Liquid cooling fusion energy consumption regulation optimization method and system based on dynamic regulation

Abstract

The present invention provides a liquid cooling fusion energy consumption regulation optimization method and system based on dynamic regulation, including: S1: obtaining computing equipment heat load data, liquid cooling system status data, air cooling system status data and system constraint data, and constructing a cooling fusion real-time status map; S2: calibrating the cooling prediction model based on the cooling fusion real-time status map. After the calibration is completed, extracting the cooling key dynamic feature vector, and obtaining short-term cooling prediction data based on the cooling prediction model and the cooling fusion real-time status map, and constructing a cooling conflict objective function; S3: obtaining a cooling random sample set, initializing the PSO algorithm based on the cooling random sample set, and generating an optimal control parameter combination; S4: controlling the cooling system in real time based on the optimal control parameter combination, and cyclically executing steps S1 to S4. The above method can find a better energy consumption control in real time while the cooling system maintains its cooling capacity.

Description

Technical Field

[0001] The present invention relates to the field of liquid cooling control technology, and in particular to a liquid cooling fusion energy consumption regulation optimization method and system based on dynamic regulation. Background Art

[0002] As electronic equipment rapidly develops toward high power and high density, the balance between cooling system performance and energy consumption has become a key factor restricting equipment operating efficiency. From server clusters in data centers to power devices in new energy vehicles and precision instruments in the industrial control field, if the heat generated by equipment during operation cannot be dissipated in a timely manner, it can lead to performance degradation at best and malfunction or even device burnout at worst. Cooling systems that combine liquid and air cooling have become the mainstream solution for high-power density heat dissipation in data centers. However, in the context of carbon conservation, the proportion of cooling energy consumption continues to rise, typically accounting for 30-40% of the total energy consumption of data centers. Existing cooling systems that combine liquid and air cooling still face challenges.

[0003] Specifically, traditional liquid-cooling and air-cooling fusion cooling systems often rely on single-objective models or PID control, which faces limitations in scenarios with dynamic load fluctuations. For example, when a GPU cluster encounters sudden computing tasks, PID control uses a more aggressive strategy to force cooling at the expense of energy efficiency, or uses a single-objective model that lacks coupled modeling of multiple control objectives, such as only minimizing the power consumption of the pump. Therefore, there is an urgent need to construct an optimization method that can find the best energy consumption control in real time while ensuring that the cooling system maintains its cooling capacity. Summary of the Invention

[0004] In view of the above-mentioned problems, in combination with the first aspect of the present invention, an embodiment of the present invention provides a method for optimizing energy consumption regulation of liquid-cooled fusion based on dynamic regulation, the method comprising:

[0005] S1: Obtain computing equipment thermal load data, liquid cooling system status data, air cooling system status data, and system constraint data, and build a real-time cooling fusion status map;

[0006] S2: Obtain a cooling prediction model and calibrate the cooling prediction model based on the cooling fusion real-time state map. After calibration, extract the cooling key dynamic feature vectors, obtain short-term cooling prediction data based on the cooling prediction model and the cooling fusion real-time state map, and construct a cooling conflict objective function.

[0007] S3: Obtain a cooling random sample set, initialize the PSO algorithm based on the cooling random sample set, and generate an optimal control parameter combination through the cooling conflict objective function;

[0008] S4: Based on the optimal control parameter combination, the cooling system is controlled in real time, and steps S1 to S4 are executed in a loop.

[0009] As a further solution of the present invention, computing equipment heat load data, liquid cooling system status data, air cooling system status data, and system constraint data are obtained, and a cooling fusion real-time status map is constructed, including:

[0010] Acquire a raw sensor signal stream based on the sensor module, acquire computing device thermal load data, liquid cooling system status data, and air cooling system status data based on the raw sensor signal stream, and acquire system constraint data based on an initial cooling system design plan;

[0011] Setting a specific time window for the computing device heat load data, the liquid cooling system status data, and the air cooling system status data, and aligning the timestamps. After the setting is completed, an original cooling status data set is constructed;

[0012] The original cooling state dataset is preprocessed to generate a spatially correlated cooling state dataset. The temporal cooling feature vector is extracted at the same time. A cooling fusion real-time state map is constructed based on the spatially correlated cooling state dataset and the temporal cooling feature vector.

[0013] As a further solution of the present invention, a data preprocessing operation is performed on the original cooling state data set to generate a spatial correlation cooling state data set, and a temporal cooling feature vector is extracted. A cooling fusion real-time state map is constructed based on the spatial correlation cooling state data set and the temporal cooling feature vector, including:

[0014] Performing a preprocessing operation on the original cooling state data set, including at least missing filling, anomaly detection, and normalization operations, and introducing system constraint data during the preprocessing operation to obtain a preprocessed cooling state data set;

[0015] Obtain a physical layout topology diagram of the cooling system, perform spatial mapping and regional association on the preprocessed cooling state dataset based on the physical layout topology diagram of the cooling system, and generate a cooling space state dataset;

[0016] Extract the cooling time series feature vector of the cooling space state data set, perform regional state fusion based on the cooling time series feature vector and the cooling space state data set, and output cooling key state indicator data and cooling low-dimensional feature data vector;

[0017] A cooling fusion real-time status map is constructed based on cooling key status indicator data and cooling low-dimensional feature data vectors.

[0018] As a further solution of the present invention, a cooling prediction model is obtained, and a calibration operation is performed on the cooling prediction model based on the cooling fusion real-time state map. After the calibration is completed, the cooling key dynamic feature vector is extracted, and short-term cooling prediction data is obtained based on the cooling prediction model and the cooling fusion real-time state map, and a cooling conflict objective function is constructed, including:

[0019] Obtain a pre-trained cooling prediction model, input the cooling low-dimensional feature data vector in the cooling fusion real-time state map of the historical window into the cooling prediction model, predict the cooling low-dimensional feature data vector of the current window based on the cooling prediction model, compare the predicted cooling low-dimensional feature data vector of the current window with the actual cooling low-dimensional feature data vector of the current window, and calibrate the cooling prediction model based on the optimization algorithm;

[0020] Import the cooling low-dimensional feature data vector of the current window into the calibrated cooling prediction model, perform relationship extraction on the cooling prediction model, obtain the cooling key dynamic feature vector, and obtain short-term cooling prediction data based on the cooling prediction model;

[0021] The cooling conflict objective function is constructed based on the cooling key dynamic feature vectors, short-term cooling prediction data and cooling fusion real-time state map.

[0022] As a further solution of the present invention, a cooling conflict objective function is constructed based on the cooling key dynamic feature vector, the short-term cooling prediction data and the cooling fusion real-time state map, including:

[0023] Generate risk penalty function based on cooling key dynamic feature vector and short-time cooling prediction data;

[0024] Energy consumption component modeling is performed based on the cooling fusion real-time state map, cooling key dynamic feature vectors and short-term cooling prediction data, and a total energy consumption model is constructed based on the energy consumption component modeling;

[0025] Based on the risk penalty function and the total energy consumption model, a preliminary cooling conflict objective function is constructed, and a trade-off weight coefficient is introduced in the process of constructing the preliminary cooling conflict objective function;

[0026] The cooling preliminary conflict objective function is subjected to constraint addition processing, and a cooling conflict objective function is obtained based on the constraint addition processing.

[0027] As a further solution of the present invention, constraint addition processing is performed on the cooling preliminary conflict objective function, and the cooling conflict objective function is obtained based on the constraint addition processing:

[0028] Acquiring hard constraint data based on a cooling fusion real-time state map, and performing hard constraint processing on the cooling preliminary conflict objective function based on the hard constraint data;

[0029] Soft constraint data is acquired, and soft constraint processing is performed on the cooling preliminary conflict objective function based on the soft constraint data.

