AI prediction and active early warning system for smart energy field station

Abstract

The application relates to the field of intelligent operation and maintenance of power systems, and discloses an AI prediction and active early warning system for a smart energy station, which obtains multi-source heterogeneous operation monitoring data; then divides the data modes, constructs independent analysis models, and corrects output continuous state deviation vectors in combination with remote signaling data; determines dynamic weights and outputs weighted state deviation vectors by analyzing the causal correlation among the modes; constructs a normal correlation matrix to calculate consistency scores and accurately determines whether the operation consistency among the modes is abnormal; constructs a model according to the determination result, performs time series deduction on the weighted state deviation vectors, and obtains state evolution trajectories in future periods; and finally, based on a strategy library, intervention deduction is performed, and an optimal intervention path and active early warning information are output. The application solves the problem of high false alarm and missed alarm rates in single dimension analysis in the prior art, realizes closed-loop management from passive alarm to active early warning intervention, and significantly improves the operation and maintenance level of the smart energy station.

Description

Technical Field

[0001] This invention relates to the field of intelligent operation and maintenance technology for power systems, and more specifically to an AI prediction and proactive early warning system for smart energy plants. Background Technology

[0002] With the accelerated construction of new power systems, smart power early warning technology has entered a period of rapid development. my country has achieved remarkable results in the field of power fault monitoring. For example, it has established the first full-voltage-level lightning fault location system. Some enterprises and research institutions have developed systems that use AI deep learning algorithms to accurately identify electrical hazards, with an early warning accuracy rate of over 98%. This has effectively shortened fault handling time in areas such as distribution network traveling wave fault location and early warning of equipment.

[0003] However, despite breakthroughs in single-dimensional monitoring and early warning, there are still significant technological shortcomings in the application of integrated smart energy stations encompassing "source, grid, load, and storage":

[0004] Most existing early warning systems analyze single devices or single types of data (such as monitoring only voltage or temperature) independently, lacking effective mining of causal relationships between multimodal data from the generation side, grid side, load side, equipment side, and environment side. Data from different modes are often processed in isolation, making it difficult to construct normal correlation matrices across modes. This results in the system's inability to accurately identify inconsistencies between operating parameters, making it highly susceptible to false alarms or missed alarms in situations of fluctuating renewable energy output or complex environments.

[0005] Existing early warning mechanisms are mostly passive alarms based on thresholds, meaning alarms are only triggered after a fault occurs or an indicator exceeds a limit. The system lacks the ability to extrapolate state evolution trajectories based on historical time-series data, making it impossible to predict the future trend of state deviation components. Furthermore, early warning information often remains at the "alarm" level, lacking intervention and extrapolation mechanisms based on a strategy library. This fails to provide maintenance personnel with proactive decision support including optimal intervention paths, making it difficult to form a complete closed-loop process of "monitoring-analysis-early warning-response." Summary of the Invention

[0006] In order to overcome the above-mentioned deficiencies of the prior art, the present invention provides an AI prediction and proactive early warning system for smart energy power plants to solve the problems existing in the background art.

[0007] This invention provides the following technical solution: an AI prediction and proactive early warning system for smart energy power plants, comprising:

[0008] Data acquisition module: Collects multi-source heterogeneous operation monitoring data of smart energy power stations and performs preprocessing. The multi-source heterogeneous operation monitoring data includes multi-dimensional real-time operation data and historical operation benchmark data.

[0009] Multimodal analysis module: Based on data type, real-time running data is divided into different modes, and independent analysis models are constructed using historical running benchmark data corresponding to each mode. The output is corrected by combining synchronously acquired remote signaling data, and a continuous state deviation vector is output.

[0010] Causal weighting module: Based on historical operating baseline data, causal correlation analysis is performed between each mode to obtain the causal influence coefficient between each mode. The dynamic weight of each mode is determined according to the causal influence coefficient. The continuous state deviation vector is weighted to obtain the weighted state deviation vector.

[0011] Consistency determination module: Constructs a normal correlation matrix between modalities based on historical operating benchmark data, calculates the consistency score between the weighted state deviation vector and the normal correlation matrix, and determines whether the operational consistency between modalities is abnormal;

[0012] State trajectory prediction module: Based on the determination result of the consistency between modes, a state trajectory prediction model is constructed based on historical operation benchmark data. The state trajectory prediction model is used to perform time-series extrapolation on the weighted state deviation vector to obtain the state evolution trajectory for future periods.

[0013] Intervention Decision and Early Warning Module: Based on the intervention strategy library, the module uses a state trajectory prediction model to perform intervention simulation on the state evolution trajectory, obtain the optimal intervention path, and output proactive early warning information.

[0014] Preferably, the data acquisition module acquires multi-source heterogeneous operation monitoring data of the smart energy station and performs preprocessing including:

[0015] By deploying IoT sensors, smart meters, and SCADA systems at the power station, and using the IEC104 or Modbus communication protocol, real-time data is collected from the generation side, grid side, load side, equipment status, and environmental meteorological data. The generation side data includes real-time power generation, cumulative power generation, photovoltaic module temperature, irradiance, and energy storage system state of charge. The load side data includes real-time load power. The grid side data includes voltage, frequency, active power, reactive power, current, three-phase voltage, and three-phase current. The equipment status data includes unit temperature, transformer temperature, circuit breaker temperature, energy storage battery temperature, equipment operating status, and circuit breaker status. The environmental meteorological data includes temperature, humidity, and irradiance. The SCADA system synchronously collects telemetry and remote signaling data via the IEC104 protocol; the remote signaling data includes circuit breaker status and protection action signals.

[0016] Simultaneously extract historical operational baseline data from the historical database for the time period corresponding to the real-time operational data;

[0017] The real-time running data is denoised and normalized, and the historical running baseline data is filled with missing values ​​and time aligned.

[0018] Preferably, the multimodal analysis module divides real-time operating data into different modalities based on data type, constructs independent analysis models using historical operating benchmark data corresponding to each modality, and outputs continuous state deviation vectors, specifically including:

[0019] Based on the data source and business attributes, the power generation data, grid data, load data, equipment status data, and environmental meteorological data are respectively divided into power generation mode, grid mode, load mode, equipment mode, and environmental mode;

[0020] Using historical operating baseline data corresponding to each mode, time series-based baseline prediction models are constructed as independent analysis models. Among them, the power grid mode model predicts the normal baseline trajectory of voltage and frequency, the equipment mode model predicts the normal baseline trajectory of unit temperature, the load mode model predicts the normal baseline trajectory of real-time load power, the power generation mode model predicts the normal baseline trajectory of unit output, and the environmental mode model predicts the normal baseline trajectory of environmental parameters.

[0021] Calculate the real-time deviation between the power generation data, grid data, load data, equipment status data, and environmental meteorological data and their respective normal baseline trajectories;

[0022] The circuit breaker change time in the remote signaling data is acquired synchronously, and the sudden change time of the real-time deviation is compared with the circuit breaker change time. If the sudden change time and the circuit breaker change time are synchronized, it is determined to be a normal operation response and filtered out. If the timing is not synchronized, the deviation fluctuation is retained as an abnormal deviation signal. The filtered and retained deviation data are integrated to obtain the effective state deviation data.

