Transformer state prediction and adaptive control method based on internet of things data collection

Abstract

The application discloses a transformer state prediction and adaptive control method based on Internet of Things data collection, comprising the following steps: step one, collecting transformer multi-source operation state data and generating multi-source state data flow; step two, extracting temperature rise change characteristics, load fluctuation characteristics, partial discharge pulse density characteristics and vibration frequency band energy characteristics, and generating state event fragments; step three, constructing a state event correlation graph; step four, generating different types of event sequences and configuring prediction structure parameters; step five, inputting the event sequences into an improved TiDE model, performing graph embedding coding, event gating processing and time span reconstruction processing, and outputting a state evolution path; step six, determining a corresponding risk level; and step seven, generating a corresponding adjustment instruction and updating the prediction structure parameters. The application realizes risk propagation prediction and adaptive control of the transformer operation state.

Description

Technical Field

[0001] This invention relates to the fields of artificial intelligence and intelligent operation and maintenance technology of power equipment, and in particular to a transformer condition prediction and adaptive control method based on Internet of Things data acquisition. Background Technology

[0002] With the rapid development of intelligent power grid equipment and IoT monitoring systems for substations, transformer operating status prediction and adaptive control have gradually become an important research direction in the field of intelligent operation and maintenance of power equipment. For monitoring transformer operating status and fault early warning, existing technologies mainly use single operating parameter analysis or traditional time-series prediction models to monitor and predict status data such as oil temperature, current, partial discharge, and gas concentration. However, the following problems commonly exist in actual operating scenarios: Traditional condition monitoring methods mostly rely on threshold judgments or short-term time-series predictions based on independent sensor data, making it difficult to effectively characterize the correlation and propagation relationships between different types of condition data. This results in insufficient ability to identify complex risk evolution processes such as partial discharge propagation, thermal anomaly propagation, and load fluctuation coupling. Existing time-series prediction models typically use fixed time windows or unified prediction paths to process multi-source condition data, which cannot adapt to the differences in propagation cycles between slowly changing and abruptly changing condition events, easily leading to deviations between prediction results and actual risk propagation processes. Existing methods lack joint modeling mechanisms for temporal correlations, co-occurrence relationships, and risk propagation relationships among condition events, making it difficult to establish risk propagation structures under complex operating conditions. Furthermore, current AI prediction methods in transformer condition prediction are mostly limited to single-path prediction, without introducing event propagation competition mechanisms and dynamic prediction structure reconstruction mechanisms. This makes it impossible to achieve dynamic competition and propagation weight adjustment between different risk propagation paths, limiting the ability to predict state evolution and adaptive control under complex operating scenarios.

[0003] Therefore, how to provide a transformer state prediction and adaptive control method based on IoT data acquisition is a problem that urgently needs to be solved by those skilled in the art. Summary of the Invention

[0004] One objective of this invention is to propose a transformer state prediction and adaptive control method based on IoT data acquisition. This invention constructs a state event association graph and combines it with an improved TiDE model. It utilizes event decoupling coding, dynamic event gating, time-series span reconstruction, and propagation path competition mechanisms to achieve dynamic modeling of risk propagation relationships among multi-source state events. At the same time, it establishes differentiated prediction paths based on different event types to improve the state evolution prediction capability for slowly changing state events and abruptly changing state events, thereby realizing risk propagation prediction and adaptive control of transformer operating state.

[0005] The transformer state prediction and adaptive control method based on Internet of Things (IoT) data acquisition according to an embodiment of the present invention includes the following steps: Step 1: Collect multi-source operating status data of the transformer during operation and generate a multi-source status data stream; Step 2: Perform sliding window segmentation on the multi-source state data stream to extract temperature rise change features, load fluctuation features, partial discharge pulse density features, vibration frequency band energy features, gas growth features, and communication delay fluctuation features, and generate state event fragments. Step 3: Based on the temporal correlation, state co-occurrence relationship, and risk propagation relationship among the state event fragments, construct a state event correlation graph; Step 4: Based on the state event association diagram, perform event decoupling processing on the state event segments to generate thermal event sequences, discharge event sequences, vibration event sequences, gas event sequences, and load event sequences, and configure predictive structural parameters based on the risk propagation state corresponding to each event sequence; Step 5: Input each event sequence and the state-event association graph into the improved TiDE model, perform graph embedding encoding on the state-event association graph, perform event gating and time span reconstruction on each event sequence, and output the state evolution path; Step Six: Determine the corresponding risk level based on the state evolution path, and establish a mapping relationship between state event fragments and risk levels; Step 7: Generate cooling adjustment instructions, load adjustment instructions, and sampling adjustment instructions based on the correspondence between risk level and control level, and update the status event association diagram and predicted structural parameters based on the controlled transformer operation data.

[0006] Optionally, step one specifically includes: Temperature acquisition nodes, current acquisition nodes, partial discharge acquisition nodes, vibration acquisition nodes, gas acquisition nodes, and environmental monitoring nodes are deployed in the transformer body, heat dissipation structure, winding structure, bushing structure, and cooling structure respectively to establish a multi-node Internet of Things acquisition network. The multi-node Internet of Things (IoT) acquisition network is used to collect data on oil temperature changes, winding temperature changes, load current fluctuations, partial discharge pulses, vibration spectrum data, dissolved gas concentration changes in oil, ambient temperature and humidity changes, and cooling device operating status data according to a unified time reference, and generates original status data frames in the order of acquisition time. Calculate the time interval between adjacent data frames based on the data acquisition time corresponding to each acquisition node, mark the data frames whose time interval exceeds the current sampling interval of the corresponding acquisition node with a delay identifier, and record the corresponding node number and delay duration. The data integrity rate is calculated based on the ratio between the number of received data packets and the number of sent data packets. Data frames with missing segments are marked with missing identifiers, and the time span and node position corresponding to the missing segments are recorded. The original state data frames are aligned to a unified time axis, and different types of state data are resampled and aligned according to the timestamp order to generate multidimensional time series data blocks. Time window association processing and corresponding position matching processing are performed on different types of state data in each multidimensional time series data block to establish the time series correlation between oil temperature change data, load current fluctuation data, partial discharge pulse data, vibration spectrum data and gas concentration change data, and generate multi-source state data stream.

