Closed cooling tower control method for dynamic dew point temperature tracking based on AI algorithm model

Abstract

The invention discloses a closed cooling tower control method for performing dynamic dew point temperature tracking based on an AI algorithm model, the control method is used for controlling a closed cooling tower, and the control method comprises the following steps: step 1, collecting data in real time; step 2, data preprocessing; step 3, in a decision-making stage of AI algorithm model reasoning, the central controller carries the AI algorithm model of a supervised learning pre-training and reinforcement learning online optimization hybrid architecture, and outputs an optimal cooling mode based on the preprocessed data; 4, cooling mode execution and mode switching are carried out, and the central controller controls all the modules to execute the corresponding cooling modes in the dry cooling mode, the evaporation cooling mode and the collaborative cooling mode according to the model instructions; 5, closed-loop feedback: monitoring and adjusting the running state, and after the cooling mode is executed, continuously monitoring the running state by the system to form closed-loop optimization; and 6, processing a special scene.

Description

Technical Field

[0001] This invention relates to the field of closed-circuit cooling tower technology, and in particular to an energy-saving closed-circuit cooling device based on dynamic adjustment of ambient dew point temperature, which is suitable for industrial cooling, process air conditioning and other scenarios, and can realize the utilization of natural cold sources and intelligent energy-saving operation throughout the year. Background Technology

[0002] While existing dew point temperature-dependent cooling tower technologies (such as negative pressure evaporation and wide-temperature-range precooling) can achieve cooling effects close to or below the dew point temperature, they have several key drawbacks in practical applications, as follows: I. The mode switching lacks intelligence and has insufficient dynamic adaptation capabilities. Current technologies generally lack high-precision real-time environmental dew point sensors, relying instead on temperature / humidity sensors to indirectly estimate the dew point. This leads to significant errors and an inability to accurately obtain current environmental dew point data, making it difficult for the system to detect extreme high / low dew point conditions and trigger targeted adjustment strategies. Closed-loop cooling strategies are mostly based on fixed parameters or a single temperature threshold, lacking dynamic optimization algorithms based on dew point, resulting in a significant decrease in efficiency under extreme high / low dew point environments.

[0003] 2. The closed-loop circuit has low coupling with dew point utilization and poor heat dissipation matching. Missing path branch: The closed working fluid dissipates heat only through a fixed air-cooled heat exchanger (or spray heat exchanger), lacking dew point-based latent heat recovery branch switching logic. In high dew point environments, the working fluid still follows the air-cooled branch, failing to prioritize the absorption of latent heat from humid air through the heat pump plate heat exchanger, resulting in wasted latent heat resources; in low dew point environments, the working fluid is still forced to follow the heat pump co-processing branch, increasing unnecessary energy consumption.

[0004] Fixed heat exchange area allocation: The heat exchange area ratio between the closed working fluid and the humid air, and the heat pump evaporator is preset to a fixed value and cannot be dynamically adjusted according to the dew point. At high dew points, it is necessary to increase the latent heat exchange area, but this cannot be achieved, resulting in insufficient latent heat absorption; at low dew points, it is necessary to increase the sensible heat area, but it is difficult to reduce the ineffective proportion of latent heat exchange.

[0005] Rigid working fluid flow regulation: The closed-loop working fluid pump operates at a fixed speed, lacking a dynamic flow adjustment mechanism linked to dew point. At low dew points, sensible heat dissipates quickly, which could reduce flow rate and lower energy consumption; however, the existing system maintains high flow rate operation, resulting in energy waste.

[0006] Third, waste heat is not effectively recovered, resulting in low energy utilization. Existing closed-loop heat exchange towers focus on sensible heat dissipation, but their technical principles do not consider latent heat recovery as a key aspect of energy utilization. They only focus on the sensible heat carried away by the evaporation of sprayed water, neglecting the latent heat released by the condensation of water vapor in humid air (the latent heat of water vapor phase change is approximately 2260 kJ / kg, several times that of sensible heat). They treat humid air as waste gas for heat dissipation rather than an energy source, lacking a thermodynamic cycle design for latent heat conversion. Furthermore, they fail to incorporate latent heat conversion technologies such as heat pump cycles, making it impossible to convert the latent heat of humid air into usable high-grade heat, resulting in the direct emission of a large amount of latent heat.

[0007] IV. Poor adaptability to high dew point environments, resulting in an imbalance between cooling efficiency and energy consumption. Evaporative cooling has inherent limitations: Traditional closed-loop cooling towers rely on the evaporation of sprayed water and the exchange of sensible heat between the air and the sprayed water, which drives evaporation. At high dew points, the air is nearly saturated (with minimal humidity difference), causing a sharp drop in evaporation. Sensible heat exchange becomes the primary heat dissipation method, but its sensible heat efficiency (20-50 W / (m²・K)) ​​is far lower than that of evaporative cooling (100-300 W / (m²・K)), making it difficult to meet target temperature requirements.

[0008] Latent heat recovery cycle defects: The latent heat of humid air is not incorporated into the core cycle. At high dew points, a large amount of latent heat is not utilized and cannot be converted into auxiliary cooling energy. The cooling gap can only be filled by an additional cooling source.

[0009] Additional cooling source non-cooperative design: Additional cooling source (such as chiller unit) operates independently from the main system without thermodynamic coupling (such as using latent heat to preheat the refrigerant). The load is too large during startup and energy consumption is not optimized.

[0010] V. Poor adaptability of feature processing to model decision-making, resulting in insufficient accuracy and stability of decision-making. Inadequate feature processing: No feature selection mechanism was established for cooling tower operation scenarios. The raw data contained redundant and low-correlation features, and there was a lack of extraction and integration of scenario-based derived features. This resulted in messy data dimensions and a low proportion of effective information in the input model, affecting decision-making efficiency and accuracy. At the same time, the strong correlation between features was not considered, which could easily lead to model overfitting.

[0011] The model architecture has flaws: a single model cannot balance the speed of decision-making with long-term optimality, lacks a collaborative mechanism between supervised learning and reinforcement learning, resulting in low decision-making efficiency in normal scenarios and failure to achieve optimal energy consumption in extreme scenarios; moreover, the model does not have self-learning capabilities and cannot dynamically adjust parameters and strategies according to seasonal changes, environmental migration, etc., resulting in poor adaptability.

[0012] Weak anti-interference capability: No effective protection mechanism has been established for problems such as sensor noise, electromagnetic interference, and data anomalies in industrial environments. The model is easily misled by interference data, leading to decision bias. At the same time, there is a lack of fault tolerance and fallback mechanism. When the sensor fails, decision failure is likely to occur, affecting the stable operation of the equipment. Summary of the Invention

[0013] The technical problem this invention aims to solve is to address the above shortcomings by providing a control method for closed-loop cooling towers based on AI algorithm models for dynamic dew point temperature tracking, thereby achieving the following objectives: Enhance the system’s dynamic adaptability and achieve intelligent mode switching based on accurate dew point data to avoid equipment damage caused by frequent switching. Optimize system coupling and matching, dynamically adjust heat exchange path, area and working fluid flow rate to make full use of energy; Establish a latent heat recovery mechanism to improve energy utilization and reduce waste heat; Enhance adaptability to high dew point environments, ensure cooling efficiency through a collaborative cooling mode, and balance energy consumption; Optimize feature processing and AI algorithm model decision architecture to improve decision accuracy, stability and adaptability, enhance anti-interference and fault tolerance capabilities, and ultimately achieve intelligent collaborative, high-efficiency energy saving and stable and reliable operation of closed cooling towers.

[0014] To solve the above technical problems, the present invention adopts the following technical solution: A control method for a closed cooling tower based on an AI algorithm model for dynamic dew point temperature tracking. The control method is used to control a closed cooling tower, which includes a central controller, a main heat exchanger, an atomization system, and a heat pump system. The main heat exchanger is connected to a user-end load heat source through a closed working fluid loop. The main heat exchanger is also connected to a plate heat exchanger, which is connected to the heat pump system. The heat pump system includes an evaporator, a compressor, a condenser, and a throttling and pressure-reducing component. The compressor is connected to the evaporator and the condenser on both sides. The condenser output pipe is connected to a plate heat exchanger. The throttling and pressure-reducing component is installed on the condenser output pipe. The evaporator is connected to an induced draft fan. The atomization system includes an atomizing pump, which is connected to an atomizing nozzle, which is positioned above the main heat exchanger. An air inlet is provided at the bottom of the main heat exchanger; The central controller is connected to a sensor array, an atomizing pump, an induced draft fan, a compressor, and valves at the air inlet. The cooling modes of the closed cooling tower include dry cooling mode, evaporative cooling mode, and synergistic cooling mode. The control method includes the following steps: Step 1, real-time data acquisition: Collect raw environmental data, raw equipment data, and raw energy efficiency data through a sensor group. The raw environmental data includes the real-time dew point temperature T1, environmental temperature T2, and relative humidity RH. The raw equipment data includes the working medium inlet temperature T_in, working medium outlet temperature T_out, working medium flow rate Q, spray water temperature, and fan speed. The raw energy efficiency data includes the energy consumption per unit cooling capacity and the mode switching frequency; Step 2, data preprocessing: Perform filtering and smoothing, normalization, 3σ principle-based outlier detection, and feature extraction on the data to eliminate interference and extract effective features; Step 3, the decision-making stage of AI algorithm model inference: The central controller is equipped with an AI algorithm model with a hybrid architecture of supervised learning pre-training and reinforcement learning online optimization, and outputs the optimal cooling mode based on the preprocessed data; Step 4, cooling mode execution and mode switching: The central controller controls each module to execute the corresponding cooling mode among the dry cooling mode, evaporative cooling mode, and collaborative cooling mode according to the model instructions. The trigger condition for the dry cooling mode is T1 < T_out; the trigger condition for the evaporative cooling mode is T1 ≈ T_out ± 2°C; the trigger condition for the collaborative cooling mode is T1 > T_out; Step 5, closed-loop feedback: operation state monitoring and adjustment: After the cooling mode is executed, the system continuously monitors the operation state to form a closed-loop optimization; Step 6, special scenario handling: Extreme high dew point where T1 is much higher than T_out: The refrigeration source starts in stages according to the dew point threshold trigger + working medium temperature compensation logic. First, start the latent heat recovery of the heat pump system. If the working medium temperature still does not meet the standard, gradually increase the refrigeration load; Extreme low dew point where T1 is much lower than T_out: Maintain the dry cooling mode.

[0015] Furthermore, the feature extraction in Step 2 includes the following steps: Step 2.1, calculate derivative features: Deduce key correlation features from the raw data; Working medium temperature deviation: The calculation formula is the current working medium outlet temperature - the target cooling temperature; Ratio of dew point to environmental temperature: The calculation formula is the real-time dew point temperature T1 / environmental temperature T2; Temperature difference between the inlet and outlet of the working medium: The calculation formula is the working medium inlet temperature T_in - working medium outlet temperature T_out; Refrigeration source load matching degree: The calculation formula is the current refrigeration source load / rated refrigeration source load; Difference between the spray water temperature and the dew point: The calculation formula is the spray water temperature - real-time dew point temperature T1; Cooling efficiency per unit energy consumption: The calculation formula is the working medium temperature deviation / energy consumption per unit cooling capacity; Step 2.2, Feature Integration: The raw environmental data, raw equipment data, and raw energy efficiency data are merged with the six derived features calculated in Step 2.1 to form an initial feature pool. Step 2.3, Feature Filtering: Form the model input feature set, evaluate the feature importance based on the pre-trained random forest classifier, remove redundant and low-relevance features, and finally retain 12 core input features.

