An edge-computing-based real-time data processing method for a wind farm

Abstract

This invention discloses a real-time data processing method for wind farms based on edge computing, belonging to the field of data processing technology. The method includes deploying edge computing nodes at each wind turbine within the wind farm and binding these nodes to the corresponding wind turbine's operational data acquisition interface. The edge computing nodes continuously collect wind turbine operational data and perform unified time alignment and structured processing to form a continuous real-time data stream. The real-time data undergoes quality verification and generates a state data sequence. Based on historical operational data, feature mapping rules are formed at the edge side to compress high-dimensional monitoring data, resulting in a low-dimensional state feature stream. This invention can significantly reduce data transmission and computational load while ensuring the real-time performance of wind farm operational data processing, and improve the accuracy of operational event identification and cause determination, thereby simultaneously improving the efficiency and reliability of wind farm operational monitoring.

Description

Technical Field

[0001] This invention relates to the field of data processing technology, and in particular to a real-time data processing method for wind farms based on edge computing. Background Technology

[0002] As renewable energy generation becomes increasingly widespread, wind farms are becoming larger, more concentrated, and more automated. Modern wind turbines are typically equipped with various sensors, such as those measuring wind speed, wind direction, vibration, temperature, current, and voltage, and their operational status is continuously monitored through wind turbine control systems (like SCADA). These systems collect massive amounts of data every second or even more frequently, providing a foundation for assessing turbine status, issuing early warnings of faults, and scheduling maintenance. However, with the expansion of wind farms and the increase in the number of turbines, the resulting operational data is more multi-dimensional, frequent, and time-sensitive. The traditional approach is to transmit all data to a centralized server or cloud for unified processing. This leads to increasingly significant problems with communication bandwidth constraints, slow processing speeds, and inadequate real-time response. To address these challenges, edge computing technology has been increasingly used in wind farm monitoring in recent years—that is, setting up edge nodes with computing capabilities near the turbines or within the farm site to allow data to be processed and preliminarily analyzed locally. This approach has indeed reduced the burden on the central system and improved response speed and reliability.

[0003] However, current edge computing-based data processing methods still have some shortcomings in practical engineering. First, many existing solutions simply collect or preprocess data without establishing a complete state modeling and event recognition mechanism at the wind turbine side, still requiring the transmission of large amounts of high-frequency, high-dimensional data back to the central platform for analysis. This results in a long processing chain, insufficient real-time performance, and the untapped potential of edge computing in reducing latency and improving efficiency. Second, most current monitoring solutions use fixed processing frequencies and analyze each wind turbine individually, failing to dynamically adjust the processing rhythm based on operating status and lacking collaborative analysis capabilities between wind turbines. This makes it difficult to distinguish between regional environmental changes and genuine problems with a single device, easily leading to a waste of computing and communication resources, and the accuracy of alarms needs improvement. Summary of the Invention

[0004] In view of the aforementioned existing problems, the present invention is proposed.

[0005] Therefore, this invention provides a real-time data processing method for wind farms based on edge computing, which solves the problems of insufficient real-time processing of wind farm operation data, low utilization efficiency of edge computing power and communication resources, and difficulty in accurately distinguishing between regional disturbances and single-machine anomalies in the prior art.

[0006] To solve the above-mentioned technical problems, the present invention provides the following technical solution: In a first aspect, the present invention provides a real-time data processing method for wind farms based on edge computing, comprising, Edge computing nodes are deployed on each wind turbine side within the wind farm, and the edge computing nodes are bound to the corresponding wind turbine's operation data acquisition interface. The edge computing nodes continuously collect wind turbine operation data and complete unified time alignment and structured processing to form a continuous real-time data stream. The real-time data is quality checked and a state data sequence is generated. Based on the historical operation data, a feature mapping rule is formed at the edge to compress the high-dimensional monitoring data, resulting in a low-dimensional state feature stream. A normal operation reference model is constructed based on low-dimensional state feature flow, and the deviation information between the actual state and the reference state is calculated during real-time operation to identify changes in the wind turbine's operating state and generate operating events. The data processing rhythm is adaptively adjusted according to the frequency of operation events to match the processing intensity with the degree of change in operation status. When adjacent wind turbines generate operation events within a similar time window, each edge computing node only shares the event-level results and performs local collaborative analysis to distinguish between regional disturbance events and single-unit abnormal events, and obtain the final event type determination result.

[0007] As a preferred embodiment of the edge computing-based real-time data processing method for wind farms described in this invention, the steps for binding the edge computing node to the corresponding wind turbine's operational data acquisition interface are as follows: Edge computing nodes are deployed on each wind turbine side within the wind farm, and each edge computing node is bound to the corresponding wind turbine one-to-one; Based on the one-to-one binding relationship, the edge computing node is bound to the corresponding wind turbine's operation data acquisition interface, enabling the edge computing node to continuously collect the corresponding wind turbine's operation data, and generating and fixing a data access configuration file on the edge computing node side to describe the acquisition points and sampling rules.

