Dynamic compensation method for evaporation loss of cooling tower

Abstract

The invention provides a dynamic compensation method for evaporation loss of a cooling tower, which comprises the following steps of: acquiring key parameters such as inlet and outlet water temperature difference, circulating water quantity, environment wet bulb temperature, fan state and water replenishing record in real time through an industrial sensor network and a historical database, and inputting the key parameters into a high-level strategy network after standardization and normalization processing; dynamically generating water-saving, energy-saving and stable priority weights and modulating a reward function structure; based on a depth deterministic strategy gradient algorithm and in combination with a multi-target weight adaptive adjustment mechanism, reasoning an optimal compensation water volume instruction to realize closed-loop dynamic compensation of evaporation loss; a cross-working-condition migration training and online fine adjustment mechanism is introduced, the model generalization and field adaptability are improved, the response speed and accuracy of a compensation strategy to multi-target collaborative optimization are improved, water resource saving, energy consumption reduction, operation safety and adaptation to different working conditions are considered, and long-term efficient operation of the cooling tower is effectively guaranteed.

Description

Technical Field

[0001] This invention relates to the field of multi-objective collaborative optimization control technology for cooling towers, and in particular to a dynamic compensation method for evaporation loss in cooling towers. Background Technology

[0002] In the current field of intelligent control of cooling tower operation and multi-objective optimization of industrial processes, mainstream technical solutions for dynamic compensation of evaporation loss in cooling towers typically include rule-driven water compensation methods, traditional predictive models (such as statistical regression and simple neural networks), and reinforcement learning strategies based on single-objective optimization. These solutions aim to achieve efficient water replenishment and energy consumption control of cooling towers, thereby improving the economy and stability of system operation. Existing publicly available technologies primarily focus on generating water replenishment commands based on evaporation predictions, using preset rules or static reward mechanisms, or weighting objectives such as water saving, energy saving, and system stability using empirical parameters to output compensation strategies. Some recent technologies are beginning to incorporate reinforcement learning frameworks, improving the accuracy of water replenishment adjustments through real-time perception of operating environment parameters and adaptive strategy optimization. For example, some reinforcement learning-based compensation systems can automatically adjust water replenishment based on the cooling tower inlet and outlet water temperature difference, circulating water volume, and environmental meteorological conditions, while also considering energy management and system safety, representing a growing trend in this field. However, current technologies generally have the following limitations: (1) Compensation strategies often adopt single-objective or static weighted reward mechanisms, which make it difficult to balance the conflict of multiple objectives such as water conservation, energy conservation and system stability. This can easily lead to an imbalance in resource allocation, resulting in a long-term neglect of one objective such as water quantity, energy consumption or system stability. (2) When the operating environment fluctuates drastically, traditional models cannot autonomously adjust the weights between objectives, which can easily lead to the failure of compensation strategies or system instability. In addition, existing reinforcement learning methods lack identification and balancing measures for competition and cooperation among multiple objectives, which poses a risk of strategy collapse and optimization bias. (3) Most solutions rely on fixed rules or static parameter configurations, which have limited generalization ability and are difficult to adapt to dynamic changes under complex working conditions such as different power plants and different climate zones, thus affecting the efficiency of water resource management and the improvement of the overall energy efficiency of the system. Summary of the Invention

[0003] In order to solve the above-mentioned technical problems, the present invention provides a method for dynamic compensation of evaporation loss in cooling towers.

[0004] The technical solution of this invention is implemented as follows: A method for dynamic compensation of evaporation loss in a cooling tower, comprising: S1: Obtain real-time environmental parameters during the operation of the cooling tower, including the inlet / outlet water temperature difference, circulating water volume, ambient wet-bulb temperature, fan operating status, and historical water replenishment records, as input conditions for multi-objective collaborative optimization decision-making. S2: Normalize and standardize the acquired real-time environmental parameters to eliminate the interference of different units on the model training process and improve the convergence efficiency of subsequent reinforcement learning algorithms. S3: Input the standardized environmental parameters into the high-level policy network, and combine them with the system operation target weight configuration file to generate a multi-objective priority vector under the current operating condition. The multi-objective priority vector includes water-saving priority, energy-saving priority, or stability priority modes. S4: Based on the multi-objective priority vector as the context modulation signal, dynamically adjust the reward function structure of the underlying action network; S5: The deep deterministic policy gradient algorithm is adopted. The multivariate input state space includes temperature gradient change rate, liquid level fluctuation trend and load fluctuation intensity as input. The policy network inference is executed to output specific compensation water volume adjustment instructions. S6: Construct a target competitiveness evaluation unit, calculate the Pareto front distance between each target, and if it is detected that a target continuously suppresses other targets for more than a set threshold, trigger the weight rebalancing mechanism to forcibly increase the short-term weight gain of the suppressed target. S7: Introducing a cross-operating condition transfer training strategy, using historical operating data from multiple power plants and climate zones for pre-training during the offline training phase to enhance the model's generalization ability under different operating conditions. S8: During system operation, an online fine-tuning mechanism is executed to update the local parameters of the pre-trained model based on the real-time feedback data of the current project, so as to improve the model's adaptability to specific field characteristics; S9: The generated compensation water volume adjustment command is transmitted to the dynamic compensation module through the API interface to ensure that the compensation command has sufficient responsiveness and robustness to realize the dynamic adjustment function of the closed-loop control system. S10: Periodically evaluate the long-term operational performance feedback of the system, dynamically evolve the weight coefficients of each objective in the reward function, form a closed-loop adaptive optimization mechanism, prevent strategy collapse and improve the overall energy efficiency of the system.

[0005] The present invention provides a dynamic compensation method for evaporation loss in cooling towers, which has the following beneficial effects: (1) This invention decouples and dynamically assigns weights to multiple optimization tasks such as water saving, energy saving and stable operation through a hierarchical strategy architecture. It adopts a meta-reward function and weight coefficients generated in real time by high-level strategies, effectively avoiding decision bias and strategy imbalance caused by a single or static reward function, realizing the balanced adaptation of the compensation strategy to the needs of each objective, and improving the flexibility and robustness of system optimization. (2) This invention, through innovative target competitiveness evaluation and Pareto front distance algorithm, can promptly identify the phenomenon of a certain target continuously suppressing other targets, leading to system optimization imbalance, and automatically triggers a weight rebalancing mechanism to temporarily increase the weight of the suppressed target and prevent the strategy from falling into single-target polarization collapse. This mechanism ensures that the system can maintain a dynamic balance among multiple operating targets in the long term under complex working conditions, significantly enhancing the stability and adaptability of the compensation strategy; (3) This invention employs the Deep Deterministic Strategy Gradient Algorithm (DDPG) in the water volume prediction and compensation command generation stage, combined with rich multivariate inputs, to achieve high-precision, low-latency dynamic compensation command output. Its closed-loop feedback mechanism and API interface enable the compensation command to have a forward-looking response capability, ensuring the timeliness and operational robustness of the dynamic adjustment of cooling tower compensation; (4) This invention significantly improves the generalization performance and practical adaptability of the model in diverse working conditions and complex environments through cross-condition transfer training (pre-training with data from multiple power plants and multiple climate zones) and online fine-tuning mechanism with real-time feedback from the field. The compensation strategy can quickly adapt to the actual dynamics on site, avoid the failure of traditional models due to changes in working conditions, and ensure the universality and long-term operational stability of the compensation mechanism in different industrial sites. Attached Figure Description

[0006] Figure 1 This is a flowchart of a dynamic compensation method for evaporation loss in a cooling tower according to the present invention; Figure 2 This is a sub-flowchart of a dynamic compensation method for evaporation loss in a cooling tower according to the present invention; Figure 3 This is another sub-flowchart of a dynamic compensation method for evaporation loss in a cooling tower according to the present invention. Detailed Implementation

[0007] Embodiments of the present invention are described in detail below, examples of which are shown in the accompanying drawings, wherein the same or similar reference numerals denote the same or similar elements or elements having the same or similar functions throughout. The embodiments described below with reference to the accompanying drawings are exemplary and are only used to explain the present invention, and should not be construed as limiting the present invention.

