A refrigeration efficient machine room intelligent scheduling and deep energy-saving control method based on AI and optimization algorithm
By employing intelligent scheduling methods based on AI and optimization algorithms, real-time data collection, demand prediction, modeling, and proactive optimization are achieved, solving the problems of high energy consumption and poor adaptability of traditional data center control systems. This enables efficient, stable operation and deep energy saving in the cooling room.
Patent Information
- Authority / Receiving Office
- CN · China
- Patent Type
- Applications(China)
- Current Assignee / Owner
- TENGSHENG HUIXIN (XIAN) INFORMATION TECHNOLOGY CO LTD
- Filing Date
- 2025-11-21
- Publication Date
- 2026-07-03
AI Technical Summary
Traditional data center control systems rely on human experience and lack objective data, resulting in high energy consumption, low efficiency, inability to adapt to dynamic changes, lack of global optimization capabilities, and existing algorithms that rely on large amounts of data lack foresight and are difficult to achieve deep energy saving.
An intelligent scheduling method based on AI and optimization algorithms is adopted. Through real-time data collection, end-point demand prediction, equipment and system modeling, proactive optimization and precise execution, combined with machine learning adaptive evolution, collaborative optimization and parameter adjustment between equipment are achieved to achieve optimal global energy efficiency.
It achieves proactive optimization of the refrigeration room system, improves the system's energy efficiency and decision-making intelligence in dynamic environments, solves the problems of energy efficiency degradation and insufficient adaptability in traditional control methods, and achieves a balance between deep energy saving and system resilience.
Smart Images
Figure CN122340756A_ABST
Abstract
Description
Technical Field
[0001] This invention relates to the field of high-efficiency computer room technology for central air conditioning refrigeration systems, and more specifically to a method for intelligent scheduling and deep energy-saving control of high-efficiency refrigeration computer rooms based on AI and optimization algorithms. Background Technology
[0002] Against the backdrop of global warming and an increasingly severe energy crisis, improving energy efficiency has become one of the most pressing global issues. The design and renovation of high-efficiency data centers in the central air conditioning sector, which accounts for a significant portion of energy consumption, is a key measure to achieve this goal.
[0003] Traditional data center control systems typically face several problems in actual operation: First, the setting of operating parameters (such as chilled water temperature, cooling water temperature, and equipment start-up / shutdown combinations) heavily relies on the experience and judgment of maintenance personnel, lacking objective data as a basis for decision-making, making it difficult to continuously maintain the system's optimal operating conditions. Second, there is often a lack of effective coordination between various equipment units, such as chillers, pumps, and cooling towers, achieving only local optimization while ignoring overall efficiency. For example, increasing the chilled water temperature to reduce the main unit's energy consumption may actually cause a significant increase in the energy consumption of the terminal pumps, ultimately leading to a decrease in the overall energy efficiency of the system. Third, the system responds slowly to dynamic changes in building load, failing to adjust operating strategies in a timely manner based on real-time and predicted loads, resulting in frequent occurrences of "over-cooling" and ineffective energy consumption. Fourth, system operating parameters are usually set and remain unchanged for a long period, while equipment performance gradually declines with seasonal changes, increased service life, and changes in the external environment. Fixed operating strategies are difficult to adapt to such dynamic changes and lack the ability for continuous optimization.
[0004] To address the above issues: 1. Traditional fixed-condition design: Utilizing line-frequency equipment, relying on manual start-up, shutdown, and switching, with simple control logic, typically only ensuring cooling needs are met under the most unfavorable conditions, resulting in extremely high energy consumption, low efficiency, severe "over-protection," and a complete lack of adaptability; 2. Variable-condition manual experience: Inverters are installed, providing basic automated control such as differential pressure and temperature control. However, optimization strategies, such as finding the operating point with the lowest overall energy consumption, require manual adjustments by engineers based on experience, leading to over-reliance on personal experience, inability to achieve global and real-time optimization, and incomplete optimization; 3. Pure data-driven passive algorithm: The system collects massive amounts of operational data such as temperature, flow rate, and power, using regression, neural network, and other algorithms to train predictive models. Control strategies are based on data correlation, using search algorithms such as genetic algorithms to find the setpoint combination with the lowest predicted energy consumption. This involves a "black box" model, questionable reliability, reliance on large amounts of high-quality data, passive response lacking foresight, and potential getting trapped in local optima. All three of these computer room designs and renovations no longer meet current energy-saving and environmental protection requirements.
[0005] To address the shortcomings of the aforementioned technologies, this patent proposes an intelligent scheduling and deep energy-saving control method for high-efficiency cooling data centers based on AI and optimization algorithms. This method transforms hardware into software, real-time control from a fixed to a real-time approach, and the system from a black-box to a white-box approach, shifting from passive to proactive control. Through a closed-loop intelligent control logic of "perception-decision-optimization-execution-learning," it comprehensively considers various factors such as load matching and system energy consumption rationality while meeting the terminal cooling load requirements. This rationally controls various devices and parameters within the system, thereby achieving a proactive, optimization-based, deep energy-saving control system for high-efficiency data centers. Summary of the Invention
[0006] In view of this, the present invention provides an intelligent scheduling and deep energy-saving control method for high-efficiency cooling room based on AI and optimization algorithms. The main idea is to model the adjustable and controllable variables of specific equipment and the instantaneous power of the equipment, and obtain the corresponding input-output characteristic curves of the equipment. Based on this, the system balances the cooling capacity constraints among the main unit, cooling water pump, cooling tower, chilled water pump, and terminal demand, while meeting user needs. The adjustable parameters of each equipment, such as on or off, input frequency, supply and return water temperature, are adjusted to complete the corresponding arrangement and combination. The optimization is achieved by minimizing the total instantaneous power of the entire system, i.e., the power consumption, under a certain combination. The optimal setting parameters are then the combination, thereby solving the problems mentioned in the background art.