[0008] As a preferred embodiment of the wind farm real-time data processing method based on edge computing described in this invention, the steps for forming a continuous real-time data stream are as follows: Based on the data access configuration file, the edge computing node collects wind turbine operation data according to a preset sampling period, performs unified time correction on the collected data and encapsulates it into structured events with quality tags, and outputs them in chronological order to form a continuous structured event stream.

[0009] As a preferred embodiment of the wind farm real-time data processing method based on edge computing described in this invention, the steps of performing quality verification on the real-time data and generating a state data sequence are as follows: Based on structured event streams, edge computing nodes align and aggregate events by timestamp to establish a continuous time index sequence, and perform missing data detection and outlier detection on the aligned sampled data to remove missing data segments and abnormal measurement point data. After determining missing and abnormal data, a state vector is constructed based on valid measurement point data, and a continuous sequence of state sample windows is generated using a sliding window method to characterize the time evolution features of the wind turbine's operating state.

[0010] As a preferred embodiment of the wind farm real-time data processing method based on edge computing described in this invention, the steps for obtaining the low-dimensional state feature flow are as follows: Based on a continuous and effective sample window sequence, edge computing nodes construct window-level feature samples for training, train a feature importance evaluation model based on the samples, sort the original measurement points and derived features by importance, select the top K features to form a feature set and solidify it into feature mapping rules. Based on the feature mapping rules, feature extraction and mapping are performed on the continuous state vector to generate and continuously output a low-dimensional state feature stream, which is used to characterize the low-dimensional evolution features of the wind turbine's operating state.

[0011] As a preferred embodiment of the wind farm real-time data processing method based on edge computing described in this invention, the steps for identifying changes in the wind turbine operating status and generating operating events are as follows: A normal operation reference model is constructed based on a predefined healthy operation range. The continuously arriving low-dimensional state features are written into a time-series buffer structure, a regression vector is constructed according to a fixed number of backtracking steps, and the model parameters are updated online through recursive least squares. Based on the normal operation reference model, the reference predicted features are compared with the actual low-dimensional state features during real-time operation. The operation deviation is calculated and change point detection is performed on the deviation. At the corresponding time, the operation event is generated and the event tag stream is output.

[0012] As a preferred embodiment of the wind farm real-time data processing method based on edge computing described in this invention, the step of adaptively adjusting the data processing rhythm according to the frequency of occurrence of operating events includes the following steps: After generating the runtime event tagging stream, a fixed-duration sliding time window is established locally. The runtime events entering the window are dynamically maintained, the number of events within the time window is counted, and the event density reflecting the degree of recent changes in the runtime status is calculated. Based on the event density, a data processing execution frequency is generated through a preset deterministic mapping relationship, and the execution frequency is written into the local scheduling table to adjust the triggering cycle of real-time processing tasks. At the same time, the event summary uploading frequency is adjusted synchronously with the execution frequency.

[0013] As a preferred embodiment of the edge computing-based real-time data processing method for wind farms described in this invention, the steps for obtaining the final event type determination result are as follows: After completing the adaptation of the operation event generation and scheduling rhythm, the edge computing nodes establish collaborative groups based on the geographical adjacency of the wind turbines, and generate event-level summaries only for the operation events within a time range consistent with the sliding time window, and exchange them within the collaborative group. The edge computing node performs consistency matching with the local events based on the received neighbor event summaries within a preset time threshold, calculates the neighborhood consistency ratio reflecting the degree of synchronous response of adjacent wind turbines, and outputs a unique event type result based on a fixed judgment threshold, including regional disturbance events and single-unit abnormal events.

[0014] As a preferred embodiment of the wind farm real-time data processing method based on edge computing described in this invention, the determined event type and event information are merged to form an event result record containing the wind turbine number, event time, and final event type, and output in chronological order.

[0015] As a preferred embodiment of the wind farm real-time data processing method based on edge computing described in this invention, when an event is determined to be a regional disturbance event, the edge computing node performs archiving and summary upload operations on the event, while maintaining the local data processing and monitoring strategy unchanged; when an event is determined to be a single-machine abnormal event, the edge computing node increases the attention level and recording granularity of the corresponding wind turbine data processing.

[0016] The beneficial effects of this invention are as follows: By deploying edge computing nodes on the wind turbine side and binding data acquisition interfaces, wind turbine operation data can be uniformly time-aligned and structured at the source, significantly shortening the data processing link and improving the real-time performance of operation status identification. Simultaneously, by performing quality verification, state modeling, and feature compression on real-time data, a low-dimensional state feature stream is formed at the edge, and a normal operation reference model is constructed, enabling rapid perception and event-level output of wind turbine operation status changes, effectively reducing the pressure of high-frequency, high-dimensional data on communication bandwidth and the central system. Furthermore, this invention introduces an adaptive processing rhythm mechanism based on operation event density, allowing data processing intensity to be dynamically adjusted according to changes in operation status, improving the utilization efficiency of edge computing power and bandwidth resources. And through cross-wind turbine event-level collaborative analysis, it distinguishes between regional disturbance events and single-unit abnormal events, improving the accuracy of operation event cause determination and the reliability of operation and maintenance decisions. Attached Figure Description

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

[0018] Figure 1 This is a flowchart of the real-time data processing method for wind farms based on edge computing in Example 1.

