Multi-modal time sequence graph fusion and adaptive big data mining analysis method and system

Abstract

The invention discloses a multi-modal time sequence graph fusion and adaptive big data mining analysis method and system, and relates to the technical field of smart power grid big data analysis. Through the design of a federated learning framework and model parameter decoupling, clients can realize multi-modal time sequence graph fusion and adaptive big data mining analysis on the premise of not sharing original local data; and a global model containing a general rule is cooperatively trained, so that the contradiction between data privacy protection and model generalization ability is effectively solved. And by reserving and specially optimizing the local private parameter set, the method ensures that the final analysis model can closely fit the unique operation characteristics of each power grid, realizes the fusion of general knowledge and local experience, greatly improves the prediction accuracy of the local abnormal state event, and also improves the prediction accuracy of the local abnormal state event. And historical experience is introduced to dynamically correct the current state, so that the reliability and practicability of an analysis result are remarkably enhanced.

Description

TECHNICAL FIELD

[0001] The present application belongs to the technical field of smart grid big data analysis, and particularly relates to a multi-modal time series graph fusion and adaptive big data mining analysis method and system. BACKGROUND

[0002] With the rapid development of smart grids, the data generated by the operation of power systems presents an explosive growth trend. These massive, heterogeneous power big data contain key information for improving the stability and predictive maintenance level of power grid operation. How to deeply mine and utilize these data has become the core issue of the intelligent upgrading of the power industry. Different regional power grids, such as mountainous area power grids dominated by hydropower and coastal city power grids dominated by new energy such as photovoltaic, have significant differences in energy structure, geographical environment, load characteristics and operation mode, and their power operation data also exhibit significantly different characteristics. The existing technology generally adopts a centralized data analysis method when processing data of different power grids. This method requires all sensitive operation data of each power grid to be uploaded to a central server for unified training, which not only brings huge data transmission and storage pressure, but also poses serious risks of data security and privacy leakage. If only isolated analysis models are deployed within a single power grid, the data of other power grids cannot be used to learn more extensive and general power system operation rules, resulting in one-sided model cognition and insufficient generalization ability, making it difficult to cope with complex and variable system states. SUMMARY

[0003] In view of the defects in the prior art, the present application provides a multi-modal time series graph fusion and adaptive big data mining analysis method and system to solve the above technical problems.

[0004] The multi-modal time series graph fusion and adaptive big data mining analysis method is applied to a federated learning system comprising one central server and at least two client systems, and comprises the following steps: S1: In the analysis model deployed on each of the client systems, the model parameters are divided into a global shared parameter set and a local private parameter set; wherein the global shared parameter set is used to learn general power system operation rules through aggregation in federated learning, and the local private parameter set is only driven by local data to optimize to adapt to the local power grid characteristics of the client and does not participate in aggregation.

[0005] S2: Perform multi-round iterative training on the analysis model. In the iterative training, each of the client systems performs the following operations: uploads the model update information corresponding to the global shared parameter set in the client system to the central server for aggregation; downloads the aggregation result from the central server to update the global shared parameter set of itself; and performs local training optimization on the local private parameter set of itself using local data.

[0006] S3: The client system acquires the local power operation data to be analyzed and performs analysis using the trained analysis model. The analysis collaboratively applies the general rules provided by the globally shared parameter set and the local characteristics adapted to the local private parameter set to determine the correlation prediction results with the target abnormal state event.

[0007] Preferably, the at least two client systems include: The first client system deployed in mountainous power grids that are mainly powered by hydropower; A second client system deployed in the power grids of coastal cities, which are mainly based on photovoltaic power.

[0008] Preferably, in the analysis model, the globally shared parameter set is used to perform the following operations: Time, space, data, and events are encoded as graph nodes to construct a temporal knowledge graph, and a temporal graph neural network is used to capture the latent patterns in the temporal knowledge graph to form general temporal features that characterize the general laws. The local private parameter set is used to perform the following operations: Receive the general timing features formed by the globally shared parameter set; By combining the existing fault operation and maintenance tags of the power grid to which the client system belongs, the general time-series characteristics are adaptively integrated and causal inference corrected; Generate personalized analysis results adapted to the local characteristics.

[0009] Preferably, the model update information is the gradient or weight change of the globally shared parameter set.

[0010] Preferably, when each client system uploads its model update information corresponding to the globally shared parameter set to the central server for aggregation, it also performs the following steps: Before uploading, the model update information is processed by homomorphic encryption or differential privacy noise addition.

[0011] Preferably, when each client system uploads its model update information corresponding to the globally shared parameter set to the central server for aggregation, it also performs the following steps: The central server performs a weighted average of the received model update information based on the local data volume, model contribution, or preset weight of each of the at least two client systems.

