Low-carbon energy efficiency optimization method for industrial cooling systems based on edge computing

By using edge computing and active ruminant detection, the problems of high energy consumption and production instability in industrial cooling systems under load fluctuations and equipment aging have been solved, achieving low-carbon energy efficiency optimization and improved production stability.

CN123117434APending Publication Date: 2026-11-17JIANGSU LANCHAUNG INFORMATION TECH SERVICESCO LTD
View PDF 0 Cites 0 Cited by

Patent Information

Application Number
CN202611270302.X
Authority / Receiving Office
CN · China
Patent Type
Applications(China)
Current Assignee / Owner
Filing Date
2026-08-20
Publication Date
2026-11-17

AI Technical Summary

Technical Problem

Existing industrial cooling systems cannot dynamically adjust to load fluctuations and equipment aging, resulting in high energy consumption, unstable production, and a lack of effective performance testing methods. Reliance on manual testing leads to a decrease in the accuracy of optimized control.

Method used

By employing an edge computing-based approach, energy efficiency models and knowledge graphs are constructed through data acquisition, cleaning, and modeling. This enables inter-device correlation analysis and load forecasting, combined with an active rumination detection mechanism for online optimization control, and dynamic adjustment of equipment operation strategies.

Benefits of technology

It achieves low-carbon energy efficiency optimization, reduces energy consumption, improves production stability and the real-time nature of equipment performance testing, and reduces the need for manual intervention.

✦ Generated by Eureka AI based on patent content.

Smart Images

  • Figure CN123117434A_ABST
    Figure CN123117434A_ABST
Patent Text Reader

Abstract

This invention discloses a low-carbon energy efficiency optimization method for industrial cooling systems based on edge computing. The method is executed on edge computing nodes deployed in industrial sites and includes the following steps: S1, collecting equipment operation data, system process data, environmental data, and production plan data of the industrial cooling system, performing data cleaning, interpolation, alignment, and normalization processing on the edge side to obtain a standardized dataset, and constructing a historical steady-state operating condition database; S2, constructing data-driven energy efficiency models for the chillers, circulating water pumps, and cooling tower fans in the cooling system based on historical operation data, characterizing the nonlinear correspondence between equipment input energy consumption and effective output; S3, predicting the system cooling load based on time-series deep learning, using an LSTM long short-term memory network to construct a cooling load prediction model, and outputting the predicted total system cooling load value for future multiple time scales; achieving the goal of energy reduction and carbon reduction.
Need to check novelty before this filing date? Find Prior Art

Description

Technical Field

[0001] This invention belongs to the field of low-carbon energy efficiency optimization of industrial cooling systems. Background Technology

[0002] Industrial cooling systems, which include core equipment such as refrigeration units, circulating water pumps, and cooling tower fans, undertake key functions such as process cooling and equipment temperature control. Their energy consumption typically accounts for 20% to 40% of the total electricity consumption of industrial enterprises, making them a core component of energy conservation and carbon reduction for enterprises.

[0003] However, many companies still use manual experience and fixed parameter control modes, such as fixed water pump operating frequency and fixed number of cooling towers to start and stop, and cannot dynamically adjust the operation strategy according to the fluctuation of process load and changes in environmental temperature and humidity.

[0004] Existing energy-saving retrofits are mostly carried out on single equipment, such as variable frequency retrofits of single chillers and speed regulation optimization of single fans, without considering the strong coupling characteristics between chillers, circulating water pumps, and cooling towers. Existing systems are mostly reactive after failures, lacking the ability to predict load trends and automatically analyze root causes, and cannot identify the risk of load increases in advance and intervene proactively, which can easily lead to fluctuations in process temperature or even production interruptions.

[0005] The cooling efficiency of core cooling towers naturally declines over time due to factors such as packing scaling, uneven water distribution, and fan blade wear. Existing technologies lack online, non-intrusive performance testing methods and rely on manual periodic shutdowns for testing. This not only affects production continuity but also has a long testing cycle, making it impossible to correct the energy efficiency model in a timely manner. Consequently, the accuracy of optimized control continues to decline as the equipment ages. Summary of the Invention

[0006] Purpose of the invention: In order to overcome the shortcomings of the existing technology, the present invention provides a low-carbon energy efficiency optimization method for industrial cooling systems based on edge computing, which achieves the effect of reducing energy consumption and carbon emissions.

[0007] Technical Solution: To achieve the above objectives, the present invention provides a low-carbon energy efficiency optimization method for industrial cooling systems based on edge computing. The method is executed on edge computing nodes deployed in industrial settings and includes the following steps:

[0008] S1. Collect equipment operation data, system process data, environmental data and production plan data of industrial cooling system, perform data cleaning, interpolation, alignment and normalization processing on the edge side to obtain standardized dataset, and build historical steady-state operating condition library.

[0009] S2. Data-driven energy efficiency models are constructed for the chiller, circulating water pump, and cooling tower fan in the cooling system based on historical operating data to characterize the nonlinear relationship between equipment input energy consumption and effective output.

[0010] S3. System cold load prediction based on time-series deep learning: A cold load prediction model is constructed using an LSTM long short-term memory network to output the predicted total system cold load value for future multiple time scales.

[0011] S4. Construct knowledge graphs of inter-device relationships and internal parameter relationships, respectively, and quantify the strength of inter-device collaborative relationships and the sensitivity of parameter influence.

