Liquid cooling heat management system control method based on deep learning

Abstract

The invention discloses a liquid cooling heat management system control method based on deep learning. The method comprises the steps that S1, the temperature, the pressure difference, the flow, the pump rotating speed and the valve opening data sequence of a liquid cooling heat management system are collected; s2, constructing a loop topological parameter set, and generating an interconnection matrix and a dissipation matrix; s3, forming an expansion state vector by the physical state and the transport delay state; s4, constructing a target embedded vector, inputting the extended state vector, the target embedded vector and the loop topological parameter set into the improved PHNN model, and outputting an energy function gradient and a port matrix; s5, condensation and over-temperature safety items are introduced to form an effective energy function gradient; s6, performing multi-step state prediction under the constraint of the topological structure to obtain a prediction state sequence; and S7, the pump rotating speed and valve opening degree control instruction limited by the actuator and the safety constraint is output. The prediction precision, the operation safety and the energy consumption control performance of the liquid cooling heat management system are improved.

Description

Technical Field

[0001] This invention relates to the field of deep learning and liquid cooling thermal management control technology, and in particular to a control method for a liquid cooling thermal management system based on deep learning. Background Technology

[0002] With the increasing demands for thermal management performance in applications such as high-power electronic devices, data centers, and new energy vehicles, liquid-cooled thermal management systems are widely used due to their high heat exchange efficiency and accurate temperature control. Liquid-cooled thermal management systems typically form a closed-loop system consisting of cold plates, heat exchangers, piping, pumps, and valves to continuously cool critical components. Their control objectives include not only stable control of the cold plate outlet temperature and loop pressure difference, but also consideration of system energy consumption and operational safety constraints such as condensation and overheating. Therefore, achieving accurate modeling, predictive control, and safety constraint coordination of liquid-cooled thermal management systems under complex operating conditions has become an important research direction in this field.

[0003] Existing control methods for liquid-cooled thermal management systems mostly employ control strategies based on empirical models or simplified physical models, such as proportional-integral-derivative (PID) control, rule-table control, or linear model predictive control. These methods typically treat the liquid-cooling loop as a lumped-parameter system, neglecting pipeline transport delays, topological differences, and the effects of multi-actuator coupling. Under conditions of rapid load changes or complex loop structures, problems such as control lag and increased temperature fluctuations easily arise. Furthermore, while some studies introduce data-driven models or neural network methods to model the system, most only treat the system as a black-box mapping, failing to explicitly incorporate the physical topological constraints and energy dissipation characteristics of the liquid-cooling loop. Consequently, the model's generalization ability and stability are limited in practical applications.

[0004] Meanwhile, existing control methods typically employ threshold judgment or external logic switching to handle safety constraints. When condensation or overheating risks are detected, intervention is achieved through limiting, forced adjustment, or switching control modes. This approach separates safety constraints from the system dynamics, making it difficult to continuously reflect the impact of safety boundaries during predictive control. This can easily lead to abrupt changes in control actions and even system oscillations. Furthermore, for the transport delay effects caused by changes in pipe length and flow rate in liquid cooling systems, existing methods mostly use fixed delays or empirical compensation strategies, which cannot be dynamically adjusted according to operating conditions, affecting prediction accuracy and control effectiveness.

[0005] Therefore, how to provide a control method for a liquid cooling thermal management system based on deep learning is a problem that urgently needs to be solved by those skilled in the art. Summary of the Invention

[0006] One objective of this invention is to propose a control method for a liquid-cooled thermal management system based on deep learning. This invention combines the loop topology and transport delay characteristics of the liquid-cooled thermal management system with a deep learning model to construct an improved PHNN model that includes physical structural constraints. The control objective and safety constraints are introduced into the energy function gradient in an endogenous manner to achieve multi-step state prediction and generation of constrained control commands, thereby simultaneously improving temperature control accuracy, operational safety, and system energy consumption control performance under complex operating conditions.

[0007] A control method for a liquid-cooled thermal management system based on deep learning according to an embodiment of the present invention includes the following steps: S1. Collect the temperature, differential pressure, flow rate, pump speed and valve opening data sequence of the liquid cooling thermal management system; S2. Construct a set of loop topology parameters based on the connection relationships between pipes, valves, heat exchangers and cold plates in the liquid cooling loop, and generate an interconnection matrix and a dissipation matrix based on the set of loop topology parameters. S3. Construct a physical state vector from the temperature, pressure difference, and flow data sequences, and combine it with the transport delay state generated based on the pipeline length parameter and the flow data sequence to form an extended state vector; S4. Construct a target embedding vector based on the target temperature of the cold plate, the target loop pressure difference and the energy consumption weight parameters, and input the extended state vector, the target embedding vector and the loop topology parameter set into the improved PHNN model, and output the energy function gradient and port matrix. S5. Construct a condensation safety item based on dew point temperature data, preset condensation threshold and cold plate outlet temperature data, and construct an over-temperature safety item based on preset temperature upper limit and cold plate outlet temperature data. Combine the condensation safety item and the over-temperature safety item with the energy function gradient to form an effective energy function gradient. S6. Based on the interconnection matrix, dissipation matrix, effective energy function gradient and port matrix, the extended state vector is updated in multiple steps using the improved PHNN model to obtain the predicted state sequence. S7. Based on the predicted state sequence, a control quantity is formed, and under the conditions of actuator and safety constraints, control commands for pump speed and valve opening are output.

