Thermal power plant data analysis and diagnosis method

Abstract

The invention 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 acquired through data acquisition, a data set is divided according to an acquisition period, the period is dynamically adjusted, a working condition difference degree is calculated to construct a gradient training set, data enhancement preprocessing is completed through improved empirical mode decomposition and wavelet threshold de-noising, an LSTM-CNN hybrid model containing a time sequence attention mechanism is constructed, and an LSTM-CNN hybrid model containing a time sequence attention mechanism is constructed. And performing real-time detection and analysis after optimizing parameters by using an improved particle swarm algorithm. According to the method, the problem that a traditional manual inspection and statistical method is insufficient in nonlinear and non-stationary data processing is solved, transient abnormity and multi-parameter coupling faults of equipment can be accurately captured, the accuracy and the real-time performance of thermal power plant operation state diagnosis are improved, and the energy efficiency and the production safety are guaranteed.

Description

TECHNICAL FIELD

[0001] The 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. BACKGROUND

[0002] As a key hub of energy conversion, the precise analysis and fault diagnosis of the operation state of a thermal power plant is crucial for ensuring energy efficiency and production safety. At present, the data analysis and diagnosis of a thermal power plant often relies on manual experience combined with traditional statistical methods, such as recording equipment parameters through regular inspection and performing trend analysis based on historical data tables. This approach has significant limitations: on the one hand, the frequency and coverage of manual inspection are limited, making it difficult to capture transient abnormalities and multi-parameter coupled faults of equipment; on the other hand, traditional statistical methods have insufficient processing capabilities for nonlinear and non-stationary operation data (such as steam turbine vibration signals and boiler thermal parameter fluctuations), often leading to missed or misjudged fault characteristics, and thus causing unplanned equipment shutdowns and energy losses. SUMMARY

[0003] The present application proposes a thermal power plant data analysis and diagnosis method to address the technical problems in the background art.

[0004] To achieve the above purpose, the technical solution adopted by the present application comprises the following steps

[0005] S1, data acquisition: acquiring real-time operation data in the production process of a thermal power plant, including boiler thermal parameters, steam turbine vibration signals, generator electrical parameters and pipe network pressure and flow data;

[0006] S2, data set division: dividing the data set according to the collection period, and the specific implementation steps are as follows:

[0007] First, a collection period T is preset, and the first 3T data is used as the initial training set for the first round of training. After each model training is completed, the training set automatically adds one period of newly collected data;

[0008] For each newly added period of data, the working condition difference degree is calculated, and it is judged whether it is a high-value data period based on the working condition difference degree;

[0009] A gradient training set is constructed. In normal model training, the training set accumulated by the full period is used as the basis for training. When high-value data is detected, additional high-value reinforcement training is performed, and a temporary small training set is formed by combining the high-value data period with the last three periods of data to make a quick fine-tuning of the model;

[0010] Finally, the collection period is dynamically adjusted based on the data fluctuation level to realize self-adaptive update frequency of the training set period;

[0011] S3, data enhancement preprocessing: processing the collected data for training to obtain feature-enhanced data;

[0012] S4, hybrid neural network model construction: constructing an LSTM-CNN hybrid model containing a time sequence attention mechanism, including an input layer, a dynamic time sequence encoding layer, a spatial feature fusion layer, an attention weight distribution layer, and an output layer;

[0013] Input layer: used to receive feature-enhanced data;

[0014] Dynamic time sequence encoding layer: used to extract time sequence long and short term dependence features through bidirectional LSTM;

[0015] Spatial feature fusion layer: used to extract device spatial layout correlation features using CNN convolution kernel;

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

[0017] Output layer: used to convert the weighted feature vector output by the attention weight distribution layer into a structured result 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 through the improved particle swarm algorithm, and obtain the trained model;

[0019] S6, real-time detection: model prediction of real-time data in the cycle for model prediction, obtaining the model prediction result;

[0020] S7, real-time analysis: analyzing the model prediction result in real-time detection for evaluating the current thermal power plant operation state.

[0021] As a preferred, the calculation method of the working condition difference degree includes feature difference measurement, weight distribution and comprehensive calculation;

[0022] Feature difference measurement: used to calculate the difference of each feature of the cycle data;

[0023] Weight distribution: using expert experience and random forest feature importance analysis results to assign weights to different feature differences;

[0024] Comprehensive calculation: summing up each feature difference according to the corresponding weight to obtain the working condition difference degree.