[0030] As a further solution of the present invention, a cooling random sample set is obtained, and a PSO algorithm is initialized based on the cooling random sample set. The PSO algorithm generates an optimal control parameter combination through a cooling conflict objective function, including:

[0031] Obtain a random sampling range based on the cooling conflict objective function, and perform sampling based on the random sampling range using a Monte Carlo algorithm to obtain a cooling random sample set;

[0032] Importing the cooling random sample set into the PSO algorithm and performing an initialization operation on the PSO algorithm, wherein the PSO algorithm selects a cooling random sample from the cooling random sample set for initialization, and obtains an initial particle position set of the cooling random sample;

[0033] The cooling conflict objective function is introduced into the PSO algorithm, and the initial particle position set is optimized and updated iteratively based on the PSO algorithm. When the update iteration converges, the optimal control parameter combination is output.

[0034] As a further solution of the present invention, a random sampling range is obtained based on the cooling conflict objective function, and a cooling random sample set is obtained by sampling based on the random sampling range using a Monte Carlo algorithm, including:

[0035] Determine a set of control variables to be optimized based on the cooling conflict objective function, obtain an upper limit of a feasibility range and a lower limit of a possibility range based on the set of control variables to be optimized, and obtain the random sampling range based on the upper limit of the feasibility range and the lower limit of the possibility range;

[0036] The random sampling range is introduced into the Monte Carlo algorithm, the total cooling random samples are obtained based on the Monte Carlo algorithm, the total cooling random samples are randomly selected based on the selection method of the Monte Carlo algorithm, and the randomly selected cooling random samples are combined into a cooling random sample set.

[0037] As a further solution of the present invention, the cooling conflict objective function is introduced into the PSO algorithm, and the initial particle position set is optimized and updated iteratively based on the PSO algorithm. When the update iteration converges, the optimal control parameter combination is output, including:

[0038] Performing particle evaluation on the initial particle position set based on the PSO algorithm;

[0039] Performing empirical updating on the initial particle position set based on the PSO algorithm;

[0040] Performing motion update on the initial particle position set based on the PSO algorithm;

[0041] When the PSO algorithm updates and iterates the initial particle position set, the convergence is checked. If converged, the update iteration is stopped and the optimal control parameter combination is output. If not converged, the update iteration is continued.

[0042] In another aspect, an embodiment of the present invention further provides a liquid cooling fusion energy consumption adjustment and optimization system based on dynamic adjustment, the system comprising:

[0043] An acquisition module, the acquisition module being used to acquire computing device thermal load data, liquid cooling system status data, air cooling system status data, and system constraint data;

[0044] A map module, which is used to construct a real-time state map of cooling and fusion;

[0045] A model module, the model module is used to extract key dynamic feature vectors of cooling and obtain short-term cooling prediction data, and to construct a cooling conflict objective function;

[0046] An optimization module, wherein the optimization module generates an optimal control parameter combination based on a PSO algorithm;

[0047] A control module, wherein the control module controls the cooling system in real time based on an optimal control parameter combination;

[0048] A loop module is used to execute steps S1 to S4 in a loop.

[0049] Based on the above aspects, the embodiment of the present application obtains the thermal load data of the computing device, the liquid cooling system status data, the air cooling system status data and the system constraint data to obtain the most basic operating parameters of the cooling system, imposes interval restrictions on the operating parameters, and constructs a cooling fusion real-time status map. Furthermore, the cooling prediction model is calibrated based on the cooling fusion real-time status map, so that the cooling prediction model that may be offline can operate in a relatively standard state after it goes online. The cooling key dynamic feature vectors and short-time cooling prediction data are extracted through the cooling prediction model, and the cooling conflict objective function is constructed based on the cooling key dynamic feature vectors and the short-time cooling prediction data, so as to model the relationship between energy consumption and cooling capacity. On this basis, a cooling random sample set is obtained through the Monte Carlo algorithm to ensure the diversity of the cooling random sample set. The optimal control parameter combination is obtained based on the PSO algorithm and the cooling conflict objective function, and this is executed in a loop, so that the method can find a better energy consumption control in real time while the cooling system maintains its cooling capacity. BRIEF DESCRIPTION OF THE DRAWINGS

[0050] Figure 1It is a schematic diagram of the execution flow of the liquid cooling fusion energy consumption regulation optimization method based on dynamic regulation provided in an embodiment of the present invention.

[0051] Figure 2 It is a schematic diagram of a liquid cooling fusion energy consumption regulation and optimization system based on dynamic regulation provided in an embodiment of the present invention. DETAILED DESCRIPTION

[0052] The drawings in the embodiments clearly and completely describe the technical solutions in the embodiments of the present invention. Obviously, the described embodiments are only some embodiments of the present invention, not all embodiments. Generally, the components of the embodiments of the present invention described and shown in the drawings herein can be arranged and designed in various different configurations.

[0053] Therefore, the following detailed description of the embodiments of the present invention provided in the accompanying drawings is not intended to limit the scope of the invention as claimed, but rather merely represents selected embodiments of the present invention. All other embodiments derived by persons of ordinary skill in the art based on the embodiments of the present invention without creative effort shall fall within the scope of protection of the present invention.

[0054] The present invention will be described in detail below with reference to the accompanying drawings. Figure 1 This is a flow chart of a liquid-cooled fusion energy consumption regulation optimization method based on dynamic adjustment provided by an embodiment of the present invention. The liquid-cooled fusion energy consumption regulation optimization method based on dynamic adjustment is introduced in detail below.

[0055] Specifically, the liquid cooling fusion energy consumption adjustment optimization method based on dynamic adjustment includes:

[0056] Step S1, obtain computing equipment heat load data, liquid cooling system status data, air cooling system status data and system constraint data, and construct a cooling fusion real-time status map.

[0057] It can be understood that the purpose of step S1 is to comprehensively and accurately perceive the thermal state, cooling capacity and energy consumption of the entire system, and to provide a more accurate and integrated data basis for subsequent modeling, prediction and optimization.

[0058] In this embodiment, step S1 specifically includes the following steps:

[0059] Step S1-1, obtain the original sensor signal stream based on the sensor module, obtain computing equipment thermal load data, liquid cooling system status data, air cooling system status data based on the original sensor signal stream, and obtain system constraint data based on the initial cooling system design plan.

[0060] It can be understood that the original sensor signal stream from the monitoring system or hardware interface is obtained through the sensor module, thereby obtaining different computing device thermal load data, liquid cooling system status data, and air cooling system status data. The computing device thermal load data may include computing device thermal load data such as power consumption data of key CPU, GPU, NPU, core temperature data, casing or heat sink temperature data, etc. In specific implementation, each key part can be sampled as needed. The liquid cooling system status data may include liquid cooling system status data such as inlet temperature data, outlet temperature data, liquid cooling flow data, pump power consumption data, etc. when the liquid cooling system is running. In specific implementation, the liquid cooling system data can be sampled as needed, and the air cooling system status data may include fan speed data, fan power consumption data, ambient temperature data, cabinet inlet air temperature data, cabinet outlet air temperature data, key air duct point data, etc. for air cooling system status data. In specific implementation, the air cooling system data can be sampled as needed.

[0061] Furthermore, the system constraint data is obtained based on the initial design plan for the cooling system. The initial design plan for the cooling system defines how the liquid cooling system is constructed and operated to meet specific heat dissipation targets and safety guarantees. For example, the design angles of the initial design plan for the cooling system may include angles such as heat load requirements and heat dissipation targets, physical structure and layout, key equipment selection and specifications, etc., and the system constraint data may include system constraint data such as temperature limit data of key components and related data of equipment operation limitations. The system constraint data can ensure that the overall system will not exceed the safety limit of the physical equipment during operation and when the method optimizes and adjusts the operating parameters of the overall system. For example, the real-time monitoring core temperature is always lower than the maximum temperature set by the system to prevent the chip of the data center server from being damaged due to overheating. At the same time, the premise for subsequent optimization through the algorithm or model is to find a specific control parameter combination that meets the lowest energy consumption under cooling conditions under the condition of satisfying the system constraint data. The system constraint data can define and limit the search space of the algorithm or model to prevent the algorithm or model from having a solution that exceeds the physical condition limit during execution.

