Intelligent control method and system for power plant cooling system based on multi-source data

Abstract

The present invention provides a method and system for intelligent control of a power plant cooling system based on multi-source data, which relates to the field of power system control technology. The method comprises the following steps: collecting real-time operating data, load forecast data, and weather forecast data; calculating performance indicators based on refrigeration efficiency and energy consumption; determining initial values ​​of control parameters through adaptive reinforcement learning; constructing a data sequence to extract dynamic characteristics; generating a reference trajectory to calculate deviation values; and generating a sequence of adjustment instructions that meet constraints. The method can realize intelligent control of the cooling system and improve system stability and energy utilization efficiency.

Description

Technical Field

[0001] The present invention relates to the field of power system control technology, and in particular to a method and system for intelligently controlling a power plant cooling system based on multi-source data. Background Art

[0002] The cooling system of a thermal power plant is a key link in the plant's energy conversion and thermal energy utilization. The operating efficiency of the cooling system directly affects the plant's overall energy consumption and economic benefits. The cooling system mainly includes equipment such as cooling towers, circulating water pumps, and condensers. The flow of circulating water between the equipment enables heat transfer and dissipation. How to optimize the operating parameters of the cooling system, improve cooling efficiency, and reduce energy consumption has become a key research direction in the power industry.

[0003] Existing technologies primarily rely on fixed empirical rules or simple feedback control, failing to fully utilize multi-source data for comprehensive decision-making and adapting to complex and changing operating environments and load conditions. Passive response modes also make it difficult to implement control at the optimal time and effectively address the impact of meteorological conditions and load fluctuations. Furthermore, the thermodynamic process of cooling systems exhibits significant nonlinear characteristics and time lag effects, making it difficult for existing technologies to capture the dynamic changes of the system. Control accuracy and response speed are limited, and the control effect is particularly poor when the system state undergoes drastic changes.

[0004] Therefore, a solution is urgently needed to solve the problems existing in the prior art. Summary of the Invention

[0005] The embodiments of the present invention provide a method and system for intelligently controlling a power plant cooling system based on multi-source data, which can at least solve some of the problems existing in the prior art.

[0006] A first aspect of an embodiment of the present invention provides an intelligent control method for a power plant cooling system based on multi-source data, comprising:

[0007] Collect real-time operating data of the power plant cooling system, power plant load forecast data and weather forecast data;

[0008] The system performance index is calculated by combining cooling efficiency and energy consumption. A parameter sensitivity analysis is performed on pre-acquired historical operating data to obtain weight coefficients. The performance index change trend for the next four hours is calculated by combining power plant load forecast data and weather forecast data. The current performance score is calculated based on real-time operating data and the weight coefficients. The exploration step size is dynamically updated based on the current performance score and performance index change trend. Initial control parameters are iteratively calculated using adaptive reinforcement learning, and the initial incremental values ​​for fan speed and circulating water flow are output.

[0009] The collected real-time operation data is constructed into a data sequence in chronological order, the mutual information value of the data sequence is calculated to determine the optimal time delay, the false neighbor ratio is calculated to determine the minimum embedding dimension, and the dynamic characteristic value sequence is extracted based on the optimal time delay and the minimum embedding dimension. The steady-state operation reference trajectory is generated and the deviation value between it and the current operation trajectory is calculated. The correction parameter is generated based on the deviation value and superimposed with the initial incremental value of the fan speed and the initial incremental value of the circulating water flow to obtain the target adjustment value corresponding to the fan speed and the circulating water flow;

[0010] The minimum interval time and maximum adjustment range of the control parameters are obtained, the target adjustment value is constrained and verified, an adjustment instruction sequence that meets the constraints is generated, and the system adjustment is executed.

[0011] In an optional embodiment,

[0012] Collecting real-time operating data of the power plant cooling system, power plant load forecast data and weather forecast data including:

[0013] Collecting power plant cooling system operating data, including cooling tower outlet water temperature, cooling tower inlet water temperature, circulating water flow, ambient temperature, ambient humidity, and fan speed;

[0014] Collecting power plant load forecast data, the load forecast data including load change forecast values ​​for the next 4 hours;

[0015] Collect weather forecast data, which includes predicted values ​​of ambient temperature and ambient humidity for the next four hours.

[0016] In an optional embodiment,

[0017] The system performance index is calculated by combining cooling efficiency and energy consumption. A parameter sensitivity analysis is performed on pre-acquired historical operating data to obtain weight coefficients. The performance index change trend for the next four hours is calculated by combining power plant load forecast data and weather forecast data. The current performance score is calculated based on real-time operating data and weight coefficients, including:

[0018] The cooling tower inlet water temperature, cooling tower outlet water temperature and wet bulb temperature are calculated to obtain the cooling efficiency. The cumulative value of the fan power and the water pump power is calculated to obtain the system energy consumption. The system performance index is constructed based on the cooling efficiency and system energy consumption.

[0019] Perform principal component analysis and dimensionality reduction on historical operation data to obtain a characteristic variable matrix. Calculate the partial correlation coefficient between each characteristic variable in the characteristic variable matrix and the system performance index. Use the ratio of the absolute value of the partial correlation coefficient to the sum of the absolute values ​​of the partial correlation coefficients as the weight coefficient.

[0020] A sliding time window method is used to construct a performance indicator prediction sequence. The previously acquired load change prediction value, ambient temperature prediction value, and ambient humidity prediction value are substituted into the performance indicator prediction sequence to calculate the performance indicator change trend for the next 4 hours.

[0021] Extract the equipment running time sequence and the corresponding performance indicator sequence, use the least squares method to fit the performance attenuation baseline curve, construct an adaptive learning rate based on the fluctuation variance of the performance indicator, project the deviation between the short-term change trend and the long-term attenuation trend of the performance indicator into the Gaussian kernel space, obtain the performance attenuation correction coefficient through iterative calculation of the kernel function, calculate the current performance indicator value based on the real-time operation data, and multiply the weighted combination result of the weight coefficient and the current performance indicator value by the performance attenuation correction coefficient to obtain the current performance score.

[0022] In an optional embodiment,

[0023] The exploration step size is dynamically updated based on the current performance score and performance indicator change trend. The initial control parameters are iteratively calculated using adaptive reinforcement learning. The output initial incremental values ​​of the fan speed and circulating water flow include:

[0024] The absolute value of the performance indicator change trend is multiplied by a preset adaptive adjustment coefficient to obtain an exponential decay factor, and the exploration step length is calculated based on the exponential decay factor and the current performance score;

[0025] Generate a random projection matrix using a standard normal distribution, generate a bias vector based on the normal distribution, multiply a system state vector constructed based on the power plant cooling system operating data by the random projection matrix and add the bias vector to obtain a feature mapping matrix using a sigmoid activation function;

[0026] Calculating the generalized inverse matrix of the feature map matrix based on singular value decomposition, setting the singular values ​​smaller than a preset threshold to zero, and then reconstructing to obtain an initial output weight matrix;

[0027] Calculating a system predicted output based on the feature mapping matrix and the initial output weight matrix, calculating a difference between the system predicted output and the actual system output to obtain a prediction error, multiplying the prediction error by a preset learning rate as a parameter adjustment, updating the initial output weight matrix to obtain a second output weight matrix, and calculating a second system output;

[0028] The exploration step size is multiplied by the gradient output by the second system to obtain an initial incremental value of the fan speed and an initial incremental value of the circulating water flow rate.

[0029] In an optional embodiment,

[0030] The collected real-time operation data is constructed into a data sequence in chronological order. The mutual information value of the data sequence is calculated to determine the optimal time delay. The false neighbor ratio is calculated to determine the minimum embedding dimension. The dynamic characteristic value sequence is extracted based on the optimal time delay and minimum embedding dimension. The steady-state operation reference trajectory is generated and the deviation value between it and the current operation trajectory is calculated. The correction parameter is generated based on the deviation value and superimposed on the initial incremental value of the fan speed and the initial incremental value of the circulating water flow. The target adjustment values ​​corresponding to the fan speed and the circulating water flow are obtained, including:

[0031] Constructing a data sequence from the real-time operation data in chronological order, calculating a marginal probability distribution and a joint probability distribution of the data sequence, calculating a mutual information value based on the marginal probability distribution and the joint probability distribution, and determining a time value corresponding to a first local minimum of the mutual information value as an optimal time delay;

[0032] Calculating the ratio of the number of false neighbor points under different embedding dimensions to the number of sampling points of the data sequence based on the false neighbor algorithm to obtain a false neighbor ratio, and determining the dimension corresponding to the first time the false neighbor ratio is less than a preset threshold as the minimum embedding dimension;

[0033] Combining the data sequence according to the optimal time delay and the minimum embedding dimension to construct a current running trajectory, and calculating the Euclidean distance between adjacent data points to obtain a dynamic characteristic value sequence;

[0034] generating a steady-state operation reference trajectory based on the historical steady-state operation data according to the optimal time delay and the minimum embedding dimension, calculating the Euclidean distance between the current operation trajectory and the steady-state operation reference trajectory to obtain a spatial deviation value, and calculating the difference between the dynamic characteristic value sequence and the dynamic characteristic value sequence corresponding to the historical steady-state operation data to obtain a dynamic deviation value;

[0035] The product of the spatial deviation value and the preset proportional coefficient is added to the product of the dynamic deviation value and the preset differential coefficient to obtain a correction parameter, the product of the correction parameter and the gradient of the spatial deviation value with respect to the fan speed is superimposed on the initial increment of the fan speed to obtain the target adjustment value corresponding to the fan speed, and the product of the correction parameter and the gradient of the spatial deviation value with respect to the circulating water flow is superimposed on the initial increment of the circulating water flow to obtain the target adjustment value corresponding to the circulating water flow.

