends the service life of the equipment, demonstrating significant energy-saving effects and promotional value.

Claims

1. An energy-saving control system for a parallel cold storage system using a single-stage compressor refrigeration unit, characterized in that: Includes the hardware layer and the algorithm layer; The hardware layer includes a compressor unit, an evaporator, a condenser, a fan, a liquid supply valve, a sensor module, and a PLC controller; The algorithm layer includes a host computer algorithm module, which includes an environment perception module, a local intelligent agent decision-making module, a global optimization module, a conflict detection module, and a control execution module. The exhaust end of the compressor unit is connected to the inlet of the condenser via an exhaust pipe. The condenser condenses the high-pressure gaseous refrigerant into a high-pressure liquid. The liquid is then distributed to each liquid supply valve via a liquid distribution device connected to the condenser outlet. The outlets of the liquid supply valves are connected to the evaporator inlets of the corresponding cold storage units. The evaporators are arranged in each cold storage unit to absorb the heat load inside the cold storage unit and vaporize the liquid refrigerant. The gas outlet of the evaporator is connected in parallel with the suction main pipe to the suction end of the compressor unit, forming a complete refrigerant cycle. The sensor module collects real-time data on cold storage temperature, compressor unit suction and discharge pressure, and power consumption, and transmits it to the host computer algorithm module. The host computer algorithm module outputs control commands through Multi-Agent Reinforcement Learning (MARL) and Cooperative Genetic Algorithm (CGA), which drive the compressor contactor, fan relay, and liquid supply valve to operate via the PLC controller.

2. The energy-saving control system for the parallel cold storage system with a single-stage compressor refrigeration unit according to claim 1, characterized in that: The specific functions of each module in the host computer algorithm module are as follows: Environmental sensing module: Collects temperature, load demand and compressor operating status of each cold storage area, and performs noise reduction, anomaly detection and normalization on the data; Local agent decision-making module: Each cold storage is configured with an independent agent, which generates local control instructions based on MARL, and the optimization objective is determined by the reward function; Global optimization module: Based on CGA, it performs global optimization on the operating status of the compressor group and the instructions of local intelligent agents. The optimization target is determined by the fitness function, with additional constraints for temperature overrun and equipment life protection. Conflict detection module: Determines conflicts between local agent instructions and global optimization results, and feeds back the corrected instructions to the reward function and fitness function for parameter updates; Control execution module: Outputs optimized control commands to the PLC controller in digital form to achieve stable temperature and energy-saving operation of the cold storage group.

3. A control method for an energy-saving control system based on the single-stage compressor parallel cold storage system as described in claim 2, characterized in that: Includes the following steps: S1: Environmental Sensing: Collects the temperature of each cold storage unit through sensor modules. Load requirements, compressor suction and discharge pressure \ The operating status and energy consumption data are preprocessed to form a state vector; S2: Local agent decision-making: Based on MARL, an independent agent is configured for each cold storage. The agent makes decisions on the start and stop actions of the fan and liquid supply valve according to the state vector, with the reward function as the optimization objective. S3: Global Cooperative Optimization: CGA is used to perform global optimization on the compressor operating status and local agent instructions, and global control instructions are generated with the fitness function as the target. S4: Conflict Detection and Coordination: Coordinate local decisions and global optimization results through conflict and feedback mechanisms to ensure decision consistency; S5: Control Execution and Feedback Update: The PLC controller executes global control commands to drive the compressor, fan, and liquid supply valve. After the cycle ends, the control effect is fed back to the reward function and fitness function to update the parameter weights and achieve closed-loop optimization.

4. The method according to claim 3, characterized in that: The reward function described in S2 is defined as follows: in, As a penalty for temperature error, , The current temperature of the cold storage. To set the target temperature; This is an estimate of the compressor power allocated proportionally to the warehouse; This is a compressor start indication function. It is set to 1 when the current action causes the compressor to switch from a stopped state to a started state, and 0 otherwise. This is a function indicating frequent device switching. It takes the value 1 when the number of switching times per unit time exceeds the threshold, and 0 otherwise. This is a temperature stability indication function; it takes the value 1 when the temperature is within the set error range, and 0 otherwise. These are the weighting coefficients.

5. The method according to claim 4, characterized in that: The The calculation methods include the following two: When reflecting the total power allocation: ,in This is the current total power of the compressor. This is the system's rated power; When reflecting the load balance among cold storage facilities: ,in This represents the average power.

6. The method according to claim 3, characterized in that: The fitness function described in S3 is: in, The normalized total energy consumption predicted for the candidate scheme; Normalize the number of candidate start / stop cycles; Accumulated for temperature violations; This is an indicator of load imbalance. For safety constraints; These are the weighting coefficients.

7. The method according to claim 3, characterized in that: The conflict and feedback mechanism described in S4 follows these rules: (1) If a local action causes the cold storage temperature to exceed the set safety boundary, a global optimization scheme will be forcibly adopted; (2) If the global scheme causes the compressor to start and stop at high frequency or violates the minimum switching cycle constraint of the equipment, the local action is executed first, and a penalty is imposed on the global fitness function; (3) The action results after conflict resolution are fed back to the reward function and fitness function to update the parameter weights and ensure that the system gradually converges.

8. The method according to claim 7, characterized in that: The temperature safety boundary constraints include the target temperature and the allowable fluctuation range. When these are exceeded, the reward function applies a strong penalty. The equipment life protection constraints include the minimum downtime of the compressor, the maximum number of starts, and the minimum switching cycle of the fan and liquid supply valve. When these are violated, the safety protection logic is triggered and the Penalty term is included in the fitness function.

9. The method according to claim 3, characterized in that: The temperature deviation term in the reward function described in S2 is expressed by the formula... Computational Recommendation =1℃, and a penalty is triggered when the temperature deviation exceeds the stability dead zone. This represents the measured temperature of the i-th cold storage unit at the current moment, which is the temperature data collected by the sensor in real time. This represents the preset target temperature for the i-th cold storage unit, serving as a reference point for temperature control. The reference scale for temperature normalization is a fixed value. This represents the temperature deviation of the i-th cold storage unit, used to reflect temperature control errors. This represents the temperature error, used in the state vector of MARL.

Images