A method for data analysis and diagnosis of a thermal power plant

Abstract

The present application belongs to the technical field of industrial data analysis and fault diagnosis, and particularly relates to a thermal power plant data analysis and diagnosis method. Real-time operation data such as boiler thermal parameters are obtained through data acquisition, data sets are divided according to the acquisition cycle, the cycle is dynamically adjusted, and the working condition difference degree is calculated to construct a gradient training set. Improved empirical mode decomposition and wavelet threshold denoising are used to complete data enhancement preprocessing, an LSTM-CNN hybrid model containing a time sequence attention mechanism is constructed, and real-time detection and analysis are carried out after the parameters are optimized by using an improved particle swarm algorithm. The method solves the problem of insufficient processing of nonlinear and non-stationary data by traditional manual inspection and statistical methods, can accurately capture transient abnormalities and multi-parameter coupling faults of equipment, and improves the accuracy and real-time performance of the operation state diagnosis of the thermal power plant, thereby ensuring energy efficiency and production safety.

Description

Technical Field

[0001] This invention belongs to the field of industrial data analysis and fault diagnosis technology, and in particular relates to a data analysis and diagnosis method for thermal power plants. Background Technology

[0002] As a crucial hub for energy conversion, the accurate analysis and fault diagnosis of the operating status of thermal power plants are essential for ensuring energy efficiency and production safety. Currently, data analysis and diagnosis in thermal power plants often rely on manual experience combined with traditional statistical methods, such as recording equipment parameters through regular inspections and then performing trend analysis based on historical data tables. This approach has significant limitations: firstly, the frequency and coverage of manual inspections are limited, making it difficult to capture transient anomalies and multi-parameter coupled faults; secondly, traditional statistical methods are insufficient for processing nonlinear and non-stationary operating data (such as turbine vibration signals and boiler thermal parameter fluctuations), often leading to missed or misjudged fault characteristics, which in turn causes unplanned equipment shutdowns and energy losses. Summary of the Invention

[0003] In view of the technical problems existing in the background art, the present invention proposes a data analysis and diagnosis method for thermal power plants.

[0004] To achieve the above objectives, the technical solution adopted by the present invention includes the following steps:

[0005] S1. Data Acquisition: Acquire real-time operating data during the production process of the thermal power plant, including boiler thermal parameters, turbine vibration signals, generator electrical parameters, and pipeline pressure and flow data;

[0006] S2. Dataset Partitioning: The dataset is divided according to the collection period. The specific implementation steps are as follows:

[0007] First, a collection period T is preset. The first round uses the first 3T of data as the initial training set. After each model training is completed, the training set is automatically accumulated with the data of the newly collected period.

[0008] For each newly added dataset, calculate the operating condition difference degree for the periodic data, and determine whether it is a high-value data period based on the operating condition difference degree;

[0009] Construct a gradient training set. During regular model training, use the training set accumulated over all cycles as the basic training. When high-value data is detected, additional high-value reinforcement training is triggered. Use the high-value data cycle combined with the data from the three most recent cycles to form a temporary small training set, and make a quick fine-tuning of the model.

[0010] Finally, the acquisition cycle is dynamically adjusted based on the data fluctuation level to achieve an adaptive update frequency for the training set.

[0011] S3. Data Augmentation Preprocessing: The collected training data is processed to obtain feature-enhanced data;

[0012] S4. Hybrid Neural Network Model Construction: Construct an LSTM-CNN hybrid model that includes a temporal attention mechanism, comprising an input layer, a dynamic temporal coding layer, a spatial feature fusion layer, an attention weight allocation layer, and an output layer.

