A method for stabilizing the outlet temperature of a condenser circulating water

Abstract

The present application relates to the technical field of intelligent control of thermal equipment, aiming at the problems of low sensitivity and insufficient real-time of traditional fouling detection and energy efficiency regulation caused by condenser structure aging, dirt accumulation and flow channel deflection, a stable control method for the outlet water temperature of condenser circulating water is proposed, the dynamic thermal inertia characteristics reflecting the actual heat transfer performance are obtained through micro-amplitude heat pulse excitation and multi-node temperature response, combined with real-time water quality parameter characteristic vector, the fouling tendency index is generated by migrating industrial boiler fouling knowledge, and multi-objective weight dynamic configuration is implemented accordingly, driving the segmented linear prediction model to output safe and efficient real-time regulation instructions. The scheme can realize adaptive optimization control of energy efficiency and anti-fouling under the fouling risk warning, and improve the safety, energy saving and intelligent response ability of the condenser operation.

Description

Technical Field

[0001] This invention relates to the field of multi-objective optimization of industrial process systems and intelligent control technology of thermal equipment, and in particular to a method for stable control of the outlet temperature of condenser circulating water. Background Technology

[0002] Current mainstream technologies in condenser circulating water system control employ methods based on PID feedback, fuzzy control, multi-model switching, and classical multi-objective optimization. These methods rely on statically weighted multi-objective functions to adjust valve openings or shell-side medium parameters through real-time temperature deviation feedback, supplemented by periodic manual water quality testing to configure chemical dosing procedures. Some high-end systems incorporate genetic algorithms, particle swarm optimization, and offline global optimization models based on operating condition clustering. The system is trending towards an intelligent control architecture that integrates heat exchange performance self-sensing, online fouling monitoring, automatic water quality management, and energy consumption control. However, existing methods set fixed rules for weight allocation in the initial design phase, lacking dynamic and flexible adjustment capabilities. Water quality parameter models rely on intermittent sampling or offline analysis, making it difficult to dynamically quantify equipment aging, micro-channel scaling, and thermal inertia characteristics in real time.

[0003] Existing technologies suffer from the following problems: the multi-objective optimization function is statically fixed, making it difficult to flexibly adjust the weights of objectives such as heat exchange efficiency, scaling inhibition, energy consumption, and equipment lifespan, and failing to achieve dynamic weight migration and real-time game coordination; water quality and scaling risk management rely on offline testing and empirical models, lacking high-speed online sensing and risk pre-intervention capabilities; the means of characterizing equipment aging and thermal inertia changes are limited, failing to provide reliable digital fingerprints of health status; and global optimization methods have high processing complexity, high computational power dependence, and poor real-time performance and on-site adaptability. Summary of the Invention

[0004] This application provides a method for stabilizing the outlet temperature of condenser circulating water, aiming to solve one of the problems or issues of the prior art mentioned in the background section.

[0005] This application provides a method for stabilizing the outlet temperature of condenser circulating water, specifically including:

[0006] S1: Inject a small, non-disturbance step thermal pulse signal into the condenser circulating water inlet side, and simultaneously collect the temperature response time-series curves of the shell-side medium side and the outlet water side to obtain the original thermal response dataset.

[0007] S2: Extract the thermal time constant distribution spectrum of the condenser based on the original thermal response dataset, and generate a fouling feature vector by combining it with the newly established condenser reference spectrum. Then, fuse the thermal time constant distribution spectrum and the fouling feature vector into a dynamic thermal inertial fingerprint database.

[0008] S3: Real-time acquisition of water quality parameters of condenser circulating water, and construction of real-time water quality feature vector;

[0009] S4: Input the real-time water quality feature vector into the transfer learning model to reuse the corrosion-deposition coupling mode features of the industrial boiler scaling database, and calculate and generate the scaling tendency index.

[0010] S5: Based on the matching degree of the dynamic thermal inertia fingerprint database, the scaling tendency index, the heat exchange efficiency decay rate of the previous cycle, and the peak and valley electricity price time period identifier, perform multi-objective weight dynamic migration and output the real-time weight coefficient group corresponding to multiple control objectives.

[0011] S6: Input the real-time weight coefficient group into a preset improved piecewise linear predictive controller to generate predicted temperature trajectory and safety boundary envelope data for future time periods;

[0012] S7: Determine the positional relationship of the actual outlet water temperature relative to the safety boundary envelope data and the changing trend of the scaling tendency index. When the preset triggering conditions are met, trigger the feedforward compensation mode to generate a large flow rate and low inlet temperature setting command; otherwise, trigger the slow-change adjustment mode to generate a fine-tuning valve opening slope command.

[0013] S8: Send the feedforward compensation mode command or the gradual adjustment mode command to the circulating water valve actuator and the shell-side medium temperature control unit for temperature control.

[0014] The method for stabilizing the outlet temperature of condenser circulating water provided in this application has the following beneficial effects:

[0015] This solution constructs a dynamic characteristic sensing mechanism based on a thermal inertial fingerprint database, combined with a scaling tendency index model, to achieve hourly online identification and trend prediction of the internal structural state and scaling risk of the condenser. It utilizes micro-amplitude thermal pulse excitation to extract the intrinsic thermal response characteristics of the equipment, identifying heat exchange performance degradation trends in advance and triggering feedforward intervention. A multi-objective weighted dynamic migration controller and an improved piecewise linear predictive control architecture are introduced, integrating thermal inertial matching degree, scaling probability index, efficiency decay rate, and electricity price signals. Parallel predictive sub-models generate the natural evolution trajectory of future outlet water temperature and the maximum allowable temperature rise safety envelope, respectively. Based on their relative positions, the system intelligently switches between gradual adjustment and feedforward compensation modes to achieve adaptive migration of control priorities. An integrated digital twin verification module constructs a sub-second closed-loop simulation verification mechanism. All control commands undergo dynamic security assessment in a virtual environment before issuance, eliminating the need for a global optimization solver or large-scale historical data training, resulting in low computational consumption and strong engineering deployability. Attached Figure Description

[0016] Figure 1 This is the main flow chart of a method for stabilizing the outlet water temperature of a condenser circulating water system;

[0017] Figure 2 This is a sub-flowchart of a method for stabilizing the outlet water temperature of a condenser circulating water system;

[0018] Figure 3 This is another sub-flowchart of a method for stabilizing the outlet water temperature of a condenser circulating water system. Detailed Implementation

[0019] 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.

[0020] 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.

[0021] like Figure 1 As shown, this application provides a method for stabilizing the outlet temperature of condenser circulating water, specifically including:

[0022] S1: Inject a small, non-disturbance step thermal pulse signal into the condenser circulating water inlet side, and simultaneously collect the temperature response time-series curves of the shell-side medium side and the outlet water side to obtain the original thermal response dataset.

[0023] S2: Extract the thermal time constant distribution spectrum of the condenser based on the original thermal response dataset, and generate a fouling feature vector by combining it with the newly established condenser reference spectrum. Then, fuse the thermal time constant distribution spectrum and the fouling feature vector into a dynamic thermal inertial fingerprint database.

[0024] S3: Real-time acquisition of water quality parameters of condenser circulating water, and construction of real-time water quality feature vector;

[0025] S4: Input the real-time water quality feature vector into the transfer learning model to reuse the corrosion-deposition coupling mode features of the industrial boiler scaling database, and calculate and generate the scaling tendency index.

[0026] S5: Based on the matching degree of the dynamic thermal inertia fingerprint database, the scaling tendency index, the heat exchange efficiency decay rate of the previous cycle, and the peak and valley electricity price time period identifier, perform multi-objective weight dynamic migration and output the real-time weight coefficient group corresponding to multiple control objectives.

[0027] S6: Input the real-time weight coefficient group into a preset improved piecewise linear predictive controller to generate predicted temperature trajectory and safety boundary envelope data for future time periods;

[0028] S7: Determine the positional relationship of the actual outlet water temperature relative to the safety boundary envelope data and the changing trend of the scaling tendency index. When the preset triggering conditions are met, trigger the feedforward compensation mode to generate a large flow rate and low inlet temperature setting command; otherwise, trigger the slow-change adjustment mode to generate a fine-tuning valve opening slope command.

[0029] S8: Send the feedforward compensation mode command or the gradual adjustment mode command to the circulating water valve actuator and the shell-side medium temperature control unit for temperature control.

[0030] Step S1: Inject a small, non-disturbance step thermal pulse signal into the condenser circulating water inlet side, and simultaneously acquire the temperature response time-series curves of the shell-side medium side and the outlet water side to obtain the raw thermal response dataset. Specifically, this includes:

[0031] S1.1: A micro thermal pulse excitation array deployed based on the condenser circulating water inlet pipe applies a micro-amplitude, non-disturbance step thermal pulse signal with a preset amplitude and controllable duration to the condenser circulating water fluid entering the condenser, generating an excited condenser circulating water inlet temperature sequence containing specific thermal disturbance characteristics.

[0032] A micro thermal pulse excitation array is uniformly deployed along the circumference of the inner wall of the condenser circulating water inlet pipe. The array consists of several independent and controllable thin-film resistance heating units. Each heating unit is tightly attached to the pipe wall through an insulating thermal conductive layer to ensure that heat energy is efficiently transferred to the fluid without affecting the integrity of the flow channel structure.

[0033] Based on the current condenser circulating water flow rate and heat capacity parameters, a step heat pulse signal control command with a preset amplitude is generated. This amplitude is set to produce a small temperature rise of 0.5 to 1.5 degrees Celsius in the inlet water temperature, so as to ensure that the disturbance to the subsequent heat exchange process is within the non-intrusive range and avoid causing drastic fluctuations in the system operating conditions.

[0034] A synchronous trigger signal is sent to the micro thermal pulse excitation array to control all heating units to start simultaneously within milliseconds. A micro-amplitude step thermal pulse with a precise and controllable duration is applied. The pulse duration is determined based on the residence time of the fluid in the detection section, and is usually set to 2 to 5 seconds to form a thermal disturbance source with a clear start time.

[0035] Real-time monitoring of the current and voltage feedback data of the heating unit, dynamic adjustment of output power using power closed-loop control, compensation for energy loss deviation caused by differences in tube wall heat capacity, ensuring that the actual heat flux of the injected fluid is consistent with the theoretical set value, thereby generating a stimulated condenser circulating water inlet temperature sequence containing specific thermal disturbance characteristics.