[0023] The effective state deviation data is encoded according to each modal dimension to construct a continuous state deviation vector containing voltage deviation component, frequency deviation component, temperature deviation component, load deviation component and environmental deviation component.

[0024] Preferably, the causal weighting module performs causal correlation analysis on each mode based on historical operating baseline data to obtain the causal influence coefficient between each mode, specifically including:

[0025] Based on historical operational baseline data corresponding to each mode, a cross-domain coupling correlation graph is constructed with each mode as a node and the causal influence path between modes as an edge. Specifically, the correlation of heat conduction is calculated using historical time series of environmental meteorological data and equipment status data; the correlation of power-frequency response is calculated using historical time series of power generation data and grid data; and the voltage-load sensitivity is calculated using historical time series of load data and grid data. Furthermore, the correlation of supply and demand balance between power generation data and load data, and the correlation of output response between environmental meteorological data and power generation data are calculated. The correlation indicators obtained from the above calculations are normalized and used as the causal influence coefficients between each mode pair, which are then filled into the corresponding edge weights of the cross-domain coupling correlation graph.

[0026] The process of determining the dynamic weights of each mode based on the causal influence coefficient and weighting the continuous state deviation vectors to obtain a weighted state deviation vector specifically includes:

[0027] For each mode, the sum of the causal influence coefficients between the mode and the other modes is used as the comprehensive influence intensity of the mode. The comprehensive influence intensity of each mode is normalized to generate a dynamic weight vector for each mode in the early warning judgment. A dynamic weight matrix is ​​constructed with the dynamic weight vector as the diagonal element.

[0028] The deviation components corresponding to each mode in the continuous state deviation vector are weighted and summed with the dynamic weight matrix to calculate the weighted state deviation vector.

[0029] Preferably, the consistency determination module constructs a normal correlation matrix between modalities based on historical operational benchmark data, calculates the consistency score between the weighted state deviation vector and the normal correlation matrix, and determines whether the operational consistency between modalities is abnormal, specifically including:

[0030] Based on the historical operating baseline data corresponding to each mode, the power-frequency response characteristics between the power generation mode and the grid mode, the voltage-load sensitivity characteristics between the grid mode and the load mode, the temperature-power correlation characteristics between the power generation mode and the equipment mode, the heat conduction response characteristics between the environmental mode and the equipment mode, and the output response characteristics between the environmental mode and the power generation mode are extracted respectively, and a 5×5 normal correlation matrix is ​​constructed. Among them, the statistical mean and covariance of the historical time series are used to characterize the correlation characteristics of each mode pair. The diagonal elements of the matrix are the operating stability index of each mode itself, and the off-diagonal elements are the correlation strength index between the mode pairs.

[0031] Based on the covariance matrix of the normal correlation matrix, the Mahalanobis distance algorithm is used to calculate the consistency distance between the weighted state deviation vector and the center of the normal baseline distribution. The consistency distance is then mapped to a consistency score in the range of 0 to 1 using an exponential decay function, where the closer the consistency score is to 1, the higher the operational consistency.

[0032] The consistency score is compared with a preset consistency judgment threshold. If the consistency score is lower than the preset consistency judgment threshold, the intermodal operation consistency is determined to be abnormal and an early warning message is triggered. The early warning message is an AI prediction early warning message generated based on the consistency judgment result. If the consistency score is not lower than the preset consistency judgment threshold, the intermodal operation consistency is determined to be normal.

[0033] Preferably, the state trajectory prediction module constructs a state trajectory prediction model based on historical operating baseline data, specifically including:

[0034] Based on the consistency score output by the consistency determination module, select the type of state trajectory prediction model:

[0035] When the consistency score is lower than the preset consistency judgment threshold, a nonlinear trajectory prediction model based on long short-term memory network is constructed; when the consistency score is not lower than the preset consistency judgment threshold, a linear trajectory prediction model based on linear regression is constructed.

[0036] Using historical time series sequences of each mode in historical operating benchmark data as training samples, the historical time series sequences are processed by a time series decomposition algorithm to separate periodic components, trend components and residual components. The periodic components, trend components and residual components are then input into the state trajectory prediction model for training to obtain the trained state trajectory prediction model.

[0037] The process of using a state trajectory prediction model to perform time-series extrapolation on the weighted state deviation vector to obtain the state evolution trajectory for future periods specifically includes:

[0038] The weighted state deviation vector is used as the current input state of the state trajectory prediction model. The model is iteratively predicted step by step according to a preset time step to obtain the initial predicted trajectory sequence.

[0039] The covariance matrix of the normal correlation matrix is ​​called, and the Mahalanobis distance of the predicted vector at each time step in the initial predicted trajectory sequence relative to the normal reference center is calculated using the covariance matrix to obtain the prediction deviation at the current time step; the prediction deviation is compared with the preset deviation judgment threshold.

[0040] If the prediction deviation exceeds the preset deviation judgment threshold, the prediction vector is projected and corrected based on the eigenvector of the normal correlation matrix, and the corrected vector is used as the iterative input for the next time period; if the prediction deviation does not exceed the preset deviation judgment threshold, the prediction vector is directly used as the iterative input for the next time period.

[0041] After the iteration is completed, the state evolution trajectory for multiple future time periods is output; the state evolution trajectory is a numerical sequence of state deviation for each mode in each future time period, and the numerical points in the numerical sequence of state deviation that exceed the preset safety boundary are marked as abnormal warning indicators.

[0042] Preferably, the intervention decision-making and early warning module performs intervention deduction on the state evolution trajectory based on the intervention strategy library to obtain the optimal intervention path and outputs proactive early warning information, specifically including:

[0043] The system captures abnormal time periods in the state evolution trajectory where the deviation component exceeds a preset safety boundary and the corresponding target mode. Based on a preset anomaly-policy mapping relationship, it calls several candidate intervention strategies that match the target mode from the intervention strategy library. The intervention strategy library includes power adjustment instructions for frequency mode anomalies, reactive power compensation instructions for voltage mode anomalies, and load transfer instructions for equipment mode temperature anomalies.

[0044] An intervention effect evaluation function is constructed, and candidate intervention strategies are substituted into the state trajectory prediction model as input variables for simulation and deduction. The expected adjustment time and required adjustment magnitude of the state evolution trajectory returning to the safety boundary after the implementation of each candidate intervention strategy are calculated.

[0045] A multi-objective optimization algorithm was adopted, with the optimization objectives of minimizing adjustment time and minimizing adjustment amplitude. Candidate intervention strategies were comprehensively scored, and the strategy with the highest score was selected as the optimal intervention path.

[0046] Based on the optimal intervention path, proactive early warning information containing the target equipment identifier, adjustment direction, and adjustment value is generated and pushed to the site monitoring interface.

[0047] The technical effects and advantages of this invention are as follows:

[0048] (1) The causal influence coefficient and dynamic weight are obtained by performing causal correlation analysis between each mode through the causal weight module. Combined with the consistency judgment module, the normal correlation matrix between modes is constructed to calculate the consistency score. This breaks the data silos of multi-source heterogeneous data and realizes the effective mining of deep correlation characteristics of multi-modal data such as power generation side, power grid side, and load side. It can accurately identify the consistency anomalies between operating parameters, effectively solve the problem of false alarms and missed alarms caused by single-dimensional analysis under the fluctuation of new energy output or complex environment, and significantly improve the accuracy and robustness of the early warning system.