[0008] Optionally, S1 includes: S11. Temperature measurements are collected at the cold plate inlet, cold plate outlet, primary side inlet, and primary side outlet of the liquid cooling thermal management system, and a temperature data sequence is formed according to the control cycle. S12. Pressure measurements are collected at the pump outlet and the loop return location in the liquid cooling loop, and a loop differential pressure data sequence is formed based on the difference between the pressure measurements. S13. Collect coolant flow rate measurements on the main circuit pipeline of the liquid cooling circuit and form a flow data sequence according to the control cycle order; S14. Read the pump speed setting value from the pump control unit, form a pump speed data sequence according to the control cycle order, and read the proportional valve opening setting value from the valve control unit, form a valve opening data sequence according to the control cycle order. S15. The temperature data sequence, the loop pressure difference data sequence, the flow rate data sequence, the pump speed data sequence, and the valve opening data sequence are time-aligned according to a unified control cycle.

[0009] Optionally, S2 includes: S21. Based on the physical connection relationship between pipes, valves, heat exchangers and cold plates in the liquid cooling circuit, construct a set of circuit nodes and a set of circuit connections. The set of circuit nodes includes cold plate nodes, heat exchanger nodes and pipe junction nodes. The set of circuit connections includes the directed fluid connection relationship between nodes. S22. Based on the set of loop nodes and the set of loop connections, construct a set of loop topology parameters, which includes a node connection matrix, connection direction identification parameters, and loop branch identification parameters. S23. Construct an interconnection matrix based on the set of loop topology parameters. The interconnection matrix is ​​formed by the transpose difference between the first parameter matrix and the first parameter matrix, and the positions of the non-zero elements in the interconnection matrix correspond one-to-one with the positions of the non-zero elements in the node connection matrix. S24. Construct a dissipation matrix based on the set of loop topology parameters. The dissipation matrix is ​​formed by the product of the second parameter matrix and the transpose of the second parameter matrix, and the distribution of non-zero elements in the dissipation matrix corresponds to the loop branch identification parameter. S25. Use the interconnection matrix and the dissipation matrix as structural parameters of the improved PHNN model, and keep the distribution of non-zero elements of the interconnection matrix and the dissipation matrix unchanged during the initialization and update of the model parameters.

[0010] Optionally, S3 includes: S31. Based on the pipe connection relationship defined in the loop topology parameter set, the liquid cooling loop is divided into multiple pipe segments with definite start and end nodes, and a corresponding pipe length parameter is associated with each pipe segment. S32. Within each control cycle, based on the flow data sequence corresponding to the pipeline segment, obtain the flow parameters corresponding to the pipeline segment in the current control cycle, and combine them with the pipeline length parameters to form the transport time parameter sequence of the pipeline segment. S33. Based on the proportional relationship between the transport time parameter sequence and the control cycle duration, the transport time parameter sequence is mapped to a discrete delay index sequence, and the discrete delay index sequence is subjected to amplitude limiting processing based on the preset delay index upper limit and delay index lower limit. S34. For each pipeline segment, according to the discrete delay index sequence, select the corresponding physical state component from the historical control cycle corresponding to the discrete delay index, and form the selected physical state components into a pipeline segment transport state set according to the pipeline segment order. S35. The pipeline segment transport state set is concatenated with the physical state vector of the current control cycle according to the connection order defined in the loop topology parameter set to form an extended state vector containing the current state component and the historical transport component.

[0011] Optionally, S4 includes: S41. Obtain the target temperature parameters of the cold plate, the target loop pressure difference parameters, and the energy consumption weight parameters, and construct the target parameter vector in the order corresponding to the temperature state component, pressure difference state component, and flow-related state component in the extended state vector. S42. Based on the physical dimensions of each parameter component in the target parameter vector, perform component-level mapping processing on the target parameter vector; S43. Based on the parameter grouping relationship of the target parameter vector, construct a target embedding matrix group, which includes a first embedding matrix corresponding to the temperature state component, a second embedding matrix corresponding to the pressure difference state component, and a third embedding matrix corresponding to the energy consumption weight. S44. Perform matrix mapping on the corresponding parameter components in the target parameter vector using the first embedding matrix, the second embedding matrix and the third embedding matrix respectively to form a target embedding vector set grouped by state category; S45. The target embedding vector set is rearranged according to the order of each state component in the extended state vector, and then concatenated with the extended state vector and the loop topology parameter set to form a model input vector with state dimension alignment. S46. Input the model input vector into the improved PHNN model, generate the parameter matrix calculation result corresponding to the model input vector inside the model, and output the energy function gradient and port matrix corresponding to the current control cycle.

[0012] Optionally, S5 includes: S51. Acquire dew point temperature data, cold plate outlet temperature data, preset condensation threshold and preset temperature upper limit within the current control cycle. S52. Determine the condensation temperature difference value based on the cold plate outlet temperature data and the dew point temperature data, and compare the condensation temperature difference value with the preset condensation threshold value to obtain the condensation violation amount; S53. Determine the over-temperature difference value based on the preset upper temperature limit and the cold plate outlet temperature data, and compare the over-temperature difference value with zero to obtain the over-temperature violation amount; S54. Map the condensation violation amount and the over-temperature violation amount according to the preset monotonic mapping relationship, and determine the condensation weight coefficient and the over-temperature weight coefficient. The preset monotonic mapping relationship is given by a piecewise linear mapping table, and the condensation weight coefficient and the over-temperature weight coefficient are respectively limited between the preset weight lower limit and the preset weight upper limit. S55. Construct a condensation safety vector and an over-temperature safety vector according to the dimensional order of the temperature state components in the extended state vector, and write the product of the condensation violation amount and the condensation weight coefficient into the dimension corresponding to the cold plate outlet temperature of the condensation safety vector, and write the product of the over-temperature violation amount and the over-temperature weight coefficient into the dimension corresponding to the cold plate outlet temperature of the over-temperature safety vector. S56. The condensation safety vector and the overtemperature safety vector are superimposed dimension by dimension on the energy function gradient to obtain the effective energy function gradient.

