Power dispatching monitoring data anomaly detection method based on artificial intelligence

Abstract

The invention discloses a power dispatching monitoring data anomaly detection method based on artificial intelligence, and relates to the field of data anomaly detection, and the method comprises the steps: firstly carrying out the multi-modal alignment and preprocessing of high-frequency time sequence monitoring data and sparse log text of power dispatching, and constructing the association mapping of heterogeneous data under a time window; a retrieval enhancement generation technology is utilized to deeply parse a log text to construct a global semantic context, and time series data is synchronously subjected to double-flow coding to extract inherent fluctuation features. Based on this, a cross-modal semantic gating feature modulation mechanism is established, and a dynamic reweighting sequential sequence is guided by using log semantics, so that when a semantic background of a specific scheduling operation is perceived, the attention on compliance data mutation is automatically reduced. And finally, performing signal reconstruction and adaptive anomaly judgment based on the feature sequence under semantic guidance. In this way, false alarms can be significantly reduced while the real fault sensing capability is guaranteed, and the intelligence and robustness of the power monitoring system are improved.

Description

Technical Field

[0001] This application relates to the field of data anomaly detection, and more specifically, to an artificial intelligence-based method for detecting anomalies in power dispatch and monitoring data. Background Technology

[0002] With the advancement of new power system construction and the high proportion of renewable energy integrated into the grid, the operating environment of power dispatching systems is becoming increasingly complex, exhibiting significant dynamism and uncertainty. As the "brain" of the power grid, the power dispatching control center aggregates massive amounts of monitoring data in real time, primarily including high-frequency sampled telemetry and teleindication time-series data, as well as dispatching logs recording system operating status and manual operations. This multi-source, heterogeneous data contains critical state information about the power grid's operation and is fundamental to ensuring the safe and stable operation of the system. Traditional anomaly detection methods often rely on fixed thresholds set by human experience or simple rule matching. Faced with the surge in data volume and drastic fluctuations in modern power grids, these methods are prone to false alarms or missed alarms, failing to meet the dual requirements of real-time performance and accuracy. Therefore, constructing an anomaly detection scheme for power dispatching monitoring data based on artificial intelligence, utilizing deep learning technology to automatically mine the spatiotemporal correlation features in the data, is of significant practical importance for improving the proactive perception of power grid faults and reducing the monitoring pressure on dispatchers.

[0003] In existing technologies, data-driven anomaly detection methods have made some progress in processing single-modal time-series data, widely employing models such as recurrent neural networks or Transformers for signal reconstruction and prediction. However, they still have significant limitations when facing the specific scenario of power dispatching. Most existing detection schemes independently process dense time-series monitoring data and sparse text log data, or simply concatenate results at the decision-making level, ignoring the deep semantic coupling between the two. Specifically, fluctuations in time-series data during power dispatching are often controlled by specific semantic contexts, such as "equipment maintenance" or "switching operations" recorded in logs. However, existing technologies lack an effective cross-modal interaction mechanism and cannot utilize sparse log semantic vectors to dynamically regulate the feature extraction process of dense time-series data. This lack of a semantic guidance mechanism means that when faced with data feature drift under known operational contexts, the model cannot automatically reduce its sensitivity to compliance fluctuations through semantic gating or weighting mechanisms. Consequently, false anomaly alarms are frequently triggered during normal dispatching operations, severely restricting the availability and reliability of intelligent monitoring systems in actual dispatching operations.

[0004] Therefore, there is an urgent need for an optimized method for detecting anomalies in power dispatch and monitoring data based on artificial intelligence. Summary of the Invention

[0005] This application is made in order to solve the above-mentioned technical problems.

[0006] According to one aspect of this application, an artificial intelligence-based method for detecting anomalies in power dispatching and monitoring data is provided, comprising: Acquire power dispatching time-series data and power dispatching log data; Multimodal data preprocessing and time alignment are performed on power dispatch time-series data and power dispatch log data to obtain windowed power dispatch time-series data and power dispatch log text; Power dispatch log text is subjected to RAG-based enhanced semantic vectorization to obtain power dispatch enhanced semantic vectors; Dual-stream encoding is performed on windowed power dispatch time-series data to obtain power dispatch time-series feature sequences; Based on the enhanced semantic vector of power dispatch, the power dispatch time-series feature sequence is subjected to semantic gating to obtain the modulated power dispatch time-series feature sequence. Based on the enhanced semantic vector of power dispatch, signal reconstruction and adaptive anomaly determination are performed on the modulated power dispatch time-series feature sequence to obtain anomaly judgment results.

[0007] Compared with existing technologies, this application provides an AI-based method for detecting anomalies in power dispatch monitoring data. First, it performs multimodal alignment and preprocessing on high-frequency time-series monitoring data and sparse log text for power dispatch, constructing a correlation mapping between heterogeneous data within the given time window. Then, it utilizes retrieval-enhanced generation techniques to deeply analyze the log text to construct a global semantic context, and simultaneously performs dual-stream encoding on the time-series data to extract inherent fluctuation features. Based on this, a cross-modal semantic gating feature modulation mechanism is established, using log semantic guidance to dynamically reweight the time-series sequence. This automatically reduces the focus on compliant data mutations when the semantic background of a specific dispatch operation is perceived. Finally, signal reconstruction and adaptive anomaly determination are performed based on the semantically guided feature sequence. This significantly reduces false alarms while ensuring the ability to detect real faults, improving the intelligence and robustness of the power monitoring system. Attached Figure Description

[0008] The above and other objects, features, and advantages of this application will become more apparent from the more detailed description of the embodiments of this application in conjunction with the accompanying drawings. The drawings are provided to further illustrate the embodiments of this application and form part of the specification. They are used together with the embodiments of this application to explain this application and do not constitute a limitation thereof. In the drawings, the same reference numerals generally represent the same components or steps.

[0009] Figure 1 This is a flowchart of an AI-based power dispatching and monitoring data anomaly detection method according to an embodiment of this application.

[0010] Figure 2 This is a data flow diagram of an AI-based power dispatching and monitoring data anomaly detection method according to an embodiment of this application.

[0011] Figure 3 This is a flowchart of sub-step S2 of the AI-based power dispatch monitoring data anomaly detection method according to an embodiment of this application.

