A thermal power plant working condition anomaly detection method based on space-time relation learning

Abstract

The application discloses a thermal power plant working condition anomaly detection method based on space-time relation learning, which comprises the following steps: obtaining macro state data and micro monitoring data of the thermal power plant; inputting the macro state data into a pre-trained state extraction model, generating state weights for representing the current running state through embedding coding and pattern mining; inputting the micro monitoring data into an anomaly detection model based on space-time relation learning, and outputting anomaly scores reflecting the equipment working condition; aligning the state weights and the anomaly scores on the time axis, and calculating the fusion result based on a fusion strategy; comparing the fusion result with a preset threshold value, if the fusion result is greater than the threshold value, determining that the current time is a working condition anomaly, otherwise, it is normal. The application overcomes the defects that the existing time series multivariate anomaly detection method is difficult to deeply mine the complex patterns of time trends and the relationship between features, does not consider macro data indicators, and has poor generalization ability in different scenes.

Description

Technical Field

[0001] This invention relates to the fields of deep neural networks, graph learning, spatiotemporal data analysis, and anomaly detection, specifically to a method for detecting anomalies in the operating conditions of thermal power plants based on spatiotemporal relationship learning. Background Technology

[0002] With the continuous growth of energy demand and the improvement of the intelligence level of the power system, the operational safety and economy of thermal power plants, as the basic guarantee unit for power supply, are receiving increasing attention. Thermal power plants are complex industrial systems composed of boilers, steam turbines, generators, and various auxiliary equipment. These equipment operate under high temperature, high pressure, and high speed conditions for extended periods, making sub-optimal or abnormal situations unavoidable. Abnormal operation of various equipment not only affects power generation efficiency and increases maintenance costs, but may also trigger unplanned outages or even safety accidents, posing a direct threat to grid stability and power supply reliability.

[0003] In actual operation, the data collected by thermal power plants mainly includes the following: (1) Microscopic monitoring data, which are collected in real time by sensors, including various key performance indicators (KPIs) such as temperature, pressure, flow, vibration, current, and power of each equipment; (2) Macroscopic status data, including daily plant-level and unit-level power generation, coal consumption, and operating rate. The above characteristic indicators are continuously collected at a certain frequency to form massive, high-dimensional, and strongly coupled multivariate time series data. The data contains both the normal operation mode of the system and the early abnormal signs and fault evolution characteristics. However, due to the complex equipment mechanism, variable operating conditions, and numerous influencing factors, traditional monitoring methods based on single indicator thresholds are often unable to effectively identify composite, gradual, or correlated anomalies, and false alarms and missed alarms occur frequently, making it difficult for power plant operation and maintenance personnel to intervene in a timely manner for accurate handling.

[0004] Furthermore, the operational data of thermal power plants exhibits significant periodicity, trends, and operational condition correlations over time, with complex physical couplings and statistical correlations among variables. Anomalies may manifest as sudden changes in a single indicator, coordinated deviations of multiple indicators, or gradual instability of the overall system's operational characteristics. Faced with such a high-dimensional, dynamic, and nonlinear data environment, developing an intelligent detection method capable of automatically, rapidly, and accurately identifying abnormal operating conditions and possessing interpretability has become an urgent need to improve the operation and maintenance level of thermal power plants, achieve predictive maintenance, and ensure safe and economical operation.

[0005] The task of detecting anomalies in thermal power plant operating conditions follows a multivariate time-series data anomaly detection paradigm. Existing related technologies mainly include traditional detection methods based on statistics and machine learning, and detection models based on deep learning techniques. Representative traditional detection methods include HBOS, KNN, and LOF based on sample distance; IFOres based on isolated tree structures; and COPOD and Fast-MCD based on distribution density. Limited by their linear and feature-channel-independent processing methods, these methods struggle to extract the joint anomaly patterns and long-term / short-term trends among multiple indicator features, thus limiting their detection effectiveness.

[0006] In recent years, a series of deep learning models have been proposed to further learn complex correlation patterns between time domains and features (spatial domains). These models are generally divided into two categories: prediction-based methods and reconstruction-based methods. Reconstruction-based anomaly detection models project samples into a latent space, learn latent representations of normal patterns, and then attempt to recover the original input to minimize the reconstruction error of normal samples. InterFusion uses a hierarchical variational autoencoder to learn the relationship between the time and spatial domains, ensuring that the latent space is random, smooth, continuous, and regularized, thus improving the robustness of the model. USAD addresses the problem that the reconstruction errors of anomalies and normal samples are close and cannot be effectively distinguished. It designs an encoder-decoder-based adversarial training framework, allowing the model to receive anomalous patterns during the training phase, thereby amplifying the reconstruction error of anomalous samples. MSCRED proposes a multi-scale convolutional recurrent encoder-decoder framework, which constructs a multi-resolution signature matrix to characterize the state evolution of the system at different time granularities and uses a deep convolutional structure to simultaneously capture spatial relationships between variables and sequential temporal features. Finally, it achieves anomaly detection based on the multi-dimensional fusion of reconstruction errors and further locates root cause features.