[0013] Optionally, S6 includes: S61. Obtain the interconnection matrix, dissipation matrix, effective energy function gradient and port matrix within the current control cycle, and use the extended state vector as the initial state vector. S62. Calculate the state change vector for the initial state vector based on the interconnection matrix, dissipation matrix, effective energy function gradient, and port matrix. S63. Perform discrete-time integration on the state change vector according to the control cycle duration to obtain the state vector corresponding to the next prediction step. S64. Take the state vector corresponding to the next prediction step as the new state vector, and repeat the state change vector calculation and discrete time integration process to form a sequence of state vectors corresponding to multiple consecutive prediction steps. S65. During repeated execution, the structural parameters of the interconnection matrix and the dissipation matrix remain unchanged, and the gradient of the effective energy function is re-determined based on the state vector corresponding to each prediction step. S66. Arrange the state vectors corresponding to the consecutive prediction steps in chronological order to form a prediction state sequence.

[0014] Optionally, S7 includes: S71. Based on the predicted state sequence, extract the temperature state component, pressure difference state component and flow-related state component corresponding to each prediction step according to the prediction step order to form a predicted state set. S72. Based on the change of state components corresponding to each prediction step in the predicted state set, determine the control increment vector corresponding to each prediction step. The control increment vector corresponds to the pump speed state component and the valve opening state component in terms of dimension. S73. Apply actuator constraint processing to the control increment vector. The actuator constraint processing includes restricting the pump speed change rate, the pump speed range, the valve opening change rate, and the valve opening range item by item to obtain the constrained control increment vector. S74. The constrained control increment vector is combined with the pump speed setpoint and valve opening setpoint of the current control cycle to obtain the pump speed control command and valve opening control command corresponding to the next control cycle. S75. The pump speed control command and the valve opening control command are output to the corresponding actuators according to the control cycle.

[0015] The beneficial effects of this invention are: This invention integrates the loop topology and transport delay characteristics of a liquid-cooled thermal management system with a PHNN model to address the issues of temperature response lag, strong coupling of multiple actuators, and difficulty in coordinating safety constraints in liquid-cooled systems under complex operating conditions. First, it constructs a set of loop topology parameters based on the actual connection relationships between pipes, valves, heat exchangers, and cold plates, thereby generating interconnection and dissipation matrices. The physical structure and energy dissipation characteristics of the liquid-cooled loop are intrinsically introduced into the model. In the state modeling stage, transport delay states are generated by combining pipe length parameters and flow data, explicitly incorporating the propagation time sequence of the coolant in the loop into the extended state vector, achieving an approximate expression of the distributed parameter characteristics. In the control objective processing... In this invention, the target temperature of the cold plate, the target loop pressure difference, and the energy consumption weight parameters are constructed into a structurally aligned target embedding vector, allowing the control target to directly participate in the construction process of the energy function gradient, thereby avoiding the problem of separation between target constraints and system dynamics in traditional methods. Regarding safety constraints, condensation risk and overheating risk are transformed into safety vectors consistent with the state dimension and intrinsically superimposed onto the energy function gradient, ensuring that the safety boundary continuously influences the system evolution direction during multi-step prediction. In the prediction and control phase, an improved PHNN model is used to perform multi-step state prediction under fixed topological constraints, and control commands constrained by actuators are formed based on the predicted state sequence, achieving smooth adjustment of pump speed and valve opening. Through the above technical solutions, this invention improves the prediction accuracy and control stability of the liquid-cooled thermal management system for load changes while ensuring system physical consistency and operational safety, balancing temperature control performance and energy consumption constraints, and exhibiting good engineering applicability and robustness. Attached Figure Description

[0016] 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:

[0017] Figure 1 This is a schematic diagram of the overall process of a deep learning-based control method for a liquid-cooled thermal management system proposed in this invention. Figure 2 This is a schematic diagram of the process of multi-step prediction and control command generation based on the improved PHNN model in this invention; Figure 3 This is a schematic diagram of the loop topology and extended state construction of the liquid-cooled thermal management system in this invention. Detailed Implementation

[0018] 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.

[0019] refer to Figure 1-3 A control method for a liquid-cooled thermal management system based on deep learning includes the following steps: S1. Collect the temperature, differential pressure, flow rate, pump speed and valve opening data sequence of the liquid cooling thermal management system; S2. Construct a set of loop topology parameters based on the connection relationships between pipes, valves, heat exchangers and cold plates in the liquid cooling loop, and generate an interconnection matrix and a dissipation matrix based on the set of loop topology parameters. S3. Construct a physical state vector from the temperature, pressure difference, and flow data sequences, and combine it with the transport delay state generated based on the pipeline length parameter and the flow data sequence to form an extended state vector; S4. Construct a target embedding vector based on the target temperature of the cold plate, the target loop pressure difference and the energy consumption weight parameters, and input the extended state vector, the target embedding vector and the loop topology parameter set into the improved PHNN model, and output the energy function gradient and port matrix. S5. Construct a condensation safety item based on dew point temperature data, preset condensation threshold and cold plate outlet temperature data, and construct an over-temperature safety item based on preset temperature upper limit and cold plate outlet temperature data. Combine the condensation safety item and the over-temperature safety item with the energy function gradient to form an effective energy function gradient. S6. Based on the interconnection matrix, dissipation matrix, effective energy function gradient and port matrix, the extended state vector is updated in multiple steps using the improved PHNN model to obtain the predicted state sequence. S7. Based on the predicted state sequence, a control quantity is formed, and under the conditions of actuator and safety constraints, control commands for pump speed and valve opening are output.

