Efficient refrigerating machine room energy-saving automatic control system based on reinforcement learning

Abstract

The invention discloses an efficient refrigerating machine room energy-saving automatic control system based on reinforcement learning, and the system comprises the following modules: an operation data collection module which is used for collecting key equipment operation parameters and generating unified time sequence data; the energy consumption modeling module constructs energy consumption mathematical models of the water chilling unit, a water pump and a cooling tower and stores the energy consumption mathematical models into a database; the cooling capacity prediction module is used for predicting the future cooling capacity demand based on the historical data and the environmental parameters; the feed-forward optimization module is used for optimizing power distribution by using an adaptive particle swarm algorithm and generating control parameters; the state correction module automatically adjusts the running state according to the running efficiency deviation condition; the cooling strategy optimization module introduces an improved DSAC algorithm to optimize a cooling control strategy; and the instruction deployment module deploys the strategy to an actual system and outputs a regulation and control instruction. The energy efficiency management level of the refrigerating machine room is improved, and integration of energy conservation and intelligent control is achieved.

Description

Technical Field

[0001] This invention relates to the field of intelligent control technology, and in particular to a high-efficiency energy-saving automatic control system for refrigeration rooms based on reinforcement learning. Background Technology

[0002] This invention relates to a high-efficiency energy-saving automatic control system for refrigeration computer rooms based on reinforcement learning, which has wide applications in the refrigeration industry. However, current refrigeration system control technology still has some problems and shortcomings. Existing refrigeration systems mostly rely on traditional control algorithms, lacking the ability to dynamically adjust and intelligently optimize based on real-time data. This results in insufficiently precise energy consumption management and an inability to automatically adjust system parameters according to load fluctuations and environmental changes.

[0003] Current optimization methods for most refrigeration systems rely on fixed control strategies, making it difficult to fully utilize multi-source data, such as chiller load, cooling water pump power, cooling tower fan frequency, and ambient wet-bulb temperature. This prevents energy efficiency optimization from comprehensively considering the influence of multiple factors. Traditional systems often neglect the interactions between equipment and the combined impact of each device's operation, resulting in the failure to maximize energy efficiency on the cooling side.

[0004] Furthermore, the control strategies of existing systems are mostly based on experience and rules, lacking sufficient adaptive capabilities. These strategies respond slowly to system load fluctuations, making it difficult to adjust system operating parameters in real time and cope with dynamically changing cooling demands and environmental conditions. Therefore, the systems are not only energy inefficient but also lag behind in response, leading to unnecessary energy consumption and a significant waste of electrical resources.

[0005] Furthermore, traditional energy-saving algorithms cannot flexibly respond to changes in different working environments and load conditions in practical applications, making it difficult to achieve personalized and intelligent control. They also lack the ability to learn and optimize for specific scenarios and cannot continuously adapt to different needs during system operation.

[0006] Therefore, how to provide a high-efficiency energy-saving automatic control system for refrigeration computer rooms based on reinforcement learning is a problem that urgently needs to be solved by those skilled in the art. Summary of the Invention

[0007] One objective of this invention is to propose a high-efficiency energy-saving automatic control system for refrigeration rooms based on reinforcement learning. This invention fully integrates key technologies such as adaptive particle swarm optimization (PSO) and an improved DSAC algorithm, constructing a core process encompassing multi-dimensional parameter acquisition, energy consumption mathematical modeling, dynamic optimization decision-making, energy efficiency feedback mechanisms, and adaptive strategy updates. This enables intelligent control and dynamic optimization of the cooling system under varying load conditions. Through the accurate calculation of cooling demand prediction and energy allocation optimization using the adaptive PSO algorithm, this invention effectively improves the energy efficiency management of chillers, chilled water pumps, cooling water pumps, and cooling towers. Simultaneously, combined with the improved DSAC algorithm, based on the cooling-side COP maximization objective, reinforcement learning is used to automatically adjust the control strategy and update the optimal control scheme in real time, thereby significantly improving the system's energy efficiency, response speed, and energy-saving effect. This method has advantages such as intelligent decision-making, adaptive optimization, and high energy efficiency, effectively solving the problems of lagging energy efficiency management, untimely response, and insufficient optimization in existing refrigeration systems.

[0008] According to an embodiment of the present invention, a high-efficiency energy-saving automatic control system for a cooling room based on reinforcement learning includes the following steps:

[0009] Run the data acquisition module to collect the operating parameters of each key device in the cooling room and form a raw dataset of the room's operating status under a unified timestamp;

[0010] The energy consumption mathematical modeling module establishes chiller unit performance curves and pump power characteristics based on the original dataset of the computer room's operating status, constructs an energy consumption mathematical model library, and stores it in the associated database;

[0011] The cooling demand forecasting module, based on historical terminal operation data and energy consumption models, combined with ambient wet-bulb temperature, establishes a terminal cooling load forecasting model and outputs the predicted cooling demand value for the next time period.

[0012] The feedforward control optimization module inputs the predicted cooling demand into the energy consumption mathematical model, calls the adaptive particle swarm optimization algorithm to calculate the optimal real-time power allocation scheme, generates feedforward control parameters, outputs the real-time power ratio result, and updates the parameter set in the associated database.

[0013] The operating status correction module establishes an efficiency deviation detection mechanism based on the updated parameter set of the associated database and the performance curve of the chiller unit. When the operating point deviates from the high efficiency range, it automatically adjusts the water temperature and flow rate or performs a change in the number of units and outputs the corrected setting status.

[0014] The cooling-side strategy optimization module uses the chiller load, cooling water pump power, cooling tower fan frequency, and ambient wet-bulb temperature as state vectors, inputs them into the improved DSAC algorithm, and continuously updates the optimal control strategy set for the number and frequency of cooling water pumps and cooling towers through interactive learning.

[0015] The command deployment and control module deploys the optimal control strategy set and updated database parameters on the cooling side to the actual control system and coordinates the output to the chilled water pump, cooling water pump and cooling tower.

[0016] Optionally, modules can be integrated using the following methods:

[0017] S1. Collect the operating data of the chiller room, calculate the current chiller unit's cooling power, define the chiller unit load by the ratio of the chiller unit to the rated cooling capacity, record it under a unified timestamp, and form the original dataset of the chiller room's operating status.

[0018] S2. Based on the original dataset of the computer room operation status, establish an energy consumption mathematical model library, perform mathematical modeling on the performance curves of chiller units, power characteristics of pumps and cooling towers, form an energy consumption mathematical model, and store it in the associated database.

[0019] S3. Based on the energy consumption mathematical model and related database, terminal operation history data and ambient wet-bulb temperature, model and simulate the terminal load data, predict the cooling demand for the next period, and output the predicted cooling demand value.

[0020] S4. Input the predicted cooling demand into the energy consumption mathematical model, and use the adaptive particle swarm optimization algorithm to calculate the real-time optimal power allocation relationship under different cooling loads, generate feedforward control parameters, update the associated database, output the real-time optimal power ratio result and update the parameter set of the associated database.

[0021] S5. Establish the optimal efficiency curve model based on the performance curve of the chiller unit and the parameter set of the updated associated database. When the operating point deviates from the high-efficiency zone, automatically adjust the temperature and flow of chilled water and cooling water or perform the addition or reduction of units. Output the corrected water temperature and number of units set and the operating status close to the optimal efficiency curve.

[0022] S6. Based on the corrected operating status, an improved DSAC algorithm is introduced on the cooling side. The optimal number of cooling water pumps and frequency control strategies are selected using the chiller load, cooling water pump power, cooling tower fan frequency and ambient wet-bulb temperature as state variables. The optimization objective is to maximize the COP on the cooling side. The algorithm repeatedly interacts with the environment to learn and update the optimal strategy set on the cooling side.