[0012] S5. Based on the cooling load forecast results, perform load over-limit judgment at both the equipment level and the system level, and identify two types of load trends: short-term increase and long-term increase.

[0013] S6. Based on knowledge graph, abnormal load cause analysis: when an upward load trend is identified, the specific cause of the load increase is automatically located by tracing back the source through the knowledge graph.

[0014] S7. The safety warning level is determined by comprehensively considering the type of load increase, the number of overloaded devices, and the severity of the cause.

[0015] S8. With the goal of minimizing the total energy consumption of the system, a mixed integer nonlinear programming problem is constructed based on the energy efficiency model and the predicted cooling load, and the optimal equipment operation scheme is obtained by solving the problem.

[0016] S9. After verifying the safety boundary of the optimal operating scheme, the scheme is sent to the field execution equipment through the industrial protocol to achieve closed-loop automatic control.

[0017] S10. Periodically calculate the deviation between the actual load and the predicted load, execute the corresponding correction strategy according to the deviation level, and quantify the cooling tower performance degradation online through the cooling tower energy efficiency active ruminant detection mechanism based on wet-bulb temperature anchor point matching, calibrate the energy efficiency model, and periodically trigger model retraining to maintain accuracy.

[0018] The cooling tower energy efficiency proactive ruminant detection mechanism based on wet-bulb temperature anchor point matching specifically involves real-time acquisition of the current ambient wet-bulb temperature, retrieval of historical anchor point operating conditions in the historical steady-state operating condition database where the wet-bulb temperature deviation is within the threshold, the time interval meets the minimum requirement, and both times are in a steady state, and recording the corresponding cooling tower's return water temperature, return water flow rate, and fan frequency benchmark parameters; under the premise of ensuring stable process cooling supply, actively adjusting the corresponding water pumps, valves, fan frequencies, and chiller operating status and parameters to reproduce the target cooling tower's inlet operating conditions to the historical anchor point's benchmark parameter level, and acquiring the current outlet water temperature after stabilization; then comparing the current outlet water temperature difference with the historical anchor point to calculate the cooling efficiency decay rate and quantify the degree of decay of the cooling tower's inherent performance; and handling according to the degree of decay: mild decay automatically calibrates the energy efficiency model and knowledge graph parameters, moderate and above decay triggers corresponding level maintenance warnings, and optimizes the cooling tower load scheduling priority.

[0019] In step S2: The chiller energy efficiency model takes the cooling load and cooling water inlet temperature as inputs and the chiller input power as output, and is constructed using a gradient boosting tree or a multilayer perceptron; the circulating water pump energy efficiency model takes the output flow rate and pipeline pressure difference as inputs and the pump input power as output, and is constructed using a multinomial regression or a random forest; the cooling tower fan energy efficiency model takes the circulating water return temperature, ambient wet-bulb temperature, and fan operating frequency as inputs and the cooling tower outlet water temperature and fan input power as outputs, and is constructed using a neural network for online dynamic calibration.

[0020] S3 includes: constructing historical cooling load lag characteristics, rolling statistical characteristics, and environmental and production-related characteristics as model inputs; adopting a two-layer LSTM network structure, performing a prediction every 5 minutes, and outputting the cooling load prediction value for the next 30 minutes to 4 hours; the model evaluation index is mean absolute percentage error <5% and coefficient of determination R²>0.94.

[0021] S5 sets four threshold levels for individual devices and the entire system: rated load, maximum safe load, warning threshold, and alert threshold. The warning threshold is 0.85 times the rated load, and the alert threshold is 0.95 times the rated load. The type of rise is determined by the duration of the trend: a duration of less than 2 hours is a short-term rise, and a duration of 2 hours or more is a long-term rise. At the same time, the strength of the upward trend is calculated to characterize the average magnitude of the exceedance.

[0022] In S6, the reasons for load increases are divided into short-term and long-term causes: short-term causes include instantaneous process cooling load peaks, equipment start-up and shutdown disturbances, sensor data noise, and transient responses to control parameter adjustments; long-term causes include production plan load increases, inherent performance degradation of cooling towers, seasonal increases in ambient temperature and humidity, and changes in downstream process parameters; among them, the inherent performance degradation of cooling towers is quantitatively determined through an active rumination detection mechanism.

[0023] In S7, the warning levels are from low to high: Blue Level I, Yellow Level II, Orange Level III, and Red Level IV. The basic warning level is first calculated based on the number of over-limit devices, and then multiplied by the severity coefficient of the cause to obtain the final level.

[0024] In S8, the objective function is to minimize the total input power; the decision variables include the start-stop status of each device, the operating frequency of the variable frequency device, and the temperature setting parameters; the constraints include cooling load demand constraints, equipment capacity constraints, process temperature constraints, equipment safe operation constraints, and active rumination avoidance constraints.

[0025] Beneficial effects: The entire process of this invention is carried out at the edge nodes on site, without relying on cloud transmission. The control response latency is reduced to the millisecond level, and it can still operate independently in a network outage environment. At the same time, production data does not leave the factory area, avoiding data security risks from the architecture level and adapting to the complex network environment of industrial sites.

[0026] By combining LSTM time-series prediction with knowledge graph root cause analysis, the risk of load increase can be identified several hours in advance and the cause can be automatically located. Through hierarchical early warning and pre-adjustment strategies, passive fault handling can be transformed into proactive risk prevention and control, effectively reducing the range of process temperature fluctuations and improving production stability.