[0019] Figure 2 This is a schematic diagram of the cross-wind turbine local collaborative sensing process in Example 1. Detailed Implementation

[0020] To make the above-mentioned objects, features and advantages of the present invention more apparent and understandable, the specific embodiments of the present invention will be described in detail below with reference to the accompanying drawings.

[0021] Many specific details are set forth in the following description in order to provide a full understanding of the invention. However, the invention may also be practiced in other ways different from those described herein, and those skilled in the art can make similar extensions without departing from the spirit of the invention. Therefore, the invention is not limited to the specific embodiments disclosed below.

[0022] Secondly, the term "one embodiment" or "embodiment" as used herein refers to a specific feature, structure, or characteristic that may be included in at least one implementation of the present invention. The phrase "in one embodiment" appearing in different places in this specification does not necessarily refer to the same embodiment, nor is it a single or selective embodiment that is mutually exclusive with other embodiments.

[0023] Example 1, referring to Figure 1 and Figure 2 This is the first embodiment of the present invention, which provides a real-time data processing method for wind farms based on edge computing, including the following steps: S1. Deploy edge computing nodes on each wind turbine side within the wind farm, and bind the edge computing nodes to the corresponding wind turbine's operation data acquisition interface. The edge computing nodes continuously collect wind turbine operation data and complete unified time alignment and structured processing to form a continuous real-time data stream. S1.1: Deploy edge computing nodes on each wind turbine side within the wind farm and bind each edge computing node to the corresponding wind turbine one-to-one; Specifically, a list of wind turbines is exported from the wind farm asset management system to generate a list to be deployed. ,in Number the wind turbine. Station number, For the installation location of the tower base control cabinet, This provides the physical / logical port information for the wind turbine's operation data acquisition interface; after completing the list, maintenance personnel carry the industrial edge computing gateway to the corresponding wind turbine's tower base control cabinet location and press... Install the gateway into the standard guide rail or fixed bracket inside the control cabinet, and connect it to the power supply and grounding terminals of the control cabinet to put the gateway into an operational state. After the gateway powers on, it reads the gateway's unique hardware identifier (such as serial number SN or MAC address MAC) and creates a new identifier locally on the gateway. A named binding directory is used to store configuration files, test point dictionaries, and runtime logs; then, basic binding information is written to this directory. This ensures that all data collected subsequently can be traced back to a specific wind turbine and a specific edge node.

[0024] S1.2: Based on the one-to-one binding relationship, the edge computing node is bound to the corresponding wind turbine's operation data acquisition interface, so that the edge computing node continuously collects the corresponding wind turbine's operation data, and generates and solidifies a data access configuration file on the edge computing node side to describe the acquisition points and sampling rules.

[0025] Specifically, after completing the writing of the basic identifiers for the edge nodes, the operations and maintenance personnel follow... The system establishes a physical link (Ethernet or fiber optic) between the gateway communication port and the wind turbine's operation data acquisition interface, and configures the link parameters (interface address, port number, link keep-alive period) on the gateway side. Then, the gateway immediately performs an "interface handshake verification", that is, sends a standard read request to the operation data acquisition interface and verifies the validity of the response. If the response is successful, the link status is set to "established", thus forming a one-to-one communication channel between the "edge node and the wind turbine acquisition interface". After the communication channel is established, the gateway imports the measurement point list from the measurement point configuration file or the station-side measurement point table exported from the wind turbine control system, and generates a measurement point dictionary TagMap according to a unified field format. The TagMap includes at least four fields: "Measurement Point ID, Measurement Point Semantics, Engineering Unit, and Sampling Period". The gateway performs a consistency check on the measurement point dictionary (checking whether the measurement point ID is duplicated, whether the sampling period is missing, and whether the unit field is empty). After the check passes, the TagMap is permanently stored in the binding directory as the sole basis for subsequent data acquisition and scheduling. The gateway merges the basic binding information, link parameters and TagMap to generate a data access configuration file, and writes fixed parameters such as "collection interface address, measurement point reading mapping relationship and sampling scheduling strategy" into the configuration file, so that the subsequent collection process can be automatically executed by relying only on this configuration file; After generating the data access configuration file, the gateway creates two types of fixed task queues locally and sets their priorities: one is the "real-time processing task thread", which is used to execute tasks with high real-time requirements such as data acquisition scheduling, timestamp writing, and event encapsulation; the other is the "cooperative communication task thread", which is used to exchange event-level summary information with adjacent wind turbine edge nodes. The real-time processing task thread is set to high priority and the cooperative communication task thread is set to low priority to ensure that the determinism of local data acquisition and real-time processing is still prioritized when computing power is tight. After the task thread and priority queue are established, the gateway performs a "binding verification sampling", which means selecting core measurement points from TagMap according to the data access configuration file and continuously reading them to check whether the read values ​​are valid and whether the sampling interval meets the sampling period set by the dictionary. After the verification is successful, the status of the edge node of this wind turbine is set to "binding completed" and the bound data access configuration file is output.

