Real-time optimization method and system for cold station based on AI self-learning

Abstract

The application provides an AI self-learning-based real-time optimization method and system for a cold station. The method comprises the following steps: in each control cycle, collecting real-time data of key operating quantities of the cold station and performing preprocessing to generate constraint contribution values and operating constraint states of each operating quantity; calculating an overall adjustment coefficient, calculating a differentiated adjustment coefficient for each key operating quantity, and calculating a set of adjustable operating spaces that are constrained in the current control cycle; taking the operating strategy executed in the last cycle as a starting point, calculating an operating decision value of the current cycle under the boundary constraint of the set of adjustable operating spaces; converting the operating decision value into device control instructions that can be issued in the current cycle through an inertial progressive updating method, and sending the device control instructions to the cold station execution device to complete closed-loop control. The application provides a structured and implementable technical path for long-term application of self-learning technology in actual cold station scenarios.

Description

Technical Field

[0001] This invention belongs to the field of AI self-learning, and in particular relates to a method and system for real-time optimization of cold sites based on AI self-learning. Background Technology

[0002] Centralized chiller plants, as one of the core energy systems in large public buildings and industrial parks, typically involve multiple refrigeration units, cooling towers, chilled water pumps, cooling water pumps, and complex piping networks and terminal load systems. They exhibit significant characteristics such as a large number of devices, multiple operational dimensions, strong coupling relationships between devices, and continuous changes in operating conditions over time. In actual engineering projects, the operating status of chiller plants is not only affected by outdoor weather conditions and building load changes, but also closely related to equipment start-up and shutdown combinations, operational intensity distribution, and historical operating conditions, resulting in significant non-stationarity and phased characteristics in system operation. Existing chiller plant control methods mainly rely on manual experience-based parameter setting, fixed logic control, or automatic control based on local feedback. Although some systems have introduced model prediction or data-driven optimization methods, they still rely on static rules or offline modeling results, making it difficult to continuously adapt to changes in operating conditions during operation. While these methods can maintain basic operation under stable load and minimal equipment status changes, in actual long-term operation, when operating conditions shift or constraints gradually change, the original control strategies often fail to adjust in a timely manner, leading to significant fluctuations in system energy efficiency levels between different operating stages. With advancements in computing power and data acquisition capabilities, some research and engineering practices have begun to explore the introduction of self-learning or intelligent optimization mechanisms into chiller plant control, aiming to gradually approach the globally optimal operating state through online adjustments. However, chiller plant systems differ fundamentally from general simulation or information systems that allow for free trial and error. Their operation is subject to significant engineering constraints and cumulative effects of execution consequences. For example, refrigeration systems exhibit significant thermal inertia, equipment operating state adjustments are lagging, the execution layer faces cycle time limitations in response to command changes, and inappropriate exploratory control behavior can impact energy consumption, operational stability, and even equipment reliability over extended timescales. In this context, without a structured representation of operational constraints and continuous constraints on control adjustment magnitudes, directly applying self-learning or intelligent optimization algorithms to real-time chiller plant control can easily lead to overly conservative strategies or overly aggressive adjustments in the field, hindering the long-term stable application of these technologies in practical chiller plant systems. Summary of the Invention

[0003] The purpose of this invention is to propose a real-time optimization method and system for cold storage plants based on AI self-learning, in order to solve the above-mentioned problems.

[0004] To achieve the above objectives, a real-time optimization method for cold storage sites based on AI self-learning is provided in a first aspect of the present invention, the method comprising the following steps:

[0005] Step 1: In each control cycle, real-time data of key operating quantities of the chiller plant are collected and preprocessed to generate constraint contribution values ​​for each operating quantity, and the operating constraint status is calculated based on the constraint contribution values ​​of each operating quantity; the operating constraint status is used to comprehensively describe the overall constraint strength within the current control cycle.

[0006] Step 2: Based on the operational constraint state, calculate the overall adjustment coefficient to represent the global scaling benchmark for differentiated tightening; based on the overall adjustment coefficient and the constraint contribution value of the operational quantity, calculate the differentiated adjustment coefficient for each key operational quantity; and use the differentiated adjustment coefficient to scale the preset nominal adjustable range of each dimension to obtain the set of constrained adjustable operational space within the current control cycle.

[0007] Step 3: Starting from the execution strategy executed in the previous cycle, under the boundary constraints of the adjustable execution space set, and combined with the adjustment suggestions given by the pre-built AI self-learning model, the execution decision value of the current cycle with limited optimization is calculated; wherein, the calculation process introduces a boundary avoidance regularization term to make the execution decision value automatically move away from the interval edge when the adjustable execution space is narrow.

[0008] Step 4: The operation decision value is converted into equipment control commands that can be issued in the current cycle through an inertial incremental update method, and then sent to the chiller station execution equipment to complete closed-loop control.

[0009] Furthermore, the key operating parameters of the chiller plant include the operating intensity of the chiller unit, the operating status of the water pumps in the water system, and the operating intensity of the cooling system.