[0027] An active rumination detection mechanism for cooling tower energy efficiency based on wet-bulb temperature anchor point matching is introduced. Using wet-bulb temperature as a unified anchor point, the inherent performance of the cooling tower is detected online without human intervention by actively reproducing the operating conditions. Efficiency degradation caused by packing scaling, fan wear, etc. can be quantified without shutdown, and the energy efficiency model can be calibrated in a timely manner. This solves the pain points of traditional manual detection being lagging and affecting production, and ensures long-term optimized control accuracy. Attached Figure Description

[0028] Figure 1 This is a schematic diagram of the overall process architecture of the energy efficiency optimization method described in this invention;

[0029] Figure 2 This is a schematic diagram of the process of the active ruminant detection mechanism for cooling tower energy efficiency based on wet-bulb temperature anchor point matching of the present invention.

[0030] Figure 3 This is a schematic diagram of the load warning and cause analysis process of the present invention. Detailed Implementation

[0031] The following is in conjunction with the appendix Figures 1 to 3 The invention will be further described below.

[0032] To achieve the above objectives, this invention provides an energy efficiency optimization method for industrial cooling systems based on edge computing. The method is executed on edge computing nodes deployed in industrial settings and specifically includes the following steps:

[0033] S1. Collect comprehensive operational data from industrial cooling systems and perform standardized preprocessing at the edge, specifically including:

[0034] Equipment operating data: input power of the chiller, inlet and outlet water temperature / flow rate on the evaporator side, inlet and outlet water temperature / flow rate on the condenser side, and operating load rate; input power, inlet and outlet pressure, output flow rate, and operating frequency of the circulating water pump; input power, operating frequency, and inlet and outlet water temperature of the cooling tower fan.

[0035] System process data: chilled water main supply / return water temperature and main flow rate; cooling water main supply / return water temperature and main flow rate.

[0036] Environmental data: ambient temperature, ambient humidity, and ambient wet-bulb temperature.

[0037] Production planning data: downstream process operating load rate and production scheduling information.

[0038] Preprocessing steps: Abnormal jump data are removed using the 3σ criterion, missing values ​​are filled by linear interpolation, multi-source data are aligned to a 5-minute time granularity, and after max-min normalization, a standardized dataset is formed and stored in the local time series database of the edge node; at the same time, steady-state operating condition data are filtered and labeled to form a historical steady-state operating condition library for subsequent anchor point matching.

[0039] S2 constructs data-driven energy efficiency models for the three core energy-consuming devices in the cooling system, accurately describing the nonlinear relationship between device input energy consumption and effective output. All models are serialized and stored in the edge node model library, supporting local real-time inference.

[0040] The chiller energy efficiency model is constructed using the chiller cooling load RT and cooling water inlet temperature as inputs and the chiller input power as output. It is fitted with the energy consumption characteristics of the chiller under different operating conditions by using gradient boosting tree (XGBoost) or multilayer perceptron (MLP).

[0041] The circulating water pump energy efficiency model is constructed using the circulating water pump output flow rate and pipeline pressure difference as inputs and the pump input power as output. It reflects the energy consumption pattern of the pump under variable frequency operation.

[0042] Cooling tower fan energy efficiency model: The model takes the circulating water return temperature, ambient wet-bulb temperature, and fan operating frequency as inputs, and the cooling tower outlet water temperature and fan input power as outputs. It is constructed using a neural network to characterize the relationship between cooling tower cooling capacity and energy consumption. The model supports online dynamic calibration through a cooling tower energy efficiency active ruminant detection mechanism based on wet-bulb temperature anchor point matching, adapting to the natural performance degradation of the equipment.

[0043] S3. A system-level cooling load prediction model is constructed using an LSTM (Long Short-Term Memory) network to predict future cooling loads at multiple time scales on the edge side, providing feedforward basis for optimization decisions. The specific steps are as follows:

[0044] 1. Construct historical cooling load lag characteristics (cooling load values ​​of the past 24 5-minute cycles), rolling statistical characteristics (12-cycle rolling mean and rolling standard deviation), and external correlation characteristics (ambient temperature and humidity, production plan load rate, and downstream process operating status) to jointly form the model input feature vector.

[0045] 2. A two-layer LSTM network structure is adopted, with the hidden layer dimension set to 64, and the output layer outputting the predicted cooling load values ​​for the next 8 cycles (corresponding to 4 hours).

[0046] 3. Divide the training, validation, and test sets into a ratio of 7:1.5:1.5. Use the Adam optimizer (learning rate 0.001) and mean squared error (MSE) as the loss function for training. Introduce an early stopping mechanism with patience=10 to prevent overfitting. The model accuracy requirements are mean absolute percentage error (MAPE) < 5% and coefficient of determination (R²) > 0.94.

[0047] 4. Edge nodes perform a prediction every 5 minutes, outputting the core prediction value for the next 30 minutes and the trend value for the next 4 hours.

[0048] S4. Construct two types of relational knowledge graphs to characterize the collaborative relationships between devices and the correlation relationships of parameters within devices, respectively, and store them in a graph database embedded in the edge nodes for anomaly root cause tracing and operational pattern mining:

[0049] Knowledge graph of inter-equipment relationships: Each individual piece of equipment within the cooling system is represented as an entity, including sets of chiller entities, circulating water pump entities, and cooling tower fan entities. Any two equipment entities e... i e j The formula for association weight is:

[0050] w ij =α·orr(L i ,L j )+β·causality(L i →L j The correlation coefficient is defined as follows: ) where corr is the Pearson correlation coefficient, reflecting the synchronicity of load changes; causality is the Granger causality test result, determining whether load fluctuations in device i are the cause of load changes in device j; α and β are weighting coefficients, with default values ​​of α=0.6 and β=0.4. Edge nodes iteratively update their correlation weights daily based on the day's operational data.