[0012] Figure 4 This is a flowchart of sub-step S4 of the AI-based power dispatch monitoring data anomaly detection method according to an embodiment of this application.

[0013] Figure 5 This is a flowchart of sub-step S5 of the AI-based power dispatching and monitoring data anomaly detection method according to an embodiment of this application.

[0014] Figure 6 This is a flowchart of sub-step S6 of the AI-based power dispatch monitoring data anomaly detection method according to an embodiment of this application. Detailed Implementation

[0015] Embodiments of this disclosure will now be described in more detail with reference to the accompanying drawings. While some embodiments of this disclosure are shown in the drawings, it should be understood that this disclosure can be implemented in various forms and should not be construed as limited to the embodiments set forth herein. Rather, these embodiments are provided to provide a more thorough and complete understanding of this disclosure. It should be understood that the accompanying drawings and embodiments of this disclosure are for illustrative purposes only and are not intended to limit the scope of protection of this disclosure.

[0016] To address the problems mentioned above in the background technology, this application proposes an artificial intelligence-based method for detecting anomalies in power dispatching and monitoring data. Figure 1 This is a flowchart of an AI-based power dispatching and monitoring data anomaly detection method according to an embodiment of this application. Figure 2 This is a data flow diagram of an AI-based power dispatching and monitoring data anomaly detection method according to an embodiment of this application. Figure 1 and Figure 2As shown, the artificial intelligence-based power dispatch monitoring data anomaly detection method includes the following steps: S1, acquiring power dispatch time-series data and power dispatch log data; S2, performing multimodal data preprocessing and time alignment on the power dispatch time-series data and power dispatch log data to obtain windowed power dispatch time-series data and power dispatch log text; S3, performing RAG-based enhanced semantic vectorization on the power dispatch log text to obtain power dispatch enhanced semantic vectors; S4, performing dual-stream encoding on the windowed power dispatch time-series data to obtain power dispatch time-series feature sequences; S5, based on the power dispatch enhanced semantic vectors, performing semantic-gated feature modulation on the power dispatch time-series feature sequences to obtain modulated power dispatch time-series feature sequences; S6, based on the power dispatch enhanced semantic vectors, performing signal reconstruction and adaptive anomaly judgment on the modulated power dispatch time-series feature sequences to obtain anomaly judgment results.

[0017] In the aforementioned AI-based power dispatch monitoring data anomaly detection method, step S1 involves acquiring power dispatch time-series data and power dispatch log data. It should be understood that during the operation of the power dispatch system, the representation of the power grid state relies on the collaborative support of high-frequency measurement data and operational records. Anomaly detection requires analysis based on complete multi-source data; the absence of any type of data will lead to incomplete detection dimensions. Therefore, this application collects power dispatch time-series data and power dispatch log data during the power dispatch process to provide a basic data source for subsequent cross-modal fusion analysis. This ensures that the data relied upon for anomaly detection covers the dynamic physical state and static operational background of the power grid, laying a data foundation for accurate anomaly identification.

[0018] Specifically, in one possible embodiment, step S1 is implemented as follows: First, through the real-time data acquisition interfaces of the power monitoring and data acquisition system and the wide-area measurement system, multi-channel high-frequency measurement data such as voltage, current, and frequency are acquired, with the data sampling interval set to milliseconds according to scheduling requirements. Then, through the log database access interface of the scheduling system, text data such as operation ticket records, scheduling instruction logs, and equipment operating status records are extracted, with all log data accompanied by timestamps accurate to the second. Finally, the power scheduling time-series data and power scheduling log data are stored in the time-series database and text database respectively, while a data acquisition status monitoring mechanism is established to ensure the integrity and real-time performance of data transmission.

[0019] In the aforementioned AI-based power dispatch monitoring data anomaly detection method, step S2 involves multimodal data preprocessing and time alignment of power dispatch time-series data and power dispatch log data to obtain windowed power dispatch time-series data and power dispatch log text. It should be understood that power dispatch time-series data contains noise and missing values, and log data exhibits textual heterogeneity. Furthermore, the time dimensions of the two types of data are not correlated, leading to distorted feature extraction and failure of cross-modal fusion if used directly. Therefore, this application further purifies the power dispatch time-series data and power dispatch log data separately, and achieves precise alignment and windowing based on the time dimension to unify the data format and eliminate time deviations. This enables heterogeneous data to form structured time window data pairs, providing standardized input for subsequent semantically guided feature modulation and ensuring the collaborative analysis effect of cross-modal data.

[0020] In particular, in one specific embodiment, Figure 3 This is a flowchart of sub-step S2 of the artificial intelligence-based power dispatching and monitoring data anomaly detection method according to an embodiment of this application. Figure 3 As shown, step S2 includes: S21, performing time-series data cleaning on the power dispatch time-series data to obtain cleaned power dispatch time-series data; S22, performing text normalization and domain terminology normalization on the power dispatch log data to obtain standard power dispatch log data; S23, performing cross-modal data frame alignment based on sliding window on the cleaned power dispatch time-series data and standard power dispatch log data to obtain windowed power dispatch time-series data and power dispatch log text.

[0021] Specifically, step S21 involves cleaning the power dispatch time-series data to obtain cleaned power dispatch time-series data. It should be understood that during the acquisition and transmission of power dispatch time-series data, factors such as sensor malfunctions and transmission interference can lead to missing values ​​and abnormal impulse noise. These data defects severely interfere with the accuracy of subsequent feature extraction. Therefore, this application employs data repair and noise suppression techniques to process the time-series data, thereby eliminating invalid information and interference components. This results in continuous, smooth time-series data that conforms to the physical laws of power grid operation, providing a reliable data foundation for subsequent dual-stream coding to extract accurate fluctuation features.

[0022] Specifically, in one possible embodiment, step S21 is implemented as follows: First, to address the data missing problem, linear interpolation is used to fill in individual missing points; for continuous missing segments, cubic spline interpolation is used for completion. This algorithm uses effective data points before and after the missing segment as model points to construct a piecewise cubic polynomial function with continuous second derivatives, thereby accurately fitting the nonlinear trend of power data and avoiding trend distortion caused by simple linear connections. Second, to address impulse noise and high-frequency noise, abnormal noise points are first identified by threshold judgment, and then the Savitzky-Golay filter is applied to smooth the data. The core mechanism of this filter is to perform local polynomial least squares fitting on the data points within a sliding window, using points on the fitted curve to replace the original points. This allows it to effectively filter out high-frequency noise while preserving the signal's waveform width, height, and other local features (such as peak shape) to the greatest extent possible, avoiding signal flattening caused by ordinary moving averages. The filter window size is determined based on the data sampling frequency and the power grid fluctuation period. Finally, the processed data is checked for consistency, and extreme values ​​that still do not meet physical constraints are removed to ensure the rationality and continuity of the cleaned data.