[0007] Prediction-based anomaly detection models focus on learning the long-term and short-term dependencies of multivariate time series to predict future trends, and then use the error between the predicted and true values ​​to identify anomalies. LSTM-NDT extracts a time series for each feature variable, performs single-step prediction using a Long Short-Term Memory network, and proposes a dynamic threshold setting method to reduce false negatives. GDN uses a graph attention network on a self-learning graph structure to extract spatial correlations between features, improving prediction and detection performance. Simultaneously, based on the prediction errors of each feature, the location of anomalies can be initially pinpointed, and then the root cause node can be found by tracing the graph topology. GTA employs a graph spatiotemporal attention mechanism to explicitly model the dynamic spatiotemporal dependencies in multivariate time series. By adaptively learning the correlation strength and temporal change patterns between nodes, it achieves accurate representation of the state of complex systems, thereby improving the sensitivity and stability of anomaly detection.

[0008] However, existing multivariate time series anomaly detection techniques have the following shortcomings in the context of thermal power plants: (1) It is difficult to fully capture the complex spatiotemporal dependencies between timestamps and features, which makes it more difficult to distinguish between abnormal and normal patterns, thus limiting their detection performance in multi-feature index detection tasks.

[0009] (2) Existing methods generally select micro-monitoring indicators as characteristic variables. However, in the operating scenario of thermal power plants, macro-state indicators have a significant impact on micro-indicators. For example, the daily coal consumption and operating rate determine the operating power of the power generation equipment, and the daily maintenance plan determines the start-up and shutdown status of the units at different times. Therefore, it is necessary to integrate macro-indicators into the anomaly detection process.

[0010] (3) For large power groups, there are usually multiple independently operating thermal power plants. In order to effectively solve the task of detecting abnormal operation status of each target power plant in the group, the existing methods are limited by weak generalization and need to be trained and deployed independently in each power plant environment, resulting in large hardware and software overhead and poor flexibility. Summary of the Invention

[0011] To overcome the shortcomings of existing time-series multivariate anomaly detection methods, such as difficulty in deeply exploring complex patterns in time trends and relationships between features, failure to consider macro-level data indicators, and poor generalization ability across different scenarios, this invention provides a method for detecting anomalies in thermal power plant operating conditions based on spatiotemporal relationship learning. This addresses the intelligent operation and maintenance needs of thermal power plants and improves power generation efficiency. Furthermore, this method can be applied simultaneously to multiple thermal power plants under a power generation group.

[0012] To solve the above-mentioned technical problems, the technical solution adopted by the present invention is as follows: In a first aspect, the present invention provides a method for detecting anomalies in the operating conditions of thermal power plants based on spatiotemporal relationship learning, the method comprising: Acquire macroscopic status data and microscopic monitoring data of thermal power plants; Macroscopic state data is input into a pre-trained state extraction model, and state weights are generated to represent the current operating state through embedding encoding and pattern mining. Microscopic monitoring data is input into an anomaly detection model based on spatiotemporal relationship learning, and an anomaly score reflecting the equipment's operating condition is output. Align the state weights and anomaly scores on the time axis, and calculate the fusion result based on the fusion strategy; The fusion result is compared with a preset threshold. If the fusion result is greater than the threshold, the current time is determined to be an abnormal operating condition; otherwise, it is normal.

[0013] Following the above technical solution, the macroscopic status data are plant-level or unit-level status indicators, including coal consumption, operating rate, load rate, on-grid electricity, and maintenance status; the microscopic monitoring data are the monitoring sensor data of each piece of equipment, including coal mills, steam turbines, generators, and water pumps.

[0014] Following the above technical solution, the steps of the state extraction model include: Preprocess macroscopic state data to generate macroscopic multivariate time series; Deep linear networks are used to embed and encode macroscopic multivariate time series data to obtain embedded and encoded representations. A large model based on the Transformer architecture is used to mine the operational mode features in the embedded encoding representation and generate state weights, as follows: For the output of the last layer of the large model The state weights are generated according to the following process. : ; In the formula, It is a non-linear activation function. and These are the learnable linear weight matrix and the bias vector, respectively.

[0015] Following the above technical solution, the steps of the anomaly detection model include: Microscopic monitoring data are preprocessed to generate microscopic multivariate time series, and a sliding time window is constructed as input; An anomaly detection model is constructed, including a prediction module, a reconstruction module, and an adversarial training mechanism; wherein: The prediction module employs convolutional networks, multi-head self-attention mechanisms, and graph attention mechanisms. It learns spatiotemporal relationships to uncover long-term and short-term time trends from a global perspective and to explore spatial relationships from both global and local perspectives. The reconstruction module uses an autoencoder to reconstruct the input time window and learn the latent representation of the data; The adversarial training mechanism uses a mini-maximum optimization strategy to perform adversarial training on the prediction module and the reconstruction module, so that the anomaly detection model can correctly distinguish between normal and abnormal samples. By combining prediction error, reconstruction error, and adversarial error, the anomaly score for the current time window is calculated.