[0007] Optionally, step two specifically involves: The multi-source state data stream is processed by sliding window segmentation in chronological order, and a window state segment corresponding to the time range is established based on the data start time and end time corresponding to each time window. Based on the oil temperature change data and winding temperature change data in each window state segment, the temperature change difference between adjacent sampling times is calculated, the temperature change difference in the same direction is accumulated, and the duration of the corresponding change is statistically analyzed to generate temperature rise change characteristics. Based on the load current fluctuation data in each window state segment, the current change amplitude between adjacent sampling times is calculated, and the number of fluctuations, fluctuation duration, and fluctuation segment location corresponding to the current change amplitude are statistically analyzed to generate load fluctuation characteristics. Based on the partial discharge pulse data in each window state segment, the number of pulses in each time window is counted, and the partial discharge pulse data is divided into segments according to the pulse amplitude range to generate partial discharge pulse density characteristics. The frequency band is divided based on the vibration spectrum data in each window state segment. The spectral amplitude in each frequency band is accumulated, and the spectral change is statistically analyzed to generate vibration frequency band energy characteristics. Based on the data on the change in dissolved gas concentration in oil in each window state segment, the change in gas concentration between adjacent sampling times is calculated, and the growth segment and growth duration corresponding to the change in gas concentration are statistically analyzed to generate gas growth characteristics. Based on the communication delay data in each window state segment, the delay change between adjacent sampling times is calculated, and the fluctuation segment and fluctuation duration corresponding to the delay change are statistically analyzed to generate communication delay fluctuation characteristics. Based on the time overlap and change synchronization relationships, the temperature rise change characteristics, load fluctuation characteristics, partial discharge pulse density characteristics, vibration frequency band energy characteristics, gas growth characteristics, and communication delay fluctuation characteristics are correlated and processed, and state event fragments are generated in the order of time windows.

[0008] Optionally, step three specifically includes: Arrange the state event segments in chronological order of their occurrence. Establish a time-related edge when the time interval between adjacent state event segments is less than the duration of the corresponding event. Matching processing is performed on the overlapping segments and segments with consistent change directions in the temperature rise change characteristics, load fluctuation characteristics, partial discharge pulse density characteristics, vibration frequency band energy characteristics, and gas growth characteristics corresponding to each state event segment to establish state co-occurrence association edges; Divide the state event segments with continuous and consistent change direction and continuous temporal sequence into risk propagation segments, and establish risk propagation association edges according to the transmission order between risk propagation segments; The initial state event association graph is constructed by using each state event fragment as a graph node and the time association edge, state co-occurrence association edge, and risk propagation association edge as graph edges. Node attribute data is generated based on the event type, event duration, event change intensity, and the number of times the event occurs per unit time for each graph node. Edge attribute data is generated based on the time interval, duration of synchronous changes, and duration of risk propagation corresponding to each graph edge. Based on the connection relationships between graph nodes and graph edges, graph structure assembly processing is performed on each graph node, each graph edge, node attribute data, and edge attribute data to generate a state event association graph.

[0009] Optionally, step four specifically involves: Read the event type identifier, event duration, event change intensity, and risk propagation direction data corresponding to each graph node in the state event association graph, and classify each state event fragment according to the event type; State event segments corresponding to temperature rise change characteristics are divided into thermal event segments, state event segments corresponding to partial discharge pulse density characteristics are divided into discharge event segments, state event segments corresponding to vibration frequency band energy characteristics are divided into vibration event segments, state event segments corresponding to gas growth characteristics are divided into gas event segments, and state event segments corresponding to load fluctuation characteristics are divided into load event segments. Based on the temporal order and risk propagation direction among the various state event segments, the thermal event segments, discharge event segments, vibration event segments, gas event segments, and load event segments are sequentially arranged to generate thermal event sequences, discharge event sequences, vibration event sequences, gas event sequences, and load event sequences. The duration, intensity of change, risk propagation span, and level of event propagation in each event sequence are statistically analyzed, and risk propagation status data for the corresponding event sequence is established. Based on the risk propagation status data corresponding to each event sequence, calculate the change in event propagation span and the change in event duration, and establish time span parameters according to the segment distribution corresponding to the change. Based on the risk propagation status data corresponding to each event sequence, the number of times the event propagation direction changes and the number of times the event propagation level changes are statistically analyzed, and event gating parameters are established according to the statistical results; The prediction structure parameters for the corresponding event sequence are configured based on the time span parameter and the event gating parameter.

[0010] Optionally, step five specifically includes: Input each event sequence and the state event association graph into the improved TiDE model; The improved TiDE model includes an event state graph embedding module, an event decoupling encoding module, a dynamic event gating module, a temporal span reconstruction module, a propagation path competition module, and a risk evolution decoding module. The event state graph embedding module takes the graph nodes, graph edges, node attribute data, and edge attribute data in the state event association graph as input, performs graph embedding encoding processing on the node attribute data and edge attribute data, and generates a state propagation embedding vector. The event decoupling encoding module takes thermal event sequence, discharge event sequence, vibration event sequence, gas event sequence and load event sequence as input, establishes corresponding event encoding channels, performs independent mapping processing on each event sequence, generates event encoding vectors, and concatenates the state propagation embedding vector with each event encoding vector according to their position correspondence to generate a fused event feature vector. The dynamic event gating module takes the fused event feature vector and event gating parameters as input, establishes event gating segments based on the number of changes in the event propagation direction and the number of changes in the event propagation level, performs gating weight allocation processing on the corresponding event channels in the fused event feature vector, and generates dynamic gating event vectors. The time-series span reconstruction module takes the dynamic gating event vector and time span parameter as input, divides the interval between consecutive events in each event sequence into segments, establishes long-span event segments and short-span event segments, and establishes prediction paths for corresponding event types based on the time span parameter. The propagation path competition module takes the event propagation span, event propagation level, event change intensity, and dynamic gating event vector corresponding to each predicted path as input, counts the number of event propagation segments, the number of propagation direction changes, the cumulative value of event change intensity, and the cumulative value of event propagation level in each predicted path, calculates the propagation competition value of the corresponding predicted path, and generates a competitive propagation path based on the propagation competition value. The risk evolution decoding module takes the competition propagation path as input, performs time-series decoding processing on the event propagation segment in each competition propagation path, generates risk evolution sub-paths, and performs weighted merging of each risk evolution sub-path to output the state evolution path.