[0051] Equipment internal parameter correlation knowledge graph: For a single piece of equipment, identify its core operating parameter entities, such as the chiller's cooling load, cooling water inlet temperature, input power, and evaporation pressure; and the cooling tower's return water temperature, wet-bulb temperature, fan frequency, and outlet water temperature. Define three types of logical relationships between parameters: causal relationship (CAUSE), constraint relationship (CONST), and computational derivation relationship (DERIVE).

[0052] Sensitivity weight calculation: parameter p ia With p ib The formula for association weight is:

[0053]

[0054] Where sensitivity is the sensitivity coefficient, which characterizes the degree of influence of a unit change in parameter a on parameter b; , For coefficients, the default value is... =0.5, =0.5.

[0055] Dynamic calibration: The sensitivity weights of cooling tower parameters are updated synchronously with the results of the active ruminant detection mechanism.

[0056] S5. Based on the cooling load forecast results, load threshold verification and upward trend identification are performed at both the equipment and system levels to provide a basis for early warning and optimization:

[0057] 1. For individual equipment and the overall cooling system, set separately:

[0058] Rated load L r : Rated output of the equipment nameplate or rated load of the system design.

[0059] Maximum safe load L m The maximum safe operating load allowed for the equipment.

[0060] Warning threshold L w =0.85×L r Warning threshold L a =0.95×L r .

[0061] The system-level threshold is the sum of the rated loads of all operating devices multiplied by the safety factor η, with a default value of 0.9.

[0062] 2. Exceedance judgment: Obtain the predicted load value for the next operating cycle, compare it with the corresponding threshold to generate an exceedance flag: 0 represents normal, 1 represents warning status, and 2 represents alert status.

[0063] 3. Trend type identification: Obtain the predicted load sequence for the next 8 periods and calculate the trend duration T. d ;

[0064]

[0065] Where 𝐼 is an indicator function, which takes the value 1 if the condition is met, and 0 otherwise.

[0066] Judgment rule: T d <4 (duration <2 hours) indicates a short-term upward trend; T d ≥4 indicates a long-term upward trend.

[0067] Simultaneously calculate the upward trend strength S t The average value representing the out-of-limit amplitude:

[0068]

[0069] S6. Analysis of abnormal load causes based on knowledge graphs: When an upward trend in load is identified, the cause of the anomaly is automatically located by reverse tracing based on the relational knowledge graph. The steps are as follows:

[0070] Identify the target equipment or subsystem experiencing overload; starting from the target entity, reverse-search all associated edges with a weight > 0.5 in the equipment knowledge graph to obtain the set of precursor equipment; verify the current load and load change rate of the precursor equipment, and investigate parameter anomalies by combining the equipment's internal parameter graph; for load anomalies related to the cooling tower, call the quantitative results of the cooling tower energy efficiency active ruminant detection mechanism based on wet-bulb temperature anchor point matching to determine whether it is caused by equipment performance degradation.

[0071] Reasons for the brief increase: instantaneous peak of process cooling load, equipment start-up and shutdown disturbances, sensor data noise, and transient response of control parameter adjustment.

[0072] Reasons for long-term increases: increased production plan load, inherent performance degradation of cooling towers, seasonal increase in ambient temperature and humidity, and changes in downstream process parameters.

[0073] S7. Establish a four-level early warning system, and comprehensively determine the early warning level based on the type of load increase, the number of overloaded devices, and the severity of the cause:

[0074] The risk levels are divided into four categories from low to high: Level I (LOW, blue, low risk), Level II (MEDIUM, yellow, medium risk), Level III (HIGH, orange, high risk), and Level IV (CRITICAL, red, extremely high risk), corresponding to numerical codes 1 to 4.

[0075] Brief increase scenario: Count the number of devices exceeding the limit N s N s =0 represents LOW, N s =1 represents MEDIUM, 2≤N s ≤4 indicates HIGH, N s >4 represents CRITICAL;

[0076] Long-term upward scenario: Counting the number of devices exceeding limits, N l N l =0 represents LOW, N l =1 represents MEDIUM, 2≤N l ≤3 represents HIGH, N l >3 is CRITICAL;

[0077] Hybrid upgrade scenario: Take the maximum value of the two scenario levels as the base level for fusion. If the total number of over-limit devices is >5, then increase by 1 level.

[0078] Severity Correction: Introduce a cause severity coefficient C sThe value ranges from 0.5 to 2.0. For example, 0.5 corresponds to sensor noise, 1.0 corresponds to equipment start-up and shutdown disturbances, 1.2 corresponds to instantaneous process peaks, 1.5 corresponds to moderate performance degradation of cooling towers, and 2.0 corresponds to severe performance degradation of cooling towers.

[0079] S8. With the goal of minimizing the total system energy consumption, based on the cooling load prediction results and equipment energy efficiency models, a mixed-integer nonlinear programming problem is constructed and solved in real time at the edge:

[0080] Objective function: Minimize the total input power of the cooling system

[0081] minE total =∑P chiller +∑P pump +∑P fan

[0082] The energy consumption of each device is calculated using the energy efficiency model in step S2, wherein the cooling tower energy efficiency model uses the latest parameters calibrated by the active ruminant mechanism.