[0016] Following the above technical solution, the prediction module includes a convolutional input layer, a temporal relationship learning layer, a spatial relationship learning layer, and a convolutional output layer; The convolutional input layer is used to perform convolution processing on the input time window data to extract long and short-term time patterns and obtain feature tensors. And serve as input for the temporal relation learning layer and the spatial relation learning layer; The temporal relationship learning layer employs a multi-head self-attention mechanism to mine long-term and short-term temporal trends from a global perspective. The output of the temporal relationship learning layer is obtained after processing by a feedforward network. ; The spatial relationship learning layer includes global learning units and local learning units. The global learning unit will... The transpose then employs a multi-head attention mechanism and a feedforward network to mine spatial relationships from a global perspective. Local learning units construct the topological structure between feature variables based on self-learning directed graphs, and use a graph attention network to aggregate information on this topological structure to learn the local spatial relationships between feature variables, ultimately obtaining the output of the spatial relationship learning layer. ; The convolutional output layer converts the output of the convolutional input layer into a single layer. Output of the time-related learning layer Output of the spatial relationship learning layer The data is fused together, and dimension matching is performed through a convolutional network to output a prediction result that matches the dimension of the prediction target.

[0017] Following the above technical solution, during adversarial training, pairs of positive and negative samples are input into the reconstruction module; the positive sample is the original time window, and the negative sample is the time window generated by the prediction module, which is composed of a part of the original time window and the prediction result of the prediction module. The adversarial training mechanism designs minimum and maximum losses for positive and negative samples respectively. The loss function is expressed as the weighted difference between the reconstruction loss of positive samples and the reconstruction loss of negative samples.

[0018] Following the above technical solution, the integration strategy includes: The state weights and anomaly scores are fused based on an exponential gating mechanism, and the fusion formula is as follows: ; In the formula, For the fusion result, For state weights, These are abnormal scores. This is the sensitivity coefficient.

[0019] Following the above technical solution, the method further includes: The generalization ability of the model in multiple thermal power plant operation scenarios is improved through model fine-tuning strategies, as follows: The anomaly detection model was pre-trained on raw thermal power plant data; When migrating to the target thermal power plant, freeze all parameters in the model except for the temporal relationship learning layer and the spatial relationship learning layer; The temporal and spatial relationship learning layers were fine-tuned using micro-monitoring data from the target thermal power plant. The finely tuned model was deployed at the target thermal power plant for detecting anomalies in the plant's operating conditions.

[0020] In a second aspect, the present invention provides a computer device / apparatus / system, including a memory, a processor, and a computer program stored in the memory, wherein the processor executes the computer program to implement the steps of the method described in the first aspect.

[0021] Thirdly, the present invention provides a computer-readable storage medium having a computer program / instructions stored thereon, which, when executed by a processor, implement the steps of the method described in the first aspect.

[0022] In summary, compared with the prior art, the above-described technical solutions conceived by this invention can achieve the following beneficial effects: (1) This invention proposes an anomaly detection model based on spatiotemporal relationship learning. The model takes the multivariate time series composed of monitoring data of various equipment in the power plant as input, and relies on the prediction module, reconstruction module and adversarial training mechanism to deeply explore the complex long-term and short-term trend changes in the time domain and the dependency relationship between the spatial domain features, so as to achieve accurate identification of equipment operating anomalies. (2) This invention breaks through the limitation of existing methods that only focus on equipment monitoring data, and uses pattern mining technology based on large models to extract special states from the macro data of power plants. These states are integrated with the anomaly scores output by the detection model in the form of weights, thereby avoiding misjudgment of operating conditions caused by changes in power plant operating conditions and improving the accuracy of anomaly detection; (3) This invention designs a model fine-tuning strategy by analyzing the inherent correlations between data from different thermal power plants. The aim is to extract universal correlations and patterns among the operating data of each power plant to improve the generalization performance of the designed anomaly detection model. Ultimately, the same model can be used to effectively realize the anomaly detection tasks of different power plants, significantly reducing resource and time costs and improving the overall detection efficiency. Attached Figure Description

[0023] Figure 1 This is an overall flowchart of a method for detecting abnormal operating conditions in thermal power plants based on spatiotemporal relationship learning, according to an embodiment of the present invention. Figure 2 This is a framework diagram of a state extraction model according to an embodiment of the present invention; Figure 3 This is a framework diagram of an anomaly detection model according to an embodiment of the present invention; Figure 4 This is a network framework diagram of a prediction module according to an embodiment of the present invention; Figure 5 This is a schematic diagram of a model fine-tuning strategy according to an embodiment of the present invention. Detailed Implementation

[0024] To make the objectives, technical solutions, and advantages of this invention clearer, the invention will be further described in detail below with reference to the accompanying drawings and embodiments. It should be understood that the specific embodiments described herein are merely illustrative and not intended to limit the invention. All other embodiments obtained by those skilled in the art based on the embodiments provided by this invention without inventive effort are within the scope of protection of this invention.

[0025] Obviously, the accompanying drawings described below are merely some examples or embodiments of the present invention. Those skilled in the art can apply the present invention to other similar scenarios based on these drawings without any inventive effort. Furthermore, it is understood that although the efforts made in this development process may be complex and lengthy, for those skilled in the art related to the content disclosed in this invention, modifications to design, manufacturing, or production based on the technical content disclosed in this invention are merely conventional technical means and should not be construed as insufficient disclosure of the present invention.

[0026] In this invention, the reference to "embodiment" means that a specific feature, structure, or characteristic described in connection with an embodiment may be included in at least one embodiment of the invention. The appearance of this phrase in various places in the specification does not necessarily refer to the same embodiment, nor is it a mutually exclusive, independent, or alternative embodiment. It will be explicitly and implicitly understood by those skilled in the art that the embodiments described in this invention may be combined with other embodiments without conflict.