[0036] In an optional embodiment,

[0037] Calculating the ratio of the number of false neighbor points under different embedding dimensions to the number of sampling points of the data sequence based on the false neighbor algorithm to obtain a false neighbor ratio, and determining the dimension corresponding to when the false neighbor ratio is less than a preset threshold for the first time as the minimum embedding dimension includes:

[0038] The data sequence is reconstructed into a phase space of a preset dimension. The Euclidean distance between the reconstructed vectors is calculated to obtain a distance matrix. The difference between the reconstructed vectors is calculated to obtain a direction vector matrix. The local covariance matrix is ​​calculated based on the reconstructed vectors. The local covariance matrix is ​​subjected to eigenvalue decomposition to obtain an eigenvector matrix and an eigenvalue matrix.

[0039] The distance ratio criterion is obtained by calculating the distance ratio of the reconstructed vector after increasing the dimension. The direction consistency criterion is obtained by calculating the inner product of the eigenvector corresponding to the maximum eigenvalue in the direction vector matrix and the eigenvector matrix. The local density ratio of adjacent reconstructed vectors is obtained by calculating the local density criterion.

[0040] Calculating the ratio of the maximum eigenvalue to the second largest eigenvalue in the eigenvalue matrix to obtain the local manifold curvature, calculating the distance ratio criterion weight and the direction consistency criterion weight based on the local manifold curvature in combination with an exponential function, taking the complement of the sum of the distance ratio criterion weight and the direction consistency criterion weight as the local density criterion weight, and calculating the spatial position weight based on the local manifold curvature, wherein the spatial position weight is inversely proportional to the local manifold curvature;

[0041] The product of the distance ratio criterion and the distance ratio criterion weight, the product of the directional consistency criterion and the directional consistency criterion weight, and the product of the local density criterion and the local density criterion weight are added together to obtain a comprehensive criterion value. The comprehensive criterion value is compared with the preset threshold to obtain the false neighbor judgment result. The product of the false neighbor judgment result and the spatial position weight is calculated to obtain the false neighbor ratio. The dimension corresponding to the first time the false neighbor ratio is less than the preset threshold is determined as the minimum embedding dimension.

[0042] In an optional embodiment,

[0043] Obtaining the minimum interval time and maximum adjustment range of the control parameters, performing constraint verification on the target adjustment value, generating an adjustment instruction sequence that meets the constraints, and executing system adjustment include:

[0044] Obtaining adjustment constraints of the control parameters, wherein the adjustment constraints include a minimum interval time and a maximum adjustment range;

[0045] Dividing the target adjustment value into time interval segments based on the minimum interval time, and constraining the amplitude of the segmented target adjustment value based on the maximum adjustment amplitude to obtain an adjustment value sequence that meets the constraint;

[0046] The adjustment value sequence is converted into an adjustment instruction sequence in chronological order, and the adjustment instruction sequence is executed to realize system adjustment.

[0047] A second aspect of an embodiment of the present invention provides an intelligent control system for a power plant cooling system based on multi-source data, comprising:

[0048] The first unit is used to collect real-time operating data of the power plant cooling system, power plant load forecast data and weather forecast data;

[0049] The second unit is used to calculate the system performance index by combining cooling efficiency and energy consumption, perform parameter sensitivity analysis on pre-acquired historical operating data to obtain weight coefficients, combine power plant load forecast data and weather forecast data to calculate the performance index change trend for the next four hours, calculate the current performance score based on real-time operating data and weight coefficients, dynamically update the exploration step size based on the current performance score and performance index change trend, and iteratively calculate the initial control parameters using adaptive reinforcement learning. The initial incremental values ​​of fan speed and circulating water flow are output;

[0050] The third unit is used to construct a data sequence from the collected real-time operation data in chronological order, calculate the mutual information value of the data sequence to determine the optimal time delay, calculate the false neighbor ratio to determine the minimum embedding dimension, extract the dynamic characteristic value sequence based on the optimal time delay and the minimum embedding dimension, generate a steady-state operation reference trajectory and calculate the deviation value between it and the current operation trajectory, generate a correction parameter based on the deviation value, and superimpose the correction parameter with the initial incremental value of the fan speed and the initial incremental value of the circulating water flow to obtain the target adjustment value corresponding to the fan speed and the circulating water flow;

[0051] The fourth unit is used to obtain the minimum interval time and maximum adjustment range of the control parameters, perform constraint verification on the target adjustment value, generate an adjustment instruction sequence that meets the constraints and execute system adjustment.

[0052] According to a third aspect of an embodiment of the present invention, an electronic device is provided, including:

[0053] A processor and a memory for storing processor-executable instructions, wherein the processor is configured to call the instructions stored in the memory to execute the aforementioned method.

[0054] According to a fourth aspect of an embodiment of the present invention, a computer-readable storage medium is provided, on which computer program instructions are stored. When the computer program instructions are executed by a processor, the method described above is implemented.

[0055] In the present invention, the system performance indicators are calculated by combining refrigeration efficiency and energy consumption, and the exploration step is dynamically updated by adaptive reinforcement learning, so as to realize the precise adjustment of the control parameters of the cooling system, reduce the energy consumption of the system, and improve the cooling effect, so that the system can maintain efficient operation under different working conditions. The system dynamic characteristics are extracted through phase space reconstruction technology, and the steady-state operation reference trajectory is generated. The deviation value is calculated and corrected in real time, and the optimal adjustment instruction sequence is generated in combination with the constraint verification mechanism, which ensures the stability and safety of the system adjustment, avoids the equipment loss caused by frequent adjustment, and extends the service life of the equipment. BRIEF DESCRIPTION OF THE DRAWINGS

[0056] Figure 1 Schematic diagram of a flow chart of a method for intelligent control of a power plant cooling system based on multi-source data according to an embodiment of the present invention;

[0057] Figure 2 Schematic diagram of the curve showing the change of false neighbor ratio with embedding dimension in the power plant cooling system. DETAILED DESCRIPTION

[0058] To make the objectives, technical solutions, and advantages of the embodiments of the present invention more clear, the technical solutions in the embodiments of the present invention will be clearly and completely described below in conjunction with the accompanying drawings in the embodiments of the present invention. Obviously, the described embodiments are only part of the embodiments of the present invention, not all of the embodiments. Based on the embodiments of the present invention, all other embodiments obtained by ordinary technicians in this field without making creative efforts shall fall within the scope of protection of the present invention.

[0059] The following specific embodiments are used to describe the technical solution of the present invention in detail. The following specific embodiments can be combined with each other, and the same or similar concepts or processes may not be described in detail in some embodiments.

[0060] Figure 1 FIG. 1 is a flow chart of an intelligent control method for a power plant cooling system based on multi-source data according to an embodiment of the present invention. Figure 1 As shown, the method includes:

[0061] Collect real-time operating data of the power plant cooling system, power plant load forecast data and weather forecast data;

[0062] The system performance index is calculated by combining cooling efficiency and energy consumption. A parameter sensitivity analysis is performed on pre-acquired historical operating data to obtain weight coefficients. The performance index change trend for the next four hours is calculated by combining power plant load forecast data and weather forecast data. The current performance score is calculated based on real-time operating data and the weight coefficients. The exploration step size is dynamically updated based on the current performance score and performance index change trend. Initial control parameters are iteratively calculated using adaptive reinforcement learning, and the initial incremental values ​​for fan speed and circulating water flow are output.

[0063] The collected real-time operation data is constructed into a data sequence in chronological order, the mutual information value of the data sequence is calculated to determine the optimal time delay, the false neighbor ratio is calculated to determine the minimum embedding dimension, and the dynamic characteristic value sequence is extracted based on the optimal time delay and the minimum embedding dimension. The steady-state operation reference trajectory is generated and the deviation value between it and the current operation trajectory is calculated. The correction parameter is generated based on the deviation value and superimposed with the initial incremental value of the fan speed and the initial incremental value of the circulating water flow to obtain the target adjustment value corresponding to the fan speed and the circulating water flow;

[0064] The minimum interval time and maximum adjustment range of the control parameters are obtained, the target adjustment value is constrained and verified, an adjustment instruction sequence that meets the constraints is generated, and the system adjustment is executed.

[0065] In an optional embodiment,

[0066] Collecting real-time operating data of the power plant cooling system, power plant load forecast data and weather forecast data including:

[0067] Collecting power plant cooling system operating data, including cooling tower outlet water temperature, cooling tower inlet water temperature, circulating water flow, ambient temperature, ambient humidity, and fan speed;

[0068] Collecting power plant load forecast data, the load forecast data including load change forecast values ​​for the next 4 hours;

[0069] Collect weather forecast data, which includes predicted values ​​of ambient temperature and ambient humidity for the next four hours.