[0020] The improved PHNN model used in this invention introduces a set of loop topology parameters and control target parameters to participate in model construction. The loop topology parameters are used to constrain the structural form of the interconnection matrix and the dissipation matrix, ensuring that the interconnection matrix always satisfies antisymmetry and the dissipation matrix always satisfies positive semi-definiteness, thereby ensuring that the system state evolution conforms to the laws of energy transfer and dissipation. At the same time, by explicitly incorporating the transport delay state into the state vector, the model can reflect the transmission time sequence of coolant in the pipeline within the discrete control cycle. In addition, by introducing the condensation safety term and the over-temperature safety term into the energy function gradient calculation process in the form of energy terms, the safety boundary constraints are endogenous to the model prediction process, thereby forming a control decision that simultaneously satisfies temperature control, energy consumption constraints, and condensation constraints in the multi-step prediction update stage.

[0021] In this embodiment, S1 includes: S11. Temperature measurements are collected at the cold plate inlet, cold plate outlet, primary side inlet, and primary side outlet of the liquid cooling thermal management system, and a temperature data sequence is formed according to the control cycle. S12. Pressure measurements are collected at the pump outlet and the loop return location in the liquid cooling loop, and a loop differential pressure data sequence is formed based on the difference between the pressure measurements. S13. Collect coolant flow rate measurements on the main circuit pipeline of the liquid cooling circuit and form a flow data sequence according to the control cycle order; S14. Read the pump speed setting value from the pump control unit, form a pump speed data sequence according to the control cycle order, and read the proportional valve opening setting value from the valve control unit, form a valve opening data sequence according to the control cycle order. S15. The temperature data sequence, the loop pressure difference data sequence, the flow rate data sequence, the pump speed data sequence, and the valve opening data sequence are time-aligned according to a unified control cycle.

[0022] In this embodiment, S2 includes: S21. Based on the physical connection relationship between pipes, valves, heat exchangers and cold plates in the liquid cooling circuit, construct a set of circuit nodes and a set of circuit connections. The set of circuit nodes includes cold plate nodes, heat exchanger nodes and pipe junction nodes. The set of circuit connections includes the directed fluid connection relationship between nodes. S22. Based on the set of loop nodes and the set of loop connections, construct a set of loop topology parameters, which includes a node connection matrix, connection direction identification parameters, and loop branch identification parameters. S23. Construct an interconnection matrix based on the set of loop topology parameters. The interconnection matrix is ​​formed by the transpose difference between the first parameter matrix and the first parameter matrix, and the positions of the non-zero elements in the interconnection matrix correspond one-to-one with the positions of the non-zero elements in the node connection matrix. S24. Construct a dissipation matrix based on the set of loop topology parameters. The dissipation matrix is ​​formed by the product of the second parameter matrix and the transpose of the second parameter matrix, and the distribution of non-zero elements in the dissipation matrix corresponds to the loop branch identification parameter. S25. Use the interconnection matrix and the dissipation matrix as structural parameters of the improved PHNN model, and keep the distribution of non-zero elements of the interconnection matrix and the dissipation matrix unchanged during the initialization and update of the model parameters.

[0023] In this invention, the set of loop nodes and the set of loop connections are not abstract graph structures, but are constructed by mapping one-to-one with the actual pipeline routing, valve placement, and installation relationship between heat exchangers and cold plates in the liquid cooling system. Non-zero elements in the node connection matrix correspond to actual fluid connection paths. The parameter matrix of the interconnection matrix only participates in parameter initialization and updating at the non-zero positions, and remains zero at the other positions, thus limiting the internal state coupling of the model to only occur between physically connected nodes. The low-rank parameter matrix of the dissipation matrix is ​​grouped and configured according to the loop branch identifier parameters, so that different loop branches correspond to different dissipation parameter sets. This structure remains unchanged during model training and online updates, thereby ensuring that the internal structure of the model is always consistent with the physical topology of the liquid cooling loop.

[0024] In this embodiment, S3 includes: S31. Based on the pipe connection relationship defined in the loop topology parameter set, the liquid cooling loop is divided into multiple pipe segments with definite start and end nodes, and a corresponding pipe length parameter is associated with each pipe segment. S32. Within each control cycle, based on the flow data sequence corresponding to the pipeline segment, obtain the flow parameters corresponding to the pipeline segment in the current control cycle, and combine them with the pipeline length parameters to form the transport time parameter sequence of the pipeline segment. S33. Based on the proportional relationship between the transport time parameter sequence and the control cycle duration, the transport time parameter sequence is mapped to a discrete delay index sequence, and the discrete delay index sequence is subjected to amplitude limiting processing based on the preset delay index upper limit and delay index lower limit. The control cycle duration is the time interval parameter for the liquid-cooled thermal management system to perform state sampling, model calculation, and control quantity update; S34. For each pipeline segment, according to the discrete delay index sequence, select the corresponding physical state component from the historical control cycle corresponding to the discrete delay index, and form the selected physical state components into a pipeline segment transport state set according to the pipeline segment order. S35. The pipeline segment transport state set is concatenated with the physical state vector of the current control cycle according to the connection order defined in the loop topology parameter set to form an extended state vector containing the current state component and the historical transport component.

[0025] Under the constraints of the loop topology parameter set, this invention divides the actual pipeline in the liquid cooling loop into several pipeline segments with defined start and end nodes, and uses the pipeline segments as the basic unit for transport modeling. For each pipeline segment, a transport time parameter sequence is formed based on the flow measurement value within the corresponding control cycle and the predetermined pipeline length parameter. The transport time parameter sequence is then discretized and mapped with the control cycle duration to obtain a delay index sequence aligned with the control cycle. The corresponding physical state component is selected from the historical control cycle through the delay index sequence, so that the formed transport state can reflect the propagation time sequence relationship of the coolant in the pipeline.