[0013] Input layer: Used to receive data after feature enhancement;

[0014] Dynamic temporal coding layer: used to extract long-short-term dependency features of time series through bidirectional LSTM;

[0015] Spatial Feature Fusion Layer: Used to extract spatial layout-related features of the device using CNN convolutional kernels;

[0016] Attention weight allocation layer: used to construct a weight matrix and dynamically weight key parameter features;

[0017] Output layer: Used to transform the weighted feature vectors output by the attention weight allocation layer into structured results that can be used for anomaly detection;

[0018] S5. Model Training: Input the training set into the hybrid neural network model, adjust the model parameters by improving the particle swarm optimization algorithm, and obtain the trained model.

[0019] S6. Real-time detection: The model uses real-time data from the period in which it is used for prediction to make predictions and obtain the prediction results.

[0020] S7. Real-time Analysis: Analyze the results of model predictions in real-time detection to assess the current operating status of the thermal power plant.

[0021] Preferably, the calculation method for the operating condition difference includes feature difference measurement, weight allocation, and comprehensive calculation.

[0022] Feature difference measure: used to calculate the difference between various features of data in a measurement period;

[0023] Weighting: Using expert experience and the results of random forest feature importance analysis, weights are assigned to the differences in different features;

[0024] Comprehensive calculation: The differences in each feature are summed according to their corresponding weights to obtain the degree of difference in working conditions.

[0025] Preferably, the minimum acquisition period is implemented by dynamically adjusting based on the data fluctuation level, including: fluctuation quantification, level division, and period adjustment;

[0026] Volatility quantization: Calculates the variance of the data sliding window within the initially set period;

[0027] Classification: Based on the calculated variance threshold, the volatility equivalence is classified.

[0028] Cycle adjustment: Calculate the volatility level index according to different volatility levels, and multiply the previous cycle length by the volatility level index to obtain the new cycle length.

[0029] Preferably, step S3, data augmentation preprocessing, involves processing the collected training data to obtain feature-enhanced data. The specific implementation of this step is as follows:

[0030] First, the training data collected was used... Multi-scale decomposition is performed, and an improved empirical mode decomposition is adopted to construct an improved intrinsic mode basis function that satisfies: ,in The decomposition scale factor is represented by k, the time shift factor is represented by m, and the modal branch index is represented by m. Represents traditional empirical mode basis functions. This is represented by an improved basis function that incorporates frequency domain constraints. The basis function fusion coefficients are adaptively adjusted by minimizing the modal aliasing exponent.

[0031] After improving empirical mode decomposition, modal components with different size branches are obtained. Calculate the cross-correlation coefficients between each component and the original data. ,in Let covariance function be used. Represents the standard deviation function;

[0032] For the filtered modal components, an adaptive noise threshold is constructed by combining wavelet thresholding for denoising. Then, soft thresholding is applied to the wavelet coefficients. ,in, The coefficients before processing. The coefficients are represented by the processed values, and the denoised modal components are reconstructed using inverse wavelet transform. ;

[0033] For the denoised modal components Mode aliasing is suppressed by an adaptive modal component screening algorithm, and the energy distribution entropy of each mode is calculated. ,in, The energy probability distribution of the denoised modes is given, where X represents the data length. An energy distribution entropy threshold is set, and non-aliased modes with energy distribution entropy less than the threshold are retained to obtain the filtered mode components. ;

[0034] Use the same improved empirical mode basis functions as during decomposition. For the filtered modal components Reconstruction is performed to obtain feature-enhanced data. The calculation method is as follows: .

[0035] Preferably, the adaptive noise threshold The calculation method is as follows: ,in Here are the wavelet transform coefficients for the modal components, and N is the number of coefficients. This represents the noise sensitivity coefficient.

[0036] Preferably, the specific implementation of bidirectional LSTM extraction of time series features in the dynamic temporal coding layer includes sequence normalization, bidirectional feature extraction, and temporal dependency fusion.

[0037] Sequence standardization: The input feature enhancement data is standardized by channel, and the mean and standard deviation of each channel are calculated to obtain the standardized sequence;

[0038] Bidirectional feature extraction: Normalized sequences are processed in parallel using forward LSTM and backward LSTM;

[0039] Temporal dependency fusion: The forward and backward hidden states are concatenated, and the temporal dependency feature vector is output after the dimensionality is compressed through a fully connected layer.