[0070] Real-time operating data for the cooling system is collected through a sensor network distributed throughout key nodes of the power plant's cooling system. The cooling tower outlet water temperature is collected using a PT100 temperature sensor with a measurement range of 0-100°C, an accuracy of ±0.1°C, and a sampling frequency of once every 30 seconds. The sensor is installed on the cooling tower outlet pipe and transmits data to the data acquisition unit via a 4-20mA signal. For example, in one case, the cooling tower outlet water temperature was 28.5°C. The cooling tower inlet water temperature is also collected using a PT100 temperature sensor, installed on the cooling tower inlet pipe with the same sampling frequency as the outlet water temperature sensor. In one case, the collected value was 38.7°C.

[0071] The circulating water flow is collected using an electromagnetic flow meter with a range of 0-15000m 3 / h, with an accuracy of ±0.5%, installed on the circulating water main pipeline, with a sampling frequency of once per minute. In actual application, the circulating water flow at a certain moment is 12500m 3 Ambient temperature and humidity are collected using an integrated temperature and humidity sensor with a temperature measurement range of -40°C to 80°C and a humidity measurement range of 0-100% RH, with an accuracy of ±0.3°C and ±2% RH, respectively. Installed in a weather station near the cooling tower, the sensor samples data once per minute. The ambient temperature at a specific moment was 32.6°C and the humidity was 65% RH.

[0072] Fan speed is acquired using Hall effect sensors or encoders with a measurement range of 0-1500 rpm and an accuracy of ±1 rpm. These sensors are mounted directly on the cooling tower fan shaft or acquired indirectly through the inverter output frequency, sampling once per minute. At a specific moment, the fan speed is 850 rpm. All sensor data is transmitted to the power plant's distributed control system or a dedicated data acquisition server via an industrial fieldbus (such as Modbus RTU or Profibus).

[0073] To ensure data quality, collected data is verified in real time. Temperature data is checked to ensure it is within a reasonable range (for example, cooling tower outlet water temperature is typically between 25°C and 35°C). Any abnormal data is flagged and interpolated and corrected based on previous and subsequent data. For example, if the cooling tower outlet water temperature suddenly jumps from 28.5°C to 5.2°C, this value is marked as abnormal and replaced with the previous 10-minute average of 28.7°C. All original and corrected data are stored in a time series database for subsequent analysis and backtracking.

[0074] Power plant load forecast data is collected through an interface with the power plant's production management system. Based on historical load data, the daily plan, and grid dispatch instructions, a four-hour load change forecast is generated with an accuracy of ±3%. Load forecast data is generated at 15-minute intervals, for a total of 16 forecast points. Load forecast data is retrieved from the production management system via the OPC UA protocol or RESTful API and is updated hourly. For example, the four-hour load forecast values ​​obtained at 13:00 are: 580 MW at 13:15, 585 MW at 13:30, 590 MW at 13:45, 595 MW at 14:00, and so on, reaching 630 MW at 17:00.

[0075] Weather forecast data is collected through professional weather service interfaces or the power plant's own weather forecast model. Forecast data includes ambient temperature and humidity forecasts for the next four hours, with forecasts calculated at 30-minute intervals, for a total of eight time points. Forecast data is obtained via HTTP / HTTPS from the weather service provider's API or from the power plant's own weather forecast system, and is updated hourly. For example, the predicted ambient temperature values ​​for the next four hours obtained at 13:00 are: 33.2°C at 13:30, 33.5°C at 14:00, 33.8°C at 14:30, 34.0°C at 15:00, 34.1°C at 16:00, 33.8°C at 16:30, and 33.5°C at 17:00. The corresponding predicted ambient humidity values ​​are: 64% RH at 13:30, 63% RH at 14:00, 62% RH at 14:30, 61% RH at 15:00, 60% RH at 15:30, 60% RH at 16:00, 61% RH at 16:30, and 62% RH at 17:00.

[0076] To ensure the reliability of data collection, a multi-redundancy design is employed. Backup sensors are configured for key data points, automatically switching to the primary sensor if it fails. The communication network utilizes redundant links to prevent single-point failures from causing data interruptions. Furthermore, a data quality monitoring mechanism is implemented to identify slowly drifting sensor failures through statistical analysis. For example, if the measured value of the cooling tower outlet water temperature sensor deviates by more than 2°C from the historical typical value for the same period, load, and environmental conditions for a sustained period of more than 30 minutes, an alarm indicating a possible sensor failure is generated.

[0077] All collected data is timestamped to ensure data consistency and timeliness. Real-time operating data, load forecast data, and weather forecast data are stored in the power plant's time-series database, providing data support for subsequent cooling system optimization and control algorithms. Data storage utilizes a tiered architecture, with hot data (the last seven days) stored in high-performance storage and cold data transferred to low-cost storage. The data is retained for three years to meet the needs of long-term data analysis and optimization.

[0078] In this embodiment, a comprehensive data collection system is constructed by simultaneously collecting cooling system operating data, load forecast data, and weather forecast data. This ensures the completeness of the data required for subsequent analysis. The collected data covers multiple dimensions and can comprehensively consider various influencing factors. By collecting ambient temperature and humidity data, the impact of the environment on cooling performance can be accurately assessed, improving control accuracy. Simultaneously, real-time collection of inlet and outlet water temperatures directly reflects cooling performance, facilitating timely adjustment of control strategies.

[0079] In an optional embodiment,

[0080] The system performance index is calculated by combining cooling efficiency and energy consumption. A parameter sensitivity analysis is performed on pre-acquired historical operating data to obtain weight coefficients. The performance index change trend for the next four hours is calculated by combining power plant load forecast data and weather forecast data. The current performance score is calculated based on real-time operating data and weight coefficients, including:

[0081] The cooling tower inlet water temperature, cooling tower outlet water temperature and wet bulb temperature are calculated to obtain the cooling efficiency. The cumulative value of the fan power and the water pump power is calculated to obtain the system energy consumption. The system performance index is constructed based on the cooling efficiency and system energy consumption.

[0082] Perform principal component analysis and dimensionality reduction on historical operation data to obtain a characteristic variable matrix. Calculate the partial correlation coefficient between each characteristic variable in the characteristic variable matrix and the system performance index. Use the ratio of the absolute value of the partial correlation coefficient to the sum of the absolute values ​​of the partial correlation coefficients as the weight coefficient.

[0083] A sliding time window method is used to construct a performance indicator prediction sequence. The previously acquired load change prediction value, ambient temperature prediction value, and ambient humidity prediction value are substituted into the performance indicator prediction sequence to calculate the performance indicator change trend for the next 4 hours.

[0084] Extract the equipment running time sequence and the corresponding performance indicator sequence, use the least squares method to fit the performance attenuation baseline curve, construct an adaptive learning rate based on the fluctuation variance of the performance indicator, project the deviation between the short-term change trend and the long-term attenuation trend of the performance indicator into the Gaussian kernel space, obtain the performance attenuation correction coefficient through iterative calculation of the kernel function, calculate the current performance indicator value based on the real-time operation data, and multiply the weighted combination result of the weight coefficient and the current performance indicator value by the performance attenuation correction coefficient to obtain the current performance score.

[0085] To construct a system performance indicator, the cooling system's temperature sensors collect data on the cooling tower's inlet water temperature (e.g., 35°C) and outlet water temperature (e.g., 28°C). The current wet-bulb temperature (e.g., 25°C) is obtained from meteorological data. Cooling efficiency can be calculated as the ratio of the cooling range to the theoretical cooling range. The cooling range is the inlet water temperature minus the outlet water temperature (35°C - 28°C = 7°C), while the theoretical cooling range is the inlet water temperature minus the wet-bulb temperature (35°C - 25°C = 10°C). Therefore, the cooling efficiency is 70%. System energy consumption is calculated by accumulating fan power and pump power. For example, if the average fan power is 50 kW and the average pump power is 30 kW over a one-hour period, the total system energy consumption is 80 kWh per hour. The system performance indicator can be defined as the ratio of cooling efficiency to normalized energy consumption. If the 80 kWh energy consumption is normalized to 0.8 (relative to a baseline value of 100 kWh), the system performance indicator is 70% / 0.8 = 87.5%.

[0086] Perform parameter sensitivity analysis on historical operating data to determine weight coefficients. Collect at least 30 days of historical operating data, including parameters such as cooling tower inlet and outlet water temperature, ambient temperature, ambient humidity, fan speed, and pump flow, along with corresponding system performance indicators. Perform principal component analysis to reduce dimensionality, for example, from the original 15 operating parameters to eight characteristic variables. Calculate the partial correlation coefficients between these eight characteristic variables and the system performance indicators. Assume that the absolute values ​​obtained are: 0.82, 0.65, 0.43, 0.38, 0.25, 0.18, 0.12, and 0.08, respectively, with a total of 2.91. Divide each absolute value by the total to obtain the weight coefficients: 0.282, 0.223, 0.148, 0.131, 0.086, 0.062, 0.041, and 0.027, respectively. The weight coefficients reflect the degree of influence of each parameter on system performance.