[0011] Optionally, step six specifically includes: Read the event propagation direction, event propagation span, event propagation level, and event change intensity in the state evolution path, and establish risk propagation segments according to the event propagation order; The duration, spread, and intensity of event changes in each risk propagation segment are statistically analyzed, and a risk evolution sequence is established in chronological order. The cumulative values ​​of event change intensity, event propagation span, and event propagation level in each risk evolution sequence are weighted and summed to calculate the corresponding risk propagation value. The corresponding numerical ranges are divided according to the magnitude of the risk transmission value, and a corresponding risk level is generated based on each numerical range. The frequency and duration of each state event segment in different risk levels are statistically analyzed, and consecutive state event segments under the corresponding risk levels are divided into risk-related segments. A table of correspondence between status event segments and risk levels is generated based on the intensity of event changes, the span of event propagation, and the duration of events in each risk-related segment; The risk levels are arranged in chronological order of risk transmission to generate a risk level sequence.

[0012] Optionally, step seven specifically includes: Read the risk level, risk propagation span, event propagation level, and event duration from the risk level sequence, and establish corresponding control levels based on the risk level; The risk propagation span, event propagation level, and event duration corresponding to each control level are weighted and accumulated to generate the control change amount under the corresponding control level, and the correspondence between control level and control change amount is established. Based on the correspondence between control level and control change, calculate the speed adjustment and running duration adjustment of the cooling device, and generate cooling adjustment command; Based on the correspondence between control levels and control changes, calculate the load current adjustment amplitude and load adjustment duration corresponding to the load current, and generate load adjustment commands. Based on the correspondence between control levels and control changes, calculate the sampling interval change and sampling duration for each acquisition node, and generate sampling adjustment instructions. The control command sequence is established by arranging the cooling adjustment command, load adjustment command, and sampling adjustment command in the order of risk propagation time. The control command sequence is sent to the corresponding cooling device, load regulating device and IoT acquisition node, and the corresponding parameter adjustments are executed.

[0013] The beneficial effects of this invention are: This invention constructs a state-event correlation graph that integrates temperature rise variation characteristics, load fluctuation characteristics, partial discharge pulse density characteristics, vibration frequency band energy characteristics, and gas growth characteristics. Combined with the collaborative design of an improved TiDE model and a risk propagation path competition mechanism, it addresses the problems of weak correlation between multi-source state data of transformers, difficulty in modeling risk propagation processes, and the fixed structure of traditional time series predictions. It proposes a state-event graph construction strategy based on time correlation, state co-occurrence relationships, and risk propagation relationships, significantly improving the ability to express the propagation correlation between multi-source state events and the ability to model risk evolution under complex operating conditions. In the event sequence processing stage, an event decoupling coding mechanism is introduced. Through independent coding processing of thermal event sequences, discharge event sequences, vibration event sequences, gas event sequences, and load event sequences, different types of state events can be represented. The model employs a method to separate and represent the propagation characteristics of state events. It introduces a dynamic event gating mechanism and a temporal span reconstruction mechanism into the prediction structure, dynamically adjusting the gating weights of corresponding event channels based on changes in the direction and level of event propagation. Furthermore, it establishes long-span and short-span prediction paths according to the interval between consecutive events, effectively enhancing the model's ability to predict the differences between slowly changing and abruptly changing state events. In the risk propagation prediction stage, a propagation path competition module is constructed. Through propagation competition values, different prediction paths are ranked competitively and their propagation weights are adjusted to avoid interference from low-correlation propagation paths on the state evolution results. Finally, a risk evolution decoding module is used to perform temporal decoding and path fusion of the competing propagation paths, outputting the corresponding risk level and control level, thus realizing risk propagation prediction, adaptive control, and dynamic structural adjustment of transformer operating states. Attached Figure Description

[0014] The accompanying drawings are provided to further illustrate the invention and form part of the specification. They are used in conjunction with embodiments of the invention to explain the invention and do not constitute a limitation thereof. In the drawings: Figure 1 This is a schematic diagram of the overall process of the transformer state prediction and adaptive control method based on Internet of Things data acquisition proposed in this invention; Figure 2 This is a schematic diagram of the improved TiDE model in the transformer state prediction and adaptive control method based on Internet of Things data acquisition proposed in this invention. Detailed Implementation

[0015] The present invention will now be described in further detail with reference to the accompanying drawings. These drawings are simplified schematic diagrams, illustrating only the basic structure of the invention, and therefore only show the components relevant to the invention.

[0016] refer to Figures 1-2 A transformer state prediction and adaptive control method based on IoT data acquisition includes the following steps: Step 1: Collect multi-source operating status data of the transformer during operation and generate a multi-source status data stream; Step 2: Perform sliding window segmentation on the multi-source state data stream to extract temperature rise change features, load fluctuation features, partial discharge pulse density features, vibration frequency band energy features, gas growth features, and communication delay fluctuation features, and generate state event fragments. Step 3: Based on the temporal correlation, state co-occurrence relationship, and risk propagation relationship among the state event fragments, construct a state event correlation graph; Step 4: Based on the state event association diagram, perform event decoupling processing on the state event segments to generate thermal event sequences, discharge event sequences, vibration event sequences, gas event sequences, and load event sequences, and configure predictive structural parameters based on the risk propagation state corresponding to each event sequence; Step 5: Input each event sequence and the state-event association graph into the improved TiDE model, perform graph embedding encoding on the state-event association graph, perform event gating and time span reconstruction on each event sequence, and output the state evolution path; Step Six: Determine the corresponding risk level based on the state evolution path, and establish a mapping relationship between state event fragments and risk levels; Step 7: Generate cooling adjustment instructions, load adjustment instructions, and sampling adjustment instructions based on the correspondence between risk level and control level, and update the status event association diagram and predicted structural parameters based on the controlled transformer operation data.

[0017] In this embodiment, step one specifically includes: Temperature acquisition nodes, current acquisition nodes, partial discharge acquisition nodes, vibration acquisition nodes, gas acquisition nodes, and environmental monitoring nodes are deployed in the transformer body, heat dissipation structure, winding structure, bushing structure, and cooling structure respectively to establish a multi-node Internet of Things acquisition network. The multi-node Internet of Things (IoT) acquisition network is used to collect data on oil temperature changes, winding temperature changes, load current fluctuations, partial discharge pulses, vibration spectrum data, dissolved gas concentration changes in oil, ambient temperature and humidity changes, and cooling device operating status data according to a unified time reference, and generates original status data frames in the order of acquisition time. Calculate the time interval between adjacent data frames based on the data acquisition time corresponding to each acquisition node, mark the data frames whose time interval exceeds the current sampling interval of the corresponding acquisition node with a delay identifier, and record the corresponding node number and delay duration. The data integrity rate is calculated based on the ratio between the number of received data packets and the number of sent data packets. Data frames with missing segments are marked with missing identifiers, and the time span and node position corresponding to the missing segments are recorded. The original state data frames are aligned to a unified time axis, and different types of state data are resampled and aligned according to the timestamp order to generate multidimensional time series data blocks. Time window association processing and corresponding position matching processing are performed on different types of state data in each multidimensional time series data block to establish the time series correlation between oil temperature change data, load current fluctuation data, partial discharge pulse data, vibration spectrum data and gas concentration change data, and generate multi-source state data stream.

