A wind turbine operation risk detection method based on end-cloud cooperation

Abstract

The application discloses a wind turbine operation risk detection method based on end-cloud cooperation, comprising the following steps: collecting operation parameter sets of the wind turbine to generate structured operation time sequence data; constructing a time sequence representation reorganization structure to generate a reorganized time sequence representation; constructing an operation risk distribution modeling framework to perform distributed risk modeling processing and generate an operation risk distribution; performing risk distribution construction processing to generate a cloud-side global evolution risk distribution prior and an end-side individual operation risk residual distribution representation; generating an operation risk detection result based on the reorganized time sequence representation and the end-side individual operation risk residual distribution representation; calculating a distribution distance index and generating a risk distribution offset trigger signal; and recalibrating the cloud-side global evolution risk distribution prior based on the risk distribution offset trigger signal and returning to the end-side. The application improves the stability and self-adaptive ability of operation risk detection under complex working conditions.

Description

Technical Field

[0001] This invention relates to the field of wind turbine operation status monitoring and fault diagnosis technology, and in particular to a wind turbine operation risk detection method based on edge-cloud collaboration. Background Technology

[0002] With the continuous expansion of wind power installed capacity, wind turbines operate continuously for extended periods in complex and variable environments. Their operational status is affected by various factors such as wind condition fluctuations, control strategy switching, and equipment aging, resulting in significant non-stationarity and uncertainty in operational risks. Effectively detecting potential risks during wind turbine operation has become a key technical issue in the safe operation and maintenance management of wind farms.

[0003] Existing wind turbine operation risk detection technologies largely rely on threshold determination, single-point indicator monitoring, or fixed-time-series modeling. These methods typically assume that operational data is stable and continuous over time, making it difficult to adapt to changes in risk evolution characteristics under frequent operating condition switching. They are prone to false alarms or missed alarms in complex operating scenarios. Furthermore, while some methods incorporate probabilistic models or distributed modeling techniques, they are often only implemented in a centralized architecture, making it difficult to simultaneously meet the real-time response requirements of the end-side and the long-term statistical modeling capabilities of the cloud.

[0004] On the other hand, existing edge-cloud collaborative monitoring solutions mostly divide tasks from the perspective of computing power or data transmission, lacking refined modeling and constraint design for the distribution structure of operational risks themselves. The coupling relationship between the cloud update mechanism and the edge risk judgment is unclear, making it difficult to achieve stable and controllable closed-loop adjustment when facing long-term drift of operational risks.

[0005] Therefore, how to provide a method for detecting operational risks of wind turbine units based on edge-cloud collaboration is a problem that urgently needs to be solved by those skilled in the art. Summary of the Invention

[0006] One objective of this invention is to propose a wind turbine operation risk detection method based on edge-cloud collaboration. This invention constructs structured runtime sequence data based on the set of operating parameters generated during wind turbine operation, and introduces a time-series representation reorganization structure driven by operating conditions. The structured runtime sequence data undergoes condition-consistent reorganization processing to form a reorganized time-series representation. Based on this, an operation risk distribution modeling framework is constructed using the time-series basic modeling mechanism of the Moirai model. Distributed risk modeling processing is performed on the reorganized time-series representation to obtain the operation risk distribution. Through edge-cloud functional decoupling, the operation risk distribution is divided into a cloud-based global evolution risk distribution prior and an edge-side individual operation risk residual distribution representation. Furthermore, this invention generates operation risk state values ​​on the edge based on forward time windows and triggers cloud-based recalibration processing in conjunction with distribution distance indicators, forming an edge-cloud collaborative operation risk detection closed loop, thereby achieving stable detection and dynamic adaptation of wind turbine operation risks.

[0007] A wind turbine operation risk detection method based on edge-cloud collaboration according to an embodiment of the present invention includes the following steps:

[0008] Collect the set of operating parameters of the wind turbine, perform time alignment processing, parse the operating condition information, and generate structured runtime sequence data;

[0009] Based on operational condition information, a time-series representation reorganization structure is constructed, and operational condition consistency reorganization processing is performed on the structured runtime time-series data to obtain the reorganized time-series representation.

[0010] A runtime risk distribution modeling framework is constructed based on the Moirai model, and distributed risk modeling processing is performed based on the recombined time series representation to obtain the runtime risk distribution;

[0011] In the operational risk distribution modeling framework, the risk distribution construction process of decoupling edge and cloud functions is performed, and the operational risk distribution is constructed as a priori representation of the global evolutionary risk distribution in the cloud and a representation of the residual distribution of individual operational risks on the edge.