[0087] A sliding time window method is used to construct a performance indicator prediction sequence. A 24-hour sliding window is selected, with a sliding step of 1 hour. Within each time window, time series data of system performance indicators and related parameters (such as load, ambient temperature, and humidity) are extracted to build a regression model. For example, at the current time of 13:00, data from the previous 24 hours is collected to build a prediction model. The predicted load values ​​(e.g., 80%, 85%, 88%, and 84%), ambient temperature (e.g., 32°C, 33°C, 34°C, and 33°C), and humidity (e.g., 60%, 58%, 55%, and 57%) for the next four hours (13:00-17:00) are substituted into the model to obtain predicted performance indicator values ​​(e.g., 86%, 84%, 82%, and 83%) for the next four hours.

[0088] To calculate the current performance score, consider the impact of device operating time on performance. A sequence of device operating time since commissioning (e.g., 0 days, 30 days, 60 days, ..., 360 days) is extracted, along with the corresponding performance indicator sequence (e.g., 100%, 98%, 97%, ..., 88%). The least squares method is used to fit a performance decay baseline curve. For example, a curve that decreases over time demonstrates the natural decay of device performance with operating time. An adaptive learning rate is also constructed, inversely proportional to the variance of the performance indicator. For example, when the performance indicator fluctuates significantly (with a variance of 2.5), the learning rate is set to 0.02; when the fluctuation is smaller (with a variance of 0.5), the learning rate is set to 0.1. The deviation between the short-term performance indicator change (e.g., the trend over the past seven days) and the long-term decay trend (predicted by the baseline curve) is calculated. For example, if the actual short-term performance decreases by 2% while the baseline curve predicts a 1% decrease, the deviation is -1%. This deviation is projected into the Gaussian kernel space, and the kernel function is iteratively calculated to obtain a performance decay correction factor. For example, 0.95 indicates that the actual performance decay is faster than expected.

[0089] The current performance index is calculated based on real-time operating data. Assuming a current cooling efficiency of 72% and a standardized energy consumption of 0.75, the current performance index is 72% / 0.75 = 96%. Substituting the real-time values ​​of each parameter into the weight coefficients, a weighted combination is calculated. For example, if the weighted performance index is 92%, multiplying this weighted performance index by the performance degradation correction factor yields a current performance score of 92% x 0.95 = 87.4%. The current performance score comprehensively considers real-time operating efficiency, energy consumption, parameter sensitivity, and equipment aging, and serves as an important basis for determining the operating status of the cooling system and optimizing control strategies.

[0090] In this embodiment, the system performance index is constructed by cooling efficiency and system energy consumption, and a comprehensive evaluation of cooling effect and energy consumption is achieved. The principal component analysis is used to reduce the dimension of historical operation data, effectively extract key characteristic variables, and reduce data redundancy. The weight coefficient is determined by the partial correlation coefficient, so that the contribution of each characteristic variable to the performance evaluation matches its actual impact. The projection iterative calculation based on the Gaussian kernel space is used to achieve adaptive adjustment of the performance attenuation correction coefficient, so that the performance evaluation is more in line with the actual operating status. The weight coefficient is weightedly combined with the current performance index value and multiplied by the performance attenuation correction coefficient. The obtained performance score reflects both the current operating status and takes into account historical performance changes, thereby ensuring the accuracy and reliability of the evaluation results.

[0091] In an optional embodiment,

[0092] The exploration step size is dynamically updated based on the current performance score and performance indicator change trend. The initial control parameters are iteratively calculated using adaptive reinforcement learning. The output initial incremental values ​​of the fan speed and circulating water flow include:

[0093] The absolute value of the performance indicator change trend is multiplied by a preset adaptive adjustment coefficient to obtain an exponential decay factor, and the exploration step length is calculated based on the exponential decay factor and the current performance score;

[0094] Generate a random projection matrix using a standard normal distribution, generate a bias vector based on the normal distribution, multiply a system state vector constructed based on the power plant cooling system operating data by the random projection matrix and add the bias vector to obtain a feature mapping matrix using a sigmoid activation function;

[0095] Calculating the generalized inverse matrix of the feature map matrix based on singular value decomposition, setting the singular values ​​smaller than a preset threshold to zero, and then reconstructing to obtain an initial output weight matrix;

[0096] Calculating a system predicted output based on the feature mapping matrix and the initial output weight matrix, calculating a difference between the system predicted output and the actual system output to obtain a prediction error, multiplying the prediction error by a preset learning rate as a parameter adjustment, updating the initial output weight matrix to obtain a second output weight matrix, and calculating a second system output;

[0097] The exploration step size is multiplied by the gradient output by the second system to obtain an initial incremental value of the fan speed and an initial incremental value of the circulating water flow rate.

[0098] The exploration step size is calculated based on the performance indicator's changing trend. The current performance score P is collected from the cooling tower controller. The performance score is a comprehensive calculation of cooling efficiency, energy consumption ratio, and stability, and ranges from 0 to 100. The performance indicator's changing trend, ΔP, is calculated—the difference between the current and previous performance scores. The exponential decay factor β is multiplied by the absolute value of the performance indicator's changing trend, |ΔP|, and the preset adaptive adjustment coefficient α (set to 0.05). A large |ΔP| value indicates that system performance is rapidly changing and requires a larger exploration step size. A small |ΔP| value indicates that system performance is stabilizing and requires a smaller exploration step size. The exploration step size ε is calculated based on the exponential decay factor β and the current performance score P using the function ε = 0.1 × (100 - P) × exp(-β). For example, when P = 85 and |ΔP| = 2, β = 0.05 × 2 = 0.1, and the exploration step size ε = 0.1 × (100 - 85) × exp (-0.1) ≈ 1.35.

[0099] The system state vector is constructed based on the operating data of the power plant's cooling system. Parameters such as the cooling tower inlet and outlet water temperatures, ambient wet-bulb and dry-bulb temperatures, circulating water flow rate, and fan power are collected to form a 10-dimensional system state vector x. For example, the system state vector x at a given moment might be [35.2, 28.7, 25.3, 32.1, 1200, 75, 0.85, 1.2, 0.95, 0.78], representing the inlet water temperature (°C), outlet water temperature (°C), wet-bulb temperature (°C), dry-bulb temperature (°C), circulating water flow rate (m³ / h), fan power (kW), and four other normalized parameters.

[0100] Subsequently, a standard normal distribution is used to generate a random projection matrix W with a matrix dimension of 50×10, where each element obeys a normal distribution with a mean of 0 and a standard deviation of 1. At the same time, a bias vector b is generated based on the normal distribution with a dimension of 50×1, where its elements obey a normal distribution with a mean of 0 and a standard deviation of 0.1. The system state vector x is multiplied by the random projection matrix W and added to the bias vector b to obtain the intermediate result Wx+b. The intermediate result is processed by the sigmoid activation function g(z)=1 / (1+exp(-z)) to obtain the feature mapping matrix H. For example, for the first element of the input vector, calculate: z1=(w 11 ×35.2+w 12 ×28.7+...+w 110 ×0.78)+b1=2.31, and through the sigmoid function we get H1=1 / (1+exp(-2.31))≈0.91.

[0101] In order to obtain the initial output weight matrix, the generalized inverse matrix of the feature map matrix H is calculated based on the singular value decomposition. The matrix H is subjected to singular value decomposition to obtain H=USV^T, where U and V are the left singular vector and right singular vector matrices, respectively, and S is the singular value diagonal matrix. The singular values ​​in S that are less than the preset threshold 0.01 are set to zero to obtain the corrected singular value matrix S'. The inverse matrix S'^+ of S' is calculated, where the non-zero diagonal elements s' i The inverse of is 1 / s' i Construct the generalized inverse matrix H^+=VS'^+U^T, which is the initial output weight matrix β0. For example, if a singular value of the feature map matrix H is 0.008, it is set to zero because it is less than the threshold 0.01; for the singular value 0.05, its corresponding value in S'^+ is 1 / 0.05=20.

[0102] The system predicted output ŷ = Hβ0 is calculated based on the feature mapping matrix H and the initial output weight matrix β0. The system predicted output is a two-dimensional vector representing the predicted values ​​of the fan speed and circulating water flow rate, respectively. The difference between the system predicted output and the actual system output y is calculated to obtain the prediction error e = y - ŷ. The product ηe of the prediction error and the preset learning rate η (set to 0.01) is used as the parameter adjustment to update the initial output weight matrix to obtain the second output weight matrix β1 = β0 + ηeH^T. The second system output ŷ1 = Hβ1 is calculated based on the updated weight matrix. For example, if the predicted output ŷ = [45.3, 1150] and the actual output y = [46.2, 1180], the prediction error e = [0.9, 30], and the parameter adjustment is [0.009, 0.3] multiplied by H^T.

[0103] Calculate the gradient ∇ŷ1 of the second system output ŷ1 with respect to the system state x. This gradient represents the sensitivity of the system output to the input state. Multiply the exploration step size ε by the gradient ∇ŷ1 to obtain the initial increments Δω for the fan speed and ΔQ for the circulating water flow rate. For example, when ε = 1.35, the component of ∇ŷ1 with respect to the fan speed is 0.42, and the component with respect to the circulating water flow rate is 0.35, the initial increments Δω for the fan speed are 1.35 × 0.42 ≈ 0.57 (Hz), and the initial increments ΔQ for the circulating water flow rate are 1.35 × 0.35 ≈ 0.47 (m³ / h).