[0026] In this embodiment, S4 includes: S41. Obtain the target temperature parameters of the cold plate, the target loop pressure difference parameters, and the energy consumption weight parameters, and construct the target parameter vector in the order corresponding to the temperature state component, pressure difference state component, and flow-related state component in the extended state vector. The target circuit pressure difference is a pressure difference set value determined based on the liquid cooling circuit structural parameters and the stable operating range of the pump and valve; S42. Based on the physical dimensions of each parameter component in the target parameter vector, perform component-level mapping processing on the target parameter vector; S43. Based on the parameter grouping relationship of the target parameter vector, construct a target embedding matrix group, which includes a first embedding matrix corresponding to the temperature state component, a second embedding matrix corresponding to the pressure difference state component, and a third embedding matrix corresponding to the energy consumption weight. S44. Perform matrix mapping on the corresponding parameter components in the target parameter vector using the first embedding matrix, the second embedding matrix and the third embedding matrix respectively to form a target embedding vector set grouped by state category; S45. The target embedding vector set is rearranged according to the order of each state component in the extended state vector, and then concatenated with the extended state vector and the loop topology parameter set to form a model input vector with state dimension alignment. S46. Input the model input vector into the improved PHNN model, generate the parameter matrix calculation result corresponding to the model input vector inside the model, and output the energy function gradient and port matrix corresponding to the current control cycle.

[0027] In the model input construction stage, this invention arranges the target temperature of the cold plate, the target loop pressure difference, and the energy consumption weight in the same order as the state components in the extended state vector. Based on this, the corresponding numerical mapping intervals are pre-defined according to the physical dimensions of each parameter, so that different types of target parameters are consistent with the corresponding state components in terms of numerical scale. Furthermore, by constructing target embedding matrices according to state categories, the target parameters are mapped to a vector space that corresponds one-to-one with the state dimension in the energy function gradient. Under the input order limited by topological parameters, they are combined with the extended state vector, so that the control target participates in the gradient construction process inside the model in a structured manner, and achieves tight coupling between target constraints and state evolution process without changing the model topological structure parameters.

[0028] In this embodiment, S5 includes: S51. Acquire dew point temperature data, cold plate outlet temperature data, preset condensation threshold and preset temperature upper limit within the current control cycle. S52. Determine the condensation temperature difference value based on the cold plate outlet temperature data and the dew point temperature data, and compare the condensation temperature difference value with the preset condensation threshold value to obtain the condensation violation amount; The preset condensation threshold is a threshold parameter corresponding to the minimum temperature difference between the cold plate outlet temperature and the dew point temperature, determined based on the allowable condensation conditions on the surface of the cold plate. When the condensation temperature difference is less than the preset condensation threshold, the condensation violation is the difference between the preset condensation threshold and the condensation temperature difference; when the condensation temperature difference is not less than the preset condensation threshold, the condensation violation is zero. S53. Determine the over-temperature difference value based on the preset upper temperature limit and the cold plate outlet temperature data, and compare the over-temperature difference value with zero to obtain the over-temperature violation amount; When the cold plate outlet temperature data is greater than the preset temperature upper limit, the over-temperature violation amount is the difference between the cold plate outlet temperature data and the preset temperature upper limit; when the cold plate outlet temperature data is not greater than the preset temperature upper limit, the over-temperature violation amount is zero. S54. Map the condensation violation amount and the over-temperature violation amount according to the preset monotonic mapping relationship, and determine the condensation weight coefficient and the over-temperature weight coefficient. The preset monotonic mapping relationship is given by a piecewise linear mapping table, and the condensation weight coefficient and the over-temperature weight coefficient are respectively limited between the preset weight lower limit and the preset weight upper limit. S55. Construct a condensation safety vector and an over-temperature safety vector according to the dimensional order of the temperature state components in the extended state vector, and write the product of the condensation violation amount and the condensation weight coefficient into the dimension corresponding to the cold plate outlet temperature of the condensation safety vector, and write the product of the over-temperature violation amount and the over-temperature weight coefficient into the dimension corresponding to the cold plate outlet temperature of the over-temperature safety vector. S56. The condensation safety vector and the overtemperature safety vector are superimposed dimension by dimension on the energy function gradient to obtain the effective energy function gradient.

[0029] This invention unifies the condensation risk and overheating risk into a safety vector with the same dimension as the extended state vector within the model. The non-zero components of the safety vector correspond only to the state related to the cold plate outlet temperature, thus limiting the scope of safety impact from spreading to irrelevant states such as flow rate and pressure difference. By directly writing the product of the violation quantity and the weight coefficient into the corresponding dimension, the safety impact is made consistent with the state dimension in terms of numerical value and can continuously participate in the gradient calculation process. The safety vector and the energy function gradient are synthesized by a dimension-by-dimensional superposition method, so that the model considers both thermal management objectives and safety boundary constraints when updating the state.

[0030] In this embodiment, S6 includes: S61. Obtain the interconnection matrix, dissipation matrix, effective energy function gradient and port matrix within the current control cycle, and use the extended state vector as the initial state vector. S62. Calculate the state change vector for the initial state vector based on the interconnection matrix, dissipation matrix, effective energy function gradient, and port matrix. S63. Perform discrete-time integration on the state change vector according to the control cycle duration to obtain the state vector corresponding to the next prediction step. S64. Take the state vector corresponding to the next prediction step as the new state vector, and repeat the state change vector calculation and discrete time integration process to form a sequence of state vectors corresponding to multiple consecutive prediction steps. S65. During repeated execution, the structural parameters of the interconnection matrix and the dissipation matrix remain unchanged, and the gradient of the effective energy function is re-determined based on the state vector corresponding to each prediction step. S66. Arrange the state vectors corresponding to the consecutive prediction steps in chronological order to form a prediction state sequence.