[0049] (2) The state trajectory prediction module constructs a state trajectory prediction model based on the consistency judgment result, and performs time-series deduction on the weighted state deviation vector to obtain the state evolution trajectory for future periods. This changes the traditional passive alarm mode based on thresholds, realizes the advance prediction of the future trend of state deviation components, and can detect hidden dangers in advance before the fault occurs or the indicator exceeds the limit, thus saving valuable fault handling time for operation and maintenance personnel and significantly improving the initiative and timeliness of early warning.

[0050] (3) The intervention decision and early warning module uses the state trajectory prediction model based on the intervention strategy library to perform intervention deduction to obtain the optimal intervention path, which breaks through the limitation of traditional early warning only staying at the "alarm" level. It can provide maintenance personnel with proactive decision support including specific adjustment range and response time, and realize the closed-loop management of the whole process of "monitoring-analysis-early warning-disposal", effectively shortening the fault handling time and reducing maintenance costs. Attached Figure Description

[0051] Figure 1 This is a system structure block diagram of the present invention. Detailed Implementation

[0052] The technical solutions of the present invention will be clearly and completely described below with reference to the accompanying drawings. In addition, the forms of the various structures described in the following embodiments are merely illustrative. The AI ​​prediction and proactive early warning system for smart energy stations involved in the present invention is not limited to the structures described in the following embodiments. All other embodiments obtained by those skilled in the art without creative effort are within the scope of protection of the present invention.

[0053] like Figure 1 The embodiment shown provides an AI prediction and proactive early warning system for smart energy power plants, including:

[0054] Data acquisition module: Collects multi-source heterogeneous operation monitoring data of smart energy power stations and performs preprocessing. The multi-source heterogeneous operation monitoring data includes multi-dimensional real-time operation data and historical operation benchmark data.

[0055] In this embodiment, the data acquisition module collects multi-source heterogeneous operation monitoring data from the smart energy station and performs preprocessing, including:

[0056] By deploying IoT sensors, smart meters, and SCADA systems at the power station, and using the IEC104 or Modbus communication protocol, real-time data is collected from the generation side, grid side, load side, equipment status, and environmental meteorological data. The generation side data includes real-time power generation, cumulative power generation, photovoltaic module temperature, irradiance, and energy storage system state of charge. The load side data includes real-time load power. The grid side data includes voltage, frequency, active power, reactive power, current, three-phase voltage, and three-phase current. The equipment status data includes unit temperature, transformer temperature, circuit breaker temperature, energy storage battery temperature, equipment operating status, and circuit breaker status. The environmental meteorological data includes temperature, humidity, and irradiance. The SCADA system synchronously collects telemetry and remote signaling data via the IEC104 protocol; the remote signaling data includes circuit breaker status and protection action signals.

[0057] Simultaneously extract historical operational baseline data from the historical database for the time period corresponding to the real-time operational data;

[0058] The real-time running data is denoised and normalized, and the historical running baseline data is filled with missing values ​​and time aligned.

[0059] It should be specifically noted that the data acquisition module acquires and preprocesses multi-source heterogeneous operation monitoring data in the following ways:

[0060] At the hardware level, the data acquisition module deploys various acquisition terminals: temperature sensors and irradiance sensors are installed in the photovoltaic array area of ​​the power station to sense the surface temperature of the photovoltaic modules and the intensity of solar irradiance; smart meters and power analyzers are installed at the inverters, energy storage converters, and grid connection points to collect real-time power generation, cumulative power generation, energy storage system state of charge, and electrical quantities such as voltage, frequency, active power, reactive power, current, three-phase voltage, and three-phase current at the grid connection point; temperature monitoring sensors and status monitoring devices are deployed on key equipment such as the power station's step-up transformers, circuit breakers, and energy storage battery compartments to collect unit temperature, transformer temperature, circuit breaker temperature, energy storage battery temperature, equipment operating status, and circuit breaker status; meteorological monitoring stations are set up around the power station to collect ambient temperature, ambient humidity, and irradiance data. All of the above sensors and smart meters are connected to the power station's data aggregation gateway via RS485 bus or Ethernet interface. The data aggregation gateway, acting as an edge computing node, performs initial aggregation and format unification of the raw signals.

[0061] At the communication protocol level, the data aggregation gateway and the site SCADA system use the Modbus TCP protocol for data interaction, while the SCADA system uses the IEC 104 communication protocol for remote data transmission with the upper-level dispatch master station or centralized control center. The IEC 104 protocol transmits telemetry data and remote signaling data simultaneously. This remote signaling data specifically includes the open / closed status of each circuit breaker and the action signals of relay protection devices. The remote signaling data is reported in the form of switch signals, and each status change triggers an upload, ensuring that changes in the status of site equipment can be captured in real time.

[0062] During real-time data acquisition, the data aggregation gateway periodically polls and collects data from the power generation side, grid side, load side, equipment status, and environmental meteorological data at a sampling frequency of no less than once per second. After each poll, the data is timestamped and temporarily stored in a local cache queue. Subsequently, the real-time data is pushed to the system's data processing layer via a publish-subscribe mechanism through a message queue middleware. Simultaneously, the data acquisition module extracts historical operating baseline data from the station's historical database, based on the time window of the current real-time data (e.g., looking back 30 days or longer from the current moment). This historical operating baseline data consists of statistical or recorded values ​​of similar monitoring parameters under the same season and operating conditions in the past. It serves as the input basis for the subsequent multimodal analysis module to build the baseline model and for the state trajectory prediction module to train.

[0063] In the data preprocessing stage, for real-time operational data, a sliding window-based median filtering method or wavelet threshold denoising method is first used to remove noise introduced by occasional sensor jumps and electromagnetic interference. Then, a minimum-maximum normalization method is used to uniformly map the data of each dimension to the interval between zero and one, eliminating the influence of different units and orders of magnitude. For historical operational baseline data, linear interpolation or a method based on the average of adjacent time periods is first used to fill in missing data points caused by sensor failures or communication interruptions. Then, a timestamp alignment method is used to uniformly align historical data from different acquisition frequencies to the same time granularity (e.g., uniformly align to a five-minute interval or a fifteen-minute interval), ensuring that the historical baseline data has consistent comparability in the time dimension, providing reliable and uniformly formatted input data for the subsequent multimodal analysis module and state trajectory prediction module.

[0064] Multimodal analysis module: Based on data type, real-time running data is divided into different modes. Independent analysis models are constructed using historical running benchmark data corresponding to each mode. The output is corrected by combining synchronously acquired remote signaling data, and a continuous state deviation vector is output.

[0065] In this embodiment, the multimodal analysis module divides real-time operating data into different modalities based on data type, constructs independent analysis models using historical operating benchmark data corresponding to each modality, and outputs continuous state deviation vectors, specifically including:

[0066] Based on the data source and business attributes, the power generation data, grid data, load data, equipment status data, and environmental meteorological data are respectively divided into power generation mode, grid mode, load mode, equipment mode, and environmental mode;

[0067] Using historical operating baseline data corresponding to each mode, time series-based baseline prediction models are constructed as independent analysis models. Among them, the power grid mode model predicts the normal baseline trajectory of voltage and frequency, the equipment mode model predicts the normal baseline trajectory of unit temperature, the load mode model predicts the normal baseline trajectory of real-time load power, the power generation mode model predicts the normal baseline trajectory of unit output, and the environmental mode model predicts the normal baseline trajectory of environmental parameters.