[0104] In this embodiment, the exponential decay factor is calculated by the performance indicator change trend and the adaptive adjustment coefficient, and then the exploration step is obtained, which realizes the adaptive optimization of the control parameter adjustment amplitude. The feature mapping matrix is ​​constructed by using the random projection matrix and the sigmoid activation function, which realizes the nonlinear mapping of the system state vector to the high-dimensional feature space. The output weight matrix is ​​dynamically updated by the prediction error and the learning rate, which realizes the online optimization of the model parameters. It can continuously adapt to the dynamic changes of the system and maintain the prediction accuracy. The exploration step is combined with the system output gradient to generate the control increment, which not only ensures the accuracy of the control direction, but also realizes the adaptive adjustment of the control amplitude, which can effectively improve the control performance of the system.

[0105] In an optional embodiment,

[0106] The collected real-time operation data is constructed into a data sequence in chronological order. The mutual information value of the data sequence is calculated to determine the optimal time delay. The false neighbor ratio is calculated to determine the minimum embedding dimension. The dynamic characteristic value sequence is extracted based on the optimal time delay and minimum embedding dimension. The steady-state operation reference trajectory is generated and the deviation value between it and the current operation trajectory is calculated. The correction parameter is generated based on the deviation value and superimposed on the initial incremental value of the fan speed and the initial incremental value of the circulating water flow. The target adjustment values ​​corresponding to the fan speed and the circulating water flow are obtained, including:

[0107] Constructing a data sequence from the real-time operation data in chronological order, calculating a marginal probability distribution and a joint probability distribution of the data sequence, calculating a mutual information value based on the marginal probability distribution and the joint probability distribution, and determining a time value corresponding to a first local minimum of the mutual information value as an optimal time delay;

[0108] Calculating the ratio of the number of false neighbor points under different embedding dimensions to the number of sampling points of the data sequence based on the false neighbor algorithm to obtain a false neighbor ratio, and determining the dimension corresponding to the first time the false neighbor ratio is less than a preset threshold as the minimum embedding dimension;

[0109] Combining the data sequence according to the optimal time delay and the minimum embedding dimension to construct a current running trajectory, and calculating the Euclidean distance between adjacent data points to obtain a dynamic characteristic value sequence;

[0110] generating a steady-state operation reference trajectory based on the historical steady-state operation data according to the optimal time delay and the minimum embedding dimension, calculating the Euclidean distance between the current operation trajectory and the steady-state operation reference trajectory to obtain a spatial deviation value, and calculating the difference between the dynamic characteristic value sequence and the dynamic characteristic value sequence corresponding to the historical steady-state operation data to obtain a dynamic deviation value;

[0111] The product of the spatial deviation value and the preset proportional coefficient is added to the product of the dynamic deviation value and the preset differential coefficient to obtain a correction parameter, the product of the correction parameter and the gradient of the spatial deviation value with respect to the fan speed is superimposed on the initial increment of the fan speed to obtain the target adjustment value corresponding to the fan speed, and the product of the correction parameter and the gradient of the spatial deviation value with respect to the circulating water flow is superimposed on the initial increment of the circulating water flow to obtain the target adjustment value corresponding to the circulating water flow.

[0112] Collect real-time operating data, such as fan speed, circulating water flow, temperature, pressure, and other parameters. Arrange this data in chronological order to construct a data sequence X(t), where t represents a point in time and X(t) represents the parameter value at that point in time. For example, during the operation of a certain power plant, the collected fan speed data is {1200, 1210, 1205, 1215, 1220, ...} rpm, with a sampling interval of 5 seconds.

[0113] To determine the optimal time delay, the mutual information value of the data sequence is calculated. For the data sequence X(t) and the delay sequence X(t+τ), the marginal probability distribution and the joint probability distribution are calculated. For example, the data range is divided into several intervals, and the number of data points in each interval is counted and divided by the total number of data points to obtain the probability distribution. For a certain wind turbine operating data, when the delay time τ is 5, 10, 15, and 20 seconds, respectively, the calculated mutual information values ​​are {0.85, 0.62, 0.58, 0.63}. It is observed that when τ = 15 seconds, the mutual information value reaches the first local minimum of 0.58, so the optimal time delay is determined to be 15 seconds.

[0114] Using the false neighbor algorithm, when the embedding dimension is d, find the nearest neighbor points of each point in the d-dimensional space, and check whether these points are still nearest neighbors in the d+1-dimensional space. If not, they are considered to be false neighbor points. The ratio of the number of false neighbor points to the total number of sampling points is calculated to obtain the false neighbor ratio. For example, for the temperature data of the circulating water system, when the embedding dimensions are 2, 3, 4, and 5 respectively, the false neighbor ratio is {0.15, 0.08, 0.03, 0.02}. Set the preset threshold to 0.05. When the embedding dimension is 4, the false neighbor ratio is less than the threshold for the first time, so the minimum embedding dimension is determined to be 4.

[0115] Using the determined optimal time delay and minimum embedding dimension, the data sequence is reconstructed into a trajectory in a high-dimensional phase space. For the time series {x(t1), x(t2), ..., x(t n )}, construct the vector {X(t1), X(t2), ..., X(t m )}, where X(t i )={x(t i ), x(t i +τ),...,x(t i +(d-1)τ)}, τ is the optimal time delay and d is the minimum embedding dimension. For example, when τ = 15 seconds and d = 4, the first reconstructed vector is {x(t1), x(t1+15), x(t1+30), x(t1+45)}.

[0116] The Euclidean distances between adjacent data points in the reconstructed phase space are calculated to obtain a sequence of dynamical eigenvalues. For example, for the reconstructed vector sequence, the Euclidean distances between each pair of adjacent vectors are calculated to obtain the sequence of dynamical eigenvalues ​​{5.2, 4.8, 6.1, 5.5, ...}.

[0117] Based on historical steady-state operating data, a steady-state reference trajectory is generated using the same optimal time delay and minimum embedding dimension. For example, data from the historical database showing normal equipment operation is extracted and reconstructed to create the reference trajectory. The Euclidean distance between the current operating trajectory and the reference trajectory is calculated to obtain the spatial deviation value. Assuming the calculated spatial deviation value is 8.3, this indicates that the current operating state deviates from the ideal steady-state state.

[0118] The difference between the current dynamic characteristic value sequence and the dynamic characteristic value sequence corresponding to the historical steady-state operation is calculated to obtain the dynamic deviation value. For example, if the current dynamic characteristic value is 5.8 and the corresponding historical steady-state value is 5.2, the dynamic deviation value is 0.6, which indicates the deviation in the dynamic change rate of the system.

[0119] Set the preset proportional coefficient Kp = 0.5 and the differential coefficient Kd = 0.3. The calculated correction parameter is the product of the spatial deviation value and the proportional coefficient, plus the product of the dynamic deviation value and the differential coefficient. That is, the correction parameter = 8.3 × 0.5 + 0.6 × 0.3 = 4.33.

[0120] Calculate the gradient of the spatial deviation value with respect to the fan speed. For example, if the fan speed needs to be adjusted by 2 rpm for each unit of spatial deviation, the gradient is 2. Add the product of the correction parameter and the gradient (4.33 × 2 = 8.66) to the initial fan speed increment (assuming it is 10 rpm) to obtain the target fan speed adjustment value of 10 + 8.66 = 18.66 rpm, indicating that the fan speed needs to be increased by 18.66 rpm.

[0121] Calculate the gradient of the spatial deviation value to the circulating water flow rate. For example, the circulating water flow rate needs to be adjusted by 0.5m for each unit of spatial deviation value. 3 / h, the gradient is 0.5. Add the product of the correction parameter and the gradient 4.33×0.5=2.17 to the initial incremental value of the circulating water flow (assuming 3m 3 / h), the target adjustment value of circulating water flow is 3+2.17=5.17m 3 / h, that is, the circulating water flow rate needs to be increased by 5.17m 3 / h.

[0122] In this embodiment, the optimal time delay is determined by calculating the mutual information value, and the optimal time scale selection for time series reconstruction is achieved. The delay time optimization method based on information theory ensures the optimal balance between the independence and correlation between the time series components in the reconstructed phase space. The false nearest neighbor algorithm is used to dynamically determine the minimum embedding dimension, avoiding the increase in computational complexity caused by selecting too high a dimension or the information loss caused by selecting too low a dimension, and realizing adaptive optimization of the phase space reconstruction dimension to ensure the reconstruction quality. At the same time, the spatial deviation value and the dynamic deviation value are calculated to achieve a comprehensive evaluation of the static and dynamic characteristics of the system state. The correction parameter is combined with the deviation gradient to generate the target adjustment value, and the coordinated adjustment of the fan speed and the circulating water flow is achieved, ensuring the directional accuracy and amplitude rationality of the control parameter adjustment.

[0123] Figure 2 This is a diagram showing the curve of the false neighbor ratio of the power plant cooling system changing with the embedding dimension. Figure 2 As shown in the figure, the solid line represents the changing trend of the false nearest neighbor ratio (FNN) with the embedding dimension, and the hollow circle is the minimum embedding dimension point. As the dimension increases, the false nearest neighbor ratio gradually decreases, indicating that the system state gradually unfolds in the high-dimensional space.