[0007] To achieve the above objectives, the present invention adopts the following technical solution:
[0008] A method for intelligent scheduling and deep energy-saving control of a cooling-efficient data center based on AI and optimization algorithms includes the following steps:
[0009] S1. Comprehensive perception and data fusion (perception): Real-time collection of multi-source and heterogeneous data from the data center, and preprocessing of the collected data;
[0010] S2. End-point demand forecasting and equipment and system modeling (decision making): End-point demand is forecasted using the classic time series forecasting algorithm SARIMAX. Simultaneously, equipment and system modeling are achieved through deep integration of physical mechanism models and data-driven algorithms.
[0011] S3. Active optimization: Using the particle swarm optimization algorithm, the optimal combination of operating parameters under the current operating conditions is actively and quickly searched in the global energy efficiency model.
[0012] S4. Precise Execution and Feedback Control (Execution): The optimal strategy is distributed to each execution device through the IoT platform to ensure precise system execution, so that the actual operating state of the system is infinitely close to the optimal set point;
[0013] S5. Machine Learning and Adaptive Evolution (Learning): Used after parameters are set, the model learns and optimizes itself, and corrects itself through feedback of results, making the model more accurate.
[0014] Preferably, in step S1, the sources and dimensions of the multi-source and heterogeneous data include: environmental parameters: outdoor temperature and humidity, weather forecast; system operating parameters: chilled water supply and return water temperature / pressure / flow rate, cooling water supply and return water temperature, equipment current / power, etc.; equipment status parameters: start / stop combination status of the main unit, chilled water pump, cooling water pump, and cooling tower, and corresponding equipment frequencies; chilled water supply and return water setpoint temperature; cooling water supply and return water setpoint temperature; unit COP (coefficient of performance), valve opening, filter differential pressure, etc. This massive amount of data provides a solid foundation for subsequent analysis and decision-making. Data preprocessing includes normalizing / standardizing the collected data.
[0015] Preferably, step S2 includes the following specific contents:
[0016] 1. End-user demand forecasting:
[0017] Using the classic time series forecasting algorithm SARIMAX, seasonality (S) and external variables (X), such as temperature, humidity, and holidays, can be well taken into account.
[0018] AR - Autoregressive: Predicts the current value using a linear combination of its past values.
[0019]
[0020] I - Difference: Transform a non-stationary time series into a stationary time series by calculating the difference between adjacent observations.
[0021]
[0022] MA - Moving Average: Improves current forecasts using a linear combination of past forecast errors.
[0023]
[0024] S - Seasonality: Based on the ARIMA model, AR, I, and MA components with seasonal cycles (such as 24 hours a day, 7 days a week, and 12 months a year) are explicitly introduced.
[0025]
[0026] X - Exogenous variables: Introducing external factors that influence the target variable as input features of the model, such as canonical variables:
[0027] Meteorological parameters: dry-bulb temperature, wet-bulb temperature, humidity, solar radiation intensity (these are extremely important exogenous variables);
[0028] Time characteristics: whether it is a weekday, holiday, or month (even if seasonality has been used, the month can provide additional information about the annual cycle).
[0029] Building information: building area, building type (although usually static, it is useful when modeling multiple buildings together).
[0030] 2. Equipment modeling (taking a water pump as an example; other equipment is similar and will not be elaborated on here):
[0031] (1) Physical mechanism model
[0032] ① Similarity Law of Water Pumps
[0033] Based on the core similarity law of the physical model of a water pump and its inherent efficiency characteristics (assuming constant efficiency):
[0034] Flow rate (Q) is directly proportional to rotational speed (n):
[0035]
[0036] The head (H) is proportional to the square of the rotational speed (n²):
[0037]
[0038] Shaft power (P_shaft) is proportional to the cube of the rotational speed (n³):
[0039]
[0040] ②From shaft power to electrical power
[0041]
[0042] ③ Establish a physical model
[0043] A. Determine the rated performance curve
[0044] Obtain the data table of (Q0, H0, η_pump0) for the rated frequency (50Hz) of this brand of water pump;
[0045] B. Calculate shaft power under rated operating conditions
[0046] For each data point (Q0, H0, η_pump0) in the table above, calculate its corresponding rated shaft power:
[0047]
[0048] C. Calculate parameters at the new frequency using the similarity law.
[0049] Assume the new frequency is f (e.g., 40Hz), and the rated frequency is f0 (50Hz). For each data point on the rated frequency curve:
[0050] Calculate new traffic:
[0051] Calculate the power of the new shaft:
[0052] Estimate the new head: (This step is used for subsequent load factor estimation)
[0053] D. Estimate the motor efficiency to obtain the electrical power.
[0054] Based on the "motor efficiency-load rate curve":
[0055] Calculate load factor:
[0056] The motor efficiency η_motor can be obtained by looking up a table or interpolating.