[0036] S1.2: Using a high-precision broadband conductivity sensing unit and a fast-response temperature sensor, the shell-side medium temperature fluctuation data and the outlet temperature fluctuation data under the action of the inlet water temperature sequence of the stimulated condenser are synchronously acquired at the millisecond level to construct a multi-node coupled temperature response raw record set.

[0037] The system receives the stimulated condenser circulating water inlet temperature sequence generated by S1.1 and uses it as a time reference trigger signal to synchronously start the high-precision broadband conductivity sensing unit and the fast-response temperature sensor array.

[0038] Platinum resistance temperature sensors deployed at the inlet and outlet of the shell-side medium and the outlet of the condenser circulating water are used to capture transient temperature fluctuation data in real time during the transmission of excited thermal pulses at a sampling frequency of milliseconds, forming the original temperature time sequence flow.

[0039] The real-time change curve of the conductivity of the condenser circulating water is synchronously acquired from the broadband conductivity sensing unit. This is used to monitor the minute fluctuations in ion mobility caused by slight thermal disturbances and to eliminate the interference of sudden changes in water quality on the temperature response.

[0040] Hardware-level timestamps are applied to the collected shell-side medium temperature fluctuation data and effluent-side temperature fluctuation data to ensure strict alignment of multi-node data on a microsecond-level time scale.

[0041] The aligned multi-source sensor data is structured and encapsulated according to spatial location index to construct a multi-node coupled temperature response raw record set containing influent excitation features, shell-side response features, and effluent response features.

[0042] S1.3: Perform timestamp alignment calibration and high-frequency noise filtering on the original temperature response record set of multi-node coupled devices to eliminate sensor sampling time delay differences and environmental electromagnetic interference, and generate a set of pure temperature response time series curves with strict time consistency.

[0043] Specifically, taking the triggering time of the step thermal pulse signal as the global time reference zero point, the timestamps of each sensor channel in the multi-node coupled temperature response raw record set are synchronously calibrated using a linear interpolation method. The sampling time of each channel is uniformly mapped to a common time axis with the reference zero point as the origin, eliminating microsecond-level timing deviations caused by the independent operation of sampling clocks of different sensors.

[0044] For the time-aligned temperature signals of each channel, a second-order low-pass Butterworth filter with a cutoff frequency of 10Hz is used to perform forward-reverse zero-phase filtering to eliminate high-frequency interference components introduced by inverter switching noise and electromagnetic radiation coupling, while avoiding the introduction of phase delay.

[0045] The filtered signal is further processed by sliding window mid-range filtering, with the window length set to 5 sampling points to remove isolated abnormal peak values ​​caused by occasional sensor drift or signal transmission packet loss, ensuring the smoothness and physical consistency of the temperature response curve.

[0046] The temperature signal sequences of each channel, after being timestamped and double filtered, are reassembled according to the original sensor spatial layout index and encapsulated into a standardized time series curve set containing the shell-side medium temperature response channel and the outlet water temperature response channel, outputting a pure temperature response time series curve set.

[0047] S1.4: Based on the pure temperature response time series curve set, calculate the dynamic temperature difference evolution trajectory of the shell-side medium side and the outlet water side relative to the inlet water temperature sequence of the excited condenser to quantify the transmission delay and attenuation characteristics of the heat pulse in the heat exchange tube bundle and form a dynamic characteristic vector of heat transfer.

[0048] The inlet water temperature sequence of the stimulated condenser circulating water, the shell-side medium temperature fluctuation data, and the outlet water temperature fluctuation data are extracted from the pure temperature response time series curves and used as the calculation benchmark for dynamic temperature difference evolution.

[0049] Using the inlet water temperature sequence of the stimulated condenser circulating water as the time reference axis, point-by-point difference calculations are performed on the shell-side medium temperature fluctuation data and the outlet water temperature fluctuation data to eliminate the interference of ambient background temperature drift on thermal response analysis.

[0050] A shell-side dynamic temperature difference evolution function is constructed to calculate the instantaneous deviation of the shell-side medium temperature relative to the inlet water temperature. This deviation reflects the heat transfer efficiency through the pipe wall into the shell-side medium.

[0051] A dynamic temperature difference evolution function on the tube side is constructed to calculate the instantaneous deviation of the water side temperature relative to the inlet water temperature. This deviation value characterizes the total heat absorption and friction loss of the condenser circulating water as it flows through the tube bundle.

[0052] By employing a sliding time window mechanism, the critical response period after the step thermal pulse is captured, and the peak arrival time in the temperature difference evolution curve is extracted and defined as the heat transfer delay time parameter.

[0053] The decay slope of the temperature difference evolution curve from the initial state to the steady state is calculated, and the decay rate of the thermal pulse energy in the heat exchange process is quantified by the exponential fitting method, which is defined as the thermal decay coefficient.

[0054] By concatenating the heat transfer delay time parameter and the heat attenuation coefficient into a vector, a dynamic feature vector of heat transfer containing both time-domain lag characteristics and amplitude-domain attenuation characteristics is formed.

[0055] S1.5: The dynamic feature vector of heat transfer is fused and encapsulated with the flow and pressure parameters under the current operating conditions in a multidimensional manner to generate a complete original thermal response dataset that characterizes the aging degree of the equipment structure, the state of dirt accumulation, and the flow channel deviation distribution.

[0056] Step S2: Extract the thermal time constant distribution spectrum of the condenser based on the original thermal response dataset, and generate a fouling feature vector by combining it with the newly established condenser reference spectrum. Then, fuse the thermal time constant distribution spectrum and the fouling feature vector into a dynamic thermal inertial fingerprint database. Specifically, this includes:

[0057] S2.1: Denoising, smoothing and time alignment calibration are performed on the shell-side medium-side temperature response time series curves and the outlet-side temperature response time series curves in the original thermal response dataset to eliminate high-frequency noise interference from the sensor and unify the sampling time reference, thereby obtaining a standard thermal pulse response sequence with high signal-to-noise ratio.

[0058] S2.2: Based on the standard thermal pulse response sequence, construct an excitation constraint matrix containing the amplitude and duration of a micro-amplitude non-perturbation step thermal pulse signal, and use the least squares method to perform parameter estimation calculations to separate the set of initial thermal time constants that reflect the aging of the equipment structure and the flow channel deviation state.

[0059] Based on the high signal-to-noise ratio standard thermal pulse response sequence obtained from S2.1, the excitation parameters of the micro-amplitude unperturbed step thermal pulse signal, including pulse amplitude and duration, are extracted.

[0060] Construct an excitation constraint matrix U, and discretize the pulse signal into a time series vector, which serves as the input excitation term for system identification.

[0061] A first-order inertial plus pure time-delay mathematical model of the condenser heat exchange process is established to describe the dynamic transmission relationship from inlet water temperature disturbance to outlet water temperature response.

[0062] Using the least squares method, the loss function is defined as the sum of squares of the differences between the model's predicted output and the actual measured output. The specific formula is as follows:

[0063]

[0064] in, Let be the actual outlet water temperature at the i-th sampling time. The model predicts the temperature, and N is the number of sampling points.

[0065] Take the partial derivative of the loss function J with respect to the parameters to be identified, set the partial derivative to zero, and obtain the optimal solution conditions for the parameter estimation.

[0066] Solve the normal equations to separate the set of initial thermal time constants that reflect the aging of the equipment structure and the flow channel deviation state.

[0067] S2.3: Perform statistical distribution fitting on the initial set of thermal time constants, and use Gaussian mixture model cluster analysis to extract the mean and variance distribution characteristics of the dominant thermal time constants in order to quantify the differences in thermal hysteresis effects in different flow channel regions, thereby generating a thermal time constant distribution spectrum that characterizes the non-uniformity of heat transfer dynamics.

[0068] The system receives the initial set of thermal time constants output from step S2.2. This set contains multiple discrete time constant values, reflecting the thermal response hysteresis characteristics of different flow channel regions in the condenser. Data cleaning is performed on the initial set of thermal time constants to remove outliers caused by transient interference from sensors, ensuring the statistical validity of the input data. A Gaussian mixture model (GMM) is constructed, setting the number of mixture components to correspond to the number of main flow channel partitions in the condenser, and initializing the mean, variance, and weight parameters of each component. The expectation-maximization algorithm is used to iteratively optimize the model parameters, calculating the posterior probability of each time constant sample belonging to each Gaussian component until the log-likelihood function converges. The mean of each converged Gaussian component is extracted as the dominant thermal time constant, characterizing the average thermal inertia level of a specific flow channel region. The variance of each Gaussian component is calculated to quantify the dispersion of the thermal response in that region, reflecting uneven flow velocity distribution or local differences in fouling thickness within the flow channel. The extracted mean and variance are combined into a multidimensional feature vector to form the thermal time constant distribution spectrum.

[0069] S2.4: Based on the thermal time constant distribution spectrum and the preset reference thermal time constant distribution spectrum of the newly built condenser, a deviation mapping calculation is performed. Using the nonlinear coupling relationship model between fouling thermal resistance and time constant drift, the spectral deviation is transformed into a normalized fouling feature vector that characterizes the current degree of fouling accumulation.

[0070] Read the thermal time constant distribution spectrum data, extract the characteristic time constant values ​​and variance weights corresponding to each cluster center, and construct the dynamic thermal inertia feature vector under the current operating condition. Call the pre-stored newly created condenser reference thermal time constant distribution spectrum, which is established based on the equipment's factory acceptance test data and characterizes the ideal heat transfer delay characteristics under fouling-free conditions. Perform spectrum deviation mapping calculation, performing point-by-point difference operations between the dynamic thermal inertia feature vector and the reference spectrum to generate a time constant drift sequence reflecting the increment of the thermal hysteresis effect. Introduce a nonlinear coupling relationship model between fouling thermal resistance and time constant drift. This model, based on heat transfer principles, describes the exponential correlation between the increase in thermal resistance caused by the increase in fouling layer thickness and the extension of thermal response time. Calculate the normalized fouling characteristic value using the following formula:

[0071]

[0072] Where F is the normalized fouling characteristic value, Δτ is the time constant drift, and k is the coupling coefficient. The drift threshold is used. The calculated fouling feature values ​​are encapsulated into a multi-dimensional vector and combined with the weight coefficients of different flow channel regions to generate a normalized fouling feature vector that characterizes the overall degree of fouling accumulation.