[0012] Preferably, when the client system acquires the local power operation data to be analyzed and performs analysis using the trained analysis model, the specific steps include: S31: Based on the target power operation and maintenance scenario corresponding to the power operation data to be analyzed, determine the past power operation and maintenance scenarios that meet the scenario commonality judgment requirements of the target power operation and maintenance scenario, and determine the preceding abnormal state event corresponding to the target abnormal state event. S32: Determine the first abnormal state development trend vector of the target abnormal state event and the second abnormal state development trend vector of each of the preceding abnormal state events; S33: Perform a trend vector integration operation on the first abnormal state development trend vector and the second abnormal state development trend vector to obtain the current abnormal state development trend vector of the target abnormal state event in the target power operation and maintenance scenario. S34: Perform joint analysis on the current abnormal state development trend vector and the global state development trend vector of the power operation data to determine the correlation prediction result.

[0013] Preferably, when performing trend vector integration operation on the first abnormal state development trend vector and the second abnormal state development trend vector in step S33, the specific steps include: S331: Based on the first abnormal state development trend vector corresponding to each of the preceding abnormal state events, determine the common value of the target abnormal state event and each of the preceding abnormal state events in the target power operation and maintenance scenario. S332: Based on the second abnormal state development trend vector corresponding to each of the preceding abnormal state events, determine the common value of the past state development trend of the target abnormal state event and each of the preceding abnormal state events in the past power operation and maintenance scenario. S333: Based on the common values ​​of past state development trends, the common values ​​of the target state development trends are corrected to obtain the current abnormal state development trend vector.

[0014] Preferably, step S333, when correcting the common value of the development trend of the target state, specifically includes the following steps: The common value of the development trend of the target state is represented as the degree of difference between the development trend vectors of each of the first abnormal states; The common values ​​of past state development trends are used as adjustment factors to adjust the degree of difference, so as to obtain the current abnormal state development trend vector.

[0015] A multimodal time-series graph fusion and adaptive big data mining and analysis system, comprising a central server and at least two client systems, wherein: The central server is configured to: receive and aggregate model update information from each of the client systems to generate an aggregation result, and distribute the aggregation result back to each of the client systems; Each of the aforementioned client systems is configured as follows: Deploy an analysis model, wherein the model parameters of the analysis model are divided into a globally shared parameter set and a local private parameter set. The globally shared parameter set is used to learn the general operating rules of the power system, and the local private parameter set is used to adapt to the local power grid characteristics of the client and does not participate in aggregation. The analysis model is subjected to multiple rounds of iterative training. During the iterative training, each client system performs the following operations: uploads the model update information corresponding to the globally shared parameter set in the client system to the central server for aggregation; downloads the aggregation results from the central server to update its own globally shared parameter set; and performs localized training optimization on its own local private parameter set using local data. The local power operation data to be analyzed is acquired, and the trained analysis model is used to determine the correlation prediction results with the target abnormal state event by coordinating the global shared parameter set and the local private parameter set.

[0016] The beneficial effects of this invention are as follows: By decoupling the federated learning framework and model parameters, each client can collaboratively train a global model containing general rules without sharing original local data, thus effectively resolving the contradiction between data privacy protection and model generalization ability. Furthermore, by retaining and specifically optimizing local private parameter sets, this method ensures that the final analysis model closely matches the unique operating characteristics of each power grid, achieving a fusion of general knowledge and local experience. This significantly improves the accuracy of predicting local abnormal state events while also introducing historical experience to dynamically correct the current state, significantly enhancing the reliability and practicality of the analysis results. Attached Figure Description

[0017] To more clearly illustrate the technical solutions in the embodiments of this application or the prior art, the drawings used in the description of the embodiments or the prior art will be briefly introduced below. Obviously, the drawings described below are only some embodiments of this application. For those skilled in the art, other drawings can be obtained based on these drawings without creative effort.

[0018] Figure 1 A flowchart illustrating the steps of the multimodal time-series graph fusion and adaptive big data mining and analysis method provided by this invention; Figure 2The flowchart illustrates the steps of a client system for acquiring and analyzing local power operation data using a trained analysis model in the multimodal time-series graph fusion and adaptive big data mining and analysis method provided by this invention. Figure 3 This is a schematic diagram of the structure of the multimodal time series graph fusion and adaptive big data mining and analysis system provided by the present invention. Detailed Implementation

[0019] In the description of this invention, it should be understood that the terms "center," "longitudinal," "lateral," "length," "width," "thickness," "upper," "lower," "front," "rear," "left," "right," "vertical," "horizontal," "top," "bottom," "inner," "outer," "clockwise," "counterclockwise," "axial," "radial," and "circumferential" indicate the orientation or positional relationship based on the orientation or positional relationship shown in the accompanying drawings. They are used only for the convenience of describing this invention and simplifying the description, and do not indicate or imply that the device or element referred to must have a specific orientation, or be constructed and operated in a specific orientation. Therefore, they should not be construed as limitations on this invention.