[0057] Calculate electrical power:
[0058] (2) Physical-data driven fusion
[0059] ① Physical model as a baseline predictor
[0060] Using the steps described in (1) above, a theoretical electric power P_physical is calculated for each running data point (f, Q_actual).
[0061] ② Data-driven model learning residuals
[0062] Collect a large amount of actual operating data: [frequency (f), flow rate (Q_actual), actual electrical power (P_actual)];
[0063] For each data point, calculate the residual: ;
[0064] To train a machine learning model (such as LightGBM), its:
[0065] Input features: frequency (f), flow rate (Q_actual), theoretical shaft power P_shaft;
[0066] Output target: the predicted residuals.
[0067] ③ Fusion Prediction
[0068] Final high-precision power prediction = P_physical(Q, f) + residual prediction(Q, f, ...)
[0069] 3. System Modeling
[0070] Based on the modeling of specific equipment, the pipeline network is taken into account to construct the resistance of the pipeline network to flow: Where H_static is the static head, a constant value. C is the system resistance coefficient, which integrates the effects of all pipe dimensions, lengths, roughness, and component resistance coefficients; it is an inherent characteristic of the system.
[0071] Preferably, in step S3, the optimal combination of operating parameters under the current operating conditions is actively and quickly searched in the global energy efficiency model using the particle swarm optimization algorithm. The basic idea is as follows:
[0072] (1) The algorithm is initialized as a group of random particles (random solutions). Each particle moves in the search space and adjusts its position according to its own experience and the experience of its companions.
[0073] (2) Each particle has a position vector and a velocity vector, as well as a fitness value determined by an optimization function (which can be similar to an objective function, which is determined by the specific device modeling results and the system modeling results).
[0074] Position matrix: An n×d matrix, where n is the number of particles and d is the problem dimension (variables to be considered, such as the start / stop status of the host, chilled water pump, cooling water pump, cooling tower and the corresponding equipment set frequency, chilled water supply and return water set temperature, cooling water supply and return water set temperature, etc.). Each row represents the position vector of a particle.
[0075] Velocity matrix: an n×d matrix, where each row represents the velocity vector of a particle.
[0076] When updating velocity and position, we can use matrix operations to update all particles at once, instead of using loops. This can significantly improve computational efficiency, especially when the number of particles and dimensionality are high.
[0077] Speed update matrix form:
[0078]
[0079] Position update matrix form:
[0080]
[0081] (3) Particles track two extreme values: individual extreme value (pbest, i.e. the best solution found by the particle itself) and global extreme value (gbest, i.e. the best solution found by the entire population so far).
[0082] For one of the particles, in each iteration, it explores only one point in the search space, which is a set of variable values in dimension d. Therefore, after each iteration, there is a new set of n candidate solutions (the current positions of all particles). At this point, the algorithm will actively obtain the global historical best solution (gbest) from this set of solutions and gradually approach the optimal solution by adjusting parameters and through multiple iterations.
[0083] Using the matrix operations above can avoid loops, significantly improve performance, and increase the number of particles or the problem dimension without affecting the code structure.
[0084] Preferably, in step S4, the optimal strategy is distributed to each execution device through the IoT platform to ensure accurate system execution and make the actual operating state of the system infinitely close to the optimal set point.
[0085] Preferably, in step S5, after the parameters are set, the model learns and optimizes itself, and corrects itself through result feedback, making the model more accurate.
[0086] A method for intelligent scheduling and deep energy-saving control of a high-efficiency cooling data center based on AI and optimization algorithms, comprising: a data acquisition module, a data preprocessing module, a terminal demand prediction module, an equipment and system modeling module, an active optimization module, a precise execution feedback control module, and a machine learning adaptive evolution module;
[0087] The data acquisition module is used to collect multi-source and heterogeneous data from the data center.
[0088] The data preprocessing module is used to preprocess the collected multi-source and heterogeneous data from the data center.
[0089] The end-user demand forecasting module is used to forecast end-user demand load in real time.
[0090] The equipment and system modeling module is used for equipment modeling of specific equipment such as water pumps, main units, and cooling towers, as well as overall system modeling, to achieve scientific mapping of source-load-driven input and output values.
[0091] The proactive optimization module is used to proactively and quickly search for the optimal combination of operating parameters under the current operating conditions in the global energy efficiency model;
[0092] The precise execution feedback control module is used to execute the strategy pushed by the active optimization module, adjust the parameters of the equipment, and make the actual operating state of the system infinitely close to the optimal set point.
[0093] The machine learning adaptive evolution module is used to enable the model to learn and optimize itself after the parameters are set, and to correct the model through feedback of results, so as to make the model more accurate.
[0094] A computer-readable storage medium having a computer program stored thereon, which, when executed by a processor, implements the aforementioned method for intelligent scheduling and deep energy-saving control of a cooling-efficient data center based on AI and optimization algorithms.
[0095] A processing terminal includes a memory and a processor. The memory stores a computer program that can run on the processor. When the processor executes the computer program, it implements the aforementioned method for intelligent scheduling and deep energy-saving control of a cooling-efficient computer room based on AI and optimization algorithms.