[0025] As a preferred, the minimum collection cycle is achieved by dynamically adjusting based on data fluctuation level, including fluctuation quantization, level division and cycle adjustment;

[0026] Fluctuation quantization: calculating the variance of the data sliding window in the initially set cycle;

[0027] Ranking: According to the calculated variance threshold, the fluctuation equivalent is divided;

[0028] Period adjustment: According to different fluctuation levels, the fluctuation level index is calculated, and the new period length is obtained by multiplying the previous period length and the fluctuation level index.

[0029] As preferred, the step S3 data enhancement preprocessing: the specific implementation of processing the collected data for training to obtain feature-enhanced data is:

[0030] First, the collected training data is processed Multi-scale decomposition is performed, improved empirical mode decomposition is adopted, and improved intrinsic mode basis function is constructed to meet: Wherein represents the decomposition scale factor, k is the time translation factor, m is the mode branch index, represents the traditional empirical mode basis function, represents the improved basis function by introducing the frequency domain constraint, is the basis function fusion coefficient, which is adaptively adjusted by minimizing the mode aliasing index;

[0031] After improved empirical mode decomposition, different size-branch mode components are obtained, and the cross-correlation coefficient of each component and the original data is calculated , wherein is the covariance function, represents the standard deviation function;

[0032] For the screened mode components, wavelet threshold denoising is combined to construct an adaptive noise threshold Then, soft threshold processing is performed on the wavelet coefficients , wherein, is the coefficient before processing, represents the coefficient after processing, and the denoised mode component is reconstructed by inverse wavelet transform ;

[0033] For the denoised mode component , the adaptive mode component screening algorithm is used to suppress mode aliasing, and the energy distribution entropy of each mode is calculated , wherein, is the energy probability distribution of the denoised mode, X represents the data length, and the energy distribution entropy threshold is set. The non-aliasing mode with energy distribution entropy less than the threshold is retained to obtain the screened mode component ;

[0034] The same improved empirical mode basis function as in the decomposition is used to the screened mode component reconstructing to obtain feature-enhanced data , the calculation method is: .

[0035] As preferred, the adaptive noise threshold The calculation method is: , wherein is the modal component wavelet transform coefficient, N is the number of coefficients, Indicates the noise sensitivity coefficient.

[0036] As preferred, the specific implementation of the bidirectional LSTM in the dynamic time sequence encoding layer to extract the time sequence features includes sequence standardization, bidirectional feature extraction and time sequence dependence fusion.

[0037] Sequence standardization: the input feature-enhanced data is divided into channels, the mean and standard deviation of each channel are calculated, and the standardized sequence is obtained;

[0038] Bidirectional feature extraction: the standardized sequence is processed by forward LSTM and backward LSTM in parallel;

[0039] Time sequence dependence fusion: the forward and backward hidden states are spliced, and the time sequence dependence feature vector is output after dimension compression by the full connection layer.

[0040] As preferred, the improved particle swarm algorithm is used in the model training to adjust the model parameters, including particle swarm coding and initialization, dynamic weight iterative optimization, adaptive learning factor adjustment and hybrid local search mechanism.

[0041] Particle swarm coding and initialization: the model parameters are coded as particle positions, the particle swarm positions and velocities are initialized based on the historical scene optimal parameter distribution, and the parameter search space is covered;

[0042] Dynamic weight iterative optimization: the inertia weight is nonlinearly decayed with the training period, and the parameter oscillation is suppressed;

[0043] Adaptive learning factor adjustment: according to the distance between the particle and the global optimal solution, the individual and social learning factors are dynamically adjusted;

[0044] Hybrid local search mechanism: the global optimal particle triggers the simulated annealing local search, and the suboptimal solution is accepted according to the Boltzmann probability, so as to avoid falling into local optimum.

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

[0046] 1. Collecting multiple types of real-time operation data of the thermal power plant, dynamically adjusting the collection period according to the fluctuation level, dividing the data set and calculating the working condition difference degree, constructing a gradient training set, and realizing the adaptive update of the training set period.

[0047] 2. Multi-scale decomposition is performed on the training data, improved empirical mode decomposition is combined with wavelet threshold denoising, and mode aliasing is suppressed by an adaptive mode component screening algorithm to obtain feature-enhanced data.