[0012] Based on the recombined time series representation and the residual distribution representation of individual operational risks at the edge, operational risk assessment processing is performed to generate operational risk detection results;

[0013] After outputting the operational risk detection results in each monitoring cycle, the distribution distance index between the residual distribution of the end-side individual operational risk in the current monitoring cycle and the prior global evolutionary risk distribution in the cloud is calculated, and a risk distribution offset trigger signal is generated.

[0014] Based on the risk distribution offset trigger signal, the prior of the global evolution risk distribution in the cloud is recalibrated and transmitted back to the edge to update the residual distribution representation of individual operational risks on the edge.

[0015] Optionally, the generation of the structured runtime sequence data includes:

[0016] The raw operating parameters generated during the operation of the wind turbine are collected to form an operating parameter set, which consists of wind condition parameters, power parameters, structural state parameters, and control state parameters.

[0017] Perform time alignment processing on all types of parameters in the set of operating parameters under a unified sampling period to generate a synchronously sampled data sequence, and obtain the time-aligned set of operating parameters.

[0018] Based on the control state parameters in the time-aligned set of operating parameters, the operating condition information is analyzed to determine the control condition type of the wind turbine. Then, the control condition type is corrected by combining the power change mode corresponding to the power parameters to obtain the operating condition information.

[0019] The time-aligned set of runtime parameters is bound to runtime condition information to generate structured runtime sequence data.

[0020] Optionally, the generation of the recombined timing representation includes:

[0021] Based on the operating condition information, the operating condition identifier corresponding to the structured runtime sequence data is read, and the operating condition dimension is divided. Time sampling points with the same operating condition information are divided into a set of time sequence segments corresponding to the same operating condition.

[0022] For each set of time segments corresponding to each operating condition, splicing is performed according to the order of time occurrence to form a recombined subsequence;

[0023] During the process of splicing recombinant subsequences, isolation constraints are applied to the time segments corresponding to different operating conditions.

[0024] The recombined subsequences are organized according to the operating condition information to construct an operating condition-driven time-series representation recombination structure. The internal representation order of the structured runtime sequence data is subjected to operating condition consistency recombination processing, and the recombined time-series representation is output.

[0025] Optionally, the generation of the operational risk distribution includes:

[0026] Based on the time series fundamental modeling mechanism of the Moirai model, a risk distribution modeling framework is constructed.

[0027] The recombined time series representation is input into the operational risk distribution modeling framework, and distributed feature modeling processing is performed on the recombined time series representation to extract distributed modeling features that characterize the evolution of operational risk.

[0028] Based on the distribution modeling characteristics, parametric modeling processing is performed to construct a set of quantiles and generate a set of risk quantile values, which are then combined to generate the operational risk distribution;

[0029] Based on the operational condition information corresponding to the recombined time series representation, cross-operational condition constraint processing is performed on the operational risk distribution, operational condition isolation constraints are applied to the risk quantile value set, and consistency correction processing is performed.

[0030] Optionally, the construction of the cloud-based global evolutionary risk distribution prior and the edge-side individual operational risk residual distribution representation includes:

[0031] In the operational risk distribution modeling framework, the operational risk distribution is parsed by the distribution parameter type, and the parameters of the operational risk distribution are divided into distribution structure parameters that characterize the long-term evolution characteristics across operating conditions and distribution offset parameters that characterize the current individual operational offset characteristics of the wind turbine.

[0032] Perform time-dimensional statistical aggregation and evolutionary smoothing updates on the risk quantile value set corresponding to the distribution structure parameters to construct a cloud-based global evolutionary risk distribution prior;

[0033] Based on the distribution offset parameter, the distribution offset calculation process is performed on the risk quantile value set corresponding to the current monitoring period and the prior global evolution risk distribution in the cloud, the distribution offset parameter is updated, and the residual distribution representation of the individual operation risk on the edge is constructed.

[0034] The prior representation of the global evolution risk distribution in the cloud and the residual distribution representation of the individual operational risk on the edge are respectively incorporated into the cloud functional structure and edge functional structure of the operational risk distribution modeling framework.

[0035] Optionally, the generation of the operational risk detection results includes:

[0036] In the end-side functional structure of the operational risk distribution modeling framework, a risk assessment input sequence within the forward time window is constructed based on the recombined time series representation and the end-side individual operational risk residual distribution representation.