[0068] Calculate the real-time deviation between the power generation data, grid data, load data, equipment status data, and environmental meteorological data and their respective normal baseline trajectories;

[0069] The circuit breaker change time in the remote signaling data is acquired synchronously, and the sudden change time of the real-time deviation is compared with the circuit breaker change time. If the sudden change time and the circuit breaker change time are synchronized, it is determined to be a normal operation response and filtered out. If the timing is not synchronized, the deviation fluctuation is retained as an abnormal deviation signal. The filtered and retained deviation data are integrated to obtain the effective state deviation data.

[0070] The effective state deviation data is encoded according to each modal dimension to construct a continuous state deviation vector containing voltage deviation component, frequency deviation component, temperature deviation component, load deviation component and environmental deviation component.

[0071] It should be specifically explained that the multimodal analysis module first divides the real-time operating data into five modes based on the data source and business attributes: real-time power generation, cumulative power generation, photovoltaic module temperature, irradiance, and energy storage system state of charge are classified as the power generation mode; voltage, frequency, active power, reactive power, current, three-phase voltage, and three-phase current are classified as the grid mode; real-time load power is classified as the load mode; unit temperature, transformer temperature, circuit breaker temperature, energy storage battery temperature, equipment operating status, and circuit breaker status are classified as the equipment mode; and temperature, humidity, and irradiance are classified as the environmental mode. For each mode, the system uses the historical operating benchmark data corresponding to that mode (i.e., historical records of the same type of monitoring parameters under the same season and operating conditions in the past) to construct a time series-based benchmark prediction model. Specifically, the power grid mode uses a Long Short-Term Memory (LSTM) network model, with historical voltage and frequency time series as input, to train and obtain normal baseline trajectory prediction curves for voltage and frequency; the equipment mode uses a sliding window regression-based time series model, with historical unit temperature time series as input, to train and obtain normal baseline trajectory prediction curves for unit temperature; the load mode uses exponential smoothing or ARIMA models, with historical real-time load power time series as input, to train and obtain normal baseline trajectory prediction curves for real-time load power; the generation mode uses a prediction method that combines physical mechanisms and data-driven approaches, with historical irradiance, component temperature, and generation power joint time series as input, to train and obtain normal baseline trajectory prediction curves for unit output; and the environmental mode uses multiple linear regression or random forest regression models, with historical environmental temperature, humidity, and irradiance time series as input, to train and obtain normal baseline trajectory prediction curves for environmental parameters. After training, each of the aforementioned independent analysis models takes real-time data from several consecutive time steps prior to the current moment (e.g., the past 24 hours or the past 48 hours) as input, and predicts forward step by step, outputting the normal baseline trajectory value for one or more future time steps. For real-time deviation calculation, a formula is used... Calculate the deviation of each mode, where The real-time monitoring value of mode i at time t. The normal baseline trajectory value predicted by the model. This represents the real-time deviation.

[0072] In terms of remote signaling data correction, the system synchronously acquires remote signaling data collected via the IEC 104 protocol from the SCADA system, focusing on extracting the circuit breaker state change time (i.e., the precise moment when the circuit breaker changes from closed to open or from open to closed). The system compares the moments of abrupt changes in the real-time deviation sequence of each mode (i.e., a jump in deviation exceeding a preset threshold between adjacent time steps, which can be set as the mean plus three times the standard deviation based on the statistical distribution of historical deviations) with the circuit breaker change time: if the time difference between the deviation change time of a certain mode and the circuit breaker change time is within the preset allowable range (e.g., within five seconds before and after), the deviation change is determined to be a normal operational response caused by the circuit breaker operation, and the deviation value corresponding to the change point is filtered out from the deviation sequence and not treated as an abnormal signal; if the time difference between the deviation change time and any circuit breaker change time exceeds the preset allowable range, the deviation fluctuation is determined to be not caused by a known operation, and the deviation value is retained as an abnormal deviation signal. After the above filtering and retention processes, each mode obtains a corrected effective deviation sequence.

[0073] In terms of constructing continuous state deviation vectors, the system uniformly encodes and splices the effective deviation data after mode correction according to mode dimensions. Specifically, the voltage deviation component and frequency deviation component of the power grid mode, the temperature deviation component of the equipment mode, the load deviation component of the load mode, and the environmental parameter deviation component of the environmental mode are arranged into a multi-dimensional vector in a preset fixed order. Each dimension corresponds to the deviation value of the corresponding mode at the current time. This multi-dimensional vector is the continuous state deviation vector, and its number of dimensions is equal to the sum of the deviation components of each mode. Each component in the vector is continuously updated with time step, forming a continuous state deviation vector sequence that evolves with time.

[0074] Causal weighting module: Based on historical operating baseline data, causal correlation analysis is performed between each mode to obtain the causal influence coefficient between each mode. The dynamic weight of each mode is determined according to the causal influence coefficient. The continuous state deviation vector is weighted to obtain the weighted state deviation vector.

[0075] In this embodiment, the causal weighting module performs causal correlation analysis between each mode based on historical operating baseline data to obtain the causal influence coefficient between each mode, specifically including:

[0076] Based on historical operational baseline data corresponding to each mode, a cross-domain coupling correlation graph is constructed with each mode as a node and the causal influence path between modes as an edge. Specifically, the correlation of heat conduction is calculated using historical time series of environmental meteorological data and equipment status data; the correlation of power-frequency response is calculated using historical time series of power generation data and grid data; and the voltage-load sensitivity is calculated using historical time series of load data and grid data. Furthermore, the correlation of supply and demand balance between power generation data and load data, and the correlation of output response between environmental meteorological data and power generation data are calculated. The correlation indicators obtained from the above calculations are normalized and used as the causal influence coefficients between each mode pair, which are then filled into the corresponding edge weights of the cross-domain coupling correlation graph.

[0077] The process of determining the dynamic weights of each mode based on the causal influence coefficient and weighting the continuous state deviation vectors to obtain a weighted state deviation vector specifically includes:

[0078] For each mode, the sum of the causal influence coefficients between the mode and the other modes is used as the comprehensive influence intensity of the mode. The comprehensive influence intensity of each mode is normalized to generate a dynamic weight vector for each mode in the early warning judgment. A dynamic weight matrix is ​​constructed with the dynamic weight vector as the diagonal element.

[0079] The deviation components corresponding to each mode in the continuous state deviation vector are weighted and summed with the dynamic weight matrix to calculate the weighted state deviation vector.