[0016] Furthermore, the specific process of step 2.3 is as follows: Step 2.3.1: Determine the evaluation index, and use the reduction in Gini coefficient natively supported by the random forest classifier as the core evaluation index; Step 2.3.2, AI algorithm model training and importance calculation; The 17 initial features after feature integration and the corresponding optimal cooling mode labels are input into the random forest classifier to complete full feature training. During model training, each decision tree records the changes in the Gini coefficient when each feature participates in the split. Finally, the average of the results of 200 decision trees is used to obtain the Gini importance score of each feature, which is normalized to the [0,1] interval. For the top 15 features in Gini score, the importance of the ranking is further calculated: each feature is randomly shuffled 10 times, and the model classification accuracy is recalculated after each shuffle. The average decrease in accuracy is calculated and used as the secondary validation score. Step 2.3.3, Feature importance ranking and threshold setting; The combined scores of the two indicators are used to calculate the overall importance score of the features, which is calculated as Gini score × 0.7 + ranking importance score × 0.3, and then sorted in descending order. Set screening thresholds: Hard threshold: the overall importance score is ≥0.05, that is, the feature contributes no less than 5% to the model decision. Features below this threshold are judged as low-value redundant features; Soft threshold: exclude strongly correlated feature pairs. Through Pearson correlation coefficient detection, if the correlation coefficient of two features is ≥0.8, retain the one with the higher score to avoid feature redundancy leading to model overfitting. Step 2.3.4: Screening, Verification, and Final Determination; Candidate features are selected based on a threshold, approximately 12-14 items, and a simplified feature set is constructed. Retrain the random forest classifier using the candidate feature set and verify the classification accuracy: If the accuracy is greater than or equal to 95% of the original full-feature model, meaning that the model performance does not significantly decrease after screening, the candidate set is retained; if the accuracy is less than 95%, 1-2 features with the second-highest scores are re-validated until the performance requirements are met; finally, 12 core input features are determined. Step 2.3.5, Feature extraction output results; The final output consists of 12 core feature vectors, covering three dimensions: environmental status, equipment operation, and energy efficiency performance. Specifically, these include: environmental features: real-time dew point temperature, ambient temperature, relative humidity, and the ratio of dew point to ambient temperature; equipment features: temperature difference between inlet and outlet of working fluid, working fluid flow rate, difference between spray water temperature and dew point, fan speed, and matching degree of cooling source load; and energy efficiency features: energy consumption per unit of cooling capacity, mode switching frequency, and cooling efficiency per unit of energy consumption.

[0017] Furthermore, step 3 includes the following steps: Step 3.1: Preliminary assessment of the pre-trained model. The 12 feature vectors are input into the pre-trained random forest classifier. The specific process is as follows: Step 3.1.1, Preliminary preparations: Feature vectors and model state are ready; Input feature vector standardization: The 12 core features have been preprocessed and normalized and aligned in dimension. The dimension alignment is arranged in a fixed order: real-time dew point temperature → ambient relative humidity → ratio of dew point to ambient temperature → temperature difference between working fluid inlet and outlet → working fluid flow rate → difference between spray water temperature and dew point → fan speed → cooling source load matching degree → energy consumption per unit cooling capacity → mode switching frequency → cooling efficiency per unit energy consumption → working fluid temperature deviation, forming a 1×12-dimensional feature vector, which is input into a pre-trained random forest classifier. Step 3.1.2, Classification reasoning of feature vectors by a single decision tree 200 decision trees independently determine the input feature vector, and the reasoning logic of each tree is consistent; Splitting rules: Each node is determined by minimizing the Gini coefficient during pre-training based on the optimal splitting threshold of the feature values. For example, if the normalized value of the working fluid temperature deviation is >0.3, it does not meet the standard. The feature vector is then assigned to the next level child node until the leaf node is reached. Leaf node output category: Each leaf node corresponds to a unique cooling mode among dry cooling / evaporative cooling / co-cooling. The leaf node category that the feature vector finally falls into is the preliminary judgment result of the decision tree. Step 3.1.3: Integrated voting of 200 decision trees; Each decision tree independently outputs one cooling mode decision, forming a set of 200 independent decision results; the number of votes for each type of cooling mode in the set is counted; according to the principle of majority rule, the mode with the most votes is selected as the ensemble decision result of the random forest. Step 3.1.4: Pattern suggestion output and confidence level labeling; Output initial pattern suggestions: The ensemble decision results are used as initial pattern suggestions and passed to the subsequent DQN reinforcement learning model; Calculate the confidence level of the model: Confidence level = (number of highest votes / total number of decision trees) × 100%, which is used as a reference for subsequent optimization of the DQN model. If the confidence level is ≥80%, the DQN model will retain the suggestion first; if the confidence level is <50%, the DQN model will focus on optimization. Step 3.1.5, Optimize the reinforcement learning model; The DQN reinforcement learning model aims to optimize long-term energy consumption by optimizing the initial suggestions. Its state space is the current sensor feature vector, and its action space consists of three cooling modes. The reward function is set as: R=α*(target temperature achievement rate)-β*(unit energy consumption)-γ*(mode switching frequency), where α, β, and γ are adjustable weight coefficients. Finally, it outputs the determined cooling mode command.

[0018] Furthermore, step 3 also includes the following steps: Step 3.2, AI algorithm self-learning mechanism, the specific process is as follows; Step 3.2.1, Weekly threshold dynamic optimization: For the random forest classifier, based on recent running data, adjust the mode switching thresholds for dry-cold / evaporative-cold / co-current-cold to adapt to seasonal or environmental humidity changes; Data filtering and preprocessing: Triggered weekly, extracting runtime data from the most recent 30 days; Filter valid data: Remove abnormal data caused by sensor failure or extreme interference, and retain records with a cooling effect compliance rate of ≥80% and no sudden changes in energy consumption data; The data were grouped into three ranges: low dew point <10℃, medium dew point 10-25℃, and high dew point >25℃. The optimal combination of mode → energy consumption → compliance rate was calculated for each group. For each dew point interval, a grid search method is used to traverse the possible threshold ranges and calculate the comprehensive energy efficiency score under different thresholds. The score = 0.6 compliance rate + 0.4 (1 - relative energy consumption). The random forest classifier was retrained using the filtered 30-day data, and the split threshold of the decision tree was fine-tuned. Step 3.2.2: Daily online fine-tuning of reinforcement learning. Based on daily operational data feedback, the network weights of the DQN model are adjusted to make the mode decision more aligned with the long-term goal of minimizing energy consumption. Experience data storage: Triggered once every 24 hours of operation, the status-action-reward data of the day is added to the experience replay pool, with a fixed capacity of 10,000 records. A first-in-first-out strategy is adopted to remove the oldest data. The newly added data is labeled to supplement extended information such as energy consumption changes, cooling effect and stability in the following hour, enriching the evaluation dimensions of the reward function. DQN model fine-tuning training: 1000 data points are randomly sampled from the experience replay pool as the fine-tuning training set. The training parameters are set as follows: learning rate = 0.001, training epochs = 50 epochs, batch size = 32. A target network and evaluation network separation mechanism is adopted. The evaluation network updates its weights based on new data, and the target network synchronizes the weights of the evaluation network every 100 steps. The reward function is dynamically adjusted: the weight coefficients of α, β, and γ are fine-tuned according to the energy consumption performance of the day. The exploration and utilization balance is achieved using an ε-greedy strategy: the ε value decreases over time, starting at 0.3 and decreasing by 0.05 every 30 days, reaching a minimum of 0.05. Step 3.2.3, Scene migration adaptation: When the device is deployed to a new climate zone, it can quickly adapt to the new scene through transfer learning, reducing the cost of retraining.

[0019] Furthermore, step 3 also includes the following steps: Step 3.3, Supplementing the AI ​​algorithm model to resist interference, constructing a four-layer protection mechanism consisting of a data layer, a model layer, a decision layer, and an anomaly handling layer. The specific process is as follows: Step 3.3.1: Redundant data acquisition and filtering / denoising at the data layer to ensure the data input to the model is authentic and stable. Multi-sensor redundancy backup and fusion: For the core input real-time dew point temperature T1, two independent dew point sensors are deployed to collect data simultaneously; under normal circumstances, the average value of the two sensor data is taken as the input; if the data of a single sensor exceeds the reasonable range or there is no response, it is automatically judged as a fault and the system switches to the valid data of the other sensor; if both sensors fail, the subsequent abnormal handling mechanism is triggered. Layered filtering processing: Basic filtering uses a 5-second sliding window filter on all sensor data. The calculation formula is: Current filter value = (Data from the previous 4 seconds + Current data) / 5, which smooths high-frequency noise; Dynamic data enhancement filtering adds Kalman filtering to data with drastic dynamic changes in working fluid temperature and refrigerant load, and corrects them through prediction-update iteration. Electromagnetic interference adaptive protection is implemented by deploying electromagnetic interference sensors to monitor the intensity of environmental electromagnetic interference in real time. An interference threshold is set, with ≥10V / m considered excessive. If the interference does not exceed the threshold, the original filtering intensity is maintained. If the interference exceeds the threshold, the sliding window is automatically expanded to 10 seconds, and the prediction weight of the Kalman filter is increased to further suppress noise. Step 3.3.2, robust training and anti-interference adaptation of the model layer, improves the model's tolerance to interference, so that the model can still output the correct decision when there is a slight deviation in the input data, and avoids being misled by the interference data; Step 3.3.3, Decision-making smoothing mechanism and switching constraints; Mode switching cooldown setting: Set a fixed cooldown period, and continuous mode switching is not allowed within 5 minutes; During the cooldown period, the processing logic continuously monitors the data. If the triggering conditions for the new mode are still met after the cooldown period ends, the switch will be executed; if the conditions disappear, the current mode will be maintained. The decision results are cross-validated. If the model recommendation is collaborative cooling, it must simultaneously meet three core characteristic conditions: dew point > target temperature, working fluid temperature deviation > 0, and the difference between spray water temperature and dew point < 2℃. If only one characteristic is met, but the other characteristics are contradictory, it is judged as an erroneous recommendation caused by interference, and it will not be output for the time being. The current model will continue to be used. Handling of verification failure: The current data is marked as suspicious data and stored in a separate database for subsequent model iteration analysis of interference patterns. Step 3.3.4, Fallback mechanism and manual intervention in the exception handling layer; Multi-sensor fallback: Set fault judgment criteria. If ≥3 core sensors fail simultaneously or the data is abnormal, i.e., exceeding the mean ± 3 times the standard deviation, the safety mode is automatically triggered. Safety mode logic: switch to evaporative cooling mode, and shut down the refrigeration source staged start logic to maintain a fixed spray volume and fan speed. Alarm mechanism: send alarm information to operation and maintenance personnel through the equipment control system, mark the location of the faulty sensor and abnormal data, and prompt timely maintenance. Decision deviation verification and manual review: Real-time calculation of the deviation between the model output and the actual operating conditions. Deviation value = (average energy consumption within 10 minutes after mode switch / average energy consumption within 10 minutes before switch) - (cooling compliance rate within 10 minutes after mode switch / cooling compliance rate within 10 minutes before switch). A deviation threshold is set. If the deviation value is ≥20%, such as a sudden increase in energy consumption after mode switch but no improvement in cooling effect, it is judged as a decision deviation and the manual review process is triggered.

[0020] Furthermore, the specific process of switching from dry cooling mode to evaporative cooling mode in step 4 is as follows: The switching trigger condition is that the dew point temperature T1 is within the working fluid outlet temperature T_out ±2℃ range for 20 consecutive seconds, and the lag threshold must be maintained for 20 seconds; in extreme environments, the deviation value will be corrected, with a deviation of +1℃ for ultra-low temperature environments <-10℃ and a deviation of ±3℃ for high temperature environments of 35~40℃. The central controller first continuously receives data such as T1 and T_out collected by the sensor group, and after confirming that all switching conditions are met, it outputs a switching command. The air inlet valve is fully open in dry cooling mode and remains fully open during the switching process to provide sufficient air for the subsequent evaporation process; The atomizing pump starts from the off state, pressurizes and draws water from the source and delivers it to the atomizing nozzle. The nozzle then opens and evenly sprays tiny droplets onto the outer surface of the finned tubes of the main heat exchanger. The induced draft fan gradually increases its speed from the low speed of the dry cooling mode, eventually creating a negative pressure environment of 0.8-0.9 atm inside the tower, guiding the outside air to flow through the outside of the finned tube and fully contact the atomized droplets; The plate heat exchanger and heat pump system remain shut down and do not participate in heat exchange during the entire switching process; The high-temperature working fluid output from the user-end load heat source still flows through the main heat exchanger along the original path. The heat exchange mode changes from sensible heat exchange with dry air to latent heat exchange with atomized droplets, and the working fluid temperature gradually approaches the dew point. The sensor array continuously monitors data such as dew point, working fluid temperature and flow rate, and negative pressure inside the tower, and feeds it back to the central controller to dynamically adjust the output power of the atomizing pump and the speed of the induced draft fan to ensure stable operation in evaporative cooling mode.

[0021] Furthermore, the specific process of switching from evaporative cooling mode to dry cooling mode in step 4 is as follows: The switching condition is that T1 < T_out - 2℃ is satisfied for 15 consecutive seconds, and the hysteresis threshold is maintained for 25 seconds; in the ultra-low temperature environment of <-10℃, the trigger threshold of dry cooling mode is relaxed to T1 < T_out + 1℃; After the central controller confirms that the switching conditions are met through sensor data, it sends a command to terminate evaporative cooling and start dry cooling. The atomizing pump stops first, and the atomizing nozzles close accordingly, stopping the spraying of liquid droplets onto the main heat exchanger to avoid wasting water resources. The induced draft fan gradually decreases from medium-high speed to low speed in dry cooling mode, maintaining only the basic airflow to meet the airflow requirements for sensible heat exchange between dry air and working fluid, while reducing energy consumption. The air inlet valve remains fully open, continuously introducing dry outside air; The plate heat exchanger and heat pump system remain shut down and do not participate in the heat exchange process. The heat exchange mode of the working fluid flowing through the main heat exchanger changes from latent heat exchange to sensible heat exchange, and cooling is achieved by the dry air absorbing heat from the working fluid. The sensor array focuses on monitoring the dry air temperature and the temperature difference between the inlet and outlet of the working fluid, and feeds the data back to the central controller to ensure that the working fluid cooling effect meets the standards, while avoiding frequent start-ups and shutdowns of components.