[0023] Specifically, step S22 involves text normalization and domain terminology standardization of the power dispatch log data to obtain standard power dispatch log data. It should be understood that power dispatch log data is entered by different operators, resulting in inconsistent text formats, inconsistent domain terminology, and redundant characters, making it difficult for the model to accurately parse the log semantics. Therefore, this application further standardizes the log text format and unifies the terminology mapping to eliminate text heterogeneity and ambiguity. This transforms unstructured logs into standardized text data, ensuring that subsequent models can accurately capture the core semantics of the logs and provide precise textual semantic support for cross-modal semantic guidance.

[0024] Specifically, in one possible embodiment, step S22 is implemented as follows: First, special characters, redundant spaces, and meaningless symbols are removed from the log, full-width characters are converted to half-width characters, and the format for date and time representation is standardized. Then, a terminology dictionary for the power dispatching field is pre-built, containing standard and synonymous expressions for core terms such as equipment names, operation types, and fault types. The log text is traversed, and non-standard terms are replaced with standard terms from the dictionary. For example, "main transformer" and "transformer" are standardized to "power transformer," and "switching operation" and "switching operation" are standardized to "switching switching operation," ensuring consistency in terminology representation.

[0025] Specifically, step S23 involves aligning the cleaned power dispatch time-series data and standard power dispatch log data across modal data frames using a sliding window to obtain windowed power dispatch time-series data and power dispatch log text. It should be understood that the cleaned time-series data is a continuous high-frequency data stream, while the standard log data is a text record at discrete time points. The two types of data differ in temporal granularity and presentation format, making it impossible to directly establish a semantic and temporal feature association. Therefore, this application further employs a sliding window mechanism to perform time-dimensional frame division and association matching of the cleaned power dispatch time-series data and standard power dispatch log data, thereby constructing cross-modal data pairs within the time window. This ensures that each time-series data segment corresponds to unique semantic background information, providing an accurate temporal association basis for subsequent semantically gating modulation of time-series features and ensuring the synchronization of semantic guidance and time-series analysis.

[0026] Specifically, in one possible embodiment, step S23 is implemented as follows: First, the length and step size of the sliding window are set. The window length is determined based on the duration of typical power grid operations and the data sampling frequency, and the step size is set to 1 / 8 of the window length to ensure data continuity. Second, using the time axis of the time series data as a reference, the sliding window is slid according to the set parameters to extract time series data segments within each window, forming windowed time series data. Then, all standard log data within each window's time interval are retrieved and concatenated into a single log text in chronological order. Windows without log data are assigned a preset no-operation flag text. Finally, the time series data and log text of each window are associated and stored to form a one-to-one corresponding cross-modal data frame.

[0027] In the aforementioned AI-based power dispatch monitoring data anomaly detection method, step S3 involves performing RAG-based enhanced semantic vectorization on the power dispatch log text to obtain an enhanced semantic vector for power dispatch. It should be understood that since power dispatch log texts are mostly brief operation records containing only core operation information, lacking key context such as the normal data fluctuation range corresponding to the operation, historical similar operation cases, and equipment characteristic parameters, encoding solely through a language model would lead to a one-sided semantic representation, failing to provide accurate contextual semantic support for subsequent time-series feature modulation. Therefore, this application further combines Retrieval Enhanced Generation (RAG) technology to supplement related knowledge from the knowledge base before semantic encoding, thereby enriching the semantic dimension of the log text. This allows the final enhanced semantic vector for power dispatch to simultaneously contain core log operation information and domain knowledge context, accurately reflecting the semantic background of the current operating scenario. This provides a reliable semantic basis for subsequent semantic gating to dynamically adjust the sensitivity of time-series features, avoiding modulation deviations caused by insufficient semantic information.

[0028] Specifically, in one possible embodiment, step S3 is implemented as follows: First, a knowledge base for power dispatching is constructed. This knowledge base covers historical operation and maintenance cases (such as maintenance records of similar equipment in the past five years), equipment technical manuals (such as current fluctuation parameters during main transformer maintenance), and dispatching operation procedures (such as standard procedures and data requirements for switching operations). The BM25 algorithm is used to establish a text index, and the BGE model is used to vectorize the knowledge base data, ensuring efficient retrieval. Then, the corresponding log text is used as the query statement to retrieve the five knowledge fragments with the highest semantic relevance from the knowledge base, such as "normal current fluctuation range ±8% during #2 main transformer maintenance" and "no abnormal alarms were triggered during maintenance of the same model of main transformer in 2023." Finally, the original log text is concatenated with the five retrieved knowledge fragments to form an enhanced prompt text. During the concatenation, the original log text is placed first, followed by each knowledge fragment in sequence. Adjacent text fragments are distinguished by a preset text separator. The enhanced prompt text is then input into the encoder of the BERT model, which has been fine-tuned in the field of power dispatching, to extract the [CLS]token vector output by the model. This vector is the power dispatching enhanced semantic vector.

[0029] The core flaw of the aforementioned mechanism lies in the coarseness of its information fusion method, namely, the indiscriminate linear text splicing of the original log and the retrieved contextual information. This flattened processing approach places the original log, which is the core of the event, on an equal footing with multiple knowledge fragments serving as auxiliary explanations. This easily leads to key, urgent anomaly signals being diluted or submerged by a large amount of lengthy and inconsistent background information. When processing this long text, the language model's attention is easily distracted, making it unable to accurately focus on the most crucial event description. More importantly, this approach ignores the inherent hierarchical relationship and causal explanation between the original log and the retrieval context. The result (the event that occurs) is the retrieval context, while the retrieval context should be the cause or background used to explain or corroborate this result. Simple splicing flattens this structured relationship into a piece of unstructured text. Furthermore, the splicing process completely discards the quantitative indicators generated during the retrieval process—the relevance score—that measure the degree of relevance between knowledge and the query, assuming all retrieval fragments are equally important, which is clearly unreasonable. These defects collectively result in a semantic vector with high information entropy and low signal-to-noise ratio, whose quality and stability are insufficient to meet the requirements of high-precision scenarios such as power dispatching.

