Power plant equipment fault prediction method based on time sequence large model

Abstract

The invention discloses a power plant equipment fault prediction method based on a time sequence large model, and the method comprises the following steps: S1, collecting and preprocessing the time sequence data of a multi-source sensor of a power plant, and generating a standardized time sequence data set; s2, constructing a time sequence large model, inputting standardized data, and outputting a future operation state predicted value; s3, comparing the running state prediction value with an actual measurement value to generate a prediction residual sequence; s4, constructing a Bayesian neural network model, inputting a prediction residual sequence, and outputting error probability distribution; s5, optimizing a Bayesian neural network structure and hyper-parameters by adopting an ant colony optimization algorithm; s6, confidence interval estimation is executed, and whether the state is a high-risk state or not is judged; and S7, outputting a running state label, and dynamically acquiring a data closed-loop updating model. According to the invention, high-precision prediction and uncertainty evaluation of the operation state of the power plant equipment are realized, so that the accuracy and response time efficiency of fault early warning are improved.

Description

Technical Field

[0001] The present invention relates to the technical field of power equipment operating status monitoring and intelligent prediction, and in particular to a power plant equipment fault prediction method based on a large time series model. Background Art

[0002] In the operation and maintenance of power systems, power plant equipment serves as a core supporting facility, and the stability of its operating status directly impacts the safety and economic viability of the power grid. With the continuous improvement of equipment intelligence and the increasing abundance of operational data collected by sensors, the power industry is focusing on how to efficiently predict equipment failures from massive amounts of multi-source data. Traditional power plant equipment fault diagnosis methods rely heavily on rule bases, expert experience, or shallow machine learning models, making it difficult to accurately capture the nonlinear evolution and temporal dependencies of equipment operation. This severely limits prediction accuracy and adaptability, especially when dealing with complex, variable, and dynamic operating conditions.

[0003] In recent years, deep learning technology has been gradually introduced into the field of equipment health prediction. Time series models such as long short-term memory (LSTM) and gated recurrent neural networks (GRU) have, to a certain extent, alleviated the problem of information forgetting. However, these models still face bottlenecks in their ability to model long time series, resulting in low training efficiency and high computational costs, making them difficult to deploy effectively in large-scale industrial scenarios. Furthermore, existing methods, which mostly rely on point predictions for model output, fail to provide uncertainty measures for operational status predictions, making it difficult to support reliable risk assessment and early warning strategy development.

[0004] Furthermore, most existing deep learning models employ fixed structures and hyperparameter configurations, requiring extensive human experience in their design and lacking the ability to adaptively tune for varying device characteristics. Commonly used grid search or random search strategies for model selection and parameter setting for complex tasks are computationally expensive and inefficient, failing to integrate intelligent optimization methods for structural-level global search. Regarding uncertainty modeling, Bayesian neural networks, as probabilistic modeling tools, offer theoretical advantages, but their complex structure and difficulty in training make efficient construction and tuning solutions unavailable in practical applications.

[0005] Therefore, building an equipment failure prediction method with strong time series modeling capabilities, support for predictive uncertainty expression, and the ability to adaptively optimize structure and parameters has become a key challenge in current technological development. This paper addresses these challenges by proposing a power plant equipment failure prediction solution based on a large time series model that integrates a Bayesian neural network and an ant colony optimization algorithm to improve the model's accuracy, robustness, and intelligent decision-making capabilities.

[0006] Therefore, how to provide a method for predicting power plant equipment failures based on a large time series model is an urgent problem that those skilled in the art need to solve. Summary of the Invention

[0007] One purpose of the present invention is to propose a method for predicting power plant equipment failures based on a large time series model. The present invention comprehensively utilizes time series modeling technology, Bayesian neural network and ant colony optimization algorithm, and describes in detail the complete process of realizing equipment failure prediction and uncertainty assessment in a multi-source time series data environment. It has the advantages of high prediction accuracy, strong adaptability and high risk judgment reliability.

[0008] A method for predicting power plant equipment failure based on a time series large model according to an embodiment of the present invention includes the following steps:

[0009] S1. Collect multi-source sensor time series data of various key equipment in the power plant, pre-process the multi-source sensor time series data, and generate a standardized time series data set;

[0010] S2. Build a large time series model, input the standardized time series data set into the large time series model to predict the equipment operating status, and output the predicted operating status values ​​of the target equipment in multiple future time steps based on the long-term dependency modeling mechanism;

[0011] S3, performing difference calculation between the operation status prediction value and the actual measurement value sequence of the target device to obtain a prediction residual sequence;

[0012] S4. Construct a Bayesian neural network model, receive the prediction residual sequence as input, and output the probability distribution parameters of the prediction error, including the prediction mean and prediction variance;

[0013] S5. Use the ant colony optimization algorithm to jointly optimize the structural parameters and training hyperparameters of the Bayesian neural network model to optimize the modeling ability of the Bayesian neural network;

[0014] S6. Use the optimized Bayesian neural network model to perform confidence interval estimation on the operating status prediction value of the time series large model, and combine it with the preset fault identification threshold to judge the equipment operating status. When the prediction lower limit is lower than the fault threshold and the variance is less than the uncertainty upper limit, it is marked as a high-risk state;

[0015] S7. Output the equipment operation status label according to the fault probability level, including four levels: normal, suspicious, critical fault, and failed. Collect new operation data to update the model parameters, and re-execute the ant colony optimization process within the preset period to achieve dynamic adaptive update and closed-loop feedback control of equipment fault prediction.