[0022] Furthermore, the specific process of switching from evaporative cooling mode to collaborative cooling mode in step 4 is as follows: The switching condition is that T1 > T_out + 2℃ is met for 10 consecutive seconds, the lag threshold is maintained for 15 seconds, and the high temperature load is responded to quickly; in the high temperature environment of 35~40℃, the collaborative cold trigger threshold is advanced to T1 > T_out + 1℃, and the collaborative cold operation is stable in the ultra-high temperature environment of ≥40℃. After the central controller confirms the switching conditions, it outputs a coordinated cooling mode command, which starts the heat pump system and adjusts the status of relevant components while maintaining the operation of the evaporative cooling core components. The atomizing pump and atomizing nozzle continue to operate, maintaining the spraying of liquid droplets onto the outside of the finned tubes of the main heat exchanger, thus sustaining the atomization and evaporation process. The induced draft fan maintains a medium-high speed, keeps the negative pressure inside the tower at 0.8-0.9 atm, and guides the hot and humid air after heat exchange in the main heat exchanger to flow to the evaporator. When the heat pump system starts from the off state: the throttling and depressurization component depressurizes and cools the high-pressure refrigerant, forming a low-temperature liquid refrigerant that flows into the plate heat exchanger; the compressor starts and compresses the gaseous refrigerant from the evaporator into a high-temperature, high-pressure gaseous state; the condenser couples with the plate heat exchanger and transfers the cooling capacity to the plate heat exchanger; The working fluid output from the user-end load heat source first flows through the main heat exchanger to exchange latent heat with the atomized droplets. After initial cooling, it enters the plate heat exchanger to exchange heat with the low-temperature refrigerant, thus achieving two-stage cooling. The air inlet valve remains fully open to provide sufficient air for evaporation and latent heat recovery of the heat pump. The sensor array monitors dew point, working fluid temperature and flow rate, and refrigerant load data in real time. The central controller dynamically adjusts the atomization amount, fan speed, and compressor load to ensure that the working fluid reaches the target temperature while avoiding full-load startup of the refrigerant.

[0023] Furthermore, the specific process of switching from collaborative cooling mode to evaporative cooling mode in step 4 is as follows: The switching condition is that T1≤T_out+2℃ is satisfied for 30 consecutive seconds, and the lag threshold is maintained for 30 seconds to ensure a smooth switch from high energy consumption mode to low energy consumption mode. Once the central controller confirms that the switching conditions have been met through sensor data, it sends a command to shut down the heat pump system and maintain the operation of the evaporative cooling core. The heat pump system gradually shuts down: the compressor stops working and no longer compresses the refrigerant; the throttling and depressurizing components stop depressurizing and cooling the refrigerant; the transfer of cold energy between the condenser and the plate heat exchanger terminates, and the plate heat exchanger stops participating in heat exchange with the working fluid; The atomizing pump and atomizing nozzle continue to operate, spraying droplets onto the outside of the finned tubes of the main heat exchanger to maintain the latent heat exchange of atomization evaporation. The induced draft fan maintains a medium-high speed to maintain negative pressure inside the tower, directly exhausting the hot and humid air that has passed through the main heat exchanger outside the tower, instead of directing it to the evaporator. The air inlet valve remains fully open to provide sufficient air for the evaporation process; The heat exchange path of the working fluid changes from two-stage cooling of main heat exchanger + plate heat exchanger to latent heat exchange only through the main heat exchanger and atomized droplets. The sensor array continuously monitors the working fluid outlet temperature, tower negative pressure, and energy consumption data. The central controller adjusts the output power of the atomizing pump and the speed of the induced draft fan to ensure that the cooling effect meets the standard, while reducing the operational fluctuations caused by the heat pump shutdown.

[0024] The present invention adopts the above technical solution and has the following technical effects compared with the prior art: 1. Significantly improved dynamic adaptability: By collecting data in real time through a high-precision dew point sensor and combining it with the dynamic optimization and clear switching rules of the AI ​​algorithm model, it accurately senses extreme high / low dew point states and realizes intelligent switching between dry cooling, evaporative cooling and synergistic cooling modes. This avoids frequent switching caused by short-term environmental fluctuations, reduces equipment wear and tear, and improves the annual energy saving rate by 30%-50%. It is adaptable to the entire ambient temperature range from -10℃ to 40℃.

[0025] 2. System Coupling and Energy Utilization Optimization: Establish a dew point-based latent heat recovery branch switching logic, a dynamic heat exchange area matching mechanism, and a working fluid flow linkage adjustment scheme to solve the problems of single path branches, fixed heat exchange area, and rigid flow regulation, so as to realize on-demand energy allocation and efficient utilization and reduce unnecessary energy consumption.

[0026] 3. Significantly improved waste heat recovery efficiency: The introduction of a heat pump thermodynamic cycle efficiently recovers the latent heat released by the condensation of water vapor in humid air (the latent heat of phase change is about 2260kJ / kg), and converts it into usable high-grade cooling capacity to assist in cooling. This reduces energy consumption during auxiliary cooling by more than 20%, significantly reduces waste heat, and improves energy utilization.

[0027] 4. Enhanced adaptability to high dew point environments: By linking the evaporation system and the heat pump recovery system in a coordinated cooling mode, a two-stage cooling mechanism is constructed, which can ensure that the cooling effect meets the standard even if the ambient dew point is higher than the target temperature; at the same time, the refrigeration source stage start-up logic is optimized to avoid high energy consumption caused by full load operation and solve the problem of imbalance between cooling efficiency and energy consumption in high dew point environments.

[0028] 5. Ensuring Decision Accuracy and Stability: Optimize model input quality through scenario-based feature extraction, redundant feature removal, and processing of highly correlated features; employ a hybrid architecture of random forest pre-training + DQN reinforcement learning optimization, coupled with a self-learning mechanism that features weekly threshold updates, daily fine-tuning, and scene migration adaptation, to improve decision accuracy and long-term optimality; construct a four-layer anti-interference mechanism and fault tolerance scheme (data layer - model layer - decision layer - anomaly handling layer) to effectively resist the effects of sensor noise, electromagnetic interference, etc. Automatically switch to backup when a single sensor fails, and trigger a safety mode when multiple sensors fail, ensuring stable equipment operation and reducing maintenance risks. Attached Figure Description

[0029] To more clearly illustrate the specific embodiments of the present invention or the technical solutions in the prior art, the accompanying drawings used in the description of the specific embodiments or the prior art will be briefly introduced below. In all the drawings, similar elements or parts are generally identified by similar reference numerals. In the drawings, the elements or parts are not necessarily drawn to scale.

[0030] Figure 1 This is a schematic diagram of the structure of the present invention; Figure 2 This is a modular structure diagram of the closed-loop cooling tower based on the present invention; Figure 3 This is a flowchart illustrating the control method logic of the present invention. Figure 4 This is a schematic diagram of the multimodal cooling mode switching of the present invention. Detailed Implementation

[0031] Examples, such as Figure 1 As shown, a closed-loop cooling tower that uses an AI algorithm model for dynamic dew point temperature tracking includes: The closed-loop working fluid circulation link includes a user-end load heat source 3, which is connected to a main heat exchanger 2 via a closed-loop working fluid loop. The main heat exchanger 2 is connected to a plate heat exchanger 4. The circulation process of the working fluid in the closed-loop working fluid circulation link is as follows: user-end load heat source 3 → closed-loop working fluid loop → main heat exchanger 2 → plate heat exchanger 4 → cooled working fluid flows back to user-end load heat source 3, forming a closed loop.

[0032] A closed-loop working fluid (such as an aqueous solution of ethylene glycol) flows through the user-end load heat source, absorbing the heat generated by the equipment operation. The working fluid temperature rises, and the high-temperature working fluid enters the finned tubes of the main heat exchanger. It exchanges heat with dry air or atomized water + air outside the tubes, and the working fluid temperature is initially reduced, achieving initial heat dissipation. The initially cooled working fluid enters the plate heat exchanger and exchanges heat with the low-temperature refrigerant delivered by the heat pump system. The working fluid temperature is further reduced to the target value, achieving secondary heat dissipation. The cooled low-temperature working fluid flows back to the user-end load heat source, and the above process is repeated.

[0033] The heat pump thermodynamic cycle link includes an evaporator 6, which is connected to a compressor 8. The compressor 8 is connected to a condenser 5, and a throttling and pressure-reducing component 9 is installed on the output pipe of the condenser 5. The thermodynamic cycle process in the heat pump thermodynamic cycle link is: evaporator 6 → compressor 8 → condenser 5 → throttling and pressure-reducing component 9 → evaporator 6, forming a closed loop of latent heat recovery and cooling capacity conversion.

[0034] The throttling and depressurization component reduces the pressure and temperature of the high-pressure refrigerant, forming a low-temperature liquid refrigerant. This liquid refrigerant flows into a plate heat exchanger to exchange heat with the closed working fluid. After absorbing heat, the refrigerant evaporates into a gaseous state. The gaseous refrigerant enters the compressor and is compressed into a high-temperature, high-pressure gaseous refrigerant, increasing its energy level. The high-temperature, high-pressure refrigerant then enters the evaporator and comes into contact with the hot, humid air discharged from the tower. The refrigerant releases heat and condenses into a liquid state. At the same time, water vapor in the hot, humid air condenses upon encountering the condenser. The liquid refrigerant then re-enters the throttling and depressurization component, repeating the phase change cycle described above.

[0035] The air circulation link includes a main heat exchanger 2, which is connected to an induced draft fan 7. The main heat exchanger 2 has an air inlet 10 for the cooling tower at its lower end. The air circulation process in the air circulation link is as follows: air inlet 10 → outside of the finned tube of the main heat exchanger 2 → hot and humid air is collected and flows to the evaporator 6 → the air after the latent heat is recovered by the evaporator 6 is discharged outside the tower, and the air flow is powered by the induced draft fan 7.

[0036] Outside air enters the tower through the air inlet. The negative pressure environment maintained by the induced draft fan guides the airflow. The air flows over the outside of the finned tubes of the main heat exchanger and exchanges heat with the working fluid inside the tubes. Dry cooling mode: Dry air directly absorbs the sensible heat of the working fluid, and the air temperature rises.

[0037] Evaporative cooling mode / cooperative cooling mode: Atomized water evaporates outside the pipe, absorbing the latent heat of the working fluid, and the air carries water vapor to form humid and hot air.

[0038] Latent heat recovery stage: After heat exchange, the humid and hot air flows to the heat pump evaporator. The water vapor in the air condenses when it encounters the cold (the refrigerant condenses and releases heat), and the latent heat is absorbed by the refrigerant.

[0039] Exhaust stage: The dry, cold air that has completed latent heat recovery is exhausted outside the tower by the induced draft fan.

[0040] The control signal link includes a central controller, which is connected to a sensor group, an atomizing pump 11, an induced draft fan 7, a compressor 8, and a valve at the air inlet 10. The central controller acquires real-time data, including dew point, working fluid temperature, and flow rate, through the sensor group and sends control commands to the atomizing pump, induced draft fan, compressor, air inlet valve, and other actuators to achieve mode switching and parameter adjustment.

[0041] The system collects ambient dew point temperature, ambient temperature, and relative humidity in real time using a dew point meter, a thermometer and hygrometer, and a wind speed sensor.

[0042] Temperature and flow sensors are used to monitor the inlet temperature, outlet temperature, and flow rate of the closed working fluid.

[0043] The system collects data such as fan speed, cooling source load, energy consumption per unit cooling capacity (kWh / ℃), and mode switching frequency through speed sensors, current sensors, and energy consumption metering sensors.

[0044] The condenser 5 is coupled to the plate heat exchanger 4 to achieve the transfer of cooling capacity.

[0045] The atomizing pump 11 is connected to the atomizing nozzle 12 via a pipe, providing atomizing water to the atomizing nozzle 12.

[0046] The evaporator 6, compressor 8, condenser 5, and throttling and pressure-reducing component 9 constitute a heat pump system.

[0047] like Figures 2 to 4 As shown, the control method for a closed-loop cooling tower based on an AI algorithm model for dynamic dew point temperature tracking includes the following steps: Step 1: Real-time data collection; Raw environmental data: Dew point temperature T1, ambient temperature T2, and relative humidity RH are collected in real time using a dew point meter, a thermometer and hygrometer, and a wind speed sensor, with a sampling frequency of 1Hz; Raw data from equipment: Monitoring working fluid inlet temperature T_in, working fluid outlet temperature T_out, working fluid flow rate Q, spray water temperature, and fan speed through temperature sensors, flow sensors, and speed sensors; Raw energy efficiency data: Energy consumption per unit cooling capacity (kWh / ℃) and mode switching frequency are collected through current sensors and energy consumption metering sensors.