[0018] In this implementation, the multi-node IoT acquisition network uses an edge clock synchronization mechanism to perform unified time reference calibration on each acquisition node. The edge clock synchronization mechanism includes master clock broadcasting, node time offset recording, and time drift correction. Each acquisition node dynamically adjusts the sampling interval according to the rate of change of the corresponding state data. Oil temperature change data and gas concentration change data use a long-period sampling method, while partial discharge pulse data and vibration spectrum data use a short-period sampling method. The original state data frame uses a combination of frame header identifier, node identifier, time identifier, and state data segment to form a data structure. Resampling alignment processing includes time axis interpolation processing, missing segment compensation processing, and asynchronous sampling point position mapping processing. Time window association processing establishes association segments according to the direction, magnitude, and duration of change of each type of state data within the sliding time window. Corresponding position matching processing establishes a multi-dimensional correspondence relationship according to the sampling position of each type of state data in the same time window.

[0019] In this embodiment, step two specifically includes: The multi-source state data stream is processed by sliding window segmentation in chronological order, and a window state segment corresponding to the time range is established based on the data start time and end time corresponding to each time window. Based on the oil temperature change data and winding temperature change data in each window state segment, the temperature change difference between adjacent sampling times is calculated, the temperature change difference in the same direction is accumulated, and the duration of the corresponding change is statistically analyzed to generate temperature rise change characteristics. Based on the load current fluctuation data in each window state segment, the current change amplitude between adjacent sampling times is calculated, and the number of fluctuations, fluctuation duration, and fluctuation segment location corresponding to the current change amplitude are statistically analyzed to generate load fluctuation characteristics. Based on the partial discharge pulse data in each window state segment, the number of pulses in each time window is counted, and the partial discharge pulse data is divided into segments according to the pulse amplitude range to generate partial discharge pulse density characteristics. The frequency band is divided based on the vibration spectrum data in each window state segment. The spectral amplitude in each frequency band is accumulated, and the spectral change is statistically analyzed to generate vibration frequency band energy characteristics. Based on the data on the change in dissolved gas concentration in oil in each window state segment, the change in gas concentration between adjacent sampling times is calculated, and the growth segment and growth duration corresponding to the change in gas concentration are statistically analyzed to generate gas growth characteristics. Based on the communication delay data in each window state segment, the delay change between adjacent sampling times is calculated, and the fluctuation segment and fluctuation duration corresponding to the delay change are statistically analyzed to generate communication delay fluctuation characteristics. Based on the time overlap and change synchronization relationships, the temperature rise change characteristics, load fluctuation characteristics, partial discharge pulse density characteristics, vibration frequency band energy characteristics, gas growth characteristics, and communication delay fluctuation characteristics are correlated and processed, and state event fragments are generated in the order of time windows.

[0020] In this implementation, the sliding window segmentation process adopts a dynamic window advancement method, and the window length is jointly adjusted according to the partial discharge pulse change frequency, load fluctuation frequency, and vibration spectrum change frequency. During the window advancement process, the overlapping segments between adjacent windows are recorded, and the continuity of the state data in the overlapping segments is checked. The frequency band division adopts the spectral energy ascending sorting method to establish the main vibration frequency band segment and the secondary vibration frequency band segment. The partial discharge pulse data establishes the pulse aggregation segment according to the pulse amplitude change direction and pulse interval time. The gas concentration change data establishes the continuous growth segment according to the concentration change rate. The change synchronization relationship is established by the consistency of the change direction of different types of features in the same time window and the overlap of change time. The state event segment uses a combination of time identifier, event type identifier, change intensity identifier, and associated segment identifier to form the event data structure.

[0021] In this embodiment, step three specifically includes: Arrange the state event segments in chronological order of their occurrence. Establish a time-related edge when the time interval between adjacent state event segments is less than the duration of the corresponding event. Matching processing is performed on the overlapping segments and segments with consistent change directions in the temperature rise change characteristics, load fluctuation characteristics, partial discharge pulse density characteristics, vibration frequency band energy characteristics, and gas growth characteristics corresponding to each state event segment to establish state co-occurrence association edges; Divide the state event segments with continuous and consistent change direction and continuous temporal sequence into risk propagation segments, and establish risk propagation association edges according to the transmission order between risk propagation segments; The initial state event association graph is constructed by using each state event fragment as a graph node and the time association edge, state co-occurrence association edge, and risk propagation association edge as graph edges. Node attribute data is generated based on the event type, event duration, event change intensity, and the number of times the event occurs per unit time for each graph node. Edge attribute data is generated based on the time interval, duration of synchronous changes, and duration of risk propagation corresponding to each graph edge. Based on the connection relationships between graph nodes and graph edges, graph structure assembly processing is performed on each graph node, each graph edge, node attribute data, and edge attribute data to generate a state event association graph.

[0022] In this implementation, the state event fragments are structured using event number, event start time, event end time, event type identifier, and event intensity identifier. Temporal association edges use a directed edge structure to record the temporal connection between preceding and subsequent state event fragments. State co-occurrence association edges use a multi-feature synchronous counting method to record the number of times different types of state event fragments occur synchronously within the same time window. Risk propagation association edges use a risk propagation chain structure to record the direction and level of risk propagation. Node attribute data uses a vectorized arrangement to record event duration, event change intensity, and event occurrence density. Edge attribute data uses segment duration, propagation span, and synchronous change ratio to form an edge feature set. The graph structure assembly process uses a joint storage method of adjacency matrix and edge index table to establish the data structure of the state event association graph.