[0073] S2.5: The normalized dirt feature vector and the thermal time constant distribution spectrum are fused and encapsulated in a multi-dimensional feature manner, and timestamp tags and operating condition identification metadata are added to construct a dynamic thermal inertial fingerprint library with version iteration capability, which serves as the sole data source for sensing the intrinsic characteristics of the equipment in subsequent control strategies.

[0074] like Figure 2 As shown, step S3: Real-time acquisition of water quality parameters of the condenser circulating water, and construction of a real-time water quality feature vector. The water quality parameters include: pH value, calcium-magnesium ion concentration ratio, bicarbonate concentration, and redox potential. Specifically, this includes:

[0075] S3.1: Perform high-frequency sampling and digital filtering on the original analog voltage signals output by the online pH sensor, conductivity sensor and redox potential sensor deployed in the condenser circulating water pipeline to obtain real-time original pH data, real-time conductivity data and real-time redox potential data after removing noise interference.

[0076] Sub-step S3.1 is located at the beginning of main step S3 and is responsible for converting the analog electrical signals output by the physical sensors into digital water quality parameters with high signal-to-noise ratio, providing a clean data foundation for subsequent ion concentration inversion.

[0077] For the millivolt or microampere level raw analog voltage signals output by the online pH sensor, conductivity sensor and oxidation-reduction potential (ORP) sensor deployed at the condenser circulating water inlet pipe, a high-precision analog-to-digital converter is used to perform high-frequency synchronous sampling, and the sampling frequency is set to 100Hz to capture the transient fluctuation characteristics in the signal and generate discrete time series data.

[0078] For the collected discrete time series data, a sliding window mid-range filter is applied to remove spike pulse noise caused by water flow turbulence or electromagnetic interference. The window length is set to 5 sampling points to preserve the step change characteristics of the signal while suppressing outliers of non-Gaussian distribution.

[0079] The median-filtered data sequence is processed by a low-pass Butterworth filter with a cutoff frequency of 2Hz to eliminate high-frequency random white noise interference, ensuring that the output signal is smooth and without phase distortion, and obtaining preliminary denoised pH, conductivity and ORP digital signals.

[0080] Using a pre-calibrated linear regression model of the sensor, the filtered digital voltage signal is converted into a physically meaningful engineering unit value, where pH value is mapped to the 0-14 range, conductivity is mapped to μS / cm, and ORP is mapped to mV, forming a raw dataset of real-time water quality parameters.

[0081] For example, in a chemical condenser circulating water system, the pH sensor output range is 0-1000mV, corresponding to pH 0-14. The system samples at a frequency of 100Hz, acquiring 100 voltage data points per second. When a peak exceeding three standard deviations is detected in the voltage sequence, a sliding window mid-range filter replaces it with the neighborhood median. Subsequently, a low-pass filter with a cutoff frequency of 2Hz further smooths the signal. If the pH sensor output voltage stabilizes at 642mV after filtering, the real-time pH value is calculated to be 9.0 according to the calibration formula pH = (V / 1000) * 14. Similarly, the conductivity sensor output voltage is processed to obtain a real-time conductivity of 850μS / cm, and the ORP sensor output voltage is processed to obtain a real-time oxidation-reduction potential of 210mV. This process effectively filters out voltage spikes caused by pump start-up and shutdown, ensuring that the water quality data input to the subsequent S3.2 step truly reflects the chemical state of the water.

[0082] S3.2: Based on the real-time conductivity raw data, the calcium-magnesium ion concentration ratio is calculated using ion concentration inversion, and the calculation result is corrected by temperature compensation using a preset hardness correction coefficient to generate standardized real-time calcium-magnesium ion concentration ratio data.

[0083] Receive the raw real-time conductivity data output from step S3.1. This data characterizes the total ion concentration level in the condenser circulating water and serves as the basic input variable for calcium and magnesium ion calculation.

[0084] By calling the preset ion concentration inversion model, a linear mapping relationship between conductivity and total dissolved solids (TDS) is established, and the total ion molar concentration in the condenser circulating water is initially estimated.

[0085] Based on the principle of Langmuir adsorption isotherms, the nonlinear deviation in the high concentration range is corrected, the contribution value of non-hardness ions is separated from the total ion concentration, and the initial calculated value reflecting the total amount of calcium and magnesium ions is extracted.

[0086] Simultaneously acquire real-time water temperature data collected by temperature sensors deployed inside the pipeline to identify the degree of influence of the current fluid thermodynamic state on the ion activity coefficient.

[0087] A temperature compensation function is constructed based on the Arrhenius equation, and the conductivity temperature correction coefficient relative to the standard reference temperature of 25 degrees Celsius is calculated. The specific formula is as follows:

[0088]

[0089] Where K is the measured conductivity, K 25 Here, α is the conductivity at standard temperature, α is the temperature compensation coefficient, and T is the real-time water temperature.

[0090] The initial calcium and magnesium ion calculation values ​​are reversed using the temperature correction coefficient to eliminate errors in ion mobility caused by water temperature fluctuations.

[0091] A preset hardness correction coefficient matrix is ​​introduced, which is derived from regression of historical water quality test data and is used to compensate for the cross-interference of coexisting ions such as sodium and potassium on conductivity.

[0092] The values, after temperature correction and cross-interference compensation, are standardized to generate dimensionless real-time calcium and magnesium ion concentration ratio data.

[0093] S3.3: Perform Nernst equation conversion on the response signal output by the integrated bicarbonate ion selective electrode, and simultaneously correlate it with the real-time pH value raw data to perform acid-base balance verification, and calculate the real-time bicarbonate concentration data.

[0094] The system receives the millivolt-level raw potential signal from the integrated bicarbonate ion-selective electrode, which characterizes the electrochemical activity of bicarbonate ions in the solution. Temperature compensation correction is applied to the raw potential signal, and a thermodynamic correlation model between the potential value and ion activity is established based on the Nernst equation to eliminate the nonlinear effect of condenser circulating water temperature fluctuations on the electrode response sensitivity.

[0095] The theoretical potential of bicarbonate ions after temperature compensation is calculated using the following formula:

[0096]

[0097] Where E is the measured potential value, E0 is the standard electrode potential, R is the ideal gas constant, T is the absolute temperature, n is the ion charge number, F is the Faraday constant, and a HCO3 This represents the activity of bicarbonate ions.

[0098] Based on the above formula, a preliminary estimate of the bicarbonate concentration is obtained through inversion calculation, and the real-time pH data generated in the previous step S3.1 is simultaneously acquired. An acid-base balance verification logic for the carbonate system is constructed, utilizing the stoichiometric constraint relationship between pH and bicarbonate concentration to identify and eliminate abnormal measurement points caused by electrode drift or interfering ions.

[0099] By incorporating Henry's Law constant and the carbonic acid dissociation constant, and combining real-time water temperature and pH values, the theoretical distribution coefficient of bicarbonate ions in aqueous solution is calculated. The preliminary estimated concentration is then weighted and fused with the theoretical distribution coefficient to correct for activity coefficient deviations caused by variations in ionic strength, generating high-precision real-time bicarbonate concentration data.

[0100] S3.4: Combine the real-time pH data, real-time calcium-magnesium ion concentration ratio data, real-time bicarbonate concentration data, and real-time redox potential data with the supplemented turbidity data and chloride ion concentration data, perform timestamp alignment and dimension normalization processing, and construct an initial water quality parameter set containing six independent variables.

[0101] The raw data of real-time pH value, calcium-magnesium ion concentration ratio, bicarbonate concentration and redox potential calculated by the previous steps are obtained and simultaneously connected to the supplementary water quality parameters collected by the online turbidity meter and chloride ion selective electrode. The supplementary water quality parameters include the turbidity data and chloride ion concentration data.

[0102] High-precision timestamp alignment processing is performed on the six-dimensional heterogeneous water quality parameters. Linear interpolation is used to resample sensor data with different sampling frequencies to a unified time reference point, eliminating timing misalignment caused by signal transmission delay.

[0103] A dimensional normalized mapping matrix is ​​constructed, and independent scaling factors and offsets are set for the logarithmic scaling characteristics of pH, the molar differences of ion concentration, and the millivolt fluctuation range of potential.

[0104] The Min-Max standardization formula is applied to linearly transform the data in each dimension, compressing the physical quantity values ​​to the [0,1] interval, thus eliminating the interference of dimensional inconsistency on the convergence of subsequent models. The specific formula is as follows:

[0105]

[0106] Where, x i Let x be the original measured value of the i-th dimension water quality parameter. min With x max These are the minimum and maximum values ​​of the parameter within the historical running window, respectively. This is the normalized standard value.

[0107] Outlier removal and smoothing filtering are performed on the normalized six-dimensional data. A sliding window midpoint filter is used to remove spike noise caused by instantaneous sensor failures, ensuring the continuity and stability of the data distribution.

[0108] According to the preset eigenvector arrangement order, the processed pH value, calcium-magnesium ratio, bicarbonate concentration, redox potential, turbidity and chloride ion concentration are sequentially encapsulated to construct an initial water quality parameter set containing six independent variables.

[0109] S3.5: Perform a multi-dimensional vector encapsulation operation based on the initial water quality parameter set, and map each standardized water quality parameter into a fixed-length array structure according to the preset feature arrangement order to generate a real-time water quality feature vector for quantifying sedimentation risk.

[0110] like Figure 3 As shown, step S4: Input the real-time water quality feature vector into the transfer learning model and reuse the corrosion-deposition coupling pattern features of the industrial boiler scaling database to calculate the scaling tendency index. The scaling tendency index represents the hourly scaling rate probability value mapped to the interval 0 to 1.

[0111] The transfer learning model is used to map the real-time water quality feature vector into a scaling tendency index. Its overall architecture includes a cascaded input layer, a feature standardization processing module, a feature extraction backbone network based on a deep domain adaptation algorithm, a convolutional neural network domain alignment network, and a fully connected mapping and probability calibration module.