[0037] Based on the risk assessment input sequence, the cumulative processing of the residual distribution of individual operational risks at the end side is performed across time steps within the forward time window to obtain the cumulative risk distribution representation within the forward time window;

[0038] Based on the cumulative risk distribution representation, the cumulative risk quantity corresponding to the forward time window is extracted, risk status mapping processing is performed, and the operational risk status value is generated as the operational risk detection result.

[0039] Optionally, the generation of the risk distribution offset trigger signal includes:

[0040] After outputting the operational risk detection results in each monitoring cycle, the residual distribution of the operational risk of the individual on the edge corresponding to the current monitoring cycle is obtained and aligned with the prior global evolutionary risk distribution in the cloud, and an aligned risk distribution pair is generated.

[0041] Based on the aligned risk distribution pairs, the difference between the risk quantile values ​​at the corresponding quantile positions of the residual distribution of individual operational risks on the end side and the prior risk distribution of global evolution in the cloud is calculated one by one in the quantile set, and a quantile difference sequence is constructed.

[0042] Weighted summation processing is performed based on the quantile difference sequence to obtain the distribution distance index;

[0043] The distribution distance indicator is subjected to time continuity detection processing within multiple consecutive monitoring periods. When the distribution distance indicator continuously exceeds the preset offset judgment threshold within the preset continuous monitoring period range, and the operation risk status value of the corresponding monitoring period meets the preset risk triggering condition, a risk distribution offset trigger signal is generated.

[0044] Optionally, the update of the end-side individual operational risk residual distribution representation includes:

[0045] In the cloud-based functional structure that runs the risk distribution modeling framework, the risk distribution offset trigger signal generated and uploaded by the edge is received as a control condition to trigger the cloud-based global evolution risk distribution prior recalibration process.

[0046] After the triggering conditions are met, the distribution structure parameters in the prior of the global evolution risk distribution in the cloud are recalibrated and prepared, and the main structure parameters of the Moirai model in the running risk distribution modeling framework are locked.

[0047] Based on the historical trend of the distribution distance index corresponding to the risk distribution offset trigger signal, parameter update processing is performed on the distribution structure parameters to obtain the recalibrated distribution structure parameters.

[0048] Based on the recalibrated distribution structure parameters, the cloud global evolution risk distribution prior is updated to form the recalibrated cloud global evolution risk distribution prior;

[0049] The updated distribution parameters corresponding to the prior of the recalibrated cloud-based global evolution risk distribution are sent back to the edge, and the residual distribution representation of the individual operational risk on the edge is synchronously updated.

[0050] The beneficial effects of this invention are:

[0051] First, this invention constructs a time-series representation reorganization structure driven by operating conditions, performs condition-consistent reorganization processing on structured runtime sequence data, and enables the reorganized time-series representation to reflect the evolutionary continuity of operating risks under the same operating conditions. This effectively avoids the information aliasing problem introduced by traditional physical time sequence modeling in scenarios with frequent operating condition switching, and improves the stability and reliability of operating risk distribution modeling under complex operating conditions.

[0052] Secondly, this invention constructs an operational risk distribution modeling framework based on the time series fundamental modeling mechanism of the Moirai model. Operational risks are represented in a distributed manner in the form of a set of quantiles and a set of risk quantile values. Through cross-operational condition constraint processing and distribution parameter type parsing, the functions of cloud-based global evolutionary risk distribution prior and end-side individual operational risk residual distribution representation are decoupled. This enables the operational risk detection process to have both the ability to characterize long-term evolutionary patterns and the ability to sensitively respond to individual operational deviations.

[0053] Furthermore, this invention introduces operational risk assessment processing under a forward time window at the edge and generates a risk distribution offset trigger signal by combining the distribution distance index. It then performs restricted recalibration processing on the distribution structure parameters in the cloud, forming a closed loop for edge-cloud collaborative operational risk detection. This enables continuous adaptive updating of operational risk detection results without relying on the return of original time-series data, thereby enhancing the robustness and engineering applicability of the system during long-term operation. Attached Figure Description

[0054] 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:

[0055] Figure 1 This is an overall flowchart of a wind turbine operation risk detection method based on edge-cloud collaboration proposed in this invention;

[0056] Figure 2 This is a schematic diagram of the timing representation reassembly structure driven by the operating conditions in this invention;

[0057] Figure 3 This is a schematic diagram of edge-cloud function decoupling and recalibration based on the operational risk distribution modeling framework in this invention. Detailed Implementation

[0058] 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.