[0016] Optionally, the multimodal biosignals specifically include electromyography, electroencephalography, heart rate and skin electrical response, which are used to reflect muscle fatigue, neural response, emotional fluctuations and concentration during the performance.

[0017] Optionally, the preprocessing of the multi-source sensor time series data specifically includes performing missing value filling, outlier removal, time alignment, denoising filtering and normalization processing on the multi-source sensor time series data.

[0018] Optionally, the S2 specifically includes:

[0019] S21, record the standardized time series data set as input sequence X = {x1, x2, ..., x T}, where x t represents the multidimensional device state vector at the t-th time step, T represents the length of the time series, and d represents the dimension of the sensor variables contained in each time step;

[0020] S22. Input the input sequence X into the time series large model. The time series large model is an informer structure based on the sparse attention mechanism. The time series large model structure includes an input encoding module, a position embedding module, a multi-head sparse attention module, a feedforward network module, and a prediction output module.

[0021] S23. In the input encoding stage, embedding mapping is performed on the input sequence to obtain the embedded representation E = {e1, e2, ..., e T}, where e t =f embed (x t ), f embed is the embedded function;

[0022] S24. In the attention calculation stage, the sparse self-attention mechanism is used to calculate the attention weight matrix A. The attention is expressed as:

[0023]

[0024] Among them, Q i =e i W Q , K j =e j W K , W Q ,W K is the trainable weight matrix, d k is the attention dimension;

[0025] S25. Calculate the context representation C = A·V based on the sparse attention matrix A and the embedding sequence E, where V = E·W V , W V is the value vector transformation matrix, C represents the context encoding result of each time step;

[0026] S26, input the context representation C into the feedforward network module and the prediction output module, and output the target device's operating state prediction value sequence in the next τ time steps

[0027]

[0028] in, Represents the multivariate predicted value of the target device at the T+kth time step.

[0029] Optionally, the S3 specifically includes:

[0030] S31, receiving a device operating state prediction value sequence output by a large time series model within a target prediction period, wherein the device operating state prediction value sequence includes multidimensional operating parameters of multiple time steps;

[0031] S32. Collecting a sequence of actual measurement values ​​of the target device within a corresponding prediction time period, wherein the sequence of actual measurement values ​​is consistent with the sequence of predicted values ​​of the device operating status in terms of the number of time steps and dimensional structure;

[0032] S33, matching the equipment operation status prediction value sequence with the actual measurement value sequence at each time step to establish a time step mapping relationship;

[0033] S34, performing the difference calculation operation between the predicted value and the actual measured value in sequence according to the time step mapping relationship;

[0034] S35, arranging the difference results at each time step in chronological order to form a complete prediction residual sequence;

[0035] S36. Encapsulate the prediction residual sequence in a unified format as a data set for Bayesian neural network model training and input.

[0036] Optionally, the S4 specifically includes:

[0037] S41, initializing the neural network model structure, setting the neural network model structure to include an input layer, at least two fully connected hidden layers, and an output layer, wherein the input layer receives the prediction residual sequence ε, and each input node corresponds to a multidimensional variable of a time step residual;

[0038] S42, for each connection weight parameter in the neural network and bias parameters Probabilistic modeling is introduced separately, the prior distribution is specified as normal distribution, and the weights satisfy Bias Satisfaction Where l represents the lth layer of the network, μ and σ 2 To initialize the hyperparameters;

[0039] S43. During each forward propagation process, each parameter is sampled from its currently estimated variational distribution, and the sampled value is used to replace the fixed weight input into the network to form the forward operation expression of the probabilistic neural network;

[0040] S44. During the training process, the optimization objective is defined as minimizing the sum of the log-likelihood loss term of the prediction residual and the Kullback-Leibler divergence term, which constitutes the loss function form under variational inference:

[0041]

[0042] Among them, θ is all network parameters, q(θ) is the variational posterior distribution of the parameters, and p(θ) is its prior distribution. is the target loss function of the Bayesian neural network, is the expectation operator under the variational distribution q(θ), log2 is the logarithmic function, KL[q(θ)‖p(θ)] is the Kullback-Leibler divergence between the variational posterior distribution q(θ) and the prior distribution p(θ);

[0043] S45. Use stochastic gradient variational inference and reparameterization techniques to train the network, sample parameters for each batch of input data, and update the mean and standard deviation of the network variational distribution;

[0044] S46. After the training is completed, multiple forward propagation samplings are performed on the input prediction residual sequence to output a set of probability distribution parameters of the residual.