[0112] The input layer is used to receive the real-time water quality feature vector, and the feature standardization processing module is used to perform standardization and normalization on each input water quality parameter to eliminate dimensional differences and output a standard water quality feature matrix (see step S4.1 for specific implementation).

[0113] The feature extraction backbone network based on the deep domain adaptation algorithm uses the maximum mean difference loss function to align the water quality feature distribution of the industrial boiler scaling database with the water quality feature distribution of the target domain condenser, extracts high-order abstract features containing crystallization kinetics and local deposition characteristics, and constructs a transfer knowledge feature library (see step S4.2 for specific implementation).

[0114] The convolutional neural network domain alignment network takes the transfer knowledge feature library and the standard water quality feature matrix as parallel inputs, extracts the local corrosion-deposition coupling fluctuation pattern on the time series through a one-dimensional convolutional layer, and introduces the maximum mean difference loss function as a regularization term to correct the feature distribution shift caused by equipment structure differences, and outputs a domain-invariant deposition risk feature representation (see step S4.3 for specific implementation).

[0115] The fully connected mapping and probability calibration module includes three fully connected hidden layers and a Sigmoid probability calibration layer, which are used to output the scaling tendency index after the domain-invariant deposition risk feature representation is sequentially processed by nonlinear mapping compression and probability calibration (see steps S4.4 to S4.5 for specific implementation).

[0116] The transfer learning model uses water quality parameter sequences and corresponding actual scale thickness measurements from an industrial boiler scaling database as labeled training data, and the water quality feature matrix collected in real time from the condenser as unlabeled data for the target domain. It performs end-to-end training on the feature extraction backbone network and the domain alignment network by jointly optimizing the mean squared error loss and the maximum mean difference domain adaptation loss. The maximum mean difference loss is used to constrain the feature distribution of the source and target domains in the hidden layer to tend towards consistency, while the mean squared error loss is used to ensure the model's regression prediction accuracy for the scaling rate in the source domain. After training, the transfer learning model successfully transfers the corrosion-deposition coupling mode features from the industrial boiler scaling database to the current condenser operating conditions, achieving accurate cross-domain prediction of the scaling tendency index.

[0117] Specifically, this includes: S4.1: Standardizing and normalizing the real-time water quality feature vector to eliminate dimensional differences in water quality parameters such as pH value, calcium-magnesium ion concentration ratio, bicarbonate concentration, and redox potential, generating a standard water quality feature matrix as the benchmark data for model input.

[0118] S4.2: Based on the standard water quality feature matrix, the deep domain adaptation algorithm is used to extract corrosion-deposition coupling mode features from the industrial boiler scaling database, and a migration knowledge feature library containing general crystallization kinetics and local flow field deposition characteristics is constructed.

[0119] The industrial boiler scaling database is a pre-constructed source domain dataset that stores water quality parameter sequences and corresponding actual scaling thickness measurements from the historical operation of industrial boilers. The water quality parameter sequences include indicators such as pH value, calcium-magnesium ion concentration ratio, bicarbonate concentration, and redox potential. The actual scaling thickness measurements are obtained by directly measuring deposits on the heating surfaces during periodic boiler shutdowns for maintenance. Each water quality parameter sequence in the database, along with its corresponding scaling thickness measurement, constitutes a labeled training sample, providing source domain supervised learning signals for the transfer learning model.

[0120] The standard water quality feature matrix, after being normalized by S4.1, is received and loaded as the target domain input data into the input layer of the pre-trained transfer learning model.

[0121] The database of industrial boiler scaling is accessed. This database stores the sequence of water quality parameters during historical operation and the corresponding actual scaling thickness measurements, forming a labeled source domain dataset.

[0122] The maximum mean difference (MMD) loss function in the deep domain adaptation algorithm is used to calculate the distance between the source domain water quality feature distribution and the target domain condenser water quality feature distribution in the regeneration kernel Hilbert space.

[0123] The MMD distance is calculated using the following formula to quantify distributional differences:

[0124]

[0125] Among them, X s and X t These are the source domain feature set and the target domain feature set, respectively, n s and n t For the sample size, For kernel functions that map to high-dimensional space, Let H be the square of the norm in the reproducing kernel Hilbert space H. For the i-th sample in the source domain feature set, Let j be the j-th sample in the feature set of the target domain.

[0126] A multilayer perceptron is constructed as a feature extractor, and the MMD loss is minimized to force the feature distributions of the source and target domains to align in the hidden layers.

[0127] High-order abstract features were extracted from the aligned hidden layers. These features implied the crystallization kinetics of calcium carbonate and the deposition characteristics of magnesium hydroxide.

[0128] By mapping the extracted high-order abstract features with corrosion potential data, the catalytic effect of corrosion products on the deposition core can be identified.

[0129] The corrosion-deposition coupling mode feature is a high-order abstract feature extracted from an industrial boiler scaling database. It implicitly contains the crystallization kinetics of calcium carbonate, the deposition characteristics of magnesium hydroxide, and the catalytic effect of corrosion products on the deposition core. This feature is obtained through the following method: First, based on a standard water quality feature matrix, the maximum mean difference loss function in the deep domain adaptation algorithm is used to align the water quality feature distribution of the source domain industrial boiler scaling database with the water quality feature distribution of the target domain condenser. Then, high-order abstract features are extracted from the aligned implicit layer. Finally, the extracted high-order abstract features are correlated and mapped with corrosion potential data to identify the catalytic effect of corrosion products on the deposition core, thus obtaining the corrosion-deposition coupling mode feature.

[0130] By integrating crystallization kinetics, local flow field deposition characteristics, and corrosion-deposition coupling modes, a migration knowledge feature base is constructed.

[0131] S4.3: Based on the migration knowledge feature library and the standard water quality feature matrix, perform cross-domain feature alignment processing to correct the feature distribution offset caused by differences in equipment structure, and generate a domain-invariant deposition risk feature representation that is adapted to the current condenser operating conditions.

[0132] The source domain high-dimensional feature maps and standard water quality feature matrices in the transfer knowledge feature library are used as parallel input channels and connected to the convolutional neural network domain alignment network. The feature extraction backbone network of the convolutional neural network domain alignment network includes a one-dimensional convolutional layer, a batch normalization layer, and a max pooling layer. The one-dimensional convolutional kernel size is set to 3×1 with a stride of 1, which is used to capture the local corrosion-deposition coupling fluctuation pattern of water quality parameters in the time series.

[0133] The Maximum Mean Discrepancy (MMD) loss function is introduced as a domain adaptation regularization term to calculate the distance between the scaling characteristic distribution of the source domain industrial boiler and the real-time water quality characteristic distribution of the target domain condenser in the regeneration kernel Hilbert space. The characteristic distribution offset is quantified by the following formula:

[0134]

[0135] Among them, X s With X t n represents source domain feature samples and target domain feature samples, respectively. s With n t For the sample size, For kernel mapping functions, and These are the i-th and j-th samples in the source domain feature samples and the target domain feature samples, respectively.

[0136] During training, the mean squared error loss and the MMD domain adaptation loss are jointly minimized, forcing the deep feature representations learned by the convolutional neural network domain alignment network to have a consistent statistical distribution between the source and target domains, thereby eliminating the feature distribution shift caused by differences in device structure.

[0137] After multiple nonlinear transformations and domain alignment constraints, the output layer generates a deposition risk feature representation vector with domain invariance. This vector eliminates equipment-specific noise and retains only the general physicochemical features that reflect the nature of scaling.

[0138] S4.4: Perform nonlinear mapping operation on the fully connected layer based on the domain-invariant deposition risk characteristic representation to obtain the original scaling rate prediction value.

[0139] The system receives a domain-invariant deposition risk feature representation generated through cross-domain feature alignment. This vector eliminates interference from equipment structural differences between industrial boilers and condensers and serves as input data for the three fully connected hidden layers in the fully connected mapping and probability calibration module. In the three fully connected hidden layers, the first hidden layer has 64 neurons and uses the ReLU activation function to perform a nonlinear transformation on the input high-dimensional abstract features to capture the complex coupling relationships between water quality parameters. The output of the first layer is used as input to the second hidden layer, which has 32 neurons. Through weighted summation and bias term adjustment, it further extracts deep semantic information reflecting the crystallization kinetics of calcium carbonate and magnesium hydroxide. A Dropout regularization mechanism is introduced, randomly discarding some neuron connections with a probability of 0.2 to prevent the model from overfitting specific operating conditions and enhance its generalization ability to water quality fluctuations. The third hidden layer has 16 neurons, compressing high-dimensional features into a low-dimensional space, focusing on the key principal components that determine the scaling rate. The output layer has a single neuron and uses a linear activation function to map the features extracted by the previous layer to continuous raw scaling rate predictions. The output of the j-th neuron in layer l is calculated using the following formula:

[0140]

[0141] in, This represents the output value of the j-th neuron in the l-th layer. The weight coefficients are used to connect the i-th neuron in layer (l-1) with the j-th neuron in layer l. Let f be the bias term for the j-th neuron in the l-th layer, f be the nonlinear activation function, and n be the number of neurons in the previous layer. The network parameters are updated using backpropagation with the mean squared error loss function to minimize the deviation between the predicted scaling rate and the historically measured deposition thickness.

[0142] S4.5: Perform Sigmoid function probability calibration on the original predicted scaling rate to compress the nonlinear deposition trend value to the range of zero to one, and finally generate a scaling tendency index that characterizes the hourly scaling rate probability value.

[0143] The system receives the original predicted scaling rate value generated in the preceding step S4.4. This value is the predicted result reflecting the deposition trend of calcium carbonate and magnesium hydroxide per unit time, output after the nonlinear mapping operation of the three fully connected hidden layers, and serves as the input data for the Sigmoid probability calibration layer. A Sigmoid activation function is constructed as a probability calibration model to map the original predicted value from the unbounded domain to a bounded probability interval. The Sigmoid calibration function is defined using the following mathematical expression:

[0144]

[0145] Where I is the output scaling tendency index, R is the input original scaling rate prediction value, k is the gain coefficient for controlling the steepness of the curve, b is the offset threshold parameter, and e is the natural logarithm base.