[0030] In particular, in another preferred embodiment, in order to solve the problems of semantic pollution, loss of hierarchical relationships and discarding of key information, and to ensure that each information unit maintains its pure semantic representation before fusion, the improved mechanism first independently encodes the original log and each retrieved knowledge fragment.

[0031] In other words, based on a knowledge base, context-independent encoding and relevance score extraction are performed on the power dispatch log text to obtain log vectors and context sets. Specifically, after obtaining Top-K knowledge fragments and their relevance scores through hybrid retrieval, a shared pre-trained language model encoder is used to process the original logs. and each piece of knowledge Perform independent deep semantic encoding.

[0032] in, This is a log vector, which is the original log. The dense vector representation obtained after encoder processing. Let be the knowledge fragment vector, which is the i-th retrieved knowledge fragment. The dense vector representation obtained after encoder processing Let C be the total number of knowledge fragments retrieved, and let C be the context set containing... A set of knowledge fragment vectors and their corresponding relevance scores.

[0033] It is understandable that in scenarios such as power dispatching where precise distinction between events and context is required, eliminating early mixing prevents the core event description from being interfered with by background knowledge, ultimately producing a clean log vector representing the original log. And a context set C containing independent vectors of each knowledge fragment and their relevance scores, lays a solid foundation for achieving hierarchical and selective fusion.

[0034] Then, after obtaining independent context vectors, in order to address the defect of discarding relevance scores in the above mechanism, the mechanism will use these scores as attention weights to perform weighted aggregation on all knowledge fragment vectors.

[0035] In other words, the context set is aggregated using relevance attention-based context vector aggregation to obtain an aggregated context vector. Specifically, a softmax function with a temperature coefficient is used to normalize the relevance scores, and all knowledge fragment vectors are weighted and summed using these weights to aggregate a single, comprehensive context vector. .

[0036] in, The aggregated context vector is a comprehensive vector formed by weighting and aggregating all knowledge fragment vectors according to their relevance scores. The attention weight for the i-th knowledge segment is calculated using the Softmax function based on its relevance score. Let be the vector of the i-th knowledge fragment. denoted as the relevance score given by the retrieval system for the i-th knowledge fragment. Temperature coefficient: A hyperparameter in the Softmax function used to control the smoothness of the weight distribution. It is an exponential function. To retrieve the total number of knowledge fragments.

[0037] This step transforms intangible relevance into tangible contribution, allowing historical cases or procedural clauses most relevant to the current scheduling log to play a dominant role at the semantic level. It uses a temperature coefficient. To dynamically adjust the model's focus: when dealing with emergency failure scenarios that require a high degree of trust in the most relevant evidence, a smaller focus is needed. A higher value can make the weight distribution more acute, focusing on the most critical knowledge; however, in handling routine operational scenarios that require integrating information from multiple sources, a larger value... This value can smooth out the weight distribution. The aim is to aggregate multiple discrete knowledge vectors of varying importance into a single aggregated context vector that accurately reflects the comprehensive background knowledge. The final aggregated vector not only contains all relevant knowledge, but its internal semantic structure has also been intelligently reconstructed based on its relevance to the original event, resulting in a significantly improved signal-to-noise ratio.

[0038] Finally, hierarchical semantic injection is performed on the log vector and the aggregated context vector to obtain the power dispatch enhanced semantic vector. That is, in order to address the core problem that simple concatenation cannot reflect the master-slave hierarchical relationship, the mechanism designs a hierarchical semantic injection operation to achieve fine correction of the original log semantics by context knowledge in a modulated rather than parallel manner.

[0039] Specifically, the aggregation context vector obtained in the previous step A dynamic injection gate is generated through two parallel fully connected layers. and a nonlinear injection vector Finally, through a gating mechanism, the injection vector is selectively and element-wise injected into the original log vector. In the process, the final enhanced semantic vector is generated. Specifically, hierarchical semantic injection is performed on the log vector and the aggregation context vector using the following formula: in, For dynamic injection gate: a gate vector generated from the aggregated context vector, with element values ​​between (0,1), used to control the degree of injection. Injection vector: A vector generated by nonlinear transformation of the aggregated context vector, used to correct the original log vector. Enhanced semantic vectors for power dispatch: The final generated enhanced semantic vectors that incorporate contextual knowledge. and This is the weight matrix of the fully connected layer. and This is the bias vector for the fully connected layer. The Sigmoid activation function compresses the output to the range (0,1). It is the hyperbolic tangent function. The Hadamard product represents element-wise multiplication of vectors.

[0040] This step collectively constructs a semantic injection module that is dynamically adjusted based on contextual content. Like a semantic scalpel driven by background knowledge, it can precisely adjust semantic components in the original log vector that might cause misjudgments based on the retrieved context (e.g., the current operation is a scheduled maintenance task), in order to produce a high-quality enhanced semantic vector. This vector retains the core information of the original log (through...). (Controlled loyalty), and further enhanced or corrected intelligently based on background knowledge (through...) (Control acceptance), thereby providing downstream tasks with stable, reliable semantic input that is deeply integrated with domain knowledge.

[0041] Through a series of interconnected technical measures, this improved mechanism successfully addresses the technical shortcomings caused by indiscriminate text concatenation in the previous mechanisms. Ultimately, this mechanism aims to generate a high-quality enhanced semantic vector. This vector is no longer a simple accumulation of original information and background knowledge; it not only contains the core information of the original log text but also creatively and selectively injects background knowledge processed with relevance weighting and hierarchical structure, possessing characteristics of high signal-to-noise ratio, high stability, and deep understanding of domain knowledge. Its fundamental purpose is to provide a lower information entropy and more semantically specific input for subsequent anomaly detection models, thereby significantly improving the accuracy and reliability of anomaly detection in complex power dispatching scenarios.