[0020] The following disclosure provides many different embodiments or examples for implementing different structures of the present invention. To simplify the disclosure of the present invention, the components and arrangements of specific examples are described below. Of course, these are merely examples and are not intended to limit the present invention.

[0021] The embodiments of the invention will now be described in detail with reference to the accompanying drawings.

[0022] like Figure 1 As shown, the multimodal time-series graph fusion and adaptive big data mining analysis method is applied to a federated learning system that includes a central server and at least two client systems, and includes the following steps: S1: In the analysis model deployed on each client system, its model parameters are divided into a global shared parameter set and a local private parameter set; wherein, the global shared parameter set is used to learn the general power system operation law through aggregation in federated learning, and the local private parameter set is only optimized by local data to adapt to the local power grid characteristics of the client and does not participate in aggregation.

[0023] S2: Perform multiple rounds of iterative training on the analysis model. During the iterative training, each client system performs the following operations: uploads the model update information corresponding to the global shared parameter set in the client system to the central server for aggregation; downloads the aggregation results from the central server to update its own global shared parameter set; and performs localized training optimization on its own local private parameter set using local data.

[0024] S3: The client system acquires the local power operation data to be analyzed and performs analysis using the trained analysis model. The analysis collaboratively applies the general rules provided by the globally shared parameter set and the local characteristics adapted to the local private parameter set to determine the correlation prediction results with the target abnormal state event.

[0025] In step S1, client systems are deployed for different power grids. For example, a first client system is deployed in a mountainous power grid dominated by hydropower, and a second client system is deployed in a coastal city power grid dominated by photovoltaic power. Each system deploys the same initial analysis model. The parameters of this model are explicitly divided into two parts: a globally shared parameter set used to encode the power system time-series knowledge graph and composed of a time-series graph neural network; and a local private parameter set corresponding to the personalized adapter used to perform localized corrections of general features. Subsequently, in the iterative training in step S2, the first client system uses its local hydropower operation data for training, calculates the gradient change of the globally shared parameter set, and uploads it to the central server after homomorphic encryption. The second client system also performs the same operation using its local photovoltaic operation data and uploads it simultaneously. After receiving the encrypted gradient, the central server performs a weighted average based on the amount of local data of each client or a preset weight to form a global aggregated gradient and distributes the results. Each client updates its own globally shared parameter set after downloading, while its local private parameter set is continuously optimized using local fault operation and maintenance tags. This process does not interact with the central server. Finally, in the localization analysis stage of step S3, when the mountain power grid needs to analyze a new set of operational data, its client system uses a pre-trained model for local analysis: the data first passes through the globally shared parameter set to extract features reflecting general electrical physics laws, and then these features are fed into the local private parameter set. The latter is then corrected by combining localized knowledge such as the start-up and shutdown modes of hydropower units and reservoir water levels specific to the mountain power grid, ultimately outputting accurate anomaly prediction results for the mountain power grid. This method, through a federated learning mechanism, allows each power grid to collaboratively train a globally shared model with better performance and stronger generalization ability without uploading any original operational data, fundamentally solving the data privacy leakage risk of centralized analysis. At the same time, through parameter set partitioning and specialized optimization of local private parameters, the model can learn and solidify the unique operational characteristics of each power grid. For example, the model can distinguish between the high power fluctuations caused by pumped storage in mountain power grids and the sudden drop in photovoltaic output caused by weather changes in coastal city power grids, avoiding misjudgments that occur when using a single model.

[0026] More specifically, the at least two client systems include: The first client system deployed in mountainous power grids that are mainly powered by hydropower; A second client system deployed in the power grids of coastal cities, which are mainly based on photovoltaic power.

[0027] To address the issue of single models being unable to adapt to different power grid characteristics, this invention deploys specific client-side systems on power grids with significant differences. For example, a first client-side system is deployed in a mountainous power grid rich in hydropower resources. During training, this system's local private parameter set will specifically learn and adapt to operating modes related to hydropower generator units, such as the unique patterns of unit start-up and shutdown, active power regulation, and the nonlinear relationship between reservoir water level and power generation. A second client-side system is deployed in a coastal city power grid with a large photovoltaic installed capacity. This system's local private parameter set will focus on learning and adapting to system characteristics dominated by intermittent renewable energy sources, such as the strong correlation between meteorological factors like solar intensity, cloud cover, and temperature and photovoltaic output, as well as the harmonic characteristics brought about by a large number of inverter devices. During federated training, the two client-side systems share global model parameters regarding the basic operating rules of the power system, while each retains and deepens its understanding of the two distinctly different operating modes of hydropower and photovoltaic power.