[0146] The initial value of the gain coefficient k is set to 1.0, and the offset b is dynamically updated based on the average predicted value of the scaling critical point in historical operating conditions to adapt to the sensitivity requirements under different water quality backgrounds. The original predicted scaling rate R is substituted into the above formula, and exponential and reciprocal calculations are performed to eliminate dimensional differences and suppress the interference of extreme outliers on control decisions. The calculated scaling tendency index I is truncated to ensure that the output value strictly converges within a closed interval of 0 to 1, where 0 represents no scaling risk and 1 represents extremely high scaling risk. Through the above Sigmoid probability calibration process, the original predicted scaling rate value is transformed into a standardized scaling tendency index, completing the mapping output of the domain-invariant deposition risk characteristic representation to the scaling tendency index by the fully connected mapping and probability calibration module.

[0147] Step S5: Based on the matching degree of the dynamic thermal inertial fingerprint database, the scaling tendency index, the heat exchange efficiency decay rate of the previous cycle, and the peak and valley electricity price time period identifier, perform multi-objective weight dynamic migration and output the real-time weight coefficient group of multiple control objectives.

[0148] The heat exchange efficiency decay rate of the previous cycle refers to the relative decrease in the actual heat exchange efficiency of the heat exchange equipment within the previous complete control cycle compared to the heat exchange efficiency at the beginning of the current cycle. This heat exchange efficiency decay rate of the previous cycle is obtained as follows: at the end of each control cycle, the actual heat exchange efficiency is calculated based on the operating data collected within that cycle, and compared with the heat exchange efficiency at the beginning of the cycle. The difference between the two is divided by the heat exchange efficiency at the beginning to obtain the heat exchange efficiency decay rate within that cycle.

[0149] The peak-valley electricity price time period identifier is a label used to distinguish the high and low levels of electricity prices during different electricity consumption periods. This identifier is obtained according to the peak-valley time period division rules published by the power supply department, which typically divide a day into three categories: peak period, flat period, and valley period. Peak period corresponds to the highest electricity consumption period, flat period corresponds to the middle-priced period, and valley period corresponds to the lowest electricity consumption period. In this embodiment, the peak-valley electricity price time period identifier is mapped to a corresponding energy consumption sensitivity coefficient to quantify the constraint strength of different time periods on instantaneous energy consumption targets: peak period is assigned a high sensitivity coefficient, flat period a medium sensitivity coefficient, and valley period a low sensitivity coefficient.

[0150] Specifically, this includes: S5.1: Obtaining the matching degree, fouling tendency index, heat exchange efficiency decay rate of the previous cycle, and peak-valley electricity price time period identifier from the dynamic thermal inertia fingerprint database, and performing normalization mapping processing to generate a standardized multi-dimensional state feature vector. It should be noted that when the system is first started or when historical data from the previous cycle is unavailable, the heat exchange efficiency decay rate of the previous cycle is initialized to 0, and a preset nominal heat exchange efficiency is used as the baseline value; after the first control cycle is completed, this value is updated to the actual calculated value based on measured data.

[0151] S5.2: Based on the standardized multidimensional state feature vector, a conflict resolution operation is performed using a fuzzy logic reasoning mechanism to calculate and generate an initial game factor that characterizes the urgency difference between the heat exchange efficiency target and the scale inhibition target under the current operating conditions.

[0152] The system receives the standardized multidimensional state feature vector generated by S5.1 and extracts the dynamic thermal inertial fingerprint database matching degree component and the scaling tendency index component as the core input variables of the fuzzy logic reasoning system.

[0153] The urgency of the heat exchange efficiency target is defined as the first output variable, and the urgency of the scale inhibition target is defined as the second output variable. Fuzzy subsets based on triangular membership functions are constructed to cover three linguistic value ranges: low, medium, and high.

[0154] Based on the preset expert knowledge base rules, when the scaling tendency index is in the high range and the thermal inertia matching degree is below the threshold, it is determined that scaling risk is dominant and the strong scaling inhibition rule cluster is activated.

[0155] When the thermal inertia matching degree is in the high range and the scaling tendency index is in the low range, the equipment is determined to be in a stable state, the high-efficiency heat exchange priority rule cluster is activated, and the scaling inhibition weight is weakened.

[0156] The Mamdani fuzzy inference mechanism is used to perform a union operation on the activated rule clusters to generate a preliminary fuzzy output set for the two objectives of heat transfer efficiency and fouling inhibition.

[0157] The centroid method is used to defuzzify the initial fuzzy output set, and numerical indices representing the urgency of the heat exchange efficiency target and the urgency of the fouling inhibition target are calculated.

[0158] An initial game factor calculation model is constructed, dividing the urgency value of scaling inhibition by the urgency value of heat exchange efficiency to quantify the relative priority difference between the two under the current operating conditions. The specific formula is as follows:

[0159]

[0160] in, As the initial game factors, This represents the critical value for scale inhibition. The value is crucial for heat exchange efficiency.

[0161] S5.3: Based on the initial game factors and combined with the peak-valley electricity price time period identifier of the power grid, a time-sensitive weight drift strategy is executed to dynamically adjust the basic priority of instantaneous energy consumption targets in the overall evaluation system and generate a time-varying benchmark weight matrix.

[0162] Receive the initial game factor and the peak-valley electricity price time period identifier output by S5.2 as the benchmark input data for time-varying weight calculation.

[0163] The peak and valley electricity price time periods of the power grid are analyzed and mapped to the corresponding energy consumption sensitivity coefficients. Peak periods are assigned a high sensitivity coefficient, flat periods are assigned a medium sensitivity coefficient, and valley periods are assigned a low sensitivity coefficient, so as to quantify the constraint strength of instantaneous energy consumption targets in different time periods.

[0164] A time-sensitive weight drift function is constructed, and the energy consumption sensitivity coefficient is nonlinearly coupled with the initial game factor. The basic priority weight of the instantaneous energy consumption target is calculated using the following formula:

[0165]

[0166] in, λ is the instantaneous energy consumption base weight, λ is the drift decay constant, β is the energy consumption sensitivity coefficient, and G is the initial game factor.

[0167] Based on the calculated instantaneous energy consumption base weights, the remaining weight allocation ratios of heat exchange efficiency and scaling inhibition intensity are derived using the complementary normalization principle, ensuring that the sum of the three weights is constant at 1, thereby generating a time-varying benchmark weight matrix containing three dimensions of values.

[0168] By employing the aforementioned time-sensitive weight drift strategy, the static game result from the previous step is transformed into a time-varying benchmark weight matrix that dynamically adapts to the economics of the power grid, thereby achieving a real-time balance between energy consumption costs and process stability.

[0169] S5.4: The time-varying benchmark weight matrix is ​​used to perform nonlinear weighted fusion processing on the initial game factors to eliminate coupling interference between multiple objectives and generate an intermediate weight distribution sequence after smoothing and filtering.

[0170] The system receives the time-varying benchmark weight matrix generated by S5.3 and the initial game factors output by S5.2 as the input data source for nonlinear weighted fusion processing.

[0171] A nonlinear coupling mapping operator based on the hyperbolic tangent function is constructed to map the scalar values ​​representing the conflict between heat transfer efficiency and fouling inhibition in the initial game factors to a dynamic adjustment coefficient space in the interval [-1, 1].

[0172] The dynamic adjustment coefficient is used to perform a reverse correction operation on the instantaneous energy consumption weight component in the time-varying reference weight matrix, thereby reducing the energy consumption priority under high conflict conditions to release control degrees of freedom.

[0173] Exponential smoothing filtering was used to perform time-series de-jittering on the corrected weight vector, with a smoothing coefficient of 0.85 set to suppress weight oscillation caused by high-frequency fluctuations in water quality parameters.

[0174] Calculate the intermediate weight distribution sequence of the three dimensions of heat exchange efficiency, fouling inhibition intensity and instantaneous energy consumption after smoothing, and ensure that the weights of each dimension have continuity and monotonicity constraints on the time axis.

[0175] S5.5: Perform summation constraint verification and proportional normalization reconstruction on the intermediate weight distribution sequence to finally output a real-time weight coefficient set that satisfies the normalization and non-negativity constraints for multiple control objectives.

[0176] The system receives an intermediate weight distribution sequence generated from previous steps. This sequence contains preliminary weight values ​​for three dimensions: heat exchange efficiency, fouling inhibition intensity, and instantaneous energy consumption. Non-negativity constraint checks are performed on the values ​​of each dimension in the intermediate weight distribution sequence. Any abnormal weight values ​​less than zero are identified and truncated, and forcibly reset to zero to eliminate invalid negative weight interference caused by fluctuations in fuzzy logic reasoning.

[0177] Calculate the sum of the weight values ​​of each dimension after nonnegativity correction, construct the weight normalization denominator, and ensure that the subsequent proportion reconstruction has a mathematically complete foundation.

[0178] The following formula is used to perform the proportional normalization reconstruction operation, mapping the weights of each dimension to a unit interval:

[0179]

[0180] Among them, w i Let be the final real-time weighting coefficient for the i-th control objective. is the intermediate state weight value of the i-th control objective after non-negativity correction, and j is the index variable for traversing the three objectives of heat exchange efficiency, fouling inhibition intensity, and instantaneous energy consumption.

[0181] The final real-time weight coefficient group is subjected to a sum constraint secondary check to verify whether the sum of the three is strictly equal to 1. If there is a floating-point operation error, a micro-compensation allocation is performed to ensure the conservation of the weight system.

[0182] The verified heat exchange efficiency weight, fouling inhibition intensity weight, and instantaneous energy consumption weight are encapsulated into a standardized real-time weight coefficient group, and then output to the downstream predictive controller after being tagged with a timestamp.

[0183] Step S6: Input the real-time weight coefficient group into the preset improved piecewise linear predictive controller to generate the predicted temperature trajectory and safety boundary envelope data for the future time period.

[0184] The improved piecewise linear predictive controller is used to generate predicted temperature trajectories and safety boundary envelope data for future time periods based on the real-time weight coefficient group. Its overall architecture includes a first parallel sub-model channel and a second parallel sub-model channel set in parallel, a dynamic gain configuration interface connected to the input of the two channels respectively, and a timing alignment and composite data structure construction module connected to the output of the two channels.