[0040] Preferably, the model training process employs an improved particle swarm optimization algorithm to adjust the model parameters, including particle swarm encoding and initialization, dynamic weight iterative optimization, adaptive learning factor adjustment, and a hybrid local search mechanism.

[0041] Particle swarm encoding and initialization: Encode model parameters into particle positions, initialize particle swarm positions and velocities based on the optimal parameter distribution of historical scenes, and cover the parameter search space;

[0042] Dynamic weight iterative optimization: Inertial weights decay non-linearly with the training cycle, suppressing parameter oscillations;

[0043] Adaptive learning factor adjustment: Dynamically adjust individual and social learning factors based on the distance between the particle and the global optimal solution;

[0044] Hybrid local search mechanism: simulated annealing is triggered for local search of the globally optimal particle, and suboptimal solutions are accepted according to Boltzmann probability to avoid getting trapped in local optima.

[0045] Compared with the prior art, the advantages and positive effects of the present invention are as follows:

[0046] 1. Collect various types of real-time operating data from thermal power plants, dynamically adjust the collection cycle according to the fluctuation level, divide the dataset accordingly, calculate the operating condition difference, construct a gradient training set, and achieve adaptive updating of the training set cycle.

[0047] 2. The training data is decomposed into multiple scales, and improved empirical mode decomposition combined with wavelet thresholding is used for denoising. The adaptive mode component screening algorithm is used to suppress mode aliasing, resulting in feature-enhanced data.

[0048] 3. Construct an LSTM-CNN hybrid model that includes a temporal attention mechanism. After processing through multiple layers such as the input layer and the dynamic temporal coding layer, the weighted feature vector is transformed into a structured result for anomaly detection.

[0049] 4. An improved particle swarm optimization algorithm is used to adjust the model parameters, including particle swarm encoding and initialization mechanisms. After optimization, the model is monitored and analyzed in real time to evaluate the operating status of the thermal power plant. Attached Figure Description

[0050] To more clearly illustrate the technical solutions of the embodiments of the present invention, the drawings used in the following description of the embodiments will be briefly introduced. Obviously, the drawings described below are some embodiments of the present invention. For those skilled in the art, other drawings can be obtained based on these drawings without creative effort.

[0051] Figure 1 This is a schematic diagram of the structural flow of a data analysis and diagnostic method for a thermal power plant. Detailed Implementation

[0052] To better understand the above-mentioned objectives, features, and advantages of the present invention, the present invention will be further described below in conjunction with the accompanying drawings and embodiments. It should be noted that, unless otherwise specified, the embodiments and features described in these embodiments can be combined with each other.

[0053] Numerous specific details are set forth in the following description in order to provide a full understanding of the invention. However, the invention may also be practiced in other ways than those described herein, and therefore the invention is not limited to the specific embodiments disclosed in the following specification.

[0054] In this example, as the core equipment for energy conversion, the operating efficiency of a thermal power plant directly affects energy utilization and production safety. Existing fault diagnosis methods for thermal power plants typically rely on manual inspections and statistical analysis based on historical data. Most of these methods are ineffective when dealing with nonlinear or non-stationary data, easily overlooking transient anomalies or multi-parameter coupled faults. Therefore, accurately and in real-time diagnosing the operating status of thermal power plants, predicting faults in advance, and avoiding unplanned shutdowns are key issues for improving the operational safety and economy of thermal power plants. Based on this, this invention proposes a data analysis and diagnosis method for thermal power plants, the specific process of which is as follows: Figure 1As shown, this method, through the comprehensive application of technologies such as data acquisition, enhanced preprocessing, and hybrid neural network models, enables real-time monitoring, fault prediction, and diagnosis of the status of thermal power plant equipment, thereby improving equipment stability and energy utilization efficiency.