[0059] refer to Figure 1-3A method for detecting operational risks of wind turbine units based on edge-cloud collaboration includes the following steps:

[0060] The system collects a set of operating parameters generated during the operation of the wind turbine, including wind condition parameters, power parameters, structural status parameters, and control status parameters. The system aligns the operating parameter set with a uniform sampling period and parses the operating condition information based on the control status parameters to generate structured runtime sequence data.

[0061] Based on the runtime condition information in the structured runtime sequence data, a runtime condition-driven time-series representation reorganization structure is constructed. Using the runtime condition-driven time-series representation reorganization structure, the internal representation sequence of the structured runtime sequence data is subjected to runtime condition consistency reorganization processing to obtain the reorganized time-series representation.

[0062] Based on the time series fundamental modeling mechanism of the Moirai model, a runtime risk distribution modeling framework is constructed. In the runtime risk distribution modeling framework, distributed risk modeling processing is performed based on the recombined time series representation to obtain the runtime risk distribution.

[0063] In the operational risk distribution modeling framework, the operational risk distribution is processed by decoupling the edge and cloud functions to construct the operational risk distribution, which is then constructed as a priori representation of the global evolutionary risk distribution in the cloud and a representation of the residual distribution of individual operational risks on the edge.

[0064] In the edge functional structure of the operational risk distribution modeling framework, based on the recombined time series representation and the edge individual operational risk residual distribution representation, the operational risk judgment process is performed according to the preset forward time window to generate operational risk status values ​​as operational risk detection results;

[0065] After outputting the operational risk detection results in each monitoring cycle, the distribution distance index between the residual distribution of the end-side individual operational risk in the current monitoring cycle and the prior global evolutionary risk distribution in the cloud is calculated. The distribution distance index is obtained by summarizing the quantile difference sequence corresponding to the preset quantile set. When the operational risk state value meets the preset risk triggering condition and the distribution distance index exceeds the preset offset judgment threshold for three consecutive monitoring cycles, a risk distribution offset trigger signal is generated.

[0066] In the cloud-based functional structure of the risk distribution modeling framework, based on the risk distribution offset trigger signal, the cloud-based global evolution risk distribution prior is recalibrated to obtain the recalibrated cloud-based global evolution risk distribution prior. The recalibrated cloud-based global evolution risk distribution prior is then transmitted back to the edge side to update the edge side's individual operational risk residual distribution representation and drive the generation of subsequent operational risk state values, forming a closed loop for edge-cloud collaborative operational risk detection.

[0067] In this embodiment, the generation of structured runtime sequence data includes:

[0068] The raw operating parameters generated during the operation of the wind turbine are collected to form an operating parameter set. The operating parameter set consists of wind condition parameters, power parameters, structural state parameters, and control state parameters. The wind condition parameters are used to characterize the wind speed and wind direction, the power parameters are used to characterize the output power change of the turbine, the structural state parameters are used to characterize the vibration and temperature of the key components of the turbine, and the control state parameters are used to characterize the real-time control state of the turbine's control system.

[0069] Time alignment processing is performed on various parameters in the set of operating parameters under a unified sampling period, so that wind condition parameters, power parameters, structural state parameters and control state parameters form a synchronously sampled data sequence under the same time base, and the time-aligned set of operating parameters is obtained.

[0070] The operating condition information is parsed based on the control status parameters in the time-aligned set of operating parameters. The control status parameters consist of pitch control status, pitch mode identifier, power limiting control status, and start / stop status identifier. The control condition type of the wind turbine is determined according to the combination of control status parameters. The control condition type is then corrected by combining the power change mode corresponding to the power parameters to obtain the operating condition information.

[0071] The time-aligned set of runtime parameters is bound to runtime condition information to generate structured runtime sequence data containing time information, runtime parameter information, and runtime condition information. This structured runtime sequence data is used for the construction and invocation of subsequent runtime condition-driven time sequence representation reassembly structures.

[0072] In this embodiment, the generation of the recombined time-series representation includes:

[0073] Based on the operational condition information already contained in the structured runtime sequence data, the operational condition identifier corresponding to the structured runtime sequence data is read to provide an operational condition index for subsequent time series reassembly processing. Based on the operational condition identifier, the structured runtime sequence data is divided into operational condition dimensions. Time sampling points with the same operational condition information are divided into a set of time series segments corresponding to the same operational condition. The time series segments corresponding to the same operational condition are non-continuously distributed in the time dimension.