[0010] Furthermore, the operational constraint state is obtained by weighted summation of the constraint contribution values ​​of each operational quantity;

[0011] The constraint contribution value is calculated and generated by a piecewise linear mapping function based on the proportion value, inflection point parameter, and slope parameter of the preprocessed key operating quantities of the chiller station. The inflection point parameter is used to represent the upper limit of the normal adjustment range of the operating quantity. The slope parameter is used to represent the rate of increase of the constraint contribution after the proportion value exceeds the inflection point. The piecewise linear mapping function is used to characterize the relationship between the constraint contribution and the proportion value.

[0012] Furthermore, the calculated overall adjustment coefficient is used to convert the running constraint state into a continuous coefficient for the scaling interval width;

[0013] The differential adjustment coefficient is calculated based on the overall adjustment coefficient, the constraint contribution value of the corresponding operating quantity, and the penalty intensity coefficient; the penalty intensity coefficient is used to adjust the degree of influence of the constraint distribution on the differential tightening of the operating space.

[0014] Specifically, when the constraint contribution value of the corresponding operating quantity is large, the penalty term amplifies and compresses the differentiated adjustment coefficient; when the constraint contribution value of the corresponding operating quantity is small, the differentiated adjustment coefficient is closer to the global adjustment scale given by the overall adjustment coefficient.

[0015] Furthermore, the step of scaling the preset nominal adjustable ranges of each dimension using the differentiated adjustment coefficient to obtain the constrained adjustable operating space set within the current control cycle is specifically as follows:

[0016] The nominal interval center and nominal half-width are pre-configured based on the aforementioned differential adjustment coefficient;

[0017] The adjustable operating space of the current control cycle is the set of intervals centered on the nominal interval center and with the product of the differential adjustment coefficient and the nominal half-width as the half-width.

[0018] Furthermore, the pre-built AI self-learning model is a fixed-structure feedforward computation structure:

[0019] The input layer receives an adjustable operating space set, which then passes through several linear transformation layers and nonlinear activation layers to output a one-dimensional performance score. At the same time, a dimension-wise adjustment amount is given on the output side through a differentiable linear header.

[0020] The model parameters are gradually corrected according to the control cycle during operation, thus achieving a self-learning effect.

[0021] Its online updates use energy efficiency evaluation values ​​obtained after operation for error-driven correction, keeping the model structure stable and the parameters updated incrementally.

[0022] Furthermore, the operational decision value is generated by an interval pruning operator based on the operational strategy executed in the previous cycle, the dimension-wise adjustment suggestions calculated by the self-learning model according to the current candidate strategy, the boundary avoidance regularization term, and the adjustable operational space set.

[0023] Furthermore, the boundary avoidance regularization term is generated based on the product of the previously executed operating strategy, the nominal interval center, and the differential adjustment coefficient and the nominal half-width, and is used to represent the special constraints in the cold station scenario:

[0024] When the convergence of the operating space causes the product of the differential adjustment coefficient and the nominal half-width to decrease, the boundary avoidance regularization term is automatically amplified, making the operating decision value more strongly pull back to the center; when the operating space is wide, the boundary avoidance regularization term is weakened, making the dimension-by-dimensional adjustment suggestions calculated by the self-learning model based on the current operating decision value amplified.

[0025] Furthermore, the incremental update method with inertia is specifically as follows:

[0026] The current cycle's operation control command value is equal to the previous cycle's operation control command value plus an update coefficient multiplied by the difference between the current operation decision value and the previous cycle's operation control command value; the update coefficient is pre-configured according to the response characteristics of different types of execution devices.

[0027] A second aspect of the present invention provides a real-time optimization system for cold storage plants based on AI self-learning, the system comprising:

[0028] The status calculation module is used to collect real-time data of key operating quantities of the cooling plant in each control cycle, preprocess the data, generate constraint contribution values ​​for each operating quantity, and calculate the operating constraint status based on the constraint contribution values ​​of each operating quantity; the operating constraint status is used to comprehensively describe the overall constraint strength within the current control cycle.

[0029] The spatial analysis module is used to calculate the overall adjustment coefficient based on the operational constraint state, which represents the global scaling benchmark for differentiated tightening; based on the overall adjustment coefficient and the constraint contribution value of the operational quantity, it calculates the differentiated adjustment coefficient for each key operational quantity; and uses the differentiated adjustment coefficient to scale the preset nominal adjustable range of each dimension to obtain the set of constrained adjustable operational space within the current control cycle.

[0030] The optimization decision module is used to calculate the operation decision value for the current cycle under the boundary constraints of the adjustable operation space set, taking the operation strategy executed in the previous cycle as the starting point and combining the adjustment suggestions given by the pre-built AI self-learning model. The calculation process introduces a boundary avoidance regularization term to make the operation decision value automatically move away from the interval edge when the adjustable operation space is narrow.

[0031] The instruction execution module is used to convert the operation decision value into equipment control instructions that can be issued in the current cycle through an inertial incremental update method, and send them to the chiller station execution equipment to complete closed-loop control.