[0096] As can be seen from the above technical solution, compared with the prior art, this invention discloses a method for intelligent scheduling and deep energy-saving control of high-efficiency cooling room based on AI and optimization algorithms. By real-time prediction of terminal demand load and modeling of specific equipment such as water pumps, main units, and cooling towers, as well as the overall system modeling, it achieves a scientific mapping of source-load-driven input and output values. Using optimization algorithms, while ensuring user needs are met, the global energy efficiency model actively and quickly searches for the optimal combination of operating parameters under the current operating conditions, minimizing total energy consumption. Simultaneously, it accurately executes optimization strategies, and feedback allows the model to self-correct and become more accurate. Specifically:
[0097] This invention enables refrigeration room systems to leap from traditional passive response to AI-driven global proactive optimization. Through closed-loop control of "perception-decision-optimization-execution-learning", it unifies the operation of discrete equipment into a collaboratively optimized whole, significantly improving the predictability, coordination and self-evolution capabilities of system control.
[0098] This invention improves the overall energy efficiency and decision-making intelligence of the system under dynamically changing environments. Its hybrid modeling method, which integrates physical mechanisms and data-driven approaches, enables the system to maintain high-precision predictions and reasonable inferences even under unfamiliar operating conditions. Its global optimization and adaptive evolution capabilities are significantly superior to existing technologies that rely on fixed strategies or local optimization.
[0099] This invention ensures that the refrigeration system maintains high efficiency and stable operation throughout its entire life cycle, fundamentally solving the energy efficiency degradation problems caused by "source-load mismatch", model inaccuracy, strategy rigidity and equipment performance degradation in traditional control methods, and achieving a balance between deep energy saving and system resilience. Attached Figure Description
[0100] To more clearly illustrate the technical solutions in the embodiments of the present invention or the prior art, the drawings used in the description of the embodiments or the prior art will be briefly introduced below. Obviously, the drawings described below are only embodiments of the present invention. For those skilled in the art, other drawings can be obtained based on the provided drawings without creative effort.
[0101] Figure 1 A schematic diagram illustrating an intelligent scheduling and deep energy-saving control method for a high-efficiency cooling data center based on AI and optimization algorithms, provided by this invention;
[0102] Figure 2 A schematic diagram of the end-user demand forecasting method provided by the present invention;
[0103] Figure 3 A schematic diagram of the device modeling method provided by the present invention;
[0104] Figure 4 , Figure 5 The diagrams shown are the flow chart and process diagram of the optimization algorithm provided by this invention.
[0105] Figure 6 A schematic diagram of a system module for a method of intelligent scheduling and deep energy-saving control of a cooling-efficient computer room based on AI and optimization algorithms provided by the present invention; Detailed Implementation
[0106] The technical solutions of the embodiments of the present invention will be clearly and completely described below with reference to the accompanying drawings. Obviously, the described embodiments are only some embodiments of the present invention, and not all embodiments. Based on the embodiments of the present invention, all other embodiments obtained by those skilled in the art without creative effort are within the scope of protection of the present invention.
[0107] This invention discloses a method for intelligent scheduling and deep energy-saving control of cooling-efficient data centers based on AI and optimization algorithms, such as... Figures 1-5 As shown, it includes the following steps:
[0108] S1. Comprehensive perception and data fusion (perception): Real-time collection of multi-source and heterogeneous data from the data center, and preprocessing of the collected data;
[0109] S2. End-point demand forecasting and equipment and system modeling (decision making): End-point demand is forecasted using the classic time series forecasting algorithm SARIMAX. Simultaneously, equipment and system modeling are achieved through deep integration of physical mechanism models and data-driven algorithms.
[0110] S3. Active optimization: Using the particle swarm optimization algorithm, the optimal combination of operating parameters under the current operating conditions is actively and quickly searched in the global energy efficiency model.
[0111] S4. Precise Execution and Feedback Control (Execution): The optimal strategy is distributed to each execution device through the IoT platform to ensure precise system execution, so that the actual operating state of the system is infinitely close to the optimal set point;
[0112] S5. Machine Learning and Adaptive Evolution (Learning): Used after parameters are set, the model learns and optimizes itself, and corrects itself through feedback of results, making the model more accurate.
[0113] To further implement the above technical solution, in step S1, the sources and dimensions of the multi-source and heterogeneous data include: environmental parameters: outdoor temperature and humidity, weather forecast; system operating parameters: chilled water supply and return water temperature / pressure / flow rate, cooling water supply and return water temperature, equipment current / power, etc.; equipment status parameters: start / stop combination status of the main unit, chilled water pump, cooling water pump, and cooling tower, and corresponding equipment frequencies; chilled water supply and return water setpoint temperature; cooling water supply and return water setpoint temperature; unit COP (coefficient of performance); valve opening; filter differential pressure, etc. This massive amount of data provides a solid foundation for subsequent analysis and decision-making.
[0114] To further implement the above technical solution, data preprocessing includes normalizing / standardizing the collected data.
[0115] Based on the standardized data center data obtained in step S1, in order to further implement the above technical solution, the specific content of step S2 is as follows:
[0116] 1. End-user demand forecasting:
[0117] Using the classic time series forecasting algorithm SARIMAX, seasonality (S) and external variables (X), such as temperature, humidity, and holidays, can be well considered. Figure 2 As shown.
[0118] AR - Autoregressive: Predicts the current value using a linear combination of its past values.