[0028] This method enables a globally shared model aggregated through federated learning to simultaneously incorporate experiences from both mountainous hydropower systems and coastal photovoltaic systems. The general principles learned are more robust and universally applicable than models trained on any single system. More importantly, during local forecasting, clients in mountainous power grids can accurately distinguish between sudden power fluctuations caused by normal pumped storage and genuine fault precursors, while clients in coastal power grids can accurately differentiate between sudden drops in photovoltaic output caused by rapidly passing clouds and abnormal grid disconnection. This targeted, personalized adaptation avoids misjudging a normal operating mode of one power grid as an abnormal state of another, fundamentally solving the problem of traditional one-size-fits-all models being incompatible with diverse power grids and having high false alarm rates. This significantly improves the accuracy of the analysis results and their practical application value.

[0029] More specifically, in the analysis model, the globally shared parameter set is used to perform the following operations: Time, space, data, and events are encoded as graph nodes to construct a temporal knowledge graph, and a temporal graph neural network is used to capture latent patterns in the temporal knowledge graph to form general temporal features that characterize the general laws.

[0030] First, when constructing the graph, the system encodes information from different dimensions into heterogeneous nodes; In one implementation, a power grid topology and state coding are established: each physical device such as substation, line, and generator is abstracted as a spatial node, and the measured values ​​such as voltage, current, and power at different times are abstracted as data nodes, and these nodes are connected by edges based on physical connection relationships and time sequence relationships.

[0031] In another approach, an operational event association encoding is established: a lightning strike, a protection action, or a dispatch command is abstracted as an event node, and periodic nodes representing time are associated with it. Simultaneously, the event nodes are connected to the device nodes they affect. After encoding, a spatiotemporal-data-event heterogeneous graph representing the multi-dimensional dynamic relationships of the power system is constructed. Subsequently, a globally shared parameter set composed of a time-series graph neural network begins to learn from this graph. Through its message passing and aggregation mechanisms, it captures potential operational patterns across different time steps and physical locations. For example, it can learn the correlation between voltage drops in a certain area's lines during summer afternoons and the photovoltaic output and air conditioning load in that area. Finally, these highly complex, nonlinear general laws are compressed and output as a set of low-dimensional, dense general time-series feature vectors.

[0032] This method utilizes temporal knowledge graphs to transform isolated, planar time-series data into a three-dimensional structure capable of expressing device topology, temporal dependencies, and causal relationships. This fundamentally solves the problem of traditional models' inability to explicitly model complex relationships in power systems. More importantly, through deep learning in temporal graph neural networks, the model can autonomously discover and quantify high-order operational patterns hidden behind massive amounts of data—patterns that are difficult to summarize solely through expert experience. Examples include the impact of concentrated flood discharge from upstream hydropower stations on the dynamic reactive power reserves of multiple downstream substations over the next few hours. This data-driven learning of general temporal features far surpasses the depth and breadth of traditional methods, providing an extremely rich and profound foundation for subsequent personalized analysis and significantly improving the overall analytical capabilities and accuracy.

[0033] The local private parameter set is used to perform the following operations: Receive the general timing features formed by the globally shared parameter set; By combining the existing fault operation and maintenance tags of the power grid to which the client system belongs, the general time-series characteristics are adaptively integrated and causal inference corrected; Generate personalized analysis results adapted to the local characteristics.

[0034] To address the issue of general models failing to adapt to local characteristics, the specific implementation of the local private parameter set in this invention focuses on the personalized modification and adaptation of general features. First, the local adapter module, comprised of this parameter set, receives the general temporal feature vector output by the upstream globally shared model. Then, this module utilizes locally stored historical data with clearly labeled parameters for targeted optimization.

[0035] In one embodiment, localized causal correction is employed: In the client-side of a mountainous power grid primarily powered by hydropower, the local adapter learns from historical data and discovers that certain alarm features in its general time-series characteristics, triggered by large power fluctuations, are all marked as normal pumped storage startups in the local historical tags. Therefore, after optimization, the local private parameter set automatically applies weighted suppression to these specific general features, achieving causal correction and thus clarifying that such fluctuations are not abnormal in this power grid.

[0036] In another embodiment, localized feature enhancement is employed: In the client-side applications of photovoltaic-dominated coastal city power grids, the local adapter learns that a large number of local line over-temperature fault labels are highly correlated with potential feature vectors representing high ambient humidity and high salt spray in general time-series features. Accordingly, a local private parameter set enhances the weights of these specific general features and integrates them with its own local fault operation and maintenance knowledge to form a highly sensitive local detector specifically targeting accelerated equipment aging patterns under salt spray environments. Finally, the feature vectors, after adaptive integration and causal correction, are fed into the output layer to generate personalized analysis results fully adapted to the local characteristics of the power grid.