[0031] In this invention, an extended state vector is used as the initial state during state prediction. Under the constraints of a fixed interconnection matrix and dissipation matrix structure, the state evolution is gradually advanced through discrete-time integration. Each prediction step redetermines the corresponding effective energy function gradient based on the current prediction state, thereby ensuring that safety constraints and target embedding remain effective throughout the prediction process. Multi-step prediction uses a sequential iterative approach to generate the state sequence, avoiding updates or retraining of model parameters. The prediction process only involves the recursive calculation of state variables and gradient values. By keeping the structural parameters unchanged and gradually updating the state-related gradients, the prediction process achieves this.

[0032] In this embodiment, S7 includes: S71. Based on the predicted state sequence, extract the temperature state component, pressure difference state component and flow-related state component corresponding to each prediction step according to the prediction step order to form a predicted state set. S72. Based on the change of state components corresponding to each prediction step in the predicted state set, determine the control increment vector corresponding to each prediction step. The control increment vector corresponds to the pump speed state component and the valve opening state component in terms of dimension. S73. Apply actuator constraint processing to the control increment vector. The actuator constraint processing includes restricting the pump speed change rate, the pump speed range, the valve opening change rate, and the valve opening range item by item to obtain the constrained control increment vector. S74. The constrained control increment vector is combined with the pump speed setpoint and valve opening setpoint of the current control cycle to obtain the pump speed control command and valve opening control command corresponding to the next control cycle. S75. The pump speed control command and the valve opening control command are output to the corresponding actuators according to the control cycle.

[0033] In the control quantity generation stage, this invention uses a multi-step predicted state sequence as input. By aligning the state changes of each predicted step, the control increment corresponds one-to-one with the pump speed and valve opening state in terms of dimension. After the control increment is formed, actuator physical constraints are introduced to limit the rate of change and the range of values, thereby ensuring that the control command meets the actuator reachability conditions at the numerical level. The control command is generated by incrementally synthesizing the current set value to avoid abruptly driving the actuator. Through the above processing, the predicted state information is transformed into a continuous and constrained control command sequence, which is directly used for the execution of subsequent control cycles.

[0034] Example 1: To verify the feasibility and effectiveness of this invention in practical engineering, the deep learning-based liquid cooling thermal management system control method proposed in this invention was applied to the liquid cooling thermal management scenario of a high-power data processing device. This device integrates multiple high-performance processing units, with each unit generating over 800W of heat under full load, placing high demands on the temperature control accuracy, response speed, and operational safety of the cooling system. The device adopts a typical closed-loop liquid cooling structure, with the liquid cooling loop including a cold plate, heat exchanger, circulating pump, proportional valve, and multiple connecting pipelines. The cooling medium is a mixture of deionized water and antifreeze. During actual operation, due to the periodic fluctuations in the computational load, the cold plate outlet temperature, loop flow rate, and differential pressure change frequently with the load. Traditional control methods are prone to problems such as temperature overshoot, response lag, and difficulty in timely mitigation of condensation risks.

[0035] In this embodiment, temperature sensors are first placed at the cold plate inlet, cold plate outlet, and primary side inlet and outlet of the heat exchanger; pressure sensors are placed at the pump outlet and loop return location; and flow sensors are placed in the main loop pipeline. Simultaneously, pump speed setpoints and valve opening setpoints are acquired in real time from the pump control unit and valve control unit. The system uses a 1-second control cycle and performs time alignment processing on various data types to form a continuous data sequence as model input. Based on the actual structure of the liquid cooling system, a loop topology parameter set including cold plate nodes, heat exchanger nodes, and pipeline junction nodes is constructed, and interconnection and dissipation matrices are generated from this set to constrain the state coupling relationships within the model.

[0036] In the state modeling process, the collected temperature, pressure difference, and flow rate data are used to construct a physical state vector. Simultaneously, based on the geometric length parameters of each pipe segment and the corresponding flow rate data, the transport time of the coolant in the pipe is calculated, and this transport time is discretized into delay indices to construct transport delay states, thus forming an extended state vector. In this way, the response hysteresis caused by the coolant propagation in the loop is explicitly introduced into the model state, enabling the model to anticipate the impact of load changes on the cold plate outlet temperature.

[0037] Regarding control target settings, based on the equipment's allowable operating temperature range, the target outlet temperature of the cold plate is set to 45℃, and the target differential pressure in the loop is set to 35kPa. Energy consumption weighting parameters are also set to balance the energy consumption impact of changes in pump speed and valve opening. These target parameters are mapped to target embedding vectors aligned with the state dimension via a target embedding matrix, and input into the improved PHNN model along with the extended state vector and loop topology parameters. The model outputs the energy function gradient and port matrix corresponding to the current control cycle, and further introduces safety constraints based on these.

[0038] During the safety constraint processing, the system acquires ambient dew point temperature data in real time and calculates the condensation temperature difference by combining it with the cold plate outlet temperature. Simultaneously, it constructs a condensation safety term based on a preset condensation threshold. On the other hand, it constructs an over-temperature safety term based on the cold plate outlet temperature and a preset upper temperature limit. The condensation safety term and the over-temperature safety term are superimposed as a vector into the energy function gradient, enabling the safety constraints to participate endogenously in subsequent multi-step prediction processes, thereby suppressing potential risks during the prediction stage.

[0039] In the prediction and control phase, based on the interconnection matrix, dissipation matrix, effective energy function gradient, and port matrix, the extended state vector is updated in multiple steps to obtain the predicted state sequence for multiple control cycles. The system generates control increments based on the changing trends of temperature, differential pressure, and flow rate in the predicted states. After applying constraints on the pump speed change rate and range, and the valve opening change rate and range, smooth pump speed and valve opening control commands are generated. These control commands are output to the actuators in real time and continue to participate in the control process as setpoints in the next control cycle.