[0062] It can be understood that the bottom-level data foundation can be obtained through step S1-1, and it is necessary to collect data foundations from computing equipment, liquid cooling systems, air cooling systems and system constraints in parallel, covering key liquid cooling system data and meeting specific heat dissipation goals and system constraints for safety assurance.

[0063] Step S1-2, setting a specific time window for the computing device thermal load data, the liquid cooling system status data and the air cooling system status data, and aligning the timestamps. After the setting is completed, constructing the original cooling status data set based on the computing device thermal load data, the liquid cooling system status data and the air cooling system status data.

[0064] Specifically, when sampling data such as thermal load data, liquid cooling system status data, and air cooling system status data from computing devices, the sampling frequencies may vary depending on the sensor type. For example, the sampling frequency for cabinet ambient temperature may be low, while the sampling frequency for CPU / GPU core temperature may be high. At any absolute point in time, the latest data points from different samples may represent states at different times. Directly comparing or fusing them may result in severe time misalignment, which can distort the results during subsequent model and algorithm calculations. Furthermore, the actual sampling process can also be subject to acquisition transmission delays and clock drift. When data is read from the sensor and transmitted across the bus and network to the acquisition server / processing node, there is an uncertain and potentially variable delay. Even if the samples are physically taken simultaneously, the timestamps of the data arriving at the processing node / acquisition server may differ significantly. Processing data based on the original arrival timestamps can result in false dynamic changes or obscure true correlations. Furthermore, if sampling uses its own local clock instead of synchronized clocks, the relative error between the timestamps of different sampled records will increase over time, resulting in the failure of long-term data correlation.

[0065] Furthermore, through step S1-2, a sliding time window of fixed length is defined, and all the original data points of the samples arriving within this window period are regarded as representing the system state at the end time or the middle time of the window. At the same time, for each data point in the window, interpolation, resampling, or nearest neighbor assignment is performed according to its original timestamp and system time base, and they are aligned to the representative time point of the window. Based on this, each different sampling data is assigned a valid timestamp under the same, unified time base, creating a consistent time perspective, and providing a stable data input for subsequent processing steps. In some possible embodiments, the sliding time window of fixed length can be set based on actual conditions.

[0066] In step S1-3, a data preprocessing operation is performed on the original cooling state data set to generate a spatially correlated cooling state data set, and a temporal cooling feature vector is extracted. A cooling fusion real-time state map is constructed based on the spatially correlated cooling state data set and the temporal cooling feature vector.

[0067] Furthermore, step S1-3 specifically includes the following steps:

[0068] Step S1-3-1, performing preprocessing operations on the original cooling state data set, including at least missing filling, anomaly detection, and normalization operations, and introducing system constraint data during the preprocessing operation to obtain a preprocessed cooling state data set.

[0069] It is understandable that the original cooling state data set may contain abnormal data such as noise and over-limit caused by abnormal detection. The original cooling state data set needs to be preprocessed through preprocessing operations to avoid the abnormal data from being imported into the remaining processing steps and ensure that the data entering the subsequent processing steps is complete and valid data.

[0070] Specifically, the original cooling state data set is tested, and system constraint data is introduced during the anomaly detection process. Each data point in the original cooling state data set is compared with the condition range restricted by the system constraint data. Obviously impossible data points are captured and eliminated. These data points may usually be caused by sampling failure, transmission error, or signal interference. The absolute change or rate of change between the value of the current data point and the value of the previous valid data point is calculated and compared with a preset maximum reasonable change threshold. The preset maximum reasonable change threshold is illustrated here. For example, assume that the CPU temperature cannot instantly rise by 20 degrees Celsius. For example, due to the inertia of the pump speed regulation of the liquid cooling system, the liquid cooling flow rate will not experience excessive instantaneous sudden changes. These data points may be caused by instantaneous interference, signal jitter, or poor sensor contact, rather than actual physical state sudden changes. Therefore, these data points are eliminated. In some possible embodiments, anomaly detection operations such as timeout checks (a detection method for identifying data flow interruptions caused by sensor disconnection, acquisition program crash, or network interruption) may also be added.

[0071] Furthermore, after these data points are removed, an intelligent filling strategy of missing fill is used to generate replacement values ​​to ensure the continuity of the data for use in subsequent processing steps. When performing specific intelligent filling, linear interpolation, sliding window statistical value filling, LSTM prediction filling and other filling methods can be used to perform missing filling operations. During specific implementation, it is necessary to formulate an intelligent filling strategy based on multiple angles such as data type, change type and importance selection. For example, although linear interpolation is performed between two adjacent valid data points, it can reflect trend changes and is relatively accurate, but it is less effective for nonlinear and rapidly changing data points, and at least two valid data points are required. Sliding window statistical value filling uses the average, median or mode of the data point in the most recent valid time window for filling. Although it is more stable, it will be introduced with a lag and will mask rapid changes.

[0072] Furthermore, after completing the preprocessing operations of missing fill and abnormality detection, normalization processing such as Softmax function can be used. The Softmax function converts each element in the channel attention weight vector into a probability value. These probability values ​​represent the relative importance of each channel in the fusion process. In specific implementation, the most appropriate normalization method can be selected based on the specific data distribution form of the preprocessed cooling state data set, or a mixed normalization method can be used to achieve the best adaptation. Here, only the Softmax function is used as an example to illustrate the normalization operation. The normalization operation can be used to convert data of different units into consistent units and normalize them to the appropriate interval according to their specific values, so as to provide comparable data for the subsequent processing steps when the algorithm and model are processed. After the preprocessing operation, the original cooling state data set that has undergone the preprocessing operation is obtained, that is, the preprocessed cooling state data set.

[0073] Step S1-3-2: Obtain a physical layout topology diagram of the cooling system, perform spatial mapping and regional association on the pre-processed cooling state data set based on the physical layout topology diagram of the cooling system, and generate a cooling space state data set.

[0074] Specifically, the physical layout topology diagram of the cooling system is a topology diagram that clearly represents the spatial position topology relationship of key heat source components, liquid cooling flow paths, fan positions, air duct directions, etc. The data points such as temperature and power consumption collected in the preprocessed cooling status data set are added with precise location identifiers for each data point based on the physical layout topology diagram of the cooling system, and the location labels are mapped to the physical areas to which they belong, that is, mapped to the specific physical locations they represent. The physical areas are represented as physical coordinates or logical location identifiers. The division of physical areas can be set based on the equipment locations of different data centers, thereby generating a preprocessed cooling status data set with spatial location information, that is, a cooling space status data set.

[0075] It is understandable that the simple numerical value of data points cannot reflect the relationship between physical distance and thermal coupling. Establishing the spatial correlation between data points is the basis for regional thermal state analysis and cooling capacity assessment. By converting abstract data point values ​​into state descriptions of physical locations, it forms the premise for system regulation, enabling it to understand the physical meaning of the data and give it important spatial dimension cognition.

[0076] In some possible embodiments, when the data center is operating normally, if there is no location information, the system will not know where the high-temperature points are, which areas need priority cooling, and which devices will be affected by traffic changes. Step S1-3-2 is a bridge connecting the physical and data models, providing spatial context for subsequent thermal state analysis, cooling efficiency evaluation, and optimization and regulation.

[0077] Step S1-3-3, extract the cooling time series feature vector of the cooling space state data set, perform regional state fusion based on the cooling time series feature vector and the cooling space state data set, and output cooling key state indicator data and cooling low-dimensional feature data vector.

