Cooling tower floating water loss prediction and analysis method

Abstract

The invention provides a cooling tower floating water loss prediction analysis method, which comprises the following steps: acquiring and cleaning multi-dimensional operation data of a cooling tower, performing working condition label classification, feature normalization and derivative feature extension, and constructing a deep neural network model to realize floating water loss prediction after dynamically adjusting a feature weight based on a working condition classification clustering result. And in the online stage, real-time working condition labels and historical error feedback are combined, and the feature weights and model parameters are dynamically optimized, so that the generalization ability and prediction accuracy of the model under variable working conditions are improved, the engineering applicability and the online adaptive optimization ability are high, and water saving and intelligent operation management of the cooling tower system are effectively supported.

Description

Technical Field

[0001] This invention relates to the field of cooling tower operation data analysis and water loss prediction technology, and in particular to a method for predicting and analyzing cooling tower runoff water loss. Background Technology

[0002] Currently, cooling tower runoff loss prediction and operating condition analysis technology is an important research direction in the field of industrial intelligent operation and maintenance, belonging to the cutting-edge technology of multidisciplinary integration, including industrial process data-driven, intelligent operating condition identification, and model adaptive prediction. In the operation and management of industrial cooling towers, accurate prediction of runoff loss is not only related to water conservation, energy efficiency improvement, and environmental compliance, but also plays a key role in equipment health status assessment and operation and maintenance decisions. Current technologies commonly employ deep learning, time series analysis, or multiple regression to predict cooling tower operational data for loss assessment. Mainstream solutions typically build offline-trained models based on historical datasets, using physical quantities such as ambient temperature, humidity, wind speed, cooling water flow rate, and packing parameters as multi-dimensional features. They then utilize fully connected neural networks (DNNs), long short-term memory networks (LSTMs), or gradient boosting trees (GBDTs) to model and predict drift water loss. Meanwhile, industry trends are gradually shifting towards data-driven operational decision-making, operational condition zoning management, and intelligent feedback optimization. Emerging research is also exploring technologies such as transfer learning, model ensembles, and cloud-edge collaborative optimization to improve prediction accuracy and broaden application scope. However, based on existing technologies, current cooling tower runoff loss prediction models still have the following prominent problems in dealing with complex and variable operating conditions, long-term continuous operation, and real-time adaptation in the field: (1) Insufficient generalization ability. Common prediction models usually adopt the "fixed weight + static operating condition assumption" approach, which only performs well within a specific operating condition range constructed from historical operating data. Once the field environment changes due to factors such as season, load, and sudden changes in equipment status, the model is prone to inaccuracy, drift, or even complete failure.

[0003] (2) Lack of online self-learning and dynamic adaptation: Most existing intelligent prediction models adopt batch offline training and periodic manual tuning and updates, which cannot achieve rapid and automatic adaptation of model prediction parameters to the continuous changes in the actual operating state of the cooling tower. This results in prediction results lagging behind the evolution of real operating conditions, making it difficult to guarantee high-precision prediction and process control in the long term.

[0004] (3) The feature weights are rigid and have weak resistance to sudden changes. The current model structure generally lacks a dynamic adjustment mechanism for the contribution of input features under different working conditions. The "one-size-fits-all" configuration of input features makes it difficult to highlight the actual impact of the weights of ambient temperature, flow rate, and filler in different working conditions. When faced with critical working condition jumps or extreme scenarios, the prediction accuracy and robustness of the model decrease significantly.

[0005] (4) The working condition label is disconnected from the feature contribution. Although some technologies introduce the concept of working condition classification or clustering partitioning, most of them are limited to data partitioning assistance. They lack the ability to effectively integrate working condition identification and dynamic weighting of input features into the entire process of model training and inference. As a result, it is impossible to achieve real-time optimization of feature contribution driven by working conditions, which limits the model's rapid adaptation to new working conditions and sudden change scenarios. Summary of the Invention

[0006] In order to solve the above-mentioned technical problems, the present invention provides a method for predicting and analyzing the loss of cooling tower drift water.

[0007] The technical solution of this invention is implemented as follows: A method for predicting and analyzing cooling tower runoff water loss, comprising: S1: Collect historical operating data of the cooling tower, including ambient temperature, humidity, wind speed, cooling water flow rate, packing status and drift water loss, and record the corresponding operating condition labels to construct a multi-dimensional operating condition feature dataset. S2: The collected raw data is imputed and normalized to eliminate the differences in the dimensions of different sensors and generate a standardized input feature matrix for subsequent model training. S3: Construct a working condition classifier based on the working condition labels, use the K-means clustering algorithm to classify the training samples into working condition categories, and generate a working condition category mapping table for dynamic adjustment of feature weights during the model training phase. S4: Construct a deep neural network model architecture, set the number of nodes in the input layer, hidden layer and output layer, initialize the connection weights of each layer, and input the standardized input feature matrix into the model to obtain the preliminary prediction output; S5: During model training, based on the current working condition category of the sample, an adaptive weighting algorithm is used to dynamically adjust the weight distribution of input features in the loss function to enhance the contribution of key features under different working conditions. S6: In the online prediction stage, the current operating data of the cooling tower is collected in real time and input into the operating condition classifier to identify the current operating condition label and call the feature weight configuration corresponding to the label to perform forward inference and generate the predicted value of drift water loss. S7: Compare the predicted value with the actual measured value, calculate the prediction error, and optimize the parameter configuration of the adaptive weight algorithm based on the error feedback information to achieve online incremental update of model weights; S8: During model inference, if a sudden change in the working condition label is detected, switch to the corresponding feature weight configuration and start the online fine-tuning mechanism to improve the model's adaptability to new working conditions. S9: Based on historical prediction errors and working condition label distribution, construct model performance evaluation indicators and dynamically adjust feature weight update strategies to optimize the model's generalization ability and prediction stability under different working conditions.