[0032] The beneficial technical effects of the present invention are at least as follows:

[0033] This invention, starting from the engineering constraints of chiller plant operation, proposes a self-learning real-time optimization method and system centered on operational constraint states. By organizing the complex and non-stationary chiller plant operation process into a clear continuous decision chain of "state representation—operational space—constrained strategy—execution command," it achieves stable implementation of intelligent optimization in the engineering system. This invention structures key operational variables of the chiller plant to form an operational constraint state representation that reflects the adjustment capacity of the current operational stage. Furthermore, it continuously maps this representation to an adjustable operational space within the current control cycle, giving the system clear and calculable adjustment boundaries at different operational stages. Based on this, the invention executes self-learning-driven real-time optimization decisions under strict constraints of the adjustable operational space, allowing the operational strategy to evolve gradually within the continuous control cycle while naturally constrained by the operational space, avoiding over-adjustment during the constraint convergence phase. At the execution level, this invention transforms the constrained optimization strategy into progressive control commands that conform to the equipment response characteristics, enabling the intelligent decision results to be smoothly executed at the chiller plant equipment side and forming a continuous closed loop. Through the above design, this invention organically combines operational constraint perception, self-learning optimization, and engineering execution mechanisms, enabling the chiller system to achieve continuous and adaptive energy efficiency optimization under complex operating conditions, while maintaining operational stability and engineering controllability. This provides a structured and implementable technical path for the long-term application of self-learning technology in actual chiller scenarios. Attached Figure Description

[0034] The present invention will be further described with reference to the accompanying drawings, but the embodiments in the drawings do not constitute any limitation on the present invention. For those skilled in the art, other drawings can be obtained based on the following drawings without creative effort.

[0035] Figure 1 This is a flowchart of the real-time optimization method for cold storage plants based on AI self-learning, as described in this invention. Detailed Implementation

[0036] Embodiments of the present invention are described in detail below. Examples of these embodiments are shown in the accompanying drawings, wherein the same or similar reference numerals denote the same or similar elements or elements having the same or similar functions throughout. The embodiments described below with reference to the accompanying drawings are exemplary and are only used to explain the present invention, and should not be construed as limiting the present invention.

[0037] like Figure 1 As shown in the embodiment of the present invention, a real-time optimization method for cold storage plants based on AI self-learning is provided, the method comprising:

[0038] Step 1: In each control cycle, real-time data of key operating quantities of the cooling plant are collected and preprocessed to generate constraint contribution values ​​for each operating quantity, and the operating constraint status is calculated based on the constraint contribution values ​​of each operating quantity; the operating constraint status is used to comprehensively describe the overall constraint strength within the current control cycle.

[0039] Specifically, within each control cycle, this step transforms the current key operational variables of the chiller plant into operational constraint state representations, used to characterize the system's tolerance for operational adjustments within that cycle. The data used comes directly from existing automated control system points in the chiller plant, read from the PLC / DDC via building automation or data center group control systems. Common communication methods include BACnet / IP, Modbus TCP, or OPC interfaces. Data acquisition is performed on a control cycle basis, for example, obtaining an operational snapshot from the control system every minute. The acquisition targets are configured according to equipment category, including operational intensity-related points for the chiller unit, operational status points for water pumps in the water system, and operational intensity points for the cooling system. The acquired data is stored in the control system cache or real-time database in timestamp and numerical form; this step only uses the data set from the current control cycle.

[0040] Furthermore, during data processing, each key operational variable is first mapped to a dimensionless proportional value to eliminate the magnitude differences between different operational variables. Each key operational variable is configured with an acceptable low and high value range during the system deployment phase; this range is derived from the equipment's allowable range, protection threshold, or existing debugging parameters. The operational variable values ​​sampled within the current control cycle are linearly mapped to a value between 0 and 1 according to this range, with out-of-bounds values ​​truncated at the boundaries. The i-th proportional value obtained after this processing is denoted as... The value reflects the position of the operational quantity relative to the acceptable range within the current cycle. Subsequently, the proportional value is mapped to a constraint contribution value through the operational constraint state model, which describes the degree of influence of the operational quantity on the overall operational constraints under the current state. The mapping adopts a piecewise linear form. When the proportional value is within the normal adjustment range, the constraint contribution changes smoothly with the proportional value. When the proportional value exceeds the configured inflection point, the constraint contribution increases linearly with the degree of excess, thus reflecting the rapid compression of the system adjustment margin when the operational quantity approaches the engineering boundary.