[0119]
[0120] in, This represents the actual observed value of the time series at the current time t, i.e., the actual cooling load of the building, with a constant term. The autoregressive coefficient represents a long-term average level or baseline of the time series, i.e., the building's base load. These are the core parameters that the model needs to estimate. The value measured at the previous time step (t-1) The value at the current time (t) The extent and direction of the impact The values at the previous two time points (t-2) were measured. The effect on the current value follows the same pattern, until... . These represent historical observations of the time series over the past p moments; they are the input features of the model. This represents the random error term or white noise at the current time t, which represents the part of the model that cannot be explained by historical data, such as sudden events, random fluctuations, etc. It is usually assumed to have a mean of 0 and a constant variance.
[0121] I - Difference: Transform a non-stationary time series into a stationary time series by calculating the difference between adjacent observations.
[0122]
[0123] in, This represents the value of the new sequence obtained after first-order difference at time t;
[0124] MA - Moving Average: Improves current forecasts using a linear combination of past forecast errors.
[0125]
[0126] Among them, the moving average coefficient These are the core parameters that the model needs to estimate. The random shock measured at the previous time step (t-1) The value at the current time (t) The extent of its lasting impact. The random shocks at the previous two time points (t-2) were measured. The sustained impact until 。;
[0127] S - Seasonality: Based on the ARIMA model, AR, I, and MA components with seasonal cycles (such as 24 hours a day, 7 days a week, and 12 months a year) are explicitly introduced.
[0128]
[0129] in It is the order of the seasonal component, and its meaning is the same as that of the non-seasonal component. same, It is a seasonal cycle. s can be s=24: a daily cycle (predicting a 24-hour pattern), or s=24*7=168: a weekly cycle (predicting a 7-day week, 24-hour day pattern), and the annual cycle can be captured by introducing external variables (such as months, holidays);
[0130] X - Exogenous variables: Introducing external factors that influence the target variable as input features of the model, such as canonical variables:
[0131] Meteorological parameters: dry-bulb temperature, wet-bulb temperature, humidity, solar radiation intensity (these are extremely important exogenous variables);
[0132] Time characteristics: whether it is a weekday, holiday, or month (even if seasonality has been used, the month can provide additional information about the annual cycle).
[0133] Building information: building area, building type (although usually static, it is useful when modeling multiple buildings together).
[0134] 3. Equipment modeling (taking a water pump as an example, specifically as follows) Figure 3 As shown, other devices are similar, so we will not go into detail here.
[0135] (1) Physical mechanism model
[0136] ① Similarity Law of Water Pumps
[0137] Based on the core similarity law of the physical model of a water pump and its inherent efficiency characteristics (assuming constant efficiency):
[0138] Flow rate (Q) is directly proportional to rotational speed (n):
[0139]
[0140] The head (H) is proportional to the square of the rotational speed (n²):
[0141]
[0142] Shaft power (P_shaft) is proportional to the cube of the rotational speed (n³):
[0143]
[0144] in, These are the known rated parameters at the rated frequency (e.g., 50Hz). These are the parameters currently in operation. For a frequency converter, the electrical frequency f is directly proportional to the motor speed n, i.e., n / n0 ≈ f / f0;
[0145] ②From shaft power to electrical power
[0146]
[0147] in, It refers to motor efficiency, which is not constant but varies with the load rate (actual power / rated power). This is the inverter efficiency, which is usually high and does not vary much, and can be approximated as a constant (e.g., 0.96-0.98).
[0148] ③ Establish a physical model
[0149] A. Determine the rated performance curve
[0150] Obtain the data table of (Q0, H0, η_pump0) for the rated frequency (50Hz) of this brand of water pump;
[0151] B. Calculate shaft power under rated operating conditions
[0152] For each data point (Q0, H0, η_pump0) in the table above, calculate its corresponding rated shaft power:
[0153]
[0154] in, is the fluid density (water ≈ 1000 kg / m³), g is the acceleration due to gravity (9.81 m / s²), 3600 is the coefficient for hours to seconds (because Q0 is in m³ / h), and 1000 is the coefficient for watts to kilowatts (the result is in kW).
[0155] C. Calculate parameters at the new frequency using the similarity law.
[0156] Assume the new frequency is f (e.g., 40Hz), and the rated frequency is f0 (50Hz). For each data point on the rated frequency curve:
[0157] Calculate new traffic:
[0158] Calculate the power of the new shaft:
[0159] Estimate the new head: (This step is used for subsequent load factor estimation)
[0160] E. Estimate the motor efficiency and obtain the electrical power.
[0161] Based on the "motor efficiency-load rate curve":
[0162] Calculate load factor:
[0163] The motor efficiency η_motor can be obtained by looking up a table or interpolating.
[0164] Calculate electrical power:
[0165] (2) Physical-data driven fusion
[0166] ① Physical model as a baseline predictor
[0167] Using the steps described in (1) above, a theoretical electric power P_physical is calculated for each running data point (f, Q_actual).
[0168] ② Data-driven model learning residuals
[0169] Collect a large amount of actual operating data: [frequency (f), flow rate (Q_actual), actual electrical power (P_actual)];
[0170] For each data point, calculate the residual: ;
[0171] To train a machine learning model (such as LightGBM), its:
[0172] Input features: frequency (f), flow rate (Q_actual), theoretical shaft power P_shaft;
[0173] Output target: the predicted residuals.
[0174] ③ Fusion Prediction
[0175] Final high-precision power prediction = P_physical(Q, f) + residual prediction(Q, f, ...)