[0008] The following disclosure provides many different embodiments or examples for implementing different structures of the invention. To simplify the disclosure, specific examples of components and arrangements are described below. Of course, these are merely examples and are not intended to limit the invention. Furthermore, reference numerals and / or letters may be repeated in different examples; such repetition is for simplification and clarity and does not in itself indicate a relationship between the various embodiments and / or arrangements discussed.

[0009] like Figure 1 As shown, the present invention provides a dynamic compensation method for evaporation loss in a cooling tower, specifically including: S1: Obtain real-time environmental parameters during the operation of the cooling tower, including the inlet / outlet water temperature difference, circulating water volume, ambient wet-bulb temperature, fan operating status, and historical water replenishment records, as input conditions for multi-objective collaborative optimization decision-making. S2: Normalize and standardize the acquired real-time environmental parameters to eliminate the interference of different units on the model training process and improve the convergence efficiency of subsequent reinforcement learning algorithms. S3: Input the standardized environmental parameters into the high-level policy network, and combine them with the system operation target weight configuration file to generate a multi-objective priority vector under the current operating condition. The multi-objective priority vector includes water-saving priority, energy-saving priority, or stability priority modes. S4: Based on the multi-objective priority vector as the context modulation signal, dynamically adjust the reward function structure of the underlying action network; S5: The deep deterministic policy gradient algorithm is adopted. The multivariate input state space includes temperature gradient change rate, liquid level fluctuation trend and load fluctuation intensity as input. The policy network inference is executed to output specific compensation water volume adjustment instructions. S6: Construct a target competitiveness evaluation unit, calculate the Pareto front distance between each target, and if it is detected that a target continuously suppresses other targets for more than a set threshold, trigger the weight rebalancing mechanism to forcibly increase the short-term weight gain of the suppressed target. S7: Introducing a cross-operating condition transfer training strategy, using historical operating data from multiple power plants and climate zones for pre-training during the offline training phase to enhance the model's generalization ability under different operating conditions. S8: During system operation, an online fine-tuning mechanism is executed to update the local parameters of the pre-trained model based on the real-time feedback data of the current project, so as to improve the model's adaptability to specific field characteristics; S9: The generated compensation water volume adjustment command is transmitted to the dynamic compensation module through the API interface to ensure that the compensation command has sufficient responsiveness and robustness to realize the dynamic adjustment function of the closed-loop control system. S10: Periodically evaluate the long-term operational performance feedback of the system, dynamically evolve the weight coefficients of each objective in the reward function, form a closed-loop adaptive optimization mechanism, prevent strategy collapse and improve the overall energy efficiency of the system.

[0010] Step S1: Obtain real-time environmental parameters during the operation of the cooling tower. These parameters include the cooling tower inlet / outlet water temperature difference, circulating water volume, ambient wet-bulb temperature, fan operating status, and historical water replenishment records, serving as input conditions for multi-objective collaborative optimization decision-making. Specifically, this includes: S1.1: Based on an industrial sensor network, collect the difference data between the inlet water temperature and the outlet water temperature during the operation of the cooling tower to obtain the temperature gradient change rate, which is used to reflect the dynamic fluctuation of cooling efficiency. Based on the high-precision temperature acquisition unit (parameters: Pt100 RTD sensor type, Class A accuracy level) in the industrial sensor network, the differential sampling method is used to realize the synchronous reading of the temperature of the cooling tower inlet pipe and outlet pipe, so as to ensure the consistency of sampling time and reduce the hysteresis deviation introduced by external interference. Furthermore, the inlet water temperature signal and the outlet water temperature signal are paired and merged under the same time base by using a timestamp alignment algorithm (parameters: sampling frequency 1Hz, time synchronization protocol IEEE1588) to generate an initial temperature difference sequence and remove instantaneous outliers caused by sensor failure or network fluctuations. Furthermore, a sliding window averaging filter method (parameters: window length 5 min, overlap rate 50%) was used to smooth the initial temperature difference sequence, and a second-order Butterworth low-pass filter (cutoff frequency 0.005 Hz) was used to filter out high-frequency noise components to obtain the dynamic change curve of the cooling tower temperature difference. Furthermore, the rate of change of the temperature gradient is calculated based on the following formula:

[0011] in, The inlet water temperature, The outlet water temperature, The time interval between adjacent samplings. This represents the rate of change of the temperature gradient. During the calculation, a difference operation is performed at each sampling point, and a smoothing curve is incorporated to improve the stability and noise resistance of the gradient calculation. Furthermore, the temperature gradient change rate is dimensionlessly eliminated by the Z-score normalization method (parameters: mean μ and standard deviation σ are calculated from historical data), resulting in a normalized temperature gradient feature vector that is adapted to the input of the subsequent high-level policy network. By using the above differential sampling, time alignment, filtering and smoothing and gradient calculation processing methods, the temperature difference between the inlet and outlet water during the operation of the cooling tower is transformed into quantifiable dynamic temperature gradient change rate data, so as to realize the real-time characterization and accurate reflection of the cooling efficiency fluctuation state. For example, in a cooling tower system with a circulating water volume of 3000 m³ / h, the sampling frequency of the Pt100 resistance temperature detector (RTD) sensors installed on the inlet and outlet pipes is set to 1 Hz. The temperature signals are time-aligned using the IEEE 1588 protocol to generate a temperature difference sequence. This sequence is then filtered and smoothed using a 5-minute moving average with a 50% window overlap to remove high-frequency noise above 0.005 Hz, outputting a temperature difference change curve. For 10 minutes of continuous operation data, the inlet water temperature is maintained between 35.2°C and 36.0°C, and the outlet water temperature is maintained between 30.5°C and 31.2°C. The temperature gradient change rate is calculated using the formula... Under the condition of 60s, the temperature gradient change rate was found to be between 0.048°C / s and 0.065°C / s. After Z-score standardization, the mean of the standardized temperature gradient feature vector was 0 and the variance was 1. It was directly input into the high-level strategy network to provide accurate dynamic fluctuation data support for cooling efficiency for multi-objective weight calculation of water saving, energy saving and stability. S1.2: Read the operating status signal of the circulating water pump in real time to obtain the current circulating water volume data, so as to construct the water volume change trend characteristics and use them as the input basis for load fluctuation intensity analysis in the compensation strategy; S1.3: Obtain the wet-bulb temperature of the air around the cooling tower through environmental temperature and humidity sensors to calculate air saturation and evaporation potential, providing environmental driving factors for compensation decisions under the water-saving priority mode; Using an ambient temperature and humidity sensor array (parameter: measurement range) The system can collect real-time temperature and relative humidity signals of the air surrounding the cooling tower, ranging from 40℃ to 80℃ and from 0%RH to 100%RH, with an accuracy of ±0.2℃ and ±2%RH, to achieve real-time monitoring of the external climate environment. Furthermore, through the wet-bulb temperature calculation algorithm (input parameter: dry-bulb temperature) relative humidity This paper proposes a method for calculating wet-bulb temperature using a combination of lookup table method and Magnus formula, achieving high-precision wet-bulb temperature estimation across multiple humidity ranges and outputting a wet-bulb temperature sequence. This is to eliminate the calculation bias of a single formula under extreme conditions; Furthermore, the air saturation calculation formula (parameter: wet-bulb temperature) is used. Dry bulb temperature Atmospheric pressure ) Calculate the saturated vapor partial pressure The formula is as follows:

[0012] in This is the saturated vapor pressure, in kPa. This is a percentage value for relative humidity, used to describe the degree of water vapor saturation in the air; Furthermore, based on the evaporation potential calculation model (parameter: wet-bulb temperature) Water temperature air saturation Evaporation potential is calculated using a heat transfer and mass transfer model. The formula is:

[0013] in To consider the overall heat transfer and mass transfer coefficients, This is the difference between water temperature and wet-bulb temperature. This is the air saturation difference, used to characterize the potential ability of air to absorb vapor; The above calculation outputs the environmental driving factor vector. As the core input for compensation decisions under the water-saving priority mode, it enables quantitative reference and dynamic adaptation to changes in external meteorological conditions; For example, in a high-humidity summer operation scenario, the ambient temperature and humidity sensor collects the dry-bulb temperature around the cooling tower as follows: ℃, relative humidity is %RH, atmospheric pressure is kPa. The wet-bulb temperature calculation module obtains the wet-bulb temperature using the Magnus formula. ℃, saturated vapor partial pressure for kPa, air saturation Calculated as kPa. The water temperature sensor simultaneously measures the circulating water temperature. = ℃, heat transfer mass transfer coefficient K is Then the evaporation potential = = Unit values ​​are integrated and input into the water-saving priority strategy model by the vector of environmental driving factors. S1.4: Based on the fan operation status monitoring module, the fan start-stop status and speed feedback signals are obtained to identify the current forced ventilation intensity of the cooling tower, which serves as a reference for the adjustment of the compensation strategy under the energy-saving priority mode. S1.5: Retrieve the historical operation record database of the water replenishment system, extract the water replenishment frequency and total water replenishment within a unit time to form historical water replenishment behavior characteristics, which are used to optimize the compensation water volume prediction benchmark under the stability priority mode.

[0014] Step S2: Normalize and standardize the acquired real-time environmental parameters to eliminate the interference of different units on the model training process and improve the convergence efficiency of subsequent reinforcement learning algorithms. Specifically, this includes: S2.1: Perform missing value detection and outlier removal on raw environmental parameters such as cooling tower inlet / outlet water temperature difference, circulating water volume, ambient wet bulb temperature, fan operating status, and historical water replenishment records to obtain a complete and reliable initial dataset. For the raw environmental parameter data acquired through the cooling tower sensor network, including the temperature difference between inlet and outlet water, circulating water volume, wet bulb temperature, fan operating status and historical water replenishment records, a missing value detection algorithm (parameter: the detection threshold is set to the maximum allowable interval time within the sampling period of each type of data) is used to identify the breakpoints or unsampled intervals in the acquired signals. Furthermore, by using a time series integrity assessment method (parameters: sliding window length is 60 seconds, and samples with less than 95% sampling within the window are considered missing), the sampling density and continuity of each parameter within adjacent time windows are quantitatively assessed, and a missing marker matrix is ​​obtained. Furthermore, a bidirectional interpolation imputation strategy is adopted (parameters: a combination of linear interpolation and spline interpolation, with interpolation selection based on the difference between adjacent sampling points) to reconstruct missing data segments and generate a preliminary data vector after imputation. Furthermore, by using an outlier detection method (parameters: 3σ principle, where σ is the sample standard deviation; the detection range is calculated independently for temperature difference, flow rate, rotation speed, and wet-bulb temperature respectively), outlier sampling points in the preliminary data vector that deviate from the current statistical mean by more than three times the standard deviation are identified, and an outlier index set is obtained. Furthermore, outlier removal and neighbor-value substitution methods (parameters: outliers within the index range are removed, and the replacement value is taken from the arithmetic mean of the two neighboring points before and after the removed point) are used to achieve smooth correction of outlier data and generate a complete and reliable initial environmental parameter dataset. By using a chain of methods including missing value detection, interpolation reconstruction, and outlier removal, the original environmental parameter signal output from the previous step S1 is transformed into a complete initial dataset that adapts to the standardization process, thereby improving the convergence efficiency of subsequent reinforcement learning algorithms under clean data conditions. For example, in a cooling tower with a circulating water volume of 6000 m³ / h, the inlet water temperature sensor showed two sampling breaks between 14:02 and 14:03 on August 5, 2024, and the circulating water volume sensor generated one out-of-threshold anomaly at 14:05. Missing value detection results showed that the temperature difference data sampling density was 92%, lower than the set 95% threshold, and the corresponding position in the missing value marker matrix was 1. Linear interpolation was used to reconstruct the missing temperature difference value at 14:02, which was 5.8℃, and spline interpolation was used to reconstruct the temperature difference value at 14:03, which was 5.9℃. In the outlier detection, the current average circulating water volume was 6000 m³ / h, with a standard deviation of 30 m³ / h. The detected sampling value at 14:05 was 6200 m³ / h, deviating from the average. Greater than = The value was identified as an outlier. This outlier was removed and replaced with the average of its nearest neighbors, resulting in a corrected circulating water volume of 5995 m³ / h. After performing the above chained processing, an initial dataset containing no missing values ​​and no outliers was generated, including temperature difference, circulating water volume, wet-bulb temperature, fan status, and water replenishment records, meeting the input requirements for subsequent Z-score standardization. S2.2: Based on the Z-score standardization method, the environmental parameters of each dimension are standardized to eliminate the dimensional differences between temperature, flow rate and state signals by normalizing the mean to zero and the variance to one, and output the standardized environmental parameter vector. S2.3: The minimum-maximum normalization method is used to map discrete variables such as wind turbine operating status and historical water replenishment records to the [0,1] interval, so as to unify their numerical expression form and form a normalized state vector compatible with continuous variables. S2.4: Perform feature scaling and dimension alignment operations on the normalized multidimensional environment parameter vector to adapt to the input layer dimension requirements of the high-level policy network and generate a standardized input tensor. S2.5: The standardized input tensor is cached in the runtime data pool, and Gaussian perturbation samples are introduced through data augmentation mechanism to improve the sample diversity and model generalization ability in the subsequent reinforcement learning training stage.