[0080] It should be noted that the causal weighting module implements causal correlation analysis and dynamic weighting processing between different modalities in the following ways:

[0081] Based on historical operational baseline data corresponding to each mode, the system constructs a cross-domain coupling correlation graph with five modes—generation, grid, load, equipment, and environment—as nodes and causal influence paths between modes as directed edges. Regarding the calculation of graph edge weights, for the correlation between environmental meteorological data and equipment status data, the system takes historical environmental temperature sequences and historical unit temperature sequences, calculating the Pearson correlation coefficient within the same time window as the heat conduction correlation. For the correlation between generation-side data and grid-side data, the system takes historical power generation sequences and historical frequency sequences, calculating the linear regression slope between power change and frequency change as the power-frequency response correlation. For the correlation between load-side data and grid-side data, the system takes historical load power sequences and historical voltage sequences, calculating the sensitivity coefficient between load change rate and voltage change rate as the voltage-load sensitivity. Simultaneously, the system calculates the supply-demand balance correlation between generation-side data and load-side data, and the output response correlation between environmental meteorological data and generation-side data.

[0082] The system performs min-max normalization on each of the above-calculated correlation indicators, mapping them to the interval between 0 and 1. The calculation formula is as follows: Where C is the original calculated index value, and These are the maximum and minimum values ​​of the indicator in historical data; the normalized values ​​are... As the causal influence coefficient between corresponding mode pairs, it is filled into the corresponding edge weight of the cross-domain coupling correlation graph.

[0083] Regarding dynamic weight determination, the system calculates the sum of all causal influence coefficients associated with each mode in the cross-domain coupling correlation graph (including outgoing edge weights where the mode influences other modes as a source and incoming edge weights where the mode is an affected node) as the comprehensive influence intensity of that mode. For example, the comprehensive influence intensity of the power generation mode is equal to the sum of its causal influence coefficients with the grid mode, load mode, and environmental mode. Then, the comprehensive influence intensities of each of the five modes are normalized to generate dynamic weight vectors for each mode in early warning determination. The calculation formula is: ,in Let be the combined influence intensity of the i-th mode. The system constructs a 5×5 dynamic weight matrix M using this dynamic weight vector, where the diagonal elements of the matrix are, in order: to Off-diagonal elements are set to zero.

[0084] In terms of weighted processing, the system will deviate from the continuous state vector. The modal deviation components are weighted and fused with the dynamic weight matrix to output a weighted state deviation vector V, calculated as follows: That is, each component of the weighted state deviation vector is equal to the corresponding original deviation component multiplied by its respective weight. For example, the voltage deviation component is corrected to... Frequency deviation component correction This process continues in the same manner, thereby achieving differentiated weighting of the deviation degree of each mode based on historical causal relationships, ensuring that key modes that have a significant impact on system stability have a higher weight in early warning judgment.

[0085] Consistency determination module: Constructs a normal correlation matrix between modalities based on historical operating benchmark data, calculates the consistency score between the weighted state deviation vector and the normal correlation matrix, and determines whether the operational consistency between modalities is abnormal.

[0086] In this embodiment, the consistency determination module constructs a normal correlation matrix between modalities based on historical operating benchmark data, calculates the consistency score between the weighted state deviation vector and the normal correlation matrix, and determines whether the operational consistency between modalities is abnormal, specifically including:

[0087] Based on the historical operating baseline data corresponding to each mode, the power-frequency response characteristics between the power generation mode and the grid mode, the voltage-load sensitivity characteristics between the grid mode and the load mode, the temperature-power correlation characteristics between the power generation mode and the equipment mode, the heat conduction response characteristics between the environmental mode and the equipment mode, and the output response characteristics between the environmental mode and the power generation mode are extracted respectively, and a 5×5 normal correlation matrix is ​​constructed. Among them, the statistical mean and covariance of the historical time series are used to characterize the correlation characteristics of each mode pair. The diagonal elements of the matrix are the operating stability index of each mode itself, and the off-diagonal elements are the correlation strength index between the mode pairs.

[0088] Based on the covariance matrix of the normal correlation matrix, the Mahalanobis distance algorithm is used to calculate the consistency distance between the weighted state deviation vector and the center of the normal baseline distribution. The consistency distance is then mapped to a consistency score in the range of 0 to 1 using an exponential decay function, where the closer the consistency score is to 1, the higher the operational consistency.

[0089] The consistency score is compared with a preset consistency judgment threshold. If the consistency score is lower than the preset consistency judgment threshold, the intermodal operation consistency is determined to be abnormal and an early warning message is triggered. The early warning message is an AI prediction early warning message generated based on the consistency judgment result. If the consistency score is not lower than the preset consistency judgment threshold, the intermodal operation consistency is determined to be normal.

[0090] It should be specifically explained that, firstly, a 5×5 normal correlation matrix is ​​constructed based on the historical operating baseline data corresponding to each mode. This matrix is ​​essentially a covariance matrix of multidimensional variables. Specifically, for the correlation between the generation mode and the grid mode, the system extracts the covariance values ​​of the historical power generation sequence and the historical frequency sequence within the same time window and fills them into the positions of the 1st row and 2nd column and the 2nd row and 1st column of the matrix. For the correlation between the grid mode and the load mode, the generation mode and the equipment mode, the environmental mode and the equipment mode, and the environmental mode and the generation mode, and so on, the calculated covariance values ​​are filled into the corresponding off-diagonal positions. For the diagonal elements of the matrix (i.e., from the 1st row and 1st column to the 5th row and 5th column), the system calculates the variance of the historical data of each mode as an indicator of the operational stability of that mode. The smaller the variance value, the more stable the operation of that mode. For the remaining off-diagonal positions, the system calculates the actual covariance value of the corresponding mode pair based on the historical operating data to ensure the integrity and positive definiteness of the covariance matrix. Based on this, the system uses the normal correlation matrix as the covariance matrix. The weighted state deviation vector at the current time is calculated using the Mahalanobis distance algorithm. Center of normal baseline distribution The consistency distance L between them is calculated using the following formula: ,in This is the weighted state deviation vector. This represents the deviation vector of each historical mode from the mean (i.e., the center of the normal baseline distribution). This is the inverse of the covariance matrix. For example, when the weighted state deviation vector is [0.5, 0.3, 0.8, 0.2, 0.4] and the historical baseline distribution center vector is [0.1, 0.1, 0.1, 0.1, 0.1], the Mahalanobis distance can be calculated using the above formula. Subsequently, the system maps the consistency distance to a consistency score Q in the interval of 0 to 1 using an exponential decay function, calculated as follows: Where e is a natural constant. When the Mahalanobis distance L is 0, the consistency score Q is 1, indicating the highest operational consistency; the larger the Mahalanobis distance L, the closer the consistency score Q is to 0. Finally, the system compares the consistency score with a preset consistency judgment threshold, which can be set to 0.7 or 0.8 based on the statistical distribution of consistency scores during historical normal operation. If the consistency score is lower than the threshold, the system determines that the operational consistency between modalities is abnormal and triggers AI prediction and warning information generated based on the consistency judgment result; if the consistency score is not lower than the threshold, the system determines that the operational consistency between modalities is normal.

[0091] State trajectory prediction module: Based on the determination results of intermodal operational consistency, a state trajectory prediction model is constructed based on historical operational benchmark data. The state trajectory prediction model is used to perform time-series extrapolation on the weighted state deviation vector to obtain the state evolution trajectory for future periods.

[0092] In this embodiment, the state trajectory prediction module constructs a state trajectory prediction model based on historical operating baseline data, specifically including:

[0093] Based on the consistency score output by the consistency determination module, select the type of state trajectory prediction model:

[0094] When the consistency score is lower than the preset consistency judgment threshold, a nonlinear trajectory prediction model based on long short-term memory network is constructed; when the consistency score is not lower than the preset consistency judgment threshold, a linear trajectory prediction model based on linear regression is constructed.