[0045] Optionally, the S5 specifically includes:

[0046] S51, set the ant colony optimization search space, encode and combine the structural parameters of the Bayesian neural network with the training hyperparameters, the structural parameters include the number of network layers L, the number of nodes in each layer N, the activation function type set A, the training hyperparameters include the learning rate η, the regularization coefficient λ, the prior distribution parameter Among them, μ0 is the mean of the prior distribution of all parameters in the Bayesian neural network, is the variance of the prior distribution of all parameters in the Bayesian neural network;

[0047] S52. Initialize the ant colony individuals and let the path of each ant individual be represented by a set of model parameter combinations:

[0048]

[0049] Where M is the number of individuals in the ant colony;

[0050] S53, define the fitness function F(P (k) ), the loss function value of the Bayesian neural network on the training set and the confidence interval coverage index γ of the validation set (k) The joint construction is the target evaluation function:

[0051]

[0052] Among them, α, β are weighting coefficients, is the loss function value, γ (k) is the proportion of the predicted confidence interval containing the true value in the validation set;

[0053] S54, according to the path selection mechanism of the ant colony algorithm, according to the individual fitness function value F(P (k) ) Update the pheromone matrix τ and the transition probability matrix π, and perform path probability update and individual selection:

[0054]

[0055] Among them, π i,j represents the probability of the i-th ant moving from state i to state j, η i,j represents the heuristic factor, Available is the set of paths consisting of all unvisited nodes that are feasible when the current ant starts from state i;

[0056] S55. Perform multiple rounds of iterations to continuously update the individual paths of the ant colony and obtain the global optimal parameter combination.

[0057] S56. Combine the optimal parameters Applied to the Bayesian neural network model, it constructs the optimized network structure and training configuration that are ultimately used for predictive residual modeling and uncertainty quantification.

[0058] Optionally, the S6 specifically includes:

[0059] S61, using the optimized Bayesian neural network model to receive the operating status prediction value output by the time series large model as input to perform residual uncertainty reasoning;

[0060] S62: Perform multiple forward propagation sampling on the predicted value of each time step to obtain multiple prediction output results for statistically analyzing the distribution characteristics of the prediction results of the time step;

[0061] S63. Based on the multiple output results, respectively calculate the predicted mean and upper and lower confidence limits for each time step to form a confidence interval estimation result of the operating status;

[0062] S64, calling the set fault identification threshold range, and comparing the lower limit of the confidence interval of each time step with the fault threshold;

[0063] S65. When the lower limit of the confidence interval is less than the preset fault identification threshold, and the prediction variance of the corresponding time step is less than the set uncertainty upper limit, mark the time step prediction as a high-risk state;

[0064] S66. Record the equipment operating status corresponding to the high-risk state time step as early warning data for use in state classification, early warning level judgment, and control system response.

[0065] The beneficial effects of the present invention are:

[0066] The present invention achieves accurate prediction of the operating status of power plant equipment and high-confidence risk judgment by constructing a fault prediction architecture that integrates a large time series model with a Bayesian neural network, and introduces an ant colony optimization algorithm to jointly tune the network structure parameters and training hyperparameters. Based on the fact that traditional models only have single-point prediction functions and are difficult to express prediction uncertainty, the present invention uses a Bayesian modeling mechanism to estimate the probability distribution of prediction errors, and can output prediction results with confidence intervals, effectively improving the ability to identify equipment failure trends early. In addition, the large time series model is based on a sparse attention mechanism and has the ability to model long-term dependent sequences, significantly enhancing the global correlation perception of multivariate time series data under complex dynamic conditions.

[0067] The ant colony optimization algorithm is used to dynamically search the number of neural network structure layers, the number of neuron nodes, the form of activation functions, and the regularization parameters, thus avoiding the problems of low efficiency and weak generalization of the traditional manual parameter adjustment method. This enables the model to realize structural adaptive configuration according to the characteristics of different power plant equipment, thereby improving the robustness and migration ability of the model during actual deployment. At the same time, in the post-processing stage of the prediction results, by setting a dynamic fault threshold and an uncertainty upper limit dual-condition judgment mechanism, not only can high-risk states be accurately marked, but also the false alarm rate and missed alarm rate can be effectively reduced. In summary, the present invention not only improves the accuracy of the prediction of the operating status of power plant equipment, but also enhances the interpretability and reliability of the model, providing strong technical support for the intelligent operation and maintenance of power systems. BRIEF DESCRIPTION OF THE DRAWINGS