[0074] For each set of time segments corresponding to each operating condition, multiple discontinuous time segments corresponding to that operating condition are spliced ​​together in chronological order to form a recombinant subsequence consistent with the operating condition.

[0075] During the recombinant subsequence splicing process, isolation constraints are applied to time segments corresponding to different operating conditions to prevent splicing time segments corresponding to different operating conditions to the same recombinant subsequence.

[0076] To avoid information leakage or temporal causality disruption introduced by the operation condition-driven time-series representation reassembly process, strict temporal causality constraints are imposed on the construction of reassembled subsequences during operation condition consistency reassembly processing. Specifically, the splicing of time-series segments executed under the same operation condition identifier is only allowed to be arranged in the order of the original timestamps, and no time segment is allowed to be inserted before its original occurrence time. This ensures that the information contained at any time position in the reassembled subsequence comes from the operation state that has been observed before that time point. At the same time, time-series segments corresponding to different operation conditions are always physically isolated during the reassembly process, and cross-operation condition information fusion is prohibited. This ensures that any reassembled subsequence only reflects the risk evolution trajectory under a single operation condition. Through the above constraints, the operation condition-driven time-series representation reassembly structure enhances the continuity of risk evolution under the same operation condition without introducing future information or cross-operation condition interference, thereby ensuring the causal consistency and interpretability of the subsequent operation risk distribution modeling process.

[0077] The recombined subsequences consistent with each operating condition are organized according to the operating condition information to construct an operating condition-driven temporal representation recombination structure. Based on the operating condition-driven temporal representation recombination structure, the internal representation sequence of the structured runtime sequence data is subjected to operating condition consistency recombination processing, and the recombined temporal representation is output so that the recombined temporal representation can be used to characterize the continuity of the evolution of operating risks under the same operating condition.

[0078] In this embodiment, the generation of the operational risk distribution includes:

[0079] Based on the time series fundamental modeling mechanism of the Moirai model, a runtime risk distribution modeling framework is constructed for risk distribution learning. The runtime risk distribution modeling framework is used to learn the mapping relationship between the recombined time series representation and the runtime risk distribution under different operating conditions.

[0080] In constructing the operational risk distribution modeling framework, the input format, output format, and training objective of the Moirai model are defined, so that the operational risk distribution modeling framework takes the recombined time series representation as the only time series input object and the distribution parameters of the operational risk distribution as the modeling output object. At the same time, operational condition-related distribution modeling constraints are set in the operational risk distribution modeling framework, so that the Moirai model can distinguish the recombined time series representations corresponding to different operational conditions during the distribution learning process, thereby learning the differentiated evolution characteristics of the operational risk distribution under different operational conditions within the same modeling framework.

[0081] The Moirai model is a probabilistic modeling model for time series data. It is based on the distributed representation of the statistical characteristics of time series in the time dimension. By jointly modeling the potential risk states of time series at different time positions, it learns the mapping relationship between time series input and risk distribution parameters. The Moirai model takes historical observation data of time series as input and models the risk uncertainty corresponding to each time position in the time series. Its output is a set of risk distribution parameters describing the future evolution trend of the time series, rather than a deterministic numerical prediction result. In this invention, the Moirai model is used to perform distribution-level modeling processing on the reconstructed time series representation. By learning the statistical evolution law of the reconstructed time series representation under different operating conditions, it obtains the operating risk distribution used to characterize the overall distribution pattern of operating risk. The operating risk distribution is used as the basic input for subsequent edge-cloud function decoupling and risk judgment processing, and is not directly used to generate operating risk state values.

[0082] The recombined time series representation is input into the operational risk distribution modeling framework. Distributed feature modeling processing is performed on the recombined time series representation in the operational risk distribution modeling framework to extract distributed modeling features used to characterize the evolution characteristics of operational risk.

[0083] In the operational risk distribution modeling framework, parametric modeling processing is performed based on distribution modeling features to construct a set of quantiles and generate a set of risk quantile values, which are then combined to generate an operational risk distribution. The operational risk distribution is jointly characterized by the set of quantiles and the set of risk quantile values.