[0015] Step S3: The standardized environmental parameters are input into the high-level policy network, and combined with the system operation target weight configuration file, a multi-objective priority vector under the current operating condition is generated. This multi-objective priority vector includes water-saving priority, energy-saving priority, or stability priority modes. Figure 2 As shown, it specifically includes: S3.1: Perform feature encoding on the standardized environmental parameters of the cooling tower, and use a fully connected neural network to perform nonlinear mapping on the input feature vector to extract abstract feature representations suitable for high-level strategy decision-making, and output feature embedding vectors; S3.2: Based on the feature embedding vector, combined with the preset priority distribution parameters in the system running target weight configuration file, the target weight mapping operation is performed using the Softmax normalization function to generate an initial multi-target priority distribution vector and output the preliminary weight allocation coefficients for the three types of targets: water saving, energy saving, and stability. Based on the high-dimensional abstract feature representation obtained from the feature embedding vector, a weighted mapping initialization method is adopted (parameter: preset priority distribution parameter in the system running target weight configuration file) to assign initial weights to the correlation strength between each feature dimension and the three types of targets: water saving, energy saving, and stability. Furthermore, by using the Softmax normalization function (parameters: feature embedding vector and initial weight assignment matrix), the original weights of each target are mapped to a probability distribution form, and a set of weight coefficients that satisfy the normalization constraint is obtained, where each coefficient is in the interval [0,1] and the sum of the three is 1; Furthermore, the temperature is adjusted using a parameter-based method (parameter: Softmax function temperature). This allows for control over the smoothness of the weight mapping process. In smaller cases, increase the weighting of higher-scoring objectives. In larger cases, the weight distribution of each objective is balanced to generate an initial multi-objective priority distribution vector; Furthermore, by utilizing matrix merging (parameters: preliminary weight coefficients for water conservation, energy conservation, and stability objectives), the multi-objective weight set is integrated into a triplet structure. , which serves as the basic input for subsequent dynamic modulation enhancement operations; Furthermore, a numerical precision correction algorithm is adopted (parameter: floating-point operation precision threshold). This allows for rounding and boundary correction of the initial weighting coefficients, ensuring that the numerical calculations meet the following requirements. Constraints; By combining mapping normalization with temperature regulation, the Softmax algorithm transforms the feature embedding vector into preliminary weight allocation coefficients for three objectives: water saving, energy saving, and stability. This enables the high-level policy network to reasonably initialize the priority of multiple objectives under the current operating conditions. For example, in an industrial circulating cooling system scenario, the feature embedding vector is [1.25, 0.85, 0.35], and the priority distribution parameter matrix preset in the system operation target weight configuration file is [[0.9, 0.5, 0.3], [0.7, 0.6, 0.4], [0.5, 0.7, 0.6]]. Through a weighted mapping initialization method, the score vector is obtained as [1.565, 1.24, 0.83]. The Softmax temperature is then set. =0.7, calculate the normalized weights of each objective: water conservation objective weights =0.47, weight of energy saving target =0.35, stability target weight =0.18. The weight set is integrated into... The triplet is used, and its sum is guaranteed to be 1 through a precision correction algorithm. Under this condition, the initial multi-objective priority distribution vector strengthens the weight of water-saving and energy-saving objectives, which helps to form a reward function structure in the underlying action network that is closer to the current environmental conditions, and realizes the stable operation of the compensation strategy while taking into account both water-saving and energy-saving. S3.3: Perform dynamic modulation enhancement operation on the initial multi-objective priority distribution vector, introduce a gating mechanism to nonlinearly adjust the weights of each objective, so as to enhance the response capability of the high-level policy network to changes in complex operating environment, and output the modulated multi-objective priority vector. S3.4: Context signal encoding operation is performed based on the modulated multi-objective priority vector. Vector concatenation technology is used to encode it into a context modulation signal, which is used as the input control parameter of the underlying action network and outputs a context control vector for dynamically adjusting the reward function structure. S3.5: Transmit the context control vector to the target competitiveness evaluation unit as a reference for the subsequent execution of the weight rebalancing mechanism, so as to ensure the balance and synergy among the targets in the multi-target collaborative optimization process, and output the priority evaluation factor for triggering the weight rebalancing mechanism.

[0016] Step S4: Based on the multi-objective priority vector as a context modulation signal, dynamically adjust the reward function structure of the underlying action network. For example... Figure 3 As shown, it specifically includes: S4.1: Based on the water-saving priority component in the multi-objective priority vector, the water-saving target sub-reward function is weighted to enhance the guiding role of water resource conservation targets in strategy learning under the current operating conditions; Water-saving priority components are extracted from the multi-objective priority vector output by the high-level policy network. As a sub-reward function for water conservation goals The core parameter input for weighted processing; Parameter binding method is adopted (parameter: energy saving priority component) , The original function coefficient set is used to establish weight control channels for various factors related to water resource consumption in the reward function. Furthermore, through a proportional coefficient mapping algorithm (parameter energy-saving priority component: Weight mapping matrix This allows for mapping water-saving priority components to the corresponding factor weights within the reward function, resulting in environmental state-reward response gain curve data. Furthermore, a normalized weighted algorithm is used (parameters: gain curve data, target range normalization factor). This process adjusts the mapped weight values ​​to the target range to prevent gradient instability in the reward function caused by high priority, and generates standardized weighting coefficients. ; Furthermore, a dynamic response modulation method is adopted (parameter: weighting coefficients). Historical Water Conservation Optimization Performance Indicator Set This function combines historical performance to perform differential correction on the current weighting coefficients, and outputs a time-adaptive water-saving weighting function expression. Through the above weighting and modulation processing, the water-saving priority component of the previous step is transformed into weighted sub-reward function data that adapts to the current environmental state, thereby realizing the guiding and enhancing effect of reinforcement learning strategy on water conservation goals. For example, in the operation scenario of an industrial cooling tower with a rated circulating water volume of 500 m³ / h, the multi-objective priority vector output by the high-level policy network is [α=0.65, β=0.20, γ=0.15], and the original definition of the water-saving objective sub-reward function is:

[0017] in, This represents the actual water replenishment volume (m³ / h). The target water replenishment volume is 450 m³ / h. The maximum allowable water replenishment volume is 500 m³ / h. After mapping using a proportional coefficient, the weighting coefficient α' = 0.65 is obtained. After normalization, it is maintained in the range of [0.6, 0.7]. Combining the water-saving optimization performance indicators in historical operation (average evaporation loss reduction rate of 15%), α' is corrected to 0.68 in the dynamic response modulation method. At this time, the weighted water-saving sub-reward function is:

[0018] The calculation result is This value has a strong water conservation orientation for the policy update of the action network during the reinforcement learning training phase, while maintaining the stability of the system temperature control. S4.2: Based on the energy-saving priority component in the multi-objective priority vector, the energy-saving target sub-reward function is weighted to improve the system's response sensitivity to energy consumption control under the current operating state; Based on the energy-saving priority component β in the multi-objective priority vector, the input objects are the energy-saving component values ​​in the context control vector output by the high-level policy network and the predefined energy-saving target sub-reward function in the low-level action network. ; A weighted multiplication method is used for processing (parameter: energy-saving priority component). , This enables the application of energy-saving priority components to the energy-saving target sub-reward function, allowing for numerical amplification or reduction of the values. Furthermore, through a weighted modulation algorithm (parameter: energy-saving priority component) Energy consumption benchmark value Real-time energy consumption monitoring value ), to achieve The internal energy efficiency deviation term is adjusted by coefficient and the first stage calculation result of the energy-saving weighted reward value is obtained; Furthermore, by using a normalization calibration method (parameters: energy-saving weighted reward value, target space normalization range [0,1]), the reward value is dimensionally aligned in the multi-objective reward function space, and the energy-saving weighted second-stage reward signal is generated; Furthermore, an exponential smoothing filter method (parameters: smoothing coefficient λ, energy-saving weighted reward sequence) is adopted to suppress reward fluctuations and generate a stable energy-saving target weighted reward output value, which serves as an input indicator for the energy control response sensitivity of the underlying action network. The weighted calculation of the energy-saving target sub-reward function is performed using the following formula:

[0019] in, The weighted sub-reward value for energy-saving targets. Energy-saving priority components generated for high-level policy networks. This represents the sub-reward value for the original energy-saving target; By using weighted and normalized processing, the energy-saving priority parameters from the previous step are transformed into executable energy control adjustment indicators, thereby improving the response sensitivity of the underlying action network to energy consumption control under the current operating conditions. For example, in the operation scenario of a cooling tower in a coal-fired power plant, the energy-saving priority component β is generated in real time by a high-level policy network, and its current value is... Original energy-saving target sub-reward value for The first-stage reward value is obtained by weighted multiplication. After normalization calibration to the [0,1] range, it still maintains Perform exponential smoothing filtering on this value, with the smoothing coefficient λ set to... Combined with the smoothed output value from the previous moment The new smoothed reward output value is calculated. The output signal is input to the underlying motion network, which significantly improves the sensitivity of energy consumption dynamic control through its compensation strategy during the current operating cycle. S4.3: Based on the stability priority component in the multi-objective priority vector, the sub-reward function of the system stability objective is weighted to enhance the system's anti-interference capability and operational stability under dynamic disturbance conditions. Based on the stable priority component in the multi-objective priority vector A weight mapping algorithm is adopted (parameters source: modulated priority vector output by the high-level policy network). ), to achieve the sub-reward function for the system stability objective. Initial weight settings; Furthermore, by using the disturbance sensitivity analysis method (parameters: historical operating state matrix, current load fluctuation intensity index), the basic sensitivity coefficient of the stability sub-reward function is quantitatively evaluated, and the stability benchmark weight correction value under dynamic disturbance conditions is obtained. Furthermore, a weighted fusion calculation method is adopted (parameter: stable priority component). (Stability benchmark weight correction value), to generate the corrected stability weighting coefficients, and generate the coefficients used for weighting the reward function. The correction vector is calculated using the following formula:

[0020] in, For stable priority components output by the high-level policy network, The disturbance sensitivity correction coefficient is calculated from the ratio of the difference between the real-time disturbance intensity and the historical stability index. Furthermore, through a normalization process (parameters: Correction vector, water-saving priority component Energy saving priority components After the correction is implemented and , The weights are constrained to ensure that the sum of the weights for all objectives equals 1, as shown in the following formula:

[0021] Furthermore, the weighted stability sub-reward function is calculated using the stability value function calculation method (parameters: current system state vector s, action vector a), as shown in the following formula:

[0022] in, The value assessment function representing the stability objective is obtained by integrating the deviation between the system state change and the control action response. By combining the above-mentioned weight mapping, perturbation sensitivity correction, normalization constraint and value function calculation, the stable priority component γ obtained in the previous step is transformed into a perturbation adaptive weighted stability reward value, thereby improving the anti-interference ability and operational stability of the underlying action network under dynamic perturbation conditions. For example, in a certain industrial circulating cooling system, the stability priority component The value is 0.35, indicating a water-saving priority. The value is 0.4, indicating energy-saving priority. The value is 0.25, and the disturbance sensitivity analysis yields a correction coefficient K = 0.1. The corrected value is then calculated using the formula. The values ​​are as follows:

[0023] Calculated = 0.385. A new constraint is introduced in the weight normalization constraint. , , The components are adjusted to ensure that the sum of the three is 1. = 0.395, = 0.22. (The rest of the text appears to be a series of numbers and symbols, possibly representing a corrupted or in Introduced as a weighting factor into the stability value function calculation, the current state vector *s* represents a temperature gradient change rate of 0.05 K / min, a liquid level fluctuation trend amplitude of 0.02 m, and an action vector *a* corresponding to a water replenishment command increment of 2.5 m³ / h. The stability value function *V(s,a)*, obtained through integration, is 0.85. This value is then compared with... Multiplying these values ​​yields a weighted stability reward of 0.32725. This weighted value is used as the input to the stability sub-reward signal during policy training iterations. S4.4: The weighted water-saving, energy-saving and stability target sub-reward functions are combined by linear weighting to generate a comprehensive meta-reward function, so as to form a unified evaluation index for multi-objective collaborative optimization; S4.5: Normalize the meta-reward function to eliminate the impact of differences in reward dimensions for different objectives on the policy optimization process, generate a normalized meta-reward function that adapts to the multi-objective space optimization requirements of reinforcement learning algorithms, and ensure the convergence stability and policy exploration efficiency of reinforcement learning algorithms in multi-objective spaces.

[0024] Step S5: A deep deterministic policy gradient algorithm is employed, using a multivariate input state space including the rate of change of temperature gradient, liquid level fluctuation trend, and load fluctuation intensity as input. Policy network inference is executed to output specific compensation water volume adjustment commands. Specifically, this includes: S5.1: Perform feature selection and dimension alignment processing on the multivariate data in the input state space. The multivariate includes the temperature gradient change rate, liquid level fluctuation trend and load fluctuation intensity, in order to obtain a standardized state vector with a unified dimension, which serves as the input condition for the deep deterministic policy gradient algorithm. S5.2: Input the standardized state vector into the action network of the deep deterministic policy gradient algorithm, and perform the estimation calculation of the action value function Q(s,a|θ) based on the reward function structure modulated by the current multi-objective priority vector to generate the optimal action policy in the current state; S5.3: The action network parameter θ is iteratively optimized through the policy gradient update mechanism. Combined with the feedback signal of the meta-reward function, the compensation water volume adjustment command of the action output is adjusted to improve the synergistic ability of the policy in water saving, energy saving and stability objectives. The input state vector consists of the standardized temperature gradient change rate, liquid level fluctuation trend, and load fluctuation intensity, combined with a multi-objective priority vector generated by a high-level policy network. The reward function structure is parameterized for control. Policy gradient update mechanism using Deep Deterministic Policy Gradient (DDPG) algorithm (parameters: learning rate η, discount factor) Batch size B), for the current action network parameters Gradient calculation is performed, and the action value function output by the Critic network is chained with the action policy predicted by the Actor network to achieve directional adjustment of parameters. Furthermore, the gradient correction is calculated using a method based on the mean squared error (MSE) objective function (input: predicted Q-value and target Q-value). and the meta-reward function The weighted feedback signal is introduced into the gradient update formula to achieve dynamic modulation of the parameter correction magnitude by different target weights, as shown in the following formula:

[0025] in, For integration The subsequent strategy performance index function, The learning rate; Furthermore, by using an adaptive moment estimation method based on the Adam optimizer (parameters: first-order moment decay rate β1, second-order moment decay rate β2), the gradient correction amount of the current iteration is smoothed by momentum and normalized by second-order variance to generate a stable parameter update step size, so as to reduce the impact of large gradient fluctuations on policy convergence. Furthermore, the gradient results of group B samples are averaged using a batch update mechanism, introducing... The degree of difference among the sub-reward components is used as an update weight adjustment factor to form a set of updated strategy parameters that balances water conservation, energy conservation and stability objectives; By updating the chain using the above parameters, the policy gradient calculation results are transformed into a set of parameters. Targeted optimization enhances the ability of the action network to output compensation water volume commands under multi-objective collaborative optimization constraints. For example, in the operating scenario of a cooling tower for a 300MW thermal power unit, the temperature gradient change rate is 0.15℃ / min, the standard deviation of the liquid level fluctuation trend is 0.03m, and the load fluctuation intensity is 85%. The multi-objective priority vector is set as α=0.5, β=0.3, and γ=0.2. =0.8, =0.6, =0.7, generated = =0.71. The DDPG algorithm was used with a learning rate η=0.001, batch size B=64, and Adam optimizer parameters β1=0.9 and β2=0.999. The Q-value bias MSE=0.005 was calculated using the Critic network to generate the gradient correction. =4.7×10^-4, introducing After adjusting the gain, the step size is 4.7 × 10^-4 × 0.71 ≈ 3.34 × 10^-4, and it converges to the stable interval of the strategy after 5 rounds of iteration. The output compensation water volume command is 12.3 m³ of water replenishment per minute; S5.4: Based on the experience playback mechanism, the historical policy execution data of the action network is sampled and trained, and the target network is used to make a stable estimate of the action value function, so as to enhance the robustness and generalization ability of the policy network in complex operating environments. S5.5: Generate the final compensation water volume adjustment command. The command is a continuous control signal that indicates the amount of water to be replenished per unit time. It serves as the control input for the dynamic compensation module to achieve closed-loop dynamic compensation control of the cooling tower evaporation loss.

[0026] Step S6: Construct a target competitiveness evaluation unit, calculate the Pareto front distance between each target, and if it is detected that a target continuously suppresses other targets beyond a set threshold, trigger a weight rebalancing mechanism to forcibly increase the short-term weight gain of the suppressed target. Specifically, this includes: S6.1: Normalize the current optimization intensity of the three types of objectives—water saving, energy saving, and system stability—output by the multi-objective priority vector to obtain a standardized objective optimization intensity vector, which serves as the input condition for objective competitiveness evaluation. S6.2: Construct a Pareto front reference set based on historical data of multi-objective optimization, and use the Euclidean distance algorithm to calculate the distance between the current objective optimization intensity vector and the Pareto front to quantify the non-dominance of the current objective combination and obtain the target Pareto front distance index; S6.3: Perform sliding window time series analysis on the Pareto front distance index of the target, and use the exponential weighted moving average algorithm to extract the competitive trend characteristics of the target in order to identify whether there is a trend evolution in which a certain target continuously suppresses other targets; S6.4: Compare the target competition trend characteristics with the preset suppression threshold. If the duration and intensity of a target continuously suppressing other targets both exceed the set threshold, it is determined to be a multi-target optimization imbalance state, triggering the weight rebalancing mechanism start signal. The target competition trend characteristics obtained from sliding window time series analysis are analyzed using a threshold comparison algorithm (parameter: suppression threshold strength). Time length threshold This enables the determination of imbalances in multi-objective optimization. Furthermore, a two-condition logic decision method is used (parameter: suppression threshold strength). Time length threshold This enables independent matching of the intensity and time components of the target competitive trend characteristics, and yields the suppression state determination matrix. Furthermore, through a conditional joint triggering mechanism (parameter: suppression state judgment matrix), the logical intersection operation of multi-target suppression intensity and suppression duration period is realized, and a suppression compliance index is generated. Furthermore, a comparison operation function is used to compare the suppression compliance index with the joint threshold. Perform a comparison and generate imbalance judgment labels; Furthermore, through the state trigger interface, the imbalance judgment label is transformed into a weight rebalancing mechanism start signal, achieving seamless connection with subsequent weight adjustment function modules; By using a judgment algorithm based on dual-condition threshold comparison and logical combination, the competitive trend features extracted in the previous step are transformed into detection signals of multi-objective optimization imbalance, thereby achieving accurate triggering of the automatic identification and rebalancing process of the suppression target. For example, in an industrial circulating cooling system, the competitive trend characteristic intensity component of the water-saving target is 0.85 (dimensionless normalized value), the time component is 72h, and the preset intensity threshold is... The time length threshold is 0.8. The time frame is 48 hours. A dual-condition logic method is used to compare the intensity and time components separately using thresholds. Both the intensity and time components are evaluated as True. A suppression compliance index of True is generated through a logical AND operation, indicating that the target meets the suppression condition. (Joint threshold) The comparison result matches successfully, generating an imbalance judgment label "True". This label is converted into a weight rebalancing mechanism activation signal via the state trigger interface, driving the subsequent dynamic weight allocation algorithm to recalculate the meta-reward function parameters α, β, and γ. After this processing, the short-term weight of the water-saving target is increased to 0.5, and the energy-saving and stability targets are adjusted to 0.25 respectively, thereby restoring the multi-objective balance in subsequent strategy optimization; S6.5: Based on the weight rebalancing mechanism, a signal is initiated to forcibly increase the short-term weight gain of the suppressed target. The α, β, and γ coefficients in the meta-reward function are recalculated using a dynamic weight allocation algorithm to restore the influence of the suppressed target in subsequent policy optimization and ensure the long-term stability of multi-objective collaborative optimization.

[0027] Step S7: Introducing a cross-operating-condition transfer training strategy, using historical operating data from multiple power plants and climate zones for pre-training during the offline training phase to enhance the model's generalization ability under different operating conditions. Specifically, this includes: S7.1: Collect and integrate historical operating data from multiple power plants and various climate regions to construct a cross-condition dataset, which serves as the input data source for transfer training; S7.2: Based on the constructed cross-working condition dataset, perform initial training on the high-level policy network and the low-level action network respectively to obtain an initial policy model with preliminary multi-working condition adaptability. S7.3: Adopt the domain adaptive algorithm to align the feature distribution during the training process, so as to reduce the data distribution differences between different power plants and climate conditions and improve the generalization performance of the model under unknown operating conditions. S7.4: Introducing a multi-task learning framework, the operational objectives of different power plants are jointly optimized as independent tasks during the pre-training stage to enhance the model's ability to respond to diverse operational objectives. S7.5: Generate a set of general policy parameters for cross-operating conditions based on the pre-trained policy model, which serves as the initial model parameters for online fine-tuning during the system deployment phase, thereby improving model convergence efficiency and operational stability.