[0176] 3. System Modeling
[0177] Based on the modeling of specific equipment, the pipeline network is taken into account to construct the resistance of the pipeline network to flow: Where H_static is the static head, a constant value. C is the system resistance coefficient, which integrates the effects of all pipe dimensions, lengths, roughness, and component resistance coefficients; it is an inherent characteristic of the system.
[0178] To further implement the above technical solution, in step S3, the optimal combination of operating parameters under the current operating conditions is actively and rapidly searched in the global energy efficiency model using the particle swarm optimization algorithm. The optimization process is as follows: Figure 4 , Figure 5 As shown. The basic idea is as follows:
[0179] (1) The algorithm is initialized as a group of random particles (random solutions). Each particle moves in the search space and adjusts its position according to its own experience and the experience of its companions.
[0180] (2) Each particle has a position vector and a velocity vector, as well as a fitness value determined by an optimization function (which can be similar to an objective function, which is determined by the specific device modeling results and the system modeling results).
[0181] Position matrix: An n×d matrix, where n is the number of particles and d is the problem dimension (variables to be considered, such as the start / stop status of the host, chilled water pump, cooling water pump, cooling tower and the corresponding equipment set frequency, chilled water supply and return water set temperature, cooling water supply and return water set temperature, etc.). Each row represents the position vector of a particle.
[0182] Velocity matrix: an n×d matrix, where each row represents the velocity vector of a particle.
[0183] When updating velocity and position, we can use matrix operations to update all particles at once, instead of using loops. This can significantly improve computational efficiency, especially when the number of particles and dimensionality are high.
[0184] Speed update matrix form:
[0185]
[0186] Where: V is the velocity matrix (n×d), X is the position matrix (n×d), Pbest is the individual extreme value matrix (n×d), each row represents the individual optimal position of a particle. Gbest is the global extreme value vector (1×d), representing the optimal position found by the entire swarm, which is expanded into an n×d matrix during calculation through a broadcast mechanism. R1 and R2 are n×d random matrices, with each element between [0,1].
[0187] Position update matrix form:
[0188]
[0189] (3) Particles track two extreme values: individual extreme value (pbest, i.e. the best solution found by the particle itself) and global extreme value (gbest, i.e. the best solution found by the entire population so far).
[0190] For one of the particles, in each iteration, it explores only one point in the search space, which is a set of variable values in dimension d. Therefore, after each iteration, there is a new set of n candidate solutions (the current positions of all particles). The algorithm then actively obtains the global historical best solution (gbest) from this set of solutions. By adjusting the parameters and iterating multiple times, it gradually approaches the optimal solution. Plotting the fitness value of gbest as a function of the number of iterations shows the optimization process.
[0191] Using the matrix operations above can avoid loops, significantly improve performance, and increase the number of particles or the problem dimension without affecting the code structure.
[0192] To further implement the above technical solution, in step S4, the optimization strategy can be issued through the IoT platform, and at the same time, each execution device drives the frequency converter, valve, equipment start and stop to ensure that the actual operating state of the system is infinitely close to the optimal set point.
[0193] To further implement the above technical solution, in step S5, after the parameters are set, the model learns and optimizes itself, and corrects itself through result feedback, making the model more accurate.
[0194] A method for intelligent scheduling and deep energy-saving control of cooling-efficient data centers based on AI and optimization algorithms, such as Figure 5 As shown, it includes: a data acquisition module, a data preprocessing module, an end-user demand forecasting module, an equipment and system modeling module, an active optimization module, a precise execution feedback control module, and a machine learning adaptive evolution module;
[0195] The data acquisition module is used to collect multi-source and heterogeneous data from the data center.
[0196] The data preprocessing module is used to preprocess the collected multi-source and heterogeneous data from the data center.
[0197] The end-user demand forecasting module is used to forecast end-user demand load in real time.
[0198] The equipment and system modeling module is used for equipment modeling of specific equipment such as water pumps, main units, and cooling towers, as well as overall system modeling, to achieve scientific mapping of source-load-driven input and output values.
[0199] The proactive optimization module is used to proactively and quickly search for the optimal combination of operating parameters under the current operating conditions in the global energy efficiency model;
[0200] The precise execution feedback control module is used to execute the strategy pushed by the active optimization module, adjust the parameters of the equipment, and make the actual operating state of the system infinitely close to the optimal set point.
[0201] The machine learning adaptive evolution module is used to enable the model to learn and optimize itself after the parameters are set, and to correct the model through feedback of results, so as to make the model more accurate.
[0202] The embodiments in this specification are described in a progressive manner, focusing on the differences between these embodiments and those of other embodiments. Similar or identical parts between embodiments can be referred to interchangeably. For the apparatuses disclosed in the embodiments, since they correspond to the methods disclosed in the embodiments, the descriptions are relatively simple; relevant parts can be referred to the method section.
[0203] The above description of the disclosed embodiments enables those skilled in the art to make or use the invention. Various modifications to these embodiments will be readily apparent to those skilled in the art, and the general principles defined herein may be implemented in other embodiments without departing from the spirit or scope of the invention. Therefore, the invention is not to be limited to the embodiments shown herein, but is to be accorded the widest scope consistent with the principles and novel features disclosed herein.