[0084] The parametric modeling process specifically involves: constructing a quantile set of multi-quantile risk functions based on distribution modeling features; jointly encoding the distribution modeling features and the quantile set to obtain a quantile conditional distribution representation; calculating the risk function output value corresponding to each quantile based on the quantile conditional distribution representation to form a multi-quantile risk function output sequence that corresponds one-to-one with the quantile set; the multi-quantile risk function output sequence is composed of the risk quantile values ​​corresponding to each quantile in the quantile set; performing monotonic consistency constraint processing on the multi-quantile risk function output sequence to ensure that the risk function output value corresponding to the larger quantile in the quantile set is not less than the risk function output value corresponding to the smaller quantile, thus obtaining a risk function output sequence that satisfies quantile monotonic consistency, which is written into the running risk distribution as a risk quantile value set, resulting in a running risk distribution jointly characterized by the quantile set and the risk quantile value set.

[0085] The multi-quantile risk function is defined as a function based on the distribution modeling features and the parameterized mapping of quantiles, and its specific calculation form is as follows:

[0086] ;

[0087] ;

[0088] in, The distribution modeling features are extracted by the operational risk distribution modeling framework based on the recombined time series representation. Indicates quantile The corresponding risk quantile is used to characterize the estimated operational risk at that risk level. Indicates quantile The corresponding parameter vector, express The transpose result, Indicates quantile The corresponding bias parameters, This represents a strictly non-negative mapping function, used to ensure that the increment of risk quantile values ​​corresponding to adjacent quantiles is non-negative, so that the risk quantile value sequence increases with quantiles. The multi-quantile risk function is used to characterize the probability distribution of operational risk at different risk levels.

[0089] The set of quantiles is set using a fixed quantile configuration. The quantile locations cover the low-risk, medium-risk, and high-risk areas of the operational risk distribution. This is used to simultaneously characterize the overall level and tail variation features of operational risk in a single distribution representation. By setting the fixed quantile set, the operational risk distributions generated under different monitoring periods and different operating conditions are ensured to be comparable within the same quantile semantic space, providing a basis for the stable calculation of subsequent distribution distance indicators.

[0090] In the operational risk distribution modeling framework, the Moirai model is limited to a distribution prior learner of the operational risk distribution, and the operational risk distribution modeling framework does not directly output operational risk state values ​​based on the Moirai model;

[0091] Based on the operational condition information corresponding to the recombined time-series representation, cross-operational condition constraint processing is performed on the operational risk distribution in the operational risk distribution modeling framework. Under the same quantile set, operational condition isolation constraints are applied to the risk quantile value set in the operational risk distribution, so that the risk quantile values ​​corresponding to different operational conditions remain independent of each other during the distribution modeling and updating process. Under the operational condition isolation constraint, consistency correction processing is performed on the risk quantile value set to obtain the cross-operational condition operational risk distribution used for subsequent edge-cloud decoupling processing.

[0092] In this embodiment, the construction of the prior representation of the cloud-based global evolution risk distribution and the residual distribution representation of the edge-side individual operational risk includes:

[0093] In the operational risk distribution modeling framework, the distribution parameter type parsing process is performed on the quantile set and the corresponding risk quantile value set in the operational risk distribution. The parameters of the operational risk distribution are divided into distribution structure parameters used to characterize the long-term evolution characteristics across operating conditions and distribution offset parameters used to characterize the current individual operational offset characteristics of the wind turbine.

[0094] The distribution structure parameter is composed of the statistical form of the risk quantile values ​​across monitoring cycles corresponding to the same quantile set of the operation risk distribution. It is used to characterize the overall structural characteristics of the operation risk distribution over time under different operating conditions. The distribution offset parameter is composed of the offset relationship between the risk quantile value set corresponding to the current monitoring cycle and the statistical form across monitoring cycles. It is used to characterize the individual offset characteristics of the current wind turbine relative to the group evolution risk distribution under a given operating condition.

[0095] Based on the distribution structure parameters, the risk quantile value set corresponding to the distribution structure parameters is subjected to time-dimensional statistical aggregation processing within multiple monitoring periods. Under the same quantile value set, the risk quantile value set is subjected to evolutionary smoothing update to complete the time-series evolution modeling of the distribution structure parameters and construct a cloud-based global evolutionary risk distribution prior. The cloud-based global evolutionary risk distribution prior is used to describe the overall distribution pattern of the wind turbine's operating risk under different operating conditions and its statistical characteristics of evolution over time, and serves as the global risk reference distribution in subsequent operating risk judgment processing.

[0096] Based on the distribution offset parameter, the distribution offset calculation is performed on the risk quantile value set corresponding to the current monitoring period and the cloud global evolution risk distribution prior under the same quantile set. The distribution offset parameter is updated, and the end-side individual operation risk residual distribution representation is constructed. The end-side individual operation risk residual distribution representation is used to characterize the individual operation risk offset state of the current wind turbine relative to the group evolution risk distribution under the given operating conditions.