[0185] The dynamic gain configuration interface is used to receive the real-time weight coefficient group, extract the heat transfer efficiency weight and map it to the scaling factor of the state transition matrix in the state space equation of the first parallel sub-model, extract the scaling inhibition intensity weight and map it to the nonlinear constraint boundary coefficient generated by the safety margin envelope in the second parallel sub-model, and perform dynamic gain configuration on the two parallel sub-model channels (see step S6.1 for specific configuration method).

[0186] The first parallel sub-model channel is based on the state-space equation after dynamic gain configuration. It uses the thermal time constant distribution spectrum in the dynamic thermal inertial fingerprint database as the structural parameter basis. Combined with the real-time inlet water temperature sequence and condenser circulating water flow data under the current operating conditions, it recursively deduces the outlet water temperature change trend under no-intervention conditions in the future time period through discrete state-space equation, and outputs the natural evolution trajectory sequence of the future outlet water temperature (see step S6.2 for specific implementation).

[0187] The second parallel sub-model channel is based on the nonlinear constraint boundary coefficient after dynamic gain configuration. It takes the scaling tendency index and condenser circulating water flow constraint parameters as input, determines the maximum allowable temperature rise threshold to prevent rapid deposition of calcium and magnesium ions through the scaling critical temperature rise threshold calculation function, and introduces a time sliding window mechanism to perform iterative calculation and smoothing filtering on each sampling point within the prediction period, and outputs the maximum temperature rise safety margin envelope data (see step S6.3 for specific implementation).

[0188] The time-series alignment and composite data structure construction module is used to perform time-axis alignment verification and resampling on the natural evolution trajectory sequence of the future effluent temperature and the maximum temperature rise safety margin envelope data, construct a composite data structure containing the predicted temperature point set and the safety boundary point set, calculate the real-time temperature difference margin and generate the risk level status identifier vector, and output standardized predicted temperature trajectory and safety boundary envelope data objects (see step S6.4 for specific implementation).

[0189] The improved piecewise linear predictive controller uses the instantaneous energy consumption weight in the real-time weight coefficient group as the switching threshold parameter for the piecewise linear control law. By weighting and combining the outputs of each sub-model, it determines the control law segment interval to be used in the current control cycle and executes the corresponding linear control law in each segment interval. The division of the control law segment interval is dynamically adjusted according to the changing trend of the grid peak-valley electricity price period identifier and the scaling tendency index. When the scaling tendency index is higher than a preset threshold, it switches to the control law segment that prioritizes scaling inhibition. When the scaling tendency index is lower than the preset threshold and the grid peak-valley electricity price period identifier indicates a valley electricity period, it switches to the control law segment that prioritizes energy saving. This achieves real-time weighted adaptive response to the three control objectives of heat exchange efficiency, scaling inhibition intensity, and instantaneous energy consumption.

[0190] Specifically, this includes: S6.1: Obtain the internal architecture definition of the improved piecewise linear predictive controller, and based on the heat exchange efficiency weight and fouling inhibition intensity weight in the real-time weight coefficient group, dynamically configure the gain of the parallel sub-model channels in the controller to generate an improved piecewise linear predictive controller instance with multi-objective adaptive capabilities.

[0191] Receive the real-time weighted coefficient set output from S5, including heat exchange efficiency weight, fouling inhibition intensity weight, and instantaneous energy consumption weight. Analyze the internal architecture definition of the improved piecewise linear predictive controller and identify the independent gain adjustment interfaces of the first parallel sub-model channel and the second parallel sub-model channel.

[0192] The heat transfer efficiency weight values ​​are extracted and mapped to scaling factors of the state transition matrix in the state space equation of the first parallel sub-model. A correction coefficient for the sensitivity of this scaling factor to the thermal inertia time constant is calculated using linear interpolation, generating dynamic gain parameters for predicting the natural temperature evolution trajectory.

[0193] The scaling inhibition intensity weight values ​​are extracted and mapped to the nonlinear constraint boundary coefficients generated by the safety margin envelope in the second parallel sub-model. An exponential decay function is used to transform the weight values ​​into a compression scaling factor for the temperature rise threshold, generating dynamic gain parameters for the safety boundary of the maximum allowable temperature rise.

[0194] The dynamic gain parameters are injected into the configuration register of the controller instance, and the internal computational logic weights of the two parallel sub-models are updated synchronously. A controller initialization self-test is performed to verify the numerical stability and control law continuity after gain configuration, ensuring no singularities are generated.

[0195] S6.2: Input the thermal time constant distribution spectrum in the dynamic thermal inertial fingerprint database into the first parallel sub-model of the configured improved piecewise linear predictive controller, use the state space equation to deduce the trend of water temperature change under no-intervention conditions in the future time period, and generate the natural evolution trajectory sequence of future water temperature.

[0196] Read the thermal time constant distribution spectrum from the dynamic thermal inertial fingerprint database, and extract the mean and variance distribution features of the dominant thermal time constants as the basis for the structural parameters of the state-space model.

[0197] Discretized linear time-invariant state-space equations are constructed to model the condenser heat exchange process as a multi-input single-output system, where the state variables are defined as the coupling vector between the pipe wall temperature field and the fluid temperature field.

[0198] The continuous state equations are discretized using the forward Euler method. A time step matching the sampling period and the thermal time constant is set to generate the discrete state transition matrix A and the control input matrix B. The specific formulas are as follows:

[0199]

[0200] Where X(k) is the state vector at time k, U(k) is the control input vector containing influent temperature and flow rate, and the elements of matrix A are composed of the reciprocal terms of the thermal time constant distribution spectrum, reflecting the thermal hysteresis effect.

[0201] Substitute the real-time inlet water temperature sequence and condenser circulating water flow data under the current operating conditions as the initial boundary conditions, perform recursive iterative calculations, and calculate the evolution values ​​of the state variables within the next N prediction steps.

[0202] By mapping the internal state vector to an observable physical quantity of effluent temperature through the output equation C, and correcting the prediction bias by combining the measurement noise covariance matrix, a sequence of natural evolution trajectory of future effluent temperature under no-intervention conditions is generated.

[0203] S6.3: The scaling tendency index, combined with the condenser circulating water flow constraint parameters, is input into the second parallel sub-model of the configured improved piecewise linear predictive controller to perform nonlinear mapping calculations to determine the critical temperature rise threshold for preventing rapid deposition of calcium and magnesium ions, and to generate the maximum temperature rise safety margin envelope data.

[0204] The scaling inhibition intensity weights from the real-time weight coefficient group output in step S5 are received and combined with the real-time water quality feature vector generated in step S3 and the current condenser circulating water flow constraint parameters as the input boundary conditions for the second parallel sub-model. The probability values ​​representing the deposition trends of calcium carbonate and magnesium hydroxide in the scaling tendency index are extracted and mapped to the thermodynamic phase transition critical temperature offset to establish a temperature safety threshold benchmark to prevent rapid ion precipitation. Based on the Langrillly saturation index LSI correction model, a scaling critical temperature rise threshold calculation function is constructed, and the maximum allowable temperature rise upper limit is determined by the following formula:

[0205]

[0206] in, Let K be the maximum allowable temperature rise threshold, WS be the empirical heat transfer correction coefficient, IG be the scaling inhibition strength weight, Cp be the specific heat capacity of the condenser circulating water, ρ be the density of the condenser circulating water, and Q be the real-time condenser circulating water flow rate. This formula shows that when the scaling tendency index IG increases, the numerator decreases, leading to a reduction in the maximum allowable temperature rise ΔTmax, thereby forcing the system to lower the outlet water temperature to inhibit scaling. Simultaneously, increasing the flow rate Q further reduces the temperature rise threshold, enhancing the flushing effect.

[0207] The calculated maximum temperature rise threshold is superimposed on the current influent temperature baseline to generate a dynamically changing effluent temperature safety upper limit curve. A time-sliding window mechanism is introduced to iteratively calculate for each sampling point within the future prediction period. Considering the lag effect of water quality parameters and the execution delay of flow regulation, the safety upper limit curve is smoothed and filtered to eliminate drastic threshold fluctuations caused by sensor noise. Based on the smoothed safety upper limit curve, a preset safety margin bandwidth is reserved downwards to generate a maximum temperature rise safety margin envelope data that includes an upper warning line and a lower optimization line.

[0208] By using the above-mentioned nonlinear mapping and dynamic envelope generation processing method, the static scaling risk index is transformed into temperature safety boundary data that evolves in real time with the operating conditions, thereby realizing physical pre-interception of the scaling acceleration process and ensuring that control actions are executed within the safety domain.

[0209] S6.4: Based on the natural evolution trajectory sequence of the future effluent temperature and the maximum temperature rise safety margin envelope data, perform time-series alignment processing to construct a composite data structure containing the predicted temperature point set and the safety boundary point set, and generate the predicted temperature trajectory and safety boundary envelope data for the future time period.

[0210] The system receives the future effluent temperature natural evolution trajectory sequence generated by S6.2, which includes the predicted temperature values ​​of discrete sampling points within the prediction time window and their corresponding timestamp indices. Simultaneously, it receives the maximum temperature rise safety margin envelope data generated by S6.3, which defines the critical temperature upper limit curve to prevent accelerated scaling and the corresponding safety threshold boundary.

[0211] Strict time axis alignment verification is performed on the natural evolution trajectory sequence and the safety margin envelope data. Linear interpolation is used to resample the two sets of data to a unified time step to ensure that there is a corresponding temperature prediction value and safety boundary threshold at each prediction time point, thus eliminating data misalignment bias caused by asynchronous model calculation.

[0212] Construct a composite data structure that includes a set of predicted temperature points and a set of safety boundary points. For each aligned time step t, extract the predicted natural evolution temperature and the maximum allowable temperature rise at that moment, and calculate the real-time temperature difference margin between the two.

[0213] The real-time temperature difference margin is calculated using the following formula:

[0214]

[0215] in, This represents the maximum allowable temperature rise at time t. The predicted temperature for natural evolution at time t. This represents the real-time temperature difference margin at time t.

[0216] Based on the calculated temperature difference margin sequence, a state identifier vector representing the system's operational risk level in the future period is generated. When the real-time temperature difference margin is greater than the preset safety threshold, it is marked as a low-risk zone; when the real-time temperature difference margin is less than or equal to the preset safety threshold but greater than zero, it is marked as a warning zone; and when the real-time temperature difference margin is less than zero, it is marked as a high-risk over-limit zone.