[0026] S1.3: Based on the data access configuration file, the edge computing node collects wind turbine operation data according to a preset sampling period, performs unified time correction on the collected data and encapsulates it into structured events with quality tags, and outputs them in chronological order to form a continuous structured event stream.

[0027] Specifically, the edge computing nodes continuously pull wind turbine operation data according to the data access configuration file and the preset sampling period in the measurement point dictionary. They read the measurement values ​​of each measurement point at the current moment from the wind turbine control system and form raw data records. During the acquisition process, the timestamps are uniformly corrected using a local monotonic clock combined with a network time synchronization mechanism. Each corrected acquisition data is encapsulated as a structured event E(t) and a quality tag q is attached. The edge computing nodes sort the structured events according to the unified timestamps and continuously output them to form a structured event stream.

[0028] Structured events Represented as: It should be noted that the original data records include the measurement point numbers. Corresponding measurement value The unified correction specifically involves compensating and correcting the local timestamp after the node periodically acquires the network time synchronization deviation, generating a unified timestamp t; the quality marker is used to indicate whether the data is a normal acquisition value, a missing value, or an outlier value.

[0029] S2. Perform quality verification on real-time data and generate a state data sequence. Based on historical operation data, form feature mapping rules on the edge side to compress high-dimensional monitoring data and obtain a low-dimensional state feature stream. S2.1: Based on the structured event stream, the edge computing nodes align and aggregate events by timestamp to establish a continuous time index sequence, and perform missing data detection and outlier detection on the aligned sampled data to remove missing data segments and abnormal measurement point data; Specifically, according to and The event stream is grouped and merged, aggregating multiple measurement point events arriving within the same timestamp t into a single sampling frame, and determining the alignment interval Δ based on the sampling period of the TagMap to establish a continuous time index sequence. After establishing the time index, the edge computing node performs a missing value check for each measurement point. After completing the missing value check, the edge computing node performs outlier checks on sampling frames that are not judged as unusable. The measurement point values ​​are checked for out-of-bounds according to the preset valid range in the TagMap; if a value exceeds the limit, the quality flag of that event is directly set to zero. Otherwise, the quality mark is set to For data that has not exceeded the limits, the rate of change of the measurement point within the adjacent sampling interval is calculated and abrupt change is judged. When the rate of change is greater than the rate of change threshold O, it is determined that the measurement point has experienced an abrupt change anomaly at time t, and the edge computing node marks the corresponding event as... This point is then removed from the construction of subsequent state samples to ensure that subsequent state vectors consist only of valid data.

[0030] rate of change The calculation formula is: In the formula, It is the measurement value of the j-th measuring point at time t. It is the time interval between two consecutive samples. It is the measurement value of the j-th measurement point at the previous sampling time t−Δ at the current time t.

[0031] It should be noted that the missing data determination involves counting the number of times a measurement point fails to be read or is marked as missing within a sliding window of length W. If the same... The number of missing values ​​exceeds the threshold. If the window is marked as unavailable and the output of the corresponding time period is stopped, the erroneous input caused by link failure or sensor failure will be blocked at the source. The outlier determination adopts a combination of "fixed engineering threshold + rate of change threshold". The rate of change threshold O is usually taken as 3 to 5 times the mean rate of change of the measurement point in the history of healthy operation data or the 95% to 99% quantile of its rate of change distribution. The basis for setting it is to ensure that it has sufficient tolerance for fluctuations in normal operating conditions, while effectively identifying abnormal sudden changes that exceed the physical inertia and control response capabilities, thereby distinguishing between sensor noise and real abnormal jumps. Fixed engineering thresholds are used to remove measurements that exceed the physically reasonable range, while change rate thresholds are used to identify non-physical jump points within a short period of time. threshold Based on a combination of the sampling period of the measurement points and the allowable duration of continuous missing measurements, it is usually set to 10% to 30% of the number of sampling points within the sliding window length W, so as to tolerate short-term communication jitter while being able to promptly identify continuous missing data caused by sensor failure or link abnormalities.

[0032] S2.2: After completing the missing and anomaly determination, a state vector is constructed based on the valid measurement point data, and a continuous state sample window sequence is generated through a sliding window method to characterize the time evolution characteristics of the wind turbine's operating state.