[0048] 3. An LSTM-CNN hybrid model containing a time sequence attention mechanism is constructed, and the weighted feature vector is converted into an abnormal detection structured result through multiple layers of processing such as an input layer and a dynamic time sequence encoding layer.

[0049] 4. An improved particle swarm algorithm is used to adjust the model parameters, including particle swarm coding and initialization mechanisms, and the optimized model is used for real-time detection and analysis to evaluate the operation state of the thermal power plant. BRIEF DESCRIPTION OF DRAWINGS

[0050] In order to more clearly illustrate the technical solutions of the embodiments of the present application, the following will briefly introduce the drawings needed in the embodiment description. Obviously, the drawings in the following description are some embodiments of the present application, and other drawings can also be obtained by those skilled in the art without creative labor.

[0051] Figure 1 It is a structure flow diagram of a thermal power plant data analysis and diagnosis method. DETAILED DESCRIPTION

[0052] In order to more clearly illustrate the technical solutions of the embodiments of the present application, the following will briefly introduce the drawings needed in the embodiment description. Obviously, the drawings in the following description are some embodiments of the present application, and other drawings can also be obtained by those skilled in the art without creative labor.

[0053] In the following description, many specific details are set forth in order to provide a thorough understanding of the present application, however, the present application can also be implemented in other ways different from those described herein, therefore, the present application is not limited to the specific embodiments disclosed in the following description.

[0054] Embodiment, thermal power plant as the core equipment of energy conversion, its operation efficiency directly affects the energy utilization rate and production safety. The existing thermal power plant fault diagnosis method usually relies on manual inspection and statistical analysis based on historical data, and most of the methods are not good when facing nonlinear and non-stationary data, and it is easy to miss the transient abnormality of the equipment or the multi-parameter coupling fault. Therefore, how to accurately and real-time diagnose the operation state of the thermal power plant, predict the fault in advance, and avoid unplanned shutdown is the key problem to improve the operation safety and economy of the thermal power plant. Based on this, the present application proposes a thermal power plant data analysis and diagnosis method, and the specific process is as follows Figure 1The method realizes real-time monitoring, fault prediction and diagnosis of the equipment state of the thermal power plant by comprehensively applying data acquisition, enhanced preprocessing, mixed neural network model and other technologies, and further improves the stability of the equipment and the energy utilization efficiency.

[0055] First, in order to realize efficient monitoring of the thermal power plant, the present application first acquires real-time operation data of the thermal power plant. These data include key parameters such as boiler thermal parameters, turbine vibration signals, generator electrical parameters, and pipe network pressure and flow. By collecting these data, the running state of the thermal power plant can be comprehensively understood, providing a basis for subsequent data processing and model training.

[0056] Then, considering efficient processing of data, the present application proposes an innovative data set division method. Unlike traditional data set division, which usually uses fixed time period data to train the model, it ignores the difference in working condition fluctuation during equipment operation. The data set may contain many irrelevant or redundant data, which may cause overfitting of the model or low training efficiency. Therefore, in the present application, the data set is divided according to the collection period. The specific implementation steps are as follows: first, preset the collection period T, and use the first 3T data as the initial training set in the first round. When each model training is completed, the training set automatically adds a new collection period data. For each new added data set period data, calculate the working condition difference degree, and judge whether it is a high value data period based on the working condition difference degree. Build a gradient training set. When training the normal model, use the training set accumulated by the full period as the basis for training. When high-value data is detected, additional high-value reinforcement training is performed using high-value data periods combined with the last three period data to form a temporary small training set for a quick fine-tuning of the model. Finally, the collection period is dynamically adjusted based on the data fluctuation level to realize adaptive update frequency of the training set period. The calculation method of the working condition difference degree includes feature difference measurement, weight allocation and comprehensive calculation. Feature difference measurement is used to calculate the difference of each feature in the period data. Weight allocation uses expert experience and random forest feature importance analysis results to allocate weights to different feature differences. Comprehensive calculation: sum the feature differences according to the corresponding weights to get the working condition difference degree. The minimum collection period is dynamically adjusted based on the data fluctuation level, including fluctuation quantization, level division and period adjustment. Fluctuation quantization: calculate the variance of the data sliding window in the initially set period. Level division: divide the fluctuation equivalent according to the calculated variance threshold. Period adjustment: according to different fluctuation levels, calculate the fluctuation level index, multiply the length of the previous period by the fluctuation level index to get the new period length.