[0037] This design allows the model to adapt to each power grid's unique operating history and environmental factors, rather than simply applying a universal standard to all power grids. The direct effect is that it not only filters out false positive alarms caused by the universal model misjudging local normal operating conditions as abnormal, but also amplifies fault precursors that are crucial locally but insensitive to the universal model. This fundamentally improves the signal-to-noise ratio of abnormal state predictions, resulting in highly accurate and reliable analysis results in actual operation and maintenance.

[0038] More specifically, the model update information is the gradient or weight change of the globally shared parameter set.

[0039] In each round of training in federated learning, each client calculates the model loss based on its local data and then obtains the gradient of each parameter in the globally shared parameter set through backpropagation. This gradient vector indicates the direction and magnitude in which the parameters should be adjusted to minimize the local loss. The client does not upload any specific power operation data; it only uploads this calculated gradient vector, which has the same dimension as the globally shared parameter set, to the central server as its "model update information." The central server aggregates the gradients from all clients, updates the global model with this aggregated gradient, and then distributes the updated global parameters to each client. This eliminates the possibility of raw data leakage at the source and greatly reduces the burden on network communication because the amount of information uploaded in each round remains constant, regardless of the size of the local data.

[0040] More specifically, when each client system uploads its model update information corresponding to the globally shared parameter set to the central server for aggregation, it also performs the following steps: Before uploading, the model update information is processed by homomorphic encryption or differential privacy noise addition.

[0041] To address the deep privacy risk that model update information might leak local data distribution characteristics, the client system encrypts the calculated gradient vector using a public key before uploading, generating encrypted gradient ciphertext. This encryption algorithm allows the central server to directly perform a weighted average aggregation of the encrypted gradients from different clients in the ciphertext state, obtaining an aggregated encrypted gradient. The server cannot decrypt any individual gradient or aggregated gradient throughout the entire process; only after distributing the final aggregation result to the client does the client use its private key to decrypt and obtain the aggregated plaintext gradient to update its local model. This dual-security mechanism fundamentally addresses the deep-seated concerns of power entities about their core operating modes being inferred, greatly enhancing their willingness to participate in federated learning and improving system security.

[0042] More specifically, when each client system uploads its model update information corresponding to the globally shared parameter set to the central server for aggregation, it also performs the following steps: The central server performs a weighted average of the received model update information based on the local data volume, model contribution, or preset weight of each of the at least two client systems.

[0043] In one implementation, weighting is based on data volume: before aggregation, the server first determines the weight of each client. The server, through statistics or client reports, determines that a mountainous hydropower client has 10TB of historical operational data, while a coastal photovoltaic client has 20TB of data. During aggregation, the server assigns a weight of 10 / (10+20)≈0.33 to the former and 20 / (10+20)≈0.67 to the latter. This means that the gradient update direction of the photovoltaic client with the larger data volume will dominate the global model update.

[0044] In another implementation, weighting is based on model contribution: after one or more rounds of training, the central server maintains a small, general validation dataset and uses it to evaluate the performance improvement of each client's model on the general task before the update. If the hydropower client's model reduces its loss function on the validation set by 0.5 in this round, while the photovoltaic client's only decreases by 0.2, the server can dynamically assign higher weights (e.g., 0.5 / (0.5+0.2)≈0.71) to the hydropower client with the greater contribution. Alternatively, preset weights can be used, such as directly assigning higher weights to backbone grid clients based on the importance and scale of the power grid.

[0045] By using dynamic weighting based on contribution, the system can intelligently identify and reward excellent systems that provide higher quality and more generalizable updates in the current training round, while suppressing systems that may provide harmful updates due to local data noise or overfitting. This mechanism directly improves the convergence speed of federated learning and the performance ceiling of the final model, making the evolution path of the global model more efficient and robust.

[0046] like Figure 2 As shown, more specifically, when the client system acquires the local power operation data to be analyzed and performs analysis using the trained analysis model, it specifically includes the following steps: S31: Based on the target power operation and maintenance scenario corresponding to the power operation data to be analyzed, determine the past power operation and maintenance scenarios that meet the scenario commonality judgment requirements of the target power operation and maintenance scenario, and determine the preceding abnormal state event corresponding to the target abnormal state event. S32: Determine the first abnormal state development trend vector of the target abnormal state event and the second abnormal state development trend vector of each of the preceding abnormal state events; S33: Perform a trend vector integration operation on the first abnormal state development trend vector and the second abnormal state development trend vector to obtain the current abnormal state development trend vector of the target abnormal state event in the target power operation and maintenance scenario. S34: Perform joint analysis on the current abnormal state development trend vector and the global state development trend vector of the power operation data to determine the correlation prediction result.