[0097] The cloud-based global evolution risk distribution prior and the edge-side individual operational risk residual distribution representation are respectively incorporated into the cloud-based functional structure and edge-side functional structure of the operational risk distribution modeling framework to support subsequent operational risk determination processing based on the edge-side individual operational risk residual distribution representation, and to provide a distribution comparison benchmark for the calculation of distribution offset triggering conditions.

[0098] In this embodiment, the generation of the operational risk detection result includes:

[0099] In the end-side functional structure of the operational risk distribution modeling framework, a risk assessment input sequence within a forward time window is constructed based on the recombined time series representation and the end-side individual operational risk residual distribution representation. The forward time window covers the risk assessment interval corresponding to multiple consecutive time steps after the current monitoring period. The risk assessment input sequence is composed of operational risk distribution information corresponding to consecutive time steps within the forward time window, which is used to continuously characterize the temporal evolution process of operational risk.

[0100] Based on the risk assessment input sequence, the cumulative processing of the residual distribution representation of the individual operational risk on the front side is performed across time steps within the forward time window to obtain the cumulative risk distribution representation within the forward time window. The cumulative processing is performed based on the risk distribution information corresponding to multiple consecutive time steps within the forward time window, which is used to eliminate the impact of risk fluctuations in a single time step on risk judgment.

[0101] Based on the cumulative risk distribution representation, the cumulative risk quantity corresponding to the forward time window is extracted under the same quantile set, and risk state mapping processing is performed based on the cumulative risk quantity to generate an operational risk state value that characterizes the comprehensive risk level within the forward time window. The operational risk state value is output as the operational risk detection result to characterize the forward operational risk state of the wind turbine in the current operating state.

[0102] In this embodiment, the generation of the risk distribution offset trigger signal includes:

[0103] After outputting the operational risk detection results in each monitoring cycle, the residual distribution representation of the end-side individual operational risk and the prior of the cloud-based global evolutionary risk distribution corresponding to the current monitoring cycle are obtained in the operational risk distribution modeling framework. The two are then aligned under the same quantile set to generate aligned risk distribution pairs for distribution difference calculation.

[0104] Based on the aligned risk distribution pairs, the difference between the residual distribution of the end-side individual operation risk and the prior risk value of the cloud global evolution risk distribution at the corresponding quantile position is calculated one by one in the quantile set, and the quantile difference sequence is constructed according to the order of the quantile set.

[0105] In the process of constructing the quantile difference sequence, the summarization weight corresponding to each quantile is set according to the risk sensitivity of the risk interval in which the quantile is located. The weight corresponding to the high-risk quantile is higher than that of the low-risk quantile, so as to enhance the system's response to abnormal risk evolution, while suppressing the interference of normal fluctuations in the low-risk interval on the distribution offset judgment result, so that the distribution distance index can more accurately reflect the structural offset of the operational risk distribution in the key risk interval.

[0106] A weighted aggregation process is performed based on the quantile difference sequence to obtain a distribution distance index that characterizes the degree of distribution shift in the current monitoring period. The aggregation weight corresponding to each quantile in the quantile difference sequence is set according to the risk sensitivity of the risk level corresponding to that quantile.

[0107] The distribution distance index is subjected to time continuity detection processing within multiple consecutive monitoring periods. When the distribution distance index continuously exceeds the preset offset judgment threshold within the preset continuous monitoring period range, and the operation risk status value of the corresponding monitoring period meets the preset risk triggering condition, a risk distribution offset trigger signal is generated. The risk distribution offset trigger signal is used as the triggering basis for the prior recalibration processing of the cloud-based global evolution risk distribution, which is used to drive the update of distribution parameters in the subsequent end-cloud collaborative operation risk detection process.

[0108] In this embodiment, the update of the end-side individual operational risk residual distribution representation includes:

[0109] In the cloud-based functional structure that runs the risk distribution modeling framework, the risk distribution offset trigger signal generated and uploaded by the edge is received as a control condition to trigger the cloud-based global evolution risk distribution prior recalibration process.

[0110] After the triggering conditions are met, a recalibration preparation process is performed on the distribution structure parameters used to characterize the long-term evolution characteristics across operating conditions in the cloud global evolution risk distribution prior. The main structure parameters of the Moirai model in the operating risk distribution modeling framework are locked, and the recalibration process is limited to the distribution parameter space of the cloud global evolution risk distribution prior.