[0055] To achieve efficient monitoring of thermal power plants, this invention first involves data acquisition to obtain real-time operational data. This data includes key parameters such as boiler thermal parameters, turbine vibration signals, generator electrical parameters, and pipeline pressure and flow rates. By collecting this data, a comprehensive understanding of the power plant's operational status can be obtained, providing a foundation for subsequent data processing and model training.

[0056] Next, considering efficient data processing, this invention proposes an innovative dataset partitioning method. Unlike traditional dataset partitioning, which typically uses data from a fixed time period for model training and ignores the variability in operating conditions during equipment operation, this method addresses the issue of datasets containing irrelevant or redundant data that could lead to overfitting or inefficient training. Therefore, this invention partitions the dataset based on the collection cycle. The specific steps are as follows: First, a collection cycle T is preset. The first round uses the first 3T of data as the initial training set. After each model training iteration, the training set is automatically accumulated with data from a newly collected cycle. For each newly added cycle of data in the dataset, the operating condition variability is calculated, and a high-value data cycle is determined based on this variability. A gradient training set is constructed. During regular model training, the training set accumulated from all cycles is used as the base training. When high-value data is detected, additional high-value reinforcement training is initiated, using the high-value data cycle combined with the most recent three cycles to form a temporary small training set for rapid model fine-tuning. Finally, the collection cycle is dynamically adjusted based on the data fluctuation level, achieving an adaptive update frequency for the training set cycle. The calculation method for the operating condition difference degree includes: feature difference measurement, weight allocation, and comprehensive calculation; feature difference measurement: used to calculate the difference of each feature in the data of the measurement period; weight allocation: using expert experience and random forest feature importance analysis results, weights are assigned to different feature differences; comprehensive calculation: the differences of each feature are weighted and summed according to their corresponding weights to obtain the operating condition difference degree. The minimum acquisition period is implemented by dynamically adjusting based on data fluctuation level, including: fluctuation quantification, level division, and period adjustment; fluctuation quantification: calculating the variance of the data sliding window within the initially set period; level division: dividing the fluctuations equally according to the calculated variance threshold; period adjustment: calculating the fluctuation level index according to different fluctuation levels, and multiplying the previous period length by the fluctuation level index to obtain the new period length.

[0057] Next, in the data preprocessing stage, this invention processes the collected training data to obtain feature-enhanced data. Specifically, this involves first processing the collected training data... Multi-scale decomposition is performed, and an improved empirical mode decomposition is adopted to construct an improved intrinsic mode basis function that satisfies: ,in The decomposition scale factor is represented by k, the time shift factor is represented by m, and the modal branch index is represented by m. Represents traditional empirical mode basis functions. This is represented by an improved basis function that incorporates frequency domain constraints. The basis function fusion coefficients are adaptively adjusted by minimizing the modal aliasing exponent.

[0058] After improving empirical mode decomposition, modal components with different size branches are obtained. Calculate the cross-correlation coefficients between each component and the original data. ,in Let covariance function be used. The standard deviation function is used to represent the noise level; for the filtered modal components, wavelet thresholding is combined with noise reduction to construct an adaptive noise threshold. , ,in Here are the wavelet transform coefficients for the modal components, and N is the number of coefficients. The function is used to calculate the median of wavelet coefficients. This represents the noise sensitivity coefficient. Then, soft thresholding is applied to the wavelet coefficients. ,in, The coefficients before processing. The coefficients are represented by the processed values, and the denoised modal components are reconstructed using inverse wavelet transform. ; for the denoised modal components Mode aliasing is suppressed by an adaptive modal component screening algorithm, and the energy distribution entropy of each mode is calculated. ,in, The energy probability distribution of the denoised modes is given, where X represents the data length. An energy distribution entropy threshold is set, and non-aliased modes with energy distribution entropy less than the threshold are retained to obtain the filtered mode components. The same improved empirical mode basis functions as those used during decomposition are employed. For the filtered modal components Reconstruction is performed to obtain feature-enhanced data. The calculation method is as follows: .