[0083] Decision variables: Discrete variables are the start / stop status (0 / 1) of each chiller, circulating water pump, and cooling tower fan; continuous variables are the operating frequency of the variable frequency equipment, the temperature difference between chilled water supply and return water, and the target temperature of cooling water.

[0084] Constraints: Cooling load demand constraint (total system cooling capacity ≥ predicted total cooling load), equipment capacity constraint (output of a single unit is within the rated range), process constraint (chilled water supply temperature ≤ process upper limit, cooling water outlet temperature ≤ condensing temperature upper limit), safety constraint, and active rumination avoidance constraint. When performing active rumination detection, priority should be given to ensuring the stability of the chilled water supply temperature, and the process red line must not be exceeded.

[0085] A mixed-integer nonlinear programming solver is used to solve the problem every 5 minutes, outputting the optimal equipment start-up and shutdown combination, the operating frequency of the variable frequency equipment, and the load distribution scheme. When performing active rumination detection, the control parameters of a single target cooling tower are temporarily adjusted, while the remaining equipment continues to operate according to the global optimal strategy.

[0086] S9. Before the optimization strategy is issued, the edge node first performs boundary verification to confirm that all control commands are within the safe operating parameters of the equipment and will not cause the system pressure or temperature to exceed the limit. Before the active rumination condition reproduction is executed, the process cooling stability is additionally verified to confirm that the chilled water supply temperature fluctuation during the adjustment process does not exceed ±0.5℃.

[0087] After verification, the edge node sends control commands to the field PLC or equipment frequency converter via industrial bus protocols such as Modbus and Profinet to realize automatic control of chiller load adjustment, water pump frequency conversion speed regulation, and fan start-up, shutdown and speed regulation; when performing active rumination detection, it sends out condition reproduction and adjustment commands for the target cooling tower separately.

[0088] Real-time acquisition of system operation data after control execution forms a control closed loop.

[0089] S10 forecast deviation handling and model dynamic iteration: After each operating cycle, the actual load L is calculated. ac With the predicted load L pr The relative deviation δ: δ=(|L ac -L pr |) / L r If δ < 5%: normal, record the deviation data and add it to the sample library; 5% ≤ δ < 10%: slight deviation, trigger model health check, verify input data quality and sensor status; 10% ≤ δ < 20%: moderate deviation, start incremental model learning, use the data of the most recent 24 hours to fine-tune the model parameters with a small learning rate of 0.0001; δ ≥ 20%: serious deviation, switch to backup rule control strategy, and at the same time notify the operation and maintenance personnel.

[0090] Regular retraining is triggered when the average deviation over three consecutive days exceeds 10%, the number of severe deviations in a single week exceeds five, or the model has been running continuously for more than 90 days. The energy efficiency model and prediction model are then updated based on the latest historical data.

[0091] Based on the above, an active ruminant detection mechanism for cooling tower energy efficiency is implemented using wet-bulb temperature anchor point matching. This mechanism uses the ambient wet-bulb temperature as the operating condition anchor point, actively reproducing historical inlet operating conditions to compare the difference in outlet water temperature, and quantifying the degree of inherent performance degradation of the cooling tower online. The specific execution process is as follows:

[0092] Step 1: The system collects the ambient wet-bulb temperature T at the current time t1 in real time. wb (t1), retrieve the historical moment t0 in the historical steady-state operating condition database that satisfies all matching conditions:

[0093] Wet-bulb temperature deviation condition: |T wb (t1)-T wb (t0)|≤ΔT wb,th , where ΔT wb,th The default value for the allowable error threshold is 0.2℃.

[0094] Time interval condition: t1-t0≥T interval,min The minimum interval can be configured according to the scenario, with a default of 7 days to ensure that the time span is sufficient to reflect changes in device performance.

[0095] Steady-state operating conditions: The system is in steady-state operation at both time t0 and time t1. The criteria for judgment are: cooling load fluctuation amplitude <5% within 15 consecutive minutes, no equipment start-up or shutdown operations, and stable process parameters.

[0096] After retrieving the historical anchor point operating conditions that meet the criteria, record the inlet operating condition baseline parameters of the target cooling tower at time t0: circulating water return water temperature T. in (t0), circulating water return flow rate F in (t0), Cooling tower fan operating frequency f fan (t0).

[0097] Step 2:

[0098] Without exceeding the process cooling constraints or affecting normal production, the edge controller performs refined proactive adjustments to reproduce the inlet conditions of the target cooling tower to the baseline parameter level at time t0 within 5-10 minutes.

[0099] By adjusting the frequency of the corresponding circulating water pump and the opening degree of the branch valve, the return water flow rate F is increased. in (t1) approaches F in (t0), the flow deviation is controlled within ±3%;

[0100] By fine-tuning the ratio of condenser-side load to cooling water bypass in the chiller, the return water temperature T is adjusted. in (t1) approaches T in (t0), the temperature deviation is controlled within ±0.3℃;

[0101] Directly adjust the frequency converter of the target cooling tower fan to make the fan operating frequency f fan (t1)=f fan (t0), the frequency deviation is controlled within ±0.5Hz.

[0102] After all inlet parameters have stabilized, maintain steady-state operation for 3 minutes to ensure that the cooling tower outlet water temperature reaches a stable value.