[0048] Step 2, Data Preprocessing; The central controller performs multi-layer processing on the raw data to eliminate interference and extract effective features: Filtering and smoothing: A 5-second sliding window filter is used to smooth sensor noise, and an additional Kalman filter is applied to dynamic data such as working fluid temperature to predict the value at the next moment and correct abnormal fluctuations. Normalization: Map data such as dew point temperature and ambient temperature to the [0,1] interval (e.g., 0℃→0, 40℃→1) to facilitate model calculation; Anomaly Handling: Outliers (such as erroneous data where dew point is greater than ambient temperature) are detected using the 3σ principle. When the value exceeds the mean ± 3 times the standard deviation, the mean of the previous 5 seconds is used instead. At the same time, through the redundancy design of dual dew point sensors, the average of the two sensors is used as the input, and the system automatically switches to the other sensor when one sensor fails. Feature extraction: Calculate derived features, including working fluid temperature deviation (current temperature - target temperature), dew point to ambient temperature ratio, etc., ultimately forming 12 core input features. The specific process is as follows: Step 2.1: Calculate derived features to deduce key association features from the original data; Based on the model's requirement to identify the linkage between dew point, working fluid state, and energy consumption, six core derived features are derived from the raw data to supplement the information dimensions of the directly collected features: Working fluid temperature deviation: The calculation formula is: current working fluid outlet temperature - target cooling temperature. It reflects the difference between the working fluid cooling effect and the target and is the core indicator for judging whether the mode needs to be adjusted. Dew point to ambient temperature ratio: The calculation formula is real-time dew point temperature T1 / ambient temperature T2. It quantifies the degree of ambient humidity saturation and helps to judge the potential for latent heat recovery. The closer the ratio is to 1, the greater the potential for latent heat recovery. The temperature difference between the inlet and outlet of the working fluid is calculated as: working fluid inlet temperature T_in - working fluid outlet temperature T_out. It reflects the heat dissipation efficiency of the working fluid in the heat exchanger and provides a basis for flow rate adjustment. Cooling source load matching degree: The calculation formula is current cooling source load / rated cooling source load, which quantifies the operating status of the cooling source and avoids waste due to overload or underload; The difference between spray water temperature and dew point: The calculation formula is spray water temperature - real-time dew point temperature T1. It is used to judge the feasibility of atomization evaporation. The larger the difference, the higher the evaporation efficiency. Unit energy consumption cooling efficiency: The calculation formula is working fluid temperature deviation / unit cooling energy consumption. It comprehensively evaluates the cost-effectiveness of current energy consumption and cooling effect, and provides an energy efficiency basis for mode optimization.

[0049] Step 2.2, Feature Integration; The three environmental raw data items, six equipment raw data items, and two energy efficiency raw data items are merged with the six derived features calculated in step 2.1 to form an initial feature pool.

[0050] Step 2.3, Feature Filtering: Form the model input feature set. Based on the feature importance evaluation of the pre-trained random forest classifier, redundant and low-relevance features are eliminated, and finally 12 core input features are retained to ensure the efficiency and accuracy of model inference.

[0051] Step 2.3.1, determine the evaluation indicators; We use the reduction in Gini coefficient, which is natively supported by the random forest classifier, as the core evaluation metric, supplemented by permutation importance cross-validation to avoid bias from a single metric.

[0052] The reduction in Gini coefficient measures the contribution of a feature to reducing data impurity during decision tree splitting. The more important the feature, the more significant the decrease in Gini coefficient during splitting, and the higher the score. Ranking Importance: By randomly shuffling the values ​​of a feature, observe the decrease in the model's classification accuracy. The greater the decrease in accuracy, the greater the influence of the feature on the decision and the higher its importance.

[0053] Step 2.3.2, AI algorithm model training and importance calculation; The 17 initial features after feature integration and the corresponding optimal cooling mode labels are input into the random forest classifier to complete full feature training. During model training, each decision tree records the changes in the Gini coefficient when each feature participates in the split. Finally, the average of the results of 200 decision trees is used to obtain the Gini importance score of each feature, which is normalized to the [0,1] interval. For the top 15 features in Gini score, the importance of the ranking is further calculated: each feature is randomly shuffled 10 times, and the model classification accuracy is recalculated after each shuffle. The average decrease in accuracy is calculated and used as the secondary validation score.

[0054] Step 2.3.3, Feature importance ranking and threshold setting; The combined scores of the two indicators are used to calculate the overall importance score of the features, which is calculated as Gini score × 0.7 + ranking importance score × 0.3, and then sorted in descending order. Set the filter threshold: Hard threshold: The overall importance score is ≥0.05, that is, the contribution of the feature to the model decision is not less than 5%. Features below this threshold are judged as low-value redundant features. Soft threshold: exclude strongly correlated feature pairs. Detect the correlation coefficient using Pearson correlation coefficient. If the correlation coefficient between two features is ≥0.8, retain the one with the higher score to avoid feature redundancy that could lead to model overfitting.

[0055] Step 2.3.4: Screening, Verification, and Final Determination; Candidate features are selected based on a threshold, approximately 12-14 items, and a simplified feature set is constructed. Retrain the random forest classifier using the candidate feature set and verify the classification accuracy: If the accuracy is greater than or equal to 95% of the original full-feature model, that is, the model performance does not decrease significantly after screening, then the candidate set is retained; If the accuracy is below 95%, re-validate 1-2 features with the second-highest scores until the performance requirements are met. Ultimately, 12 core input features were identified to ensure a balance between model performance and inference efficiency.

[0056] Step 2.3.5, Feature extraction output results; The final output consists of 12 core feature vectors, covering three dimensions: environmental status, equipment operation, and energy efficiency performance. Specifically, these include: Environmental parameters: real-time dew point temperature, ambient temperature, relative humidity, and the ratio of dew point to ambient temperature; Equipment-related parameters: working fluid inlet and outlet temperature difference, working fluid flow rate, spray water temperature and dew point difference, fan speed, and cooling source load matching degree; Energy efficiency categories: energy consumption per unit cooling capacity, mode switching frequency, and cooling efficiency per unit energy consumption.

[0057] The feature selection in this invention focuses on balancing model performance and inference efficiency, and is achieved through a four-step closed loop: indicator setting, model training, ranking and selection, and verification and confirmation. This approach offers the following significant advantages: 1. The evaluation indicators are more comprehensive, resulting in higher screening accuracy; We employ a dual-indicator weighted evaluation of Gini coefficient reduction and permutation importance. This approach considers both the direct contribution of features to model decision-making (Gini coefficient) and the verification of feature irreplaceability (permutation importance), avoiding the bias of a single indicator. For example, a feature may not be highly relevant, but if its permutation importance score is high (significantly reducing accuracy after shuffling), it will still be retained to ensure that core features are not overlooked.

[0058] 2. The screening logic is more aligned with industrial scenarios, and the screening process is deeply integrated with the needs of industrial cooling scenarios, making it more adaptable; First, supplement the scenario-based information by calculating derived features (such as the ratio of dew point to ambient temperature and unit energy consumption cooling efficiency), and then filter them to ensure that the features are strongly correlated with the core linkage relationship between dew point, working fluid state and energy consumption. The soft threshold design (removing redundant features with a correlation coefficient ≥ 0.8) avoids feature redundancy that could lead to model overfitting, while retaining key dimensions of the scenario (such as environmental, equipment, and energy efficiency features, each with their own coverage) to adapt to the complexity of the dynamic operation of the cooling tower.

[0059] 3. Balancing model performance and inference efficiency, a post-screening validation mechanism is used to achieve a balance, resulting in greater practicality; Hard thresholds ensure that the feature contribution is not less than 5%, eliminate low-value features, reduce model computation, and meet the real-time control requirements of closed cooling towers to complete parameter adjustment within 5-10 seconds. The accuracy of the filtered model should be ≥ 95% of the original full-feature model to ensure no significant performance degradation and avoid errors in cooling mode decision-making due to feature deletion. For example, if a high dew point environment is misjudged, the collaborative cooling mode should be switched.

[0060] 4. The screening process is highly interpretable, with clear quantitative standards and logic for each step, making it easy to implement in industry; The feature importance score can be directly quantified and normalized to the [0,1] interval, so that operation and maintenance personnel can clearly know the decision weight of core features such as real-time dew point temperature and working fluid temperature deviation. The screening threshold can be adjusted according to the actual operating scenario to flexibly adapt to different climate zones. For example, in southern regions with high dew points, the threshold can be fine-tuned to retain more latent heat-related characteristics, reducing the difficulty of industrial implementation.

[0061] 5. Deeply collaborates with subsequent AI algorithm models, optimizing the entire process; Feature selection is based on a pre-trained random forest classifier. The selection process fully incorporates the splitting rules of decision trees, such as prioritizing real-time dew point temperature splitting at the root node, to ensure that the core features selected are highly matched with the model's inference logic. When optimizing the DQN reinforcement learning model later, it can also iterate quickly based on the simplified features, improving the efficiency of the entire process control.

[0062] Step 3, the decision-making stage of AI algorithm model inference: The central controller is equipped with an AI algorithm model with a hybrid architecture of supervised learning pre-training + reinforcement learning online optimization. Based on the pre-processed data, it outputs the optimal cooling mode. The specific process is as follows: Step 3.1: Preliminary assessment of the pre-trained model. The 12 feature vectors are input into a pre-trained random forest classifier. This model is trained using 500,000 historical data points, simulated data, and experimental data, achieving an accuracy of ≥92%. It initially outputs suggestions for dry cooling / evaporative cooling / co-cooling modes. The specific process is as follows: Step 3.1.1, Preliminary preparations: Feature vectors and model state are ready; Input feature vector standardization: The 12 core features have been preprocessed, normalized, mapped, and aligned in a fixed order: real-time dew point temperature → ambient relative humidity → ratio of dew point to ambient temperature → temperature difference between inlet and outlet of working fluid → working fluid flow rate → difference between spray water temperature and dew point → fan speed → cooling source load matching degree → energy consumption per unit cooling capacity → mode switching frequency → cooling efficiency per unit energy consumption → working fluid temperature deviation. This forms a 1×12-dimensional feature vector, which is then input into a pre-trained random forest classifier.

[0063] The feature vectors have no missing or outlier values, ensuring the reliability of the model input.

[0064] Step 3.1.2, Classification reasoning of feature vectors by a single decision tree Two hundred decision trees independently determine the input feature vector, and the reasoning logic of each tree is consistent. Taking a single tree as an example: Root node splitting: The root node splitting feature is the real-time dew point temperature T1. Based on the normalized value of the real-time dew point temperature T1 in the feature vector, the node enters the corresponding left / right child node. For example, if the normalized value of T1 is >0.6, the node enters the high dew point branch; if it is ≤0.6, the node enters the medium / low dew point branch.

[0065] Intermediate node iterative splitting: Subsequent nodes split sequentially according to feature importance priority. The core splitting features and order are as follows: Layer 2: Working fluid temperature deviation → to determine whether the cooling effect meets the standard; Layer 3: Dew point to ambient temperature ratio → quantifies humidity saturation level; Layer 4: The difference between the spray water temperature and the dew point → Determine the feasibility of evaporative cooling; Subsequent layers: characteristics such as cooling source load matching degree and unit energy consumption cooling efficiency are used to gradually narrow down the classification range.

[0066] Splitting rules: Each node is determined by minimizing the Gini coefficient during pre-training based on the optimal splitting threshold of the feature values. For example, if the normalized value of the working fluid temperature deviation is >0.3, it is not up to standard. The feature vector is then assigned to the next level child node until the leaf node is reached.

[0067] Leaf node output category: Each leaf node corresponds to a unique cooling mode among dry cooling / evaporative cooling / co-cooling. The leaf node category into which the feature vector finally falls is the preliminary judgment result of the decision tree.

[0068] Step 3.1.3: Integrated voting of 200 decision trees; Each decision tree independently outputs one cooling mode determination result, forming a set of 200 independent determination results; The number of votes for each type of cooling mode in the statistics set is as follows: dry cooling mode has 60 votes, evaporative cooling mode has 110 votes, and synergistic cooling mode has 30 votes. Following the principle of majority rule, the pattern with the most votes is selected as the ensemble result for the random forest.

[0069] Step 3.1.4: Pattern suggestion output and confidence level labeling; Output initial pattern suggestions: The ensemble decision results are used as initial pattern suggestions and passed to the subsequent DQN reinforcement learning model.