[0008] The cooling tower drift water loss prediction and analysis method provided in this application has the following beneficial effects: (1) In the model training and inference stages, this invention dynamically adjusts the contribution of key input features to the configuration most suitable for the current working condition through working condition classification and weight mapping, thereby significantly improving the model's generalization ability to scenarios such as sudden changes in working conditions and long-term drift. (2) This invention achieves continuous optimization of model weights with actual working conditions and prediction performance through online fine-tuning driven by error feedback. The model can detect sudden changes in working condition labels in real time, automatically switch the optimal weight configuration and make rapid fine-tuning, effectively solving the problem of prediction distortion caused by changes in working conditions, and ensuring that the prediction results are always highly consistent with the actual operation. (3) This invention enables high-frequency multi-source monitoring data to fully reflect the dynamic details of the working conditions through sliding window segmentation, automatic correction of outliers, feature engineering expansion and strict time sequence alignment. The input features are comprehensive, without redundancy and without noise interference. Multi-dimensional normalization and feature expansion ensure that different physical quantities have a unified mapping space, which improves the effective learning efficiency of the neural network and the consistency of the prediction results. (4) The working condition classification and weight allocation scheme of the present invention can adapt to various complex scenarios such as environment, load, equipment aging and human intervention, so that the prediction model can not only meet the accuracy requirements of niche working conditions such as high humidity, high temperature, extreme wind speed and equipment parameter fluctuation, but also be widely used in various factory cooling towers. Especially in scenarios such as sudden abnormalities and high fluctuating loads, the model can still maintain excellent prediction stability through incremental self-learning. Attached Figure Description

[0009] Figure 1 This is a flowchart of a cooling tower drift water loss prediction and analysis method according to the present invention; Figure 2 This is a sub-flowchart of a cooling tower drift water loss prediction and analysis method according to the present invention; Figure 3 This is another sub-flowchart of the cooling tower drift water loss prediction and analysis method of the present invention. Detailed Implementation

[0010] Embodiments of the present invention are described in detail below, examples of which are shown in the accompanying drawings, wherein the same or similar reference numerals denote the same or similar elements or elements having the same or similar functions throughout. The embodiments described below with reference to the accompanying drawings are exemplary and are only used to explain the present invention, and should not be construed as limiting the present invention.

[0011] The following disclosure provides many different embodiments or examples for implementing different structures of the invention. To simplify the disclosure, specific examples of components and arrangements are described below. Of course, these are merely examples and are not intended to limit the invention. Furthermore, reference numerals and / or letters may be repeated in different examples; such repetition is for simplification and clarity and does not in itself indicate a relationship between the various embodiments and / or arrangements discussed.

[0012] like Figure 1 As shown, this application provides a method for predicting and analyzing cooling tower runoff water loss, specifically including: S1: Collect historical operating data of the cooling tower, including ambient temperature, humidity, wind speed, cooling water flow rate, packing status and drift water loss, and record the corresponding operating condition labels to construct a multi-dimensional operating condition feature dataset. S2: The collected raw data is imputed and normalized to eliminate the differences in the dimensions of different sensors and generate a standardized input feature matrix for subsequent model training. S3: Construct a working condition classifier based on the working condition labels, use the K-means clustering algorithm to classify the training samples into working condition categories, and generate a working condition category mapping table for dynamic adjustment of feature weights during the model training phase. S4: Construct a deep neural network model architecture, set the number of nodes in the input layer, hidden layer and output layer, initialize the connection weights of each layer, and input the standardized input feature matrix into the model to obtain the preliminary prediction output; S5: During model training, based on the current working condition category of the sample, an adaptive weighting algorithm is used to dynamically adjust the weight distribution of input features in the loss function to enhance the contribution of key features under different working conditions. S6: In the online prediction stage, the current operating data of the cooling tower is collected in real time and input into the operating condition classifier to identify the current operating condition label and call the feature weight configuration corresponding to the label to perform forward inference and generate the predicted value of drift water loss. S7: Compare the predicted value with the actual measured value, calculate the prediction error, and optimize the parameter configuration of the adaptive weight algorithm based on the error feedback information to achieve online incremental update of model weights; S8: During model inference, if a sudden change in the working condition label is detected, switch to the corresponding feature weight configuration and start the online fine-tuning mechanism to improve the model's adaptability to new working conditions. S9: Based on historical prediction errors and working condition label distribution, construct model performance evaluation indicators and dynamically adjust feature weight update strategies to optimize the model's generalization ability and prediction stability under different working conditions.