[0047] When a client obtains a piece of local power operation data to be analyzed, for example, to analyze the abnormal risk of a transformer under summer peak load, in step S31, the system automatically searches its historical operation and maintenance database to find all historical scenarios that meet the criteria of "summer peak load" and are related to "transformer," and further locates the preceding abnormal state events such as "frequent start-stop of cooling fans" and "slightly excessive oil temperature" that commonly occurred before the final anomaly in these scenarios. Next, in step S32, the system uses a trained model to quantify these events: it encodes the current transformer oil temperature data sequence to be analyzed, forming a first abnormal state development trend vector representing "oil temperature developing from normal to slightly excessive"; at the same time, it also encodes the preceding event of "frequent start-stop of cooling fans" in the historical scenarios, forming its corresponding second abnormal state development trend vector. Subsequently, in step S33, the system performs a trend vector integration operation. The first embodiment is a time-weighted fusion: the second vector representing "frequent cooling fan start-stop" is weighted and fused with the first vector representing "current oil temperature rise." Since historical experience shows that the former is a strong precursor to the latter, the fused "current abnormal state development trend vector" will have stronger directionality. The second embodiment is a multi-precursor event fusion: if multiple preceding events such as "increased partial discharge signal" exist in the history, their respective second vectors are integrated with the first vector to form a more comprehensive current trend vector that incorporates experience from multiple fault evolution paths. Finally, in step S34, the system jointly analyzes this vector representing "micro-fault development trends inferred from historical experience" with the global state development trend vector previously output by the global shared parameter set, which represents "the current overall operating state of the power grid." For example, if the global vector shows that the entire regional power grid is in a vulnerable state of high load and low voltage, then the risk weight of this micro-trend vector pointing to transformer overheating will be significantly amplified, and the final "correlation prediction result" output by the model will be: the transformer is very likely to overheat and trip under the current global operating conditions.

[0048] This method, by incorporating historical operation and maintenance scenarios and preceding events, enables the model to understand that faults do not occur out of thin air but follow a specific evolutionary path, thus solving the problem that traditional methods cannot model the fault development process. More importantly, by jointly analyzing the micro-level fault development trend and the macro-level global operating status, the model can make accurate judgments. It can distinguish between a small equipment indicator anomaly when the overall power grid is operating smoothly and the same indicator anomaly when the power grid is in a high-risk state.

[0049] More specifically, when performing the trend vector integration operation on the first abnormal state development trend vector and the second abnormal state development trend vector in step S33, the following steps are included: S331: Based on the first abnormal state development trend vector corresponding to each of the preceding abnormal state events, determine the common value of the target abnormal state event and each of the preceding abnormal state events in the target power operation and maintenance scenario. S332: Based on the second abnormal state development trend vector corresponding to each of the preceding abnormal state events, determine the common value of the past state development trend of the target abnormal state event and each of the preceding abnormal state events in the past power operation and maintenance scenario. S333: Based on the common values ​​of past state development trends, the common values ​​of the target state development trends are corrected to obtain the current abnormal state development trend vector.

[0050] In step S331, the system compares the current target event to be analyzed with each identified historical preceding event, calculating their "common value of target state development trend" in the current scenario. The first embodiment calculates vector space similarity: the system calculates the cosine similarity between the current oil temperature rise trend vector and the oil temperature rise trend vector that immediately followed the "frequent start-stop of cooling fan" event in the past. This common value quantifies how similar the "current oil temperature rise pattern" is to the "historical oil temperature rise pattern after fan start-stop." Next, in step S332, the system backtracks and calculates the "common value of past state development trends." The second embodiment calculates the consistency of historical evolution patterns: the system calculates the consistency or correlation strength between the development trend vector of the historical preceding event "frequent start-stop of cooling fan" and the "oil temperature rise" trend vector it ultimately leads to. This common value represents the certainty or typicality of the evolution from the precursor "fan start-stop" to the consequence "oil temperature rise" in historical experience. Finally, in step S333, the system performs a core correction operation: adjusting the current pattern similarity calculated in step S331 based on the historical certainty calculated in step S332. Specifically, if historical experience indicates that "fan start-stop" is a very strong and inevitable precursor to "oil temperature rise," then even if the current oil temperature rise pattern deviates slightly from the historical pattern, the system will still amplify the influence of this precursor, ultimately generating a "current abnormal state development trend vector" that is significantly enhanced by the strong historical causal chain.

[0051] This process quantifies the "credibility" and "correlation strength" of historical experience by calculating common values ​​from the past. It then uses this "credibility" to adjust the "similarity" between currently observed phenomena and historical patterns, fundamentally solving the adaptation problem when historical experience doesn't perfectly match the current reality. It can intelligently distinguish which historical experiences have strong guiding significance for the current situation and which are merely weakly correlated or even coincidental. This ensures that the final "current abnormal state development trend vector" both absorbs the essence of historical experience and fully respects the uniqueness of current data.