[0059] The following describes the construction of a hybrid neural network model: an LSTM-CNN hybrid model incorporating a temporal attention mechanism, including an input layer, a dynamic temporal coding layer, a spatial feature fusion layer, an attention weight allocation layer, and an output layer. The input layer receives the feature-enhanced data; the dynamic temporal coding layer extracts long-short-term dependency features from the time series using bidirectional LSTM; the spatial feature fusion layer extracts spatial layout-related features using CNN convolutional kernels; the attention weight allocation layer constructs a weight matrix based on thermodynamic physics rules to dynamically weight key parameter features; and the output layer transforms the weighted feature vector output by the attention weight allocation layer into a structured result suitable for anomaly detection. Specifically, in constructing the LSTM-CNN hybrid model with a temporal attention mechanism, the input layer receives the preprocessed data after feature enhancement, laying the foundation for subsequent processing. The dynamic temporal coding layer utilizes a bidirectional LSTM to first standardize the input data by channel, then processes the standardized sequence in parallel using forward and backward LSTMs to extract long-term and short-term dependency features of the time series. Finally, the forward and backward hidden states are concatenated and compressed through a fully connected layer to output a temporal dependency feature vector. The spatial feature fusion layer uses CNN convolutional kernels to perform convolution operations on equipment spatial layout-related data, extracting spatial correlation features and achieving effective fusion of equipment spatial information. The attention weight allocation layer constructs a weight matrix, dynamically weighting features based on the importance of different parameters, making the model focus more on key parameters. The output layer transforms the weighted feature vector output by the attention weight allocation layer to generate structured results that can be directly used for anomaly detection, such as the probability of anomalies in equipment operating status, thereby achieving accurate diagnosis of the operating status of the thermal power plant.

[0060] The constructed model is then trained by inputting the training set into the hybrid neural network model. An improved particle swarm optimization (PSO) algorithm is used to adjust the model parameters, resulting in the trained model. This improved PSO algorithm adjusts the model parameters through particle swarm encoding and initialization, dynamic weight iterative optimization, adaptive learning factor adjustment, and a hybrid local search mechanism. Particle swarm encoding and initialization encodes model parameters as particle positions. Based on the optimal parameter distribution in historical scenarios, the particle swarm positions and velocities are initialized to cover the parameter search space. Dynamic weight iterative optimization uses non-linear decay of inertial weights over the training cycle to suppress parameter oscillations. Adaptive learning factor adjustment dynamically adjusts individual and social learning factors based on the distance between the particle and the global optimum. The hybrid local search mechanism triggers simulated annealing for local searches on the globally optimal particle, accepting suboptimal solutions according to Boltzmann probability to avoid getting trapped in local optima. Specifically, the improved particle swarm optimization (PSO) algorithm first performs PSO encoding and initialization during model parameter tuning. The parameters of the hybrid neural network model (such as weights and biases) are encoded as position vectors of particles in the search space. Based on the optimal parameter distribution patterns verified in historical scenarios, the position and velocity of the particle swarm are dynamically initialized, ensuring that the initial particle swarm uniformly covers the entire parameter search space and improving the comprehensiveness of the initial search. During dynamic weight iterative optimization, the inertia weight is set to decay non-linearly with the training cycle. A larger inertia weight is maintained in the early stages of training to enhance global search capability. As training progresses, the inertia weight gradually decreases to suppress parameter oscillations and focus on local fine-grained search, balancing search efficiency at different stages. During adaptive learning factor adjustment, the distance between the particle's current position and the global optimum is calculated in real time. When the distance is large, the social learning factor is increased to encourage particles to converge towards the global optimum; when the distance is small, the individual learning factor is increased to strengthen the particle's own exploration ability. By dynamically adjusting the weights of individual and social learning factors, the algorithm avoids getting trapped in local optima. The hybrid local search mechanism targets the globally optimal particle, triggering simulated annealing to perform a local search. It accepts suboptimal solutions according to the Boltzmann probability criterion and updates directly when a better solution is found. If a suboptimal solution is encountered, it is accepted with a certain probability, achieving a balance between optimization accuracy and escaping local optima. Ultimately, the efficiency and accuracy of model parameter optimization are improved through the synergy of multiple mechanisms.