[0103] Step 3:

[0104] The cooling tower outlet water temperature T at time t1 under steady-state conditions was collected. out (t1), and the reference value of the outlet water temperature T at time t0. out (t0) Comparison, calculate two core attenuation indicators:

[0105] Water outlet temperature deviation:

[0106] ΔT out =T out (t1)-T out (t0)

[0107] Cooling efficiency degradation rate:

[0108] η d =(ΔT out ) / (T in (t0)-T wb (t0))

[0109] in This represents the theoretical cooling width of the cooling tower under this operating condition. and The higher the value, the better the cooling tower performs. to The more severe the inherent performance degradation within the range, the higher the degree of degradation, such as packing scaling, uneven water distribution, and reduced fan efficiency.

[0110] Step 4:

[0111] Based on the attenuation index values, tiered measures are implemented to achieve linkage between model calibration and operation and maintenance early warning:

[0112] when <0.3℃: It is determined that there is no significant degradation in the performance of the cooling tower, and the original energy efficiency model parameters of the cooling tower remain unchanged;

[0113] When 0.3℃≤ <0.8℃: Determines slight performance degradation, automatically corrects the outlet water temperature prediction bias term of the cooling tower energy efficiency model, and synchronously updates the sensitivity weights of relevant cooling tower parameters in the equipment's internal parameter knowledge graph to offset the impact of performance degradation on optimization accuracy;

[0114] When 0.8℃≤ <1.5℃: Determined as moderate performance degradation, triggering a yellow maintenance warning, pushing operation and maintenance tips to check the status of cooling tower packing and water distributor blockage, while increasing the data weight of incremental learning of the cooling tower model to speed up model adaptation;

[0115] when ≥1.5℃: This indicates severe performance degradation and triggers an orange maintenance warning. It is recommended to arrange a shutdown for cleaning and maintenance as soon as possible. At the same time, optimize the scheduling engine to temporarily reduce the operating priority of this cooling tower. During global optimization, prioritize the allocation of load to cooling towers with better performance to ensure the overall energy efficiency of the system.

[0116] Scenarios and System Deployment:

[0117] This embodiment optimizes an industrial cooling system supporting the production process of a chemical plant. The system includes three centrifugal chillers, each with a rated cooling capacity of 1000 RT; four variable frequency circulating cooling water pumps, each with a rated flow rate of 2000 m³ / h; and six variable frequency cooling tower fans, each with an air volume of 30000 m³ / h. It is responsible for cooling all the plant's process equipment. The original system was controlled manually based on experience, with an average annual energy consumption of approximately 8.2 million kWh. The cooling towers required manual shutdown every six months to check the packing condition, necessitating reduced load operation during the inspection period, which impacted production efficiency.

[0118] This method is based on edge computing nodes deployed in the control room of the chiller plant. These edge nodes use domestically produced industrial-grade processors, equipped with 8GB of RAM and 256GB of local storage, and support industrial protocols such as ModbusTCP and Profinet. The system's physical architecture consists of three layers:

[0119] Sensing layer: Temperature sensors, flow sensors, pressure sensors, power acquisition modules, outdoor ambient temperature and humidity meters, and wet-bulb thermometers are deployed on various devices and pipelines to collect data in all dimensions, with a wet-bulb temperature sampling accuracy of ±0.1℃.

[0120] Edge computing layer: It carries all functions such as data preprocessing, energy efficiency modeling, load forecasting, knowledge graph, optimization solution, early warning analysis, active rumination detection, and control output, and is the core computing unit of the system.

[0121] Execution layer: Field PLCs and frequency converters and controllers of various equipment receive instructions from edge nodes and execute equipment adjustments.

[0122] Implementation process:

[0123] 1. The perception layer collects raw data every minute, covering chiller power and temperature parameters, water pump flow and frequency parameters, cooling tower temperature and fan parameters, system main pipe temperature, humidity and flow rate, plant area ambient temperature, humidity and wet-bulb temperature, and downstream production unit load rate. Edge nodes perform preprocessing on the raw data: outliers are removed using the 3σ criterion, missing data is completed using linear interpolation, and the data is uniformly aligned to a 5-minute granularity. After normalization, the data is stored in the local time-series database. At the same time, historical steady-state operating conditions are filtered out using a steady-state determination algorithm to construct a dedicated database for anchor point matching.

[0124] 2. Based on the system's historical operating data over the past year, energy efficiency models for three types of equipment were trained respectively:

[0125] Refrigeration unit energy efficiency model: The inputs are cooling load and cooling water inlet temperature, and the output is input power. The XGBoost algorithm is used. The MAPE of the test set is 3.2%, and R²=0.96.

[0126] Circulating water pump energy efficiency model: The inputs are output flow rate and pipeline pressure difference, and the output is input power. The model adopts quadratic polynomial regression, and the goodness of fit R² = 0.97.

[0127] Cooling tower fan energy efficiency model: The inputs are return water temperature, ambient wet-bulb temperature, and fan frequency, and the outputs are outlet water temperature and fan power. It adopts a two-layer MLP network, and the initial outlet water temperature prediction error is <0.5℃. It supports online calibration with active rumination mechanism.

[0128] 3. A two-layer LSTM network was used to construct a cooling load prediction model. The input included historical cooling load, ambient temperature and humidity, and production load rate for the past 24 five-minute periods. The output was the predicted value for the next 8 periods. The model batch size was 32, the training epochs were 100, the early shutdown patience was 10, and the final test set MAPE was 4.1%, with R²=0.95.