[0027] Unless otherwise defined, the technical or scientific terms used in this invention shall have the ordinary meaning understood by one of ordinary skill in the art to which this invention pertains. The terms "a," "an," "an," "the," and similar words used in this invention do not indicate quantity limitation and may indicate singular or plural. The terms "comprising," "including," "having," and any variations thereof used in this invention are intended to cover non-exclusive inclusion; for example, a process, method, system, product, or device that includes a series of steps or modules (units) is not limited to the listed steps or units, but may also include steps or units not listed, or may include other steps or units inherent to these processes, methods, products, or devices. The terms "connected," "linked," "coupled," and similar words used in this invention are not limited to physical or mechanical connections, but may include electrical connections, whether direct or indirect. "A plurality" used in this invention refers to two or more. "And / or" describes the relationship between related objects, indicating that three relationships may exist; for example, "A and / or B" can represent: A alone, A and B simultaneously, and B alone. The character " / " generally indicates that the preceding and following objects have an "or" relationship. The terms "first," "second," and "third" used in this invention are merely to distinguish similar objects and do not represent a specific ordering of the objects.

[0028] This invention provides a method for detecting anomalies in the operating conditions of thermal power plants based on spatiotemporal relationship learning, specifically a method applicable to thermal power plants with different installed capacities. This method mainly includes the following four aspects: Content 1: Using macroscopic state data as input, design a state extraction model and generate state weights based on embedding encoding and pattern mining; Content 2: Using microscopic monitoring data as input, design an anomaly detection model based on spatiotemporal relationship learning; Content 3: Align the outputs of Content 1 and Content 2 on the timestamp, merge them, and obtain the exception tags; Content 4: A fine-tuning strategy is proposed to improve the generalization of the detection method in multiple thermal power plant operation scenarios.

[0029] Content 1 is achieved through the following steps: (1) Preprocess the macroscopic state data; (2) Embedding encoding is performed on the preprocessed data; (3) Use a large model based on the Transformer architecture to mine special patterns in the encoded representation and generate state weights.

[0030] Content 2 is achieved through the following steps: (1) Preprocess the micro-monitoring data to form a standard multivariate time series; (2) The prediction module is implemented by learning spatiotemporal relationships through convolutional networks, multi-head attention mechanisms and graph attention mechanisms; (3) Use an autoencoder to further extract special patterns to realize the reconstruction module; (4) Introduce an adversarial training mechanism to optimize the prediction module and the reconstruction module based on a minimization-maximization strategy; (5) Comprehensive prediction, reconstruction, and counter-error output of abnormal scores.

[0031] Content 3 is achieved through the following steps: (1) Align macro data and micro monitoring data on timestamps; (2) Combine the weights output by Content 1 with the outlier scores output by Content 2; (3) Compare the fusion result with the threshold to generate anomaly judgment labels.

[0032] Content 4 is achieved through the following steps: (1) Pre-train the anomaly detection model on the original power plant data; (2) Freeze specific model parameters according to the selected fine-tuning strategy; (3) Fine-tune the remaining model parameters based on the target power plant data; (4) Apply the trained model to the target power plant scenario.

[0033] Figure 1 This is an overall flowchart of an anomaly detection method for thermal power plant operating conditions based on spatiotemporal relationship learning, as proposed in this invention. For the 1) micro-monitoring data of various equipment and 2) macro-state data at the plant and unit levels collected during the operation of the thermal power plant, standardized multivariate time series are first generated through preprocessing methods such as missing value imputation and timestamp alignment. Then, the macro-data is input into a state extraction model, where state weights are generated through embedding encoding and pattern mining. For the micro-data, it is input into an anomaly detection model, where anomaly scores are obtained through multi-angle spatiotemporal relationship learning. Finally, the macro-data and micro-data are aligned in terms of timestamps, the state weights and anomaly scores are fused, and anomaly determination is achieved by comparing the obtained results with a predefined threshold.

[0034] Content 1: Using macroscopic state data as input, design a state extraction model and generate state weights based on embedding encoding and pattern mining.

[0035] For anomaly detection in thermal power plants, the targets are specific equipment such as turbines, pumps, and coal mills, and anomaly detection methods are mainly designed based on monitoring sensor data from these devices. However, macroscopic state data at the plant or unit level, such as grid-connected power generation, operating rate, load factor, and maintenance status, have a significant impact on monitoring data. For example, when the operating rate of a unit decreases, the operating power of the corresponding equipment also decreases. This leads to significant differences in the trends, magnitudes, and relationships between indicators of monitoring data collected under special operating conditions (shutdown, maintenance, standby, partial standby) compared to steady-state operation. Consequently, even if all equipment is operating under normal conditions, monitoring data under special operating conditions is easily misjudged as abnormal by various detection methods due to its differences from steady-state data. Therefore, to achieve more accurate anomaly detection adapted to different operating conditions, this solution designs a state extraction model to process the macroscopic state data of the power plant. By mining the implicit patterns, different operating states are distinguished, and finally, state weights are generated to provide a correction benchmark for subsequent anomaly detection of microscopic monitoring data. The model framework diagram for this content is attached. Figure 2 .

[0036] (1) Preprocess the macroscopic state data.