[0068] The accompanying drawings are used to provide a further understanding of the present invention and constitute a part of the specification. Together with the embodiments of the present invention, they are used to explain the present invention and do not constitute a limitation of the present invention. In the accompanying drawings:

[0069] Figure 1 This is a flow chart of a method for predicting power plant equipment failure based on a large time series model proposed by the present invention;

[0070] Figure 2 This is a structural diagram of the prediction residual probability modeling and output confidence interval of the Bayesian neural network of the power plant equipment failure prediction method based on the time series large model proposed by the present invention. DETAILED DESCRIPTION

[0071] The present invention will now be described in further detail with reference to the accompanying drawings, which are simplified schematic diagrams that illustrate the basic structure of the present invention in a schematic manner.

[0072] refer to Figure 1 and Figure 2 A method for predicting power plant equipment failure based on a time series large model includes the following steps:

[0073] S1. Collect multi-source sensor time series data of various key equipment in the power plant, pre-process the multi-source sensor time series data, and generate a standardized time series data set;

[0074] S2. Build a large time series model, input the standardized time series data set into the large time series model to predict the equipment operating status, and output the predicted operating status values ​​of the target equipment in multiple future time steps based on the long-term dependency modeling mechanism;

[0075] S3, performing difference calculation between the operation status prediction value and the actual measurement value sequence of the target device to obtain a prediction residual sequence;

[0076] S4. Construct a Bayesian neural network model, receive the prediction residual sequence as input, and output the probability distribution parameters of the prediction error, including the prediction mean and prediction variance;

[0077] S5. Use the ant colony optimization algorithm to jointly optimize the structural parameters and training hyperparameters of the Bayesian neural network model to optimize the modeling ability of the Bayesian neural network;

[0078] S6. Use the optimized Bayesian neural network model to perform confidence interval estimation on the operating status prediction value of the time series large model, and combine it with the preset fault identification threshold to judge the equipment operating status. When the prediction lower limit is lower than the fault threshold and the variance is less than the uncertainty upper limit, it is marked as a high-risk state;

[0079] S7. Output the equipment operation status label according to the fault probability level, including four levels: normal, suspicious, critical fault, and failed. Collect new operation data to update the model parameters, and re-execute the ant colony optimization process within the preset period to achieve dynamic adaptive update and closed-loop feedback control of equipment fault prediction.

[0080] In this embodiment, the multimodal biosignals specifically include electromyography, electroencephalography, heart rate, and galvanic skin response, which are used to reflect muscle fatigue, neural response, emotional fluctuations, and concentration during the performance.

[0081] In this embodiment, the preprocessing of the multi-source sensor time series data specifically includes performing missing value filling, outlier removal, time alignment, denoising filtering and normalization processing on the multi-source sensor time series data.

[0082] In this embodiment, S2 specifically includes:

[0083] S21, record the standardized time series data set as input sequence X = {x1, x2, ..., x T}, where x t represents the multidimensional device state vector at the t-th time step, T represents the length of the time series, and d represents the dimension of the sensor variables contained in each time step;

[0084] S22. Input the input sequence X into the time series large model. The time series large model is an informer structure based on the sparse attention mechanism. The time series large model structure includes an input encoding module, a position embedding module, a multi-head sparse attention module, a feedforward network module, and a prediction output module.

[0085] S23. In the input encoding stage, embedding mapping is performed on the input sequence to obtain the embedded representation E = {e1, e2, ..., e T}, where e t =f embed (x t ), f embed is the embedded function;

[0086] S24. In the attention calculation stage, the sparse self-attention mechanism is used to calculate the attention weight matrix A. The attention is expressed as:

[0087]

[0088] Among them, Q i =e i W Q , K j =e j W K , W Q ,W K is the trainable weight matrix, d k is the attention dimension;

[0089] S25. Calculate the context representation C = A·V based on the sparse attention matrix A and the embedding sequence E, where V = E·W V , W V is the value vector transformation matrix, C represents the context encoding result of each time step;

[0090] S26, input the context representation C into the feedforward network module and the prediction output module, and output the target device's operating state prediction value sequence in the next τ time steps

[0091]

[0092] in, Represents the multivariate predicted value of the target device at the T+kth time step.

[0093] In this embodiment, S3 specifically includes:

[0094] S31, receiving a device operating state prediction value sequence output by a large time series model within a target prediction period, wherein the device operating state prediction value sequence includes multidimensional operating parameters of multiple time steps;

[0095] S32. Collecting a sequence of actual measurement values ​​of the target device within a corresponding prediction time period, wherein the sequence of actual measurement values ​​is consistent with the sequence of predicted values ​​of the device operating status in terms of the number of time steps and dimensional structure;

[0096] S33, matching the equipment operation status prediction value sequence with the actual measurement value sequence at each time step to establish a time step mapping relationship;

[0097] S34, performing the difference calculation operation between the predicted value and the actual measured value in sequence according to the time step mapping relationship;

[0098] S35, arranging the difference results at each time step in chronological order to form a complete prediction residual sequence;

[0099] S36. Encapsulate the prediction residual sequence in a unified format as a data set for Bayesian neural network model training and input.