[0070] Calculate the confidence level of the model: Confidence level = (number of highest votes / total number of decision trees) × 100%. This is used as a reference for subsequent optimization of the DQN model. If the confidence level is ≥80%, the DQN model will retain this suggestion first; if the confidence level is <50%, the DQN model will focus on optimization.

[0071] Step 3.1.5, Optimize the reinforcement learning model; The DQN reinforcement learning model aims to optimize long-term energy consumption by optimizing the initial suggestions. Its state space is the current sensor feature vector, and its action space consists of three cooling modes. The reward function is set as: R=α*(target temperature achievement rate)-β*(unit energy consumption)-γ*(mode switching frequency), where α, β, and γ are adjustable weight coefficients. Finally, it outputs the determined cooling mode command.

[0072] Step 3.2, AI algorithm self-learning mechanism, the specific process is as follows; Step 3.2.1, Weekly threshold dynamic optimization: For the random forest classifier, optimize the mode switching threshold. Based on recent running data, adjust the mode switching thresholds for dry cold / evaporative cold / co-current cold to adapt to seasonal or environmental humidity changes.

[0073] Data filtering and preprocessing: Triggered weekly, extracting runtime data from the most recent 30 days; Filter valid data: Remove abnormal data caused by sensor failure or extreme interference, and retain records with a cooling effect compliance rate of ≥80% and no sudden changes in energy consumption data; The data were grouped into three ranges: low dew point <10℃, medium dew point 10-25℃, and high dew point >25℃. The optimal combination of mode → energy consumption → compliance rate was calculated for each group.

[0074] Threshold recalculation: For each dew point interval, a grid search method is used to traverse the possible threshold ranges and calculate the comprehensive energy efficiency score under different thresholds. The score = 0.6 compliance rate + 0.4 (1 - relative energy consumption). The threshold with the highest overall energy efficiency score is selected as the new switching threshold. For example, if the original high dew point threshold is 25°C, and the score is higher at 26°C, the threshold will be automatically updated to 26°C.

[0075] Model parameter update: The random forest classifier was retrained using the filtered 30-day data, and the split threshold of the decision tree was fine-tuned. Verify the update effect: Use the updated random forest classifier to backtrack and infer the data from the past 7 days. If the overall energy efficiency score improves by ≥5%, retain the new threshold and model parameters; if the improvement is less than 5%, roll back to the original threshold and only record the current data for reference in the next iteration.

[0076] Step 3.2.2: Daily online fine-tuning of reinforcement learning. For the DQN model, optimize the long-term energy consumption optimization strategy. Through daily operation data feedback, adjust the network weights of the DQN model to make the mode decision more in line with the long-term energy consumption minimum target and balance the cooling effect and energy saving requirements.

[0077] Experience data storage: Triggered once every 24 hours, the status-action-reward data of the day is added to the experience replay pool, with a fixed capacity of 10,000 records. A first-in-first-out strategy is adopted to remove the oldest data. Add annotations to the newly added data: supplement extended information such as energy consumption changes, cooling effect, and stability over the next hour, enriching the evaluation dimensions of the reward function.

[0078] DQN model fine-tuning training: 1000 data points were randomly sampled from the experience replay pool as the fine-tuning training set, and the training parameters were set as follows: learning rate = 0.001, training epochs = 50 epochs, batch size = 32; A target network and evaluation network separation mechanism is adopted: the evaluation network updates its weights based on new data, and the target network synchronizes the weights of the evaluation network every 100 steps to avoid training oscillations; The reward function is dynamically adjusted: the weight coefficients of α, β, and γ are fine-tuned according to the energy consumption performance of the day. For example, if the energy consumption is high on a given day, the weight of β is increased; if the mode switching is too frequent, the weight of γ is increased.

[0079] Balance between exploration and utilization: An ε-greedy strategy is adopted: the ε value decreases with running time, initially 0.3, decreasing by 0.05 every 30 days, and finally reaching a minimum of 0.05; When the ε value is high in the early stage of operation: focus on exploring new strategies, randomly select the mode with a probability of 10%-30% to avoid the model getting trapped in local optima; When the ε value is low after stable operation: mainly utilize the known optimal strategy, and only explore new strategies with a 5% probability to ensure operational stability.

[0080] Step 3.2.3, Scene migration adaptation: When the device is deployed to a new climate zone (such as moving from a low dew point area in the north to a high dew point area in the south), it can quickly adapt to the new scene through transfer learning, reducing the cost of retraining.

[0081] New scenario data collection: After the equipment is deployed to a new scenario, the transfer learning process is automatically triggered, collecting the first 1,000 operational data points, covering the core characteristics of the new environment such as dew point, temperature, and humidity. Labeling new data: The optimal pattern for the first 50 data points is confirmed manually to ensure the accuracy of data labels.

[0082] Model transfer fine-tuning: Based on the original pre-trained random forest classifier + DQN model, the parameters of the bottom feature extraction layer are frozen, and only the top decision layer, such as the last 3 split nodes of the random forest and the output layer of DQN, are unfrozen. Fine-tuning was performed using 1000 new scene data points, with 20 training rounds and a learning rate of 0.0005, allowing the model to quickly learn the feature-pattern mapping relationship of the new scene.

[0083] Adaptation effect verification: After fine-tuning, the model is validated using 200 new scenario data points. If the model decision accuracy is ≥90% and the energy consumption is ≥8% lower than the original model, the model is considered to have successfully adapted. If the accuracy is less than 90%, collect 500 more data points to supplement the training until the validation requirements are met, ensuring that the model can still run efficiently in new scenarios.

[0084] Step 3.3, Anti-interference Supplementation of AI Algorithm Model: Constructing a four-layer protection mechanism consisting of a data layer, a model layer, a decision layer, and an anomaly handling layer. This mechanism addresses issues such as sensor data noise, environmental interference, and equipment failure, ensuring the stability and accuracy of model decisions through multi-dimensional measures. The specific process is as follows: Step 3.3.1: Redundant acquisition and filtering noise reduction at the data layer improves data reliability from the source, eliminates noise and anomalies in the original sensor data, and ensures that the data input to the model is real and stable.

[0085] Multi-sensor redundancy backup and fusion, the specific process is as follows: Dual sensor deployment for key data: For the core input real-time dew point temperature T1, two independent dew point sensors are deployed to collect data simultaneously; Data fusion rules: Under normal circumstances, the average value of the data from the two sensors is taken as the input; if the data from a single sensor exceeds the reasonable range (e.g., dew point > ambient temperature + 5℃) or there is no response, it is automatically judged as a fault and the valid data from the other sensor is switched; if both sensors are faulty, the subsequent abnormal handling mechanism is triggered.

[0086] The hierarchical filtering process is as follows: Basic filtering: A 5-second sliding window filter is applied to all sensor data (such as ambient temperature and working fluid flow rate). The calculation formula is: Current filter value = (data from the previous 4 seconds + current data) / 5, which smooths high-frequency noise. Dynamic data enhancement filtering: For data with drastic dynamic changes, such as working fluid temperature and refrigerant load, an additional Kalman filter is applied, and corrections are made through prediction-update iteration. The prediction phase uses data from the previous moment and equipment operating characteristics to predict the theoretical value at the current moment. During the update phase, the Kalman gain is calculated based on the actual values ​​collected by the current sensor, the predicted values ​​are corrected, and the final filtering result is obtained, reducing the impact of instantaneous fluctuations.

[0087] Electromagnetic interference adaptive protection, the specific process is as follows: Deploy electromagnetic interference sensors to monitor the intensity of environmental electromagnetic interference in real time; Set an interference threshold; ≥10V / m is considered excessive. If the interference is within the limit, maintain the original filtering strength. If the interference exceeds the limit, automatically expand the sliding window to 10 seconds and increase the prediction weight of the Kalman filter to further suppress noise.

[0088] Step 3.3.2, robust training and anti-interference adaptation of the model layer, improves the model's tolerance to interference, so that the model can still output the correct decision when there is a slight deviation in the input data, and avoids being misled by the interference data.

[0089] Robust training with noisy data follows the specific process: During the model pre-training phase, 10% artificial noise data is added to the training dataset to simulate sensor interference scenarios. Environmental characteristics: values ​​that randomly increase or decrease by ±5%; Equipment-related characteristics: values ​​that randomly increase or decrease by ±8%; Energy efficiency characteristics: values ​​that randomly increase or decrease by ±10%; Retrain the random forest and DQN models using a noisy dataset to allow the models to learn the mapping relationship between normal data → noisy data → correct patterns, thereby reducing their sensitivity to small fluctuations in the data.

[0090] The feature importance-weighted anti-interference method is as follows: Based on feature importance scores, higher weights are assigned to core features and lower weights to secondary features; During inference, if secondary feature data shows abnormal fluctuations, such as a sudden increase in the frequency of pattern switching, the model automatically reduces its influence weight on the decision to avoid secondary interference data dominating the pattern determination.

[0091] Step 3.3.3: Decision-making smoothing mechanism and switching restrictions to avoid erroneous decisions caused by interference, prevent frequent mode switching triggered by brief interference, protect the equipment and maintain stable cooling effect.

[0092] The mode switching cooldown period setting process is as follows: Set a fixed cooling period: continuous mode switching is not allowed within 5 minutes. For example, after dry cooling → evaporative cooling, even if the sensor data meets the conditions for evaporative cooling → synergistic cooling within 5 minutes, the switching will not be performed. Cooldown period processing logic: Continuously monitor data. If the conditions for triggering the new mode are still met after the cooldown period ends, then switch to the new mode; otherwise, maintain the current mode.

[0093] The decision results are cross-validated, and the specific process is as follows: Preliminary pattern suggestions for random forest output include adding feature consistency checks: Verification logic: If the mode suggestion is collaborative cooling, it must simultaneously meet three core characteristic conditions: dew point > target temperature, working fluid temperature deviation > 0, and spray water temperature difference from dew point < 2℃; if only one characteristic is met, but the other characteristics are contradictory, it is judged as an incorrect suggestion caused by interference, and will not be output for the time being, and the current mode will continue to be used. Verification failure handling: Mark the current data as suspicious data and store it in a separate database for subsequent model iteration analysis of interference patterns.

[0094] Step 3.3.4, the fallback mechanism of the exception handling layer and manual intervention are used to deal with severe interference or failure. When the interference exceeds the model's self-processing capacity, the safety plan is activated to avoid equipment shutdown or cooling failure.

[0095] The multi-sensor failure fallback process is as follows: Set fault judgment criteria: If ≥3 core sensors fail simultaneously or the data is abnormal, i.e., exceeding the mean ± 3 times the standard deviation, the safety mode will be automatically triggered. Safety mode logic: Switch to evaporative cooling mode, and at the same time turn off the staged start logic of the cooling source, and maintain a fixed spray volume and fan speed; Alarm mechanism: The equipment control system sends alarm information to maintenance personnel, indicating the location of faulty sensors and abnormal data, prompting timely repair.

[0096] The specific process for decision deviation verification and manual review is as follows: Deviation between real-time calculation model output and actual operating conditions: Deviation value = (Average energy consumption within 10 minutes after mode switch / Average energy consumption within 10 minutes before switch) - (Cooling compliance rate within 10 minutes after mode switch / Cooling compliance rate within 10 minutes before switch). Set a deviation threshold: If the deviation value is ≥20%, such as a sudden increase in energy consumption after mode switching but no improvement in cooling effect, it is judged as a decision deviation and a manual review process is triggered. Review and processing: Suspend the model's self-decision-making, maintain the current mode, and wait for the operation and maintenance personnel to confirm whether to adjust the mode or investigate the source of interference; after the review is approved, the model resumes self-decision-making, and the review result is added to the training data as a label to optimize the subsequent anti-interference ability.

[0097] The AI ​​algorithm model decision-making process in this invention is designed specifically for the characteristics of industrial cooling scenarios using closed-loop cooling towers. Its core advantages are reflected in five dimensions: architectural synergy, dynamic decision-making, scenario adaptability, robustness, and interpretability. A detailed analysis follows: 1. Architecture Design: A hybrid architecture is used to balance decision accuracy and long-term optimality. A hybrid architecture of random forest pre-training and DQN reinforcement learning optimization is adopted to achieve accurate initial judgment and iterative optimization. Random Forest pre-trained model: trained based on 500,000 historical / simulated / experimental data, with an accuracy of ≥92%, quickly outputs preliminary pattern suggestions for dry cooling / evaporative cooling / cooperative cooling, avoiding blind exploration in the early stages of reinforcement learning and ensuring the accuracy of the decision-making basis; DQN reinforcement learning model: With the goal of optimizing long-term energy consumption, it optimizes the initial suggestions, balances cooling effect, energy saving requirements and equipment wear, and solves the limitations of static decision-making by single supervised learning. Architecture linkage: The collaboration between the two is achieved through confidence threshold, which ensures decision-making efficiency in normal scenarios and solves the dynamic adaptation problem in extreme scenarios.