[0037] For status indicators collected from the centralized control platform, such as coal consumption, unit utilization rate, load rate, on-grid electricity, and maintenance status, missing values ​​are first filled using linear interpolation. Then, the collection frequency for different indicators is standardized (generally once per hour). For indicators with high collection frequency, downsampling is used for sparsity reduction; while for indicators with low collection frequency, resampling is used for filling. For numerical indicators, normalization is used to scale them to a uniform interval to facilitate subsequent model calculations; while for categorical indicators, one-hot encoding is used for discretization. After preprocessing, the macroscopic status data presents a standard multivariate time series. For sampling points Data , contains Individual indicator characteristics.

[0038] (2) Embedding encoding is performed on the preprocessed data.

[0039] With sampling points The processed state data As input to the state extraction model, a deep linear network is first used for preliminary processing. The purpose is to project various indicators into the same representation space, ultimately obtaining the embedded encoding representation of the state data in this space. This scheme uses a multilayer perceptron to implement the deep linear network. The calculation process is expressed as follows: ; in, As a non-linear activation function, the output is limited to the interval (-1, 1). and These are the learnable linear weight matrix and bias vector, respectively. The output of the last layer is used as the encoded representation. This information is then input into the subsequent network modules of the state extraction model.

[0040] (3) Use a large model based on the Transformer architecture to mine special patterns in the encoded representation and generate state weights.

[0041] The encoded representation generated in the previous step needs further analysis to uncover hidden patterns, thereby distinguishing and quantifying different operational states. This solution employs a large model based on a multi-layered Transformer structure to implement these operations, effectively utilizing the powerful deep relationship extraction capabilities of the large model. For the output of the last layer... The final state weights are generated according to the following process. : ; in, For another non-linear activation function, the output is limited to the interval [0, +∞]. and These are the learnable linear weight matrix and the bias vector, respectively, which transform the representation vector output by the large model into a weight scalar.

[0042] For the same power plant or unit, the state weights obtained under special operating conditions such as standby, shutdown, and maintenance should be less than those under steady-state operating conditions. The more the operating conditions deviate from steady state, the closer the state weights are to 0.

[0043] Content 2: Using microscopic monitoring data as input, design an anomaly detection model based on spatiotemporal relationship learning.

[0044] This paper proposes an anomaly detection model for monitoring data of micro-devices. The model mainly consists of three parts: a prediction module, a reconstruction module, and an adversarial mechanism. It ultimately outputs an anomaly score corresponding to a specific time window. The model framework diagram is attached. Figure 3 .

[0045] (1) Preprocess the micro-monitoring data to form a standard multivariate time series.

[0046] The preprocessing of micro-monitoring data is similar to that of macro-state data. For sensor indicators collected in real time from equipment such as coal mills, steam turbines, generators, and water pumps, missing values ​​are first filled in, and then the collection frequency of different indicators is standardized (generally once per minute). For indicators with high collection frequency, downsampling is used for sparsity reduction; while for indicators with low collection frequency, resampling is used for filling. After preprocessing, the micro-monitoring data presents a standard multivariate time series. For sampling points Data , contains Individual indicator characteristics. Furthermore, [the following will be discussed / discussed / informed]. The sampling point is divided into a sliding time window that corresponds one-to-one with the sampling point and used as the input to the anomaly detection model. The corresponding time window is , which includes Data from each sampling point.

[0047] (2) The prediction module is implemented by learning spatiotemporal relationships through convolutional networks, multi-head attention mechanism and graph attention mechanism.

[0048] To delve deeper into the complex temporal trends and spatial (inter-indicator feature) correlations in multivariate time series data, a multi-angle spatiotemporal attention learning network is designed as the prediction module in the anomaly detection model. Its framework is shown in the appendix. Figure 4 It consists of three components: a convolutional input-output layer, a temporal relation learning layer, and a spatial relation learning layer.

[0049] ① Convolutional Input / Output Layer: To extract long and short-term time patterns and reduce the impact of noise, this scheme first performs a moving average operation to expand the original data into multiple channels. Then, the input time series is windowed... Split into inclusion Historical data of each sampling point and contain Prediction target for each sampling point The system consists of two parts, using a two-dimensional convolutional network as the input and output layers to match the dimensions. The result after processing by the convolutional input layer is shown below. As input for subsequent time and space relationship learning layers.

[0050] ② Temporal Relationship Learning Layer: Convolutional input-output layers can only learn the correlation between adjacent time stamps, but cannot capture the relationship between arbitrary sequence positions within the entire window. Therefore, this scheme designs a temporal relationship learning layer. It adopts a multi-head self-attention mechanism, using timestamps as objects, to mine long-term and short-term time change trends from a global perspective. In this layer, we first... The feature representation space is enriched by splitting it into h attention heads. For each attention head... i The self-attention calculation process is represented as follows: ; in, , , All are learnable linear transformation matrices. Subsequently, the calculation results of each attention head are reassembled, and after further processing by a feedforward network, the output of the temporal relationship learning layer is obtained. .

[0051] ③ Spatial Relationship Learning Layer: In addition to temporal relationships, complex spatial dependencies also exist among the feature variables in the data. Therefore, this scheme designs a spatial relationship learning layer. It employs multi-head attention and graph attention mechanisms to mine spatial relationships from global and local perspectives, respectively.