[0078] Specifically, the data points of the first few continuous windows form a short-term historical sequence, and on the data points of the current time window and the short-term historical sequence, the selected key data points (for example, power consumption data, core temperature data, inlet temperature data, outlet temperature data, etc., not all data need to be calculated, only the key data that plays a key role in the actual operation process) are calculated, and the rate of change is calculated based on the representative value of the current window data point - the representative value of the previous window data point / time interval. A positive value indicates an increase, and a negative value indicates a decrease. Then, the arithmetic mean or median is calculated for the data points of the current window and the data points of the previous N windows to smooth the instantaneous fluctuations, which is used to reveal the main trends of the data in the recent period and capture dynamic characteristics. The instantaneous value can only reflect the current data point, while the main trend can affect the future state and control decision.

[0079] In some possible embodiments, taking the CPU core temperature time series data of a computing node in a server cluster as an example, assuming that the temperature samples of the node in three consecutive time windows are [time t1: 70°C, time t2: 71°C, time t3: 72.3°C], the time window is 1 second, the instantaneous change rate (current value - previous value) / time interval is (72.3-71.0) / 1=+1.3°C / s, the arithmetic mean is (70.0+71.0+72.3) / 3=71.1°C, or the median is 71°C in [70.0, 71.0, 72.31], which is used to smooth instantaneous fluctuations and reflect the recent trend level. At the same time, the position identifiers are bound, and the final generated cooling time series feature vector is [72.3, 1.3, 71.1 / 71] (here it is assumed that the cooling time series feature vector is obtained after processing in the previous steps and is not expressed as normalization. The actual dimension and value are determined by the specific situation).

[0080] Furthermore, the input of regional state fusion is the cooling time series feature vector and the cooling space state data set. After the input, the key cooling state indicator data is obtained, and the multiple power consumption data of the same physical area are averaged or weighted averaged, such as according to the number of cores or area, to obtain regional heat load data, thereby quantifying the local heat source intensity and providing a basis for the allocation of cooling resources. The higher the heat load, the greater the required cooling capacity. Then, the weighted average of the heat load data of the computing equipment such as the temperature data of multiple cores in the physical area, the temperature data of the shell or heat sink, etc. is used to obtain regional temperature characterization data, thereby reflecting the actual thermal state. Then, the temperature difference in the physical area is used in combination with the liquid cooling system state data and the air cooling system state data to estimate the rough unit power consumption cooling capacity or temperature rise margin, and use it as the regional cooling efficiency or regional cooling margin, thereby quantifying the cooling safety margin, tightly. Next, the energy consumption of all physical areas is summarized and calculated, and the current maximum temperature and key temperature difference are identified. Finally, the cooling key state indicator data is constructed based on the data in the cooling time series feature vector. The cooling key state indicator data at least includes data such as regional heat load data, regional temperature characterization data, regional cooling efficiency or regional cooling margin, energy consumption, etc. At the same time, the cooling key state indicator data is organized into the cooling low-dimensional feature data vector required for subsequent use. The data and dimensions contained in the cooling low-dimensional feature data vector need to be adaptively adjusted according to the actual computing power and processing capacity. The cooling low-dimensional feature data vector at least includes the key data in the cooling key state indicator data. Through regional state fusion, a large amount of raw data can be refined into low-dimensional, high-value information reflecting the regional or overall level thermal state, cooling capacity and energy consumption. The goal is to provide a comprehensive and concise state snapshot for the subsequent use.

[0081] Step S1-3-4, constructing a cooling fusion real-time state map based on the cooling key state indicator data and the cooling low-dimensional feature data vector.

[0082] It can be understood that the cooling key state indicator data and the cooling low-dimensional feature data vector can be processed based on a method such as a spatial mapping engine to generate a cooling fusion real-time state map. The cooling fusion real-time state map may include visual maps such as a heat load distribution map, a characterization distribution map, a cooling capacity and flow distribution map, an energy consumption distribution map, a constraint boundary map, and at least contains a cooling low-dimensional feature data vector, wherein the heat load distribution map is used to represent the real-time power consumption density map of each physical area, using a numerical or visual mapping method, the temperature distribution map is used to reflect the thermal state of each physical area, and the hot spot position and characterization value are marked in the map, the cooling capacity and flow distribution map is used to represent the cooling capacity of the liquid cooling system and the air cooling system, the liquid cooling circuit flow and wind speed distribution and other visual data, the energy consumption distribution map is used to represent the overall real-time energy consumption in the form of a bar graph or percentage, the constraint boundary map is used to identify the distance between the current value and the key constraints, such as the maximum temperature and other data, and the cooling low-dimensional feature data vector is used to provide a direct calculation basis for subsequent model prediction or optimization calculation.

[0083] Furthermore, the cooling fusion real-time status map not only organizes the collected, processed and fused information into a structured knowledge expression, but also provides a visual reference for operation and maintenance personnel in practical applications.

[0084] In this embodiment, raw data is collected in parallel, preprocessed to verify data quality, spatial mapping is used to establish relationships, dynamic features are used to enhance prediction capabilities, intelligent fusion is used to refine states, and graphs are used to construct complete representations. This ensures that modeling and optimization in subsequent steps are based on accurate, real-time, integrated, and unified high-fidelity state information, which is the prerequisite for the entire closed-loop regulation and control.

[0085] Step S2: obtain the cooling prediction model, calibrate the cooling prediction model based on the cooling fusion real-time state map, extract the key dynamic feature vector of cooling after the calibration is completed, and obtain short-term cooling prediction data based on the cooling prediction model and the cooling fusion real-time state map to construct the cooling conflict objective function.

[0086] In this embodiment, step S2 specifically includes the following steps:

[0087] Step S2-1, obtain a pre-trained cooling prediction model, input the cooling low-dimensional feature data vector in the cooling fusion real-time state map of the historical window into the cooling prediction model, predict the cooling low-dimensional feature data vector of the current window based on the cooling prediction model, check the predicted cooling low-dimensional feature data vector of the current window with the actual cooling low-dimensional feature data vector of the current window, and calibrate the cooling prediction model based on the optimization algorithm.

[0088] Specifically, during the historical operation of the data center, relevant historical data can be collected, and the required cooling prediction model can be built based on the historical data. The model is only started at startup. If it is not started on a daily basis, it exists in the form of an offline model. During the construction process, a low-order physical model can be selected. Its physical interpretability is low, but the time for parameter calibration may be relatively delayed. Alternatively, a data-driven model can be selected, such as the linear autoregressive model ARX, the nonlinear autoregressive model NARX, the lightweight neural network, the support vector machine regression SVR and other data-driven models. These models can fit complex nonlinear data, but require historical data, and generalization needs to be guaranteed. Alternatively, a hybrid model can be selected, such as a hybrid model of a physical model + a data-driven model, which has advantages in balancing physics and data. In the actual implementation process, the model to be used can be selected based on the actual situation. Choosing the appropriate model type is a trade-off between performance and speed. The model must run quickly under limited resources. The physical model provides basic stability, and the data-driven model enhances adaptability. The hybrid model has advantages in balancing physics and data, but coupling is more difficult.

[0089] Furthermore, the cooling low-dimensional feature data vector in the cooling fusion real-time state map of the historical window is input into the cooling prediction model. When the data in the cooling low-dimensional feature data vector is imported into the cooling prediction model, it can be selected based on the execution function of the cooling prediction model and the actual situation, that is, the short-term historical state sequence of [t-1, t-2, ..., tk], k can be the most recent several sampling data points, and the cooling low-dimensional feature data vector of the historical window is imported into the cooling prediction model so that the cooling low-dimensional feature data vector can output the predicted cooling low-dimensional feature data vector. Subsequently, the cooling low-dimensional feature data vector of the current window can be compared with the predicted cooling low-dimensional feature data vector to calculate the prediction error of the model. After the prediction error occurs, an optimization algorithm can be used to calibrate the prediction error of the cooling prediction model. For example, recursive least squares or exponential forgetting least squares can be used, which are suitable for linear and quasi-linear models, or gradient descent can be used to fine-tune the parameters of the nonlinear model. For hybrid models, a few backpropagations can be used to update the model, but a special model structure needs to be designed to achieve fast online convergence.