[0013] Step S1: Collect historical operating data of the cooling tower, including ambient temperature, humidity, wind speed, cooling water flow rate, packing condition, and drift water loss, and record corresponding operating condition labels to construct a multi-dimensional operating condition feature dataset. Specifically, this includes: S1.1: Based on the industrial cooling tower operation monitoring system, acquire historical operation data of the cooling tower in different operation cycles. The operation data includes ambient temperature, humidity, wind speed, cooling water flow rate, packing state parameters and drift water loss measurement values ​​to construct a basic data sample set. Based on the industrial cooling tower operation monitoring system, a multi-channel data acquisition interface is configured to realize the synchronous acquisition of ambient temperature, humidity, wind speed, cooling water flow, packing status and drift water loss. Furthermore, through the digital sensor data bus protocol (parameters: Modbus-TCP / IP or Industrial Ethernet), high-speed, low-latency transmission of raw data from various sensors is achieved, and a set of measurement values ​​with acquisition timestamps is obtained; Furthermore, a multi-source data buffering and aggregation algorithm (parameters: ring buffer size = 1024 records, refresh cycle = 1s) is adopted to centrally store the data streams output by different sensors to form a snapshot of the original data containing multi-dimensional operating parameters. Furthermore, based on the operating cycle definition of the equipment management module, the cycle segmentation function is called to classify and group the cached raw data according to the operating cycle identifier, ensuring that each group of data corresponds to the complete operating conditions of the cooling tower. Furthermore, a metadata binding mechanism is adopted to attach structured description fields such as collection location, device ID, and time range to each data group, forming a set of basic data sample records that can be uniquely indexed; Through the above data acquisition and organization algorithms, the acquisition results of the previous step are transformed into a basic data sample set covering all operating parameters and identification information, so as to realize the reliable acquisition and structured storage of raw data of different operating cycles of the cooling tower. For example, in the operation monitoring system of a cooling tower in a chemical plant, the following sensors are configured: ambient temperature sensor (measurement range: -40℃~85℃, accuracy ±0.5℃), humidity sensor (range: 0%~100%RH, accuracy ±2%RH), three-cup anemometer (range: 0~60m / s, accuracy ±0.3m / s), electromagnetic flowmeter (range: 0~5000m³ / h, accuracy ±0.5%), packing condition monitoring module (based on infrared ranging, accuracy ±1mm), and drift water loss measurement device (ultrasonic level method, accuracy ±0.1mm). All sensors upload data in real time via a Modbus-TCP network at a sampling frequency of 1Hz. The data acquisition server uses a circular buffer to temporarily store the most recent 1024 records and groups them in 600-second cycles. Each set of data is bound to the device ID, factory location, and time range information and then stored in a PostgreSQL database. This method can maintain a data loss rate of 0 when running continuously for 24 hours and ensure that all running cycle data can be accurately retrieved and called for subsequent timestamp alignment and feature extraction processing. S1.2: Perform timestamp alignment on the acquired historical operating data of the cooling tower, and use a linear interpolation algorithm to synchronize the time series data to eliminate the data misalignment problem caused by inconsistent acquisition frequencies of multiple source sensors, and generate a time-aligned multidimensional data matrix; For the timestamp information output by different sensors in the historical data of cooling tower operation, a unified time reference setting method (parameter: reference cooling tower main control system clock) is adopted to realize the reference time determination and time sequence axis construction of multi-source data; Furthermore, by using a time difference calculation method (parameter: the offset between the actual sampling time of each sensor data and the reference time), the precise mapping of each sensor data sampling point on the main axis time is achieved, and a time offset dataset is obtained. Furthermore, a linear interpolation algorithm (parameters: interpolation interval length, endpoint data values) is used to fill in missing sampling points and reconstruct the data stream corresponding to the time offset dataset, and to generate a time-synchronized feature variable sequence. Furthermore, by using a multidimensional data matrix merging algorithm (parameters: feature variable type, time index key value), the matrix splicing of feature variables of various dimensions under a unified time scale is realized, and a preliminary time-aligned multidimensional data matrix is ​​obtained; Furthermore, a time synchronization verification method is adopted (parameter: maximum allowable time deviation threshold). This allows for the verification of the time accuracy of the initial time-aligned multidimensional data matrix and the generation of time synchronization accuracy indicators. By using the timestamp alignment and linear interpolation methods described above, the raw data from multiple sources of sensors are transformed into a multi-dimensional data matrix with time consistency meeting the preset accuracy requirements, thereby ensuring the time consistency of input data in the subsequent outlier detection and condition labeling stages. For example, in a cooling tower operation monitoring scenario at a chemical plant, six types of sensor data are selected: ambient temperature, humidity, wind speed, cooling water flow rate, packing condition, and drift water loss. The main control system clock is set to a second count reference starting at midnight each day, and the sampling frequencies are as follows: , , Inconsistent frequencies. The time difference between the humidity sensor and the reference time is calculated. to Within the specified range, linear interpolation is used to fill in missing time points, generating a continuous sampling sequence. Using the time index as the key, the six feature variables are concatenated on a unified second scale. The daily operation matrix of dimension, after synchronization verification, has its time deviation controlled within a certain range. Within this range. This matrix is ​​input into the outlier detection module of the subsequent sliding window, improving detection accuracy by approximately [percentage missing]. And ensure consistency of the time dimension during the labeling of working conditions; S1.3: Perform outlier detection and correction on the time-aligned multidimensional data matrix. Use a sliding window-based Z-score normalization algorithm to identify and correct outlier data points that exceed the 3σ range, so as to improve the reliability and consistency of the original data. S1.4: Based on the operating environment and process parameters of the cooling tower, define the operating condition classification standard, which includes seasonal characteristics, load level, ambient temperature and humidity range and wind speed variation range, to form a structured chemical condition labeling system. S1.5: According to the defined operating condition classification criteria, perform operating condition labeling on each historical operation record, and use a rule-based decision tree algorithm to classify and judge the combination of input features to generate a structured training sample set with operating condition labels. S1.6: The labeled structured training sample set is segmented according to time windows, and a sliding window mechanism is used to generate continuous time series sample segments to enhance the model's ability to learn dynamic working condition change trends. For the structured training sample set with completed working condition labeling, a time window segmentation method (parameters: window length L, step size Δ) is used to realize dynamic working condition data segmentation based on fixed-length sampling units; Furthermore, by using a sliding window mechanism (parameters: window length L, window step size Δ), a point-by-point sliding truncation operation is performed on the time series to achieve continuous overlapping segmentation of the samples and obtain a set of sequence segments covering different time positions, thereby enhancing the ability of the sample set to capture short-term and long-term operating condition changes. Furthermore, during the segmentation process, a timestamp index mapping algorithm (parameters: start index, end index, step size Δ) is used to perform sequence index mapping on each segment, thereby achieving the traceability of the segments on the original timeline and generating an index-segment mapping table; Furthermore, the generated sequence fragment set is sorted and numbered, and a sequence order encoding algorithm (parameter: encoding rule R) is used to generate a unique segment identifier for each fragment, so as to facilitate the association operation in the subsequent feature expansion and working condition label calling process; By using sliding window segmentation and index encoding, the structured training sample set output in the previous step is transformed into multiple time series sample segments with temporal continuity and overlap, thereby improving the learning ability of deep neural networks to learn dynamic working condition change trends during the training phase. For example, in the structured training sample set with operating condition labels collected by the industrial cooling tower monitoring system, the window length L is set to... Minutes, step size Δ set to Minutes, the sliding window algorithm indexes from the start time of the sample set. Initially, the iteratively truncates to a length of... Minute-long time series segments, scrolling forward each time. Overlapping segments are formed within minutes. During processing, a time index mapping table establishes a correspondence between the start and end timestamps of each segment and the segment number, for example, the segment number... Corresponding timestamp range Fragment number Corresponding interval The set of sample fragments processed by sliding window segmentation, in a scenario covering a week's worth of running data, generates a number of fragments reaching [number missing]. Each segment, combined with the original operating condition label, is used for subsequent feature expansion and model training and verification, which improves the model's fitting and prediction accuracy of cooling tower operating condition change trends occurring at different time scales. S1.7: Perform feature expansion operations on the segmented continuous time series sample segments to generate derived features such as first-order difference, moving average, and exponentially weighted moving average based on the original input features, so as to enrich the dimension of the model input features; S1.8: Bind the expanded input features to the corresponding working condition labels to construct a multi-dimensional working condition feature dataset. The dataset includes the input feature vector, the corresponding working condition label, and the target output value (drift water loss), which is used for the training and validation of the subsequent deep neural network model.