[0042] In the aforementioned AI-based power dispatch monitoring data anomaly detection method, step S4 involves dual-stream encoding of the windowed power dispatch time-series data to obtain a power dispatch time-series feature sequence. It should be understood that power dispatch time-series data has strong temporal correlation; single-direction encoding (e.g., only from past to present) cannot capture the potential correlation between the current moment and future moments. For example, current fluctuations before a switching operation need to be combined with subsequent operation information to be judged as normal; forward encoding alone will miss such reverse dependencies. Therefore, this application further adopts a bidirectional encoding architecture, capturing temporal dependencies from both the forward and reverse directions, and then fusing them to form complete features, thereby covering the full contextual information of the time-series data. This allows the features of each time step to simultaneously include historical influences and future correlations, avoiding feature bias caused by the lack of unidirectional information, and providing a more accurate temporal feature foundation for subsequent semantic gating modulation.

[0043] In particular, in one specific embodiment, Figure 4 This is a flowchart of sub-step S4 of the artificial intelligence-based power dispatching and monitoring data anomaly detection method according to an embodiment of this application. Figure 4 As shown, step S4 includes: S41, performing forward temporal dependency encoding on the windowed power dispatching time-series data to obtain a forward hidden state sequence; S42, performing backward temporal dependency encoding on the windowed power dispatching time-series data to obtain a backward hidden state sequence; S43, performing bidirectional contextual feature fusion on the forward hidden state sequence and the backward hidden state sequence to obtain a power dispatching time-series feature sequence.

[0044] Specifically, step S41 involves forward temporal dependency encoding of the windowed power dispatch time-series data to obtain a forward hidden state sequence. It should be understood that the core change patterns of power dispatch time-series data propagate forward over time; for example, a slow decrease in voltage due to load growth, or a gradual increase in current caused by equipment heating. Such trends require the cumulative impact of historical data on the current moment to be accurately captured. Without forward dependency encoding, the forward evolution characteristics of the time series will be lost. Therefore, this application further employs a unidirectional GRU to process the data forward over time to extract the temporal dependencies from the beginning of the sequence to the current moment. This yields a forward hidden state sequence that reflects the impact of historical data on the current state, clearly characterizing the forward evolution patterns of the time-series data and providing fundamental historical context information for subsequent bidirectional fusion.

[0045] Specifically, in one possible embodiment, step S41 is implemented as follows: First, the initial hidden state of the forward GRU is initialized as a zero vector. Then, each feature vector of the windowed time series data is input sequentially from time step 1 to L. In this embodiment, the hidden layer dimension of the GRU unit is preset to 128 dimensions, which ensures feature capacity while avoiding overfitting. At each step, the current input and the forward hidden state of the previous time step are passed to the GRU unit, and the 128-dimensional forward hidden state of the current time step is generated through internal reset and update gates. Correspondingly, the bidirectional context feature fusion in the subsequent steps will generate a power dispatching time series feature sequence with a dimension of 256 (128×2).

[0046] Specifically, step S42 involves performing backward temporal dependency encoding on the windowed power dispatch time-series data to obtain a backward hidden state sequence. It should be understood that reverse correlation patterns exist in the power dispatch time-series data. For example, a voltage drop after equipment failure may be triggered in advance by subsequent protection actions, and current fluctuations before a switching operation need to be considered in conjunction with the stable state after the operation to be determined as normal. Forward encoding alone cannot capture such dependencies that influence the present from the future. Therefore, this application further employs a unidirectional GRU to process the data in reverse time to extract the reverse temporal dependencies from the end of the sequence to the current moment. This allows for the acquisition of a backward hidden state sequence that reflects the correlation between future data and the current state, supplementing the reverse context missed by forward encoding and avoiding the loss of dependencies due to a unidirectional perspective.

[0047] Specifically, in one possible embodiment, step S42 is implemented as follows: First, the initial hidden state of the backward GRU is initialized as a zero vector to ensure that there is no interference from future information at the beginning. Then, the windowed time series data is arranged in reverse order from L to 1 according to time steps and input into the GRU unit in sequence. At each step, the current reverse input and the backward hidden state of the next step are passed in. Then, the key future information is dynamically filtered through the gating mechanism of the GRU to generate the backward hidden state at the current time. The dimension of this state is consistent with that of the forward hidden state, both being a preset single hidden layer dimension. For example, in the reverse processing, the state at time t will incorporate the information at time t+1. After traversing all reverse time steps, the backward hidden state sequence consistent with the original time order is output. Each state contains reverse time series information from the corresponding time to the end of the sequence.

[0048] Specifically, step S43 involves performing bidirectional contextual feature fusion on the forward and backward hidden state sequences to obtain a power dispatching time-series feature sequence. It should be understood that the forward and backward hidden state sequences respectively carry forward and reverse time-series information. Using them alone or simply concatenating them would result in a fragmentation of the two types of information, failing to form a complete context for each time step. For example, current fluctuations at a certain moment require consideration of both historical load trends and future operational plans to determine if they are normal; fragmented information leads to biased feature judgments. Therefore, this application further fuses the bidirectional hidden states dimensionally by time step to integrate the historical and future contexts of each time step. This allows the features of each time step to simultaneously contain both positive dependencies and negative correlations, forming a more comprehensive and accurate time-series feature sequence, providing complete time-series information support for subsequent semantic gating modulation.

[0049] Specifically, in one possible embodiment, step S43 is implemented as follows: First, it is confirmed that the length of both the forward and backward hidden state sequences is L, and the dimension of each hidden state is a preset single hidden layer dimension, ensuring dimension matching. Then, for each time step t (from 1 to L), the forward and backward hidden state vectors at that time are concatenated along the feature dimension. For example, the forward vector of the single hidden layer dimension is concatenated with the backward vector of the same dimension to form a fusion vector with twice the preset single hidden layer dimension. Then, all fusion vectors are integrated in time step order to form a power dispatching time-series feature sequence of length L and dimension twice the preset single hidden layer dimension, ensuring that the features at each time step contain complete bidirectional contextual information.

[0050] In the aforementioned AI-based power dispatch monitoring data anomaly detection method, step S5 involves semantically gating the power dispatch time-series feature sequence based on the enhanced semantic vector to obtain the modulated power dispatch time-series feature sequence. It should be understood that since some significant fluctuations in the power dispatch time-series feature sequence originate from compliant operations, such as equipment maintenance and switching, analyzing only the time-series features would misjudge such compliant fluctuations as anomalies. The enhanced semantic vector already contains operational background information and needs to be used to guide the screening of time-series features. Therefore, this application further uses a semantic gating mechanism to adjust the weights of the time-series feature sequence using the enhanced semantic vector, thereby distinguishing between compliant fluctuations and features corresponding to real anomalies. This effectively suppresses time-series fluctuation features related to compliant operations while retaining the anomaly features corresponding to real faults, providing an accurate feature basis for subsequent anomaly judgment, significantly reducing false alarms, and improving the robustness of the detection system.