[0023] S7. Deploy the optimal strategy set for the cooling side and the parameter set of the updated relational database to the actual operating environment, and coordinate and output frequency adjustment commands to the chilled water pump, cooling water pump and cooling tower.

[0024] Optionally, S1 specifically includes:

[0025] S11. Install temperature sensors, pressure sensors and power acquisition devices at various measuring points in the chiller room. Synchronously trigger the data acquisition devices through integrated control signals to read the inlet water temperature, outlet water temperature and cooling tower outlet water temperature of the chiller unit respectively.

[0026] S12. Read the operating frequency signals of the chilled water pump and cooling water pump, combine them with the instantaneous chilled water flow rate output by the flow sensor, record the chilled water flow rate data, collect the power consumption data in the computer room power distribution cabinet and the wet-bulb temperature parameters output by the outdoor environment wet-bulb temperature sensor, and stamp all kinds of data with a unified timestamp under the same clock.

[0027] S13. Calculate the current chiller unit's cooling power based on the chilled water flow rate and the temperature difference between the inlet and outlet water of the chiller unit. Divide the calculation result by the unit's rated cooling capacity to obtain the chiller unit load. Summarize and store the temperature, frequency, electrical energy, wet-bulb temperature, and calculated chiller unit load in the order of timestamps to generate the original dataset of the computer room's operating status.

[0028] Optionally, S2 specifically includes:

[0029] S21. Read the original dataset of the computer room operation status, pair the chiller unit inlet and outlet water temperature, chilled water flow rate and power consumption as samples according to the timestamp, label the current chiller unit load and the corresponding ambient wet-bulb temperature, remove missing samples and samples that exceed the preset threshold, and obtain the sample sequence.

[0030] S22. Divide the sample sequence into low load, medium load and high load ranges according to a preset threshold, fit the chiller performance curves using the least squares method, calculate the residual value of each curve, and if the error exceeds the threshold, adjust the parameters according to the preset ratio until the error meets the condition to generate the chiller performance curve.

[0031] S23. Read the frequency, flow rate, and power consumption of the chilled water pump and the cooling water pump, pair them according to timestamp to form a pump sample sequence, group them according to frequency interval to fit the power characteristics, and output the power characteristic parameters.

[0032] S24. Read the cooling tower fan frequency, outlet water temperature, ambient wet-bulb temperature and power consumption, pair them according to timestamp to form a cooling tower sample sequence, group and fit the power characteristics and heat transfer response according to the wet-bulb temperature and fan frequency range, and output the power characteristic parameters.

[0033] S25. Combine the performance curves of the chiller unit, the power characteristics of the water pump, and the power characteristics of the cooling tower. The energy consumption output is the sum of the power consumption of the chiller unit, the power consumption of the water pump, and the power consumption of the cooling tower. Record the input and output fields and constraint boundaries to form an energy consumption mathematical model.

[0034] S26. Use the reserved sample to verify the energy consumption mathematical model, calculate the difference between the output and the sample power, count the error, adjust the parameters until the error falls within the threshold, generate version numbers and timestamps for the performance curves, water pump power characteristics and cooling tower power characteristics, write the parameters, intervals, error thresholds and version numbers into the associated database, execute the energy consumption mathematical model to be stored in the database and create an index.

[0035] Optionally, S3 specifically includes:

[0036] S31. Based on the energy consumption mathematical model, terminal operation history data and ambient wet-bulb temperature data in the associated database, extract parameters such as chiller load, chilled water flow rate, cooling tower outlet water temperature, operating frequency of chilled water pump and cooling water pump, power consumption and ambient wet-bulb temperature, and pair them into samples according to timestamps.

[0037] S32. Use the LSTM model to train the prepared historical sample data to learn the time-series relationship between chiller unit load, chilled water flow rate, cooling tower outlet water temperature, chilled water pump and cooling water pump operating frequency, power consumption, ambient wet-bulb temperature and terminal load.

[0038] S33. Input the current chiller unit load, chilled water flow rate, cooling tower outlet water temperature, chilled water pump and cooling water pump operating frequency, power consumption, and ambient wet-bulb temperature into the trained LSTM model, and output the predicted cooling demand value.

[0039] Optionally, S4 specifically includes:

[0040] S41. Input the predicted cooling demand along with the current chiller load, chilled water flow rate, cooling tower outlet water temperature and ambient wet-bulb temperature into the energy consumption mathematical model. Read the chiller performance curve set, water pump power characteristic parameters and cooling tower power characteristic parameters, and set the time step and rolling window length for the optimization calculation.

[0041] S42. Define the particle swarm encoding method, and use the number of chiller units, chilled water outlet temperature, chilled water pump operating frequency, cooling water pump operating frequency and cooling tower fan operating frequency as decision variables, and set upper limit, lower limit and rate of change constraints for each variable.

[0042] S43. Set the optimization objective as minimum energy consumption, call the energy consumption mathematical model to calculate the energy consumption of chiller, water pump and cooling tower, and multiply the values ​​of cooling demand deviation, efficiency curve deviation, temperature exceeding the limit and frequency fluctuation exceeding the preset range by the preset corresponding penalty factor to obtain the penalty item. Multiply each penalty item by the corresponding preset weight and accumulate them, and add them to the basic energy consumption objective function. That is, energy consumption plus penalty item form a comprehensive evaluation index value.

[0043] S44. Initialize the particle swarm by generating an initial particle set using a combination of historical best solutions, rule baseline schemes and random samples, ensuring that at least one particle meets the predicted cooling demand and energy consumption constraints.

[0044] S45. Filter out past control records that are the same as the current ambient wet-bulb temperature and time period from historical operation data, extract the power allocation scheme with the lowest energy consumption as the historical optimal particle, and use the power of chiller, water pump and cooling tower as the first component of the initial solution according to the proportion, calculate the corresponding power combination, and include it as the second component in the particle set.

[0045] S46. Using a uniform random sampling method, generate several different power combination particles in the particle set so that all particle combinations meet the predicted cooling demand value. Perform energy consumption model calculation on all particles, remove particles that do not meet the boundary conditions of maximum load of chiller unit, upper limit of total power, and cooling tower outlet temperature limit, and select the optimal particle with the lowest energy consumption and satisfied constraints as the scheme as the current global optimal solution.

[0046] S47. Substitute the current particle swarm power allocation result into the energy consumption mathematical model, calculate the sum of the power consumption of the chiller, chilled water pump, cooling water pump and cooling tower fan, form an energy consumption evaluation list, compare the current energy consumption of each optimal particle with the historical optimal energy consumption value, update the individual optimal and global optimal positions, and record the optimal power allocation parameters.

[0047] S48. Calculate the difference between the optimal energy consumption of the current round and the previous round. If it is lower than the set threshold, stop the iteration. Otherwise, adjust the inertia weight and learning factor according to the preset ratio, update the particle velocity and position, obtain the new power allocation parameters, repeat the optimization until convergence, output the optimal power allocation result, and write it as the feedforward control parameter to update the parameter set of the associated database.

[0048] Optionally, the initialization of the particle swarm specifically includes:

[0049] Read the historical optimal power allocation solution from the previous optimization process from the associated database, including the frequency parameters, flow values ​​and load allocation ratios of chiller units, chilled water pumps, cooling water pumps and cooling towers, and construct the historical optimal particle as one of the initial particles.

[0050] Set a baseline rule control strategy, set the chiller unit frequency to a constant reference value, load the chilled water pump and cooling water pump proportionally, adjust the cooling tower frequency inversely proportional to the ambient wet-bulb temperature, generate rule control strategy particles, and add them to the initial particle set;

[0051] Based on the predicted cooling demand, reasonable upper and lower frequency boundaries and power allocation constraints are preset. Multiple candidate particles are randomly sampled and generated in the parameter space. The corresponding basic energy consumption objective function value is calculated for each particle, and it is determined whether the cooling demand coverage constraint and energy consumption threshold limit are met.