[0041] The operational constraint state characterization is obtained by weighted summation of the constraint contribution values ​​of each operational quantity, and its calculation form is as follows:

[0042] ;

[0043] in, It indicates the operational constraint status and is used to comprehensively describe the overall constraint strength within the current control cycle; The number of critical runtime variables involved in the calculation; The weighting coefficient for the i-th critical operational quantity is used to reflect the relative impact of different operational quantities on the overall constraints. This coefficient is configured based on equipment combination and operational experience during system deployment. The constraint contribution value for the i-th runtime quantity; The proportion of the i-th critical operational quantity is obtained by interval mapping from real-time sampled values; This is a piecewise linear mapping function used to characterize the relationship between constraint contribution and scale value; its definition is:

[0044] ;

[0045] in, This is the inflection point parameter, representing the upper limit of the normal adjustment range of this operating volume; The slope parameter represents the rate at which the constraint contribution increases after the proportional value exceeds the inflection point. These parameters are configured during the system deployment phase to reflect the sensitivity of different workloads to operational consequences.

[0046] For example, within a typical control cycle, assuming the following is selected... Key operational quantities, after being scaled up, are obtained as follows: , , The configuration parameters are , , , , , The weight is , , Based on the mapping relationship, we can obtain... , , Further calculations yielded This result indicates that the second operational variable has a significant impact on the overall constraints, resulting in a high level of operational constraint characterization, which subsequent steps can use to generate the corresponding adjustable operational space.

[0047] Step 2: Based on the operational constraint state, calculate the overall adjustment coefficient to represent the global scaling benchmark for differentiated tightening; based on the overall adjustment coefficient and the constraint contribution value of the operational quantity, calculate the differentiated adjustment coefficient for each key operational quantity; and use the differentiated adjustment coefficient to scale the preset nominal adjustable range of each dimension to obtain the set of constrained adjustable operational space within the current control cycle.

[0048] Specifically, within each control cycle, the operational constraint state representation output from Step 1 is converted into an adjustable operational space defining "how much adjustment is allowed in the current cycle and the maximum adjustment range." Step 1 has already compressed the critical operational quantities of the chiller plant into two levels of information: on the one hand, using... On the one hand, it expresses the overall strength of constraints in the current cycle; on the other hand, it uses... This step analyzes the distribution differences of constraints across different operational dimensions. Determine the "overall degree of tightening". The decision on "the order and magnitude of tightening in different dimensions" ultimately forms a runtime boundary that can be directly invoked in the next step, so that subsequent real-time optimization and self-learning are naturally bounded by engineering boundaries during computation.

[0049] The input is the output of the previous step, including... as well as . It is the operational constraint state obtained by mapping key operational quantities in the previous cycle; This is the set of constraint contribution values ​​for each key operational quantity in the same mapping process, used to characterize the structural differences in the source of constraints. Here, the subscript 'i' is consistent with step one, representing the i-th key operational dimension involved in constraint representation; this dimension corresponds to the operational intensity characteristics of a type of controllable object during engineering implementation, such as the operational intensity dimension on the host side, the operational intensity dimension on the water system side, or the operational intensity dimension on the cooling side. The goal of the analysis is to generate an "allowable variation ratio" for each dimension, and then, based on this, shrink the nominal adjustment range configured in the deployment phase into an adjustable operational space for the current control cycle.

[0050] Furthermore, the first step of the analysis is... Calculate the overall adjustment coefficient , Its function is to transform the "constraint strength" into a continuous coefficient that can be directly used to scale the interval width. This mapping employs a bounded monotonic relationship, making... Continuous variation under different operating conditions avoids abrupt jumps in the operating space between adjacent control cycles, thus better aligning with the characteristics of cooling plants, which exhibit thermal inertia and operational lag. The overall regulation coefficient is calculated as follows:

[0051] ;

[0052] in, This represents the operational constraint state output from step one; This is the overall adjustment coefficient, serving as the global scaling benchmark for subsequent differential tightening in this step.

[0053] Furthermore, the second step of the analysis will and Fusion, calculating the differential adjustment coefficient for each dimension. The unique characteristic of cold-station scenarios lies in the fact that when overall constraints are strong, the priority should be to tighten the "most sensitive dimensions" rather than proportionally shrinking all dimensions. Otherwise, a situation may arise where some sensitive dimensions still allow for significant adjustments, while some insensitive dimensions are over-tightened. Therefore, This achieves the effect of "tightening sensitive dimensions first, then tightening insensitive dimensions" through a proportional structure with a penalty term: when When the value is large, the penalty term is amplified and compressed. ;when When smaller, Closer The given global adjustment scale. This relationship can be directly calculated as a constant on the controller or platform side during implementation, meeting the real-time calculation requirements within the control cycle. The calculation form of the differential adjustment coefficient is as follows:

[0054] ;

[0055] in, Let be the differential adjustment coefficient for the i-th dimension within the current control period; This is the contribution value of the i-th constraint output in step one; The penalty intensity coefficient configured for the deployment phase is used to adjust the degree of influence of constraint distribution on the differential tightening of the operating space. It is usually configured as a fixed value by the commissioning personnel based on the equipment sensitivity and on-site strategy preferences. For the reason The overall adjustment coefficient obtained.