[0014] Step S2 involves imputing missing values ​​and normalizing the collected raw data to eliminate differences in the dimensions of different sensors, and generating a standardized input feature matrix for subsequent model training. Specifically, this includes: S2.1: Identify and mark missing values ​​in the collected raw data of cooling tower operation to determine the timestamp position and missing proportion of missing samples in each sensor data stream, and obtain the distribution characteristics of missing values; S2.2: Based on the distribution characteristics of the missing values, a linear interpolation algorithm is used to fill in the missing samples in the time series data to restore the continuity of the sensor data stream and generate the filled time series dataset; S2.3: Perform maximum and minimum value statistical analysis on the features of each dimension in the imputed time series dataset to calculate the original value range of each sensor variable and obtain the feature scale distribution parameters; S2.4: Based on the feature scale distribution parameters, the minimum-maximum normalization method is used to perform linear transformation on each input feature to map all feature values ​​to the [0,1] interval, generating a normalized feature vector set; Based on the feature scale distribution parameters, a minimum-maximum normalization method is used (parameter: feature minimum value). eigenvalue Original eigenvalues This ensures that all input features maintain dimensional consistency and proportional coordination during the linear mapping process; Furthermore, using the normalization formula: This method limits the original feature values ​​to the interval [0,1] and obtains a normalized feature vector set. ; Furthermore, the normalization transformation is performed on the entire time series feature matrix through a batch processing mechanism, while maintaining the consistency of normalization parameters between samples at different times, so as to ensure the comparability and statistical stability of the data distribution. Furthermore, by employing an independent channel processing strategy, minimum-maximum mapping is performed on each sensor feature dimension using its own scale distribution parameters, thus avoiding the impact of normalization bias between features of different dimensions on model training performance. By using the min-max normalization method, the feature matrix after the previous step is filled in is transformed into a normalized feature vector set, thereby achieving consistent mapping of the multi-dimensional feature numerical domain and lightweight input standardization in the edge model deployment environment. For example, in the winter operation of a cooling tower, the ambient temperature range collected by the sensor is -5℃ to 15℃, humidity range is 50% to 90%, wind speed range is 0.5 m / s to 5.0 m / s, cooling water flow rate range is 1200 m³ / h to 3500 m³ / h, and packing condition parameter range is 0.8 to 1.0. The minimum and maximum values ​​of each characteristic dimension are set as follows: , (temperature), , (Humidity), and so on. For a sample with a temperature of 10℃ and a humidity of 70%, the normalized temperature value is calculated using the normalization formula: The result is 0.75; normalized humidity value: The result is 0.5. After performing this process, the normalized feature vector remains within the [0,1] interval, enabling direct numerical comparison between different physical quantities and significantly improving the stability and convergence speed of the model input during the training and inference phases; S2.5: Arrange the normalized feature vector set in chronological order and combine it into a standardized input feature matrix of uniform dimension, which can be used as the input data format for training deep neural network models.

[0015] Step S3: Construct a working condition classifier based on the working condition labels, use the K-means clustering algorithm to classify the training samples into working condition categories, and generate a working condition category mapping table for dynamic adjustment of feature weights during model training. Figure 2 As shown, it specifically includes: S3.1: Perform feature extraction processing on the operating condition tags in the collected historical operating data of the cooling tower to obtain the original operating condition feature vector set for operating condition classification. The operating condition tags include ambient temperature and humidity, wind speed, cooling water flow rate and packing status information. S3.2: Based on the original working condition feature vector set, the principal component analysis (PCA) algorithm is used to perform feature dimensionality reduction processing to extract a representative low-dimensional working condition feature space, reduce the complexity of subsequent clustering calculations, and retain the main working condition difference information. S3.3: In the low-dimensional working condition feature space, the K-means clustering algorithm is used to perform unsupervised clustering analysis on the training samples, initial cluster centers are set and the clustering results are iteratively optimized to divide several working condition category clusters with significant differences. In the low-dimensional working condition feature space, the K-means clustering algorithm (parameters: the number of clusters k is determined by the difference analysis of the working condition feature space, the number of iteration steps is set to 100-500, and the convergence threshold is set to 1×10-5 to 1×10-6) is used to realize unsupervised clustering analysis of the training samples. Furthermore, the distance from each sample to the current cluster center is calculated using the Euclidean distance calculation method (parameter: the distance metric is the squared Euclidean distance of the standardized feature vector space), resulting in a distance matrix between the sample and each cluster center. The Euclidean distance calculation formula is as follows: in, Let be the Euclidean distance from the i-th sample to the j-th cluster center. Let m be the normalized value of the i-th sample on the m-th feature dimension. Let M be the value of the j-th cluster center on the m-th feature dimension, where M is the number of feature dimensions; Furthermore, by iteratively updating the cluster centers (parameter: in each iteration, the centers are recalculated based on the mean of the current cluster members, and the stopping condition is the maximum number of iterations or the rate of change of the distance matrix is ​​less than a threshold), the clustering results are converged and optimized, and a final stable set of cluster centers is obtained. Furthermore, the clustering effect is quantitatively evaluated using the sum of squared errors within clusters (SSE) calculation method; Furthermore, the separation degree between different clusters is evaluated by the inter-cluster distance analysis method. The ratio of the average inter-cluster distance to the average intra-cluster distance is used as the separation degree index to select the optimal cluster number configuration. The K-means clustering algorithm is used to transform the low-dimensional working condition feature space generated in the previous sub-step into a cluster of different working condition categories, thereby classifying the working condition categories of the input samples and providing a basis for the generation of the subsequent working condition category mapping table and the dynamic adjustment of feature weights. For example, in a predictive analysis scenario for an industrial cooling tower, a low-dimensional operating condition feature space dataset is taken after dimensionality reduction. The feature dimension M=5, the sample size N=10000, and the number of clusters k=6 is determined by difference analysis. The maximum number of iterations is set to 300, and the convergence threshold is 1×10⁻⁶. -5 The distance matrix from each sample to the current cluster center is obtained by Euclidean distance calculation. The matrix size is 10000×6. When iterating and updating the cluster center, the center vector is recalculated in each round using the mean of the current cluster members. The initial SSE is 4500. The optimization iteration is until the SSE drops to 2300 and the rate of change is 0.0008, and then convergence is achieved. Finally, 6 stable cluster centers are obtained. Each center contains the feature mean vector of the corresponding working condition category. The inter-cluster separation is 3.5, which meets the requirement of working condition category discrimination in the model training stage. S3.4: Based on the K-means clustering results, the correlation between the distribution characteristics of samples within each cluster and the working condition labels is statistically analyzed. The clustering results are semantically annotated in conjunction with the experience of domain experts to generate a set of working condition category labels with physical meaning. S3.5: Based on the mapping relationship between the working condition category label set and the original working condition features, construct a working condition category mapping table, which is used to identify the current working condition category according to the real-time input features during the model training and prediction stages, and call the corresponding feature weight configuration. Based on the mapping relationship between the working condition category label set and the original working condition features, a multi-dimensional feature index matching algorithm (parameters: working condition category label set, original working condition feature variable values) is adopted to achieve a unique binding between each working condition category label and its corresponding original working condition feature vector, and to form an initial working condition-feature mapping pair. Furthermore, by using a hash mapping construction method (parameters: category label key value, feature vector hash value), a fast retrieval of the working condition category to the feature weight configuration scheme is achieved, and an efficient mapping table storage structure is obtained; Furthermore, a feature weight vector normalization algorithm (parameters: working condition category index, weight coefficient set) is adopted to standardize the value range of the feature weight coefficient vector corresponding to each working condition category in the mapping table to a normal value. The interval is defined, and a weight configuration table that can be directly used in the training process is generated; Furthermore, by establishing a bidirectional index (parameters: working condition category label, feature weight configuration ID), a bidirectional fast retrieval capability is achieved from real-time input features to working condition category labels and then to feature weight configuration, and the final version of the working condition category mapping table is generated. By generating and optimizing the working condition category mapping table, the clustering results of the previous step are transformed into efficient working condition-feature weight index data that can be directly called in the model training and prediction stages, so as to achieve real-time weight switching and stable prediction effect when the model faces changing working conditions. For example, in an application scenario of a cooling tower operation dataset, the operating condition category label set includes six physical semantic operating condition types, such as "high temperature and high humidity," "low temperature and low humidity," and "high wind speed and low load." The original operating condition feature dimensions are five (ambient temperature, humidity, wind speed, cooling water flow rate, and packing blockage rate). For the "high temperature and high humidity" category, a feature weight vector is calculated. Then, normalization is performed to obtain normalized weights. (Since the sum is 1, no scaling is needed), it is stored in a hash map table, with the key being the category label and the value being the weight vector. The input feature vector is used in the real-time prediction stage. After matching the "high temperature and high humidity" label, the corresponding normalized weight configuration is directly called to perform deep neural network inference and output the predicted value of drift water loss. The mean square error of the prediction is reduced by 15% compared with the no weight switching strategy, which significantly improves the adaptability and stability of the model under sudden working conditions.