[0100] In this embodiment, the S4 specifically includes:

[0101] S41, initializing the neural network model structure, setting the neural network model structure to include an input layer, at least two fully connected hidden layers, and an output layer, wherein the input layer receives the prediction residual sequence ε, and each input node corresponds to a multidimensional variable of a time step residual;

[0102] S42, for each connection weight parameter in the neural network and bias parameters Probabilistic modeling is introduced separately, the prior distribution is specified as normal distribution, and the weights satisfy Bias Satisfaction Where l represents the lth layer of the network, μ and σ 2 To initialize the hyperparameters;

[0103] S43. During each forward propagation process, each parameter is sampled from its currently estimated variational distribution, and the sampled value is used to replace the fixed weight input into the network to form the forward operation expression of the probabilistic neural network;

[0104] S44. During the training process, the optimization objective is defined as minimizing the sum of the log-likelihood loss term of the prediction residual and the Kullback-Leibler divergence term, which constitutes the loss function form under variational inference:

[0105]

[0106] Among them, θ is all network parameters, q(θ) is the variational posterior distribution of the parameters, and p(θ) is its prior distribution. is the target loss function of the Bayesian neural network, is the expectation operator under the variational distribution q(θ), log2 is the logarithmic function, KL[q(θ)‖p(θ)] is the Kullback-Leibler divergence between the variational posterior distribution q(θ) and the prior distribution p(θ);

[0107] S45. Use stochastic gradient variational inference and reparameterization techniques to train the network, sample parameters for each batch of input data, and update the mean and standard deviation of the network variational distribution;

[0108] S46. After the training is completed, multiple forward propagation samplings are performed on the input prediction residual sequence to output a set of probability distribution parameters of the residual.

[0109] In this embodiment, the S5 specifically includes:

[0110] S51, set the ant colony optimization search space, encode and combine the structural parameters of the Bayesian neural network with the training hyperparameters, the structural parameters include the number of network layers L, the number of nodes in each layer N, the activation function type set A, the training hyperparameters include the learning rate η, the regularization coefficient λ, the prior distribution parameter Among them, μ0 is the mean of the prior distribution of all parameters in the Bayesian neural network, is the variance of the prior distribution of all parameters in the Bayesian neural network;

[0111] S52. Initialize the ant colony individuals and let the path of each ant individual be represented by a set of model parameter combinations:

[0112]

[0113] Where M is the number of individuals in the ant colony;

[0114] S53, define the fitness function F(P (k)), the loss function value of the Bayesian neural network on the training set and the confidence interval coverage index γ of the validation set (k) The joint construction is the target evaluation function:

[0115]

[0116] Among them, α, β are weighting coefficients, is the loss function value, γ (k) is the proportion of the predicted confidence interval containing the true value in the validation set;

[0117] S54, according to the path selection mechanism of the ant colony algorithm, according to the individual fitness function value F(P (k) ) Update the pheromone matrix τ and the transition probability matrix π, and perform path probability update and individual selection:

[0118]

[0119] Among them, π i,j represents the probability of the i-th ant moving from state i to state j, η i,j represents the heuristic factor, Available is the set of paths consisting of all unvisited nodes that are feasible when the current ant starts from state i;

[0120] S55. Perform multiple rounds of iterations to continuously update the individual paths of the ant colony and obtain the global optimal parameter combination.

[0121] S56. Combine the optimal parameters Applied to the Bayesian neural network model, it constructs the optimized network structure and training configuration that are ultimately used for predictive residual modeling and uncertainty quantification.

[0122] In this embodiment, S6 specifically includes:

[0123] S61, using the optimized Bayesian neural network model to receive the operating status prediction value output by the time series large model as input to perform residual uncertainty reasoning;

[0124] S62: Perform multiple forward propagation sampling on the predicted value of each time step to obtain multiple prediction output results for statistically analyzing the distribution characteristics of the prediction results of the time step;

[0125] S63. Based on the multiple output results, respectively calculate the predicted mean and upper and lower confidence limits for each time step to form a confidence interval estimation result of the operating status;

[0126] S64, calling the set fault identification threshold range, and comparing the lower limit of the confidence interval of each time step with the fault threshold;

[0127] S65. When the lower limit of the confidence interval is less than the preset fault identification threshold, and the prediction variance of the corresponding time step is less than the set uncertainty upper limit, mark the time step prediction as a high-risk state;

[0128] S66. Record the equipment operating status corresponding to the high-risk state time step as early warning data for use in state classification, early warning level judgment, and control system response.