[0023] In this embodiment, step four specifically includes: Read the event type identifier, event duration, event change intensity, and risk propagation direction data corresponding to each graph node in the state event association graph, and classify each state event fragment according to the event type; State event segments corresponding to temperature rise change characteristics are divided into thermal event segments, state event segments corresponding to partial discharge pulse density characteristics are divided into discharge event segments, state event segments corresponding to vibration frequency band energy characteristics are divided into vibration event segments, state event segments corresponding to gas growth characteristics are divided into gas event segments, and state event segments corresponding to load fluctuation characteristics are divided into load event segments. Based on the temporal order and risk propagation direction among the various state event segments, the thermal event segments, discharge event segments, vibration event segments, gas event segments, and load event segments are sequentially arranged to generate thermal event sequences, discharge event sequences, vibration event sequences, gas event sequences, and load event sequences. The duration, intensity of change, risk propagation span, and level of event propagation in each event sequence are statistically analyzed, and risk propagation status data for the corresponding event sequence is established. Based on the risk propagation status data corresponding to each event sequence, calculate the change in event propagation span and the change in event duration, and establish time span parameters according to the segment distribution corresponding to the change. Based on the risk propagation status data corresponding to each event sequence, the number of times the event propagation direction changes and the number of times the event propagation level changes are statistically analyzed, and event gating parameters are established according to the statistical results; The prediction structure parameters for the corresponding event sequence are configured based on the time span parameter and the event gating parameter.

[0024] In this implementation, the event decoupling process uses a combination of an event type index table and a risk propagation path mapping method to classify and store state event fragments. Thermal event sequences record continuous temperature rise segments using a long-span arrangement; discharge event sequences record pulse aggregation segments using a high-frequency, short-span arrangement; vibration event sequences record vibration propagation segments using frequency band change order; gas event sequences record gas diffusion segments using concentration increase order; and load event sequences record load change segments using load fluctuation amplitude order. The risk propagation span is recorded using the number of graph edges between the risk propagation start and end nodes. The event propagation level is divided using the propagation path depth in the state event association graph. The time span parameter is recorded segmented using the interval duration of consecutive events in the corresponding event sequence. The event gating parameter is a set of parameters consisting of the event propagation direction change density and the propagation level change density.

[0025] In this embodiment, step five specifically includes: Input each event sequence and the state event association graph into the improved TiDE model; The improved TiDE model includes an event state graph embedding module, an event decoupling encoding module, a dynamic event gating module, a temporal span reconstruction module, a propagation path competition module, and a risk evolution decoding module. The event state graph embedding module takes the graph nodes, graph edges, node attribute data, and edge attribute data in the state event association graph as input, performs graph embedding encoding processing on the node attribute data and edge attribute data, and generates a state propagation embedding vector. The event decoupling encoding module takes thermal event sequence, discharge event sequence, vibration event sequence, gas event sequence and load event sequence as input, establishes corresponding event encoding channels, performs independent mapping processing on each event sequence, generates event encoding vectors, and concatenates the state propagation embedding vector with each event encoding vector according to their position correspondence to generate a fused event feature vector. The dynamic event gating module takes the fused event feature vector and event gating parameters as input, establishes event gating segments based on the number of changes in the event propagation direction and the number of changes in the event propagation level, performs gating weight allocation processing on the corresponding event channels in the fused event feature vector, and generates dynamic gating event vectors. The time-series span reconstruction module takes the dynamic gating event vector and time span parameter as input, divides the interval between consecutive events in each event sequence into segments, establishes long-span event segments and short-span event segments, and establishes prediction paths for corresponding event types based on the time span parameter. The propagation path competition module takes the event propagation span, event propagation level, event change intensity, and dynamic gating event vector corresponding to each predicted path as input, counts the number of event propagation segments, the number of propagation direction changes, the cumulative value of event change intensity, and the cumulative value of event propagation level in each predicted path, calculates the propagation competition value of the corresponding predicted path, and generates a competitive propagation path based on the propagation competition value. The risk evolution decoding module takes the competition propagation path as input, performs time-series decoding processing on the event propagation segment in each competition propagation path, generates risk evolution sub-paths, and performs weighted merging of each risk evolution sub-path to output the state evolution path.

[0026] In this implementation, the event state graph embedding module uses a joint encoding method of node index table and edge index table to establish the graph structure input of the state event association graph; the event decoupling encoding module uses an independent parameter mapping method to record the propagation characteristics corresponding to thermal events, discharge events, vibration events, gas events, and load events respectively; the dynamic event gating module establishes gating segments according to the change density of event propagation direction and the change density of propagation level, and performs segment-level weight adjustment on the vector arrangement order of different event channels; the time span reconstruction module establishes time span segments according to the interval duration of continuous events, and performs separate arrangement of long span paths and short span paths according to the time span segments; the propagation path competition module establishes a predicted path sorting sequence using propagation competition value, and adjusts the proportion of propagation vectors corresponding to each predicted path according to the sorting sequence; the risk evolution decoding module establishes path connection relationships using event propagation direction and propagation span, and generates a state evolution path sequence according to the connection relationships; The improved TiDE model is built on the temporal encoding and decoding structure of the TiDE model. It retains the TiDE model's processing methods for performing low-dimensional mapping, temporal feature compression, and multi-step prediction decoding on time series data. It also adopts the collaborative processing of the encoding and decoding channels to process event sequence data and perform temporal prediction processing on state event sequences. The improved TiDE model adds an event state graph embedding module, a dynamic event gating module, a temporal span reconstruction module, and a propagation path competition module to the TiDE model. The event state graph embedding module uses the state-event association graph to establish state propagation embedding vectors; the dynamic event gating module adjusts the gating weights of corresponding event channels according to changes in event propagation direction and propagation level; the temporal span reconstruction module establishes long-span prediction paths and short-span prediction paths according to the duration of consecutive event intervals; and the propagation path competition module performs competitive ranking processing on different prediction paths according to the propagation competition value. The improved TiDE model uses a state-event association graph to establish event propagation relationships and adjusts the prediction path structure by combining the event propagation direction, propagation level, and propagation span. Different types of event sequences are independently encoded and the propagation weights corresponding to the prediction paths are adjusted according to the propagation competition value. Long-span prediction paths and short-span prediction paths correspond to slowly changing state events and abruptly changing state events, respectively, improving the correspondence between state evolution paths and actual risk propagation processes.

[0027] In this embodiment, step six specifically includes: Read the event propagation direction, event propagation span, event propagation level, and event change intensity in the state evolution path, and establish risk propagation segments according to the event propagation order; The duration, spread, and intensity of event changes in each risk propagation segment are statistically analyzed, and a risk evolution sequence is established in chronological order. The cumulative values ​​of event change intensity, event propagation span, and event propagation level in each risk evolution sequence are weighted and summed to calculate the corresponding risk propagation value. The corresponding numerical ranges are divided according to the magnitude of the risk transmission value, and a corresponding risk level is generated based on each numerical range. The frequency and duration of each state event segment in different risk levels are statistically analyzed, and consecutive state event segments under the corresponding risk levels are divided into risk-related segments. A table of correspondence between status event segments and risk levels is generated based on the intensity of event changes, the span of event propagation, and the duration of events in each risk-related segment; The risk levels are arranged in chronological order of risk transmission to generate a risk level sequence.