[0051] In particular, in one specific embodiment, Figure 5 This is a flowchart of sub-step S5 of the artificial intelligence-based power dispatching and monitoring data anomaly detection method according to an embodiment of this application. Figure 5 As shown, step S5 includes: S51, performing linear transformation and nonlinear activation on the power dispatch enhanced semantic vector to obtain a semantic gating vector; S52, based on the semantic gating vector, performing semantic guided modulation on the power dispatch time-series feature sequence to obtain a modulated power dispatch time-series feature sequence.

[0052] Specifically, step S51 involves performing a linear transformation and nonlinear activation on the power dispatch enhanced semantic vector to obtain a semantic gating vector. It should be understood that because the dimension of the power dispatch enhanced semantic vector does not match the feature dimension of a single time step in the power dispatch time-series feature sequence, and the numerical range of the semantic vector is not explicitly constrained, it cannot be directly used as a gating signal to regulate the time-series features. Therefore, this application further performs a linear transformation and nonlinear activation sequentially on the enhanced semantic vector to adapt it to the time-series feature dimension and assign it gating properties. This results in a semantic gating vector whose dimension perfectly matches a single time step of the time-series feature and whose value is in the 0-1 range. This ensures that the vector can accurately control the passage or suppression of each time-series feature dimension, providing an adapted and effective control signal for subsequent time-step feature modulation.

[0053] Specifically, in one possible embodiment, step S51 is implemented as follows: First, the power dispatching enhanced semantic vector is input into a fully connected layer. The parameters of this fully connected layer are optimized through model training, and its output dimension is preset to be consistent with the feature dimension of a single time step of the power dispatching time-series feature sequence, thus achieving dimensional adaptation of the semantic vector. Next, the intermediate vector output by the fully connected layer is input into a Sigmoid activation function. This function compresses the value of each element in the vector to the 0-1 range. Values ​​close to 0 correspond to suppressed features, and values ​​close to 1 correspond to preserved features. Finally, a semantically gated vector with gating function is generated, ensuring that it can be directly used for subsequent feature regulation.

[0054] Specifically, in step S52, based on the semantic gating vector, the power dispatch time-series feature sequence is semantically guided to obtain the modulated power dispatch time-series feature sequence. It should be understood that since the semantic gating vector is a single-dimensional vector, while the power dispatch time-series feature sequence contains features from multiple time steps, simply superimposing them cannot achieve precise feature control at each time step, resulting in some time-step compliance fluctuations not being effectively suppressed. Therefore, this application further applies the semantic gating vector to each time step of the time-series feature sequence to adjust the intensity of the time-series features point-by-point. This allows the time-series features at each time step to be filtered under semantic guidance, specifically suppressing fluctuations corresponding to compliant operations while fully preserving features corresponding to genuine anomalies. This ensures that the modulated feature sequence can be directly used for subsequent signal reconstruction and anomaly detection, providing high-quality feature input for accurate detection.

[0055] Specifically, in one possible embodiment, step S52 is implemented as follows: First, the semantic gating vector is broadcast in the time dimension. Based on the number of time steps in the power dispatch time-series feature sequence, the single semantic gating vector is copied and expanded to form a gating matrix with the same dimension as the time-series feature sequence. The gating vector corresponding to each time step is the same as the original semantic gating vector. Next, the broadcast gating matrix is ​​multiplied element-wise with the power dispatch time-series feature sequence. The feature vector of each time step is scaled according to the element value of the corresponding gating vector. The feature dimension corresponding to compliant fluctuations is suppressed because the gating value is close to 0, while the abnormal feature dimension is retained because the gating value is close to 1. Finally, the modulated power dispatch time-series feature sequence is obtained.

[0056] In the aforementioned AI-based power dispatch monitoring data anomaly detection method, step S6 involves reconstructing the modulated power dispatch time-series feature sequence based on enhanced semantic vectors for power dispatch, and then adaptively determining anomalies to obtain an anomaly judgment result. It should be understood that because the modulated power dispatch time-series feature sequence resides in a high-dimensional latent space, it cannot be directly compared with the original time-series data to determine anomalies. Furthermore, the tolerance for anomalies varies under different semantic backgrounds (such as maintenance or normal operation), and a fixed threshold judgment can lead to false alarms or missed alarms. Therefore, this application further reconstructs the latent features into signals of the original data dimension, and then dynamically adjusts the judgment threshold in conjunction with enhanced semantic vectors to achieve semantic-adaptive anomaly identification. This allows the anomaly judgment standard to adapt to the current operating scenario while restoring the signal morphology, accurately capturing real faults while avoiding misjudging compliant operational fluctuations as anomalies, thus improving the reliability and scenario adaptability of the judgment result.

[0057] In particular, in one specific embodiment, Figure 6 This is a flowchart of sub-step S6 of the artificial intelligence-based power dispatching and monitoring data anomaly detection method according to an embodiment of this application. Figure 6 As shown, step S6 includes: S61, reconstructing the modulated power dispatch time-series feature sequence to obtain reconstructed power dispatch time-series data; S62, calculating the reconstruction error between the windowed power dispatch time-series data and the reconstructed power dispatch time-series data to obtain a reconstruction score; S63, inputting the power dispatch enhanced semantic vector into a small feedforward neural network to obtain an adaptive threshold coefficient; S64, generating the anomaly judgment result based on the comparison between the adaptive threshold coefficient and the reconstruction score.

[0058] Specifically, step S61 involves reconstructing the modulated power dispatch time-series feature sequence to obtain reconstructed power dispatch time-series data. It should be understood that since the modulated power dispatch time-series feature sequence is a feature representation of a high-dimensional latent space, its dimension does not match the original power dispatch time-series data (such as physical quantity channels like voltage and current). Therefore, anomalies cannot be directly judged by data differences and need to be mapped back to the original data space. Thus, this application further employs a decoder network to perform inverse mapping processing on the latent features, thereby restoring a signal form consistent with the dimension of the original time-series data. This allows for the acquisition of reconstructed data that can be directly compared with the windowed original time-series data, providing a basis for subsequent error quantification of the anomaly degree and ensuring that anomaly judgment has clear physical data support.