[0052] More specifically, in step S333, when correcting the common value of the development trend of the target state, the following steps are included: The common value of the development trend of the target state is represented as the degree of difference between the development trend vectors of each of the first abnormal states; The common values ​​of past state development trends are used as adjustment factors to adjust the degree of difference, so as to obtain the current abnormal state development trend vector.

[0053] In the first step of the correction process, the system explicitly represents the "common value of the target state development trend" as a kind of "difference degree." One calculation method uses vector distance: the system calculates the Euclidean or Mahalanobis distance between the trend vector of the current target event being analyzed and the oil temperature rise trend vector triggered by a previous event in history. This distance value is the "difference degree," directly quantifying the degree to which the current development pattern deviates from the historical typical pattern. Next, in the second step of the correction, the system uses the previously calculated "common value of past state development trends" as an "adjustment factor" to directly adjust this "difference degree" mathematically. One calculation method uses difference degree scaling adjustment: the system uses a high "common value of past state development trends" representing a strong historical causal relationship as a multiplier factor to multiply the "difference degree" calculated in the first step. If the historical correlation is extremely strong, this adjustment factor will significantly reduce the currently observed difference degree. Its physical meaning is: "Although the current pattern is not completely consistent with history, given the high certainty of historical experience, we should believe that the current event is being forcibly pulled into the historical evolutionary trajectory." Conversely, if the historical correlation is weak, the difference is hardly adjusted, indicating that historical experience has no reference value. The adjusted difference is reintegrated to form a highly accurate "current abnormal state development trend vector" that combines the strength of historical experience with the deviation of the current pattern.

[0054] By deconstructing current similarity into differences between the present and the past, and using historical certainty as a lever to adjust these differences, the model no longer simply pulls the vector closer or further away. Instead, it determines the extent to which it tolerates or corrects deviations between current trends and historical patterns based on the strength of historical experience. This enables the model to make reasonable judgments when faced with new situations that are highly similar to but slightly different from historical experience, neither blindly following history nor ignoring subtle changes in current data.

[0055] like Figure 3 As shown, the multimodal time-series graph fusion and adaptive big data mining and analysis system includes a central server and at least two client systems, wherein: The central server is configured to: receive and aggregate model update information from each of the client systems to generate an aggregation result, and distribute the aggregation result back to each of the client systems; Each of the aforementioned client systems is configured as follows: Deploy an analysis model, wherein the model parameters of the analysis model are divided into a globally shared parameter set and a local private parameter set. The globally shared parameter set is used to learn the general operating rules of the power system, and the local private parameter set is used to adapt to the local power grid characteristics of the client and does not participate in aggregation. The analysis model is subjected to multiple rounds of iterative training. During the iterative training, each client system performs the following operations: uploads the model update information corresponding to the globally shared parameter set in the client system to the central server for aggregation; downloads the aggregation results from the central server to update its own globally shared parameter set; and performs localized training optimization on its own local private parameter set using local data. The local power operation data to be analyzed is acquired, and the trained analysis model is used to determine the correlation prediction results with the target abnormal state event by coordinating the global shared parameter set and the local private parameter set.

[0056] Finally, it should be noted that the above embodiments are only used to illustrate the technical solutions of the present invention, and not to limit them. Although the present invention has been described in detail with reference to the foregoing embodiments, those skilled in the art should understand that modifications can still be made to the technical solutions described in the foregoing embodiments, or equivalent substitutions can be made to some or all of the technical features therein. Such modifications or substitutions do not cause the essence of the corresponding technical solutions to deviate from the scope of the technical solutions of the embodiments of the present invention, and they should all be covered within the scope of the claims and specification of the present invention.

Claims

1. A multimodal temporal atlas fusion and adaptive big data mining analysis method, characterized in that, The method is applied to a federated learning system comprising a central server and at least two client systems, and specifically comprises the following steps: S1: in an analysis model deployed on each of the client systems, dividing model parameters into a globally shared parameter set and a locally private parameter set; wherein the globally shared parameter set is used to learn general power system operation rules through aggregation in federated learning, and the locally private parameter set is only driven by local data for optimization to adapt to the local grid characteristics of the client and does not participate in aggregation; S2: performing multiple rounds of iterative training on the analysis model, wherein in the iterative training, each of the client systems performs the following operations: uploading model update information corresponding to the globally shared parameter set in the client system to the central server for aggregation; downloading aggregation results from the central server to update the globally shared parameter set of itself; and performing localized training optimization on the locally private parameter set of itself using local data; S3: the client system obtains local power operation data to be analyzed, and uses the trained analysis model to perform analysis, which cooperates the general rules provided by the globally shared parameter set with the local characteristics adapted by the locally private parameter set to determine a correlation prediction result with a target abnormal state event.