[0095] Using historical time series sequences of each mode in historical operating benchmark data as training samples, the historical time series sequences are processed by a time series decomposition algorithm to separate periodic components, trend components and residual components. The periodic components, trend components and residual components are then input into the state trajectory prediction model for training to obtain the trained state trajectory prediction model.

[0096] The process of using a state trajectory prediction model to perform time-series extrapolation on the weighted state deviation vector to obtain the state evolution trajectory for future periods specifically includes:

[0097] The weighted state deviation vector is used as the current input state of the state trajectory prediction model. The model is iteratively predicted step by step according to a preset time step to obtain the initial predicted trajectory sequence.

[0098] The covariance matrix of the normal correlation matrix is ​​called, and the Mahalanobis distance of the predicted vector at each time step in the initial predicted trajectory sequence relative to the normal reference center is calculated using the covariance matrix to obtain the prediction deviation at the current time step; the prediction deviation is compared with the preset deviation judgment threshold.

[0099] If the prediction deviation exceeds the preset deviation judgment threshold, the prediction vector is projected and corrected based on the eigenvector of the normal correlation matrix, and the corrected vector is used as the iterative input for the next time period; if the prediction deviation does not exceed the preset deviation judgment threshold, the prediction vector is directly used as the iterative input for the next time period.

[0100] After the iteration is completed, the state evolution trajectory for multiple future time periods is output; the state evolution trajectory is a numerical sequence of state deviation for each mode in each future time period, and the numerical points in the numerical sequence of state deviation that exceed the preset safety boundary are marked as abnormal warning indicators.

[0101] It should be specifically noted that the state trajectory prediction module predicts the state evolution trajectory for future time periods in the following ways:

[0102] The system first selects the prediction model type based on the consistency score output by the consistency determination module. Specifically, when the consistency score is lower than a preset threshold (e.g., 0.7), it indicates that the consistency between modes has become abnormal. The system constructs a nonlinear trajectory prediction model based on a Long Short-Term Memory (LSTM) network. This model includes an input layer, two LSTM layers each containing 64 hidden units, and a fully connected output layer, which is used to capture the nonlinear evolution of the deviation vector under abnormal operating conditions. When the consistency score is not lower than the preset threshold (e.g., 0.7), it indicates that the consistency between modes is normal. The system constructs a linear trajectory prediction model based on linear regression, which extrapolates the future state by weighting the historical linear trend of the state deviation vector.

[0103] During the model training phase, the system utilizes historical time-series sequences of five modalities—power generation, power grid, load, equipment, and environment—from historical operational baseline data as training samples. To improve prediction accuracy, the STL time-series decomposition algorithm (i.e., a seasonal-trend decomposition algorithm based on local weighted regression) is employed to process each historical time-series sequence, separating it into periodic components (e.g., the periodic fluctuations in photovoltaic power generation during the day), trend components (e.g., the long-term trend of equipment temperature slowly increasing with the seasons), and residual components (e.g., random noise or sudden disturbances). The system concatenates these three components and uses them as input features for training the state trajectory prediction model, obtaining the trained model parameters.

[0104] During the time-series deduction phase, the system will deviate the weighted state vector at the current moment. As the initial input state to the model, iterative predictions are made time-by-time according to a preset time step (e.g., 15 minutes). The iterative formula is expressed as: ,in This represents the selected prediction model function. Each iteration uses the prediction vector output by the model from the previous time period. As input for the next time period, the initial predicted trajectory sequence is obtained by iterating continuously for, for example, 24 time periods (i.e., 6 hours).

[0105] During the iteration process, the system introduces a correction mechanism to prevent error accumulation. After each time period's prediction is completed, the system calls the covariance matrix corresponding to the normal correlation matrix. Calculate the prediction vector for this period. With normal reference center vector Mahalanobis distance between The calculation formula is: The system uses the calculated Mahalanobis distance as the prediction deviation at the current moment and compares it with a preset deviation threshold. If the prediction deviation exceeds the preset threshold (e.g., a Mahalanobis distance of 2.3 is greater than the threshold of 2.0), the system projects the prediction vector onto the eigenvectors of the covariance matrix. Specifically, the prediction vector is projected onto the normal correlation subspace composed of the first k principal eigenvectors to filter out abnormal drift components. The corrected vector is then used to correct the deviation. The original prediction vector is used as the input for the next time period. If the prediction deviation does not exceed the threshold (e.g., Mahalanobis distance is 1.2), the original prediction vector is used directly as the input for the next time period.

[0106] After all iterations are completed, the system outputs the state evolution trajectory for the next 24 time periods (6 hours in total). This trajectory consists of a sequence of state deviation values ​​for each of the five modes in each time period. The system further applies graded warning labels to this trajectory. Values ​​in the sequence where the deviation value of any mode exceeds a preset safety boundary (e.g., temperature deviation exceeding 5 degrees Celsius or voltage deviation exceeding 7% of the rated value) are marked as red abnormality warning labels. Values ​​where the deviation value is within 80% to 100% of the safety boundary are marked as yellow caution labels, forming a complete state evolution trajectory output with graded warning labels.

[0107] Intervention Decision and Early Warning Module: Based on the intervention strategy library, the module uses a state trajectory prediction model to perform intervention simulation on the state evolution trajectory, obtain the optimal intervention path, and output proactive early warning information.

[0108] In this embodiment, the intervention decision-making and early warning module performs intervention deduction on the state evolution trajectory based on the intervention strategy library, obtains the optimal intervention path, and outputs proactive early warning information, specifically including:

[0109] The system captures abnormal time periods in the state evolution trajectory where the deviation component exceeds a preset safety boundary and the corresponding target mode. Based on a preset anomaly-policy mapping relationship, it calls several candidate intervention strategies that match the target mode from the intervention strategy library. The intervention strategy library includes power adjustment instructions for frequency mode anomalies, reactive power compensation instructions for voltage mode anomalies, and load transfer instructions for equipment mode temperature anomalies.

[0110] An intervention effect evaluation function is constructed, and candidate intervention strategies are substituted into the state trajectory prediction model as input variables for simulation and deduction. The expected adjustment time and required adjustment magnitude of the state evolution trajectory returning to the safety boundary after the implementation of each candidate intervention strategy are calculated.

[0111] A multi-objective optimization algorithm was adopted, with the optimization objectives of minimizing adjustment time and minimizing adjustment amplitude. Candidate intervention strategies were comprehensively scored, and the strategy with the highest score was selected as the optimal intervention path.

[0112] Based on the optimal intervention path, proactive early warning information containing the target equipment identifier, adjustment direction, and adjustment value is generated and pushed to the site monitoring interface.

[0113] It should be specifically noted that the intervention decision-making and early warning module generates the optimal intervention path and outputs proactive early warning information in the following ways:

[0114] The system first scans the state evolution trajectory time by time, capturing abnormal time periods when the deviation component of any mode exceeds the preset safety boundary and the corresponding target mode. For example, when the state evolution trajectory shows that the temperature deviation component of the equipment mode exceeds the safety boundary by 5 degrees Celsius for 5 consecutive time periods from time 3 to time 7, the system marks the equipment mode as the target mode, and records the abnormal start time as time period 3 and the abnormal duration as 5 time periods.