[0059] Specifically, in one possible embodiment, step S61 is implemented as follows: First, the initial hidden state of the unidirectional GRU decoder is initialized to a zero vector to ensure that the initial reconstruction is bias-free. Then, the modulated time-series feature sequence is input into the decoder at time steps 1 to L. The decoder learns the mapping relationship between latent features and the original signal through its internal gate control mechanism and outputs a hidden state sequence. Next, the hidden state sequence output by the decoder is input into a fully connected layer. The output dimension of this fully connected layer is preset to be consistent with the number of channels in the original windowed power dispatching time-series data, such as voltage and current channels, and the dimension mapping is completed through linear transformation. Finally, reconstructed power dispatching time-series data that perfectly matches the length and number of channels of the original windowed time-series data is output, ensuring that the two can be directly compared at the element-level.

[0060] Specifically, in step S62, the reconstruction error between the windowed power dispatch time-series data and the reconstructed power dispatch time-series data is calculated to obtain a reconstruction score. It should be understood that, since the reconstructed power dispatch time-series data is theoretically an ideal signal without anomalies and conforming to the semantic context, the degree of difference between it and the original windowed time-series data can directly reflect whether there are anomalies in the original data. The smaller the difference, the closer the original data is to a normal state; the larger the difference, the higher the probability of anomaly. Therefore, this application further quantifies the difference between the two to obtain a reconstruction score, thereby objectively characterizing the degree of anomaly in the original data. In this way, the abstract concept of "whether it is abnormal or not" can be transformed into a specific numerical indicator, avoiding subjective judgment bias and providing a precise and quantifiable basis for subsequent anomaly determination.

[0061] Specifically, in one possible embodiment, step S62 is implemented as follows: First, it is confirmed that the length and number of channels of the windowed power dispatching time-series data and the reconstructed data are completely consistent to ensure the validity of the error calculation. Next, for each time step, the difference between the original data vector and the reconstructed data vector across all channels is calculated. The square of the Euclidean distance is used to measure the comprehensive error of a single time step, covering deviations in all physical quantities such as voltage and current. Then, the single-step errors of all time steps are summed and divided by the total number of time steps to obtain the average error of the entire window. This average error is the reconstruction score; the larger the score, the more significant the deviation between the original data and the ideal normal data.

[0062] Specifically, in step S63, the enhanced semantic vector of power dispatch is input into a small feedforward neural network to obtain an adaptive threshold coefficient. It should be understood that the tolerance of power dispatch systems for anomalies varies under different semantic backgrounds. For example, in equipment maintenance scenarios, data fluctuations are large, requiring a higher tolerance threshold for normal fluctuations; in daily operation scenarios, fluctuations are small, requiring a lower threshold. A fixed threshold coefficient cannot adapt to all scenarios. Therefore, this application further utilizes a small feedforward neural network to transform the enhanced semantic vector into scenario-adaptive threshold coefficients, thereby dynamically adjusting the anomaly judgment criteria. This allows the threshold coefficient to change in real time with the semantic background, ensuring that a reasonable judgment threshold can be set under different operating scenarios, effectively reducing false alarms or missed alarms caused by fixed thresholds.

[0063] Specifically, in one possible embodiment, step S63 is implemented as follows: First, a small feedforward neural network with two hidden layers is constructed. The input layer dimension is consistent with the dimension of the power dispatch enhancement semantic vector, and the output layer dimension is 1 (a single threshold coefficient). Next, the enhancement semantic vector is input into the feedforward neural network. The first hidden layer uses the ReLU activation function to extract semantic features, the second hidden layer continues to optimize the features through linear transformation, and the output layer uses the Softplus activation function to ensure that the output threshold coefficient is non-negative, conforming to the physical meaning that a larger coefficient indicates higher tolerance. Then, the weights and biases of the feedforward neural network are optimized through model training, enabling the network to accurately output appropriate threshold coefficients according to the semantics. For example, when the input is a maintenance-related semantic vector, a larger coefficient is output; when the input is a normal operation semantic vector, a smaller coefficient is output.

[0064] Specifically, step S64 generates the anomaly judgment result based on the comparison between the adaptive threshold coefficient and the reconstruction score. It should be understood that since anomaly judgment cannot be completed solely based on the reconstruction score, and the same score has different meanings in different scenarios (e.g., a high score may be normal in a maintenance scenario but abnormal in a routine scenario), a dynamic judgment standard needs to be determined by combining the adaptive threshold coefficient. Therefore, this application further generates a final judgment by comparing the coefficient and the score, thereby achieving scenario-adaptive anomaly identification. This ensures that the judgment result fully reflects the current operating context, neither omitting real faults (judging anomaly when the score is significantly higher than the dynamic threshold) nor misjudging compliance fluctuations as anomalies (judging normality when the score is lower than the dynamic threshold), thus improving the engineering practicality of the detection system.

[0065] Specifically, in one possible embodiment, step S64 is implemented as follows: First, a global baseline threshold is obtained. This threshold is not manually specified but determined by performing a grid search strategy on the validation set. Specifically, candidate thresholds are traversed within a preset numerical range, and the F1 score (harmonic mean of precision and recall) corresponding to each candidate threshold is calculated. The value corresponding to maximizing the F1 score is confirmed as the global baseline threshold, thereby completing threshold optimization based on automatically balancing the false negative rate and false positive rate. Next, a dynamic threshold is calculated by multiplying the adaptive threshold coefficient by the global baseline threshold to obtain a judgment threshold adapted to the current semantic scenario. For example, in a maintenance scenario, the coefficient is large, and the dynamic threshold is higher than the baseline value; in a daily scenario, the coefficient is small, and the dynamic threshold is close to the baseline value. Then, the reconstruction score is compared with the dynamic threshold. If the reconstruction score is greater than the dynamic threshold, the current windowed data is determined to be abnormal, and an abnormal judgment result is output; if the reconstruction score is less than or equal to the dynamic threshold, the data is determined to be normal, and a normal judgment result is output, along with interpretable information such as the reconstruction score and dynamic threshold, for maintenance personnel to review.