[0129] Example 1:

[0130] To verify the feasibility of this invention, it was applied to a large coal-fired power plant in a certain province. The plant's maintenance department faced the challenge of identifying operational anomalies in key equipment, such as boilers, steam turbines, and generators, in advance. This was particularly true in environments with sustained high temperatures, where multi-source sensor data frequently fluctuated. Traditional prediction methods based on models such as thresholds or support vector machines were unable to fully capture the temporal evolution of equipment, often only issuing low-confidence warnings one to two hours before a fault occurred. This resulted in high false alarm and missed alarm rates, severely impacting operational efficiency.

[0131] To enhance intelligent equipment monitoring capabilities, the power plant implemented the "Power Plant Equipment Fault Prediction Method Based on a Time Series Large Model," a method proposed in this paper. A pilot system was deployed using boiler A, turbine B, generator C, condenser D, and pump station E as test objects. The system connected six types of high-frequency sampling sensors, including vibration, temperature, current, and voltage. Using an informer-structured time series large model, the system established a multi-time-step prediction model for the operating status of each device type over the next 12 hours. The predicted outputs were then modeled using a Bayesian neural network residual, and the risk level was determined using uncertainty metrics and confidence interval assessment mechanisms.

[0132] To verify the performance of the model, the maintenance team compared the actual performance of the method of the present invention with that of the traditional prediction algorithm from June 1, 2024 to July 31, 2024. The results show that the method of the present invention is significantly superior to the traditional method in multiple indicators. In terms of prediction accuracy, the average accuracy of the present invention reaches 91.2%, while the traditional method is only 77.8%; in terms of advance warning time, the average accuracy of the present invention can reach 4.2 hours, which is about 2.5 times higher than that of the traditional method; the false alarm rate and the missed alarm rate are controlled below 7% and below 6% respectively, which is significantly better than the traditional algorithm which is generally above 15%; at the same time, since the model structure is obtained through automatic search by ant colony optimization, the overall training time is shortened to 30 to 40 minutes, which significantly improves the deployment efficiency.

[0133] This example demonstrates the significant advantages of this invention in improving fault prediction accuracy, early warning timeliness, and risk reliability. It is particularly suitable for real-world environments with complex data, numerous devices, and variable power plant operating conditions, providing strong technical support for intelligent operation and maintenance of power systems. The system has been operating stably in the pilot project area for over two months, successfully warning of five potential equipment anomalies during this period, all of which were promptly addressed, preventing actual losses.

[0134] Table 1 Comparison of power plant equipment failure prediction performance

[0135]

[0136] The performance indicators listed in Table 1 systematically compare the actual performance of traditional methods and the proposed method in predicting typical power plant equipment failures. The analysis shows that the proposed method significantly outperforms traditional methods in multiple key dimensions, demonstrating its technical advantages in complex time series modeling, risk assessment accuracy, and deployment efficiency.

[0137] First, in terms of prediction accuracy, the proposed method achieved over 89% accuracy across all five types of equipment, reaching a maximum of 94%. Conventional methods generally achieved accuracy rates between 75% and 80%. For steam turbine B, for example, the conventional method achieved an accuracy of 0.755, while the proposed method improved this to 0.912. This demonstrates that the combined modeling of a large time series model and a Bayesian neural network can more effectively capture the underlying dynamics of equipment operation.

[0138] In terms of early warning time, the method of the present invention has an average advance warning capability of 3.5 to 5.0 hours, while traditional methods generally only have an advance warning capability of 1.2 to 2.0 hours, an improvement of more than 2 times. For example, for generator C, the traditional model can only issue an alarm 1.89 hours in advance on average, while the method of the present invention can issue a high-confidence warning 4.25 hours in advance, providing more sufficient emergency response time for power plant operations and maintenance.

[0139] The present invention also demonstrates significant advantages in terms of false alarm and missed alarm rates. While traditional methods typically experience a false alarm rate exceeding 15%, the present invention limits this to 4% to 8%. The missed alarm rate is also reduced from the traditional 12% to 18% to between 3% and 7%, significantly reducing wasted maintenance resources and potential risk omissions, and improving the stability and reliability of the prediction system. For example, in boiler A, the traditional method had a false alarm and missed alarm rate of 0.186 and 0.157, respectively, while the present invention reduced these rates to 0.063 and 0.045, respectively.

[0140] In terms of the width of the prediction confidence interval, the present invention outputs more concentrated confidence estimates through Bayesian modeling, reducing the interval width to between 1.2 and 2.0, while the interval width of traditional methods is generally between 5.4 and 6.1, indicating that the present invention not only improves the prediction accuracy, but also enhances the controllability of the uncertainty of the prediction output.