[0129] Edge nodes perform a rolling forecast every 5 minutes, outputting core forecast values ​​for the next 30 minutes for decision optimization and trend values ​​for the next 4 hours for risk warning.

[0130] 4. A knowledge graph of the cooling system's equipment was constructed, containing 13 equipment entities. Based on historical data calculations, the association weight between chiller c1 and its corresponding cooling water pump p1 is 0.82 (strong coupling), and the association weight between cooling tower fan f1 and chiller c1 is 0.71. Simultaneously, an internal parameter graph of the equipment was constructed. The sensitivity weight for the relationship between cooling tower wet-bulb temperature and outlet water temperature is 0.72, and the sensitivity weight for the relationship between fan frequency and outlet water temperature is 0.45. These sensitivity weights are updated monthly based on the results of active ruminant detection.

[0131] 5. When the system has been running for three months, the ambient wet-bulb temperature stabilizes at 22.1℃ one morning. The system automatically retrieves a steady-state anchor point with the same wet-bulb temperature that existed 28 days ago, which meets the minimum interval requirement and triggers active rumination detection.

[0132] The inlet parameters of cooling tower f2 under anchor point condition t0 are: return water temperature 32.5℃, return water flow rate 1800m³ / h, fan frequency 40Hz, corresponding to an outlet water temperature of 26.8℃.

[0133] The edge controller performs active adjustment, restoring the inlet parameters of cooling tower f2 to the reference value within 8 minutes. After steady-state operation, the current outlet water temperature is measured to be 27.7℃. The outlet water temperature deviation ΔT is calculated. out =0.9℃, the cooling efficiency decay rate is about 12.7%, which is judged as moderate performance degradation.

[0134] The system automatically triggers a yellow maintenance warning, prompting the user to check the scaling condition of the packing material in cooling tower f2. At the same time, it corrects the outlet water temperature bias term in the energy efficiency model of this cooling tower and updates the sensitivity of the corresponding parameters in the knowledge graph. The global optimization engine appropriately reduces the load allocation ratio of this cooling tower and prioritizes scheduling other cooling towers with better performance.

[0135] 6. The system performs real-time threshold judgment and trend identification based on prediction results. On a certain workday afternoon, the ambient temperature continued to rise. The system predicted that the total load of the chiller would remain above the warning threshold for the next 3 hours, classifying it as a long-term upward trend with a base level of MEDIUM. Through knowledge graph tracing and active rumination detection results, factors causing equipment performance degradation were ruled out. The cause was determined to be seasonal load growth due to rising ambient temperature, with a severity coefficient of 1.0, resulting in a final warning level of Yellow Level II. The system automatically pushes the warning information to maintenance personnel and simultaneously starts the backup cooling tower fans in advance to pre-adjust cooling capacity and prevent further load exceedances.

[0136] 7. Edge nodes perform a global optimization solution every 5 minutes, with the objective of minimizing the total system power. Constraints include chilled water supply temperature ≤7℃, cooling water outlet temperature ≤32℃, and equipment operating range limitations. Under typical operating conditions, the optimal solution is output as follows: Two chillers are activated at load rates of 75% and 82% respectively; three cooling water pumps are activated at frequencies of 42Hz, 45Hz, and 40Hz respectively; and four cooling tower fans are activated at frequencies of 38Hz, 42Hz, 35Hz, and 40Hz respectively. After safety verification, control commands are sent to the field PLC via Modbus TCP to automatically adjust the equipment operating status. During active rumination detection, the target cooling tower parameters are adjusted individually, while the remaining equipment maintains globally optimal operation.

[0137] 8. The system calculates prediction deviations daily, with an average daily deviation of approximately 3.5%. After 6 months of operation, combining the results of proactive rumination detection with deviation statistics, it was found that energy efficiency degradation was caused by scaling on heat exchanger tubes in some equipment and aging of cooling tower packing, leading to an average deviation of 11.2% over three consecutive days. The system automatically triggered incremental learning, using data from the past 30 days to fine-tune the energy efficiency model parameters, and the deviation dropped back to 4.3% after correction. The system automatically performs a full retraining every quarter, updating the model synchronously with performance degradation data detected by proactive rumination, ensuring long-term optimization accuracy.

[0138] The above description is only a preferred embodiment of the present invention. It should be noted that for those skilled in the art, several improvements and modifications can be made without departing from the principle of the present invention, and these improvements and modifications should also be considered within the scope of protection of the present invention.

Claims

1. A low-carbon energy efficiency optimization method for industrial cooling systems based on edge computing, characterized in that, The method is executed on edge computing nodes deployed in industrial sites and includes the following steps: S1. Collect equipment operation data, system process data, environmental data and production plan data of industrial cooling system, perform data cleaning, interpolation, alignment and normalization processing on the edge side to obtain standardized dataset, and build historical steady-state operating condition library. S2. Data-driven energy efficiency models are constructed for the chiller, circulating water pump, and cooling tower fan in the cooling system based on historical operating data to characterize the nonlinear relationship between equipment input energy consumption and effective output. S3. System cold load prediction based on time-series deep learning: A cold load prediction model is constructed using an LSTM long short-term memory network to output the predicted total system cold load value for future multiple time scales. S4. Construct knowledge graphs of inter-device relationships and internal parameter relationships, respectively, and quantify the strength of inter-device collaborative relationships and the sensitivity of parameter influence. S5. Based on the cooling load forecast results, perform load over-limit judgment at both the equipment level and the system level, and identify two types of load trends: short-term increase and long-term increase. S6. Based on knowledge graph, abnormal load cause analysis: when an upward load trend is identified, the specific cause of the load increase is automatically located by tracing back the source through the knowledge graph. S7. The safety warning level is determined by comprehensively considering the type of load increase, the number of overloaded devices, and the severity of the cause. S8. With the goal of minimizing the total energy consumption of the system, a mixed integer nonlinear programming problem is constructed based on the energy efficiency model and the predicted cooling load, and the optimal equipment operation scheme is obtained by solving the problem.