[0016] Step S4: Construct a deep neural network model architecture, set the number of nodes in the input layer, hidden layer, and output layer, initialize the connection weights of each layer, and input the standardized input feature matrix into the model to obtain preliminary prediction output. Figure 3 As shown, it specifically includes: S4.1: Perform dimensionality analysis on the standardized input feature matrix in the multidimensional operating condition feature dataset, and determine the number of input layer nodes of the deep neural network model based on the number of input variables of the cooling tower operating characteristics, so as to construct a network input interface that matches the dimension of the input features; Dimensional analysis is performed on the standardized input feature matrix in the multidimensional working condition feature dataset. A feature dimension extraction algorithm (parameters: standardized feature matrix F, number of samples N) is used to quantify and statistically analyze the matrix column dimensions and obtain the total number of input variables K. Furthermore, by using a variable importance filtering method (parameters: total number of input variables K, working condition label mapping table T), the contribution of features under the current working condition is evaluated, and a working condition correlation matrix R is generated to help confirm the type and number of features that the input port needs to access. Furthermore, the feature mapping construction module (parameters: total dimension K, correlation matrix R) is called to map each input feature to the index space of the network input node and generate a node mapping vector M to ensure that the structure of the input layer corresponds completely to the feature dimension. Furthermore, based on the node mapping vector M and the input dimension K, the formula is used... Determine the number of nodes N in the input layer of the deep neural network model, where To standardize the number of effective column dimensions in the feature matrix, ensure that the input layer structure is consistent with the dimensions of the dataset; By setting the number of nodes in the input layer, the dimensional analysis results from the previous step are transformed into physical parameters of the network architecture, achieving a precise correspondence between the model input and the characteristic variables of the cooling tower operation. For example, in three months of historical cooling tower operation data, the standardized feature matrix has a dimension of 9000 samples and 12 columns. The Pearson correlation coefficient between each column of features and the operating condition label is calculated, yielding contribution coefficients for ambient temperature (0.85), humidity (0.78), wind speed (0.69), cooling water flow rate (0.92), and packing condition (0.81). Variables with coefficients below 0.6 are filtered out, retaining 10 effective features. The node mapping vector M is [EnvTemp, Humidity, WindSpeed, WaterFlow, Filler, Diff1, Avg7, EWMA3, LoadLevel, Season], corresponding to an input layer node count N=10. During model testing, an input layer configuration of N=10 achieves a one-to-one correspondence between operating condition features and network nodes. Verification results show that the prediction mean square error is reduced by 15% compared to N=12. S4.2: Based on the number of input layer nodes and the number of output target variables, the number of hidden layer nodes is initialized using an empirical formula method or a trial-and-error method to achieve a balance between model expressive power and computational resource consumption, and to generate a preliminary description of the hidden layer topology. S4.3: Based on the standardized input feature matrix and the initial hidden layer topology, the Xavier initialization algorithm is used to initialize the connection weights of each layer of the neural network to ensure that the output variance of each neuron is within a reasonable range, thereby obtaining the initial weight matrix set. S4.4: The standardized input feature matrix is ​​propagated forward through the initialized input layer and hidden layer in sequence. The ReLU activation function is used to perform a nonlinear transformation on the output of the hidden layer to generate the activation value matrix of each hidden layer neuron. S4.5: Perform a linear weighted summation and bias term addition operation on the output of the last hidden layer, and perform a final mapping transformation through the output layer activation function to generate a preliminary prediction output of the cooling tower runoff water loss.