[0052] Particles that do not meet the cooling requirements or energy consumption constraints are removed, while those that meet the conditions are retained, with priority given to retaining the optimal basic energy consumption particles with the lowest energy consumption.

[0053] Particles that meet the conditions are combined into an initial particle swarm, and a velocity vector is initialized for each particle. Particle swarm parameters with inertia weight and learning factor are set to complete the initial population preparation.

[0054] The particle with the smallest basic energy consumption objective function value in the current particle set is marked as the initial global optimal particle, and the power allocation scheme and corresponding objective function value are recorded.

[0055] Optionally, S5 specifically includes:

[0056] S51. Match the performance curve of the chiller unit with the parameter set of the updated associated database, read the chiller unit load, power consumption, inlet and outlet water temperature and ambient wet-bulb temperature under the current operating status, and generate the current operating point according to the timestamp.

[0057] S52. Locate the current operating point in the performance curve model, determine whether the point is within the optimal efficiency range, if it deviates, record the direction and magnitude of the deviation under the current load, generate the deviation correction target value, adjust the chilled water temperature setpoint and chilled water pump operating frequency according to the deviation correction target value, so that the chilled water outlet temperature and flow rate return to the high efficiency range, and update the current operating point.

[0058] S53. Synchronously adjust the cooling water inlet temperature and cooling water pump operating frequency. When the cooling side temperature difference is insufficient or the ambient wet bulb temperature is high, increase the cooling fan speed or start the cooling tower in advance. If the chilled water and cooling water parameters still deviate from the high efficiency range, then according to the current load and cooling demand forecast, perform the preset number of chiller units to add or remove units and adjust the current number of units.

[0059] S54. Record the operating status of the chiller unit after each adjustment, update the number of units, water temperature, flow rate and energy consumption records in the associated database, and output the operating status that is close to the optimal efficiency curve.

[0060] Optionally, S6 specifically includes:

[0061] S61. Read the chiller unit load, cooling water pump power, cooling tower fan frequency and ambient wet bulb temperature under the corrected operating state, combine them into a state vector according to the current time, and use it as the state input of the improved DSAC algorithm. Combine the number of cooling water pumps, operating frequency and the number and frequency of cooling tower fans into an action space. Select a set of control strategies from the action space as the current action vector for each decision.

[0062] S62. Based on the combination of the current state vector and the candidate action vector, calculate the corresponding cooling-side COP value in the energy consumption mathematical model, feed it back as an immediate reward value to the evaluation network of the improved DSAC algorithm, and store the state-action-Q value triplet in the experience replay pool.

[0063] S63. Using historical execution records, select the control commands for cooling water pump frequency and cooling tower number that were executed at the previous moment from the experience replay pool, encode them into action vectors, and concatenate the state vector and action vector to form a state-action pair.

[0064] S64. Input the state-action pair into two structurally independent Q-value estimation networks, use a multilayer perceptron to perform a weighted summation with preset weights and add a bias term, activate each layer using the ReLU function, and output the energy consumption distribution of the state-action pair under each Q network.

[0065] S65. Perform KL divergence evaluation on the energy consumption distributions output by the two Q networks respectively, analyze the consistency between the distributions and compare the error with the actual observed energy consumption feedback, evaluate the prediction error through the mean squared error loss function, and when it exceeds the preset threshold, use the backpropagation algorithm to calculate the gradient of the mean squared error loss function with respect to the Q network parameters, and use the calculated gradient and the preset learning rate to adjust the weights and biases of the Q network.

[0066] S66. Input the current state vector into the improved DSAC algorithm, output the continuous action space sampling distribution of the combination of cooling water pump frequency and cooling tower number in the current state, and use the reparameter technique to sample the action space to generate multiple control action candidates.

[0067] S67. For each action in the candidate action set, input two updated Q networks to estimate the energy consumption distribution, read the distribution mean and take the minimum value as the Q value score index, select the action with the largest Q value as the current control strategy, execute the current control strategy to adjust the cooling water pump frequency and the number of cooling towers, and write the current state, final action, actual energy consumption result and next state into the experience pool. At the same time, write the currently executed optimal cooling side strategy into the strategy table, update the optimal cooling side strategy set and output the current optimal cooling side strategy.

[0068] Optionally, S7 specifically includes: deploying the optimal control strategy set for the cooling side obtained through the improved DSAC algorithm, and the parameter set of the associated updated database, to the actual operating environment; coordinating the operating status of each device based on real-time monitored operating data; and outputting frequency adjustment commands to the chilled water pump, cooling water pump, and cooling tower.

[0069] The beneficial effects of this invention are:

[0070] This invention proposes a high-efficiency energy-saving automatic control system for chiller rooms based on reinforcement learning. Addressing issues such as insufficient energy efficiency optimization, lag in response to load fluctuations, fixed control strategies, and imprecise energy management in existing chiller system control, it innovatively introduces an adaptive particle swarm optimization algorithm and an improved DSAC algorithm to achieve intelligent optimization and dynamic adaptive adjustment of cooling-side energy efficiency. Through the adaptive particle swarm optimization algorithm, the system can accurately adjust equipment power allocation dynamically based on real-time cooling demand prediction and energy efficiency models, thereby effectively improving the collaborative working efficiency of chillers, chilled water pumps, cooling water pumps, and cooling towers, and minimizing energy consumption. Simultaneously, the improved DSAC algorithm optimizes the control strategy through reinforcement learning, enabling the system to continuously adjust control decisions based on actual operating conditions, thereby improving the system's adaptability and energy efficiency management accuracy in changing environments. Attached Figure Description

[0071] The accompanying drawings are provided to further illustrate the invention and form part of the specification. They are used in conjunction with embodiments of the invention to explain the invention and do not constitute a limitation thereof. In the drawings:

[0072] Figure 1 This is a structural diagram of a high-efficiency energy-saving automatic control system for refrigeration computer rooms based on reinforcement learning, as proposed in this invention.

[0073] Figure 2 This is a flowchart of the cooling demand prediction and energy efficiency optimization based on the adaptive particle swarm optimization algorithm proposed in this invention.

[0074] Figure 3 This is a flowchart of the cooling-side optimal strategy generation and control optimization based on the improved DSAC algorithm proposed in this invention. Detailed Implementation

[0075] The present invention will now be described in further detail with reference to the accompanying drawings. These drawings are simplified schematic diagrams, illustrating only the basic structure of the invention, and therefore only show the components relevant to the invention.

[0076] refer to Figure 1-3 A high-efficiency energy-saving automatic control system for refrigeration computer rooms based on reinforcement learning includes the following steps:

[0077] Run the data acquisition module to collect the operating parameters of each key device in the cooling room and form a raw dataset of the room's operating status under a unified timestamp;

[0078] The energy consumption mathematical modeling module establishes chiller unit performance curves and pump power characteristics based on the original dataset of the computer room's operating status, constructs an energy consumption mathematical model library, and stores it in the associated database;

[0079] The cooling demand forecasting module, based on historical terminal operation data and energy consumption models, combined with ambient wet-bulb temperature, establishes a terminal cooling load forecasting model and outputs the predicted cooling demand value for the next time period.

[0080] The feedforward control optimization module inputs the predicted cooling demand into the energy consumption mathematical model, calls the adaptive particle swarm optimization algorithm to calculate the optimal real-time power allocation scheme, generates feedforward control parameters, outputs the real-time power ratio result, and updates the parameter set in the associated database.

[0081] The operating status correction module establishes an efficiency deviation detection mechanism based on the updated parameter set of the associated database and the performance curve of the chiller unit. When the operating point deviates from the high efficiency range, it automatically adjusts the water temperature and flow rate or performs a change in the number of units and outputs the corrected setting status.