[0061] Finally, the real-time data used in the prediction cycle is used for model prediction to obtain the model prediction results. The model prediction results in real-time detection are then analyzed to assess the current operating status of the thermal power plant.

[0062] The above description is merely a preferred embodiment of the present invention and is not intended to limit the present invention in any other way. Any person skilled in the art may make changes or modifications to the above-disclosed technical content to create equivalent embodiments for application in other fields. However, any simple modifications, equivalent changes, and modifications made to the above embodiments based on the technical essence of the present invention without departing from the scope of the present invention shall still fall within the protection scope of the present invention.

Claims

1. A data analysis and diagnostic method for thermal power plants, characterized in that, Includes the following steps: S1. Data Acquisition: Acquire real-time operating data during the production process of the thermal power plant, including boiler thermal parameters, turbine vibration signals, generator electrical parameters, and pipeline pressure and flow data; S2. Dataset Partitioning: The dataset is divided according to the collection period. The specific implementation steps are as follows: First, a collection period T is preset. The first round uses the first 3T of data as the initial training set. After each model training is completed, the training set is automatically accumulated with the data of the newly collected period. For each newly added dataset, calculate the operating condition difference degree for the periodic data, and determine whether it is a high-value data period based on the operating condition difference degree; Construct a gradient training set. During regular model training, use the training set accumulated over all cycles as the basic training. When high-value data is detected, additional high-value reinforcement training is triggered. Use the high-value data cycle combined with the data from the three most recent cycles to form a temporary small training set, and make a quick fine-tuning of the model. Finally, the acquisition cycle is dynamically adjusted based on the data fluctuation level to achieve an adaptive update frequency for the training set. S3. Data Augmentation Preprocessing: The collected training data is processed to obtain feature-enhanced data; S4. Hybrid Neural Network Model Construction: Construct an LSTM-CNN hybrid model that includes a temporal attention mechanism, comprising an input layer, a dynamic temporal coding layer, a spatial feature fusion layer, an attention weight allocation layer, and an output layer. Input layer: Used to receive data after feature enhancement; Dynamic temporal coding layer: used to extract long-short-term dependency features of time series through bidirectional LSTM; Spatial Feature Fusion Layer: Used to extract spatial layout-related features of the device using CNN convolutional kernels; Attention weight allocation layer: used to construct a weight matrix and dynamically weight key parameter features; Output layer: Used to transform the weighted feature vectors output by the attention weight allocation layer into structured results that can be used for anomaly detection; S5. Model Training: Input the training set into the hybrid neural network model, adjust the model parameters by improving the particle swarm optimization algorithm, and obtain the trained model. S6. Real-time detection: The model uses real-time data from the period in which it is used for prediction to make predictions and obtain the prediction results. S7. Real-time analysis: Analyze the results of model predictions in real-time detection to assess the current operating status of the thermal power plant; The specific implementation of step S3, data augmentation preprocessing, which involves processing the collected training data to obtain feature-enhanced data, is as follows: First, the training data collected was used... Multi-scale decomposition is performed, and an improved empirical mode decomposition is adopted to construct an improved intrinsic mode basis function that satisfies: ,in The decomposition scale factor is represented by k, the time shift factor is represented by m, and the modal branch index is represented by m. Represents traditional empirical mode basis functions. This is represented by an improved basis function that incorporates frequency domain constraints. The basis function fusion coefficients are adaptively adjusted by minimizing the modal aliasing exponent. After improving empirical mode decomposition, modal components with different size branches are obtained. Calculate the cross-correlation coefficients between each component and the original data. ,in Let covariance function be used. Represents the standard deviation function; For the filtered modal components, an adaptive noise threshold is constructed by combining wavelet thresholding for denoising. Then, soft thresholding is applied to the wavelet coefficients. ,in, The coefficients before processing. The coefficients are represented by the processed values, and the denoised modal components are reconstructed using inverse wavelet transform. ; For the denoised modal components Mode aliasing is suppressed by an adaptive modal component screening algorithm, and the energy distribution entropy of each mode is calculated. ,in, The energy probability distribution of the denoised modes is given, where X represents the data length. An energy distribution entropy threshold is set, and non-aliased modes with energy distribution entropy less than the threshold are retained to obtain the filtered mode components. ; Use the same improved empirical mode basis functions as during decomposition. For the filtered modal components Reconstruction is performed to obtain feature-enhanced data. The calculation method is as follows: .