[0090] It is understandable that during the real-time operation of the cooling system, the system status and cooling performance will change dynamically. However, if the model is offline and started, the model parameters will be outdated. By calibrating and fine-tuning the model with real-time data, the model can more accurately reflect the true dynamic behavior of the current cooling system, improve short-term prediction accuracy, enhance robustness, and provide a more reliable basis for subsequent optimization.

[0091] Step S2-2: import the cooling low-dimensional feature data vector of the current window into the calibrated cooling prediction model, perform relationship extraction on the cooling prediction model, obtain the cooling key dynamic feature vector, and obtain short-term cooling prediction data based on the cooling prediction model.

[0092] Specifically, the key dynamic feature vector of cooling serves as the decision-making basis for real-time optimization control. Essentially, it is a low-dimensional, high-information-density quantitative feature package extracted from a complex dynamic model. It provides efficient guidance for the subsequent optimization algorithm, namely the MC-PSO (Monte Carlo algorithm and particle swarm optimization), and can direct the search direction of control parameters to ensure the optimal balance between energy consumption and cooling in a short period of time.

[0093] Furthermore, when extracting the key dynamic feature vector of cooling, the cooling low-dimensional feature data vector of the current window is imported into the calibrated cooling prediction model. At this time, the cooling prediction model is operated based on the cooling low-dimensional feature data vector, and the cooling prediction model is linearized and sensitively analyzed at the current working point, that is, the cooling low-dimensional feature data vector of the current window, so as to quantify the efficiency and direction of the instantaneous influence of small variables or disturbances on key target quantities. The control / disturbance sensitivity of the current working point can be calculated by methods such as numerical perturbation method, for example, the effect of enhancing the unit flow on reducing the hotspot temperature, the effect of increasing the unit fan speed on reducing the hotspot temperature, the rate of increase of the hotspot temperature caused by the increase in load, etc. It can also be calculated by high-order derivative methods such as Hessian approximation, but it will be more time-consuming. Then the dominant dynamic time constant and mode are extracted to reveal the nature of the response speed, understand the inertia size and time scale of temperature change and power consumption adjustment, and avoid subsequent optimization from making decisions such as acting too fast or acting too slowly and lagging. The analysis is performed using methods such as model-dominated extreme point analysis, which is only applicable to linearized models, or using the general method of step response analysis. Next, the degree of constraint boundary proximity and safety margin are dynamically evaluated, and risk characteristics such as the residual margin consumption rate are extracted. This allows for dynamic perception of the current and short-term future approach to hard constraints, such as the maximum limit temperature, and quantifies the degree of danger of the approach, thereby guiding the balance between resource allocation and risk avoidance. Finally, the real-time energy efficiency of the cooling system is quantified, characterizing the energy cost of increasing the cooling intensity near the current operating point, as well as the potential optimal energy consumption operating point, which is used to provide a key basis for the trade-off between energy consumption and cooling. This can be performed through methods such as local energy efficiency gradient calculation. After obtaining multi-dimensional and multi-type key dynamic characteristics through the above analysis, they are structured and dimensionalized, and packaged into a compact low-dimensional data package for real-time optimization. This ensures that the information density of the low-dimensional data package is high and it can be easily parsed and utilized by subsequent optimization algorithms.

[0094] Understandably, the complete cooling prediction model may be too complex to be directly embedded in a real-time optimization algorithm, such as the subsequent PSO. By compressing and refining the cooling prediction model, a set of key feature vectors for optimizing control decisions is formed, such as sensitivity, energy efficiency gradient data, and other data. These key feature vectors will provide key gradient information and dynamic information, providing guidance for the subsequent PSO search direction.

[0095] Furthermore, after the extraction of the key dynamic feature vectors of cooling is completed, short-term cooling prediction data is obtained based on the cooling prediction model. The short-term cooling prediction data is expressed as the predicted cooling demand and predicted temperature in the future short-term window, which can usually be expressed as the increased equivalent cooling flow / air volume required to offset the predicted load temperature rise. Through the processing of the cooling prediction model, the future temperature change trend is predicted to keep the current cooling intensity unchanged, showing the upcoming cooling shortage risk points and potential energy saving opportunities. The output prediction demand quantitatively tells how much additional cooling capacity will be needed in the future, so that the optimization can make immediate adjustments before the problem occurs.

[0096] Step S2-3: constructing a cooling conflict objective function based on the cooling key dynamic feature vector, the short-term cooling prediction data, and the cooling fusion real-time state map.

[0097] Furthermore, step S2-3 specifically includes the following steps:

[0098] Step S2-3-1, generating a risk penalty function based on the cooling key dynamic feature vector and the short-time cooling prediction data.

[0099] Specifically, the cooling set temperature is used as the most basic penalty item. When the hot spot temperature exceeds the cooling set temperature, a positive penalty is generated, namely the static margin penalty. When the hot spot temperature is lower than the cooling set temperature, no penalty is generated. The static margin penalty can simply, intuitively and directly reflect the risk of the current state, but it lacks foresight and is slow to reflect the risk of rapid temperature rise or prediction. Therefore, a dynamic margin penalty is needed to assist in this. The short-term cooling prediction data can be used to evaluate the predicted maximum temperature while maintaining the current cooling intensity, and to penalize potential over-temperature in the future. At the same time, the temperature change rate is taken into account, and an additional change penalty coefficient is imposed on the current status of rapid temperature rise. Because the faster the temperature rises, the faster the risk accumulates, and active intervention is needed, the dynamic margin penalty is combined with the static margin penalty and appropriate penalty weights are added to generate a risk penalty function = Static margin penalty + dynamic margin penalty (i.e. Overtemperature Penalty+ Change penalty), where A lower constant, such as 0.2, can be set as a basic guarantee. A moderate constant, such as 0.3, can be set to quickly respond to emergencies. A higher constant, such as 0.5, can be set to avoid forward-looking risks. The above is the construction of the risk penalty function in the form of typical linear weighting. In some possible embodiments, the risk penalty function can also be constructed using nonlinear or piecewise functions, such as exponential functions or piecewise linear functions. The piecewise linear function is used as an example here. The construction angle can be to set multiple temperature intervals. The closer to the cooling setting, the greater the penalty slope.

[0100] It can be understood that the cooling set temperature is the highest priority constraint. Through this step, it can be converted into a differentiable or derivative penalty term. Based on the core idea that costs should be incurred when the current or predicted temperature exceeds the cooling set temperature, it is constructed based on the cooling key dynamic feature vector and short-term cooling prediction data.

[0101] Step S2-3-2, perform energy consumption component modeling based on the cooling fusion real-time state map, cooling key dynamic feature vectors and short-term cooling prediction data, and construct a total energy consumption model based on the energy consumption component modeling.

[0102] Specifically, based on the cooling low-dimensional feature data vector, cooling key dynamic feature vector and short-time cooling prediction data contained in the cooling fusion real-time state map, the liquid cooling pump power consumption and air cooling fan power consumption are modeled. The model is modeled as a function related to the pump speed and flow rate, and the model is modeled as a function of the fan speed and air volume. The modeling is usually carried out using the method of approximate cubic relationship and physical principles. After the modeling is completed, the energy efficiency characteristics in the cooling key dynamic feature vector (such as energy efficiency gradient data) are used to perform initial parameter correction of the model parameters, and the total energy consumption of the model is calculated.

[0103] Step S2-3-3, constructing a preliminary conflict objective function for cooling based on the risk penalty function and the total energy consumption model, and introducing a trade-off weight coefficient in the process of constructing the preliminary conflict objective function for cooling.

[0104] In some possible embodiments, the cooling preliminary conflict objective function can be constructed by the most commonly used linear weighted summation, or an adaptive dynamically adjusted adjustment factor can be added based on the risk level. Here, the cooling preliminary conflict objective function is constructed in the form of linear weighted summation for illustration. The cooling preliminary conflict objective function can be equal to (1-x)×risk penalty function+x×total energy consumption model. For the introduced trade-off weight coefficient x, optionally, x can be selected to be equal to 0.5, which is used to balance the cooling temperature and energy consumption. In specific use, the trade-off weight coefficient can also be adjusted based on the required situation, such as adjusting it to 0.7, which means it is more inclined to save energy consumption, or adjusting it to 0.3, which means it is more inclined to pursue the cooling temperature.