[0082] The cooling-side strategy optimization module uses the chiller load, cooling water pump power, cooling tower fan frequency, and ambient wet-bulb temperature as state vectors, inputs them into the improved DSAC algorithm, and continuously updates the optimal control strategy set for the number and frequency of cooling water pumps and cooling towers through interactive learning.

[0083] The command deployment and control module deploys the optimal control strategy set and updated database parameters on the cooling side to the actual control system and coordinates the output to the chilled water pump, cooling water pump and cooling tower.

[0084] In this embodiment, the modules are interconnected using the following method:

[0085] S1. Collect operating data of the chiller room, including parameters such as chiller inlet and outlet water temperature, cooling tower outlet water temperature, operating frequency of chilled water pump and cooling water pump, power consumption, and ambient wet-bulb temperature. Calculate the current chiller cooling power based on chilled water flow rate and chiller inlet and outlet water temperature difference, and define the chiller load by the ratio of chilled water flow rate to rated cooling capacity. Record the data under a unified timestamp to form the original dataset of the chiller room operating status.

[0086] S2. Based on the original dataset of the computer room operation status, establish an energy consumption mathematical model library, perform mathematical modeling on the performance curves of chiller units, power characteristics of pumps and cooling towers, form an energy consumption mathematical model, and store it in the associated database.

[0087] S3. Based on the energy consumption mathematical model and related database, terminal operation history data and ambient wet-bulb temperature, model and simulate the terminal load data, predict the cooling demand for the next period, and output the predicted cooling demand value.

[0088] S4. Input the predicted cooling demand into the energy consumption mathematical model, and use the adaptive particle swarm optimization algorithm to calculate the real-time optimal power allocation relationship under different cooling loads, generate feedforward control parameters, update the associated database, output the real-time optimal power ratio result and update the parameter set of the associated database.

[0089] S5. Establish the optimal efficiency curve model based on the performance curve of the chiller unit and the parameter set of the updated associated database. When the operating point deviates from the high-efficiency zone, automatically adjust the temperature and flow of chilled water and cooling water or perform the addition or reduction of units. Output the corrected water temperature and number of units set and the operating status close to the optimal efficiency curve.

[0090] S6. Based on the corrected operating status, an improved DSAC algorithm is introduced on the cooling side. The optimal number of cooling water pumps and frequency control strategies are selected using the chiller load, cooling water pump power, cooling tower fan frequency and ambient wet-bulb temperature as state variables. The optimization objective is to maximize the COP on the cooling side. The algorithm repeatedly interacts with the environment to learn and update the optimal strategy set on the cooling side.

[0091] S7. Deploy the optimal strategy set for the cooling side and the parameter set of the updated relational database to the actual operating environment, and coordinate and output frequency adjustment commands to the chilled water pump, cooling water pump and cooling tower.

[0092] This implementation significantly improves the energy efficiency and intelligence level of chiller rooms. By collecting operational data from key equipment, a unified energy consumption mathematical model is constructed to accurately characterize cooling power and system load. An adaptive particle swarm optimization algorithm is introduced, combined with predicted cooling demand, to achieve a feedforward control strategy for optimal power allocation, improving the system's responsiveness under dynamic loads. Furthermore, a dynamic adjustment mechanism for water temperature and flow rate ensures that the operating point always aligns with the optimal efficiency zone. On the cooling side, an improved DSAC algorithm is employed to automatically learn and optimize cooling strategies under complex operating conditions, aiming to maximize COP and achieve intelligent scheduling of cooling equipment. Finally, the optimization results are deployed to the actual system through a control module, coordinating and controlling chilled water pumps, cooling water pumps, and cooling towers. This invention has significant advantages in improving overall operating efficiency, reducing energy consumption, and enhancing control accuracy, making it suitable for energy-saving retrofits and intelligent operation and maintenance scenarios in large chiller rooms.

[0093] In this embodiment, S1 specifically includes:

[0094] S11. Install temperature sensors, pressure sensors and power acquisition devices at various measuring points in the chiller room. Synchronously trigger the data acquisition devices through integrated control signals to read the inlet water temperature, outlet water temperature and cooling tower outlet water temperature of the chiller unit respectively.

[0095] S12. Read the operating frequency signals of the chilled water pump and cooling water pump, combine them with the instantaneous chilled water flow rate output by the flow sensor, record the chilled water flow rate data, collect the power consumption data in the computer room power distribution cabinet and the wet-bulb temperature parameters output by the outdoor environment wet-bulb temperature sensor, and stamp all kinds of data with a unified timestamp under the same clock.

[0096] S13. Calculate the current chiller unit's cooling power based on the chilled water flow rate and the temperature difference between the inlet and outlet water of the chiller unit. Divide the calculation result by the unit's rated cooling capacity to obtain the chiller unit load. Summarize and store the temperature, frequency, electrical energy, wet-bulb temperature, and calculated chiller unit load in the order of timestamps to generate the original dataset of the computer room's operating status.

[0097] In this embodiment, S2 specifically includes:

[0098] S21. Read the original dataset of the computer room operation status, pair the chiller unit inlet and outlet water temperature, chilled water flow rate and power consumption as samples according to the timestamp, label the current chiller unit load and the corresponding ambient wet-bulb temperature, remove missing samples and samples that exceed the preset threshold, and obtain the sample sequence.

[0099] S22. Divide the sample sequence into low load, medium load and high load ranges according to a preset threshold, fit the chiller performance curves using the least squares method, calculate the residual value of each curve, and if the error exceeds the threshold, adjust the parameters according to the preset ratio until the error meets the condition to generate the chiller performance curve.

[0100] S23. Read the frequency, flow rate, and power consumption of the chilled water pump and the cooling water pump, pair them according to timestamp to form a pump sample sequence, group them according to frequency interval to fit the power characteristics, and output the power characteristic parameters.

[0101] S24. Read the cooling tower fan frequency, outlet water temperature, ambient wet-bulb temperature and power consumption, pair them according to timestamp to form a cooling tower sample sequence, group and fit the power characteristics and heat transfer response according to the wet-bulb temperature and fan frequency range, and output the power characteristic parameters.

[0102] S25. Combine the performance curves of the chiller unit, the power characteristics of the water pump, and the power characteristics of the cooling tower. The energy consumption output is the sum of the power consumption of the chiller unit, the power consumption of the water pump, and the power consumption of the cooling tower. Record the input and output fields and constraint boundaries to form an energy consumption mathematical model.

[0103] S26. Use the reserved sample to verify the energy consumption mathematical model, calculate the difference between the output and the sample power, count the error, adjust the parameters until the error falls within the threshold, generate version numbers and timestamps for the performance curves, water pump power characteristics and cooling tower power characteristics, write the parameters, intervals, error thresholds and version numbers into the associated database, execute the energy consumption mathematical model to be stored in the database and create an index.

[0104] In this embodiment, S3 specifically includes:

[0105] S31. Based on the energy consumption mathematical model, terminal operation history data and ambient wet-bulb temperature data in the associated database, extract parameters such as chiller load, chilled water flow rate, cooling tower outlet water temperature, operating frequency of chilled water pump and cooling water pump, power consumption and ambient wet-bulb temperature, and pair them into samples according to timestamps.

[0106] S32. Use the LSTM model to train the prepared historical sample data to learn the time-series relationship between chiller unit load, chilled water flow rate, cooling tower outlet water temperature, chilled water pump and cooling water pump operating frequency, power consumption, ambient wet-bulb temperature and terminal load.

[0107] S33. Input the current chiller unit load, chilled water flow rate, cooling tower outlet water temperature, chilled water pump and cooling water pump operating frequency, power consumption, and ambient wet-bulb temperature into the trained LSTM model, and output the predicted cooling demand value.