[0124] When the embedding dimension is 1, the FNN ratio reaches a high of 0.32, indicating significant overlap of system states in one-dimensional space. As the dimension increases to 6, the FNN ratio drops to 0.03, falling below the preset threshold of 0.05 (dashed line) for the first time. Therefore, the minimum embedding dimension is determined to be 6. At this point, the system states are fully expanded in the 6-dimensional phase space, accurately reflecting the system's dynamic characteristics. This method, based on the false nearest neighbor algorithm, avoids the increased computational complexity caused by selecting too high a dimensionality or the information loss caused by selecting too low a dimensionality, achieving adaptive optimization of the phase space reconstruction dimensionality.

[0125] In an optional embodiment,

[0126] Calculating the ratio of the number of false neighbor points under different embedding dimensions to the number of sampling points of the data sequence based on the false neighbor algorithm to obtain a false neighbor ratio, and determining the dimension corresponding to when the false neighbor ratio is less than a preset threshold for the first time as the minimum embedding dimension includes:

[0127] The data sequence is reconstructed into a phase space of a preset dimension. The Euclidean distance between the reconstructed vectors is calculated to obtain a distance matrix. The difference between the reconstructed vectors is calculated to obtain a direction vector matrix. The local covariance matrix is ​​calculated based on the reconstructed vectors. The local covariance matrix is ​​subjected to eigenvalue decomposition to obtain an eigenvector matrix and an eigenvalue matrix.

[0128] The distance ratio criterion is obtained by calculating the distance ratio of the reconstructed vector after increasing the dimension. The direction consistency criterion is obtained by calculating the inner product of the eigenvector corresponding to the maximum eigenvalue in the direction vector matrix and the eigenvector matrix. The local density ratio of adjacent reconstructed vectors is obtained by calculating the local density criterion.

[0129] Calculating the ratio of the maximum eigenvalue to the second largest eigenvalue in the eigenvalue matrix to obtain the local manifold curvature, calculating the distance ratio criterion weight and the direction consistency criterion weight based on the local manifold curvature in combination with an exponential function, taking the complement of the sum of the distance ratio criterion weight and the direction consistency criterion weight as the local density criterion weight, and calculating the spatial position weight based on the local manifold curvature, wherein the spatial position weight is inversely proportional to the local manifold curvature;

[0130] The product of the distance ratio criterion and the distance ratio criterion weight, the product of the directional consistency criterion and the directional consistency criterion weight, and the product of the local density criterion and the local density criterion weight are added together to obtain a comprehensive criterion value. The comprehensive criterion value is compared with the preset threshold to obtain the false neighbor judgment result. The product of the false neighbor judgment result and the spatial position weight is calculated to obtain the false neighbor ratio. The dimension corresponding to the first time the false neighbor ratio is less than the preset threshold is determined as the minimum embedding dimension.

[0131] Obtain the data sequence to be analyzed. A data sequence is a signal collected by a sensor, such as a sequence of time-varying physical quantities like temperature, pressure, and speed. For example, a vibration signal collected during the operation of a device might contain 10,000 sampling points at a sampling frequency of 1000 Hz.

[0132] For the acquired data sequence, a reconstruction vector is constructed in the phase space of a preset dimension. The preset dimension can be set to a larger value, such as 10 dimensions, so that the minimum possible embedding dimension can be covered in the subsequent calculation process. The reconstruction vector is constructed using the time delay method, selecting an appropriate time delay value τ (such as τ = 10 determined by the mutual information method), and taking every τ point in the original data sequence as a component of the reconstruction vector. For example, for the data sequence {x1, x2, ..., x 10000}, the first m-dimensional reconstruction vector is {x1, x 1+τ , x 1+2τ ,...,x 1+(m-1)τ}.

[0133] After constructing the reconstruction vectors, calculate the Euclidean distance between them to obtain the distance matrix D. For any two reconstruction vectors X_i and X_j, their Euclidean distance is the square root of the sum of the squares of the differences between their components. For example, the distance D(i, j) between X_i and X_j is 0.85 units. Calculating the differences between the reconstruction vectors yields the direction vector matrix V. A direction vector represents a vector pointing from one reconstruction vector to another.

[0134] Calculate the local covariance matrix C based on the reconstructed vectors. For each reconstructed vector X_i, select its k nearest neighbors (e.g., k = 20) and calculate the covariance relationship between these points and X_i to form the local covariance matrix C_i. Perform eigenvalue decomposition on the local covariance matrix to obtain the eigenvector matrix E and the eigenvalue matrix Λ. The eigenvalues ​​are arranged in descending order. For example, the largest eigenvalue λ1 of a point is 2.35, and the second largest eigenvalue λ2 is 1.12.

[0135] The distance ratio criterion R_d is calculated by calculating the distance ratio of the reconstructed vectors after increasing the dimension. When increasing the dimension from m to m+1, for each pair of neighboring points X_i and X_j, calculate the ratio of their new distance in m+1 dimensions to their distance in m dimensions. For example, if the distance ratio for a pair of points is 1.75, it means that the distance has increased significantly in the higher dimension.

[0136] Compute the inner product of the direction vector matrix and the eigenvector corresponding to the largest eigenvalue in the eigenvector matrix to obtain the directional consistency criterion R_a. This criterion measures the consistency of the direction vector with the local principal direction and takes a value in the range [-1, 1]. A value of 0.92 indicates high directional consistency.

[0137] The local density criterion R_d is calculated by calculating the local density ratio of adjacent reconstructed vectors. For each reconstructed vector X_i, select its k nearest neighbors and calculate the point density within the sphere radius r. Then, compare the density ratio of X_i to its nearest neighbor X_j in the corresponding local area. For example, a density ratio of 0.78 indicates little density change.

[0138] The ratio of the largest eigenvalue to the second largest eigenvalue in the eigenvalue matrix is ​​calculated to obtain the local manifold curvature κ. For example, κ = 2.10 indicates that the local region exhibits obvious one-dimensional manifold characteristics. Based on the local manifold curvature κ, an exponential function is used to calculate the distance ratio criterion weight w_d and the directional consistency criterion weight w_a. For example, w_d = 0.6 and w_a = 0.3. The complement of the sum of the distance ratio criterion weight and the directional consistency criterion weight is used as the local density criterion weight w_ρ. In this example, w_ρ = 0.1.

[0139] The spatial position weight w_s is calculated based on the local manifold curvature κ. The spatial position weight is inversely proportional to the local manifold curvature; the greater the curvature, the smaller the weight. For example, when κ = 2.10, we get w_s = 0.48.

[0140] Add the product of the distance ratio criterion R_d and its weight w_d, the product of the directional consistency criterion R_a and its weight w_a, and the product of the local density criterion R_ρ and its weight w_ρ to obtain the comprehensive criterion value J. For example, J = 0.85. Compare the comprehensive criterion value J with the preset threshold θ. For example, if θ = 0.7, then J > θ, and the point is determined to be a false neighbor. Calculate the product of the false neighbor determination result and the spatial position weight w_s to obtain the weighted false neighbor result.

[0141] Repeat the above calculation for all pairs of neighbor points, and calculate the ratio of the number of false neighbor points to the number of sample points in the data sequence to obtain the false neighbor ratio FNN for the current dimension. For example, when m = 3 dimensions, FNN = 0.12, and when m = 4 dimensions, FNN = 0.03. Set the false neighbor ratio threshold ε = 0.05. The dimension corresponding to the first time FNN falls below ε is determined as the minimum embedding dimension. In this example, the minimum embedding dimension is 4.

[0142] In this embodiment, by constructing a distance matrix, a direction vector matrix and a local covariance matrix, feature extraction of the reconstructed vector in multiple dimensions such as Euclidean distance, direction and local distribution is achieved, local manifold features are extracted based on eigenvalue decomposition, and a complete local structure evaluation system is constructed through distance ratio criterion, direction consistency criterion and local density criterion, which can accurately identify local geometric structure features in phase space, introduce spatial position weights that are inversely proportional to the local manifold curvature, reduce the weight of false neighbor judgment in areas with larger manifold curvature, and improve the accuracy of dimensionality estimation. By determining the proportion of false neighbors through a combination of comprehensive criterion values ​​and spatial position weights, a robust estimation of the minimum embedding dimension is achieved, effectively avoiding the interference of local features on dimensionality estimation, and improving the reliability of dimensionality determination.

[0143] In an optional embodiment,

[0144] Obtaining the minimum interval time and maximum adjustment range of the control parameters, performing constraint verification on the target adjustment value, generating an adjustment instruction sequence that meets the constraints, and executing system adjustment include:

[0145] Obtaining adjustment constraints of the control parameters, wherein the adjustment constraints include a minimum interval time and a maximum adjustment range;

[0146] Dividing the target adjustment value into time interval segments based on the minimum interval time, and constraining the amplitude of the segmented target adjustment value based on the maximum adjustment amplitude to obtain an adjustment value sequence that meets the constraint;

[0147] The adjustment value sequence is converted into an adjustment instruction sequence in chronological order, and the adjustment instruction sequence is executed to realize system adjustment.