[0105] It can be understood that the cooling preliminary conflict objective function is the cooling conflict objective function that has not been constrained and is the predecessor of the cooling conflict objective function.

[0106] Step S2-3-4, performing constraint addition processing on the preliminary cooling conflict objective function, and obtaining the cooling conflict objective function based on the constraint addition processing.

[0107] Specifically, the constraint addition processing includes hard constraint data and soft constraint data. Hard constraint processing is performed on the cooling preliminary conflict objective function based on the hard constraint data, and soft constraint processing is performed on the cooling preliminary conflict objective function based on the soft constraint data.

[0108] Furthermore, hard constraint data is obtained based on the real-time state map of cooling fusion. Hard constraint data is system constraint data, which is suitable for constraining some physical limits (such as maximum pump speed, maximum fan speed) and temperature red lines. Through hard constraints, it can be ensured that the subsequent PSO does not consider the optimization that exceeds the physical limits at all, thus avoiding high-risk regulation in the final control.

[0109] Furthermore, soft constraint data is obtained. Soft constraint data may include desired operating range (e.g., avoiding pumps or fans from operating in inefficient areas), wear range (e.g., large adjustments to pumps or fans), etc. Soft constraints can improve the availability of the final control and ensure that PSO generates physically feasible and smoothly executed solutions during optimization.

[0110] It can be understood that the cooling preliminary conflict objective function processed by hard constraints and soft constraints is the cooling conflict objective function, which ensures that the subsequent optimization results are within a safe range and are within a range that meets physical feasibility and operational requirements. The cooling conflict objective function is a direct quantification of the balance between energy consumption and cooling temperature, which can map complex problems into a mathematical problem. It is the evaluation criterion for the subsequent MC-PSO to conduct optimization and evaluation of each potential control strategy, and also provides a clear search direction for the PSO particle swarm.

[0111] Step S3: obtaining a cooling random sample set, and performing an initialization operation on the PSO algorithm based on the cooling random sample set. The PSO algorithm generates an optimal control parameter combination through a cooling conflict objective function.

[0112] In this embodiment, step S3 includes:

[0113] Step S3-1: obtaining a random sampling range based on the cooling conflict objective function, and performing sampling based on the random sampling range using a Monte Carlo algorithm to obtain a cooling random sample set.

[0114] Furthermore, step S3-1 specifically includes the following steps:

[0115] Step S3-1-1, determine the set of control variables to be optimized based on the cooling conflict objective function, obtain the upper limit of the feasibility range and the lower limit of the possibility range based on the set of control variables to be optimized, and obtain the random sampling range based on the upper limit of the feasibility range and the lower limit of the possibility range.

[0116] Specifically, the set of control variables to be optimized is determined by the hard constraint data and soft constraint data taken before cooling the conflicting objective function. For example, the set of control variables to be optimized can be expressed as [flow rate, fan speed, ...]. Then, the upper limit and lower limit of the feasible range of each control variable in the set of control variables to be optimized are clarified, and the upper limit and lower limit of the feasible range form the sampling range for random sampling of the subsequent Monte Carlo algorithm. By obtaining the random sampling range, the definition of the parameter space range that can be searched and has boundary constraints is determined.

[0117] Step S3-1-2, importing the random sampling range into the Monte Carlo algorithm, obtaining the total cooling random samples based on the Monte Carlo algorithm, randomly selecting the total cooling random samples based on the selection method of the Monte Carlo algorithm, and combining the randomly selected cooling random samples into a cooling random sample set.

[0118] It can be understood that the Monte Carlo algorithm is an efficient numerical random method based on random sampling. The random sampling range is imported into the Monte Carlo algorithm. By using the Monte Carlo sampling in the Monte Carlo algorithm, all the control variables in the random sampling range are used as the sum of cooled random samples. Random sample points are generated uniformly or preferentially in the feasible domain. This can not only fully explore the solution space, but also provide a diverse population for the initial particle position set of the subsequent PSO algorithm, thereby improving the global search capability of the PSO algorithm.

[0119] In some possible embodiments, Monte Carlo sampling may use a selection method such as the most basic random sampling to generate N cooling random samples through a uniform distribution or a quasi-random sequence, or may randomly sample the N cooling random samples through a method such as boundary reinforcement sampling, which increases the sample density near the boundary constraint, or increases the preference for sampling in hot spot risk areas, such as based on the predicted cooling temperature, speculating areas that may cause overheating, and appropriately increasing the number of sample points in the area, so that the subsequent PSO algorithm pays special attention to some risk points.

[0120] It can be understood that Monte Carlo sampling is the core step in the Monte Carlo algorithm. By widely covering the entire feasible domain of random sampling, it can avoid the PSO algorithm from falling into the local area during initialization, and can explore the area near the constraint boundary, thereby expanding the diversity of the PSO algorithm. The cooled random sample set obtained by Monte Carlo sampling meets the constraint conditions and provides a basic initialization particle position for the subsequent PSO algorithm.

[0121] Step S3-2: importing the cooling random sample set into the PSO algorithm and performing an initialization operation on the PSO algorithm. The PSO algorithm selects cooling random samples from the cooling random sample set for initialization, and obtains an initial particle position set of the cooling random samples.

[0122] It can be understood that the PSO algorithm, namely the particle swarm optimization algorithm, is a parallel random optimization algorithm based on swarm intelligence. It uses a swarm of particles, namely candidate solutions, to collaboratively search for the optimal solution, namely the optimal control parameter combination, in the solution space. Each particle can dynamically adjust its movement direction and speed according to its own historical optimal position and the historical optimal position of the group, and quickly converge to the optimal solution through iterative updates. The algorithm is good at handling nonlinear, multi-peak, and constrained optimization problems, such as cooling system optimization problems, and can improve the efficiency of dynamic optimization.

[0123] Specifically, the PSO algorithm is initialized, its algorithm parameters are set based on the specific cooling system, the cooling random sample set is used as the initial particle position, and the speed of all particles is initialized.

[0124] Step S3-3: importing the cooling conflict objective function into the PSO algorithm, optimizing the initial particle position set based on the PSO algorithm and performing update iterations, and outputting the optimal control parameter combination after the update iterations converge.

[0125] Furthermore, particle evaluation is performed on the initial particle position set based on the PSO algorithm.

[0126] Specifically, a particle evaluation is performed on each particle in the initial particle position set to check whether each dimension of the particle is within the constraint boundary. If so, the next particle is evaluated and the particle is introduced into the cooling conflict objective function to obtain its evaluation value. If not, the particle can be directly marked as invalid. In some possible embodiments, a method such as particle boundary repair can be used to repair particles that are not within the constraint boundary, that is, a dual evaluation of real-time and prediction is performed on the particles based on the cooling conflict function and the cooling prediction model, or the particles are forced to be pulled back to the boundary constraint. Excessive discussion will not be given here, and the specific method used can be determined based on actual conditions during execution.

[0127] It can be understood that by evaluating the initial particle position set, particles can be forced to fly in a safe and physically feasible space, and the quality of particles can be evaluated. Particles that are outside the task boundary or predicted to cause overheating can be completely eliminated through invalid marking. Particles can also be repaired through repair methods, but the repair method can easily lead the PSO algorithm into a suboptimal solution or fall into local stagnation.

[0128] Furthermore, the initial particle position set is updated empirically and motion-based based on the PSO algorithm.