Claims
1. A method for intelligent scheduling and deep energy-saving control of a high-efficiency cooling data center based on AI and optimization algorithms, characterized in that, Includes the following steps: S1. Comprehensive perception and data fusion (perception): Real-time collection of multi-source and heterogeneous data from the data center, and preprocessing of the collected data; S2. End-point demand forecasting and equipment and system modeling (decision making): End-point demand is forecasted using the classic time series forecasting algorithm SARIMAX. Simultaneously, equipment and system modeling are achieved through deep integration of physical mechanism models and data-driven algorithms. S3. Proactive Optimization: Utilizing the particle swarm optimization algorithm, the system proactively and rapidly searches for the optimal combination of operating parameters under the current operating conditions within the global energy efficiency model. S4. Precise Execution and Feedback Control (Execution): The optimal strategy is distributed to each execution device through the IoT platform to ensure precise system execution, so that the actual operating state of the system is infinitely close to the optimal set point; S5. Machine Learning and Adaptive Evolution (Learning): After the parameters are set, the model learns and optimizes itself, and corrects itself through feedback of results, making the model more accurate.
2. The method for intelligent scheduling and deep energy-saving control of a high-efficiency cooling data center based on AI and optimization algorithms as described in claim 1, characterized in that, In step S1, the sources and dimensions of the multi-source and heterogeneous data include: environmental parameters: outdoor temperature and humidity, weather forecasts; system operating parameters: chilled water supply and return water temperature / pressure / flow rate, cooling water supply and return water temperature, equipment current / power, etc.; equipment status parameters: start / stop combination status of the main unit, chilled water pump, cooling water pump, and cooling tower, and corresponding equipment frequencies, chilled water supply and return water setpoint temperatures, cooling water supply and return water setpoint temperatures, unit COP (coefficient of performance), valve opening, filter differential pressure, etc. This massive amount of data provides a solid foundation for subsequent analysis and decision-making.
3. The method for intelligent scheduling and deep energy-saving control of a high-efficiency cooling data center based on AI and optimization algorithms as described in claim 1, characterized in that, In step S1, data preprocessing includes normalizing / standardizing the collected data.
4. The method for intelligent scheduling and deep energy-saving control of a high-efficiency cooling data center based on AI and optimization algorithms as described in claim 1, characterized in that, Step S2 includes the following specific contents: (1) End-user demand forecasting: Using the classic time series forecasting algorithm SARIMAX, seasonality (S) and external variables (X), such as temperature, humidity, and holidays, can be well taken into account. AR (Autoregressive): Predicts the current value using a linear combination of its past values. I-Difference: Transforms a non-stationary time series into a stationary time series by calculating the difference between adjacent observations. MA (Moving Average): Improves current forecasts using a linear combination of past forecast errors. Y_t=c+ε_t-θ1ε_{t-1}-θ2ε_{t-2}-...-θ_qε_{tq} S-Seasonalization: Based on the ARIMA model, AR, I, and MA components with seasonal cycles (such as 24 hours a day, 7 days a week, and 12 months a year) are explicitly introduced. SARIMAX(p,d,q)(P,D,Q)[s] X-Exogenous variables: Introducing external factors that influence the target variable as input features of the model, such as canonical variables: Meteorological parameters: dry-bulb temperature, wet-bulb temperature, humidity, solar radiation intensity (these are extremely important exogenous variables); Time characteristics: whether it is a weekday, holiday, or month (even if seasonality has been used, the month can provide additional information about the annual cycle); Building information: building area, building type (although usually static, it is useful when modeling multiple buildings together). (2) Equipment modeling (taking water pumps as an example, other equipment is similar and will not be elaborated here): ①Physical mechanism model A. Similarity Law of Water Pumps Based on the core similarity laws of the physical model of a water pump and its inherent efficiency characteristics (assuming constant efficiency): Flow rate (Q) is directly proportional to rotational speed (n): Q / Q0=n / n0 Head (H) and the square of rotational speed (n) 2 Proportional to: H / H0=(n / n0) 2 Shaft power (P_shaft) versus the cube of the rotational speed (n) 3 Proportional to: P_shaft / P_shaft0=(n / n0) 3 B. From shaft power to electrical power P_electric=P_shaft / (η_motor*η_inverter) C. Establishing a physical model a. Determine the rated performance curve Obtain the data table of (Q0, H0, η_pump0) at the rated frequency (50Hz) of this brand of water pump; b. Calculate shaft power under rated operating conditions For each data point (Q0, H0, η_pump0) in the table above, calculate its corresponding rated shaft power: P_shaft0=(ρ*g*Q0*H0) / (3600*1000*η_pump0) c. Apply the similarity law to calculate parameters at the new frequency. Assuming the new frequency is f (e.g., 40Hz), the rated frequency f o (50Hz). For each data point on the rated curve: Calculate the new flow: Q = Q0 * (f / f0) Calculate the power of the new shaft: P_shaft = P_shaft0 * (f / f0) 3 Estimate the new head: H = H0 * (f / f0) 2 (This step is used for subsequent load factor estimation) d. Estimate the motor efficiency to obtain the electrical power. Based on the "motor efficiency-load rate curve": Calculate the load factor: Load factor = P_shaft / (motor rated power) The motor efficiency η_motor can be obtained by looking up a table or interpolating. Calculate electrical power: P_electric = P_shaft / (η_motor * η_inverter) ②Physics-Data Driven Fusion A. Physical model as a baseline predictor Using step ① above, calculate a theoretical electric power P_physical for each running data point (f,Q_actual). B. Data-driven model learning residuals Collect a large amount of actual operating data: [frequency (f), flow rate (Q_actual), actual electrical power (P_actual)]; For each data point, calculate the residual: Residual = P_actual - P_physical; Training a machine learning model (such as LightGBM), That: Input features: frequency (f), flow rate (Q_actual), theoretical shaft power P_shaft; Output target: the predicted residuals. C. Fusion Prediction Final high-precision power prediction = P_physical(Q,f) + residual prediction(Q,f,...) (3) System Modeling Based on the modeling of specific equipment, the pipeline network is taken into consideration, and the resistance of the pipeline network to the flow is constructed: H_system = H_static + C*Q 2 Where H_static is the static head, a constant value. C is the system resistance coefficient, which integrates the effects of all pipe dimensions, lengths, roughness, and component resistance coefficients; it is an inherent characteristic of the system.