[0056] Furthermore, to obtain Subsequently, the adjustable operating space is generated by shrinking the "center and half-width of the nominal interval". In engineering, the nominal interval center is pre-configured for each dimension. With nominal half width These originate from the normal operation strategy boundaries determined during the project commissioning phase (e.g., the conventional adjustable range corresponding to each dimension) and remain stable during operation. The adjustable operating space of the current control cycle is thus defined by... Centered on, with This is a set of half-width intervals; the execution or optimization side only needs to restrict candidate policies to this interval. For example: within a certain control cycle... At a moderate level When taking a medium scale, if a certain dimension If significantly larger, then It will further decrease under the effect of the penalty term, corresponding to the half-width of the interval. Consequently, it shrinks significantly; while another dimension Smaller, then Closer The contraction within the range is relatively mild. Taking the value within one cycle as an example, if... ,but If taken And the constraint contribution values ​​of two certain dimensions are respectively and ,but , In the same nominal half-width Under this configuration, the adjustable half-width of the second dimension will be smaller than that of the first dimension, thereby tightening the "more sensitive operating links" to a safe and acceptable adjustment range earlier, forming an operating space that conforms to the distribution characteristics of the constraints on the site of the cooling plant.

[0057] Furthermore, the output of this step is the adjustable operating space of the current control cycle. In engineering implementation, this can be output as a set of adjustable operating spaces by dimension. This set is directly used as a constraint for real-time optimization and self-learning decision-making in the next step. By... Mapped to overall adjustment coefficient and combined Generate differential adjustment coefficient This step injects two types of information, namely "constraint strength and constraint distribution", into the adjustable runtime space, so that the runtime space has both overall convergence characteristics and priority tightening characteristics for sensitive dimensions, providing clear and engineering-executable boundary inputs for subsequent policy calculations.

[0058] Step 3: Starting from the execution strategy executed in the previous cycle, under the boundary constraints of the adjustable running space set, and combined with the adjustment suggestions given by the pre-built AI self-learning model, calculate the running decision value of the current cycle with limited optimization; wherein, the calculation process introduces a boundary avoidance regularization term to make the running decision value automatically move away from the interval edge when the adjustable running space is narrow.

[0059] Specifically, in each control cycle, this step uses the adjustable operating space output from step two as constraint input to complete the real-time optimization decision for the cooling plant and introduces a self-learning mechanism to iteratively correct the decision. Step two has transformed the operational constraint state representation from step one into a set of adjustable operating spaces. ,in and The nominal runtime range configuration from the deployment phase. The constraint analysis results are derived from the current control cycle. The operational characteristics of the cold station scenario determine that there is thermal inertia and response lag between adjacent control cycles. The operating strategy needs to evolve smoothly within continuous cycles. Therefore, this step generates candidate strategies and performs constrained updates within the operating space, starting from the "strategy executed in the previous cycle". This ensures that the output strategy satisfies the operating boundary of the current cycle while maintaining the continuity of strategy changes, facilitating stable execution in the actual data center control link.

[0060] The input is the set of adjustable operating space intervals for the current control cycle. and the operating strategy already executed in the previous control cycle. (This strategy, as an internal state during system operation, is stored in the group control platform or controller cache; it belongs to the historical execution records naturally generated by the control system and can be directly read.) When constructing candidate strategies within the operating space, using... Based on the nominal center The pullback mechanism ensures that candidate points remain near the center of the current period interval, thus reducing the probability of hitting the boundary during convergence. Subsequently, the self-learning mechanism provides dimension-by-dimensional adjustment suggestions for the candidate strategy and implements strong constraints through interval pruning of the operating space.

[0061] Furthermore, the computational carrier of the self-learning mechanism adopts a parameterized performance mapping model to map candidate operating strategies to a performance score and output adjustment direction and magnitude suggestions for each operating dimension. In engineering implementation, this model can be configured as a fixed-structure feedforward computation structure: the input layer receives candidate operating strategy vectors, which then pass through several linear transformation layers and nonlinear activation layers to output a one-dimensional performance score, while providing dimension-wise adjustment amounts on the output side through a differentiable linear header. The model parameters are gradually corrected according to the control cycle during operation, forming a self-learning effect; its online update uses the energy efficiency evaluation value obtained after operation for error-driven correction, keeping the model structure stable and the parameters updated incrementally. Since the goal of the cooling station is to "improve the overall energy efficiency and maintain stability within the boundary", this step introduces a "boundary avoidance regularization term" for the cooling station scenario during strategy update. This regularization term causes the strategy to automatically move away from the interval edge when the adjustable operating space is narrow, thereby reducing the probability of touching the protection threshold or causing fluctuations; when the operating space is wide, the regularization term naturally weakens, allowing the strategy to make fuller use of the adjustable range.