[0040] To verify the beneficial effects of this invention, a comparative experiment was conducted for 72 hours of continuous operation, comparing the method of this invention with a traditional PID-based liquid cooling control method. During the experiment, the calculated load fluctuated periodically between 30% and 100%, and the ambient temperature varied between 20℃ and 28℃. The experimental results are shown in Table 1.

[0041] Table 1. Comparison of Operating Performance of Liquid Cooling Thermal Management System under Different Control Methods

[0042] As shown in Table 1, this invention demonstrates significant advantages in terms of cold plate outlet temperature control accuracy, response speed, and safety. Under the same load conditions, the method of this invention reduces the maximum deviation of the cold plate outlet temperature by approximately 66% and shortens the temperature stabilization time by more than 50%, effectively suppressing temperature overshoot caused by transport delay. Simultaneously, through target embedding and energy gradient adjustment mechanisms, the average pump speed and overall system power consumption are significantly reduced while maintaining temperature control accuracy, demonstrating excellent energy consumption control. Regarding safety, since condensation and over-temperature risks are introduced into the state update process during the prediction phase, no condensation risk triggering or over-temperature alarm occurred during the experiment, significantly improving system operational stability.

[0043] In summary, this embodiment demonstrates that the deep learning-based liquid cooling thermal management system control method proposed in this invention can achieve forward-looking prediction and smooth control of the liquid cooling system state under complex load variation conditions. It improves temperature control accuracy while taking into account energy consumption and operational safety, and has strong engineering practical value.

[0044] 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 control method for a liquid-cooled thermal management system based on deep learning, characterized in that, Includes the following steps: S1. Collect the temperature, differential pressure, flow rate, pump speed and valve opening data sequence of the liquid cooling thermal management system; S2. Construct a set of loop topology parameters based on the connection relationships between pipes, valves, heat exchangers and cold plates in the liquid cooling loop, and generate an interconnection matrix and a dissipation matrix based on the set of loop topology parameters. S3. Construct a physical state vector from the temperature, pressure difference, and flow data sequences, and combine it with the transport delay state generated based on the pipeline length parameter and the flow data sequence to form an extended state vector; S4. Construct a target embedding vector based on the target temperature of the cold plate, the target loop pressure difference and the energy consumption weight parameters, and input the extended state vector, the target embedding vector and the loop topology parameter set into the improved PHNN model, and output the energy function gradient and port matrix. S5. Construct a condensation safety item based on dew point temperature data, preset condensation threshold and cold plate outlet temperature data, and construct an over-temperature safety item based on preset temperature upper limit and cold plate outlet temperature data. Combine the condensation safety item and the over-temperature safety item with the energy function gradient to form an effective energy function gradient. S6. Based on the interconnection matrix, dissipation matrix, effective energy function gradient and port matrix, the extended state vector is updated in multiple steps using the improved PHNN model to obtain the predicted state sequence. S7. Based on the predicted state sequence, a control quantity is formed, and under the conditions of actuator and safety constraints, control commands for pump speed and valve opening are output.

2. The control method for a liquid-cooled thermal management system based on deep learning according to claim 1, characterized in that, S1 includes: S11. Temperature measurements are collected at the cold plate inlet, cold plate outlet, primary side inlet, and primary side outlet of the liquid cooling thermal management system, and a temperature data sequence is formed according to the control cycle. S12. Pressure measurements are collected at the pump outlet and the loop return location in the liquid cooling loop, and a loop differential pressure data sequence is formed based on the difference between the pressure measurements. S13. Collect coolant flow rate measurements on the main circuit pipeline of the liquid cooling circuit and form a flow data sequence according to the control cycle order; S14. Read the pump speed setting value from the pump control unit, form a pump speed data sequence according to the control cycle order, and read the proportional valve opening setting value from the valve control unit, form a valve opening data sequence according to the control cycle order. S15. The temperature data sequence, the loop pressure difference data sequence, the flow rate data sequence, the pump speed data sequence, and the valve opening data sequence are time-aligned according to a unified control cycle.

3. The control method for a liquid-cooled thermal management system based on deep learning according to claim 1, characterized in that, S2 includes: S21. Based on the physical connection relationship between pipes, valves, heat exchangers and cold plates in the liquid cooling circuit, construct a set of circuit nodes and a set of circuit connections. The set of circuit nodes includes cold plate nodes, heat exchanger nodes and pipe junction nodes. The set of circuit connections includes the directed fluid connection relationship between nodes. S22. Based on the set of loop nodes and the set of loop connections, construct a set of loop topology parameters, which includes a node connection matrix, connection direction identification parameters, and loop branch identification parameters. S23. Construct an interconnection matrix based on the set of loop topology parameters. The interconnection matrix is ​​formed by the transpose difference between the first parameter matrix and the first parameter matrix, and the positions of the non-zero elements in the interconnection matrix correspond one-to-one with the positions of the non-zero elements in the node connection matrix. S24. Construct a dissipation matrix based on the set of loop topology parameters. The dissipation matrix is ​​formed by the product of the second parameter matrix and the transpose of the second parameter matrix, and the distribution of non-zero elements in the dissipation matrix corresponds to the loop branch identification parameter. S25. Use the interconnection matrix and the dissipation matrix as structural parameters of the improved PHNN model, and keep the distribution of non-zero elements of the interconnection matrix and the dissipation matrix unchanged during the initialization and update of the model parameters.