[0098] 2. Decision-making logic, dynamic self-learning, constructing a three-layer self-learning mechanism of weekly + daily + scenario migration to achieve full lifecycle adaptation, adapting to all scenarios and the entire lifecycle: Weekly threshold dynamic optimization: Based on the recent 30 days of data, the threshold is grouped into low / medium / high dew point intervals, and the threshold is switched by grid search update mode to adapt to seasonal or humidity changes. Daily reinforcement learning fine-tuning: The experience replay pool is updated daily, and the weights of the DQN model and the reward function are fine-tuned to match the real-time running status. Scene migration adaptation: When the device is deployed to a new climate zone, the underlying feature layer is frozen and the top decision layer is fine-tuned through transfer learning. Only 1,000 new data points are needed for rapid adaptation, avoiding the high cost of retraining. Long-term optimal orientation: The reward function incorporates the mode switching frequency to avoid decisions that achieve short-term targets but result in high energy consumption / high losses in the long term, thus adapting to the long-term operation needs of industrial equipment.

[0099] 3. Scenario adaptation: Deeply adapts to the full operating conditions of closed-circuit cooling towers, with more refined decision-making logic and precise coverage of extreme environments and complex operating conditions: Optimization for extreme environments: Ultra-high temperature ≥40℃: Collaborative cooling mode is activated first, evaporative cooling deviation value is adjusted to +4℃, and stable high-load cooling is achieved; Ultra-low temperature <-10℃: The trigger threshold for dry cooling mode is relaxed to avoid atomization and icing, while reducing fan energy consumption; T1 has an extremely high dew point that is much higher than T_out: the cooling source starts in stages, reducing energy consumption by 25%-35% compared to traditional full-load start-up; Fault-tolerant design: Through dual-sensor redundancy + anomaly handling layer fallback, automatic switching to backup is performed when a single sensor fails, and evaporative cooling safety mode is triggered when multiple sensors fail, ensuring that decision-making does not fail; Operating condition segmentation decision: Based on 12 core features, the system segments scenarios according to the relationship between dew point and working fluid temperature, humidity saturation, etc., and outputs accurate modes to avoid one-size-fits-all decisions.

[0100] 4. Robustness: A four-layer anti-interference mechanism—data layer, model layer, decision layer, and anomaly handling layer—is constructed to ensure stable decision-making and adapt to complex industrial environments. Data layer: Dual sensor redundancy fusion + hierarchical filtering to suppress noise and instantaneous fluctuations; Model layer: During the pre-training stage, 10% artificial noise data is added, and core features are weighted to resist interference and improve the model's tolerance to data bias. Decision-making level: Mode switching cooling-off period + feature consistency verification to avoid accidental switching due to short-term interference; Anomaly handling layer: Decision bias verification + multi-sensor fault fallback, to deal with severe interference or failure and avoid equipment risks.

[0101] 5. Interpretability and industrial applicability, balancing quantitative transparency and flexible adaptation, better meeting the needs of industrial operation and maintenance, and reducing operation and maintenance costs: Strong interpretability: The decision-making process of the random forest is traceable, and the importance of core features is quantified, enabling operation and maintenance personnel to clearly understand the key influencing factors; The screening threshold and the weight of the reward function are adjustable parameters that can be flexibly configured according to user needs; Low implementation cost: Only 12 core features are retained after feature screening, and the model has a fast inference speed, meeting the real-time requirement of parameter adjustment for industrial equipment within 5 - 10 seconds; Scenario transfer learning only requires a small amount of new data and does not require full retraining, reducing the adaptation cost across regions and operating conditions; The anomaly handling mechanism is connected to the manual review process, conforming to the actual operation logic of industrial operation and maintenance and reducing the personnel learning cost.

[0102] In summary, the present invention deeply binds AI decision-making with the technical characteristics of closed-circuit cooling towers and industrial operation scenarios, solves the balance between accuracy and optimality through a hybrid architecture, solves the full-life cycle adaptation through a self-learning mechanism, and solves the industrial environment stability through an anti-interference mechanism, ultimately achieving the decision-making goals of intelligent collaboration, high energy efficiency, and stable reliability, which is also its core competitiveness compared with existing general-purpose AI decision-making.

[0103] Step 4, execution and switching, operating in a multi-modal cooling mode. The central controller controls each module to execute the corresponding cooling mode according to the model instructions and strictly follows the switching rules to ensure stable operation; Step 4.1, the operating logics of the three cooling modes of dry cooling mode, evaporative cooling mode, and collaborative cooling mode are specifically as Figure 3 shown: The triggering condition for the dry cooling mode is T1 < T_out, and the operating logic is to close the atomization system, open the dry air inlet valve, and directly exchange heat between the dry air and the finned tube main heat exchanger, with no evaporation loss, and preferentially save water resources; The triggering condition for the evaporative cooling mode is T1 ≈ T_out ± 2°C, and the operating logic is to start the atomization pump and the negative pressure fan, maintain a negative pressure of 0.8 - 0.9 atm inside the tower, and the atomized water evaporates on the outer surface of the main heat exchanger to make the working medium temperature approach the dew point; The triggering condition for the collaborative cooling mode is T1 > T_out, and the operating logic is to link the evaporation system and the waste heat recovery heat pump: the humid hot air enters the heat pump evaporator to release latent heat; the heat pump condenser transfers the cold quantity to the plate heat exchanger; the closed-circuit working medium sequentially passes through the main heat exchanger → the plate heat exchanger for double-stage cooling to reach the target temperature; at the same time, the refrigeration source starts in stages according to the dew point threshold and the working medium temperature, gradually increasing the load to avoid full-load operation.

[0104] Step 4.2, the mode switching rules for dry cooling mode, evaporative cooling mode, and synergistic cooling mode, to avoid equipment damage and cooling fluctuations, the switching must meet the time window and hysteresis threshold requirements, such as... Figure 4 As shown; Step 4.2.1, switch from dry cooling mode to evaporative cooling mode. The specific process is as follows: The switching trigger condition is that the dew point temperature T1 is within the working fluid outlet temperature T_out ±2℃ range for 20 consecutive seconds, and the lag threshold must be maintained for 20 seconds; in extreme environments, the deviation value will be corrected, with a deviation of +1℃ for ultra-low temperature environments <-10℃ and a deviation of ±3℃ for high temperature environments of 35~40℃. The central controller first continuously receives data such as T1 and T_out collected by the sensor group, and after confirming that all switching conditions are met, it outputs a switching command. The valve at air inlet 10 is fully open in dry cooling mode and remains fully open during the switching process to provide sufficient air for the subsequent evaporation process; The atomizing pump 11 starts from the off state, pressurizes and draws water from the source and delivers it to the atomizing nozzle 12. The nozzle then opens and evenly sprays tiny droplets onto the outer surface of the finned tubes of the main heat exchanger 2. The induced draft fan 7 gradually increases its speed from the low speed of the dry cooling mode, eventually forming a negative pressure environment of 0.8-0.9 atm inside the tower, guiding the outside air to flow through the outside of the finned tube and fully contact the atomized droplets; Plate heat exchanger 4 and the heat pump system remain shut down and do not participate in heat exchange during the entire switching process; The high-temperature working fluid output by the user-end load heat source 3 still flows through the main heat exchanger 2 along the original path. The heat exchange mode changes from sensible heat exchange with dry air to latent heat exchange with atomized droplets, and the working fluid temperature gradually approaches the dew point. The sensor array continuously monitors data such as dew point, working fluid temperature and flow rate, and negative pressure inside the tower, and feeds it back to the central controller to dynamically adjust the output power of the atomizing pump and the speed of the induced draft fan to ensure stable operation in evaporative cooling mode.

[0105] Step 4.2.2, switch from evaporative cooling mode to dry cooling mode. The specific process is as follows: The switching condition is that T1 < T_out - 2℃ is satisfied for 15 consecutive seconds, and the hysteresis threshold is maintained for 25 seconds; in the ultra-low temperature environment of <-10℃, the trigger threshold of dry cooling mode is relaxed to T1 < T_out + 1℃; After the central controller confirms that the switching conditions are met through sensor data, it sends a command to terminate evaporative cooling and start dry cooling. The atomizing pump 11 stops first, and the atomizing nozzle 12 closes accordingly, stopping the spraying of liquid droplets onto the main heat exchanger 2 to avoid wasting water resources. The induced draft fan 7 gradually decreases from medium-high speed to low speed in dry-cooling mode, maintaining only the basic airflow to meet the airflow requirements for sensible heat exchange between dry air and working fluid, while reducing energy consumption. The valve at air inlet 10 remains fully open, continuously introducing dry outside air; Plate heat exchanger 4 and the heat pump system remain shut down and do not participate in the heat exchange process; The heat exchange mode of the working fluid flowing through the main heat exchanger 2 changes from latent heat exchange to sensible heat exchange, and the cooling is achieved by the dry air absorbing the heat of the working fluid. The sensor array focuses on monitoring the dry air temperature and the temperature difference between the inlet and outlet of the working fluid, and feeds the data back to the central controller to ensure that the working fluid cooling effect meets the standards, while avoiding frequent start-ups and shutdowns of components.

[0106] Step 4.2.3, switch from evaporative cooling mode to collaborative cooling mode. The specific process is as follows: The switching condition is that T1 > T_out + 2℃ is met for 10 consecutive seconds, the lag threshold is maintained for 15 seconds, and the high temperature load is responded to quickly; in the high temperature environment of 35~40℃, the collaborative cold trigger threshold is advanced to T1 > T_out + 1℃, and the collaborative cold operation is stable in the ultra-high temperature environment of ≥40℃. After the central controller confirms the switching conditions, it outputs a coordinated cooling mode command, which starts the heat pump system and adjusts the status of relevant components while maintaining the operation of the evaporative cooling core components. The atomizing pump 11 and atomizing nozzle 12 continue to operate, maintaining the spraying of liquid droplets onto the outside of the finned tubes of the main heat exchanger 2, thus sustaining the atomization and evaporation process. The induced draft fan 7 maintains a medium-high speed to maintain a negative pressure of 0.8-0.9 atm inside the tower, while guiding the hot and humid air after heat exchange in the main heat exchanger to the evaporator 6. The heat pump system starts from the off state: the throttling and depressurizing component 9 depressurizes and cools the high-pressure refrigerant, forming a low-temperature liquid refrigerant that flows into the plate heat exchanger 4; the compressor 8 starts and compresses the gaseous refrigerant from the evaporator 6 into a high-temperature and high-pressure gaseous state; the condenser 5 is coupled to the plate heat exchanger 4 and transfers the cooling capacity to the plate heat exchanger. The working fluid output from the user-end load heat source 3 first flows through the main heat exchanger 2 and exchanges latent heat with the atomized droplets. After initial cooling, it enters the plate heat exchanger 4 and exchanges heat with the low-temperature refrigerant for a second time, thus achieving two-stage cooling. The air inlet valve 10 remains fully open to provide sufficient air for evaporation and latent heat recovery of the heat pump. The sensor array monitors dew point, working fluid temperature and flow rate, and refrigerant load data in real time. The central controller dynamically adjusts the atomization amount, fan speed, and compressor load to ensure that the working fluid reaches the target temperature while avoiding full-load startup of the refrigerant.

[0107] Step 4.2.4, switch from collaborative cooling mode to evaporative cooling mode. The specific process is as follows: The switching condition is that T1≤T_out+2℃ is satisfied for 30 consecutive seconds, and the lag threshold is maintained for 30 seconds to ensure a smooth switch from high energy consumption mode to low energy consumption mode. Once the central controller confirms that the switching conditions have been met through sensor data, it sends a command to shut down the heat pump system and maintain the operation of the evaporative cooling core. The heat pump system gradually shuts down: the compressor 8 stops working and no longer compresses the refrigerant; the throttling and depressurizing component 9 stops depressurizing and cooling the refrigerant; the transfer of cold energy between the condenser 5 and the plate heat exchanger 4 terminates, and the plate heat exchanger 4 stops participating in the heat exchange of the working fluid; The atomizing pump 11 and atomizing nozzle 12 continue to operate, spraying droplets onto the outside of the finned tubes of the main heat exchanger 2 to maintain the latent heat exchange of atomization evaporation. The induced draft fan 7 maintains a medium-high speed to maintain negative pressure inside the tower, and directly discharges the hot and humid air that has passed through the main heat exchanger outside the tower, instead of guiding it to the evaporator 6. The air inlet valve 10 remains fully open to provide sufficient air for the evaporation process; The heat exchange path of the working fluid changes from two-stage cooling of main heat exchanger + plate heat exchanger to latent heat exchange only through the main heat exchanger and atomized droplets. The sensor array continuously monitors the working fluid outlet temperature, tower negative pressure, and energy consumption data. The central controller adjusts the output power of the atomizing pump and the speed of the induced draft fan to ensure that the cooling effect meets the standard, while reducing the operational fluctuations caused by the heat pump shutdown.