[0062] Furthermore, the constrained optimization operation strategy for the current control cycle is obtained through the following update method:

[0063] ;

[0064] in, This represents the final operational decision value adopted by the i-th operational dimension within the control period t; The operating decision values ​​that were executed in the previous control cycle are directly read from the historical execution records of the control system; and The nominal operating interval center and half-width used in step two are derived from the deployment phase configuration; The differential adjustment coefficient obtained from step two is used to make the half-width of the adjustable range for the current cycle [value missing]. ; The dimension-wise adjustment suggestions calculated by the self-learning model based on the current candidate strategy are derived from the output header of the feedforward computation structure. This is the step size factor, configured during the deployment phase, used to limit the update magnitude within a single cycle; This is the boundary avoidance coefficient, configured during the deployment phase, used to adjust the intensity of the "pullback towards the center of the interval"; It is a small positive constant, configured during the deployment phase, to avoid [further issues]. In extremely rare cases, this can result in excessively large proportional terms. For the interval pruning operator, ensure that the update result strictly falls within the running space given in step two. Inside. In the update item This reflects the special constraints of the cold station scenario: when the operating space converges, leading to... When the scale factor decreases, it automatically increases, making the strategy pull back towards the center more strongly; when the operating space is wider, the scale factor decreases, making the self-learning suggestions more effective. This allows for a more comprehensive functioning of the system, thus naturally binding the effect of "whether or not to allow stronger learning intervention" to the runtime space output from step two.

[0065] Within a typical control cycle, assuming the nominal center of a certain operational dimension is... The nominal half-width is Step two analysis yields Then the adjustable half-width of this dimension in the current period is If the previous cycle's execution value If it is close to the edge of the interval, then If the absolute value is large, the boundary avoidance term will account for a higher proportion in the update, making... Move towards the center of the interval; if It is already close to If this value is small, then the strategy update is mainly driven by self-learning suggestions. This allows for more detailed optimization near a safe location. This behavior aligns with common-sense control practices at cold plant sites: when the system is near a boundary, it should prioritize returning to a safe zone, and once within the safe zone, more thorough fine-tuning and optimization are permitted.

[0066] Furthermore, the output of this step is the constrained optimization operation strategy for the current control cycle. This strategy satisfies the runtime constraints given in Step 2 dimension by dimension, and, with the help of self-learning suggestions, generates progressive optimization results over consecutive cycles. This output will be directly used in the next step to generate and execute the chiller plant operation control instruction set.

[0067] Step 4: The operation decision value is converted into equipment control commands that can be issued in the current cycle through an inertial incremental update method, and then sent to the chiller station execution equipment to complete closed-loop control.

[0068] Specifically, in each control cycle, this step involves implementing the constrained optimization strategy output from step three. At the executable layer, generate and issue cooling plant operation control commands. And complete the effective execution within the current cycle. The execution objects at the chiller plant site usually include chiller controller, variable frequency pump controller, cooling side control unit, etc. These objects have rhythm and inertia characteristics in accepting command changes. Therefore, this step adopts the execution link of "strategy target → command progressive approximation → execution", so that the continuous evolution of the strategy layer can be smoothly inherited by the equipment side.

[0069] Furthermore, the control command generation adopts the first-order inertial / exponential smoothing concept commonly used in discrete control. Its original form can be traced back to the "first-order inertial element" execution model in classical control engineering: the response of the executed quantity to the target quantity has inertia, suitable for being represented by a discrete update where "the current value approaches the target value proportionally." The engineering modification made in this application to this classical concept is: the execution cycle time and device response characteristics of each dimension are uniformly converted into a single configuration parameter. This allows the variation range within each control cycle to be directly configured according to the device type, and continuously transmits the executed instruction value from the previous cycle as a status variable, thereby enabling the strategy layer output to be displayed correctly. The changes are stably translated into instructions acceptable to the device. The incremental update method with inertia has the following specific update formula:

[0070] ;

[0071] in, This refers to the value of the i-th execution control instruction issued to the execution layer within the control cycle t. The control command values ​​that have been issued and taken effect in the previous control cycle are directly read from the command history cache of the group control platform or controller; The value of the restricted optimization strategy output in step three; The update coefficients for the execution layer are configured as fixed values ​​during the deployment phase based on the device's response characteristics. Different devices can use different coefficients; for example, a smaller coefficient is used for host-type devices to reflect stronger inertia, while a medium coefficient is used for pump-type devices to reflect faster response. Both sides of the formula represent the setpoint values ​​for the same controlled object, and the values ​​within parentheses represent the differences between similar values. As a scaling factor, the update result remains consistent with the instruction value.

[0072] It is important to note that the logical relationship between the formulas is reflected in the process of "mapping first, then progressively, and then issuing":

[0073] First, according to the binding relationship... Align the target setpoint semantics to the corresponding point; then calculate using the above update formula. To achieve asymptotic approximation of the goal; and finally... Write the control system point and wait for equipment confirmation of effectiveness. To demonstrate the feasibility of the calculation, a calculation example within a control cycle is given: Assume that the instructions executed in the previous cycle for a certain dimension i are... The target given by this cycle strategy is This dimension configuration The instruction issued in this cycle is: If the strategy objective for the next cycle is still... and If it remains unchanged, it will continue to approach the target using the same mechanism in subsequent cycles, thus decomposing the changes at the strategy level into executions over multiple cycles and reducing the volatility risk caused by sudden changes on the execution side. During the execution phase, the group control platform uses existing control interfaces to... Write to the corresponding point and record the writing timestamp, point readback value and execution confirmation status to form a debugging record; in engineering implementations that require the retention of supporting materials, the controller point log and the group control platform distribution log can be directly exported as the record basis for the issuance and execution of instructions.