4. The control method for a liquid-cooled thermal management system based on deep learning according to claim 1, characterized in that, S3 includes: S31. Based on the pipe connection relationship defined in the loop topology parameter set, the liquid cooling loop is divided into multiple pipe segments with definite start and end nodes, and a corresponding pipe length parameter is associated with each pipe segment. S32. Within each control cycle, based on the flow data sequence corresponding to the pipeline segment, obtain the flow parameters corresponding to the pipeline segment in the current control cycle, and combine them with the pipeline length parameters to form the transport time parameter sequence of the pipeline segment. S33. Based on the proportional relationship between the transport time parameter sequence and the control cycle duration, the transport time parameter sequence is mapped to a discrete delay index sequence, and the discrete delay index sequence is subjected to amplitude limiting processing based on the preset delay index upper limit and delay index lower limit. S34. For each pipeline segment, according to the discrete delay index sequence, select the corresponding physical state component from the historical control cycle corresponding to the discrete delay index, and form the selected physical state components into a pipeline segment transport state set according to the pipeline segment order. S35. The pipeline segment transport state set is concatenated with the physical state vector of the current control cycle according to the connection order defined in the loop topology parameter set to form an extended state vector containing the current state component and the historical transport component.

5. The control method for a liquid-cooled thermal management system based on deep learning according to claim 1, characterized in that, S4 includes: S41. Obtain the target temperature parameters of the cold plate, the target loop pressure difference parameters, and the energy consumption weight parameters, and construct the target parameter vector in the order corresponding to the temperature state component, pressure difference state component, and flow-related state component in the extended state vector. S42. Based on the physical dimensions of each parameter component in the target parameter vector, perform component-level mapping processing on the target parameter vector; S43. Based on the parameter grouping relationship of the target parameter vector, construct a target embedding matrix group, which includes a first embedding matrix corresponding to the temperature state component, a second embedding matrix corresponding to the pressure difference state component, and a third embedding matrix corresponding to the energy consumption weight. S44. Perform matrix mapping on the corresponding parameter components in the target parameter vector using the first embedding matrix, the second embedding matrix and the third embedding matrix respectively to form a target embedding vector set grouped by state category; S45. The target embedding vector set is rearranged according to the order of each state component in the extended state vector, and then concatenated with the extended state vector and the loop topology parameter set to form a model input vector with state dimension alignment. S46. Input the model input vector into the improved PHNN model, generate the parameter matrix calculation result corresponding to the model input vector inside the model, and output the energy function gradient and port matrix corresponding to the current control cycle.

6. The control method for a liquid-cooled thermal management system based on deep learning according to claim 1, characterized in that, S5 includes: S51. Acquire dew point temperature data, cold plate outlet temperature data, preset condensation threshold and preset temperature upper limit within the current control cycle. S52. Determine the condensation temperature difference value based on the cold plate outlet temperature data and the dew point temperature data, and compare the condensation temperature difference value with the preset condensation threshold value to obtain the condensation violation amount; S53. Determine the over-temperature difference value based on the preset upper temperature limit and the cold plate outlet temperature data, and compare the over-temperature difference value with zero to obtain the over-temperature violation amount; S54. Map the condensation violation amount and the over-temperature violation amount according to the preset monotonic mapping relationship, and determine the condensation weight coefficient and the over-temperature weight coefficient. The preset monotonic mapping relationship is given by a piecewise linear mapping table, and the condensation weight coefficient and the over-temperature weight coefficient are respectively limited between the preset weight lower limit and the preset weight upper limit. S55. Construct a condensation safety vector and an over-temperature safety vector according to the dimensional order of the temperature state components in the extended state vector, and write the product of the condensation violation amount and the condensation weight coefficient into the dimension corresponding to the cold plate outlet temperature of the condensation safety vector, and write the product of the over-temperature violation amount and the over-temperature weight coefficient into the dimension corresponding to the cold plate outlet temperature of the over-temperature safety vector. S56. The condensation safety vector and the overtemperature safety vector are superimposed dimension by dimension on the energy function gradient to obtain the effective energy function gradient.

7. The control method for a liquid-cooled thermal management system based on deep learning according to claim 1, characterized in that, S6 includes: S61. Obtain the interconnection matrix, dissipation matrix, effective energy function gradient and port matrix within the current control cycle, and use the extended state vector as the initial state vector. S62. Calculate the state change vector for the initial state vector based on the interconnection matrix, dissipation matrix, effective energy function gradient, and port matrix. S63. Perform discrete-time integration on the state change vector according to the control cycle duration to obtain the state vector corresponding to the next prediction step. S64. Take the state vector corresponding to the next prediction step as the new state vector, and repeat the state change vector calculation and discrete time integration process to form a sequence of state vectors corresponding to multiple consecutive prediction steps. S65. During repeated execution, the structural parameters of the interconnection matrix and the dissipation matrix remain unchanged, and the gradient of the effective energy function is re-determined based on the state vector corresponding to each prediction step. S66. Arrange the state vectors corresponding to the consecutive prediction steps in chronological order to form a prediction state sequence.

8. The control method for a liquid-cooled thermal management system based on deep learning according to claim 1, characterized in that, S7 includes: S71. Based on the predicted state sequence, extract the temperature state component, pressure difference state component and flow-related state component corresponding to each prediction step according to the prediction step order to form a predicted state set. S72. Based on the change of state components corresponding to each prediction step in the predicted state set, determine the control increment vector corresponding to each prediction step. The control increment vector corresponds to the pump speed state component and the valve opening state component in terms of dimension. S73. Apply actuator constraint processing to the control increment vector. The actuator constraint processing includes restricting the pump speed change rate, the pump speed range, the valve opening change rate, and the valve opening range item by item to obtain the constrained control increment vector. S74. The constrained control increment vector is combined with the pump speed setpoint and valve opening setpoint of the current control cycle to obtain the pump speed control command and valve opening control command corresponding to the next control cycle. S75. The pump speed control command and the valve opening control command are output to the corresponding actuators according to the control cycle.

Images