[0017] Step S5: During model training, based on the current sample's operating condition category, an adaptive weighting algorithm is used to dynamically adjust the weight distribution of input features in the loss function to enhance the contribution of key features under different operating conditions. Specifically, this includes: S5.1: Perform joint analysis on the input feature vector and the working condition category label, and calculate the initial weight value of each input feature under the current working condition based on the feature importance evaluation algorithm to generate the initial configuration table of feature weights; S5.2: Based on the initial configuration table of feature weights and the model prediction output error, the gradient descent algorithm is used to backpropagate and update the neural network parameters, and a feature weighting term is introduced into the loss function to enhance the influence of key features on the model training process. S5.3: Perform feature sensitivity analysis on the operating condition category of the current training sample, and use the Local Interpretability Model (LIME) to interpret the model prediction results in order to identify the key feature set that has the greatest impact on the output under this operating condition; S5.4: Based on the identification results of the key feature set, a dynamic weight adjustment strategy is adopted to update the weight coefficients of each feature in the loss function in real time, so as to improve the feature selection ability and prediction stability of the model under the current working conditions. The input conditions are the identified key feature set of the current working condition and the corresponding initial weight configuration, and the execution object is the weight term of the loss function during the training phase of the deep neural network; A weighted gradient update method (parameters: current feature weight vector, prediction error gradient matrix) is used to adjust the coefficients of each feature weight in the loss function in real time, resulting in a weighted gradient correction matrix. Furthermore, by using a feature contribution normalization algorithm (parameters: feature importance score, feature weight dynamic range constraint), the magnitude of the weighted gradient correction matrix is ​​limited, and a feature weight update vector that satisfies the constraint is generated. Furthermore, an exponential smoothing update strategy is employed (parameter: smoothing factor). (Historical weight coefficients) to achieve a smooth transition of feature weights, reduce drastic fluctuations during training, and obtain a smoothed weight coefficient matrix; Furthermore, the weight adjustment amount is calculated using the following formula through the error-weighted adjustment algorithm: in, For the first Feature weight increment, For learning rate, For feature importance scores, This is the gradient value of the feature. This represents the prediction error value; By using the above update formula, the feature contribution analysis results of the previous step are transformed into real-time updated data of the weight coefficients of the loss function, thereby optimizing the feature selection and improving the prediction stability of the model under the current working conditions. For example, under high load conditions in summer, the key feature set of a cooling tower includes ambient temperature (importance score of 1). ), humidity (importance score) Cooling water flow rate (importance score) Set the learning rate Prediction error The gradient values ​​of each feature are as follows: , , Applying the above formula, the increment of the temperature feature weight is: The humidity feature weight increment is The increment of traffic feature weights is After the update, the weight coefficients increased steadily, and the mean squared error of the model's prediction under this working condition decreased during the training period. The predicted stability factor has increased to over 90%. This verifies the effectiveness of the dynamic weight adjustment strategy in practical applications; S5.5: Normalize the updated feature weight configuration and feed it back into the next training iteration to form a closed-loop feature weight optimization mechanism, continuously improving the model's generalization ability and adaptability under different working conditions.

[0018] Step S6: In the online prediction stage, the current operating data of the cooling tower is collected in real time and input into the operating condition classifier to identify the current operating condition label. Then, the feature weight configuration corresponding to the label is called for forward inference to generate a predicted value for drift water loss. Specifically, this includes: S6.1: In the online prediction phase, real-time data collection is performed on the operating status of the cooling tower to obtain multi-dimensional operating parameters, including ambient temperature, humidity, wind speed, cooling water flow rate, and packing status, in order to construct the input feature vector of the current operating condition. S6.2: Input the input feature vector into the pre-trained operating condition classifier, perform operating condition category identification of the current sample based on the K-means clustering algorithm, and output the corresponding operating condition label to identify the specific operating condition type of the current cooling tower operation. S6.3: Based on the working condition label output by the working condition classifier, query the pre-built feature weight configuration mapping table to obtain the feature weight configuration parameters corresponding to the working condition label, so as to use them in the forward inference process of the deep neural network model; S6.4: Input the input feature vector and the corresponding feature weight configuration parameters into the deep neural network model, perform forward propagation calculation, and generate a preliminary prediction value of the cooling tower runoff water loss, so as to estimate the loss under the current operating conditions; The input feature vectors identified and mapped by the working condition classifier are weighted element-wise with the corresponding feature weight configuration parameters. The matrix multiplication method is used (parameters: input feature matrix size m×1, weight matrix size m×n) to adjust the contribution of different features in the network input layer. Furthermore, the matrix weighted result is input into the input layer of the deep neural network, and the forward propagation algorithm (parameters: activation function type is ReLU, number of hidden layers is 3, and number of nodes is 64, 32, and 16 respectively) is used to realize the cross-layer transfer of feature information and obtain the activation value matrix of each hidden layer. Furthermore, through weighted summation and nonlinear transformation of each hidden layer, the activation value matrix of the previous layer is multiplied and accumulated with the connection weights and bias parameters of the current layer. The calculation formula is as follows: in, This is the output vector of the current hidden layer. This is the weight matrix for the current layer. Activate the output of the previous layer. For the current layer bias, For activation functions; Furthermore, by performing a linear mapping calculation between the output vector of the last hidden layer and the output layer weight matrix, and employing a regression-type output structure without activation functions, the network output is converted into a preliminary predicted value of cooling tower runoff water loss. The calculation formula is as follows: in, These are preliminary forecast values. This is the output layer weight matrix. For the final hidden layer output, For output layer bias; Through the above matrix weighting, layer-by-layer forward propagation and output layer linear mapping, the input features after feature weight adjustment are transformed into quantifiable prediction data, so as to achieve a preliminary estimate of the cooling tower runoff water loss under the current operating conditions. For example, under the high-load operation condition of a power plant in summer, the collected input feature vector has a length of 6, including ambient temperature (35℃), humidity (80%), wind speed (3m / s), cooling water flow rate (2500m³ / h), packing resistance coefficient (0.8), and historical average drift water loss (15m³ / h), with corresponding weight configuration parameters of [0.25, 0.3, 0.15, 0.1, 0.1, 0.1]. After element-wise weighting, the vector is input into the network, with a first hidden layer of 64 nodes, a second hidden layer of 32 nodes, and a third hidden layer of 16 nodes. The weight matrix is ​​initialized using Xavier, and the activation function is ReLU for all layers. After matrix multiplication and accumulation, and nonlinear transformation, the output layer uses a 1-node linear activation to obtain preliminary prediction results. m³ / h. After error correction in the subsequent S6.5 post-processing stage, the prediction error was reduced to 2%, achieving stable prediction capability under this high load condition; S6.5: Post-process the preliminary prediction value, combine it with historical error feedback information to correct the prediction result, and output the final prediction result of cooling tower runoff water loss to improve the accuracy and stability of online prediction.