[0074] This invention also provides a real-time optimization system for cold storage plants based on AI self-learning, the system comprising:

[0075] The status calculation module is used to collect real-time data of key operating quantities of the cooling plant in each control cycle, preprocess the data, generate constraint contribution values ​​for each operating quantity, and calculate the operating constraint status based on the constraint contribution values ​​of each operating quantity; the operating constraint status is used to comprehensively describe the overall constraint strength within the current control cycle.

[0076] The spatial analysis module is used to calculate the overall adjustment coefficient based on the operational constraint state, which represents the global scaling benchmark for differentiated tightening; based on the overall adjustment coefficient and the constraint contribution value of the operational quantity, it calculates the differentiated adjustment coefficient for each key operational quantity; and uses the differentiated adjustment coefficient to scale the preset nominal adjustable range of each dimension to obtain the set of constrained adjustable operational space within the current control cycle.

[0077] The optimization decision module is used to calculate the operation decision value for the current cycle under the boundary constraints of the adjustable operation space set, taking the operation strategy executed in the previous cycle as the starting point and combining the adjustment suggestions given by the pre-built AI self-learning model. The calculation process introduces a boundary avoidance regularization term to make the operation decision value automatically move away from the interval edge when the adjustable operation space is narrow.

[0078] The instruction execution module is used to convert the operation decision value into equipment control instructions that can be issued in the current cycle through an inertial incremental update method, and send them to the chiller station execution equipment to complete closed-loop control.

[0079] Those skilled in the art will understand that, for the sake of convenience and brevity, the specific working processes of the systems, devices, and units described above can be referred to the corresponding processes in the foregoing method embodiments, and will not be repeated here.

[0080] In the embodiments provided in this application, it should be understood that the disclosed systems, apparatuses, and methods can be implemented in other ways. For example, the apparatus embodiments described above are merely illustrative; for instance, the division of units is only a logical functional division, and in actual implementation, there may be other division methods. For example, multiple units or components may be combined or integrated into another system, or some features may be ignored or not executed. Furthermore, the coupling or direct coupling or communication connection shown or discussed may be through some interfaces; the indirect coupling or communication connection of apparatuses or units may be electrical, mechanical, or other forms.

[0081] If the aforementioned functions are implemented as software functional units and sold or used as independent products, they can be stored in a computer-readable storage medium. Based on this understanding, the technical solution of this application, in essence, or the part that contributes to the prior art, or a portion of the technical solution, can be embodied in the form of a software product. This computer software product is stored in a storage medium and includes several instructions to cause a computer device (which may be a personal computer, server, or network device, etc.) to execute all or part of the steps of the methods described in the various embodiments of this application. The aforementioned storage medium includes various media capable of storing program code, such as USB flash drives, portable hard drives, read-only memory (ROM), random access memory (RAM), magnetic disks, or optical disks.

[0082] Although embodiments of the invention have been shown and described, those skilled in the art will understand that various changes, modifications, substitutions and variations can be made to these embodiments without departing from the principles and spirit of the invention, the scope of which is defined by the claims and their equivalents.