[0129] Specifically, the experience of the initial particle position set is updated based on the individual experience and global experience in the PSO algorithm. The optimal position discovered by the particle itself is retained through individual experience, and the optimal position discovered by the entire group is recorded through global experience. Through this process, the initial particle position set (i.e., the particle swarm) can be driven to move to a better area, guiding the entire group to converge to the global optimal or highest quality approximate solution. The motion engine of the PSO algorithm drives the particles to update their position and speed. The update process includes inertia, cognitive, and social terms. The inertia term is used to retain the last movement direction, the cognitive term is closer to the individual's historical best position, and the social term is closer to the group's historical best position. The motion engine of the PSO algorithm can be set, such as reducing the size of the inertia term to strengthen local search, reducing the social term to reduce the problem of being attracted by the global optimal and ignoring the local problem, or increasing the inertia term to enhance the global exploration ability, and increasing the social term to converge to the potential better area faster.

[0130] When the PSO algorithm updates and iterates the initial particle position, the convergence is checked. If converged, the update iteration is stopped and the optimal control parameter combination is output. If not converged, the update iteration is continued.

[0131] In some possible embodiments, the convergence of the update iteration can be set based on requirements such as the maximum number of iterations, convergence of the objective function, reaching the global optimal value, particle swarm convergence, timeout, etc. By setting the convergence of the update iteration, the cycle of the PSO algorithm can be controlled to avoid unnecessary redundant computing consumption. Taking the maximum number of iterations that are commonly used as an example, real-time performance can be ensured and efficiency can be improved. After convergence, the optimal control parameter combination can be output, and the optimal control parameters can be denormalized, such as by using linear denormalization, Z-Score denormalization, etc., mapped to the original physical range, and its data scale is restored to the actual physical value of the project, and finally a control parameter instruction that can be sent is generated.

[0132] Step S4: Control the cooling system in real time based on the optimal control parameter combination, and execute steps S1 to S4 in a loop.

[0133] Specifically, a physical bus such as CAN, RS485, etc. can be used to send the optimal control parameter combination to the cooling system in the form of a control instruction, and the cooling system can be regulated in real time. Then, step S1 is entered to start the next round of cyclic regulation operation, so as to achieve real-time control of the cooling system by finding the optimal energy consumption while maintaining the cooling capacity of the cooling system.

[0134] Figure 2 A schematic diagram of a liquid-cooled fusion energy consumption regulation and optimization system based on dynamic adjustment provided by some embodiments of the present application, which can realize the idea of ​​the present application, is shown. The liquid-cooled fusion energy consumption regulation and optimization method based on dynamic adjustment is introduced in detail below.

[0135] Specifically, the liquid cooling fusion energy consumption regulation and optimization system based on dynamic regulation includes:

[0136] An acquisition module, the acquisition module being used to acquire computing device thermal load data, liquid cooling system status data, air cooling system status data, and system constraint data;

[0137] A map module, which is used to construct a real-time state map of cooling and fusion;

[0138] A model module, the model module is used to extract key dynamic feature vectors of cooling and obtain short-term cooling prediction data, and to construct a cooling conflict objective function;

[0139] An optimization module, wherein the optimization module generates an optimal control parameter combination based on a PSO algorithm;

[0140] A control module, wherein the control module controls the cooling system in real time based on an optimal control parameter combination;

[0141] A loop module is used to execute steps S1 to S4 in a loop.

[0142] The specific usage and function of this embodiment are described below:

[0143] This method can obtain the most basic operating parameters of the cooling system by obtaining computing equipment thermal load data, liquid cooling system status data, air cooling system status data and system constraint data, impose interval restrictions on the operating parameters, and construct a cooling fusion real-time state map. Furthermore, the cooling prediction model is calibrated based on the cooling fusion real-time state map, so that the cooling prediction model that may be offline can operate in a relatively standard state after going online. The cooling key dynamic feature vector and short-time cooling prediction data are extracted through the cooling prediction model, and a cooling conflict objective function is constructed based on the cooling key dynamic feature vector and the short-time cooling prediction data, so as to model the relationship between energy consumption and cooling capacity. On this basis, a cooling random sample set is obtained through the Monte Carlo algorithm to ensure the diversity of the cooling random sample set. The optimal control parameter combination is obtained based on the PSO algorithm and the cooling conflict objective function, and this is executed cyclically, so that the method can find a better energy consumption control in real time while the cooling system maintains its cooling capacity.

[0144] In addition, an embodiment of the present invention further provides an electronic device, including:

[0145] At least one processor; and a memory communicatively connected to the at least one processor; wherein the memory stores instructions executable by the at least one processor, and the instructions are executed by the at least one processor to enable the at least one processor to perform the method proposed in the first embodiment of the present invention.

[0146] The following is a detailed introduction to the various components of electronic equipment:

[0147] The term "processor" is the control center of an electronic device and can be a single processor or a collective term for multiple processing elements. For example, the processor can be one or more central processing units (CPUs), an application-specific integrated circuit (ASIC), or one or more integrated circuits configured to implement the first embodiment of the present invention, such as one or more digital signal processors (DSPs) or one or more field programmable gate arrays (FPGAs).

[0148] The processor can execute various functions of the electronic device by running or executing software programs stored in the memory and calling data stored in the memory.

[0149] The memory is used to store the software program for executing the solution of the present invention, and the execution is controlled by the processor. The specific implementation method can refer to the above method embodiment and will not be repeated here.

[0150] The memory may be a read-only memory (ROM) or other type of static storage device capable of storing static information and instructions, a random access memory (RAM) or other type of dynamic storage device capable of storing information and instructions, an electrically erasable programmable read-only memory (EEPROM), a compact disc read-only (CD-ROM), or other optical disc storage, optical disc storage (including compact discs, laser discs, optical discs, digital versatile discs, Blu-ray discs, etc.), a magnetic disk storage medium or other magnetic storage device, or any other medium capable of carrying or storing desired program code in the form of instructions or data structures and accessible by a computer, but is not limited thereto. The memory may be integrated with the processor or exist independently and be coupled to the processor via an interface circuit of the electronic device, and this is not specifically limited in the embodiments of the present invention.

[0151] The above embodiments can be implemented in whole or in part via software, hardware (e.g., circuits), firmware, or any other combination thereof. When implemented using software, the above embodiments can be implemented in whole or in part in the form of a computer program product. The computer program product comprises one or more computer instructions or computer programs. When loaded or executed on a computer, the processes or functions described in accordance with the embodiments of the present invention are fully or partially performed. The computer can be a general-purpose computer, a special-purpose computer, a computer network, or other programmable device. The computer instructions can be stored in a computer-readable storage medium or transferred from one computer-readable storage medium to another. For example, the computer instructions can be transferred from one website, computer, server, or data center to another website, computer, server, or data center via wireless communication (e.g., infrared, wireless, microwave, etc.). The computer-readable storage medium can be any available medium accessible by a computer, or a data storage device such as a server or data center that contains a collection of one or more available media. The available medium can be magnetic media (e.g., floppy disks, hard disks, magnetic tapes), optical media (e.g., DVDs), or semiconductor media. The semiconductor media can be a solid-state drive.

[0152] It should be understood that the term "and / or" as used herein simply describes an association between related objects, indicating that three possible relationships exist. For example, "A and / or B" can represent the existence of A alone, the existence of both A and B, or the existence of B alone. A and B can be singular or plural. Furthermore, the character " / " as used herein generally indicates an "or" relationship between the related objects, but it may also indicate an "and / or" relationship. For specific understanding, please refer to the context.

[0153] It should be understood that in the embodiments of the present invention, the size of the serial numbers of the above-mentioned processes does not mean the order of execution. The execution order of each process should be determined by its function and internal logic, and should not constitute any limitation on the implementation process of the embodiments of the present invention.

[0154] The embodiments described above are only used to illustrate the technical solutions of the present invention, rather than to limit the same. Although the present invention has been described in detail with reference to the aforementioned embodiments, those skilled in the art should understand that they can still modify the technical solutions described in the aforementioned embodiments, or make equivalent replacements for some of the technical features therein. These modifications or replacements do not deviate the essence of the corresponding technical solutions from the spirit and scope of the technical solutions of the various embodiments of the present invention, and should all be included in the scope of protection of the present invention.