[0141] Finally, in terms of model training time, due to the introduction of ant colony optimization to jointly tune the structure and parameters, the method of the present invention shortens the training time to 30 to 40 minutes, which is better than the traditional method of 45 to 60 minutes of training time, and is conducive to achieving rapid iteration and periodic model updates in actual deployment.

[0142] In summary, the tabular data fully verifies the technical effectiveness and practical value of the method of the present invention in the task of fault prediction of key equipment in power plants. It is particularly suitable for operating status monitoring systems in industrial scenarios that have high requirements for high reliability, strong adaptability and real-time feedback capabilities.

[0143] The above description is only a preferred specific embodiment of the present invention, but the scope of protection of the present invention is not limited thereto. Any technician familiar with the technical field, within the technical scope disclosed by the present invention, who makes equivalent replacements or changes based on the technical solution and inventive concept of the present invention, should be covered by the scope of protection of the present invention.

Claims

1. A method for predicting power plant equipment failure based on a large time series model, characterized in that: The steps include: S1. Collect multi-source sensor time series data of various key equipment in the power plant, pre-process the multi-source sensor time series data, and generate a standardized time series data set; S2. Build a large time series model, input the standardized time series data set into the large time series model to predict the equipment operating status, and output the predicted operating status values ​​of the target equipment in multiple future time steps based on the long-term dependency modeling mechanism; S3, performing difference calculation between the operation status prediction value and the actual measurement value sequence of the target device to obtain a prediction residual sequence; S4. Construct a Bayesian neural network model, receive the prediction residual sequence as input, and output the probability distribution parameters of the prediction error, including the prediction mean and prediction variance; S5. Use the ant colony optimization algorithm to jointly optimize the structural parameters and training hyperparameters of the Bayesian neural network model to optimize the modeling ability of the Bayesian neural network; S6. Use the optimized Bayesian neural network model to perform confidence interval estimation on the operating status prediction value of the time series large model, and combine it with the preset fault identification threshold to judge the equipment operating status. When the prediction lower limit is lower than the fault threshold and the variance is less than the uncertainty upper limit, it is marked as a high-risk state; S7. Output the equipment operation status label according to the fault probability level, including four levels: normal, suspicious, critical fault, and failed. Collect new operation data to update the model parameters, and re-execute the ant colony optimization process within the preset period to achieve dynamic adaptive update and closed-loop feedback control of equipment fault prediction.

2. The method for predicting power plant equipment failure based on a time series large model according to claim 1, characterized in that: The multimodal biosignals specifically include electromyography, electroencephalography, heart rate, and skin electrical response, which are used to reflect muscle fatigue, neural response, emotional fluctuations, and concentration during the performance.

3. The method for predicting power plant equipment failure based on a time series large model according to claim 1, characterized in that: The preprocessing of the multi-source sensor time series data specifically includes performing missing value filling, outlier removal, time alignment, denoising filtering and normalization processing on the multi-source sensor time series data.

4. The method for predicting power plant equipment failure based on a time series large model according to claim 1, characterized in that: The S2 specifically includes: S21, record the standardized time series data set as input sequence X = {x1, x2, ..., x T }, where x t represents the multidimensional device state vector at the t-th time step, T represents the length of the time series, and d represents the dimension of the sensor variables contained in each time step; S22. Input the input sequence X into the time series large model. The time series large model is an informer structure based on the sparse attention mechanism. The time series large model structure includes an input encoding module, a position embedding module, a multi-head sparse attention module, a feedforward network module, and a prediction output module. S23. In the input encoding stage, embedding mapping is performed on the input sequence to obtain the embedded representation E = {e1, e2, ..., e T }, where e t =f embed (x t ), f embed is the embedded function; S24. In the attention calculation stage, the sparse self-attention mechanism is used to calculate the attention weight matrix A. The attention is expressed as: Among them, Q i =e i W Q , K j =e j W K , W Q ,W K is the trainable weight matrix, d k is the attention dimension; S25. Calculate the context representation C = A·V based on the sparse attention matrix A and the embedding sequence E, where V = E·W V , W V is the value vector transformation matrix, C represents the context encoding result of each time step; S26, input the context representation C into the feedforward network module and the prediction output module, and output the target device's operating state prediction value sequence in the next τ time steps in, Represents the multivariate predicted value of the target device at the T+kth time step.

5. The method for predicting power plant equipment failure based on a time series large model according to claim 1, characterized in that: The S3 specifically includes: S31, receiving a device operating state prediction value sequence output by a large time series model within a target prediction period, wherein the device operating state prediction value sequence includes multidimensional operating parameters of multiple time steps; S32. Collecting a sequence of actual measurement values ​​of the target device within a corresponding prediction time period, wherein the sequence of actual measurement values ​​is consistent with the sequence of predicted values ​​of the device operating status in terms of the number of time steps and dimensional structure; S33, matching the equipment operation status prediction value sequence with the actual measurement value sequence at each time step to establish a time step mapping relationship; S34, performing the difference calculation operation between the predicted value and the actual measured value in sequence according to the time step mapping relationship; S35, arranging the difference results at each time step in chronological order to form a complete prediction residual sequence; S36. Encapsulate the prediction residual sequence in a unified format as a data set for Bayesian neural network model training and input.