[0108] In this embodiment, S4 specifically includes:

[0109] S41. Input the predicted cooling demand along with the current chiller load, chilled water flow rate, cooling tower outlet water temperature and ambient wet-bulb temperature into the energy consumption mathematical model. Read the chiller performance curve set, water pump power characteristic parameters and cooling tower power characteristic parameters, and set the time step and rolling window length for the optimization calculation.

[0110] S42. Define the particle swarm encoding method, and use the number of chiller units, chilled water outlet temperature, chilled water pump operating frequency, cooling water pump operating frequency and cooling tower fan operating frequency as decision variables, and set upper limit, lower limit and rate of change constraints for each variable.

[0111] S43. Set the optimization objective as minimum energy consumption, call the energy consumption mathematical model to calculate the energy consumption of chiller, water pump and cooling tower, and multiply the values ​​of cooling demand deviation, efficiency curve deviation, temperature exceeding the limit and frequency fluctuation exceeding the preset range by the preset corresponding penalty factor to obtain the penalty item. Multiply each penalty item by the corresponding preset weight and accumulate them, and add them to the basic energy consumption objective function. That is, energy consumption plus penalty item form a comprehensive evaluation index value.

[0112] S44. Initialize the particle swarm by generating an initial particle set using a combination of historical best solutions, rule baseline schemes and random samples, ensuring that at least one particle meets the predicted cooling demand and energy consumption constraints.

[0113] S45. Filter out past control records that are the same as the current ambient wet-bulb temperature and time period from historical operation data, extract the power allocation scheme with the lowest energy consumption as the historical optimal particle, and use the power of chiller, water pump and cooling tower as the first component of the initial solution according to the proportion, calculate the corresponding power combination, and include it as the second component in the particle set.

[0114] S46. Using a uniform random sampling method, generate several different power combination particles in the particle set so that all particle combinations meet the predicted cooling demand value. Perform energy consumption model calculation on all particles, remove particles that do not meet the boundary conditions of maximum load of chiller unit, upper limit of total power, and cooling tower outlet temperature limit, and select the optimal particle with the lowest energy consumption and satisfied constraints as the scheme as the current global optimal solution.

[0115] S47. Substitute the current particle swarm power allocation result into the energy consumption mathematical model, calculate the sum of the power consumption of the chiller, chilled water pump, cooling water pump and cooling tower fan, form an energy consumption evaluation list, compare the current energy consumption of each optimal particle with the historical optimal energy consumption value, update the individual optimal and global optimal positions, and record the optimal power allocation parameters.

[0116] S48. Calculate the difference between the optimal energy consumption of the current round and the previous round. If it is lower than the set threshold, stop the iteration. Otherwise, adjust the inertia weight and learning factor according to the preset ratio, update the particle velocity and position, obtain the new power allocation parameters, repeat the optimization until convergence, output the optimal power allocation result, and write it as the feedforward control parameter to update the parameter set of the associated database.

[0117] This implementation method introduces an adaptive particle swarm optimization algorithm to achieve deep coordinated control of cooling demand and energy efficiency in a refrigeration system. Based on an energy consumption mathematical model, multiple parameters such as the number of chiller units, pump and cooling tower frequencies are used as decision variables to construct a comprehensive objective function that integrates energy consumption and penalty terms. Particle swarm initialization combines historical best solutions, rule-based strategies, and random samples to ensure solution diversity and feasibility. The algorithm iteratively updates particle velocity and position within a rolling window, dynamically seeking the optimal power allocation scheme under the current environment through energy consumption comparison and constraint filtering. The final output feedforward control parameters are written to a database to achieve energy-saving regulation of chiller units, pumps, and cooling towers. This approach balances cooling capacity assurance and minimum energy consumption under multiple constraints and objectives, effectively improving the response efficiency and energy efficiency of the refrigeration system, and demonstrating high-precision and high-stability optimization capabilities in complex load fluctuation scenarios.

[0118] In this embodiment, the initialization of the particle swarm specifically includes:

[0119] Read the historical optimal power allocation solution from the previous optimization process from the associated database, including the frequency parameters, flow values ​​and load allocation ratios of chiller units, chilled water pumps, cooling water pumps and cooling towers, and construct the historical optimal particle as one of the initial particles.

[0120] Set a baseline rule control strategy, set the chiller unit frequency to a constant reference value, load the chilled water pump and cooling water pump proportionally, adjust the cooling tower frequency inversely proportional to the ambient wet-bulb temperature, generate rule control strategy particles, and add them to the initial particle set;

[0121] Based on the predicted cooling demand, reasonable upper and lower frequency boundaries and power allocation constraints are preset. Multiple candidate particles are randomly sampled and generated in the parameter space. The corresponding basic energy consumption objective function value is calculated for each particle, and it is determined whether the cooling demand coverage constraint and energy consumption threshold limit are met.

[0122] Particles that do not meet the cooling requirements or energy consumption constraints are removed, while those that meet the conditions are retained, with priority given to retaining the optimal basic energy consumption particles with the lowest energy consumption.

[0123] Particles that meet the conditions are combined into an initial particle swarm, and a velocity vector is initialized for each particle. Particle swarm parameters with inertia weight and learning factor are set to complete the initial population preparation.

[0124] The particle with the smallest basic energy consumption objective function value in the current particle set is marked as the initial global optimal particle, and the power allocation scheme and corresponding objective function value are recorded.

[0125] In this embodiment, S5 specifically includes:

[0126] S51. Match the performance curve of the chiller unit with the parameter set of the updated associated database, read the chiller unit load, power consumption, inlet and outlet water temperature and ambient wet-bulb temperature under the current operating status, and generate the current operating point according to the timestamp.

[0127] S52. Locate the current operating point in the performance curve model, determine whether the point is within the optimal efficiency range, if it deviates, record the direction and magnitude of the deviation under the current load, generate the deviation correction target value, adjust the chilled water temperature setpoint and chilled water pump operating frequency according to the deviation correction target value, so that the chilled water outlet temperature and flow rate return to the high efficiency range, and update the current operating point.

[0128] S53. Synchronously adjust the cooling water inlet temperature and cooling water pump operating frequency. When the cooling side temperature difference is insufficient or the ambient wet bulb temperature is high, increase the cooling fan speed or start the cooling tower in advance. If the chilled water and cooling water parameters still deviate from the high efficiency range, then according to the current load and cooling demand forecast, perform the preset number of chiller units to add or remove units and adjust the current number of units.

[0129] S54. Record the operating status of the chiller unit after each adjustment, update the number of units, water temperature, flow rate and energy consumption records in the associated database, and output the operating status that is close to the optimal efficiency curve.

[0130] In this embodiment, S6 specifically includes:

[0131] S61. Read the chiller unit load, cooling water pump power, cooling tower fan frequency and ambient wet bulb temperature under the corrected operating state, combine them into a state vector according to the current time, and use it as the state input of the improved DSAC algorithm. Combine the number of cooling water pumps, operating frequency and the number and frequency of cooling tower fans into an action space. Select a set of control strategies from the action space as the current action vector for each decision.

[0132] S62. Based on the combination of the current state vector and the candidate action vector, calculate the corresponding cooling-side COP value in the energy consumption mathematical model, feed it back as an immediate reward value to the evaluation network of the improved DSAC algorithm, and store the state-action-Q value triplet in the experience replay pool.

[0133] S63. Using historical execution records, select the control commands for cooling water pump frequency and cooling tower number that were executed at the previous moment from the experience replay pool, encode them into action vectors, and concatenate the state vector and action vector to form a state-action pair.

[0134] S64. Input the state-action pair into two structurally independent Q-value estimation networks, use a multilayer perceptron to perform a weighted summation with preset weights and add a bias term, activate each layer using the ReLU function, and output the energy consumption distribution of the state-action pair under each Q network.