[0115] Subsequently, the system calls several candidate intervention strategies matching the target mode from the intervention strategy library based on the preset anomaly-strategy mapping relationship. Specifically, when the target mode is the frequency mode, the system calls candidate strategies including increasing the energy storage discharge power by 100 kW, reducing the photovoltaic inverter output by 50 kW, and starting the backup diesel generator, etc.; when the target mode is the voltage mode, the system calls candidate strategies including engaging the capacitor bank to compensate for reactive power by 200 kVar, adjusting the tap changer of the on-load tap changer to increase the tap by 2 levels, and engaging the static var compensator, etc.; when the target mode is an abnormal equipment mode temperature, the system calls candidate strategies including transferring the load to the backup transformer, reducing the unit operating load to 80% of the rated power, and starting the auxiliary cooling fan, etc.

[0116] In terms of intervention effect evaluation, the system constructs an intervention effect evaluation function. This function takes the adjustment value of the candidate intervention strategy as the input variable and uses the previously constructed state trajectory prediction model (i.e., AI prediction model) as the simulation engine. The system substitutes the specific parameters of each candidate strategy into the AI ​​prediction model for simulation. For example, after substituting the strategy of "increasing the energy storage discharge power by 100 kilowatts" into the model, the model continuously applies the adjustment amount of this strategy from the current moment. The simulation calculates that the expected adjustment time for the state evolution trajectory to return to within the safety boundary is 4 time periods, i.e., 1 hour, and the required adjustment magnitude is a 15% decrease in the state of charge of the energy storage system. Similarly, the adjustment time and adjustment magnitude of the other candidate strategies are simulated to obtain their respective adjustment times.

[0117] In terms of strategy selection, the system employs a multi-objective decision-making method, with the optimization objectives of minimizing adjustment duration and adjustment magnitude. A comprehensive scoring function is constructed to score all candidate intervention strategies. Specifically, the adjustment duration and adjustment magnitude of each strategy are normalized and then weighted and summed (the weights can be set as 0.6 for adjustment duration and 0.4 for adjustment magnitude). The strategy with the highest score is the optimal intervention path. For example, if the strategy of "reducing the unit's operating load to 80% of its rated power" has a comprehensive score of 0.85, which is higher than the other strategies, it is selected as the optimal intervention path.

[0118] Finally, the system generates proactive early warning information based on the optimal intervention path. This information includes the target equipment identifier (e.g., main transformer No. 2), the adjustment direction (e.g., reducing the load), and the adjustment value (e.g., reducing to 80% of the rated power, i.e., from 1000 kW to 800 kW). The system then pushes this early warning information to the operation and maintenance personnel in real time through the AI ​​assistant interaction window of the station monitoring interface in the form of pop-up windows and audible and visual alarms, thereby realizing a closed loop of proactive intervention decision-making based on AI prediction.

[0119] In conclusion, the above description is only a preferred embodiment of the present invention and is not intended to limit the present invention. Any modifications, equivalent substitutions, improvements, etc., made within the spirit and principles of the present invention should be included within the protection scope of the present invention.

[0120] The above description is merely a specific embodiment of this application, but the scope of protection of this application is not limited thereto. Any variations or substitutions that can be easily conceived by those skilled in the art within the scope of the technology disclosed in this application should be included within the scope of protection of this application. Therefore, the scope of protection of this application should be determined by the scope of the claims.

Claims

1. An AI-based prediction and proactive early warning system for smart energy power plants, characterized in that: include: Data acquisition module: Collects multi-source heterogeneous operation monitoring data of smart energy power stations and performs preprocessing. The multi-source heterogeneous operation monitoring data includes multi-dimensional real-time operation data and historical operation benchmark data. Multimodal analysis module: Based on data type, real-time running data is divided into different modes, and independent analysis models are constructed using historical running benchmark data corresponding to each mode. The output is corrected by combining synchronously acquired remote signaling data, and a continuous state deviation vector is output. Causal weighting module: Based on historical operating baseline data, causal correlation analysis is performed between each mode to obtain the causal influence coefficient between each mode. The dynamic weight of each mode is determined according to the causal influence coefficient. The continuous state deviation vector is weighted to obtain the weighted state deviation vector. Consistency determination module: Constructs a normal correlation matrix between modalities based on historical operating benchmark data, calculates the consistency score between the weighted state deviation vector and the normal correlation matrix, and determines whether the operational consistency between modalities is abnormal; State trajectory prediction module: Based on the determination result of the consistency between modes, a state trajectory prediction model is constructed based on historical operation benchmark data. The state trajectory prediction model is used to perform time-series extrapolation on the weighted state deviation vector to obtain the state evolution trajectory for future periods. Intervention Decision and Early Warning Module: Based on the intervention strategy library, the module uses the state trajectory prediction model to perform intervention simulation on the state evolution trajectory, obtain the optimal intervention path, and output proactive early warning information.

2. The AI ​​prediction and proactive early warning system for smart energy power plants according to claim 1, characterized in that, The data acquisition module collects multi-source heterogeneous operation monitoring data from smart energy power stations and performs preprocessing, including: By deploying IoT sensors, smart meters, and SCADA systems at the power station, and using the IEC104 or Modbus communication protocol, real-time data is collected from the generation side, grid side, load side, equipment status, and environmental meteorological data. The generation side data includes real-time power generation, cumulative power generation, photovoltaic module temperature, irradiance, and energy storage system state of charge. The load side data includes real-time load power. The grid side data includes voltage, frequency, active power, reactive power, current, three-phase voltage, and three-phase current. The equipment status data includes unit temperature, transformer temperature, circuit breaker temperature, energy storage battery temperature, equipment operating status, and circuit breaker status. The environmental meteorological data includes temperature, humidity, and irradiance. The SCADA system synchronously collects telemetry and remote signaling data via the IEC104 protocol; the remote signaling data includes circuit breaker status and protection action signals. Simultaneously extract historical operational baseline data from the historical database that corresponds to the time period of the real-time operational data; The real-time running data is denoised and normalized, and the historical running baseline data is filled with missing values ​​and time aligned.

3. The AI ​​prediction and proactive early warning system for smart energy plants according to claim 2, characterized in that, The multimodal analysis module divides real-time operating data into different modalities based on data type, constructs independent analysis models using historical operating benchmark data corresponding to each modality, and outputs continuous state deviation vectors, specifically including: Based on the data source and business attributes, the power generation data, grid data, load data, equipment status data, and environmental meteorological data are respectively divided into power generation mode, grid mode, load mode, equipment mode, and environmental mode; Using historical operating baseline data corresponding to each mode, time series-based baseline prediction models are constructed as independent analysis models. Among them, the power grid mode model predicts the normal baseline trajectory of voltage and frequency, the equipment mode model predicts the normal baseline trajectory of unit temperature, the load mode model predicts the normal baseline trajectory of real-time load power, the power generation mode model predicts the normal baseline trajectory of unit output, and the environmental mode model predicts the normal baseline trajectory of environmental parameters. Calculate the real-time deviation between the power generation data, grid data, load data, equipment status data, and environmental meteorological data and their respective normal baseline trajectories; The circuit breaker change time in the remote signaling data is acquired synchronously, and the sudden change time of the real-time deviation is compared with the circuit breaker change time. If the sudden change time and the circuit breaker change time are synchronized, it is determined to be a normal operation response and filtered out. If the timing is not synchronized, the deviation fluctuation is retained as an abnormal deviation signal. The filtered and retained deviation data are integrated to obtain the effective state deviation data. The effective state deviation data is encoded according to each modal dimension to construct a continuous state deviation vector containing voltage deviation component, frequency deviation component, temperature deviation component, load deviation component and environmental deviation component.