[0052] The learning process from a global perspective is similar to that of the time-related learning layer. It's worth noting that the attention calculation here focuses on feature variables, therefore it's necessary to first... The transpose is then processed by a multi-head attention and feedforward network to obtain the spatial global learning output. .

[0053] Compared to global correlations, some feature variables exhibit stronger local dependencies, such as different attribute metrics of the same device within a server. Therefore, a local spatial attention unit is designed in this layer. This unit utilizes a self-learning directed graph to construct topological relationships between nodes, and then employs a Graph Attention Network (GAT) to learn local spatial relationships. The computation result of the GAT is used as the output of the spatial relationship learning layer. Specifically, for the feature tensor... For the feature vectors of all N indicators, embedding representations are further learned, and similarity calculations are performed on the obtained N embedding representations. Based on this, using the N indicators as graph nodes, the k nodes with the highest similarity to the embedding representation of each node are selected as neighbors, and the similarity value is used as the corresponding edge weight to construct a directed graph structure.

[0054] The outputs of the convolutional input layer and the temporal and spatial relationship learning layer are fused together and then processed by the convolutional output layer to obtain the output of the prediction module. The model loss function for the prediction module is expressed as: This refers to the difference between the predicted value and the actual target.

[0055] (3) Use an autoencoder to further extract special patterns and realize the reconstruction module.

[0056] To further learn the latent representation patterns of the data, this scheme designs an autoencoder (AE) to implement the temporal window reconstruction process. It consists of a pair of encoders and decoders. The encoder maps the original input to a high-dimensional representation space to further mine hidden data patterns and obtain latent representations. The decoder will then The code decodes back to the original space to obtain the reconstructed output. In this model, the encoder and decoder are symmetric networks, each consisting of three fully connected layers.

[0057] The loss function of the reconstruction module is expressed as: This refers to the difference between the reconstructed output and the original input.

[0058] (4) Introduce an adversarial training mechanism to optimize the prediction module and the reconstruction module based on the mini-maximum strategy.

[0059] The adversarial mechanism addresses the phenomenon that overfitting may occur during model training when optimizing the Adversarial Component (AE) by minimizing the reconstruction error. In other words, for any input, the trained reconstruction module tends to reconstruct it completely without loss. However, this leads to insignificant reconstruction errors for anomalous samples during the inference phase, making it difficult to correctly distinguish anomalies. This problem can be solved if the reconstruction module can determine whether the input sample is normal before reconstruction. Therefore, this scheme uses a minimization-maximization optimization strategy to adversarially train the prediction and reconstruction modules, combining their learning characteristics to overcome performance bottlenecks.

[0060] Specifically, in the adversarial mechanism, this scheme inputs pairs of positive and negative samples into the reconstruction module. The positive sample is the original time window. The negative samples are the windows generated by the prediction module. It consists of raw data and predicted output It is pieced together. Subsequently, the model optimization process based on the adversarial mechanism can be represented as designing minimum and maximum losses for positive and negative samples respectively, which can be combined as: ; In the formula, and For weights.

[0061] This process aims to train the model's ability to distinguish between normal and abnormal inputs. Since the optimization objective is no longer to minimize the reconstruction loss for any input, the phenomenon of the model failing to distinguish abnormalities due to overfitting of the reconstruction module is avoided. Furthermore, the adversarial optimization process further strengthens the spatiotemporal relationship learning ability of the prediction module.

[0062] (5) Comprehensive prediction, reconstruction, and counter-error output of abnormal scores.

[0063] During the training phase, the prediction, reconstruction, and adversarial components of the anomaly detection model are optimized sequentially according to their respective loss functions. In the inference phase, data from a specific time window is used... As the detection target, the prediction error is calculated based on the input and output of each part according to the following formula. Reconstruction error and combat error : ; ; .

[0064] Within each section, the error of each feature indicator is first aggregated independently across the sampling points within the window. The aggregation operation AGG can choose to take the maximum value or the average value. Then, by summing these three error components, the input time window is finally obtained. Corresponding abnormal scores .

[0065] Content 3: Align the outputs of the state extraction model and the anomaly detection model on the timestamp, and obtain the anomaly label after fusion.

[0066] (1) Align macro data and micro monitoring data on timestamps.

[0067] To combine the state weights output by the state extraction model with the anomaly scores output by the anomaly detection model, they first need to be aligned on the time axis. Since the acquisition frequency of macroscopic state data (typically 1 hour) is much lower than that of microscopic monitoring data (typically 1 minute), the same macroscopic state data time window is used... Its sampling period includes all microscopic monitoring data time windows Alignment. That is, all outlier scores within this period. They will share the same state weights. .

[0068] (2) Integrate state weights with anomaly scores.

[0069] In the process of fusing state weights and anomaly scores, to reduce the false alarm rate of anomalies under special operating conditions and the false negative rate of anomalies under steady state, this scheme designs a fusion method based on an exponential gating mechanism, expressed as: ; in, For the fusion result, For state weights, These are abnormal scores. This is the sensitivity coefficient. When the state weight approaches 0, it indicates that the power plant is in a special operating state such as shutdown, maintenance, or standby. In this case, the abnormal score of the equipment monitoring data is compressed to avoid false detections. When the state weight is large, it indicates that the power plant is in steady-state operation. In this case, the abnormal score is appropriately amplified to avoid missed detections.