[0028] Step S8: During system operation, an online fine-tuning mechanism is executed to update the local parameters of the pre-trained model based on real-time feedback data from the current project, thereby improving the model's adaptability to specific on-site characteristics. Specifically, this includes: S8.1: Collect and cache real-time feedback data from the cooling tower operation site. The real-time feedback data includes the actual evaporation loss after compensation execution, water replenishment response time, water level recovery curve, and energy efficiency change index, as the input sample set for the online fine-tuning mechanism. S8.2: Perform feature extraction and normalization on the collected real-time feedback data, and calculate the temporal statistical features of state transition based on the sliding window mechanism to obtain standardized online training samples for subsequent incremental learning of the model. S8.3: Based on the extracted standardized samples, a local parameter update strategy is executed. An online incremental learning algorithm is used to optimize the gradient backpropagation of the Actor-Critic structure of the underlying action network to obtain a strategy parameter update vector that adapts to the current working conditions. S8.4: Perform model performance evaluation on the updated strategy parameters. Calculate the stability factor of the strategy update based on the deviation index between the current compensation command and the actual response result to determine whether to accept the parameter update. S8.5: Based on the stability factor evaluation results, if the policy update meets the convergence threshold and does not cause system fluctuations, the updated policy parameters are written to the runtime model cache to replace the original policy network parameters, so as to realize the closed-loop iterative update of the online fine-tuning mechanism.

[0029] Step S9: The generated compensation water volume adjustment command is transmitted to the dynamic compensation module via the API interface to ensure that the compensation command has sufficient responsiveness and robustness to achieve the dynamic adjustment function of the closed-loop control system. Specifically, this includes: S9.1: Based on the compensation water volume adjustment command output by the water volume prediction module, the command data format is standardized and encapsulated to adapt to the input interface protocol requirements of the dynamic compensation module. S9.2: Construct a RESTful API communication interface, using JSON format as the data transmission carrier, and serialize the encapsulated compensation water volume adjustment instructions to improve the compatibility and parsability of cross-module data interaction. S9.3: Introduce a message queue mechanism in the API interface to asynchronously transmit and schedule the serialized compensation instructions. Perform instruction queuing and priority sorting based on the RabbitMQ protocol to improve the system's responsiveness in high-concurrency scenarios. S9.4: Deploy an instruction verification unit on the dynamic compensation module side to verify the integrity and rationality of the received compensation water volume adjustment instructions, and detect data transmission errors based on the CRC check algorithm to enhance the robustness of instruction execution; S9.5: If the verification passes, the compensation water volume adjustment command will be converted into a control signal and input to the actuator. Based on the PLC controller, the valve opening adjustment or water supply pump start / stop control will be executed to realize the dynamic compensation adjustment function of the closed-loop control system for evaporation loss. Based on the dynamic compensation module, the verification unit determines that the input conditions for the compensation water volume adjustment command are complete and reasonable. The PLC controller instruction mapping method (parameters: instruction content, target actuator type, execution mode identifier) ​​is adopted to convert the logical instructions predicted by the high-level module into a control data format that can be recognized by the low-level hardware. Furthermore, by executing a parameter decomposition algorithm (parameters: target valve number, target pump number, adjustment value, execution duration), the control target and control amplitude of the compensation water volume adjustment command are decomposed into fine-grained components, and the independent control components of each actuator are obtained. Furthermore, an analog-to-analog conversion method (parameters: adjustment value, PLC output module range, control signal type) is adopted to map the adjustment component to the corresponding voltage or current signal amplitude and generate a continuous analog control signal matrix to drive the actuator to achieve precise action; Furthermore, by applying the pulse width modulation (PWM) generation method (parameters: target action duration, control accuracy requirements), the execution components that require intermittent driving are converted into PWM pulse sequences, generating a control signal structure that meets the timing requirements of the PLC digital output port, ensuring that the actuator switches on and off at a specified rhythm. Furthermore, by combining the actuator feedback status acquisition method (parameters: valve opening sensor output, pump motor speed sensor output), the actual action effect and the expected action command can be compared in real time, and an operation deviation index can be generated for the system to make strategy corrections for the next action cycle in closed-loop control. Through the above PLC controller mapping and signal generation processing method, the verified compensation water volume adjustment command is converted into a control signal that can directly drive the field hardware actuator, so as to realize the dynamic compensation adjustment target of the closed-loop control system for evaporation loss. For example, in an industrial circulating cooling system, the verified compensation water volume adjustment command includes a valve opening adjustment value of 45% and a water pump start-stop cycle of 10 seconds on / 20 seconds off. In this step, the PLC output module range is selected as 4~20mA, and the parameter decomposition algorithm is executed to map the 45% opening to... mA analog output, i.e. A mA current signal drives the electric valve to a specified opening. The start-stop cycle of the water supply pump is achieved using a pulse width modulation (PWM) method, with a duty cycle of [value missing]. = The corresponding PWM pulse high-level duration is 10 seconds, and the low-level duration is 20 seconds. During execution, the valve opening sensor feedback value is 44.8%, and the pump speed sensor feedback shows that the rated speed of 2900 rpm is maintained during the high-level phase. The operating deviation indicators are within the allowable range, and the closed-loop control enters the next cycle. This embodiment fully verifies the accuracy of control signal generation and mapping processing, ensuring the response speed and operational stability of the compensation water volume adjustment.

[0030] Step S10: Periodically evaluate the long-term operational performance feedback of the system, dynamically evolve the weight coefficients of each objective in the reward function, form a closed-loop adaptive optimization mechanism, prevent strategy collapse, and improve the overall system efficiency. Specifically, this includes: S10.1: Based on the long-term operational performance feedback data of the system, the execution effect of water saving, energy saving and stable operation objectives in the multi-objective collaborative optimization mechanism is quantitatively evaluated to obtain the long-term contribution index sequence of each objective; S10.2: The sliding time window statistical method is used to perform dynamic trend analysis on the long-term contribution index sequence, and the fluctuation characteristics of each target in different operating cycles are extracted to generate a target priority evolution trend vector. S10.3: Input the target priority evolution trend vector into the weight evolution model, and dynamically adjust the coefficients of α, β, and γ in the current reward function based on the improved exponential smoothing algorithm to obtain the updated multi-objective weight allocation scheme; S10.4: Based on the updated multi-objective weight allocation scheme, the meta-reward function structure in the high-level policy network is reconstructed to generate a dynamic reward function expression that adapts to the current running state; S10.5: The reconstructed dynamic reward function expression is fed back to the underlying action network to drive the deep deterministic policy gradient algorithm to update the compensation policy output, so as to achieve continuous adaptive optimization and stability assurance of the compensation policy under closed-loop control.

[0031] The technical solution of the present invention has been described above with reference to the preferred embodiments shown in the accompanying drawings. However, it will be readily understood by those skilled in the art that the scope of protection of the present invention is obviously not limited to these specific embodiments. Without departing from the principles of the present invention, those skilled in the art can make equivalent changes or substitutions to the relevant technical features, and the technical solutions after these changes or substitutions will all fall within the scope of protection of the present invention.

[0032] The above description is merely a preferred embodiment of the present invention and is not intended to limit the invention. Various modifications and variations can be made to the present invention by those skilled in the art. Any modifications, equivalent substitutions, improvements, etc., made within the spirit and rules of the present invention should be included within the scope of protection of the present invention.