[0057] Then, in the data preprocessing, the collected data for training is processed to obtain feature-enhanced data. The specific implementation is as follows: first, the collected data for training is preprocessed to remove noise and outliers, and then the feature extraction is performed. Multi-scale decomposition is performed, and an improved intrinsic mode basis function is constructed to meet: wherein represents a decomposition scale factor, k is a time translation factor, and m is a modal branch index, represents a traditional empirical mode basis function, represents an improved basis function with frequency domain constraints, is a basis function fusion coefficient, which is adaptively adjusted by minimizing the modal aliasing index;

[0058] After improved empirical mode decomposition, modal components of different sizes-branches are obtained , and the cross-correlation coefficients of each component and the original data are calculated wherein is a covariance function, represents a standard deviation function; for the screened modal components, wavelet threshold denoising is combined to construct an adaptive noise threshold , wherein is a wavelet transform coefficient of the modal component, and N is the number of coefficients, is a median function used to calculate the wavelet coefficient, represents a noise sensitivity coefficient. Then the wavelet coefficients are processed by soft thresholding wherein, is the coefficient before processing, represents the coefficient after processing, and the denoised modal component is reconstructed by inverse wavelet transform ; for the denoised modal component , modal aliasing is suppressed by an adaptive modal component screening algorithm, and the energy distribution entropy of each modal is calculated wherein, is the energy probability distribution of the denoised modal, X represents the data length, and an energy distribution entropy threshold is set to retain non-aliasing modes with energy distribution entropy less than the threshold, thereby obtaining the screened modal components ; the same improved empirical mode basis function as in decomposition is used to reconstruct the screened modal components , thereby obtaining the feature-enhanced data , and the calculation method is: .

[0059] The following is the construction of the hybrid neural network model: an LSTM-CNN hybrid model containing a time sequence attention mechanism is constructed, including an input layer, a dynamic time sequence encoding layer, a spatial feature fusion layer, an attention weight distribution layer, and an output layer; the input layer is used to receive the data after feature enhancement; the dynamic time sequence encoding layer is used to extract the long and short term dependence features of the time sequence by bidirectional LSTM; the spatial feature fusion layer is used to extract the device spatial layout correlation features by using CNN convolution kernel; the attention weight distribution layer is used to dynamically weight the key parameter features by constructing a weight matrix based on the thermodynamic physical rules; the output layer is used to convert the weighted feature vector output by the attention weight distribution layer into a structured result that can be used for anomaly detection. Specifically, when constructing the LSTM-CNN hybrid model containing the time sequence attention mechanism, the input layer is responsible for receiving the data after feature enhancement preprocessing, laying the foundation for subsequent processing. The dynamic time sequence encoding layer uses bidirectional LSTM to first standardize the input data by channel, then processes the standardized sequence in parallel through forward and backward LSTM, extracts the long and short term dependence features of the time sequence, finally splices the forward and backward hidden states and compresses the dimensions through a fully connected layer, and outputs the time sequence dependence feature vector. The spatial feature fusion layer uses CNN convolution kernel to perform convolution operation on the device spatial layout related data, extracts the spatial correlation features therein, and realizes effective fusion of the device spatial information. The attention weight distribution layer constructs a weight matrix, dynamically weights the importance of different parameter features, so that the model pays more attention to the key parameters. The output layer converts the weighted feature vector output by the attention weight distribution layer to generate a structured result that can be directly used for anomaly detection, such as the abnormal probability of the device running state, so as to realize accurate diagnosis of the running state of the thermal power plant.