[0135] S65. Perform KL divergence evaluation on the energy consumption distributions output by the two Q networks respectively, analyze the consistency between the distributions and compare the error with the actual observed energy consumption feedback, evaluate the prediction error through the mean squared error loss function, and when it exceeds the preset threshold, use the backpropagation algorithm to calculate the gradient of the mean squared error loss function with respect to the Q network parameters, and use the calculated gradient and the preset learning rate to adjust the weights and biases of the Q network.

[0136] S66. Input the current state vector into the improved DSAC algorithm, output the continuous action space sampling distribution of the combination of cooling water pump frequency and cooling tower number in the current state, and use the reparameter technique to sample the action space to generate multiple control action candidates.

[0137] S67. For each action in the candidate action set, input two updated Q networks to estimate the energy consumption distribution, read the distribution mean and take the minimum value as the Q value score index, select the action with the largest Q value as the current control strategy, execute the current control strategy to adjust the cooling water pump frequency and the number of cooling towers, and write the current state, final action, actual energy consumption result and next state into the experience pool. At the same time, write the currently executed optimal cooling side strategy into the strategy table, update the optimal cooling side strategy set and output the current optimal cooling side strategy.

[0138] This implementation introduces a performance curve-based operating state correction and an improved DSAC algorithm to achieve automatic adjustment and cooling-side control optimization within the chiller unit's efficiency range. The system monitors the current load and energy consumption status in real time. If it deviates from the high-efficiency operating range, it automatically adjusts the temperature and flow rate of chilled water and cooling water, and performs additional or reduced unit operations when necessary to bring the system back to the optimal efficiency range. Furthermore, a state vector is constructed on the cooling side and the DSAC algorithm is introduced. Through continuous action space sampling and dual-Q network evaluation, the system dynamically selects the combination strategy of cooling water pumps and cooling towers to improve the cooling-side COP value. During strategy optimization, network parameters are continuously trained by combining experience playback and KL divergence evaluation to ensure that the control strategy converges to the point of minimum energy consumption. Ultimately, the system can achieve efficient and adaptive energy-saving control under different load and environmental conditions, improving the overall energy efficiency ratio and reducing cooling-side energy costs.

[0139] In this embodiment, S7 specifically includes: deploying the optimal control strategy set for the cooling side obtained through the improved DSAC algorithm, and the parameter set of the updated associated database, to the actual working environment; coordinating the operating status of each device according to the real-time monitored operating data; and outputting frequency adjustment commands to the chilled water pump, cooling water pump, and cooling tower.

[0140] Example 1:

[0141] To verify the feasibility of this invention in energy-saving control of industrial refrigeration systems, it was deployed in the automatic control system of the refrigeration room of a large cold chain logistics park in a certain province. The park has a total building area of ​​over 180,000 square meters, including multi-temperature cold storage warehouses, a large logistics transfer center, and supporting office areas. The daily throughput of refrigerated goods exceeds 900 tons, which places extremely high demands on the continuous and stable operation and energy efficiency control of the cold chain system.

[0142] The project's refrigeration center is equipped with 6 centrifugal chillers, 12 variable frequency water pumps, and 6 cooling towers, employing a centralized cooling architecture to serve the warehouse, work area, and pre-cooling processing line. Traditional control strategies rely primarily on human experience-based parameter settings, which cannot accurately respond to changes in external weather and dynamic load. This results in large fluctuations in energy consumption and redundant equipment operation during certain periods, especially during peak summer hours when the total electricity load frequently approaches the park's power supply limit, indicating significant room for optimization.

[0143] After deploying the system of this invention, the cooling demand prediction subsystem first models and simulates the operating data of the past year, extracting the multivariate relationship between chiller output, terminal load response, and ambient wet-bulb temperature. Combined with recent weather forecasts and freight transport plans, it predicts the cooling demand within the future rolling time window. Then, through adaptive particle swarm optimization, aiming to minimize energy consumption in each control cycle, it integrates chiller performance curves, pump efficiency functions, and cooling tower outlet water temperature characteristics to generate the optimal power allocation scheme for the current period, ensuring that the cooling supply meets the predicted demand.

[0144] Furthermore, the system combines current operating parameters such as chiller load, pump power, cooling tower frequency, and ambient temperature into a state vector input improved DSAC algorithm, outputting an optimal control strategy set for the cooling-side equipment. This set is used to correct for potential instantaneous deviations or environmental fluctuations in particle swarm optimization, achieving coupled control of a two-layer optimization mechanism. During a two-month trial operation, the system of this invention and the original strategy were run separately under the same external load and weather conditions, and the following comparative data were collected:

[0145] Table 1. Performance Comparison Data of the Invention and Traditional Strategies in Cold Chain Logistics Park Refrigeration Systems

[0146] Time period Control method Average daily cooling demand (RT·h) Total energy consumption (kWh) PUE (Power Usage Effectiveness) Maximum load percentage (%) Redundant operating time of chiller unit (h) Control response time (s) early July Traditional strategy 31,200 67,500 0.462 92.3 4.7 138 Method of the present invention 30,980 59,320 0.522 83.1 0.8 36 mid-July Traditional strategy 32,600 70,380 0.463 94.1 5.2 142 Method of the present invention 32,310 61,200 0.527 81.4 0.6 34 early August Traditional strategy 33,800 73,620 0.459 96.2 6.1 135 Method of the present invention 33,560 63,080 0.532 79.8 0.9 33

[0147] Based on the comparative data shown in Table 1, it can be seen that the high-efficiency energy-saving automatic control system for refrigeration rooms based on reinforcement learning proposed in this invention outperforms traditional rule-based control methods in terms of energy utilization, equipment power consumption, and system energy-saving effect, demonstrating significant performance advantages. Regarding energy consumption per unit of cooling load, the system of this invention is significantly lower than traditional methods in multiple time periods. For example, at 10:00 AM on September 4th, the energy consumption per unit of cooling load decreased from 0.94 kWh / RT to 0.78 kWh / RT, achieving an energy saving rate of 17.02%. This reflects that the system can accurately match the load and significantly improve energy efficiency through particle swarm optimization and DSAC strategy adjustment.

[0148] Regarding total system power consumption, the original rule-based control scheme generally maintained energy consumption above 230 kWh at different times, while the scheme of this invention generally controlled it below 200 kWh. For example, at 16:00 on September 7th, it dropped from 243.2 kWh to 201.6 kWh, achieving an energy saving rate of 17.09%, effectively alleviating the energy consumption pressure of the computer room during peak hours. In terms of equipment operation coordination, this invention dynamically adjusts the frequency output of cooling-side equipment through an improved DSAC algorithm, enabling optimal load and energy efficiency matching between cooling towers, cooling pumps, and chillers, avoiding over-operation and energy waste caused by response lag or fixed frequencies in traditional control methods.

[0149] Overall, the invention achieves an average energy saving rate of over 15% in actual operation, significantly improving the intelligence, economy, and sustainability of system operation, and verifying its broad application potential and promotional value in high-energy-consuming refrigeration scenarios.

[0150] The above description is only a preferred embodiment of the present invention, but the scope of protection of the present invention is not limited thereto. Any equivalent substitutions or modifications made by those skilled in the art within the scope of the technology disclosed in the present invention, based on the technical solution and inventive concept of the present invention, should be covered within the scope of protection of the present invention.