[0019] Step S7: Compare the predicted value with the actual measured value, calculate the prediction error, and optimize the parameter configuration of the adaptive weight algorithm based on the error feedback information to achieve online incremental update of the model weights. Specifically, this includes: S7.1: Calculate the difference between the predicted value and the actual measured value of the cooling tower drift water loss to obtain the prediction error sequence, which serves as the input for online model optimization; S7.2: Based on the predicted error sequence, the mean and variance of the error are calculated using the sliding window statistical method to quantify the prediction stability of the current model and generate an error distribution feature vector; S7.3: Perform gradient backpropagation algorithm on the error distribution feature vector to calculate the error gradient of each input feature under the current working condition label, so as to identify the feature dimension that contributes more to the prediction error. S7.4: Based on the error gradient information, the adaptive learning rate optimization algorithm is used to update the parameters of the feature weight allocation strategy and generate an updated feature weight configuration matrix to enhance the model's adaptability to the current working conditions. Based on error gradient information, an adaptive learning rate optimization algorithm is adopted (parameter: initial learning rate). Minimum learning rate Maximum learning rate Regulatory factors This enables dynamic parameter updates for feature weight allocation strategies. Furthermore, through the adaptive learning rate calculation module, based on the error gradient vector... The update rate of each feature weight is evaluated dimension by dimension, and the learning rate is adjusted using the following formula: in The `clip` function is used to limit the learning rate within a preset range for the i-th feature, where `i` is the update learning rate. The absolute value of the error gradient. As a regulating factor; Furthermore, the current weights are adjusted using the feature weight update formula: in Let be the current weight value of the i-th feature. By introducing a learning rate that is adjusted dimension by dimension, we can ensure that the update magnitude of the weights of features with larger gradients is appropriately reduced, while the update magnitude of the weights of features with smaller gradients is moderately increased, thereby balancing the adaptation speed and stability. Furthermore, the updated feature weight vector is normalized to ensure that the relative proportions of each weight are consistent under different working conditions. By using an adaptive learning rate optimization method, the error gradient information from the previous step is transformed into a dynamic configuration matrix of each feature weight and its corresponding learning rate, thereby improving the model's predictive adaptability under the current working conditions. For example, in a scenario where a cooling tower operates under high load during the summer, the input features include ambient temperature. ℃, humidity %, wind speed m / s, cooling water flow rate m³ / h, packing resistance coefficient The error gradient vector is Set the initial learning rate. = , = , = , = The learning rate matrix is ​​obtained through calculation using the formula. Adjust the original weight vector using the feature weight update formula. And normalized to the new weight configuration Online prediction validation shows that the adjusted model reduced the mean absolute percentage error under this operating condition. This achieves a simultaneous improvement in predictive stability and adaptability; S7.5: Perform model parameter freezing and caching operations on the updated feature weight configuration matrix to complete the online incremental update of model weights, and store the update results in the local model parameter library of the edge device for subsequent prediction and inference calls.

[0020] Step S8: During model inference, if a sudden change in the working condition label is detected, the system switches to the corresponding feature weight configuration and initiates an online fine-tuning mechanism to improve the model's adaptability to new working conditions. Specifically, this includes: S8.1: Based on the input feature data of the current inference stage, the working condition is identified by the working condition classifier to obtain the working condition label of the current sample; S8.2: Compare the working condition label of the current sample with the working condition label of the previous inference cycle. If the two are inconsistent, it is determined that the working condition label has changed abruptly, so as to trigger the feature weight configuration switching mechanism. S8.3: Based on the mutated working condition label, retrieve the corresponding feature weight vector from the pre-stored feature weight configuration table to replace the connection weight configuration between the current model input layer and hidden layer; S8.4: After the feature weights are switched, a fine-tuning training set is built based on the current sample and its neighboring historical samples to perform online fine-tuning optimization of local model parameters; S8.5: The incremental learning algorithm is used to locally update the model parameters to minimize the prediction error under the current working conditions, and the updated model parameters and weight configuration are stored in the online model cache for use in subsequent inference cycles.

[0021] Step S9: Based on historical prediction errors and the distribution of working condition labels, construct model performance evaluation indicators and dynamically adjust the feature weight update strategy to optimize the model's generalization ability and prediction stability under different working conditions. Specifically, this includes: S9.1: Perform time window sliding slicing on the historical prediction error sequence and the corresponding working condition label to obtain the working condition-error correlation feature vector, which is used for subsequent performance evaluation index construction; S9.2: Based on the aforementioned working condition-error correlation feature vector, the Pearson correlation coefficient calculation method is used to calculate the correlation coefficient between each working condition category and the prediction error, in order to identify key working condition types that have a significant impact on model performance; S9.3: Based on the distribution density and error contribution of the key working condition types, construct a weighted comprehensive performance evaluation index. The evaluation index includes working condition coverage, error volatility and model stability factor, which are used to quantify the adaptability of the model under different working conditions. S9.4: Based on the dynamic change trend of the comprehensive performance evaluation index, a fuzzy logic controller is used to adaptively adjust the learning rate, forgetting factor and weight decay parameters in the feature weight update strategy, so as to realize the online optimization of the model feature contribution adjustment mechanism. S9.5: Apply the adaptively adjusted feature weight update strategy to the online incremental training process of the model, generate the updated feature weight configuration file, and write it into the weight call cache during the model inference stage to improve the prediction accuracy and robustness of the model under new or sudden conditions.

[0022] The technical solution of the present invention has been described above with reference to the preferred embodiments shown in the accompanying drawings. However, it will be readily understood by those skilled in the art that the scope of protection of the present invention is obviously not limited to these specific embodiments. Without departing from the principles of the present invention, those skilled in the art can make equivalent changes or substitutions to the relevant technical features, and the technical solutions after these changes or substitutions will all fall within the scope of protection of the present invention.

[0023] The above description is merely a preferred embodiment of the present invention and is not intended to limit the invention. Various modifications and variations can be made to the present invention by those skilled in the art. Any modifications, equivalent substitutions, improvements, etc., made within the spirit and rules of the present invention should be included within the scope of protection of the present invention.

Claims

1. A method for predicting and analyzing water loss from cooling towers, characterized in that, Includes the following steps: S1: Collect historical operating data of the cooling tower and record the corresponding operating condition labels to construct a multi-dimensional operating condition feature dataset; S2: Perform missing value imputation and normalization on the collected raw data to generate a standardized input feature matrix; S3: Construct a working condition classifier based on the working condition labels, classify the training samples into working condition categories, and generate a working condition category mapping table; S4: Construct a deep neural network model architecture, set the number of nodes in the input layer, hidden layer and output layer, initialize the connection weights of each layer, and input the standardized input feature matrix into the model to obtain the preliminary prediction output; S5: Based on the current working condition category of the sample, an adaptive weighting algorithm is used to dynamically adjust the weight distribution of input features in the loss function; S6: Real-time acquisition of the current operating data of the cooling tower and input into the operating condition classifier, identification of the current operating condition label, and calling the feature weight configuration corresponding to the label for forward inference to generate the predicted value of drift water loss; S7: Compare the predicted value of drift water loss with the actual measured value, calculate the prediction error, and optimize the parameter configuration of the adaptive weighting algorithm based on the error feedback information; S8: During model inference, if a sudden change in the working condition label is detected, switch to the corresponding feature weight configuration and start the online fine-tuning mechanism; S9: Based on historical prediction errors and operating condition label distribution, construct model performance evaluation indicators and dynamically adjust feature weight update strategies.