[0060] The constructed model is trained below. The training set is input into the hybrid neural network model, and the model parameters are adjusted by the improved particle swarm algorithm to obtain the trained model. The improved particle swarm algorithm for adjusting the model parameters includes particle swarm coding and initialization, dynamic weight iterative optimization, adaptive learning factor adjustment, and hybrid local search mechanism. Particle swarm coding and initialization: encode the model parameters as particle positions, initialize the particle swarm positions and velocities based on the optimal parameter distribution of the historical scene, and cover the parameter search space; dynamic weight iterative optimization: the inertia weight nonlinearly decays with the training period, and the parameter oscillation is suppressed; adaptive learning factor adjustment: dynamically adjust the individual and social learning factors according to the distance between the particle and the global optimal solution; hybrid local search mechanism: trigger the simulated annealing local search for the global optimal particle, accept the suboptimal solution according to the Boltzmann probability, and avoid falling into the local optimum. Specifically, in the model parameter adjustment of the improved particle swarm algorithm, first, the particle swarm coding and initialization are performed. The parameters (such as weights, biases, etc.) of the hybrid neural network model are encoded as the position vectors of particles in the search space. Based on the verified optimal parameter distribution law in the historical scene, the positions and velocities of the particle swarm are dynamically initialized to make the initial particle swarm uniformly cover the entire parameter search space, improving the comprehensiveness of the initial search. In the dynamic weight iterative optimization, the inertia weight is set to nonlinearly decay with the training period. In the early stage of training, a larger inertia weight is maintained to enhance the global search ability. As the training progresses, the inertia weight gradually decreases to suppress parameter oscillation and focus on local fine search, balancing the search efficiency at different stages. In the adaptive learning factor adjustment process, the distance between the current position of the particle and the global optimal solution is calculated in real time. When the distance is large, the social learning factor is increased to encourage the particle to gather in the direction of 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 is prevented from falling into local optimum. The hybrid local search mechanism triggers the simulated annealing algorithm for local search for the global optimal particle, accepts the suboptimal solution according to the Boltzmann probability criterion, and directly updates when a better solution is found. If a suboptimal solution is encountered, it is accepted with a certain probability, balancing the optimization accuracy and jumping out of the local optimum. Finally, the real-time data in the period for model prediction are used for model prediction to obtain the model prediction results, and the model prediction results in real-time detection are analyzed to evaluate the current operation state of the thermal power plant.

[0061] Finally, the real-time data in the period for model prediction are used for model prediction to obtain the model prediction results, and the model prediction results in real-time detection are analyzed to evaluate the current operation state of the thermal power plant.

[0062] The above merely describes preferred embodiments of the present application, but is not intended to limit the present application to other forms, any person skilled in the art can make changes or modifications to the above disclosed technical contents into equivalent embodiments with equivalent changes, and apply to other fields, but any simple modification, equivalent change and modification made to the above embodiments according to the technical essence of the present application without departing from the technical solution content of the present application still belongs to the protection scope of the present application technical solution.

Claims

1. A thermal power plant data analysis and diagnosis method, characterized in that: The following steps are involved: S1. Data acquisition: Acquire real-time operating data during the thermal power plant's production process, including boiler thermal parameters, turbine vibration signals, generator electrical parameters, and pipe network pressure and flow data; S2. Dataset division: Divide the data set according to the collection period. The specific implementation steps are as follows: First, preset the collection period T. In the first round, use the first 3T data as the initial training set. After each model training, the training set automatically accumulates the newly collected data of one period. For each newly added data set, calculate the working condition difference and determine whether it is a high-value data period based on the working condition difference. Construct a gradient training set. During regular model training, the training set accumulated over all cycles is used as the base training set. When high-value data is detected, additional high-value reinforcement training is initiated. The high-value data cycle is combined with the data from the three most recent cycles to form a temporary small training set, allowing for quick fine-tuning of the model. Finally, the acquisition cycle is dynamically adjusted based on the data fluctuation level to achieve the adaptive update frequency of the training set cycle; S3, data enhancement preprocessing: Process the collected data for training to obtain feature-enhanced data; S4. Hybrid neural network model construction: Build an LSTM-CNN hybrid model with temporal attention mechanism, including input layer, dynamic temporal encoding layer, spatial feature fusion layer, attention weight allocation layer, and output layer; Input layer: used to receive feature-enhanced data; Dynamic temporal encoding layer: used to extract long-term and short-term dependency features of time series through bidirectional LSTM; Spatial feature fusion layer: used to extract device spatial layout related features using CNN convolution kernel; Attention weight allocation layer: used to construct a weight matrix and dynamically weight key parameter features; Output layer: used to convert the weighted feature vector output by the attention weight allocation layer into a structured result 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 algorithm, and obtain the trained model; S6. Real-time detection: Perform model prediction on the real-time data in the period used for model prediction to obtain the model prediction result; S7. Real-time analysis: Analyze the results of model prediction in real-time detection to evaluate the current operating status of the thermal power plant.