[0066] In summary, the AI-based power dispatch monitoring data anomaly detection method based on embodiments of this application is explained. First, it performs multimodal alignment and preprocessing on high-frequency time-series monitoring data and sparse log text for power dispatch, constructing a correlation mapping between heterogeneous data within the given time window. Then, it utilizes retrieval-enhanced generation technology to deeply analyze the log text to construct a global semantic context, and simultaneously performs dual-stream encoding on the time-series data to extract inherent fluctuation features. Based on this, a cross-modal semantic gating feature modulation mechanism is established, using log semantic guidance to dynamically reweight the time-series sequence, thereby automatically reducing attention to compliant data mutations when the semantic background of a specific dispatch operation is perceived. Finally, signal reconstruction and adaptive anomaly determination are performed based on the semantically guided feature sequence. This significantly reduces false alarms while ensuring the ability to perceive real faults, improving the intelligence and robustness of the power monitoring system.

[0067] As described above, the AI-based power dispatch monitoring data anomaly detection method according to the embodiments of this application can be implemented in various wireless terminals, such as servers with AI-based power dispatch monitoring data anomaly detection algorithms. In one possible implementation, the AI-based power dispatch monitoring data anomaly detection method according to the embodiments of this application can be integrated into the wireless terminal as a software module and / or hardware module. For example, the AI-based power dispatch monitoring data anomaly detection method can be a software module in the operating system of the wireless terminal, or it can be an application developed for the wireless terminal; of course, the AI-based power dispatch monitoring data anomaly detection method can also be one of many hardware modules of the wireless terminal.

[0068] Alternatively, in another example, the AI-based power dispatch monitoring data anomaly detection method and the wireless terminal can also be separate devices, and the AI-based power dispatch monitoring data anomaly detection method can be connected to the wireless terminal via wired and / or wireless networks, and transmit interactive information in accordance with an agreed data format.

Claims

1. An artificial intelligence-based power dispatch monitoring data anomaly detection method, characterized in that, The method comprises the following steps: obtaining power dispatch timing data and power dispatch log data; performing multi-modal data preprocessing and time alignment on the power dispatch timing data and the power dispatch log data to obtain windowed power dispatch timing data and power dispatch log text; performing RAG-based enhanced semantic vectorization on the power dispatch log text to obtain power dispatch enhanced semantic vectors; performing double-flow coding on the windowed power dispatch timing data to obtain power dispatch timing feature sequences; based on the power dispatch enhanced semantic vectors, performing semantic gate-based feature modulation on the power dispatch timing feature sequences to obtain modulated power dispatch timing feature sequences; based on the power dispatch enhanced semantic vectors, performing signal reconstruction and adaptive anomaly judgment on the modulated power dispatch timing feature sequences to obtain anomaly judgment results. 2.The AI-based power dispatch monitoring data anomaly detection method of claim 1, wherein, The method of performing multi-modal data preprocessing and time alignment on the power dispatch timing data and the power dispatch log data to obtain windowed power dispatch timing data and power dispatch log text comprises the following steps: performing timing data cleaning on the power dispatch timing data to obtain cleaned power dispatch timing data; performing text normalization and domain term normalization on the power dispatch log data to obtain standard power dispatch log data; performing cross-modal data frame alignment based on a sliding window on the cleaned power dispatch timing data and the standard power dispatch log data to obtain windowed power dispatch timing data and power dispatch log text. 3.The AI-based power dispatch monitoring data anomaly detection method of claim 1, wherein, The method of performing RAG-based enhanced semantic vectorization on the power dispatch log text to obtain power dispatch enhanced semantic vectors comprises the following steps: based on a knowledge base, performing context-independent coding and correlation score extraction on the power dispatch log text to obtain a log vector and a context set; performing context vector aggregation based on correlation attention on the context set to obtain an aggregated context vector; performing hierarchical semantic injection on the log vector and the aggregated context vector to obtain power dispatch enhanced semantic vectors. 4.The AI-based power dispatch monitoring data anomaly detection method of claim 3, wherein, The method of performing hierarchical semantic injection on the log vector and the aggregated context vector to obtain power dispatch enhanced semantic vectors comprises the following steps: performing hierarchical semantic injection on the log vector and the aggregated context vector according to the following formula: wherein, is a dynamic injection gate, is an injection vector, is a power schedule enhancement semantic vector, and is a weight matrix of a fully connected layer, and is a bias vector of a fully connected layer, is a sigmoid activation function, is a Hadamard product, is a log vector, is an aggregated context vector. 5.The AI-based power dispatch monitoring data anomaly detection method of claim 1, wherein, The method of performing double-flow coding on the windowed power dispatch timing data to obtain power dispatch timing feature sequences comprises the following steps: performing forward timing dependency coding on the windowed power dispatch timing data to obtain a forward hidden state sequence; performing backward timing dependency coding on the windowed power dispatch timing data to obtain a backward hidden state sequence; performing bidirectional context feature fusion on the forward hidden state sequence and the backward hidden state sequence to obtain power dispatch timing feature sequences. 6.The AI-based power dispatch monitoring data anomaly detection method of claim 1, wherein, The method of performing semantic gate-based feature modulation on the power dispatch timing feature sequences based on the power dispatch enhanced semantic vectors to obtain modulated power dispatch timing feature sequences comprises the following steps: performing linear transformation and nonlinear activation on the power dispatch enhanced semantic vectors to obtain a semantic gate vector; based on the semantic gate vector, performing semantic-guided modulation on the power dispatch timing feature sequences to obtain modulated power dispatch timing feature sequences. 7.The AI-based power dispatch monitoring data anomaly detection method of claim 1, wherein, Based on the power dispatch enhanced semantic vector, signal reconstruction and adaptive anomaly determination are performed on the modulated power dispatch time sequence feature sequence to obtain an anomaly determination result, including: Signal reconstruction is performed on the modulated power dispatch time sequence feature sequence to obtain reconstructed power dispatch time sequence data; A reconstruction error between the windowed power dispatch time sequence data and the reconstructed power dispatch time sequence data is calculated to obtain a reconstruction score; The power dispatch enhanced semantic vector is input into a small feedforward neural network to obtain an adaptive threshold coefficient; Based on a comparison between the adaptive threshold coefficient and the reconstruction score, the anomaly determination result is generated.

Images