6. The method for predicting power plant equipment failure based on a time series large model according to claim 1, characterized in that: The S4 specifically includes: S41, initializing the neural network model structure, setting the neural network model structure to include an input layer, at least two fully connected hidden layers, and an output layer, wherein the input layer receives the prediction residual sequence ε, and each input node corresponds to a multidimensional variable of a time step residual; S42, for each connection weight parameter in the neural network and bias parameters Probabilistic modeling is introduced separately, the prior distribution is specified as normal distribution, and the weights satisfy Bias Satisfaction Where l represents the lth layer of the network, μ and σ 2 To initialize the hyperparameters; S43. During each forward propagation process, each parameter is sampled from its currently estimated variational distribution, and the sampled value is used to replace the fixed weight input into the network to form the forward operation expression of the probabilistic neural network; S44. During the training process, the optimization objective is defined as minimizing the sum of the log-likelihood loss term of the prediction residual and the Kullback-Leibler divergence term, which constitutes the loss function form under variational inference: Among them, θ is all network parameters, q(θ) is the variational posterior distribution of the parameters, and p(θ) is its prior distribution. is the target loss function of the Bayesian neural network, is the expectation operator under the variational distribution q(θ), log2 is the logarithmic function, KL[q(θ)‖p(θ)] is the Kullback-Leibler divergence between the variational posterior distribution q(θ) and the prior distribution p(θ); S45. Use stochastic gradient variational inference and reparameterization techniques to train the network, sample parameters for each batch of input data, and update the mean and standard deviation of the network variational distribution; S46. After the training is completed, multiple forward propagation samplings are performed on the input prediction residual sequence to output a set of probability distribution parameters of the residual.

7. The method for predicting power plant equipment failure based on a time series large model according to claim 1, characterized in that: The S5 specifically includes: S51, set the ant colony optimization search space, encode and combine the structural parameters of the Bayesian neural network with the training hyperparameters, the structural parameters include the number of network layers L, the number of nodes in each layer N, the activation function type set A, the training hyperparameters include the learning rate η, the regularization coefficient λ, the prior distribution parameter Among them, μ0 is the mean of the prior distribution of all parameters in the Bayesian neural network, is the variance of the prior distribution of all parameters in the Bayesian neural network; S52. Initialize the ant colony individuals and let the path of each ant individual be represented by a set of model parameter combinations: Where M is the number of individuals in the ant colony; S53, define the fitness function F(P (k) ), the loss function value of the Bayesian neural network on the training set and the confidence interval coverage index γ of the validation set (k) The joint construction is the target evaluation function: Among them, α, β are weighting coefficients, is the loss function value, γ (k) is the proportion of the predicted confidence interval containing the true value in the validation set; S54, according to the path selection mechanism of the ant colony algorithm, according to the individual fitness function value F(P (k) ) Update the pheromone matrix τ and the transition probability matrix π, and perform path probability update and individual selection: Among them, π i,j represents the probability of the i-th ant moving from state i to state j, η i,j represents the heuristic factor, Available is the set of paths consisting of all unvisited nodes that are feasible when the current ant starts from state i; S55. Perform multiple rounds of iterations to continuously update the individual paths of the ant colony and obtain the global optimal parameter combination. S56. Combine the optimal parameters Applied to the Bayesian neural network model, it constructs the optimized network structure and training configuration that are ultimately used for predictive residual modeling and uncertainty quantification.

8. The method for predicting power plant equipment failure based on a time series large model according to claim 1, characterized in that: The S6 specifically includes: S61, using the optimized Bayesian neural network model to receive the operating status prediction value output by the time series large model as input to perform residual uncertainty reasoning; S62: Perform multiple forward propagation sampling on the predicted value of each time step to obtain multiple prediction output results for statistically analyzing the distribution characteristics of the prediction results of the time step; S63. Based on the multiple output results, respectively calculate the predicted mean and upper and lower confidence limits for each time step to form a confidence interval estimation result of the operating status; S64, calling the set fault identification threshold range, and comparing the lower limit of the confidence interval of each time step with the fault threshold; S65. When the lower limit of the confidence interval is less than the preset fault identification threshold, and the prediction variance of the corresponding time step is less than the set uncertainty upper limit, mark the time step prediction as a high-risk state; S66. Record the equipment operating status corresponding to the high-risk state time step as early warning data for use in state classification, early warning level judgment, and control system response.

Images