[0033] Specifically, a state vector is constructed based on the removed data, and a sliding window method is used to combine the state vectors of W consecutive time steps into a sample window. A continuous sequence of valid sample windows is generated in chronological order.

[0034] It should be noted that the state vector construction steps are as follows: The effective values ​​of each measurement point at the same time are aligned and concatenated according to a unified timestamp t to form a state vector sample. If a measurement point is determined to be missing or abnormal at time t, the corresponding component of that measurement point will not participate in the splicing at this time, thus obtaining a state vector composed of valid measurement points.

[0035] S2.3: Based on the continuous effective sample window sequence, edge computing nodes construct window-level feature samples for training, and train a feature importance evaluation model based on the samples. The original measurement points and derived features are ranked by importance, and the top K features are selected to form a feature set and solidified into feature mapping rules. Specifically, a training window set is selected from the sequence of consecutive valid sample windows within the most recent D days in chronological order, and each sample window is mapped to a window-level feature vector. , for each Generate corresponding tags To obtain the training dataset ; Edge computing nodes use training datasets locally Train the random forest model, and calculate each feature after the model is trained. Importance and with Sort all features in descending order; After calculating the importance of all features, the importance vector and the ranking index table RankList are stored in a fixed manner. The top K features are selected from high to low according to the ranking results to form a feature set. and will It is solidified into a "feature mapping rule".

[0036] Importance is defined using the decrease in average impurity as follows: Where T represents the number of decision trees in the random forest. Let represent the set of all nodes in the t-th tree that are partitioned using feature j. This represents the percentage of samples at node n, used to measure the weight of that node's contribution to the overall sample count. This represents the decrease in impurity resulting from partitioning at node n using feature j.

[0037] It should be noted that, Each dimension corresponds to an original measurement point or a derived statistic (such as mean, variance, extreme value, slope, etc.) calculated from that measurement point within the window, thereby converting the "time series sample within the window" into a "fixed-length feature sample that can be directly input into the random forest". Label The system prioritizes taking the "healthy / unhealthy" records or alarm mappings from the operation and maintenance system for that time period. If the wind turbine cannot obtain a tag during this phase, a fixed rule of "stable operation phase = healthy" is used to automatically generate a tag to ensure that the tag generation mechanism is uniquely determined and does not introduce multiple branch selections. The rule includes at least "the retained measurement point ID / derived feature ID, the corresponding location index, the unit and dimension information, and the missing processing flag". The value of K is set to a fixed value by the available computing power of the edge node and the target output bandwidth during the deployment phase. In this embodiment, K is fixed and does not change with the running phase to avoid frequent changes in the rule during the online phase, which would cause logical instability.

[0038] S2.4: Perform feature extraction and mapping on the continuous state vector according to the feature mapping rules, generate and continuously output a low-dimensional state feature stream to characterize the low-dimensional evolution features of the wind turbine's operating state.

[0039] Specifically, after the feature mapping rules are solidified, the edge computing nodes enter online operation mode: receiving continuous state vector samples. ,in accordance with right Perform feature extraction and concatenation to generate low-dimensional feature vectors. .

[0040] The mapping relationship is expressed as: In the formula, Represents a feature mapping operator used to transform the original high-dimensional state vector. Extraction The corresponding components are concatenated in the order of the index to form a low-dimensional vector.

[0041] It should be noted that the impurity can be expressed as Gini impurity or information entropy, and in this embodiment, the same impurity measure is used to ensure that the calculation process is consistent.

[0042] S3. Construct a normal operation reference model based on low-dimensional state feature flow, and calculate the deviation information between the actual state and the reference state during real-time operation to identify changes in the wind turbine's operating state and generate operating events. S3.1: Construct a normal operation reference model based on a predefined healthy operation range, write the continuously arriving low-dimensional state features into a time-series buffer structure, construct a regression vector according to a fixed number of backtracking steps, and update the model parameters online through recursive least squares. Specifically, a known healthy initialization interval is defined locally (this interval can be determined by operation and maintenance records or stable operation periods after startup; the rules are fixed and multi-branching is not introduced). Within this interval, continuously arriving low-dimensional feature vectors are... Write the data sequentially into the time-series buffer queue, and construct the regression vector after the buffer queue length reaches the number of lookback steps M. Reference features for predicting the next time step ; In obtaining the actual low-dimensional features With predictive features Then, online updates using recursive least squares are performed to gradually align the model with the normal operating mode of the wind turbine. The relationship between construction and prediction is as follows: In the formula, To fix the number of replay steps, Let be the parameter matrix to be estimated; Perform an online update using recursive least squares, expressed as: In the formula, For the gain vector, Let covariance matrix be the variance matrix. Forgetting factor, used to achieve a balance between stability and adaptation. It is the parameter estimation error covariance matrix (previous time step). These are the model parameters from the previous time step. It is the model parameter vector / matrix at the current moment.