2. The multimodal time series graph fusion and adaptive big data mining analysis method of claim 1, wherein, The at least two client systems comprise: a first client system deployed in a mountainous area power grid dominated by hydropower; a second client system deployed in a coastal city power grid dominated by photovoltaic power. 3.The multi-modal time series graph fusion and adaptive big data mining analysis method of claim 1, wherein, In the analysis model, the globally shared parameter set is used to perform the following operations: encoding time, space, data, and events as graph nodes to construct a time series knowledge graph, and using a time series graph neural network to capture potential patterns in the time series knowledge graph to form general time series features representing the general rules; the locally private parameter set is used to perform the following operations: receiving the general time series features formed by the globally shared parameter set; integrating and causally inferring the general time series features in combination with existing fault operation and maintenance labels of the power grid to which the client system belongs; generating individualized analysis results adapted to the local characteristics.

4. The multimodal time series graph fusion and adaptive big data mining analysis method of claim 1, wherein, The model update information is a gradient or weight change amount of the globally shared parameter set.

5. The multimodal time series graph fusion and adaptive big data mining analysis method of claim 4, wherein, When each client system uploads its model update information corresponding to the globally shared parameter set to the central server for aggregation, it further performs the following steps: Before uploading, the model update information is homomorphically encrypted or differentially private noise added.

6. The multimodal time series graph fusion and adaptive big data mining analysis method of claim 1, wherein, When each client system uploads its model update information corresponding to the globally shared parameter set to the central server for aggregation, it further performs the following steps: The central server performs weighted averaging on the received model update information according to the local data amount, model contribution, or preset weight of each of the at least two client systems.

7. The multimodal time series graph fusion and adaptive big data mining analysis method of claim 6, wherein, When the client system obtains local power operation data to be analyzed and uses the trained analysis model to perform analysis, it specifically comprises the following steps: S31: determining, according to a target power operation and maintenance scene corresponding to the power operation data to be analyzed, a past power operation and maintenance scene conforming to a scene commonality judgment requirement of the target power operation and maintenance scene, and determining a precursor abnormal state event corresponding to the target abnormal state event; S32: determining a first abnormal state development trend vector of the target abnormal state event and a second abnormal state development trend vector of each of the precursor abnormal state events; S33: performing a trend vector integration operation on the first abnormal state development trend vector and the second abnormal state development trend vector to obtain a current abnormal state development trend vector of the target abnormal state event in the target power operation and maintenance scene; S34: jointly analyzing the current abnormal state development trend vector and a global state development trend vector of the power operation data to determine the correlation prediction result.

8. The multimodal time series graph fusion and adaptive big data mining analysis method of claim 7, wherein, The trend vector integration operation on the first abnormal state development trend vector and the second abnormal state development trend vector in the step S33 specifically includes the following steps: S331: determining a target state development trend commonality value of the target abnormal state event and each of the precursor abnormal state events in the target power operation and maintenance scene based on the first abnormal state development trend vector corresponding to each of the precursor abnormal state events; S332: determining a past state development trend commonality value of the target abnormal state event and each of the precursor abnormal state events in the past power operation and maintenance scene based on the second abnormal state development trend vector corresponding to each of the precursor abnormal state events; S333: correcting the target state development trend commonality value according to the past state development trend commonality value to obtain the current abnormal state development trend vector.

9. The multimodal time series graph fusion and adaptive big data mining analysis method of claim 8, wherein, The correction of the target state development trend commonality value in the step S333 specifically includes the following steps: expressing the target state development trend commonality value as a difference degree between the first abnormal state development trend vectors; adjusting the difference degree by taking the past state development trend commonality value as an adjustment factor to obtain the current abnormal state development trend vector.

10. A multimodal temporal atlas fusion and adaptive big data mining analysis system, characterized in that, The system includes a central server and at least two client systems, wherein: the central server is configured to receive and aggregate model update information from each of the client systems to generate an aggregation result, and distribute the aggregation result back to each of the client systems; each of the client systems is configured to: deploy an analysis model, wherein model parameters of the analysis model are divided into a global shared parameter set and a local private parameter set, the global shared parameter set is used to learn general power system operation rules, and the local private parameter set is used to adapt to local grid characteristics of the client and does not participate in aggregation; Performing multiple rounds of iterative training on the analysis model, in each round of iterative training, each of the client systems performs the following operations: uploading model update information corresponding to the global shared parameter set in the client system to the central server for aggregation; downloading the aggregation result from the central server to update the global shared parameter set of itself; and performing local training optimization on the local private parameter set of itself using local data; Obtaining local power operation data to be analyzed, and using the trained analysis model to determine a correlation prediction result with a target abnormal state event by coordinating the global shared parameter set and the local private parameter set.

Images