[0070] (3) Compare the fusion result with the threshold to generate anomaly judgment labels.

[0071] For a specific sampling point Based on the fusion results obtained from the above process This serves as the final anomaly score. The threshold for determining anomalies can be obtained based on statistical patterns and prior knowledge. ,Will and If a comparison is made, Greater than Then determine the sampling point If an abnormal operating condition occurs, it is considered normal; otherwise, it is considered normal.

[0072] Since anomaly detection is a binary classification task, cross-entropy loss can be used to optimize the state extraction and anomaly detection models. Based on this, a certain number of anomaly-free samples can be collected, and the anomaly detection model can be optimized using methods such as minimizing prediction, reconstruction, and adversarial loss, thereby helping the model capture the pattern representation of normal operating conditions.

[0073] Content 4: A fine-tuning strategy is proposed to improve the generalization of the anomaly detection method in multiple thermal power plant operation scenarios.

[0074] In the task of detecting anomalies in the operating status of thermal power plants, the generalization ability of the adopted method is a very important performance indicator. Large power generation groups often have multiple thermal power plants. If a model is deployed and debugged separately for each power plant, it will consume a lot of hardware and software resources and time costs, which does not meet the flexibility requirements of the detection task. Therefore, under this condition, the generalization ability of the designed method should be improved as much as possible, so that the same model can be deployed simultaneously in all thermal power plant scenarios.

[0075] For the anomaly detection scheme designed in this invention, a crucial factor affecting generalization is the spatiotemporal relationship learning capability of the detection model. For different thermal power plants, variations in equipment configuration, operating strategies, and external environments lead to significant differences in the temporal trends of operational data and the correlations between indicators. Therefore, the temporal and spatial relationship learning layers of the anomaly detection model are the target parameters for improving generalization, while other components and parameters are considered secondary parameters that are not significantly related to generalization.

[0076] Based on the above settings, this section designs a pre-trained model fine-tuning strategy to improve the generalization ability of the anomaly detection method in this scheme in multiple thermal power plant operation scenarios. The process is shown in the appendix. Figure 5 This strategy first identifies a primary power plant as a benchmark, collects data from it, and pre-trains the model. Then, the pre-trained model is transferred to other power plant scenarios for fine-tuning. For a specific target power plant, during fine-tuning, the collected data is only used to update the parameters of the temporal and spatial relationship learning layers in the anomaly detection model, while the remaining parameters are frozen and not used in training. This strategy allows the same model to be applied to anomaly detection scenarios in multiple thermal power plants with minimal training overhead, while meeting the requirements for performance generalization.

[0077] Furthermore, the present invention also provides a computer device / apparatus / system, including a memory, a processor, and a computer program stored in the memory, wherein the processor executes the computer program to implement the steps of the above-described method.

[0078] The present invention also provides a computer-readable storage medium having a computer program / instructions stored thereon, which, when executed by a processor, implement the steps of the above-described method.

[0079] This invention is described with reference to flowchart illustrations and / or block diagrams of methods, systems, and computer program products according to embodiments of the invention. It should be understood that each block of the flowchart illustrations and / or block diagrams, as well as combinations of blocks in the flowchart illustrations and / or block diagrams, can be implemented by computer program instructions. These computer program instructions can be provided to a processor of a general-purpose computer, a special-purpose computer, an embedded processor, or other programmable data processing apparatus to produce a machine such that the instructions, which execute via the processor of the computer or other programmable data processing apparatus, create means for implementing the functions specified in one or more blocks of the flowchart illustrations and / or block diagrams.

[0080] These computer program instructions may also be stored in a computer-readable storage medium capable of directing a computer or other programmable data processing device to function in a particular manner, such that the instructions stored in the computer-readable storage medium produce an article of manufacture including instruction means that implement the functions specified in one or more flowcharts and / or one or more blocks in a block diagram.

[0081] These computer program instructions may also be loaded onto a computer or other programmable data processing apparatus to cause a series of operational steps to be performed on the computer or other programmable data processing apparatus to produce a computer-implemented process, thereby providing steps for implementing the functions specified in one or more flowcharts and / or one or more blocks in a block diagram.

[0082] It should be noted that, depending on the implementation needs, the various steps / components described in this invention can be broken down into more steps / components, or two or more steps / components or parts of the operation of steps / components can be combined into new steps / components to achieve the purpose of this invention.

[0083] Those skilled in the art will readily understand that the above are merely preferred embodiments of the present invention and are not intended to limit the present invention. Any modifications, equivalent substitutions, and improvements made within the spirit and principles of the present invention should be included within the scope of protection of the present invention.

Claims

1. A method for detecting anomalies in the operating conditions of thermal power plants based on spatiotemporal relationship learning, characterized in that, The method includes: Acquire macroscopic status data and microscopic monitoring data of thermal power plants; Macroscopic state data is input into a pre-trained state extraction model, and state weights are generated to represent the current running state through embedding encoding and pattern mining. Microscopic monitoring data is input into an anomaly detection model based on spatiotemporal relationship learning, and an anomaly score reflecting the equipment's operating condition is output. Align the state weights and anomaly scores on the time axis, and calculate the fusion result based on the fusion strategy; The fusion result is compared with a preset threshold. If the fusion result is greater than the threshold, the current time is determined to be an abnormal operating condition; otherwise, it is normal.