[0028] In this implementation, risk propagation segments are recorded using a continuous propagation chain between the event propagation start node and the propagation termination node; the risk evolution sequence is established using a combination of event propagation time order and propagation level order; the cumulative value of event change intensity is formed by accumulating the event change intensity in each risk propagation segment according to the propagation order; the cumulative value of event propagation span is formed by accumulating the number of graph edges in the risk propagation path; the cumulative value of event propagation level is formed by accumulating the level numbers corresponding to each propagation level; the numerical interval corresponding to the risk propagation value is established using the segment distribution after sorting the risk propagation values; the correspondence table between state event fragments and risk levels uses a mapping data structure formed by combining event type, risk level, propagation span, and duration; and the risk level sequence is arranged and stored using a combined index of timestamp and propagation level.

[0029] In this embodiment, step seven specifically includes: Read the risk level, risk propagation span, event propagation level, and event duration from the risk level sequence, and establish corresponding control levels based on the risk level; The risk propagation span, event propagation level, and event duration corresponding to each control level are weighted and accumulated to generate the control change amount under the corresponding control level, and the correspondence between control level and control change amount is established. Based on the correspondence between control level and control change, calculate the speed adjustment and running duration adjustment of the cooling device, and generate cooling adjustment command; Based on the correspondence between control levels and control changes, calculate the load current adjustment amplitude and load adjustment duration corresponding to the load current, and generate load adjustment commands. Based on the correspondence between control levels and control changes, calculate the sampling interval change and sampling duration for each acquisition node, and generate sampling adjustment instructions. The control command sequence is established by arranging the cooling adjustment command, load adjustment command, and sampling adjustment command in the order of risk propagation time. The control command sequence is sent to the corresponding cooling device, load regulating device and IoT acquisition node, and the corresponding parameter adjustments are executed.

[0030] In this implementation, the control level is established by combining risk level, risk propagation span, and event propagation hierarchy; the control change is formed by segment-weighted accumulation of the change corresponding to the risk propagation span, event duration, and event propagation hierarchy; the cooling adjustment command is formed by combining the fan speed adjustment value and cooling duration to form the cooling control parameter; the load adjustment command is formed by combining the load current adjustment value and load adjustment time to form the load control parameter; the sampling adjustment command is formed by combining the sampling interval change value and sampling duration to form the sampling control parameter; the control command sequence is established by combining the risk propagation time order and the risk propagation hierarchy order; the control command sending process uses the control number, sending time identifier, and execution status identifier to form the command data structure, and executes the parameter adjustment according to the corresponding equipment number.

[0031] Example 1: To verify the feasibility of this invention in practice, it was applied to a scenario of predictive and adaptive control of the operating status of an oil-immersed main transformer in a 220kV smart substation. This substation is located in a high-load industrial park, where the transformer operates in an environment with frequent load fluctuations, large ambient temperature variations, and a high risk of partial discharge. Traditional operation monitoring systems mainly rely on oil temperature threshold alarms and periodic manual inspections to monitor the transformer's operating status. In actual operation, this often results in problems such as delayed identification of partial discharge propagation, difficulty in predicting the propagation of load anomalies, and a lack of correlation analysis between different state parameters. This leads to delays in cooling system adjustments and an inability to dynamically intervene in the risk propagation process in a timely manner.

[0032] In this embodiment, temperature acquisition nodes, current acquisition nodes, partial discharge acquisition nodes, vibration acquisition nodes, and gas acquisition nodes are deployed on the transformer body, winding structure, heat dissipation structure, and bushing structure, respectively. Through an IoT network, real-time data on oil temperature changes, load current fluctuations, partial discharge pulses, vibration spectrum data, and dissolved gas concentration changes in the oil are collected, generating a multi-source state data stream with corresponding time sequence. The system segments the multi-source state data stream using a dynamic sliding window approach, extracting temperature rise change characteristics, load fluctuation characteristics, partial discharge pulse density characteristics, and vibration frequency band energy characteristics. State event segments are generated based on the temporal overlap and synchronization relationships between different state characteristics. Subsequently, graph structure modeling is performed on the temporal correlation, state co-occurrence, and risk propagation relationships between state event segments to form a state event correlation graph.

[0033] During state prediction, thermal event sequences, discharge event sequences, vibration event sequences, gas event sequences, and load event sequences are input into the improved TiDE model. A state propagation embedding vector is established through the event state graph embedding module. Simultaneously, the dynamic event gating module dynamically adjusts the gating weights of different event channels according to the change density of event propagation direction and the change density of propagation level. The temporal span reconstruction module establishes long-span prediction paths and short-span prediction paths according to the duration of continuous event intervals to handle slowly changing state events and abruptly changing state events, respectively. The propagation path competition module sorts the propagation competition values ​​of different prediction paths and dynamically adjusts the proportion of the propagation vector corresponding to each prediction path, thereby generating state evolution paths and corresponding risk levels.

[0034] When the system detects a continuous increase in partial discharge pulse density, accompanied by an increased rate of oil temperature change and abnormal energy accumulation in the vibration frequency band, the system automatically raises the corresponding risk level and generates cooling and load adjustment commands. The cooling system increases the fan speed and cooling duration according to the control commands, while the load adjustment device simultaneously reduces the corresponding load current amplitude. At the same time, the acquisition nodes automatically shorten the partial discharge sampling interval and vibration sampling interval to improve the acquisition accuracy of the abnormal state propagation process. After control is completed, the system continues to acquire post-control operating status data and updates the state event correlation diagram and predicted structural parameters, achieving dynamic closed-loop adjustment of the risk propagation process.

[0035] To verify the practical application effect of the present invention, the present invention was compared with the traditional fixed time window prediction method and the traditional LSTM prediction method. The test was conducted continuously for 30 days, and the accuracy of risk propagation identification, the duration of early warning of anomalies, the false alarm rate, and the load adjustment response time were statistically analyzed. The experimental results are shown in Table 1.