[0148] Obtain the adjustment constraints of the control parameters. The constraints mainly include the minimum interval time and the maximum adjustment range. The minimum interval time refers to the shortest time interval that must be maintained between two consecutive adjustment operations. This parameter is usually determined by hardware characteristics or business requirements. For example, in a temperature control system, the actuator requires at least 10 seconds to complete an adjustment action and stabilize, so the minimum interval time is set to 10 seconds. The maximum adjustment range refers to the maximum change allowed in a single adjustment, which is usually determined by the system safety or stability requirements. For example, in the same temperature control system, to avoid system oscillation, a single temperature adjustment should not exceed 2°C. These parameters can be obtained by reading the system configuration file, querying the database, or entering through the user interface.

[0149] After obtaining the constraints, the target adjustment value is segmented into time intervals based on the minimum interval. The current system state and target adjustment value are obtained, and the total adjustment amount is calculated. Assume that the current temperature is 25°C and the target temperature is 18°C, with a total adjustment amount of 7°C. The adjustment process is divided into multiple time periods based on the minimum interval and the estimated total time required to complete the adjustment. If the minimum interval is 10 seconds and the estimated total time is 70 seconds, the adjustment process can be divided into seven time periods, each corresponding to one adjustment operation. Generate a sequence of time points, such as [t0, t1, t2, t3, t4, t5, t6, t7], where t0 is the start time and the interval between adjacent time points is at least 10 seconds.

[0150] After time segmentation, the target adjustment value after segmentation is constrained based on the maximum adjustment amplitude to obtain a sequence of adjustment values ​​that meet the constraints. The ideal adjustment value is calculated for each time segment. If the total adjustment amount is evenly distributed, each time segment should be adjusted by 1°C. However, due to the maximum adjustment amplitude limit (for example, 2°C), it is necessary to ensure that the adjustment amount in each time segment does not exceed this limit. Furthermore, considering the nonlinear characteristics of actual adjustment, different adjustment strategies need to be adopted at different stages.

[0151] By comprehensively considering both time and amplitude constraints, a sequence of adjustment values ​​that satisfies these constraints is generated. For example, for a temperature drop from 25°C to 18°C, with a maximum adjustment amplitude of 2°C, the following adjustment value sequence is generated: [25°C, 23°C, 21°C, 20°C, 19°C, 18.5°C, 18°C]. This ensures that each adjustment does not exceed the maximum adjustment amplitude, and that the adjustment amplitude gradually decreases as the target value is approached to avoid overshoot.

[0152] Once the adjustment value sequence is determined, it is converted into a sequence of adjustment instructions in chronological order. Each adjustment instruction contains the execution time and the target adjustment value. For example, an instruction might be "adjust the temperature to 23°C at time t1." These abstract instructions also need to be converted into commands executable by specific devices, which may include information such as the device number, operation code, and parameter values.

[0153] The generated sequence of adjustment instructions is executed to achieve system adjustment. During the execution process, the corresponding adjustment instructions are sent to the execution device at the preset time points. At the same time, the actual execution effect is continuously monitored and feedback data is collected. If the actual adjustment effect deviates from the expected one, the subsequent instruction sequence is adjusted in real time. For example, if the actual temperature at a certain point in time is 22°C instead of the expected 23°C, the subsequent instructions are adjusted accordingly to ensure that the target value of 18°C ​​is ultimately achieved.

[0154] In this embodiment, by setting the constraints of the minimum interval time and the maximum adjustment amplitude, system fluctuations caused by too frequent adjustment or too large an adjustment amplitude are effectively prevented. Time segmentation is performed based on the minimum interval time, and amplitude constraints are performed in combination with the maximum adjustment amplitude, thereby realizing segmented optimization of the adjustment process. The adjustment value sequence that meets the constraints is converted into an adjustment instruction sequence, ensuring the timing and standardization of the control instructions, effectively reducing the mechanical stress of the equipment, and extending the service life of the equipment.

[0155] A second aspect of an embodiment of the present invention provides an intelligent control system for a power plant cooling system based on multi-source data, comprising:

[0156] The first unit is used to collect real-time operating data of the power plant cooling system, power plant load forecast data and weather forecast data;

[0157] The second unit is used to calculate the system performance index by combining cooling efficiency and energy consumption, perform parameter sensitivity analysis on pre-acquired historical operating data to obtain weight coefficients, combine power plant load forecast data and weather forecast data to calculate the performance index change trend for the next four hours, calculate the current performance score based on real-time operating data and weight coefficients, dynamically update the exploration step size based on the current performance score and performance index change trend, and iteratively calculate the initial control parameters using adaptive reinforcement learning. The initial incremental values ​​of fan speed and circulating water flow are output;

[0158] The third unit is used to construct a data sequence from the collected real-time operation data in chronological order, calculate the mutual information value of the data sequence to determine the optimal time delay, calculate the false neighbor ratio to determine the minimum embedding dimension, extract the dynamic characteristic value sequence based on the optimal time delay and the minimum embedding dimension, generate a steady-state operation reference trajectory and calculate the deviation value between it and the current operation trajectory, generate a correction parameter based on the deviation value, and superimpose the correction parameter with the initial incremental value of the fan speed and the initial incremental value of the circulating water flow to obtain the target adjustment value corresponding to the fan speed and the circulating water flow;

[0159] The fourth unit is used to obtain the minimum interval time and maximum adjustment range of the control parameters, perform constraint verification on the target adjustment value, generate an adjustment instruction sequence that meets the constraints and execute system adjustment.

[0160] According to a third aspect of an embodiment of the present invention, an electronic device is provided, including:

[0161] A processor and a memory for storing processor-executable instructions, wherein the processor is configured to call the instructions stored in the memory to execute the aforementioned method.

[0162] According to a fourth aspect of an embodiment of the present invention, a computer-readable storage medium is provided, on which computer program instructions are stored. When the computer program instructions are executed by a processor, the method described above is implemented.

[0163] The present invention may be a method, an apparatus, a system and / or a computer program product. The computer program product may include a computer-readable storage medium carrying computer-readable program instructions for executing various aspects of the present invention.

[0164] Finally, it should be noted that the above embodiments are only used to illustrate the technical solutions of the present invention, rather than to limit it. Although the present invention has been described in detail with reference to the above embodiments, those skilled in the art should understand that they can still modify the technical solutions described in the above embodiments, or replace some or all of the technical features therein with equivalents. However, these modifications or replacements do not cause the essence of the corresponding technical solutions to deviate from the scope of the technical solutions of the embodiments of the present invention.

Claims

1. An intelligent control method for a power plant cooling system based on multi-source data, characterized in that: include: Collect real-time operating data of the power plant cooling system, power plant load forecast data and weather forecast data; The system performance index is calculated by combining cooling efficiency and energy consumption. A parameter sensitivity analysis is performed on pre-acquired historical operating data to obtain weight coefficients. The performance index change trend for the next four hours is calculated by combining power plant load forecast data and weather forecast data. The current performance score is calculated based on real-time operating data and the weight coefficients. The exploration step size is dynamically updated based on the current performance score and performance index change trend. Initial control parameters are iteratively calculated using adaptive reinforcement learning, and the initial incremental values ​​for fan speed and circulating water flow are output. The collected real-time operation data is constructed into a data sequence in chronological order, the mutual information value of the data sequence is calculated to determine the optimal time delay, the false neighbor ratio is calculated to determine the minimum embedding dimension, and the dynamic characteristic value sequence is extracted based on the optimal time delay and the minimum embedding dimension. The steady-state operation reference trajectory is generated and the deviation value between it and the current operation trajectory is calculated. The correction parameter is generated based on the deviation value and superimposed with the initial incremental value of the fan speed and the initial incremental value of the circulating water flow to obtain the target adjustment value corresponding to the fan speed and the circulating water flow; The minimum interval time and maximum adjustment range of the control parameters are obtained, the target adjustment value is constrained and verified, an adjustment instruction sequence that meets the constraints is generated, and the system adjustment is executed.

2. The method according to claim 1, characterized in that Collecting real-time operating data of the power plant cooling system, power plant load forecast data and weather forecast data including: Collecting power plant cooling system operating data, including cooling tower outlet water temperature, cooling tower inlet water temperature, circulating water flow, ambient temperature, ambient humidity, and fan speed; Collecting power plant load forecast data, the load forecast data including load change forecast values ​​for the next 4 hours; Collect weather forecast data, which includes predicted values ​​of ambient temperature and ambient humidity for the next four hours.