[0043] It should be noted that, The basis for this is to balance the stability and adaptability of the model by adjusting the weight of historical samples. In engineering practice, the weight is usually set to 0.95 to 0.99 to ensure that the model converges smoothly during normal operation while maintaining adaptability to slow changes in wind turbine operating conditions.

[0044] S3.2: Based on the normal operation reference model, the reference predicted features are compared with the actual low-dimensional state features during real-time operation, the operation deviation is calculated and the change point detection is performed on the deviation, and the operation event is generated and the event tag stream is output at the corresponding time.

[0045] Specifically, in normal operation, the reference model is based on its internal state parameters. Generate predictive features Then, the edge computing node compares the predicted feature with the actual low-dimensional state feature at the same time. By comparing and calculating the deviation of the actual operating state from the reference operating state, changes in the wind turbine's operating state can be identified: Specifically, based on actual low-dimensional features Calculate the residual vector with reference predicted features and its strength scalar To transform the "structural changes in deviation over time" into triggerable runtime events, edge computing nodes pair Perform CUSUM change point detection and calculate cumulative statistics. ; when When a change in running state occurs at time t, a running event is generated directly at that time. The event content should at least include the event occurrence time t and the residual strength. Statistics The event is recorded along with the wind turbine identifier and continuously output in chronological order to form an event tagging stream.

[0046] Calculate the residual vector and its strength scalar The expression is: ; Calculate cumulative statistics The expression is: In the formula, This serves as the baseline mean of the deviation intensity under healthy operating conditions, calculated by edge computing nodes from historical data during model initialization or daily low-risk periods. Sequence statistics are obtained and solidified. This is a drift tolerance parameter used to filter out small fluctuations (calibrated and fixed by the fluctuation level of the healthy segment).

[0047] It should be noted that, Based on the residual intensity during the healthy operation phase The normal fluctuation level is usually set as a small offset near the mean of that period to suppress random fluctuations. The alarm threshold is based on cumulative statistics from historical health data. The distribution characteristics are used to calibrate the system in conjunction with the target false alarm rate, and the running event is only triggered when the residual shows a continuous structural change.

[0048] S4. Adaptively adjust the data processing rhythm according to the frequency of operation events to match the processing intensity with the degree of change in operation status. When adjacent wind turbines generate operation events within a similar time window, each edge computing node only shares the event-level results and performs local collaborative analysis to distinguish between regional disturbance events and single-unit abnormal events, and obtain the final event type determination result.

[0049] S4.1: After generating the running event tagging stream, establish a fixed-duration sliding time window locally, dynamically maintain the running events entering the window, count the number of events within the time window, and calculate the event density that reflects the degree of recent changes in running status. Specifically, a local creation with a length of The edge computing node constructs a sliding time window buffer and writes the timestamps of events entering the window into the buffer sequentially. Whenever a new sampling time t is reached or a new event is generated, the edge computing node immediately removes all event records earlier than t−ΔT from the buffer and retains the event set within the most recent ΔT time range, thereby obtaining the number of events N(t) within the window. After completing the window update, the edge computing node calculates the event density λ(t) to quantify the degree of recent changes in the running state.

[0050] The expression for the event density λ(t) is: .

[0051] S4.2: Based on the event density, generate the data processing execution frequency through a preset deterministic mapping relationship, and write the execution frequency into the local scheduling table to adjust the triggering cycle of real-time processing tasks. At the same time, adjust the event summary uploading frequency in sync with the execution frequency.

[0052] Specifically, edge computing nodes directly map the event density λ(t) to the execution frequency of subsequent data processing. To ensure that the mapping relationship is definite, feasible, and free from scheduling jitter caused by frequency abrupt changes, this embodiment uses a fixed monotonic linear limiting mapping and uses a limiting operator to constrain the execution frequency within the allowable range. The execution frequency is written into the local scheduling table, and the triggering cycle of the real-time processing task is reconfigured using this scheduling table, so that subsequent data processing links are aligned with the target frequency. The event digest is executed periodically; simultaneously, to ensure that the upload bandwidth matches the edge computing power, the event digest upload frequency is adjusted accordingly. Binding, that is, when During the upgrade, the frequency of summary submission is increased simultaneously. The uploading frequency is reduced simultaneously when the frequency is reduced, thereby achieving closed-loop coordination between computing power and bandwidth.

[0053] Execution frequency The expression is: In the formula, It is a limiting operator used to constrain the output to a certain value. Within the specified range, to avoid a sudden surge in events that could lead to uncontrollable frequency, and These are the lowest and highest processing frequencies, respectively. It is the mapping gain, used to adjust the magnitude of the impact of event density on processing frequency.

[0054] It should be noted that, The value is set based on the maximum frequency change slope that the edge node can withstand and the statistical fluctuation range of the event density, and is generally set to... Smooth coverage as event density varies from 0 to typical high values. The scope, in engineering, is often according to Calibration is performed, among which This represents the density of events during periods of high activity in historical data.