Claims

1. A method for dynamic compensation of evaporation loss in a cooling tower, characterized in that, Includes the following steps: S1: Obtain real-time environmental parameters during the operation of the cooling tower, including the cooling tower inlet / outlet water temperature difference, circulating water volume, ambient wet-bulb temperature, fan operating status, and historical water replenishment records; S2: Normalize and standardize the real-time environmental parameters to generate standardized environmental parameters; S3: Input the standardized environmental parameters into the high-level strategy network, and combine them with the system operation target weight configuration file to generate a multi-objective priority vector under the current operating conditions. The multi-objective priority vector includes water-saving priority, energy-saving priority, or stability priority modes. S4: Based on the multi-objective priority vector as the context modulation signal, dynamically adjust the reward function structure of the underlying action network; S5: Employs a deep deterministic policy gradient algorithm, taking a multivariable input state space as input, executing policy network inference, and outputting specific compensation water volume adjustment instructions; S6: Construct a target competitiveness evaluation unit, calculate the Pareto front distance between each target, and if it is detected that a target continuously suppresses other targets for more than a set threshold, trigger the weight rebalancing mechanism. S7: Introduces a cross-condition transfer training strategy, using historical operating data from multiple power plants and climate zones for pre-training during the offline training phase; S8: During operation, an online fine-tuning mechanism is executed to update the local parameters of the pre-trained model based on the real-time feedback data of the current project. S9: Transmit the generated compensation water volume adjustment command to the dynamic compensation module.

2. The method for dynamic compensation of evaporation loss in a cooling tower according to claim 1, characterized in that, Following step S9, the following is also included: S10: Periodically evaluate long-term operational performance feedback, dynamically evolve the weight coefficients of each objective in the reward function, and form a closed-loop adaptive optimization mechanism.

3. The method for dynamic compensation of evaporation loss in a cooling tower according to claim 1, characterized in that, Step S1 specifically includes: Based on an industrial sensor network, the difference between the inlet and outlet water temperatures during the operation of the cooling tower is collected to obtain the rate of change of the temperature gradient. The operating status signal of the circulating water pump is read in real time to obtain the current circulating water volume data and construct the water volume change trend characteristics; The wet-bulb temperature of the air surrounding the cooling tower is obtained by using an ambient temperature and humidity sensor, and the air saturation and evaporation potential are calculated. Based on the fan operation status monitoring module, the start-stop status and speed feedback signals of the fan are obtained, and the current forced ventilation intensity of the cooling tower is identified. The historical operation records of the water replenishment system are retrieved from the database, and the water replenishment frequency and total amount per unit time are extracted to form historical water replenishment behavior characteristics.

4. The method for dynamic compensation of evaporation loss in a cooling tower according to claim 3, characterized in that, Step S1 further includes synchronously acquiring inlet and outlet water temperatures based on an industrial sensor network, using a Pt100 resistance temperature detector (RTD) sensor with Class A accuracy and a sampling frequency of 1 Hz, and aligning the signal time using the IEEE 1588 time synchronization protocol. Data smoothing is performed using differential sampling, sliding window averaging filtering (window length 5 min, overlap rate 50%), and second-order Butterworth low-pass filtering (cutoff frequency 0.005 Hz), and a standardized temperature gradient feature vector is obtained through Z-score normalization.

5. The method for dynamic compensation of evaporation loss in a cooling tower according to claim 1, characterized in that, Step S2 specifically includes: Missing values ​​and outlier removal are performed on the original environmental parameters to obtain a complete and reliable initial dataset; The environmental parameters of each dimension are standardized using a standardization method, and the standardized environmental parameter vector is output. The minimum-maximum normalization method is used to perform a unified interval mapping on discrete variables to form a normalized state vector; Perform feature scaling and dimension alignment operations on the normalized state vector to generate a standardized input tensor; The standardized input tensor is cached in the runtime data pool, and Gaussian perturbation samples are introduced through a data augmentation mechanism.

6. The method for dynamic compensation of evaporation loss in a cooling tower according to claim 1, characterized in that, Step S3 specifically includes: The standardized environmental parameters of the cooling tower are subjected to feature encoding. A fully connected neural network is used to perform nonlinear mapping on the input feature vector to extract abstract feature representations suitable for high-level strategy decision-making and output feature embedding vectors. Based on the feature embedding vector, combined with the preset priority distribution parameters in the system operation target weight configuration file, the target weight mapping operation is performed using the Softmax normalization function to generate an initial multi-target priority distribution vector and output the preliminary weight allocation coefficients for the three types of targets: water saving, energy saving, and stability. A dynamic modulation enhancement operation is performed on the initial multi-target priority distribution vector, and a gating mechanism is introduced to nonlinearly adjust the weights of each target, outputting the modulated multi-target priority vector. Based on the modulated multi-target priority vector, a context signal encoding operation is performed, and the vector concatenation technique is used to encode it into a context modulation signal, and the context control vector is output. The context control vector is transmitted to the target competitiveness evaluation unit, and the priority evaluation factor is output.

7. The method for dynamic compensation of evaporation loss in a cooling tower according to claim 6, characterized in that, Step S3 further includes obtaining an abstract feature embedding vector through feature encoding, combining it with preset weight distribution parameters, allocating initial weights for three types of objectives—water saving, energy saving, and stability—through Softmax normalization mapping, and using parameter temperature adjustment and numerical precision correction algorithms to ensure that the weight allocation meets normalization constraints.

8. The method for dynamic compensation of evaporation loss in a cooling tower according to claim 1, characterized in that, Step S4 specifically includes: Based on the water-saving priority component in the multi-objective priority distribution vector, the sub-reward function for water-saving objectives is weighted. Based on the energy-saving priority components in the multi-objective priority vector, the energy-saving target sub-reward function is weighted. Based on the stable priority component in the multi-objective priority vector, the stability objective sub-reward function is weighted. The weighted water-saving target sub-reward function, energy-saving target sub-reward function and stability target sub-reward function are combined by linear weighting to generate a comprehensive meta-reward function; The meta-reward function is normalized to generate a normalized meta-reward function.

9. The method for dynamic compensation of evaporation loss in a cooling tower according to claim 1, characterized in that, Step S5 specifically includes: Feature selection and dimension alignment are performed on the multivariate data in the input state space to obtain a standardized state vector with a unified dimension. The standardized state vector is input into the action network of the deep deterministic policy gradient algorithm. Based on the reward function structure modulated by the current multi-objective priority vector, the action value function is estimated and calculated to generate the optimal action policy in the current state. The action network parameters are iteratively optimized through a policy gradient update mechanism, and the compensation water volume adjustment command of the action output is adjusted in combination with the feedback signal of the meta-reward function. Based on the experience playback mechanism, the historical policy execution data of the action network is sampled and trained, and the target network is used to make a stable estimate of the action value function. The final compensation water volume adjustment command is generated. The command is a continuous control signal that indicates the amount of water to be replenished per unit time, and serves as the control input for the dynamic compensation module.

10. A method for dynamic compensation of evaporation loss in a cooling tower according to claim 9, characterized in that, The multivariate data includes the rate of change of temperature gradient, the trend of liquid level fluctuation, and the intensity of load fluctuation.

Images