2. A thermal power plant data analysis and diagnosis method according to claim 1, characterized in that: The calculation method of the working condition difference is: Including feature difference measurement, weight allocation and comprehensive calculation; Feature difference measurement: used to calculate the difference between each feature of the measurement period data; Weight assignment: Use expert experience and random forest feature importance analysis results to assign weights to different feature differences; Comprehensive calculation: sum up the differences of each feature according to the corresponding weight to obtain the working condition difference.

3. A thermal power plant data analysis and diagnosis method according to claim 1, characterized in that: The minimum collection period is dynamically adjusted based on the data fluctuation level, including: fluctuation quantification, level division and period adjustment; Volatility quantification: Calculate the variance of the sliding window of data within the initially set period; Classification: Volatility is divided into different levels based on the calculated variance threshold; 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. A thermal power plant data analysis and diagnosis method according to claim 1, characterized in that: The specific implementation of step S3 data enhancement preprocessing: processing the collected data for training to obtain feature-enhanced data is as follows: First, the training data for Perform multi-scale decomposition, use improved empirical mode decomposition, and construct improved intrinsic mode basis functions to meet the following requirements: ,in represents the decomposition scale factor, k is the time translation factor, m is the modal branch index, represents the traditional empirical mode basis functions, represents the improved basis function that introduces frequency domain constraints. is the basis function fusion coefficient, which is adaptively adjusted by minimizing the modal aliasing index; After improving the empirical mode decomposition, modal components of different size branches are obtained , calculate the mutual correlation coefficient between each component and the original data ,in is the covariance function, represents the standard deviation function; For the filtered modal components, combined with wavelet threshold denoising, an adaptive noise threshold is constructed , and then perform soft threshold processing on the wavelet coefficients ,in, is the coefficient before treatment, Represents the processed coefficients, and the denoised modal components are reconstructed by inverse wavelet transform ; The modal components after denoising , suppress modal aliasing through adaptive modal component screening algorithm, and calculate the energy distribution entropy of each mode ,in, is the energy probability distribution of the denoised mode, X represents the data length, set the energy distribution entropy threshold, retain the non-aliased mode with energy distribution entropy less than the threshold, and obtain the filtered modal component ; The same improved empirical mode basis function as that used in decomposition is used , for the modal components after screening Reconstruct and obtain feature-enhanced data , calculated as: .

5. A thermal power plant data analysis and diagnosis method according to claim 4, characterized in that: The adaptive noise threshold The calculation method is: ,in is the wavelet transform coefficient of the modal component, N is the number of coefficients, Represents the noise sensitivity coefficient.

6. A thermal power plant data analysis and diagnosis method according to claim 1, characterized in that: The specific implementation of the bidirectional LSTM in the dynamic time series encoding layer to extract time series features includes sequence normalization, bidirectional feature extraction and time series dependency fusion; Sequence normalization: perform channel-wise normalization on the input feature-enhanced data, calculate the mean and standard deviation of each channel, and obtain a standardized sequence; Bidirectional feature extraction: normalized sequences are processed in parallel by forward LSTM and backward LSTM; Temporal dependency fusion: The forward and backward hidden states are concatenated, and the dimension is compressed through a fully connected layer to output a temporal dependency feature vector.

7. The thermal power plant data analysis and diagnosis method according to claim 1, characterized in that: In the step S5, the improved particle swarm algorithm is used to adjust the model parameters, including particle swarm encoding and initialization, dynamic weight iterative optimization, adaptive learning factor adjustment and hybrid local search mechanism; Particle swarm encoding and initialization: Encode model parameters as particle positions, initialize particle swarm positions and velocities based on the optimal parameter distribution of historical scenarios, and cover the parameter search space; Dynamic weight iterative optimization: The inertia weight decays nonlinearly with the training cycle, suppressing parameter oscillation; Adaptive learning factor adjustment: Dynamically adjust individual and social learning factors based on the distance between particles and the global optimal solution; Hybrid local search mechanism: Trigger simulated annealing local search for the global optimal particle, accept suboptimal solutions according to Boltzmann probability, and avoid falling into local optimality.

Images