[0217] The timestamp, predicted temperature value, safety boundary value, temperature difference margin, and risk level identifier are encapsulated into a standardized predicted temperature trajectory and safety boundary envelope data object, which serves as the direct input basis for the mode switching logic determination in step S7.

[0218] Step S7: Determine the positional relationship between the actual outlet water temperature and the safety boundary envelope data, as well as the changing trend of the scaling tendency index. When the preset triggering conditions are met, trigger the feedforward compensation mode to generate a high-flow-rate, low-inlet-temperature setting command; otherwise, trigger the gradual adjustment mode to generate a fine-tuning valve opening slope command. Specifically, this includes:

[0219] S7.1: Obtain the predicted temperature trajectory and safety boundary envelope data for the future time period generated by the previous steps, as well as the real-time updated fouling tendency index. Perform spatiotemporal alignment processing on the predicted temperature trajectory and safety boundary envelope data to construct a real-time state comparison vector containing the dimensions of temperature deviation and rate of change.

[0220] S7.2: Based on the real-time state comparison vector, perform multi-threshold logic discrimination processing, calculate the approximation index of the actual outlet water temperature relative to the upper edge of the safety boundary envelope, and generate a scaling risk mutation flag bit by combining the time series difference results of the scaling tendency index, and output a mode switching trigger signal that indicates whether the current operating condition is in the energy-saving priority zone or the scaling inhibition zone.

[0221] The predicted effluent temperature sequence and the data along the upper edge of the safety boundary envelope are extracted from the real-time state comparison vector. The difference between the two at each sampling time is calculated to construct a temperature approximation time series. A sliding window mean filter is applied to the temperature approximation time series to eliminate instantaneous measurement noise interference and generate a smoothed relative temperature safety margin.

[0222] Historical time-series data of the scaling tendency index are obtained, and the index change within adjacent sampling periods is calculated using first-order forward differencing to construct a scaling risk change rate sequence. A dynamic threshold judgment logic is set for the scaling risk change rate sequence. When the change rate exceeds a preset mutation threshold, the current moment is marked as a scaling risk mutation point, and a binary scaling risk mutation flag is generated.

[0223] The smoothed relative temperature safety margin and the scaling risk mutation flag are input into the multi-threshold logic discrimination module. The energy-saving priority zone is defined as having a relative temperature safety margin greater than the first threshold and a scaling risk mutation flag of zero; the scaling inhibition zone is defined as having a relative temperature safety margin less than the second threshold or a scaling risk mutation flag of one.

[0224] Based on the judgment result, a mode switching trigger signal is generated. If the energy-saving priority zone condition is met, a low-level trigger signal is output; if the scaling inhibition zone condition is met, a high-level trigger signal is output.

[0225] S7.3: In response to the scaling inhibition zone state indicated by the mode switching trigger signal, a preset feedforward compensation control law is invoked. Based on the strength level of the scaling risk mutation flag, the condenser circulating water flow rate increment coefficient and the shell-side medium inlet temperature adjustment range are calculated to generate a high-flow-rate, low-inlet-temperature setting command aimed at physically suppressing the scaling acceleration process.

[0226] The receiver mode switching trigger signal indicates the scale suppression zone status, and the current scale risk mutation flag strength level value is read. Based on the scale risk mutation flag strength level, the preset feedforward compensation control law function library is invoked to match the corresponding high-response gain coefficient matrix.

[0227] Based on the current thermal time constant distribution spectrum in the dynamic thermal inertial fingerprint database, the condenser circulating water flow rate increment coefficient is calculated to offset the predicted natural temperature rise trend. The condenser circulating water flow rate increment coefficient is calculated using the following formula:

[0228]

[0229] Where ΔQ is the coefficient of increase in condenser circulating water flow rate. This is the feedforward gain coefficient. To predict the outlet water temperature, The upper limit of the safety boundary temperature, The thermal inertia time constant, Specific heat capacity.

[0230] Based on the shell-side medium inlet temperature setpoint, the temperature reduction range is calculated to enhance the physical suppression effect of the heat exchange temperature difference driving force. A linear interpolation method is used to determine the temperature reduction step size based on the rate of change of the fouling tendency index, generating a shell-side medium inlet temperature reduction command.

[0231] The condenser circulating water flow rate increment coefficient is converted into an opening increment command for the electric regulating valve, which is then superimposed on the current valve opening reference value to form a high-flow-rate regulation action sequence. The shell-side medium inlet temperature reduction magnitude is converted into the target setpoint of the temperature control unit to form a low inlet temperature control command.

[0232] The high flow rate regulation action sequence and low inlet temperature control command are encapsulated in a time-space synchronization manner, and a scaling inhibition priority mode identifier is added to generate a high flow rate low inlet temperature setting command.

[0233] S7.4: In response to the energy-saving priority zone state indicated by the mode switching trigger signal, a gradient-limited gradual adjustment strategy is enabled. The valve opening fine-tuning slope is calculated based on the deviation distance between the actual outlet water temperature and the data center of the safety boundary envelope, and a fine-tuning valve opening slope command is generated to maintain thermal inertia stability and avoid frequent operation.

[0234] The receiver mode switching trigger signal indicates the energy-saving priority zone status, and the real-time deviation between the actual outlet water temperature measurement value and the data center line of the safety boundary envelope is extracted.

[0235] Based on the historical thermal time constant distribution spectrum matched in the dynamic thermal inertial fingerprint database, the system thermal response hysteresis coefficient under the current operating condition is determined as the physical constraint boundary of the valve regulation rate.

[0236] A gradient-limited control law containing a deviation proportional term and a rate of change differential term is constructed, and the theoretical valve opening adjustment amount is calculated to prevent water flow oscillation caused by excessive adjustment.

[0237] The fine-tuning valve opening slope command is calculated using the following formula:

[0238]

[0239] Where K is the final output valve opening slope command. The valve opening reference value of the previous cycle is given, e(t) is the deviation between the actual outlet water temperature and the center line of the safety boundary, α is the proportional gain coefficient, β is the differential gain coefficient, and α and β are limited by the maximum gradient threshold set by the thermal response hysteresis coefficient.

[0240] The calculated theoretical opening slope is subjected to dead zone filtering. When the absolute value of the slope is less than the preset minimum action threshold, it is forcibly set to zero to suppress high-frequency jitter of the actuator.

[0241] The effective slope value after dead zone filtering is mapped to the duty cycle increment of the pulse width modulation signal of the electric control valve, generating a standardized fine-tuning valve opening slope command.

[0242] S7.5: Standardize and encapsulate the generated high-flow-rate low-inlet-temperature setting command or the fine-tuning valve opening slope command, and add current timestamp and control mode identifier metadata to form a standard control action data package that can be directly parsed by the downstream digital twin verification module.

[0243] Step S8: Send the feedforward compensation mode command or the gradual adjustment mode command to the circulating water valve actuator and the shell-side medium temperature control unit for temperature control. In this embodiment, before sending the feedforward compensation mode command or the gradual adjustment mode command to the circulating water valve actuator and the shell-side medium temperature control unit, the method further includes: inputting the feedforward compensation mode command or the gradual adjustment mode command into the digital twin verification module for closed-loop simulation verification. Specifically, this includes:

[0244] S8.1: Obtain the feedforward compensation mode command or the gradual adjustment mode command, as well as the current condenser operating status parameters. Utilize industrial real-time data mapping technology to convert the physical side condenser circulating water flow setpoint, shell-side medium inlet temperature setpoint, and valve opening slope command into a virtual control vector that can be recognized by the digital twin verification module, generating a high-fidelity simulation excitation signal to drive the virtual condenser model.

[0245] S8.2: Based on the high-fidelity simulation excitation signal, the digital twin verification module with embedded thermodynamic equations is driven to perform sub-second multi-physics coupled closed-loop simulation calculations to simulate the flow field distribution and heat transfer process of condenser circulating water flowing through the heat exchange tube bundle, and generate a virtual operating status dataset containing the dynamic evolution curve of shell-side pressure in the future short time domain and the change trajectory of the net positive suction head at the pump inlet.

[0246] S8.3: Perform extreme value search and threshold comparison processing on the virtual operation status dataset, extract the peak pressure data in the shell side pressure dynamic evolution curve and the lowest net positive suction head data in the pump inlet net positive suction head change trajectory, and make logical judgments with the preset shell side pressure resistance limit threshold and the pump critical cavitation margin threshold respectively to generate a risk judgment flag bit that represents the execution safety.

[0247] Receive the virtual operating status dataset output by the digital twin verification module. This dataset contains a sequence of dynamic evolution curves of shell-side pressure and a sequence of changes in the net positive suction head at the pump inlet in the short time domain.

[0248] A sliding window extreme value search process is performed on the shell-side pressure dynamic evolution curve sequence. All discrete pressure sampling points within the prediction time window are traversed, and the global maximum peak pressure value is identified and extracted as a key indicator for evaluating the integrity of the shell structure.

[0249] The reverse extreme value retrieval process is performed on the trajectory sequence of the net positive suction head at the pump inlet to locate the lowest valley point of the net positive suction head within the prediction period. The minimum net positive suction head value is extracted as the core criterion for determining whether cavitation has occurred in the centrifugal pump.

[0250] The system calls upon a pre-set equipment safety threshold database to read the condenser shell-side design pressure limit threshold and the condenser circulating water pump critical net positive suction head (NPSH) threshold, ensuring that the comparison benchmark meets the technical specifications of the current equipment model.

[0251] A pressure safety logic judgment unit is constructed, which compares the extracted peak pressure value with the shell-side pressure resistance limit threshold. If the peak pressure is greater than the pressure resistance limit threshold, a high-pressure risk flag is generated; otherwise, a pressure safety flag is generated.

[0252] A cavitation safety logic judgment unit is constructed, which compares the extracted minimum net positive suction head value with the critical cavitation margin threshold of the water pump. If the minimum net positive suction head value is less than the critical cavitation margin threshold, a cavitation risk flag is generated; otherwise, an anti-cavitation safety flag is generated.