2. The low-carbon energy efficiency optimization method for industrial cooling systems based on edge computing according to claim 1, characterized in that: S9. After verifying the safety boundary of the optimal operating scheme, the scheme is sent to the field execution equipment through the industrial protocol to achieve closed-loop automatic control. S10. Periodically calculate the deviation between the actual load and the predicted load, execute the corresponding correction strategy according to the deviation level, and quantify the cooling tower performance degradation online through the cooling tower energy efficiency active ruminant detection mechanism based on wet-bulb temperature anchor point matching, calibrate the energy efficiency model, and periodically trigger model retraining to maintain accuracy.

3. The low-carbon energy efficiency optimization method for industrial cooling systems based on edge computing according to claim 2, characterized in that: In S10, the cooling tower energy efficiency active ruminant detection mechanism based on wet-bulb temperature anchor point matching specifically involves real-time acquisition of the current ambient wet-bulb temperature, searching the historical steady-state operating condition database for historical anchor point operating conditions where the wet-bulb temperature deviation is within the threshold, the time interval meets the minimum requirement, and both times are in a steady state, and recording the corresponding cooling tower's return water temperature, return water flow rate, and fan frequency reference parameters; under the premise of ensuring stable process cooling supply, by actively adjusting the corresponding water pumps, valves, fan frequencies, and chiller operating status and operating parameters, the inlet operating conditions of the target cooling tower are reproduced to the historical anchor point reference parameter level, and the current outlet water temperature is acquired after stabilization; Then, by comparing the current water temperature difference with the historical anchor point, the cooling efficiency decay rate is calculated, and the degree of decay of the inherent performance of the cooling tower is quantified. According to the degree of decay, the energy efficiency model and knowledge graph parameters are automatically calibrated for mild decay, and maintenance warnings of the corresponding level are triggered for moderate and above decay, and the priority of cooling tower load scheduling is optimized.

4. The low-carbon energy efficiency optimization method for industrial cooling systems based on edge computing according to claim 1, characterized in that: In step S2: The chiller energy efficiency model takes the cooling load and cooling water inlet temperature as inputs and the chiller input power as output, and is constructed using a gradient boosting tree or a multilayer perceptron; the circulating water pump energy efficiency model takes the output flow rate and pipeline pressure difference as inputs and the pump input power as output, and is constructed using a multinomial regression or a random forest; the cooling tower fan energy efficiency model takes the circulating water return temperature, ambient wet-bulb temperature, and fan operating frequency as inputs and the cooling tower outlet water temperature and fan input power as outputs, and is constructed using a neural network for online dynamic calibration.

5. The low-carbon energy efficiency optimization method for industrial cooling systems based on edge computing according to claim 1, characterized in that: S3 specifically includes: constructing historical cooling load lag characteristics, rolling statistical characteristics, and environmental and production-related characteristics as model inputs; adopting a two-layer LSTM network structure, performing a prediction every 5 minutes, and outputting the cooling load prediction value for the next 30 minutes to 4 hours; the model evaluation index is mean absolute percentage error <5% and coefficient of determination R²>0.

94.

6. The low-carbon energy efficiency optimization method for industrial cooling systems based on edge computing according to claim 1, characterized in that: S5 sets four threshold levels for individual devices and the entire system: rated load, maximum safe load, warning threshold, and alert threshold. The warning threshold is 0.85 times the rated load, and the alert threshold is 0.95 times the rated load. The type of rise is determined by the duration of the trend: a duration of less than 2 hours is a short-term rise, and a duration of 2 hours or more is a long-term rise. At the same time, the strength of the upward trend is calculated to characterize the average magnitude of the exceedance.

7. The low-carbon energy efficiency optimization method for industrial cooling systems based on edge computing according to claim 1, characterized in that: In S6, the reasons for load increase are divided into short-term and long-term reasons: short-term reasons include instantaneous process cooling load peaks, equipment start-up and shutdown disturbances, sensor data noise, and transient response of control parameter adjustments; long-term reasons include production plan load increases, inherent performance degradation of cooling towers, seasonal increase in ambient temperature and humidity, and changes in downstream process parameters; among them, the inherent performance degradation of cooling towers is quantitatively determined through an active ruminant detection mechanism.

8. The low-carbon energy efficiency optimization method for industrial cooling systems based on edge computing according to claim 1, characterized in that: In S7, the warning levels are from low to high: Blue Level I, Yellow Level II, Orange Level III, and Red Level IV. The basic warning level is first calculated based on the number of over-limit devices, and then multiplied by the severity coefficient of the cause to obtain the final level.

9. The low-carbon energy efficiency optimization method for industrial cooling systems based on edge computing according to claim 1, characterized in that: In S8, the objective function is to minimize the total input power; the decision variables include the start-stop status of each device, the operating frequency of the variable frequency device, and the temperature setting parameters; the constraints include cooling load demand constraints, equipment capacity constraints, process temperature constraints, equipment safe operation constraints, and active rumination avoidance constraints.