[0108] Step 5, Closed-loop feedback: Operational status monitoring and adjustment. After the cooling mode is executed, the system continuously monitors the operational status to form a closed-loop optimization. Status monitoring: Real-time tracking of whether the working fluid outlet temperature meets the standard, whether the operating parameters of each module are normal, and whether the energy consumption is within a reasonable range; Dynamic adjustment: When the dew point changes, the control loop completes the adjustment of module parameters within 5-10 seconds to cope with instantaneous environmental fluctuations; Data recording and iteration: Record the running data of the current mode for monthly model iteration updates and continuous optimization of control strategies.

[0109] Step 6, special scenario handling; T1 is much higher than T_out's extreme high dew point: the cooling source is triggered by the dew point threshold and the working fluid temperature compensation logic is used to start in stages. First, the heat pump latent heat recovery is started. If the working fluid temperature is still not up to standard, the cooling load is gradually increased, which reduces energy consumption by 25%-35% compared to the traditional full load start. T1 is far below the extreme low dew point of T_out: maintaining dry cooling mode, reducing fan energy consumption, reducing spray water consumption by 15%-20%, and saving electricity by 10%-15%; Sensor failure: In case of single sensor failure, redundancy backup is activated; in case of multiple sensor failures, the system switches to evaporative cooling safety mode and alarms are triggered to ensure uninterrupted operation of the equipment.

[0110] The description of this invention is given for illustrative and descriptive purposes only and is not intended to be exhaustive or to limit the invention to the forms disclosed. Many modifications and variations will be apparent to those skilled in the art. The embodiments were chosen and described in order to better illustrate the principles and practical application of the invention and to enable those skilled in the art to understand the invention and design various embodiments with various modifications suitable for a particular purpose.

Claims

1. A control method for a closed-loop cooling tower based on an AI algorithm model for dynamic dew point temperature tracking, characterized in that: The control method is used to control a closed-circuit cooling tower, which includes a central controller, a main heat exchanger (2), an atomization system, and a heat pump system. The main heat exchanger (2) is connected to a user-side load heat source (3) through a closed-circuit working medium loop. The main heat exchanger (2) is also connected to a plate heat exchanger (4), and the plate heat exchanger (4) is connected to the heat pump system; The heat pump system includes an evaporator (6), a compressor (8), a condenser (5), and a throttling and pressure-reducing component (9). The compressor (8) is connected to the evaporator (6) and the condenser (5) on both sides. The output pipeline of the condenser (5) is connected to the plate heat exchanger (4). The throttling and pressure-reducing component (9) is installed on the output pipeline of the condenser (5). The evaporator (6) is connected to a draft fan (7); The atomization system includes an atomization pump (11), and the atomization pump (11) is connected to an atomization nozzle (12). The atomization nozzle (12) is arranged above the main heat exchanger (2); An air inlet (10) is provided below the main heat exchanger (2); The central controller is connected to a sensor group, an atomization pump (11), a draft fan (7), a compressor (8), and a valve of the air inlet (10); The cooling modes of the closed-circuit cooling tower include a dry cooling mode, an evaporative cooling mode, and a cooperative cooling mode; The control method includes the following steps: Step 1, real-time data acquisition. The central controller collects environmental raw data, equipment raw data, and energy efficiency raw data through the sensor group. The environmental raw data are the real-time dew point temperature T1, environmental temperature T2, and relative humidity RH. The equipment raw data are the working medium inlet temperature T_in, working medium outlet temperature T_out, working medium flow rate Q, spray water temperature, and fan speed. The energy efficiency raw data are the energy consumption per unit cooling capacity and the mode switching frequency; Step 2, data preprocessing. The data is subjected to filtering and smoothing processing, normalization processing, 3σ principle-based outlier detection processing, and feature extraction processing to eliminate interference and extract effective features; Step 3, the decision-making stage of AI algorithm model inference. The central controller is equipped with an AI algorithm model with a hybrid architecture of supervised learning pre-training and reinforcement learning online optimization, and outputs the optimal cooling mode based on the preprocessed data; Step 4, cooling mode execution and mode switching. The central controller controls each module to execute the corresponding cooling mode in the dry cooling mode, evaporative cooling mode, and cooperative cooling mode according to the model instructions. The trigger condition for the dry cooling mode is T1 < T_out; the trigger condition for the evaporative cooling mode is T1 ≈ T_out ± 2°C; the trigger condition for the cooperative cooling mode is T1 > T_out; Step 5, closed-loop feedback: operation status monitoring and adjustment. After the cooling mode is executed, the system continuously monitors the operation status to form a closed-loop optimization; Step 6, special scenario processing: extremely high dew point where T1 is much higher than T_out: the refrigeration source starts in stages according to the dew point threshold trigger + working medium temperature compensation logic. First, start the latent heat recovery of the heat pump system. If the working medium temperature still does not meet the standard, gradually increase the refrigeration load; extremely low dew point where T1 is much lower than T_out: maintain the dry cooling mode.

2. The control method for a closed cooling tower based on an AI algorithm model for dynamic dew point temperature tracking as described in claim 1, characterized in that: The feature extraction in step 2 includes the following steps: Step 2.1, calculate derivative features and deduce key associated features from the raw data; Working fluid temperature deviation: The calculation formula is current working fluid outlet temperature - target cooling temperature; Dew point to ambient temperature ratio: The calculation formula is real-time dew point temperature T1 / ambient temperature T2; The temperature difference between the inlet and outlet of the working fluid is calculated as: working fluid inlet temperature T_in - working fluid outlet temperature T_out; Cooling source load matching degree: The calculation formula is current cooling source load / rated cooling source load; The difference between spray water temperature and dew point is calculated as: spray water temperature - real-time dew point temperature T1. Unit energy consumption cooling efficiency: The calculation formula is working fluid temperature deviation / unit cooling energy consumption; Step 2.2, Feature Integration: The raw environmental data, raw equipment data, and raw energy efficiency data are merged with the six derived features calculated in Step 2.1 to form an initial feature pool. Step 2.3, Feature Filtering: Form the model input feature set, evaluate the feature importance based on the pre-trained random forest classifier, remove redundant and low-relevance features, and finally retain 12 core input features.

3. The control method for a closed cooling tower based on an AI algorithm model for dynamic dew point temperature tracking as described in claim 2, characterized in that: The specific process of step 2.3 is as follows: Step 2.3.1: Determine the evaluation index, and use the reduction in Gini coefficient natively supported by the random forest classifier as the core evaluation index; Step 2.3.2, AI algorithm model training and importance calculation; The 17 initial features after feature integration and the corresponding optimal cooling mode labels are input into the random forest classifier to complete full feature training. During model training, each decision tree records the changes in the Gini coefficient when each feature participates in the split. Finally, the average of the results of 200 decision trees is used to obtain the Gini importance score of each feature, which is normalized to the [0,1] interval. For the top 15 features in Gini score, the importance of the ranking is further calculated: each feature is randomly shuffled 10 times, and the model classification accuracy is recalculated after each shuffle. The average decrease in accuracy is calculated and used as the secondary validation score. Step 2.3.3, Feature importance ranking and threshold setting; The combined scores of the two indicators are used to calculate the overall importance score of the features, which is calculated as Gini score × 0.7 + ranking importance score × 0.3, and then sorted in descending order. Set a screening threshold: a hard threshold, with a comprehensive importance score ≥ 0.05, meaning that the feature contributes no less than 5% to the model's decision. Features below this threshold are considered low-value redundant features. A soft threshold is used to exclude strongly correlated feature pairs. The Pearson correlation coefficient is used for detection. If the correlation coefficient between two features is ≥0.8, the feature with the higher score is retained to avoid feature redundancy that could lead to model overfitting. Step 2.3.4: Screening, Verification, and Final Determination; Candidate features are selected based on a threshold, approximately 12-14 items, and a simplified feature set is constructed. Retrain the random forest classifier using the candidate feature set and verify the classification accuracy: If the accuracy is greater than or equal to 95% of the original full-feature model, meaning that the model performance does not significantly decrease after screening, the candidate set is retained; if the accuracy is less than 95%, 1-2 features with the second highest scores are re-validated until the performance requirements are met; finally, 12 core input features are determined. Step 2.3.5, Feature extraction output results; The final output consists of 12 core feature vectors, covering three dimensions: environmental status, equipment operation, and energy efficiency performance. Specifically, these include: environmental features: real-time dew point temperature, ambient temperature, relative humidity, and the ratio of dew point to ambient temperature; equipment features: temperature difference between inlet and outlet of working fluid, working fluid flow rate, difference between spray water temperature and dew point, fan speed, and matching degree of cooling source load; and energy efficiency features: energy consumption per unit of cooling capacity, mode switching frequency, and cooling efficiency per unit of energy consumption.

4. The control method for a closed cooling tower based on an AI algorithm model for dynamic dew point temperature tracking as described in claim 3, characterized in that: Step 3 includes the following steps: Step 3.1: Preliminary assessment of the pre-trained model. The 12 feature vectors are input into the pre-trained random forest classifier. The specific process is as follows: Step 3.1.1, Preliminary preparations: Feature vectors and model state are ready; Input feature vector standardization: The 12 core features have been preprocessed and normalized and aligned in dimension. The dimension alignment is arranged in a fixed order: real-time dew point temperature → ambient relative humidity → ratio of dew point to ambient temperature → temperature difference between working fluid inlet and outlet → working fluid flow rate → difference between spray water temperature and dew point → fan speed → cooling source load matching degree → energy consumption per unit cooling capacity → mode switching frequency → cooling efficiency per unit energy consumption → working fluid temperature deviation, forming a 1×12-dimensional feature vector, which is input into a pre-trained random forest classifier. Step 3.1.2, Classification reasoning of feature vectors by a single decision tree 200 decision trees independently determine the input feature vector, and the reasoning logic of each tree is consistent; Splitting rules: Each node is determined by minimizing the Gini coefficient during pre-training based on the optimal splitting threshold of the feature values. For example, if the normalized value of the working fluid temperature deviation is >0.3, it does not meet the standard. The feature vector is then assigned to the next level child node until the leaf node is reached. Leaf node output category: Each leaf node corresponds to a unique cooling mode among dry cooling / evaporative cooling / co-cooling. The leaf node category that the feature vector finally falls into is the preliminary judgment result of the decision tree. Step 3.1.3: Integrated voting of 200 decision trees; Each decision tree independently outputs one cooling mode decision, forming a set of 200 independent decision results; the number of votes for each type of cooling mode in the set is counted; according to the principle of majority rule, the mode with the most votes is selected as the ensemble decision result of the random forest. Step 3.1.4: Pattern suggestion output and confidence level labeling; Output initial pattern suggestions: The ensemble decision results are used as initial pattern suggestions and passed to the subsequent DQN reinforcement learning model; Calculate the confidence level of the model: Confidence level = (number of highest votes / total number of decision trees) × 100%, which is used as a reference for subsequent optimization of the DQN model. If the confidence level is ≥80%, the DQN model will retain the suggestion first; if the confidence level is <50%, the DQN model will focus on optimization. Step 3.1.5, Optimize the reinforcement learning model; The DQN reinforcement learning model aims to optimize long-term energy consumption by optimizing the initial suggestions. Its state space is the current sensor feature vector, and its action space consists of three cooling modes. The reward function is set as: R=α*(target temperature achievement rate)-β*(unit energy consumption)-γ*(mode switching frequency), where α, β, and γ are adjustable weight coefficients. Finally, it outputs the determined cooling mode command.