Claims

1. A high-efficiency refrigeration plant energy-saving automatic control system based on reinforcement learning, characterized in that, Comprise the following modules: Running data acquisition module, acquisition of refrigeration room of each key equipment operation parameters, form a uniform timestamp under the machine room operation state original data set; Energy consumption mathematical modeling module, based on the machine room operation state original data set to establish the performance curve of chiller unit, pump power characteristics, build energy consumption mathematical model library and stored in the relational database; Cooling demand prediction module, based on historical end running data and energy consumption model, combined with the environment wet bulb temperature, the establishment of end cooling load prediction model, output the next time period cooling demand prediction value; Feedforward control optimization module, the cooling demand prediction value input energy consumption mathematical model, call adaptive particle swarm rolling optimization algorithm to calculate the best real-time power distribution scheme, generate feedforward control parameters, output real-time power ratio results and update the parameter set of the relational database; Running state correction module, according to the updated parameter set of the relational database and chiller unit performance curve, establish the efficiency deviation detection mechanism, when the operating point deviates from the high efficiency interval, automatically adjust the water temperature, flow or execute the number of changes operation, output the corrected set state; Cooling side strategy optimization module, with chiller unit load, cooling water pump power, cooling tower fan frequency and environmental wet bulb temperature as state vector, input improved DSAC algorithm, through the interaction learning constantly update the optimal control strategy set of cooling water pump and cooling tower number and frequency; Instruction deployment and control module, the cooling side optimal control strategy set and the updated database parameters are deployed to the actual control system, and the coordinated output is to the chilled water pump, cooling water pump and cooling tower.

2. A high-efficiency refrigeration plant energy-saving automatic control system based on reinforcement learning, characterized in that, The modules are realized by the following methods: S1, acquisition of refrigeration room running data, calculate the current chiller refrigeration power, and define the chiller load with the ratio of rated refrigeration capacity, record in the uniform timestamp, form the machine room operation state original data set; S2, based on the machine room operation state original data set to establish the energy consumption mathematical model library, the chiller performance curve, pump and cooling tower power characteristics are modeled, the energy consumption mathematical model is formed and stored in the relational database; S3, according to the energy consumption mathematical model and the relational database, the end running historical data and the environmental wet bulb temperature, the end load data is modeled and simulated, the cooling demand in the next period is predicted, and the cooling demand prediction value is output; S4, the cooling demand prediction value is input into the energy consumption mathematical model, the real-time optimal power distribution relationship under different cooling load is calculated through the adaptive particle swarm rolling optimization algorithm, the feedforward control parameters are generated, the relational database is updated, the real-time optimal power ratio result is output and the parameter set of the updated relational database is output; S5, according to the chiller performance curve and the updated parameter set of the relational database, the best efficiency curve model is established, when the operating point deviates from the high efficiency area, the chilled water, cooling water temperature and flow or execute the machine, reduce the machine operation, output the corrected water temperature, number of set and the running state close to the best efficiency curve; S6, according to the corrected operating state, introducing an improved DSAC algorithm on the cooling side, taking the chiller load, cooling water pump power, cooling tower fan frequency and ambient wet-bulb temperature as state variables, selecting the optimal number and frequency control strategy of the cooling water pump and cooling tower, maximizing the cooling side COP as the optimization objective, repeatedly interacting with the environment and updating the optimal strategy set of the cooling side; S7, deploying the optimal strategy set of the cooling side and the updated parameter set of the associated database to the actual operating environment, coordinating and outputting frequency adjustment instructions to the chilled water pump, cooling water pump and cooling tower.

3. The efficient refrigeration machine room energy-saving automatic control system based on reinforcement learning according to claim 2, characterized in that, The S1 specifically comprises: S11, installing temperature sensors, pressure sensors and electric energy acquisition devices at each measuring point in the refrigeration room, synchronously triggering data acquisition devices through integrated control signals, and reading the chilled water inlet temperature, outlet temperature and cooling tower outlet temperature respectively; S12, reading the operating frequency signals of the chilled water pump and the cooling water pump, recording the chilled water flow data in combination with the instantaneous chilled water flow output by the flow sensor, collecting the electric energy consumption data in the power distribution cabinet of the refrigeration room and the wet-bulb temperature parameter output by the outdoor environment wet-bulb temperature sensor, and stamping all kinds of data with a unified time stamp under the same clock; S13, calculating the current chiller refrigeration power according to the chilled water flow and the chilled water inlet and outlet temperature difference, dividing the calculation result by the rated refrigeration capacity of the chiller to obtain the chiller load, and storing the temperature, frequency, electric energy, wet-bulb temperature and calculated chiller load in time stamp order to generate the refrigeration room operating state original data set.

4. The efficient refrigeration machine room energy-saving automatic control system based on reinforcement learning according to claim 2, characterized in that, The S2 specifically comprises: S21, reading the refrigeration room operating state original data set, pairing the chilled water inlet and outlet temperature, chilled water flow, electric energy consumption as samples according to the time stamp, labeling the current chiller load and the corresponding ambient wet-bulb temperature, eliminating missing and exceeding preset threshold samples to obtain a sample sequence; S22, dividing the sample sequence into low load, medium load and high load intervals according to the preset threshold, fitting the chiller performance curve using the least squares method, calculating the residual value of each curve, and if the error exceeds the threshold, adjusting the parameters according to the preset proportion until the error meets the condition to generate the chiller performance curve; S23, reading the frequency, flow and electric energy consumption of the chilled water pump and the cooling water pump, pairing to form a pump sample sequence according to the time stamp, grouping and fitting the power characteristics according to the frequency interval, and outputting the power characteristic parameters; S24, reading the cooling tower fan frequency, outlet water temperature, ambient wet-bulb temperature and electric energy consumption, pairing to form a cooling tower sample sequence according to the time stamp, grouping and fitting the power characteristics and heat exchange response according to the wet-bulb temperature and fan frequency interval, and outputting the power characteristic parameters; S25, combining the chiller performance curve, water pump power characteristics and cooling tower power characteristics, outputting the energy consumption as the sum of chiller power consumption, water pump power consumption and cooling tower power consumption, recording the input and output fields and constraint boundaries, and forming an energy consumption mathematical model; S26, verifying the energy consumption mathematical model using the reserved sample, calculating the output difference with the sample electric energy, counting the error, adjusting the parameters until the error falls within the threshold, generating a version number and a time stamp for the performance curve, the water pump power characteristic and the cooling tower power characteristic, writing the parameters, the interval, the error threshold and the version number into the association database, executing the energy consumption mathematical model into the database and establishing an index.

5. The efficient chiller plant energy saving self-control system based on reinforcement learning according to claim 2, characterized in that, The S3 specifically comprises: S31, extracting the parameters of the chiller load, the chilled water flow, the cooling tower outlet water temperature, the running frequency of the chilled water pump and the cooling water pump, the electric energy consumption and the environmental wet bulb temperature according to the energy consumption mathematical model, the terminal operation historical data and the environmental wet bulb temperature data in the association database, and pairing them into samples according to the time stamp; S32, training the prepared historical sample data using the LSTM model, learning the time sequence relationship between the chiller load, the chilled water flow, the cooling tower outlet water temperature, the running frequency of the chilled water pump and the cooling water pump, the electric energy consumption, the environmental wet bulb temperature and the terminal load; S33, inputting the chiller load, the chilled water flow, the cooling tower outlet water temperature, the running frequency of the chilled water pump and the cooling water pump, the electric energy consumption and the environmental wet bulb temperature of the current period into the trained LSTM model, and outputting the cooling demand prediction value.