2. The data analysis and diagnosis method for a thermal power plant according to claim 1, characterized in that, The calculation method for the difference in operating conditions is as follows: This includes feature difference measurement, weight allocation, and comprehensive calculation; Feature difference measure: used to calculate the difference between various features of data in a measurement period; Weighting: Using expert experience and the results of random forest feature importance analysis, weights are assigned to the differences in different features; Comprehensive calculation: The differences in each feature are summed according to their corresponding weights to obtain the degree of difference in working conditions.

3. The data analysis and diagnosis method for a thermal power plant according to claim 1, characterized in that, The data collection cycle is dynamically adjusted based on the data fluctuation level, including fluctuation quantification, level classification, and cycle adjustment. Volatility quantization: Calculates the variance of the data sliding window within the initially set period; Classification: Based on the calculated variance threshold, the volatility equivalence is classified. Cycle adjustment: Calculate the volatility level index according to different volatility levels, and multiply the previous cycle length by the volatility level index to obtain the new cycle length.

4. The data analysis and diagnosis method for a thermal power plant according to claim 1, characterized in that, The adaptive noise threshold The calculation method is as follows: ,in Here are the wavelet transform coefficients for the modal components, and N is the number of coefficients. This represents the noise sensitivity coefficient.

5. The data analysis and diagnosis method for a thermal power plant according to claim 1, characterized in that, The specific implementation of bidirectional LSTM extraction of time series features in the dynamic temporal coding layer includes sequence normalization, bidirectional feature extraction, and temporal dependency fusion. Sequence standardization: The input feature enhancement data is standardized by channel, and the mean and standard deviation of each channel are calculated to obtain the standardized sequence; Bidirectional feature extraction: Normalized sequences are processed in parallel using forward LSTM and backward LSTM; Temporal dependency fusion: The forward and backward hidden states are concatenated, and the temporal dependency feature vector is output after the dimensionality is compressed through a fully connected layer.

6. The data analysis and diagnosis method for a thermal power plant according to claim 1, characterized in that, In step S5, model training employs an improved particle swarm optimization algorithm to adjust model parameters, including particle swarm encoding and initialization, dynamic weight iterative optimization, adaptive learning factor adjustment, and a hybrid local search mechanism. Particle swarm encoding and initialization: Encode model parameters into particle positions, initialize particle swarm positions and velocities based on the optimal parameter distribution of historical scenes, and cover the parameter search space; Dynamic weight iterative optimization: Inertial weights decay non-linearly with the training cycle, suppressing parameter oscillations; Adaptive learning factor adjustment: Dynamically adjust individual and social learning factors based on the distance between the particle and the global optimal solution; Hybrid local search mechanism: simulated annealing is triggered for local search of the globally optimal particle, and suboptimal solutions are accepted according to Boltzmann probability to avoid getting trapped in local optima.

Images