2. The method for detecting abnormal operating conditions in thermal power plants based on spatiotemporal relationship learning according to claim 1, characterized in that, Macro-level status data are plant-level or unit-level status indicators, including coal consumption, operating rate, load rate, on-grid electricity, and maintenance status; micro-level monitoring data are monitoring sensor data of each piece of equipment, including coal mills, steam turbines, generators, and water pumps.

3. The method for detecting abnormal operating conditions in thermal power plants based on spatiotemporal relationship learning according to claim 1, characterized in that, The steps of the state extraction model include: Preprocess macroscopic state data to generate macroscopic multivariate time series; Deep linear networks are used to embed and encode macroscopic multivariate time series data to obtain embedded and encoded representations. A large model based on the Transformer architecture is used to mine the operational mode features in the embedded encoding representation and generate state weights, as follows: For the output of the last layer of the large model The state weights are generated according to the following process. : ； In the formula, It is a non-linear activation function. and These are the learnable linear weight matrix and the bias vector, respectively.

4. The method for detecting abnormal operating conditions in thermal power plants based on spatiotemporal relationship learning according to claim 1, characterized in that, The steps of the anomaly detection model include: Microscopic monitoring data are preprocessed to generate microscopic multivariate time series, and a sliding time window is constructed as input; An anomaly detection model is constructed, including a prediction module, a reconstruction module, and an adversarial training mechanism; wherein: The prediction module employs convolutional networks, multi-head self-attention mechanisms, and graph attention mechanisms. It learns spatiotemporal relationships to uncover long-term and short-term time trends from a global perspective and to explore spatial relationships from both global and local perspectives. The reconstruction module uses an autoencoder to reconstruct the input time window and learn the latent representation of the data; The adversarial training mechanism uses a mini-maximum optimization strategy to perform adversarial training on the prediction module and the reconstruction module, so that the anomaly detection model can correctly distinguish between normal and abnormal samples. By combining prediction error, reconstruction error, and adversarial error, the anomaly score for the current time window is calculated.

5. The method for detecting abnormal operating conditions in thermal power plants based on spatiotemporal relationship learning according to claim 4, characterized in that, The prediction module includes a convolutional input layer, a temporal relation learning layer, a spatial relation learning layer, and a convolutional output layer; The convolutional input layer is used to perform convolution processing on the input time window data to extract long and short-term time patterns and obtain feature tensors. And serve as input for the temporal relation learning layer and the spatial relation learning layer; The temporal relationship learning layer employs a multi-head self-attention mechanism to mine long-term and short-term temporal trends from a global perspective. The output of the temporal relationship learning layer is obtained after processing by a feedforward network. ; The spatial relationship learning layer includes global learning units and local learning units. The global learning unit will... The transpose then employs a multi-head attention mechanism and a feedforward network to mine spatial relationships from a global perspective. Local learning units construct the topological structure between feature variables based on self-learning directed graphs, and use a graph attention network to aggregate information on this topological structure to learn the local spatial relationships between feature variables, ultimately obtaining the output of the spatial relationship learning layer. ; The convolutional output layer converts the output of the convolutional input layer into a single layer. Output of the time-related learning layer Output of the spatial relationship learning layer The data is fused together, and dimension matching is performed through a convolutional network to output a prediction result that matches the dimension of the prediction target.

6. The method for detecting abnormal operating conditions in thermal power plants based on spatiotemporal relationship learning according to claim 4, characterized in that, In adversarial training, pairs of positive and negative samples are input into the reconstruction module; the positive sample is the original time window, and the negative sample is the time window generated by the prediction module, which is composed of a part of the original time window and the prediction result of the prediction module. The adversarial training mechanism designs minimum and maximum losses for positive and negative samples respectively. The loss function is expressed as the weighted difference between the reconstruction loss of positive samples and the reconstruction loss of negative samples.

7. The method for detecting abnormal operating conditions in thermal power plants based on spatiotemporal relationship learning according to claim 1, characterized in that, Integration strategies include: The state weights and anomaly scores are fused based on an exponential gating mechanism, and the fusion formula is as follows: ； In the formula, For the fusion result, For state weights, These are abnormal scores. This is the sensitivity coefficient.

8. The method for detecting abnormal operating conditions in thermal power plants based on spatiotemporal relationship learning according to claim 5, characterized in that, The method also includes: The generalization ability of the model in multiple thermal power plant operation scenarios is improved through model fine-tuning strategies, as follows: The anomaly detection model was pre-trained on raw thermal power plant data; When migrating to the target thermal power plant, freeze all parameters in the model except for the temporal relationship learning layer and the spatial relationship learning layer; The temporal and spatial relationship learning layers were fine-tuned using micro-monitoring data from the target thermal power plant. The finely tuned model was deployed at the target thermal power plant for detecting anomalies in the plant's operating conditions.

9. A computer device / equipment / system, comprising a memory, a processor, and a computer program stored in the memory, characterized in that, The processor executes the computer program to implement the steps of the method according to any one of claims 1 to 8.

10. A computer-readable storage medium having a computer program / instructions stored thereon, characterized in that, When the computer program / instructions are executed by the processor, they implement the steps of the method according to any one of claims 1 to 8.

Images