[0036] Table 1. Comparison of the operational performance of different transformer condition prediction methods

[0037] As shown in Table 1, the method of this invention significantly outperforms traditional threshold monitoring and LSTM prediction methods in terms of risk propagation identification accuracy, anomaly warning capability, and control response capability. Specifically, the invention achieves a risk propagation identification accuracy of 96.8%, an improvement of 18.4% compared to traditional threshold monitoring methods and 9.2% compared to LSTM prediction methods; the partial discharge anomaly identification accuracy reaches 97.3%, indicating that this invention can effectively identify the discharge propagation process under complex operating conditions. Regarding early warning capability, the risk propagation early warning time of this invention reaches 47 minutes, significantly higher than the 8 minutes and 21 minutes of traditional methods, indicating that this invention can identify risk evolution trends earlier. Meanwhile, the anomaly false alarm rate of this invention is only 1.7%, and the risk propagation path misjudgment rate is only 2.4%, indicating that the propagation path competition mechanism can effectively reduce the interference of invalid propagation paths on the prediction results. In terms of control response, the cooling adjustment response time and load adjustment response time of this invention are reduced to 3.1 seconds and 2.8 seconds, respectively, enabling faster adaptive control.

[0038] The above description is only a preferred embodiment of the present invention, but the scope of protection of the present invention is not limited thereto. Any equivalent substitutions or modifications made by those skilled in the art within the scope of the technology disclosed in the present invention, based on the technical solution and inventive concept of the present invention, should be covered within the scope of protection of the present invention.

Claims

1. A transformer state prediction and adaptive control method based on Internet of Things data collection, characterized in that, Includes the following steps: Step 1: Collect multi-source operating status data of the transformer during operation and generate a multi-source status data stream; Step 2: Perform sliding window segmentation on the multi-source state data stream to extract temperature rise change features, load fluctuation features, partial discharge pulse density features, vibration frequency band energy features, gas growth features, and communication delay fluctuation features, and generate state event fragments. Step 3: Based on the temporal correlation, state co-occurrence relationship, and risk propagation relationship among the state event fragments, construct a state event correlation graph; Step 4: Based on the state event association diagram, perform event decoupling processing on the state event segments to generate thermal event sequences, discharge event sequences, vibration event sequences, gas event sequences, and load event sequences, and configure predictive structural parameters based on the risk propagation state corresponding to each event sequence; Step 5: Input each event sequence and the state-event association graph into the improved TiDE model, perform graph embedding encoding on the state-event association graph, perform event gating and time span reconstruction on each event sequence, and output the state evolution path; Step Six: Determine the corresponding risk level based on the state evolution path, and establish a mapping relationship between state event fragments and risk levels; Step 7: Generate cooling adjustment instructions, load adjustment instructions, and sampling adjustment instructions based on the correspondence between risk level and control level, and update the status event association diagram and predicted structural parameters based on the controlled transformer operation data.

2. The transformer state prediction and adaptive control method based on Internet of Things data collection according to claim 1, characterized in that, Step one specifically involves: Temperature acquisition nodes, current acquisition nodes, partial discharge acquisition nodes, vibration acquisition nodes, gas acquisition nodes, and environmental monitoring nodes are deployed in the transformer body, heat dissipation structure, winding structure, bushing structure, and cooling structure respectively to establish a multi-node Internet of Things acquisition network. The multi-node Internet of Things (IoT) acquisition network is used to collect data on oil temperature changes, winding temperature changes, load current fluctuations, partial discharge pulses, vibration spectrum data, dissolved gas concentration changes in oil, ambient temperature and humidity changes, and cooling device operating status data according to a unified time reference, and generates original status data frames in the order of acquisition time. Calculate the time interval between adjacent data frames based on the data acquisition time corresponding to each acquisition node, mark the data frames whose time interval exceeds the current sampling interval of the corresponding acquisition node with a delay identifier, and record the corresponding node number and delay duration. The data integrity rate is calculated based on the ratio between the number of received data packets and the number of sent data packets. Data frames with missing segments are marked with missing identifiers, and the time span and node position corresponding to the missing segments are recorded. The original state data frames are aligned to a unified time axis, and different types of state data are resampled and aligned according to the timestamp order to generate multidimensional time series data blocks. Time window association processing and corresponding position matching processing are performed on different types of state data in each multidimensional time series data block to establish the time series correlation between oil temperature change data, load current fluctuation data, partial discharge pulse data, vibration spectrum data and gas concentration change data, and generate multi-source state data stream.

3. The transformer state prediction and adaptive control method based on Internet of Things data collection according to claim 1, characterized in that, Step two specifically involves: The multi-source state data stream is processed by sliding window segmentation in chronological order, and a window state segment corresponding to the time range is established based on the data start time and end time corresponding to each time window. Based on the oil temperature change data and winding temperature change data in each window state segment, the temperature change difference between adjacent sampling times is calculated, the temperature change difference in the same direction is accumulated, and the duration of the corresponding change is statistically analyzed to generate temperature rise change characteristics. Based on the load current fluctuation data in each window state segment, the current change amplitude between adjacent sampling times is calculated, and the number of fluctuations, fluctuation duration, and fluctuation segment location corresponding to the current change amplitude are statistically analyzed to generate load fluctuation characteristics. Based on the partial discharge pulse data in each window state segment, the number of pulses in each time window is counted, and the partial discharge pulse data is divided into segments according to the pulse amplitude range to generate partial discharge pulse density characteristics. The frequency band is divided based on the vibration spectrum data in each window state segment. The spectral amplitude in each frequency band is accumulated, and the spectral change is statistically analyzed to generate vibration frequency band energy characteristics. Based on the data on the change in dissolved gas concentration in oil in each window state segment, the change in gas concentration between adjacent sampling times is calculated, and the growth segment and growth duration corresponding to the change in gas concentration are statistically analyzed to generate gas growth characteristics. Based on the communication delay data in each window state segment, the delay change between adjacent sampling times is calculated, and the fluctuation segment and fluctuation duration corresponding to the delay change are statistically analyzed to generate communication delay fluctuation characteristics. Based on the time overlap and change synchronization relationships, the temperature rise change characteristics, load fluctuation characteristics, partial discharge pulse density characteristics, vibration frequency band energy characteristics, gas growth characteristics, and communication delay fluctuation characteristics are correlated and processed, and state event fragments are generated in the order of time windows.