2. The method for predicting and analyzing cooling tower runoff water loss according to claim 1, characterized in that, Step S1 specifically includes: Based on the industrial cooling tower operation monitoring system, historical operation data of the cooling tower in different operating cycles are obtained to construct a basic data sample set; The historical running data is subjected to timestamp alignment, and the time series data is synchronized using a linear interpolation algorithm to generate a time-aligned multidimensional data matrix. Outlier detection and correction are performed on the time-aligned multidimensional data matrix, and an outlier data point is identified and corrected using a sliding window-based standardization algorithm. Based on the operating environment and process parameters of cooling towers, we define operating condition classification standards and form a structured operating condition labeling system. According to the operating condition classification criteria, an operating condition labeling operation is performed on each historical operation record. A rule-based decision tree algorithm is used to classify and judge the combination of input features, and a structured training sample set with operating condition labels is generated. The structured training sample set is segmented according to time windows, and a sliding window mechanism is used to generate continuous time series sample segments. Perform feature expansion operations on the segmented continuous time series sample segments to generate derived features based on the original input features; The expanded input features are bound to the corresponding operating condition labels to construct a multidimensional operating condition feature dataset.

3. The method for predicting and analyzing cooling tower runoff water loss according to claim 2, characterized in that, The historical operating data includes ambient temperature, humidity, wind speed, cooling water flow rate, packing condition parameters, and drift water loss measurements.

4. The method for predicting and analyzing cooling tower runoff water loss according to claim 1, characterized in that, Step S2 specifically includes: Missing values ​​in the collected raw data of cooling tower operation are identified and marked, the timestamp positions and missing proportions of missing samples in the data streams of each sensor are determined, and the distribution characteristics of missing values ​​are obtained. Based on the missing value distribution characteristics, a linear interpolation algorithm is used to fill in the missing samples in the time series data to generate a filled time series dataset. Perform maximum and minimum value statistical analysis on the features of each dimension in the filled time series dataset, calculate the original value range of each sensor variable, and obtain the feature scale distribution parameters; Based on the aforementioned feature scale distribution parameters, the min-max normalization method is used to perform linear transformation on each input feature, mapping all feature values ​​to a unified interval to generate a normalized feature vector set. The normalized feature vector set is arranged in chronological order and combined into a standardized input feature matrix of uniform dimension.

5. The method for predicting and analyzing cooling tower runoff water loss according to claim 1, characterized in that, Step S3 specifically includes: Feature extraction processing is performed on the operating condition labels in the collected historical operating data of the cooling tower to obtain the original operating condition feature vector set; Based on the original working condition feature vector set, principal component analysis algorithm is used to perform feature dimensionality reduction processing to extract a representative low-dimensional working condition feature space. In the low-dimensional working condition feature space, the K-means clustering algorithm is used to perform unsupervised clustering analysis on the training samples, setting initial cluster centers and iteratively optimizing the clustering results; Based on the clustering results, the distribution characteristics of samples within each cluster and their correlation with the working condition labels are statistically analyzed. The clustering results are then semantically annotated using domain expert experience to generate a set of working condition category labels. Based on the mapping relationship between the working condition category label set and the original working condition features, a working condition category mapping table is constructed.

6. The method for predicting and analyzing cooling tower drift water loss according to claim 5, characterized in that, In step S3, when the K-means clustering algorithm is used for classifying work conditions, it first uses principal component analysis to perform feature dimensionality reduction on the original work condition features, and then determines the number of clusters k based on the difference analysis of the work condition feature space. The number of iterations is 100-500, and the convergence threshold is set to 1×10⁻⁶. -5 Up to 1×10 -6 .

7. The method for predicting and analyzing cooling tower runoff water loss according to claim 1, characterized in that, Step S4 specifically includes: Dimensional analysis is performed on the standardized input feature matrix in the multidimensional operating condition feature dataset. The number of input variables of the cooling tower operating characteristics is used to determine the number of input layer nodes of the deep neural network model and to construct a network input interface that matches the dimension of the input features. Based on the number of input layer nodes and the number of output target variables, the number of hidden layer nodes is initialized and configured to generate a preliminary hidden layer topology description. Based on the standardized input feature matrix and the initial hidden layer topology, the connection weights of each layer of the neural network are initialized to obtain the initial weight matrix set; The standardized input feature matrix is ​​propagated forward through the initialized input layer and hidden layer in sequence. The ReLU activation function is used to perform a nonlinear transformation on the output of the hidden layer to generate the activation value matrix of each hidden layer neuron. A linear weighted summation and bias term addition operation is performed on the output of the last hidden layer. The final mapping transformation is then performed through the output layer activation function to generate a preliminary prediction output of the cooling tower runoff water loss.

8. The method for predicting and analyzing cooling tower runoff water loss according to claim 7, characterized in that, In step S4, the number of input layer nodes of the deep neural network model is set according to the number of effective features, and the number of hidden layer nodes is initialized and configured using empirical formulas and trial algorithms.

9. The method for predicting and analyzing cooling tower runoff water loss according to claim 1, characterized in that, Step S5 specifically includes: The input feature vector and the working condition category label are jointly analyzed. The initial weight value of each input feature under the current working condition is calculated based on the feature importance evaluation algorithm, and the initial configuration table of feature weights is generated. Based on the initial configuration table of feature weights and the model prediction output error, the gradient descent algorithm is used to backpropagate and update the neural network parameters, and a feature weighting term is introduced into the loss function. Perform feature sensitivity analysis on the working condition category of the current training sample, use a local interpretability model to interpret the model prediction results, and identify the key feature set that has the greatest impact on the output under this working condition; Based on the identification results of the key feature set, a dynamic weight adjustment strategy is adopted to update the weight coefficients of each feature in the loss function in real time. The updated feature weight configuration is normalized and fed back into the next training iteration, forming a closed-loop feature weight optimization mechanism.

10. The method for predicting and analyzing cooling tower runoff water loss according to claim 1, characterized in that, In step S6, the input features collected in real time are used to determine the working conditions by the working condition classifier, and the corresponding feature weight configuration is searched and loaded. The weighted result is input into the neural network to obtain the preliminary prediction value of drift water loss. The prediction value is then corrected by error feedback and the final result is output.

Images