[0055] S4.3: After completing the generation of running events and the adaptive scheduling rhythm, the edge computing nodes establish collaborative groups based on the geographical adjacency of the wind turbines, and generate event-level summaries only for running events within a time range consistent with the sliding time window, and exchange them within the collaborative group; Specifically, the wind turbine geographical adjacency table is read, a collaborative group is established, and a neighbor message buffer is created locally. The time span of the buffer is consistent with the length ΔT of the sliding time window buffer to ensure strict alignment between collaborative analysis and scheduling window. When the local machine receives a new running event from the event tagging stream within the most recent ΔT, it immediately generates a summary message for the event and triggers a cooperative broadcast. The summary message is fixed as the minimum set of event-level fields, including the wind turbine number, the time of the event, and the residual strength scalar corresponding to the event. The edge computing node only sends the summary message to all its neighbors, while continuously receiving summary messages from its neighbors within the same ΔT window and writing them into the buffer according to the sending wind turbine number. The S4.4 edge computing node performs consistency matching with the local event based on the received neighbor event summary within a preset time threshold, calculates the neighborhood consistency ratio reflecting the degree of synchronous response of adjacent wind turbines, and outputs a unique event type result according to a fixed judgment threshold. Specifically, after completing the sending of local event summaries and collecting neighbor messages, time consistency matching is performed on the neighbor message buffer using the local event time as a reference, and statistics are collected within the time threshold. Calculate the number of neighbors whose events "occur synchronously" with the local event, and calculate the neighborhood consistency ratio. ; According to a fixed judgment threshold Output a unique event type conclusion: when When this occurs, it indicates that adjacent wind turbines have experienced synchronization events within a similar time window. The edge computing node classifies its local event as a "regional disturbance event" and marks the type as [missing information]. (Corresponding to the spatial synchronization of environmental disturbances such as gusts / wakes); when When this indicates that the event is mainly concentrated on a single wind turbine, the edge computing node classifies the local event as a "single-machine abnormal event" and marks the type as [missing information]. ; The determined event type is merged with the original event information to form an event result record that includes the wind turbine number, event time, event intensity, statistics, and final event type, and is output in chronological order.

[0056] Neighborhood consistency ratio The calculation method is as follows: In the formula, It is the moment of event for neighbor J. This is the moment of the local event. It is the number of neighbors. The time consistency threshold is used to cover the time offset caused by edge communication delays and wind field disturbance propagation. In this embodiment, it is taken as an engineering-usable range of 5 to 10 seconds.

[0057] It should be noted that the time threshold Based on the typical propagation delay of gust / wake disturbances between adjacent wind turbines and the communication delay between edge nodes, the setting is generally taken in the range of 5 to 10 seconds. Determination threshold The statistical calibration is based on the synchronous occurrence ratio of regional disturbance events in historical operating data to ensure that a regional disturbance is only determined when most adjacent wind turbines respond simultaneously. S4.5: When an event is determined to be a regional disturbance event, the edge computing node performs archiving and summary upload operations on the event, while maintaining the local data processing and monitoring strategy unchanged; when an event is determined to be a single-machine abnormal event, the edge computing node increases the attention level and recording granularity of the corresponding wind turbine data processing.

[0058] Specifically, when an event is determined to be a regional disturbance event, the edge computing node marks the event as an environment-related operational event, only archives and uploads a summary of the event, and maintains the current local processing and monitoring strategy of the wind turbine to avoid unnecessary responses to the normal group operation status. When an event is determined to be a single-machine abnormal event, the edge computing node marks the event as a device-related abnormal symptom event and increases the attention level and recording granularity of the subsequent operation data of the wind turbine, providing more refined status information support for the operation and maintenance system or manual diagnosis.

[0059] It should be noted that the above embodiments are only used to illustrate the technical solutions of the present invention and are not intended to limit it. Although the present invention has been described in detail with reference to preferred embodiments, those skilled in the art should understand that modifications or equivalent substitutions can be made to the technical solutions of the present invention without departing from the spirit and scope of the technical solutions of the present invention, and all such modifications or substitutions should be covered within the scope of the claims of the present invention.

Claims

1. A real-time data processing method for wind farms based on edge computing, characterized in that: include, Edge computing nodes are deployed on each wind turbine side within the wind farm, and the edge computing nodes are bound to the corresponding wind turbine's operation data acquisition interface. The edge computing nodes continuously collect wind turbine operation data and complete unified time alignment and structured processing to form a continuous real-time data stream. The real-time data is quality checked and a state data sequence is generated. Based on the historical operation data, a feature mapping rule is formed at the edge to compress the high-dimensional monitoring data, resulting in a low-dimensional state feature stream. A normal operation reference model is constructed based on low-dimensional state feature flow, and the deviation information between the actual state and the reference state is calculated during real-time operation to identify changes in the wind turbine's operating state and generate operating events. The data processing rhythm is adaptively adjusted according to the frequency of operation events to match the processing intensity with the degree of change in operation status. When adjacent wind turbines generate operation events within a similar time window, each edge computing node only shares the event-level results and performs local collaborative analysis to distinguish between regional disturbance events and single-unit abnormal events, and obtain the final event type determination result.