[0253] The high-pressure risk flag and the cavitation risk flag are combined by performing a logical OR operation. If either flag indicates the presence of risk, the final risk assessment flag is determined to be in a dangerous state. If both flags indicate safety, the final risk assessment flag is determined to be in a safe state.

[0254] S8.4: Execute control instruction filtering operation according to the logical state of the risk judgment flag. If the risk judgment flag indicates that there is a risk of shell pressure exceeding the limit or water pump cavitation, discard the current feedforward compensation mode instruction or the gradual adjustment mode instruction and trigger the alarm reset mechanism. If the risk judgment flag indicates that it is safe, retain the original instruction and encapsulate it into a final execution control package that has passed safety verification, and output a set of control instructions to be issued with engineering robustness.

[0255] The system receives the risk assessment flag generated in step S8.3. This flag contains a logical combination of the shell-side pressure safety status bit and the pump cavitation risk status bit. The system parses the binary code of the risk assessment flag. If the high-order bit indicates that the shell-side pressure exceeds the limit or the low-order bit indicates that the net positive suction head is below the critical threshold, the system determines that the current control command has an execution risk. The system triggers the command discard mechanism, immediately blocking the transmission path of feedforward compensation mode commands or gradual adjustment mode commands to the actuator, preventing physical equipment from entering an unsafe working area. The system simultaneously activates the alarm reset mechanism, generating an abnormal event log containing fault type codes, a snapshot of current operating parameters, and a timestamp, and sends an audible and visual alarm signal to the host computer to prompt operator intervention. If all bits of the risk assessment flag indicate a safe state, the original control command data packet is retained. The retained control command undergoes integrity verification to confirm that the condenser circulating water flow rate setpoint, shell-side medium inlet temperature setpoint, and valve opening slope are all within the preset physical limits. The verified control command is encapsulated into a final execution control packet conforming to the industrial fieldbus protocol standard, with an added safety verification pass identifier and serial number.

[0256] S8.5: The set of control commands to be issued, which has been verified by security, is sent to the condenser circulating water valve actuator and the shell-side medium temperature control unit through the industrial fieldbus protocol. The electric regulating valve is driven to perform physical action response according to the fine-tuning valve opening slope command or the large flow low inlet temperature setting command, so as to perform temperature control and realize the physical implementation of the multi-objective collaborative optimization strategy.

[0257] For those skilled in the art, various other corresponding changes and modifications can be made based on the technical solutions and concepts described above, and all such changes and modifications should fall within the protection scope of the claims of this invention.

[0258] Unless otherwise defined, the technical or scientific terms used herein shall have the ordinary meaning as understood by one of ordinary skill in the art to which this application pertains. The terms “first,” “second,” “third,” and similar terms used in this patent application specification and claims do not indicate any order, quantity, or importance, but are merely used to distinguish different components. Similarly, the terms “an” or “a” and similar terms do not indicate a quantity limitation, but rather indicate the presence of at least one. The terms “comprising” or “including” and similar terms mean that the element or object preceding “comprising” or “including” encompasses the element or object listed following “comprising” or “including” and its equivalents, and do not exclude other elements or objects. The “multiple” mentioned in the embodiments of this application refers to two or more. A and / or B indicate three possibilities: A; B; and A and B.

[0259] The above description is merely an exemplary embodiment of this application, but the scope of protection of this application is not limited thereto. Any person skilled in the art can easily conceive of various equivalent modifications or substitutions within the technical scope disclosed in this application, and such modifications or substitutions should all be covered within the scope of protection of this application. Therefore, the scope of protection of this application should be determined by the scope of the claims.

Claims

1. A method for stabilizing the outlet temperature of condenser circulating water, characterized in that, include: A small, non-disturbance step thermal pulse signal is injected into the condenser circulating water inlet side, and the temperature response time-series curves of the shell-side medium side and the outlet water side are collected simultaneously to obtain the original thermal response dataset. Based on the original thermal response dataset, the thermal time constant distribution spectrum of the condenser is extracted, and a fouling feature vector is generated by combining it with the newly established condenser reference spectrum. The thermal time constant distribution spectrum and the fouling feature vector are then fused into a dynamic thermal inertial fingerprint database. Real-time water quality parameters of condenser circulating water are acquired, and a real-time water quality feature vector is constructed. The real-time water quality feature vector is input into the transfer learning model to reuse the corrosion-deposition coupling mode features of the industrial boiler scaling database, and a scaling tendency index is calculated and generated. Based on the matching degree of the dynamic thermal inertia fingerprint database, the scaling tendency index, the heat exchange efficiency decay rate of the previous cycle, and the peak and valley electricity price time period identifier, multi-objective weight dynamic migration is performed, and real-time weight coefficient groups corresponding to multiple control objectives are output. The real-time weight coefficient group is input into a preset improved piecewise linear predictive controller to generate predicted temperature trajectory and safety boundary envelope data for future time periods. Determine the positional relationship of the actual outlet water temperature relative to the safety boundary envelope data and the changing trend of the scaling tendency index. When the preset triggering conditions are met, trigger the feedforward compensation mode to generate a large flow rate and low inlet temperature setting command; otherwise, trigger the slow-change adjustment mode to generate a fine-tuning valve opening slope command. The feedforward compensation mode command or the gradual adjustment mode command is sent to the circulating water valve actuator and the shell-side medium temperature control unit for temperature control.

2. The method for stabilizing the outlet temperature of condenser circulating water according to claim 1, characterized in that, Before sending the feedforward compensation mode command or the gradual adjustment mode command to the circulating water valve actuator and the shell-side medium temperature control unit, the method further includes: inputting the feedforward compensation mode command or the gradual adjustment mode command into the digital twin verification module for closed-loop simulation verification.

3. The method for stabilizing the outlet temperature of condenser circulating water according to claim 1, characterized in that, The water quality parameters include: pH value, calcium-magnesium ion concentration ratio, bicarbonate concentration, and redox potential.

4. The method for stabilizing the outlet temperature of condenser circulating water according to claim 3, characterized in that, Step S3 specifically includes: The original analog voltage signals output by the online pH sensor, conductivity sensor and redox potential sensor deployed in the condenser circulating water pipeline are sampled at high frequency and digitally filtered to obtain real-time original pH data, real-time original conductivity data and real-time original redox potential data after removing noise interference. Based on the real-time conductivity raw data, the calcium-magnesium ion concentration ratio is calculated using ion concentration inversion, and the calculation result is corrected by temperature compensation using a preset hardness correction coefficient to generate standardized real-time calcium-magnesium ion concentration ratio data. The response signal output by the integrated bicarbonate ion selective electrode is processed by Nernst equation conversion, and the acid-base balance is verified by synchronously correlating it with the real-time pH value raw data, and the real-time bicarbonate concentration data is calculated. The real-time pH data, real-time calcium-magnesium ion concentration ratio data, real-time bicarbonate concentration data, and real-time redox potential data are combined with the supplemented turbidity data and chloride ion concentration data, and then time-stamped and dimensionally normalized to construct an initial set of water quality parameters. Based on the initial set of water quality parameters, a multi-dimensional vector encapsulation operation is performed to map each standardized water quality parameter into a fixed-length array structure according to a preset feature arrangement order, thereby generating the real-time water quality feature vector.

5. The method for stabilizing the outlet temperature of condenser circulating water according to claim 1, characterized in that, The transfer learning model includes: a cascaded input layer, a feature normalization processing module, a feature extraction backbone network based on a deep domain adaptation algorithm, a convolutional neural network domain alignment network, and a fully connected mapping and probability calibration module.

6. The method for stabilizing the outlet temperature of condenser circulating water according to claim 1, characterized in that, Step S4 specifically includes: The real-time water quality feature vector is standardized and normalized to generate a standard water quality feature matrix; Based on the standard water quality feature matrix, corrosion-deposition coupling pattern features are extracted from the industrial boiler scaling database to construct a migration knowledge feature library. Based on the aforementioned migration knowledge feature library and standard water quality feature matrix, cross-domain feature alignment processing is performed to generate a domain-invariant sedimentation risk feature representation. Based on the domain-invariant deposition risk characteristic representation, a fully connected layer nonlinear mapping operation is performed to obtain the original scaling rate prediction value; The original predicted scaling rate is subjected to Sigmoid function probability calibration to generate the scaling tendency index.

7. The method for stabilizing the outlet temperature of condenser circulating water according to claim 1, characterized in that, The multiple control objectives include: heat exchange efficiency, fouling inhibition intensity, and instantaneous energy consumption.

8. The method for stabilizing the outlet temperature of condenser circulating water according to claim 7, characterized in that, Step S5 specifically includes: The matching degree, fouling tendency index, heat exchange efficiency decay rate of the previous cycle, and peak and valley electricity price period identifier of the dynamic thermal inertia fingerprint database are obtained and normalized mapping is performed to generate a standardized multi-dimensional state feature vector. Based on the standardized multidimensional state feature vector, conflict resolution operation is performed using a fuzzy logic reasoning mechanism to calculate and generate the initial game factor. Based on the initial game factors and combined with the peak-valley electricity price time period identifier of the power grid, a time-sensitive weight drift strategy is executed to generate a time-varying benchmark weight matrix. The initial game factors are nonlinearly weighted and fused using the time-varying benchmark weight matrix to generate an intermediate weight distribution sequence. Perform summation constraint verification and proportional normalization reconstruction on the intermediate state weight distribution sequence, and output the real-time weight coefficient group.

9. The method for stabilizing the outlet temperature of condenser circulating water according to claim 1, characterized in that, The improved piecewise linear predictive controller includes a first parallel sub-model channel and a second parallel sub-model channel configured in parallel, a dynamic gain configuration interface connected to the input terminals of the two channels respectively, and a timing alignment and composite data structure construction module connected to the output terminals of the two channels.

10. The method for stabilizing the outlet temperature of condenser circulating water according to claim 1, characterized in that, The preset triggering conditions include: the preset triggering conditions are that the relative safety margin of temperature is less than or equal to the second threshold and the scaling risk mutation flag is 1. The relative safety margin of temperature is obtained by the difference between the actual effluent temperature and the upper edge of the safety boundary envelope after sliding window mean filtering. The scaling risk mutation flag is generated by comparing the first-order forward difference result of the historical time series of the scaling tendency index with the preset mutation threshold.

Images