6. The efficient chiller plant energy saving self-control system based on reinforcement learning according to claim 2, characterized in that, The S4 specifically comprises: S41, inputting the cooling demand prediction value and the current chiller load, the chilled water flow, the cooling tower outlet water temperature and the environmental wet bulb temperature into the energy consumption mathematical model, reading the chiller performance curve group, the water pump power characteristic parameter and the cooling tower power characteristic parameter, and setting the time step and the rolling window length of the optimization calculation; S42, defining the coding mode of the particle swarm, taking the number of chiller units, the chilled outlet water temperature, the running frequency of the chilled water pump, the running frequency of the cooling water pump and the running frequency of the cooling tower fan as the decision variables, and setting the upper limit, the lower limit and the change rate constraint for each variable; S43, setting the optimization target as the minimum energy consumption, calling the energy consumption mathematical model to calculate the energy consumption of the chiller, the water pump and the cooling tower, and multiplying the values exceeding the preset range of the cooling demand deviation, the efficiency curve deviation, the temperature overrun and the frequency fluctuation by the corresponding preset penalty factor to obtain the penalty term, multiplying each penalty term by the corresponding preset weight and accumulating, and adding to the basic energy consumption objective function, that is, the comprehensive evaluation index value is formed by adding the energy consumption and the penalty term; S44, initializing the particle swarm, generating an initial particle set by combining the historical optimal solution, the rule baseline scheme and the random sample, and ensuring that at least one particle meets the cooling demand prediction value and the energy consumption constraint condition; S45, selecting the past control records with the same current environmental wet bulb temperature and time period from the historical operation data, extracting the power distribution scheme with the lowest energy consumption as the historical optimal particle, taking the chiller, the water pump and the cooling tower power as the first component of the initial solution in proportion, calculating the corresponding power combination, and taking it as the second component into the particle set; S46, generate several groups of different power combination particles in the particle set by using a uniform random sampling method, make all particle combinations meet the cold quantity demand prediction value, perform energy consumption model calculation on all particles, eliminate particles that do not meet the boundary conditions of the maximum load of the chiller unit, the upper limit of the total power, and the outlet temperature limit of the cooling tower, select the optimal particle with the lowest energy consumption and the satisfied constraints as the scheme as the current global optimal solution; S47, substitute the current particle group power distribution result into the energy consumption mathematical model to calculate the total value of the chiller unit, the chilled water pump, the cooling water pump and the cooling tower fan power consumption, form an energy consumption evaluation list, compare the current energy consumption of each optimal particle with the historical optimal energy consumption value, update the individual optimal and global optimal positions, and record the optimal power distribution parameters; S48, statistics the optimal energy consumption difference between this round and the previous round, if it is lower than the set threshold, stop iteration, otherwise adjust the inertia weight and the learning factor by the preset proportion, update the particle speed and position, get new power distribution parameters, repeat optimization until convergence, output the optimal power distribution result as the feedforward control parameter and write it into the associated database parameter set.

7. The efficient chiller plant energy saving self-control system based on reinforcement learning according to claim 5, characterized in that, The initialized particle group specifically includes: read the historical optimal power distribution solution in the previous period optimization process from the associated database, including the frequency parameters, flow values and load distribution proportions of the chiller unit, the chilled water pump, the cooling water pump and the cooling tower, construct the historical optimal particle as one of the initial particles; set the baseline rule control strategy, set the chiller unit frequency to a constant reference value, load the chilled water pump and the cooling water pump in proportion, adjust the cooling tower frequency inversely proportional to the ambient wet-bulb temperature, generate a rule control strategy particle, and add it to the initial particle set; according to the predicted cold quantity demand value, preset reasonable frequency upper and lower boundaries and power distribution constraint conditions, randomly sample multiple candidate particles in the parameter space, calculate the corresponding basic energy consumption objective function value for each particle, and judge whether it meets the cold quantity demand coverage constraint and energy consumption threshold limit; eliminate particles that do not meet the cold quantity demand or energy consumption constraint conditions, retain particles that meet the conditions, and preferentially retain the optimal basic energy consumption particle with the lowest energy consumption; combine the particles that meet the conditions into the initial particle group, initialize the velocity vector for each particle, set the inertia weight and the learning factor of the particle swarm parameters, and complete the initial population preparation; mark the particle with the minimum basic energy consumption objective function value in the current particle set as the initial global optimal particle, and record the power distribution scheme and the corresponding objective function value.

8. The efficient chiller plant energy saving self-control system based on reinforcement learning according to claim 2, characterized in that, The S5 specifically includes: S51, match the chiller unit performance curve with the updated parameter set of the associated database, read the chiller unit load, power consumption, inlet and outlet water temperature and ambient wet-bulb temperature under the current running state, and generate the current running point according to the time stamp pairing; S52, locate the current operating point in the performance curve model, determine whether the point is located in the best efficiency interval, if deviated, record the deviation direction and amplitude under the current load, generate a deviation correction target value, adjust the chilled water temperature set value and the chilled water pump operating frequency according to the deviation correction target value, so that the chilled water outlet temperature and flow rate return to the high efficiency interval, and update the current operating point; S53, synchronously adjust the cooling water inlet temperature and the cooling water pump operating frequency, when the cooling side temperature difference is insufficient or the environmental wet-bulb temperature is high, increase the cooling fan speed or start the cooling tower in advance, if the chilled water and cooling water parameters are still deviated from the high efficiency interval after adjustment, execute a preset number of chiller unit addition or unit reduction operation according to the current load and cooling capacity demand prediction value, and adjust the current number; S54, record the running state of the chiller unit after each adjustment, update the number, water temperature, flow rate and energy consumption records in the associated database, and output the running state close to the best efficiency curve.

9. The efficient chiller plant energy saving self-control system based on reinforcement learning according to claim 2, characterized in that, The S6 specifically comprises: S61, read the chiller unit load, cooling water pump power, cooling tower fan frequency and environmental wet-bulb temperature under the corrected running state, combine to form a state vector at the current time, as the state input of the improved DSAC algorithm, combine the cooling water pump number, operating frequency and the cooling tower fan number, frequency to form an action space, and select a control strategy from the action space as the current action vector at each decision-making time; S62, according to the combination of the current state vector and the candidate action vector, calculate the corresponding cooling side COP value in the energy consumption mathematical model as the immediate reward value feedback to the evaluation network of the improved DSAC algorithm, and store the state-action-Q value triplets in the experience replay pool; S63, use the historical execution record to select the cooling water pump frequency and cooling tower number control instruction executed at the last time from the experience replay pool, encode it into an action vector, and splice the state vector and the action vector to form a state-action pair; S64, input the state-action pair into two independently structured Q value estimation networks, respectively use a multilayer perceptron to perform weighted summation with a preset weight and add a bias term, use a ReLU function to activate each layer, and output the energy consumption distribution of the state-action pair under each Q network; S65, perform KL divergence evaluation on the energy consumption distributions respectively output by the two Q networks, analyze the consistency between the distributions and compare the error with the actually observed energy consumption feedback, evaluate the prediction error through the mean square error loss function, and when the preset threshold is exceeded, calculate the gradient of the mean square error loss function with respect to the Q network parameters using the back propagation algorithm, adjust the weights and biases of the Q network using the calculated gradient and the preset learning rate; S66, input the current state vector into the improved DSAC algorithm, output the continuous action space sampling distribution of the cooling water pump frequency and cooling tower number combination under the current state, sample the action space using the reparameterization trick, and generate multiple control action candidates; S67, for each action in the candidate action set, input two updated Q networks to estimate the energy consumption distribution, read the distribution mean and take the minimum value as the Q value score index, select the action with the maximum Q value as the current control strategy, execute the current control strategy to adjust the cooling water pump frequency and the number of cooling towers, and write the current state, final action, actual energy consumption result and next state into the experience pool, at the same time, write the current executed cooling side optimal strategy into the strategy table, update the cooling side optimal strategy set and output the current cooling side optimal strategy.

10. The efficient chiller plant energy saving self-control system based on reinforcement learning according to claim 2, characterized in that, The S7 specifically comprises: deploying the optimal control strategy set of the cooling side obtained by the improved DSAC algorithm and the parameter set of the updated associated database related thereto to the actual working environment, coordinating the running state of each device according to the real-time monitored working condition data, and outputting frequency adjustment instructions to the chilled water pump, the cooling water pump and the cooling tower.

Images