Claims

1. A liquid cooling fusion energy consumption adjustment optimization method based on dynamic adjustment, characterized in that: The method comprises: S1: Obtain computing equipment thermal load data, liquid cooling system status data, air cooling system status data, and system constraint data, and build a real-time cooling fusion status map; S2: Obtain a cooling prediction model and calibrate the cooling prediction model based on the cooling fusion real-time state map. After calibration, extract the cooling key dynamic feature vectors and obtain short-term cooling prediction data based on the cooling prediction model and the cooling fusion real-time state map to construct a cooling conflict objective function, including: Obtain a pre-trained cooling prediction model, input the cooling low-dimensional feature data vector in the cooling fusion real-time state map of the historical window into the cooling prediction model, predict the cooling low-dimensional feature data vector of the current window based on the cooling prediction model, compare the predicted cooling low-dimensional feature data vector of the current window with the actual cooling low-dimensional feature data vector of the current window, and calibrate the cooling prediction model based on the optimization algorithm; Import the cooling low-dimensional feature data vector of the current window into the calibrated cooling prediction model, perform relationship extraction on the cooling prediction model, obtain the cooling key dynamic feature vector, and obtain short-term cooling prediction data based on the cooling prediction model; The cooling conflict objective function is constructed based on the cooling key dynamic feature vector, short-term cooling prediction data and cooling fusion real-time state map; Generate risk penalty function based on cooling key dynamic feature vector and short-time cooling prediction data; Energy consumption component modeling is performed based on the cooling fusion real-time state map, cooling key dynamic feature vectors and short-term cooling prediction data, and a total energy consumption model is constructed based on the energy consumption component modeling; Based on the risk penalty function and the total energy consumption model, a preliminary cooling conflict objective function is constructed, and a trade-off weight coefficient is introduced in the process of constructing the preliminary cooling conflict objective function; performing constraint addition processing on the preliminary cooling conflict objective function, and obtaining a cooling conflict objective function based on the constraint addition processing; S3: Obtain a cooling random sample set, initialize the PSO algorithm based on the cooling random sample set, and generate an optimal control parameter combination through the cooling conflict objective function; S4: Based on the optimal control parameter combination, the cooling system is controlled in real time, and steps S1 to S4 are executed in a loop.

2. The liquid cooling fusion energy consumption adjustment and optimization method based on dynamic adjustment according to claim 1 is characterized in that: Obtain computing equipment thermal load data, liquid cooling system status data, air cooling system status data, and system constraint data, and build a real-time cooling fusion status map, including: Acquire a raw sensor signal stream based on the sensor module, acquire computing device thermal load data, liquid cooling system status data, and air cooling system status data based on the raw sensor signal stream, and acquire system constraint data based on an initial cooling system design plan; Setting a time window for the computing device heat load data, the liquid cooling system status data, and the air cooling system status data, and aligning the timestamps. After the setting is completed, constructing an original cooling status data set; The original cooling state dataset is preprocessed to generate a spatially correlated cooling state dataset. The temporal cooling feature vector is extracted at the same time. A cooling fusion real-time state map is constructed based on the spatially correlated cooling state dataset and the temporal cooling feature vector.

3. The liquid cooling fusion energy consumption adjustment and optimization method based on dynamic adjustment according to claim 2 is characterized in that: Perform data preprocessing on the original cooling state data set to generate a spatially correlated cooling state data set, and extract the temporal cooling feature vector. Based on the spatially correlated cooling state data set and the temporal cooling feature vector, a cooling fusion real-time state map is constructed, including: Performing a preprocessing operation on the original cooling state data set, including at least missing filling, anomaly detection, and normalization operations, and introducing system constraint data during the preprocessing operation to obtain a preprocessed cooling state data set; Obtain a physical layout topology diagram of the cooling system, perform spatial mapping and regional association on the preprocessed cooling state dataset based on the physical layout topology diagram of the cooling system, and generate a cooling space state dataset; Extract the cooling time series feature vector of the cooling space state data set, perform regional state fusion based on the cooling time series feature vector and the cooling space state data set, and output cooling key state indicator data and cooling low-dimensional feature data vector; A cooling fusion real-time status map is constructed based on cooling key status indicator data and cooling low-dimensional feature data vectors.

4. The method for optimizing liquid cooling fusion energy consumption based on dynamic regulation according to claim 1 is characterized in that: The cooling preliminary conflict objective function is subjected to constraint addition processing, and the cooling conflict objective function is obtained based on the constraint addition processing: Acquiring hard constraint data based on a cooling fusion real-time state map, and performing hard constraint processing on the cooling preliminary conflict objective function based on the hard constraint data; Soft constraint data is acquired, and soft constraint processing is performed on the cooling preliminary conflict objective function based on the soft constraint data.

5. The method for optimizing liquid cooling fusion energy consumption based on dynamic regulation according to claim 1 is characterized in that: A cooling random sample set is obtained, and a PSO algorithm is initialized based on the cooling random sample set. The PSO algorithm generates an optimal control parameter combination through a cooling conflict objective function, including: Obtain a random sampling range based on the cooling conflict objective function, and perform sampling based on the random sampling range using a Monte Carlo algorithm to obtain a cooling random sample set; Importing the cooling random sample set into the PSO algorithm and performing an initialization operation on the PSO algorithm, wherein the PSO algorithm selects a cooling random sample from the cooling random sample set for initialization, and obtains an initial particle position set of the cooling random sample; The cooling conflict objective function is introduced into the PSO algorithm, and the initial particle position set is optimized and updated iteratively based on the PSO algorithm. When the update iteration converges, the optimal control parameter combination is output.

6. The method for optimizing liquid cooling fusion energy consumption based on dynamic regulation according to claim 5 is characterized in that: A random sampling range is obtained based on the cooling conflict objective function, and a cooling random sample set is obtained by sampling based on the random sampling range using a Monte Carlo algorithm, including: Determine a set of control variables to be optimized based on the cooling conflict objective function, obtain an upper limit of a feasibility range and a lower limit of a possibility range based on the set of control variables to be optimized, and obtain the random sampling range based on the upper limit of the feasibility range and the lower limit of the possibility range; The random sampling range is introduced into the Monte Carlo algorithm, the total cooling random samples are obtained based on the Monte Carlo algorithm, the total cooling random samples are randomly selected based on the selection method of the Monte Carlo algorithm, and the randomly selected cooling random samples are combined into a cooling random sample set.

7. The method for optimizing liquid cooling fusion energy consumption based on dynamic regulation according to claim 6 is characterized in that: The cooling conflict objective function is introduced into the PSO algorithm, and the initial particle position set is optimized and updated iteratively based on the PSO algorithm. When the update iteration converges, the optimal control parameter combination is output, including: Performing particle evaluation on the initial particle position set based on the PSO algorithm; Performing empirical updating on the initial particle position set based on the PSO algorithm; Performing motion update on the initial particle position set based on the PSO algorithm; When the PSO algorithm updates and iterates the initial particle position set, the convergence is checked. If converged, the update iteration is stopped and the optimal control parameter combination is output. If not converged, the update iteration is continued.

8. A liquid cooling fusion energy consumption adjustment and optimization system based on dynamic adjustment, used to implement the method according to any one of claims 1 to 7, characterized in that: The system comprises: An acquisition module, the acquisition module being used to acquire computing device thermal load data, liquid cooling system status data, air cooling system status data, and system constraint data; A map module, which is used to construct a real-time state map of cooling and fusion; A model module, the model module is used to extract key dynamic feature vectors of cooling and obtain short-term cooling prediction data, and to construct a cooling conflict objective function; An optimization module, wherein the optimization module generates an optimal control parameter combination based on a PSO algorithm; A control module, wherein the control module controls the cooling system in real time based on an optimal control parameter combination; A loop module is used to execute steps S1 to S4 in a loop.

Images