5. The control method for a closed cooling tower based on an AI algorithm model for dynamic dew point temperature tracking as described in claim 3, characterized in that: Step 3 also includes the following steps: Step 3.2, AI algorithm self-learning mechanism, the specific process is as follows; Step 3.2.1, Weekly threshold dynamic optimization: For the random forest classifier, based on recent running data, adjust the mode switching thresholds for dry-cold / evaporative-cold / co-current-cold to adapt to seasonal or environmental humidity changes; Data filtering and preprocessing: Triggered weekly, extracting runtime data from the most recent 30 days; Filter valid data: Remove abnormal data caused by sensor failure or extreme interference, and retain records with a cooling effect compliance rate of ≥80% and no sudden changes in energy consumption data; The data were grouped into three ranges: low dew point <10℃, medium dew point 10-25℃, and high dew point >25℃. The optimal combination of mode → energy consumption → compliance rate was calculated for each group. For each dew point interval, a grid search method is used to traverse the possible threshold ranges and calculate the comprehensive energy efficiency score under different thresholds. The score = 0.6 compliance rate + 0.4 (1 - relative energy consumption). The random forest classifier was retrained using the filtered 30-day data, and the split threshold of the decision tree was fine-tuned. Step 3.2.2: Daily online fine-tuning of reinforcement learning. Based on daily operational data feedback, the network weights of the DQN model are adjusted to make the mode decision more aligned with the long-term goal of minimizing energy consumption. Experience data storage: Triggered once every 24 hours of operation, the status-action-reward data of the day is added to the experience replay pool, with a fixed capacity of 10,000 records. A first-in-first-out strategy is adopted to remove the oldest data. The newly added data is labeled to supplement extended information such as energy consumption changes, cooling effect and stability in the following hour, enriching the evaluation dimensions of the reward function. DQN model fine-tuning training: 1000 data points are randomly sampled from the experience replay pool as the fine-tuning training set. The training parameters are set as follows: learning rate = 0.001, training epochs = 50 epochs, batch size = 32. A target network and evaluation network separation mechanism is adopted. The evaluation network updates its weights based on new data, and the target network synchronizes the weights of the evaluation network every 100 steps. The reward function is dynamically adjusted: the weight coefficients of α, β, and γ are fine-tuned according to the energy consumption performance of the day. The exploration and utilization balance is achieved using an ε-greedy strategy: the ε value decreases over time, starting at 0.3 and decreasing by 0.05 every 30 days, reaching a minimum of 0.

05. Step 3.2.3, Scene migration adaptation: When the device is deployed to a new climate zone, it can quickly adapt to the new scene through transfer learning, reducing the cost of retraining.

6. The control method for a closed cooling tower based on an AI algorithm model for dynamic dew point temperature tracking as described in claim 3, characterized in that: Step 3 also includes the following steps: Step 3.3, Supplementing the AI ​​algorithm model to resist interference, constructing a four-layer protection mechanism consisting of a data layer, a model layer, a decision layer, and an anomaly handling layer. The specific process is as follows: Step 3.3.1: Redundant data acquisition and filtering / denoising at the data layer to ensure the data input to the model is authentic and stable. Multi-sensor redundancy backup and fusion: For the core input real-time dew point temperature T1, two independent dew point sensors are deployed to collect data simultaneously; under normal circumstances, the average value of the two sensor data is taken as the input; if the data of a single sensor exceeds the reasonable range or there is no response, it is automatically judged as a fault and the system switches to the valid data of the other sensor; if both sensors fail, the subsequent abnormal handling mechanism is triggered. Layered filtering processing: Basic filtering uses a 5-second sliding window filter on all sensor data. The calculation formula is: Current filter value = (Data from the previous 4 seconds + Current data) / 5, which smooths high-frequency noise; Dynamic data enhancement filtering adds Kalman filtering to data with drastic dynamic changes in working fluid temperature and refrigerant load, and corrects them through prediction-update iteration. Electromagnetic interference adaptive protection is implemented by deploying electromagnetic interference sensors to monitor the intensity of environmental electromagnetic interference in real time. An interference threshold is set, with ≥10V / m considered excessive. If the interference does not exceed the threshold, the original filtering intensity is maintained. If the interference exceeds the threshold, the sliding window is automatically expanded to 10 seconds, and the prediction weight of the Kalman filter is increased to further suppress noise. Step 3.3.2, robust training and anti-interference adaptation of the model layer, improves the model's tolerance to interference, so that the model can still output the correct decision when there is a slight deviation in the input data, and avoids being misled by the interference data; Step 3.3.3, Decision-making smoothing mechanism and switching constraints; Mode switching cooldown setting: Set a fixed cooldown period, and continuous mode switching is not allowed within 5 minutes; During the cooldown period, the processing logic continuously monitors the data. If the triggering conditions for the new mode are still met after the cooldown period ends, the switch will be executed; if the conditions disappear, the current mode will be maintained. The decision results are cross-validated. If the model recommendation is collaborative cooling, it must simultaneously meet three core characteristic conditions: dew point > target temperature, working fluid temperature deviation > 0, and the difference between spray water temperature and dew point < 2℃. If only one characteristic is met, but the other characteristics are contradictory, it is judged as an erroneous recommendation caused by interference, and it will not be output for the time being. The current model will continue to be used. Handling of verification failure: The current data is marked as suspicious data and stored in a separate database for subsequent model iteration analysis of interference patterns. Step 3.3.4, Fallback mechanism and manual intervention in the exception handling layer; Multi-sensor fallback: Set fault judgment criteria. If ≥3 core sensors fail simultaneously or the data is abnormal, i.e., exceeding the mean ± 3 times the standard deviation, the safety mode is automatically triggered. Safety mode logic: switch to evaporative cooling mode, and shut down the refrigeration source staged start logic to maintain a fixed spray volume and fan speed. Alarm mechanism: send alarm information to operation and maintenance personnel through the equipment control system, mark the location of the faulty sensor and abnormal data, and prompt timely maintenance. Decision deviation verification and manual review: Real-time calculation of the deviation between the model output and the actual operating conditions. Deviation value = (average energy consumption within 10 minutes after mode switch / average energy consumption within 10 minutes before switch) - (cooling compliance rate within 10 minutes after mode switch / cooling compliance rate within 10 minutes before switch). A deviation threshold is set. If the deviation value is ≥20%, such as a sudden increase in energy consumption after mode switch but no improvement in cooling effect, it is judged as a decision deviation and the manual review process is triggered.

7. The control method for a closed-loop cooling tower based on an AI algorithm model for dynamic dew point temperature tracking as described in claim 1, characterized in that: The specific process for switching from dry cooling mode to evaporative cooling mode in step 4 is as follows: The switching trigger condition is that the dew point temperature T1 is within the working fluid outlet temperature T_out ±2℃ range for 20 consecutive seconds, and the lag threshold must be maintained for 20 seconds; in extreme environments, the deviation value will be corrected, with a deviation of +1℃ for ultra-low temperature environments <-10℃ and a deviation of ±3℃ for high temperature environments of 35~40℃. The central controller first continuously receives data such as T1 and T_out collected by the sensor group, and after confirming that all switching conditions are met, it outputs a switching command. The valve of the air inlet (10) is fully open in dry cooling mode and remains fully open during the switching process to provide sufficient air for the subsequent evaporation process; The atomizing pump (11) starts from the off state, pressurizes and draws water from the source and delivers it to the atomizing nozzle (12). The nozzle then opens and sprays tiny droplets evenly onto the outer surface of the finned tube of the main heat exchanger (2). The induced draft fan (7) gradually increases from the low speed of the dry cooling mode to eventually form a negative pressure environment of 0.8-0.9 atm in the tower, guiding the outside air to flow through the outside of the finned tube and fully contact the atomized droplets; The plate heat exchanger (4) and the heat pump system remain shut down throughout the switching process and do not participate in heat exchange; The high-temperature working fluid output by the user-end load heat source (3) still flows through the main heat exchanger (2) along the original path. The heat exchange mode changes from sensible heat exchange with dry air to latent heat exchange with atomized droplets, and the working fluid temperature gradually approaches the dew point. The sensor array continuously monitors data such as dew point, working fluid temperature and flow rate, and negative pressure inside the tower, and feeds it back to the central controller to dynamically adjust the output power of the atomizing pump and the speed of the induced draft fan to ensure stable operation in evaporative cooling mode.

8. The control method for a closed-loop cooling tower based on an AI algorithm model for dynamic dew point temperature tracking as described in claim 1, characterized in that: The specific process for switching from evaporative cooling mode to dry cooling mode in step 4 is as follows: The switching condition is that T1 < T_out - 2℃ is satisfied for 15 consecutive seconds, and the hysteresis threshold is maintained for 25 seconds; in the ultra-low temperature environment of <-10℃, the trigger threshold of dry cooling mode is relaxed to T1 < T_out + 1℃; After the central controller confirms that the switching conditions are met through sensor data, it sends a command to terminate evaporative cooling and start dry cooling. The atomizing pump (11) stops first, and the atomizing nozzle (12) closes accordingly, stopping the spraying of liquid droplets onto the main heat exchanger (2) to avoid wasting water resources; The induced draft fan (7) is gradually reduced from medium-high speed to low speed in dry-cooling mode, maintaining only the basic airflow to meet the airflow requirements for sensible heat exchange between dry air and working fluid, while reducing energy consumption. The valve of the air inlet (10) remains fully open, continuously introducing dry outside air; The plate heat exchanger (4) and the heat pump system remain shut down and do not participate in the heat exchange process; The heat exchange mode of the working fluid flowing through the main heat exchanger (2) changes from latent heat exchange to sensible heat exchange, and the cooling is achieved by absorbing the heat of the working fluid through dry air; The sensor array focuses on monitoring the dry air temperature and the temperature difference between the inlet and outlet of the working fluid, and feeds the data back to the central controller to ensure that the working fluid cooling effect meets the standards, while avoiding frequent start-ups and shutdowns of components.

9. The control method for a closed cooling tower based on an AI algorithm model for dynamic dew point temperature tracking as described in claim 1, characterized in that: The specific process for switching from evaporative cooling mode to collaborative cooling mode in step 4 is as follows: The switching condition is that T1 > T_out + 2℃ is met for 10 consecutive seconds, the lag threshold is maintained for 15 seconds, and the high temperature load is responded to quickly; in the high temperature environment of 35~40℃, the collaborative cold trigger threshold is advanced to T1 > T_out + 1℃, and the collaborative cold operation is stable in the ultra-high temperature environment of ≥40℃. After the central controller confirms the switching conditions, it outputs a coordinated cooling mode command, which starts the heat pump system and adjusts the status of relevant components while maintaining the operation of the evaporative cooling core components. The atomizing pump (11) and atomizing nozzle (12) continue to operate, spraying droplets onto the outside of the finned tubes of the main heat exchanger (2) to maintain the atomization evaporation process; The induced draft fan (7) maintains a medium-high speed and maintains a negative pressure of 0.8-0.9 atm inside the tower, while guiding the humid and hot air after the main heat exchanger has completed heat exchange to the evaporator (6). The heat pump system starts from the off state: the throttling and depressurizing component (9) depressurizes and cools the high-pressure refrigerant, forming a low-temperature liquid refrigerant that flows into the plate heat exchanger (4); the compressor (8) starts and compresses the gaseous refrigerant from the evaporator (6) into a high-temperature and high-pressure gaseous state; the condenser (5) is coupled to the plate heat exchanger (4) and transfers the cooling capacity to the plate heat exchanger; The working fluid output from the user-end load heat source (3) first flows through the main heat exchanger (2) and exchanges latent heat with the atomized droplets. After initial cooling, it enters the plate heat exchanger (4) and exchanges heat with the low-temperature refrigerant for a second time, thus achieving dual-stage cooling. The air inlet (10) valve remains fully open to provide sufficient air for evaporation and latent heat recovery of the heat pump; The sensor array monitors dew point, working fluid temperature and flow rate, and refrigerant load data in real time. The central controller dynamically adjusts the atomization amount, fan speed, and compressor load to ensure that the working fluid reaches the target temperature while avoiding full-load startup of the refrigerant.

10. The control method for a closed-loop cooling tower based on an AI algorithm model for dynamic dew point temperature tracking as described in claim 1, characterized in that: The specific process for switching from collaborative cooling mode to evaporative cooling mode in step 4 is as follows: The switching condition is that T1≤T_out+2℃ is satisfied for 30 consecutive seconds, and the lag threshold is maintained for 30 seconds to ensure a smooth switch from high energy consumption mode to low energy consumption mode. Once the central controller confirms that the switching conditions have been met through sensor data, it sends a command to shut down the heat pump system and maintain the operation of the evaporative cooling core. The heat pump system is gradually shut down: the compressor (8) stops working and no longer compresses the refrigerant; The throttling and depressurizing component (9) stops depressurizing and cooling the refrigerant; the transfer of cold energy between the condenser (5) and the plate heat exchanger (4) ends, and the plate heat exchanger (4) stops participating in the heat exchange of the working fluid. The atomizing pump (11) and atomizing nozzle (12) continue to operate, spraying droplets onto the outside of the finned tubes of the main heat exchanger (2) to maintain the latent heat exchange of atomization evaporation; The induced draft fan (7) maintains a medium-high speed, maintains negative pressure inside the tower, and directly discharges the humid and hot air after passing through the main heat exchanger outside the tower, no longer guiding it to the evaporator (6). The air inlet (10) valve remains fully open to provide sufficient air for the evaporation process; The heat exchange path of the working fluid changes from two-stage cooling of main heat exchanger + plate heat exchanger to latent heat exchange only through the main heat exchanger and atomized droplets. The sensor array continuously monitors the working fluid outlet temperature, tower negative pressure, and energy consumption data. The central controller adjusts the output power of the atomizing pump and the speed of the induced draft fan to ensure that the cooling effect meets the standard, while reducing the operational fluctuations caused by the heat pump shutdown.

Images