3. The method according to claim 1, characterized in that The system performance index is calculated by combining cooling efficiency and energy consumption. A parameter sensitivity analysis is performed on pre-acquired historical operating data to obtain weight coefficients. The performance index change trend for the next four hours is calculated by combining power plant load forecast data and weather forecast data. The current performance score is calculated based on real-time operating data and weight coefficients, including: The cooling tower inlet water temperature, cooling tower outlet water temperature and wet bulb temperature are calculated to obtain the cooling efficiency. The cumulative value of the fan power and the water pump power is calculated to obtain the system energy consumption. The system performance index is constructed based on the cooling efficiency and system energy consumption. Perform principal component analysis and dimensionality reduction on historical operation data to obtain a characteristic variable matrix. Calculate the partial correlation coefficient between each characteristic variable in the characteristic variable matrix and the system performance index. Use the ratio of the absolute value of the partial correlation coefficient to the sum of the absolute values ​​of the partial correlation coefficients as the weight coefficient. A sliding time window method is used to construct a performance indicator prediction sequence. The previously acquired load change prediction value, ambient temperature prediction value, and ambient humidity prediction value are substituted into the performance indicator prediction sequence to calculate the performance indicator change trend for the next 4 hours. Extract the equipment running time sequence and the corresponding performance indicator sequence, use the least squares method to fit the performance attenuation baseline curve, construct an adaptive learning rate based on the fluctuation variance of the performance indicator, project the deviation between the short-term change trend and the long-term attenuation trend of the performance indicator into the Gaussian kernel space, obtain the performance attenuation correction coefficient through iterative calculation of the kernel function, calculate the current performance indicator value based on the real-time operation data, and multiply the weighted combination result of the weight coefficient and the current performance indicator value by the performance attenuation correction coefficient to obtain the current performance score.

4. The method according to claim 1, wherein The exploration step size is dynamically updated based on the current performance score and performance indicator change trend. The initial control parameters are iteratively calculated using adaptive reinforcement learning. The output initial incremental values ​​of the fan speed and circulating water flow include: The absolute value of the performance indicator change trend is multiplied by a preset adaptive adjustment coefficient to obtain an exponential decay factor, and the exploration step size is calculated based on the exponential decay factor and the current performance score; Generate a random projection matrix using a standard normal distribution, generate a bias vector based on the normal distribution, multiply a system state vector constructed based on the power plant cooling system operating data by the random projection matrix and add the bias vector to obtain a feature mapping matrix using a sigmoid activation function; Calculating the generalized inverse matrix of the feature map matrix based on singular value decomposition, setting the singular values ​​smaller than a preset threshold to zero, and then reconstructing to obtain an initial output weight matrix; Calculating a system predicted output based on the feature mapping matrix and the initial output weight matrix, calculating a difference between the system predicted output and the actual system output to obtain a prediction error, multiplying the prediction error by a preset learning rate as a parameter adjustment, updating the initial output weight matrix to obtain a second output weight matrix, and calculating a second system output; The exploration step size is multiplied by the gradient output by the second system to obtain an initial incremental value of the fan speed and an initial incremental value of the circulating water flow rate.

5. The method according to claim 1, wherein The collected real-time operation data is constructed into a data sequence in chronological order. The mutual information value of the data sequence is calculated to determine the optimal time delay. The false neighbor ratio is calculated to determine the minimum embedding dimension. The dynamic characteristic value sequence is extracted based on the optimal time delay and minimum embedding dimension. The steady-state operation reference trajectory is generated and the deviation value between it and the current operation trajectory is calculated. The correction parameter is generated based on the deviation value and superimposed on the initial incremental value of the fan speed and the initial incremental value of the circulating water flow. The target adjustment values ​​corresponding to the fan speed and the circulating water flow are obtained, including: Constructing a data sequence from the real-time operation data in chronological order, calculating a marginal probability distribution and a joint probability distribution of the data sequence, calculating a mutual information value based on the marginal probability distribution and the joint probability distribution, and determining a time value corresponding to a first local minimum of the mutual information value as an optimal time delay; Calculating the ratio of the number of false neighbor points under different embedding dimensions to the number of sampling points of the data sequence based on the false neighbor algorithm to obtain a false neighbor ratio, and determining the dimension corresponding to the first time the false neighbor ratio is less than a preset threshold as the minimum embedding dimension; Combining the data sequence according to the optimal time delay and the minimum embedding dimension to construct a current running trajectory, and calculating the Euclidean distance between adjacent data points to obtain a dynamic characteristic value sequence; generating a steady-state operation reference trajectory based on the historical steady-state operation data according to the optimal time delay and the minimum embedding dimension, calculating the Euclidean distance between the current operation trajectory and the steady-state operation reference trajectory to obtain a spatial deviation value, and calculating the difference between the dynamic characteristic value sequence and the dynamic characteristic value sequence corresponding to the historical steady-state operation data to obtain a dynamic deviation value; The product of the spatial deviation value and the preset proportional coefficient is added to the product of the dynamic deviation value and the preset differential coefficient to obtain a correction parameter, the product of the correction parameter and the gradient of the spatial deviation value with respect to the fan speed is superimposed on the initial increment of the fan speed to obtain the target adjustment value corresponding to the fan speed, and the product of the correction parameter and the gradient of the spatial deviation value with respect to the circulating water flow is superimposed on the initial increment of the circulating water flow to obtain the target adjustment value corresponding to the circulating water flow.

6. The method according to claim 5, characterized in that Calculating the ratio of the number of false neighbor points under different embedding dimensions to the number of sampling points of the data sequence based on the false neighbor algorithm to obtain a false neighbor ratio, and determining the dimension corresponding to when the false neighbor ratio is less than a preset threshold for the first time as the minimum embedding dimension includes: The data sequence is reconstructed into a phase space of a preset dimension. The Euclidean distance between the reconstructed vectors is calculated to obtain a distance matrix. The difference between the reconstructed vectors is calculated to obtain a direction vector matrix. The local covariance matrix is ​​calculated based on the reconstructed vectors. The local covariance matrix is ​​subjected to eigenvalue decomposition to obtain an eigenvector matrix and an eigenvalue matrix. The distance ratio criterion is obtained by calculating the distance ratio of the reconstructed vector after increasing the dimension. The direction consistency criterion is obtained by calculating the inner product of the eigenvector corresponding to the maximum eigenvalue in the direction vector matrix and the eigenvector matrix. The local density ratio of adjacent reconstructed vectors is obtained by calculating the local density criterion. Calculating the ratio of the maximum eigenvalue to the second largest eigenvalue in the eigenvalue matrix to obtain the local manifold curvature, calculating the distance ratio criterion weight and the direction consistency criterion weight based on the local manifold curvature in combination with an exponential function, taking the complement of the sum of the distance ratio criterion weight and the direction consistency criterion weight as the local density criterion weight, and calculating the spatial position weight based on the local manifold curvature, wherein the spatial position weight is inversely proportional to the local manifold curvature; The product of the distance ratio criterion and the distance ratio criterion weight, the product of the directional consistency criterion and the directional consistency criterion weight, and the product of the local density criterion and the local density criterion weight are added together to obtain a comprehensive criterion value. The comprehensive criterion value is compared with the preset threshold to obtain the false neighbor judgment result. The product of the false neighbor judgment result and the spatial position weight is calculated to obtain the false neighbor ratio. The dimension corresponding to the first time the false neighbor ratio is less than the preset threshold is determined as the minimum embedding dimension.

7. The method according to claim 1, characterized in that Obtaining the minimum interval time and maximum adjustment range of the control parameters, performing constraint verification on the target adjustment value, generating an adjustment instruction sequence that meets the constraints, and executing system adjustment include: Obtaining adjustment constraints of the control parameters, wherein the adjustment constraints include a minimum interval time and a maximum adjustment range; Dividing the target adjustment value into time interval segments based on the minimum interval time, and constraining the amplitude of the segmented target adjustment value based on the maximum adjustment amplitude to obtain an adjustment value sequence that meets the constraint; The adjustment value sequence is converted into an adjustment instruction sequence in chronological order, and the adjustment instruction sequence is executed to realize system adjustment.

8. An intelligent control system for a power plant cooling system based on multi-source data, used to implement the method according to any one of claims 1 to 7, characterized in that: include: The first unit is used to collect real-time operating data of the power plant cooling system, power plant load forecast data and weather forecast data; The second unit is used to calculate the system performance index by combining cooling efficiency and energy consumption, perform parameter sensitivity analysis on pre-acquired historical operating data to obtain weight coefficients, combine power plant load forecast data and weather forecast data to calculate the performance index change trend for the next four hours, calculate the current performance score based on real-time operating data and weight coefficients, dynamically update the exploration step size based on the current performance score and performance index change trend, and iteratively calculate the initial control parameters using adaptive reinforcement learning. The initial incremental values ​​of fan speed and circulating water flow are output; The third unit is used to construct a data sequence from the collected real-time operation data in chronological order, calculate the mutual information value of the data sequence to determine the optimal time delay, calculate the false neighbor ratio to determine the minimum embedding dimension, extract the dynamic characteristic value sequence based on the optimal time delay and the minimum embedding dimension, generate a steady-state operation reference trajectory and calculate the deviation value between it and the current operation trajectory, generate a correction parameter based on the deviation value, and superimpose the correction parameter with the initial incremental value of the fan speed and the initial incremental value of the circulating water flow to obtain the target adjustment value corresponding to the fan speed and the circulating water flow; The fourth unit is used to obtain the minimum interval time and maximum adjustment range of the control parameters, perform constraint verification on the target adjustment value, generate an adjustment instruction sequence that meets the constraints and execute system adjustment.

9. An electronic device, characterized in that: include: processor; a memory for storing processor-executable instructions; The processor is configured to call the instructions stored in the memory to execute the method according to any one of claims 1 to 7.

10. A computer-readable storage medium having computer program instructions stored thereon, characterized in that: When the computer program instructions are executed by a processor, the method according to any one of claims 1 to 7 is implemented.

Images