5. The method for intelligent scheduling and deep energy-saving control of a high-efficiency cooling data center based on AI and optimization algorithms according to claim 1, characterized in that, In step S3, the optimal combination of operating parameters under the current operating conditions is actively and quickly searched in the global energy efficiency model using the particle swarm optimization algorithm. The basic idea is as follows: (1) The algorithm is initialized as a group of random particles (random solutions). Each particle moves in the search space and adjusts its position according to its own experience and the experience of its companions. (2) Each particle has a position vector and a velocity vector, as well as a fitness value determined by an optimization function (which can be similar to an objective function, determined by the specific device modeling results and the system modeling results). Position matrix: an n×d matrix, where n is the number of particles and d is the problem dimension (variables to be considered, such as the start / stop status of the host, chilled water pump, cooling water pump, cooling tower and the corresponding equipment set frequency, chilled water supply and return water set temperature, cooling water supply and return water set temperature, etc.). Each row represents the position vector of a particle. Velocity matrix: an n×d matrix, where each row represents the velocity vector of a particle. When updating velocity and position, we can use matrix operations to update all particles at once, instead of using loops. This can significantly improve computational efficiency, especially when the number of particles and dimensionality are high. Speed update matrix form: V=w·V+c1·R1·(Pbest-X)+c2·R2·(Gbest-X) Position update matrix form: X = X + V (3) Particles track two extreme values: the individual extreme value (pbest, i.e. the best solution found by the particle itself) and the global extreme value (gbest, i.e. the best solution found by the entire population so far). For one of the particles, in each iteration, it explores only one point in the search space, which is a set of variable values in dimension d. Therefore, after each iteration, there is a new set of n candidate solutions (the current positions of all particles). At this point, the algorithm will actively obtain the global historical best solution (gbest) from this set of solutions and gradually approach the optimal solution by adjusting parameters and through multiple iterations. Using the matrix operations above can avoid loops, significantly improve performance, and increase the number of particles or the problem dimension without affecting the code structure.
6. The method for intelligent scheduling and deep energy-saving control of a high-efficiency cooling data center based on AI and optimization algorithms according to claim 1, characterized in that, In step S4, the optimal strategy is distributed to each execution device through the IoT platform to ensure accurate system execution and make the actual operating state of the system infinitely close to the optimal set point.
7. The method for intelligent scheduling and deep energy-saving control of a high-efficiency cooling data center based on AI and optimization algorithms according to claim 1, characterized in that, In step S5, after the parameters are set, the model learns and optimizes itself, and corrects itself through feedback of results, making the model more accurate.
8. A method for intelligent scheduling and deep energy-saving control of a high-efficiency cooling data center based on AI and optimization algorithms, characterized in that, The method for intelligent scheduling and deep energy-saving control of a high-efficiency cooling room based on AI and optimization algorithms according to any one of claims 1-7 includes: a data acquisition module, a data preprocessing module, a terminal demand prediction module, an equipment and system modeling module, an active optimization module, a precise execution feedback control module, and a machine learning adaptive evolution module. The data acquisition module is used to collect multi-source and heterogeneous data from the data center. The data preprocessing module is used to preprocess the collected multi-source and heterogeneous data from the data center. The end-user demand forecasting module is used to forecast end-user demand load in real time. The equipment and system modeling module is used for equipment modeling of specific equipment such as water pumps, main units, and cooling towers, as well as overall system modeling, to achieve scientific mapping of source-load-driven input and output values. The proactive optimization module is used to proactively and quickly search for the optimal combination of operating parameters under the current operating conditions in the global energy efficiency model; The precise execution feedback control module is used to execute the strategy pushed by the active optimization module, adjust the parameters of the equipment, and make the actual operating state of the system infinitely close to the optimal set point. The machine learning adaptive evolution module is used to enable the model to learn and optimize itself after the parameters are set, and to correct the model through feedback of the results, so as to make the model more accurate.
9. A computer-readable storage medium having a computer program stored thereon, characterized in that, When executed by a processor, the computer program implements the AI-based and optimization algorithm-based intelligent scheduling and deep energy-saving control method for refrigerated high-efficiency computer rooms as described in any one of claims 1-7.
10. A processing terminal, comprising a memory and a processor, wherein the memory stores a computer program executable on the processor, characterized in that, When the processor executes the computer program, it implements the AI-based and optimization algorithm-based intelligent scheduling and deep energy-saving control method for refrigerated high-efficiency computer rooms as described in any one of claims 1-7.