2. The wind farm real-time data processing method based on edge computing as described in claim 1, characterized in that: The steps for binding the edge computing node to the corresponding wind turbine's operational data acquisition interface are as follows: Edge computing nodes are deployed on each wind turbine side within the wind farm, and each edge computing node is bound to the corresponding wind turbine one-to-one; Based on the one-to-one binding relationship, the edge computing node is bound to the corresponding wind turbine's operation data acquisition interface, enabling the edge computing node to continuously collect the corresponding wind turbine's operation data, and generating and fixing a data access configuration file on the edge computing node side to describe the acquisition points and sampling rules.

3. The wind farm real-time data processing method based on edge computing as described in claim 2, characterized in that: The steps to form a continuous real-time data stream are as follows: Based on the data access configuration file, the edge computing node collects wind turbine operation data according to a preset sampling period, performs unified time correction on the collected data and encapsulates it into structured events with quality tags, and outputs them in chronological order to form a continuous structured event stream.

4. The wind farm real-time data processing method based on edge computing as described in claim 3, characterized in that: The steps for performing quality verification on real-time data and generating a state data sequence are as follows: Based on structured event streams, edge computing nodes align and aggregate events by timestamp to establish a continuous time index sequence, and perform missing data detection and outlier detection on the aligned sampled data to remove missing data segments and abnormal measurement point data. After determining missing and abnormal data, a state vector is constructed based on valid measurement point data, and a continuous sequence of state sample windows is generated using a sliding window method to characterize the time evolution features of the wind turbine's operating state.

5. The wind farm real-time data processing method based on edge computing as described in claim 4, characterized in that: The steps to obtain the low-dimensional state feature flow are as follows: Based on a continuous and effective sample window sequence, edge computing nodes construct window-level feature samples for training, train a feature importance evaluation model based on the samples, sort the original measurement points and derived features by importance, select the top K features to form a feature set and solidify it into feature mapping rules. Based on the feature mapping rules, feature extraction and mapping are performed on the continuous state vector to generate and continuously output a low-dimensional state feature stream, which is used to characterize the low-dimensional evolution features of the wind turbine's operating state.

6. The wind farm real-time data processing method based on edge computing as described in claim 5, characterized in that: The steps for identifying changes in the wind turbine's operating status and generating operating events are as follows: A normal operation reference model is constructed based on a predefined healthy operation range. The continuously arriving low-dimensional state features are written into a time-series buffer structure, a regression vector is constructed according to a fixed number of backtracking steps, and the model parameters are updated online through recursive least squares. Based on the normal operation reference model, the reference predicted features are compared with the actual low-dimensional state features during real-time operation. The operation deviation is calculated and change point detection is performed on the deviation. At the corresponding time, the operation event is generated and the event tag stream is output.

7. The wind farm real-time data processing method based on edge computing as described in claim 6, characterized in that: The steps for adaptively adjusting the data processing rhythm based on the frequency of occurrence of runtime events are as follows: After generating the runtime event tagging stream, a fixed-duration sliding time window is established locally. The runtime events entering the window are dynamically maintained, the number of events within the time window is counted, and the event density reflecting the degree of recent changes in the runtime status is calculated. Based on the event density, a data processing execution frequency is generated through a preset deterministic mapping relationship, and the execution frequency is written into the local scheduling table to adjust the triggering cycle of real-time processing tasks. At the same time, the event summary uploading frequency is adjusted synchronously with the execution frequency.

8. The wind farm real-time data processing method based on edge computing as described in claim 7, characterized in that: The steps to obtain the final event type determination result are as follows: After completing the adaptation of the operation event generation and scheduling rhythm, the edge computing nodes establish collaborative groups based on the geographical adjacency of the wind turbines, and generate event-level summaries only for the operation events within a time range consistent with the sliding time window, and exchange them within the collaborative group. The edge computing node performs consistency matching with the local events based on the received neighbor event summaries within a preset time threshold, calculates the neighborhood consistency ratio reflecting the degree of synchronous response of adjacent wind turbines, and outputs a unique event type result based on a fixed judgment threshold, including regional disturbance events and single-unit abnormal events.

9. The wind farm real-time data processing method based on edge computing as described in claim 8, characterized in that: The determined event type and event information are merged to form an event result record containing the wind turbine number, event time, and final event type, and then output in chronological order.

10. The wind farm real-time data processing method based on edge computing as described in claim 9, characterized in that: When an event is determined to be a regional disturbance event, the edge computing node performs archiving and summary upload operations on the event, while maintaining the local data processing and monitoring strategy unchanged; when an event is determined to be a single-machine anomaly event, the edge computing node increases the attention level and recording granularity of the corresponding wind turbine data processing.

Images