4. The transformer state prediction and adaptive control method based on Internet of Things data collection of claim 1, characterized in that, Step three specifically involves: Arrange the state event segments in chronological order of their occurrence. Establish a time-related edge when the time interval between adjacent state event segments is less than the duration of the corresponding event. Matching processing is performed on the overlapping segments and segments with consistent change directions in the temperature rise change characteristics, load fluctuation characteristics, partial discharge pulse density characteristics, vibration frequency band energy characteristics, and gas growth characteristics corresponding to each state event segment to establish state co-occurrence association edges; Divide the state event segments with continuous and consistent change direction and continuous temporal sequence into risk propagation segments, and establish risk propagation association edges according to the transmission order between risk propagation segments; The initial state event association graph is constructed by using each state event fragment as a graph node and the time association edge, state co-occurrence association edge, and risk propagation association edge as graph edges. Node attribute data is generated based on the event type, event duration, event change intensity, and the number of times the event occurs per unit time for each graph node. Edge attribute data is generated based on the time interval, duration of synchronous changes, and duration of risk propagation corresponding to each graph edge. Based on the connection relationships between graph nodes and graph edges, graph structure assembly processing is performed on each graph node, each graph edge, node attribute data, and edge attribute data to generate a state event association graph.

5. The transformer state prediction and adaptive control method based on Internet of Things data collection according to claim 1, characterized in that, Step four specifically involves: Read the event type identifier, event duration, event change intensity, and risk propagation direction data corresponding to each graph node in the state event association graph, and classify each state event fragment according to the event type; State event segments corresponding to temperature rise change characteristics are divided into thermal event segments, state event segments corresponding to partial discharge pulse density characteristics are divided into discharge event segments, state event segments corresponding to vibration frequency band energy characteristics are divided into vibration event segments, state event segments corresponding to gas growth characteristics are divided into gas event segments, and state event segments corresponding to load fluctuation characteristics are divided into load event segments. Based on the temporal order and risk propagation direction among the various state event segments, the thermal event segments, discharge event segments, vibration event segments, gas event segments, and load event segments are sequentially arranged to generate thermal event sequences, discharge event sequences, vibration event sequences, gas event sequences, and load event sequences. The duration, intensity of change, risk propagation span, and level of event propagation in each event sequence are statistically analyzed, and risk propagation status data for the corresponding event sequence is established. Based on the risk propagation status data corresponding to each event sequence, calculate the change in event propagation span and the change in event duration, and establish time span parameters according to the segment distribution corresponding to the change. Based on the risk propagation status data corresponding to each event sequence, the number of times the event propagation direction changes and the number of times the event propagation level changes are statistically analyzed, and event gating parameters are established according to the statistical results; The prediction structure parameters for the corresponding event sequence are configured based on the time span parameter and the event gating parameter.

6. The transformer state prediction and adaptive control method based on Internet of Things data collection of claim 1, characterized in that, Step five specifically involves: Input each event sequence and the state event association graph into the improved TiDE model; The improved TiDE model includes an event state graph embedding module, an event decoupling encoding module, a dynamic event gating module, a temporal span reconstruction module, a propagation path competition module, and a risk evolution decoding module. The event state graph embedding module takes the graph nodes, graph edges, node attribute data, and edge attribute data in the state event association graph as input, performs graph embedding encoding processing on the node attribute data and edge attribute data, and generates a state propagation embedding vector. The event decoupling encoding module takes thermal event sequence, discharge event sequence, vibration event sequence, gas event sequence and load event sequence as input, establishes corresponding event encoding channels, performs independent mapping processing on each event sequence, generates event encoding vectors, and concatenates the state propagation embedding vector with each event encoding vector according to their position correspondence to generate a fused event feature vector. The dynamic event gating module takes the fused event feature vector and event gating parameters as input, establishes event gating segments based on the number of changes in the event propagation direction and the number of changes in the event propagation level, performs gating weight allocation processing on the corresponding event channels in the fused event feature vector, and generates dynamic gating event vectors. The time-series span reconstruction module takes the dynamic gating event vector and time span parameter as input, divides the interval between consecutive events in each event sequence into segments, establishes long-span event segments and short-span event segments, and establishes prediction paths for corresponding event types based on the time span parameter. The propagation path competition module takes the event propagation span, event propagation level, event change intensity, and dynamic gating event vector corresponding to each predicted path as input, counts the number of event propagation segments, the number of propagation direction changes, the cumulative value of event change intensity, and the cumulative value of event propagation level in each predicted path, calculates the propagation competition value of the corresponding predicted path, and generates a competitive propagation path based on the propagation competition value. The risk evolution decoding module takes the competition propagation path as input, performs time-series decoding processing on the event propagation segment in each competition propagation path, generates risk evolution sub-paths, and performs weighted merging of each risk evolution sub-path to output the state evolution path.

7. The transformer state prediction and adaptive control method based on Internet of Things data collection according to claim 1, characterized in that, Step six specifically involves: Read the event propagation direction, event propagation span, event propagation level, and event change intensity in the state evolution path, and establish risk propagation segments according to the event propagation order; The duration, spread, and intensity of event changes in each risk propagation segment are statistically analyzed, and a risk evolution sequence is established in chronological order. The cumulative values ​​of event change intensity, event propagation span, and event propagation level in each risk evolution sequence are weighted and summed to calculate the corresponding risk propagation value. The corresponding numerical ranges are divided according to the magnitude of the risk transmission value, and a corresponding risk level is generated based on each numerical range. The frequency and duration of each state event segment in different risk levels are statistically analyzed, and consecutive state event segments under the corresponding risk levels are divided into risk-related segments. A table of correspondence between status event segments and risk levels is generated based on the intensity of event changes, the span of event propagation, and the duration of events in each risk-related segment; The risk levels are arranged in chronological order of risk transmission to generate a risk level sequence.

8. The transformer state prediction and adaptive control method based on IoT data acquisition according to claim 1, characterized in that, Step seven specifically involves: Read the risk level, risk propagation span, event propagation level, and event duration from the risk level sequence, and establish corresponding control levels based on the risk level; The risk propagation span, event propagation level, and event duration corresponding to each control level are weighted and accumulated to generate the control change amount under the corresponding control level, and the correspondence between control level and control change amount is established. Based on the correspondence between control level and control change, calculate the speed adjustment and running duration adjustment of the cooling device, and generate cooling adjustment command; Based on the correspondence between control levels and control changes, calculate the load current adjustment amplitude and load adjustment duration corresponding to the load current, and generate load adjustment commands. Based on the correspondence between control levels and control changes, calculate the sampling interval change and sampling duration for each acquisition node, and generate sampling adjustment instructions. The control command sequence is established by arranging the cooling adjustment command, load adjustment command, and sampling adjustment command in the order of risk propagation time. The control command sequence is sent to the corresponding cooling device, load regulating device and IoT acquisition node, and the corresponding parameter adjustments are executed.

Images