Claims

1. A real-time optimization method for cold storage sites based on AI self-learning, characterized in that, The method includes: Step 1: In each control cycle, real-time data of key operating quantities of the chiller plant are collected and preprocessed to generate constraint contribution values ​​for each operating quantity. The operating constraint status is then calculated based on these values. The operating constraint status is obtained by weighted summation of the constraint contribution values ​​of each operating quantity. This status comprehensively describes the overall constraint strength within the current control cycle. The constraint contribution values ​​are calculated using a piecewise linear mapping function to represent the proportions of each key operating quantity. These proportions are obtained by linearly mapping the values ​​of each operating quantity to a preset interval between 0 and 1. Step 2: Based on the operational constraint state, calculate the overall adjustment coefficient to represent the global scaling benchmark for differentiated tightening; based on the overall adjustment coefficient and the constraint contribution value of the operational quantity, calculate the differentiated adjustment coefficient for each key operational quantity; and use the differentiated adjustment coefficient to scale the preset nominal adjustable range of each dimension to obtain the set of constrained adjustable operational space within the current control cycle; wherein, each dimension includes the operational intensity dimension that can be associated with the host side, the operational intensity dimension of the water system side, or the operational intensity dimension of the cooling side; The overall adjustment coefficient is used to convert the running constraint state into a continuous coefficient for the scaling interval width; The differential adjustment coefficient is calculated based on the overall adjustment coefficient, the constraint contribution value of the corresponding operating quantity, and the penalty intensity coefficient; the penalty intensity coefficient is used to adjust the degree of influence of the constraint distribution on the differential tightening of the operating space. Step 3: Starting from the execution strategy of the previous cycle, under the boundary constraints of the adjustable operating space set, and combined with the adjustment suggestions given by the pre-built AI self-learning model, the operating decision value for the current cycle's constrained optimization is calculated; wherein, the calculation process introduces a boundary avoidance regularization term to ensure that the operating decision value automatically moves away from the interval edge when the adjustable operating space is narrow; wherein, the pre-built AI self-learning model is a fixed-structure feedforward computation structure: The input layer receives an adjustable operating space set, which then passes through several linear transformation layers and nonlinear activation layers to output a one-dimensional performance score. At the same time, a dimension-wise adjustment amount is given on the output side through a differentiable linear header. The model parameters are gradually corrected according to the control cycle during operation, thus achieving a self-learning effect. Its online updates use energy efficiency evaluation values ​​obtained after operation for error-driven correction, keeping the model structure stable and the parameters updated incrementally; The running decision value is generated by the interval pruning operator based on the running strategy executed in the previous cycle, the dimension-wise adjustment suggestions calculated by the self-learning model according to the current candidate strategy, the boundary avoidance regularization term, and the adjustable running space set. The boundary avoidance regularization term is generated based on the product of the previously executed operating strategy, the nominal interval center, and the differential adjustment coefficient and the nominal half-width, and is used to represent the special constraints in the cold station scenario: When the convergence of the operating space causes the product of the differential adjustment coefficient and the nominal half-width to decrease, the boundary avoidance regularization term is automatically amplified, making the operating decision value more strongly pull back to the center; when the operating space is wide, the boundary avoidance regularization term is weakened, making the dimension-by-dimensional adjustment suggestions calculated by the self-learning model based on the current operating decision value amplified. Step 4: The operation decision value is converted into equipment control commands that can be issued in the current cycle through an inertial incremental update method, and then sent to the chiller station execution equipment to complete closed-loop control.

2. The real-time optimization method for cold storage stations based on AI self-learning according to claim 1, characterized in that, The key operating parameters of the chiller plant include the operating intensity of the chiller, the operating status of the water pumps in the water system, and the operating intensity of the cooling system.

3. The real-time optimization method for cold storage stations based on AI self-learning according to claim 1, characterized in that, The constraint contribution value is calculated and generated by a piecewise linear mapping function based on the proportion value, inflection point parameter, and slope parameter of the preprocessed key operating quantities of the chiller station. The inflection point parameter is used to represent the upper limit of the normal adjustment range of the operating quantity. The slope parameter is used to represent the rate of increase of the constraint contribution after the proportion value exceeds the inflection point. The piecewise linear mapping function is used to characterize the relationship between the constraint contribution and the proportion value.

4. The real-time optimization method for cold storage stations based on AI self-learning according to claim 1, characterized in that, The step of scaling the preset nominal adjustable ranges of each dimension using the differential adjustment coefficients to obtain the constrained adjustable operating space set within the current control cycle is as follows: The nominal interval center and nominal half-width are pre-configured based on the aforementioned differential adjustment coefficient; The adjustable operating space of the current control cycle is the set of intervals centered on the nominal interval center and with the product of the differential adjustment coefficient and the nominal half-width as the half-width.

5. The real-time optimization method for cold storage stations based on AI self-learning according to claim 1, characterized in that, The incremental update method with inertia is specifically as follows: The current cycle's operation control command value is equal to the previous cycle's operation control command value plus an update coefficient multiplied by the difference between the current operation decision value and the previous cycle's operation control command value; the update coefficient is pre-configured according to the response characteristics of different types of execution devices.

6. A system for implementing the AI-based self-learning real-time optimization method for cold storage plants as described in claim 1, characterized in that, The system includes: The status calculation module is used to collect real-time data of key operating quantities of the cooling plant in each control cycle, preprocess the data, generate constraint contribution values ​​for each operating quantity, and calculate the operating constraint status based on the constraint contribution values ​​of each operating quantity; the operating constraint status is used to comprehensively describe the overall constraint strength within the current control cycle. The spatial analysis module is used to calculate the overall adjustment coefficient based on the operational constraint state, which represents the global scaling benchmark for differentiated tightening; based on the overall adjustment coefficient and the constraint contribution value of the operational quantity, it calculates the differentiated adjustment coefficient for each key operational quantity; and uses the differentiated adjustment coefficient to scale the preset nominal adjustable range of each dimension to obtain the set of constrained adjustable operational space within the current control cycle. The optimization decision module is used to calculate the operation decision value for the current cycle under the boundary constraints of the adjustable operation space set, taking the operation strategy executed in the previous cycle as the starting point and combining the adjustment suggestions given by the pre-built AI self-learning model. The calculation process introduces a boundary avoidance regularization term to make the operation decision value automatically move away from the interval edge when the adjustable operation space is narrow. The instruction execution module is used to convert the operation decision value into equipment control instructions that can be issued in the current cycle through an inertial incremental update method, and send them to the chiller station execution equipment to complete closed-loop control.

Images