4. The AI ​​prediction and proactive early warning system for smart energy plants according to claim 3, characterized in that, The causal weighting module performs causal correlation analysis on each mode based on historical operating baseline data to obtain the causal influence coefficients between each mode, specifically including: Based on historical operational baseline data corresponding to each mode, a cross-domain coupling correlation graph is constructed with each mode as a node and the causal influence path between modes as an edge. Specifically, the correlation of heat conduction is calculated using historical time series of environmental meteorological data and equipment status data; the correlation of power-frequency response is calculated using historical time series of power generation data and grid data; and the voltage-load sensitivity is calculated using historical time series of load data and grid data. Furthermore, the correlation of supply and demand balance between power generation data and load data, and the correlation of output response between environmental meteorological data and power generation data are calculated. The correlation indicators obtained from the above calculations are normalized and used as the causal influence coefficients between each mode pair, which are then filled into the corresponding edge weights of the cross-domain coupling correlation graph. The process of determining the dynamic weights of each mode based on the causal influence coefficient and weighting the continuous state deviation vectors to obtain a weighted state deviation vector specifically includes: For each mode, the sum of the causal influence coefficients between the mode and the other modes is used as the comprehensive influence intensity of the mode. The comprehensive influence intensity of each mode is normalized to generate a dynamic weight vector for each mode in the early warning judgment. A dynamic weight matrix is ​​constructed with the dynamic weight vector as the diagonal element. The deviation components corresponding to each mode in the continuous state deviation vector are weighted and summed with the dynamic weight matrix to calculate the weighted state deviation vector.

5. The AI ​​prediction and proactive early warning system for smart energy plants according to claim 4, characterized in that, The consistency determination module constructs a normal correlation matrix between modalities based on historical operational benchmark data, calculates the consistency score between the weighted state deviation vector and the normal correlation matrix, and determines whether the operational consistency between modalities is abnormal. Specifically, this includes: Based on the historical operating baseline data corresponding to each mode, the power-frequency response characteristics between the power generation mode and the grid mode, the voltage-load sensitivity characteristics between the grid mode and the load mode, the temperature-power correlation characteristics between the power generation mode and the equipment mode, the heat conduction response characteristics between the environmental mode and the equipment mode, and the output response characteristics between the environmental mode and the power generation mode are extracted respectively, and a 5×5 normal correlation matrix is ​​constructed. Among them, the statistical mean and covariance of the historical time series are used to characterize the correlation characteristics of each mode pair. The diagonal elements of the matrix are the operating stability index of each mode itself, and the off-diagonal elements are the correlation strength index between the mode pairs. Based on the covariance matrix of the normal correlation matrix, the Mahalanobis distance algorithm is used to calculate the consistency distance between the weighted state deviation vector and the center of the normal baseline distribution. The consistency distance is then mapped to a consistency score in the range of 0 to 1 using an exponential decay function, where the closer the consistency score is to 1, the higher the operational consistency. The consistency score is compared with a preset consistency judgment threshold. If the consistency score is lower than the preset consistency judgment threshold, the intermodal operation consistency is determined to be abnormal and an early warning message is triggered. The early warning message is an AI prediction early warning message generated based on the consistency judgment result. If the consistency score is not lower than the preset consistency judgment threshold, the intermodal operation consistency is determined to be normal.

6. The AI ​​prediction and proactive early warning system for smart energy plants according to claim 5, characterized in that, The state trajectory prediction module constructs a state trajectory prediction model based on historical operational baseline data, specifically including: Based on the consistency score output by the consistency determination module, select the type of state trajectory prediction model: When the consistency score is lower than the preset consistency judgment threshold, a nonlinear trajectory prediction model based on long short-term memory network is constructed; when the consistency score is not lower than the preset consistency judgment threshold, a linear trajectory prediction model based on linear regression is constructed. Using historical time series sequences of each mode in historical operating benchmark data as training samples, the historical time series sequences are processed by a time series decomposition algorithm to separate periodic components, trend components and residual components. The periodic components, trend components and residual components are then input into the state trajectory prediction model for training to obtain the trained state trajectory prediction model. The process of using a state trajectory prediction model to perform time-series extrapolation on the weighted state deviation vector to obtain the state evolution trajectory for future periods specifically includes: The weighted state deviation vector is used as the current input state of the state trajectory prediction model. The model is iteratively predicted step by step according to a preset time step to obtain the initial predicted trajectory sequence. The covariance matrix of the normal correlation matrix is ​​called, and the Mahalanobis distance of the predicted vector at each time step in the initial predicted trajectory sequence relative to the normal reference center is calculated using the covariance matrix to obtain the prediction deviation at the current time step; the prediction deviation is compared with the preset deviation judgment threshold. If the prediction deviation exceeds the preset deviation judgment threshold, the prediction vector is projected and corrected based on the eigenvector of the normal correlation matrix, and the corrected vector is used as the iterative input for the next time period; if the prediction deviation does not exceed the preset deviation judgment threshold, the prediction vector is directly used as the iterative input for the next time period. After the iteration is completed, the state evolution trajectory for multiple future time periods is output; the state evolution trajectory is a numerical sequence of state deviation for each mode in each future time period, and the numerical points in the numerical sequence of state deviation that exceed the preset safety boundary are marked as abnormal warning indicators.

7. The AI ​​prediction and proactive early warning system for smart energy plants according to claim 6, characterized in that, The intervention decision-making and early warning module performs intervention simulation on the state evolution trajectory based on the intervention strategy library, obtains the optimal intervention path, and outputs proactive early warning information, specifically including: The system captures abnormal time periods in the state evolution trajectory where the deviation component exceeds a preset safety boundary and the corresponding target mode. Based on a preset anomaly-policy mapping relationship, it calls several candidate intervention strategies that match the target mode from the intervention strategy library. The intervention strategy library includes power adjustment instructions for frequency mode anomalies, reactive power compensation instructions for voltage mode anomalies, and load transfer instructions for equipment mode temperature anomalies. An intervention effect evaluation function is constructed, and candidate intervention strategies are substituted into the state trajectory prediction model as input variables for simulation and deduction. The expected adjustment time and required adjustment magnitude of the state evolution trajectory returning to the safety boundary after the implementation of each candidate intervention strategy are calculated. A multi-objective optimization algorithm was adopted, with the optimization objectives of minimizing adjustment time and minimizing adjustment amplitude. Candidate intervention strategies were comprehensively scored, and the strategy with the highest score was selected as the optimal intervention path. Based on the optimal intervention path, proactive early warning information containing the target equipment identifier, adjustment direction, and adjustment value is generated and pushed to the site monitoring interface.

Images