[0111] Based on the historical trend of the distribution distance index corresponding to the risk distribution offset trigger signal, parameter update processing is performed on the distribution structure parameter so that the updated distribution structure parameter minimizes the long-term offset of the end-side individual operational risk residual distribution representation in multiple monitoring periods under the same quantile set, and the recalibrated distribution structure parameter is obtained.

[0112] Based on the recalibrated distribution structure parameters, the cloud-based global evolution risk distribution prior is updated to form the recalibrated cloud-based global evolution risk distribution prior. The recalibration process does not involve the original structured runtime sequence data, does not involve the reorganization of the time sequence representation, and does not involve the adjustment of the main structure of the runtime risk distribution modeling framework.

[0113] The cloud-based global evolution risk distribution prior recalibration process is not a retraining or online learning of the operational risk distribution modeling framework, but a restricted update process performed within the distribution parameter space. The goal of the recalibration process is to reduce the long-term systematic shift of the end-side individual operational risk residual distribution representation relative to the cloud-based global evolution risk distribution prior over multiple monitoring periods, while maintaining the stability of the original operational risk distribution structure.

[0114] During recalibration, the cloud only adjusts the distribution structure parameters used to characterize the long-term evolution characteristics across operating conditions, without modifying the distribution mapping structure based on the Moirai model in the operating risk distribution modeling framework. This avoids the uncertainty caused by model structure drift. The updated distribution parameters generated during recalibration are sent back in incremental form to synchronously correct the individual operating risk residual distribution representation on the edge side. This allows the edge risk judgment to gradually return to the updated group risk distribution benchmark while maintaining real-time performance, thus achieving stable evolution and long-term consistency maintenance of operating risk distribution under edge-cloud collaboration.

[0115] The updated distribution parameters corresponding to the prior of the recalibrated cloud-based global evolution risk distribution are sent back to the edge. On the edge, the individual operational risk residual distribution representation is synchronously updated based on the updated distribution parameters to drive the generation of subsequent operational risk state values, thus forming a closed loop for edge-cloud collaborative operational risk detection.

[0116] Example 1:

[0117] To verify the feasibility of this invention in practice, it was applied to a continuous operation risk detection scenario for wind turbines. In this scenario, wind turbines operate under frequent wind fluctuations and intensive switching of operating conditions. The turbines alternate between power-limiting control, pitch regulation, and start-stop switching, resulting in significant non-stationary characteristics in operating parameters over time. Traditional risk detection methods based on fixed time sequences or single thresholds are easily affected by operating condition switching in such scenarios, leading to large fluctuations in the operational risk assessment results and making it difficult to accurately reflect the true risk level of the turbine under specific operating conditions. This embodiment addresses this problem by verifying the ability of the method of this invention to stably detect the operational risks of wind turbines under complex operating conditions.

[0118] In practical applications, the system continuously collects a set of operating parameters generated during the operation of the wind turbine. These parameters include wind condition parameters, power parameters, structural state parameters, and control state parameters, all of which are time-aligned within a unified sampling period. Based on the time-aligned control state parameters, the system parses the operating condition information, which distinguishes between different control conditions such as power-limited operation, conventional power generation operation, low-load operation, and start-stop transition states. Subsequently, the time-aligned operating parameter set is bound to the operating condition information to form structured runtime sequence data containing time information, operating parameter information, and operating condition information, providing a unified data foundation for subsequent risk modeling.

[0119] Based on structured runtime sequence data, the system constructs a runtime condition-driven temporal representation reorganization structure. This structure no longer relies solely on physical time continuity as its organizational method; instead, it performs runtime condition-consistent reorganization processing on the structured runtime sequence data according to runtime condition information. Under the same runtime condition, discontinuous time-series segments distributed at different time locations are reorganized into reorganized subsequences consistent with the runtime condition, with time-series segments corresponding to different runtime conditions remaining isolated from each other during the reorganization process. The reorganized temporal representation formed in this way can continuously reflect the evolution of operational risks under the same runtime condition, effectively avoiding interference caused by frequent runtime condition switching on risk modeling.

[0120] After obtaining the recombined time-series representation, the system constructs an operational risk distribution modeling framework based on the time-series fundamental modeling mechanism of the Moirai model. This framework uses the recombined time-series representation as input, extracts distributional modeling features reflecting the evolutionary characteristics of operational risks through distributed risk modeling processing, and further generates an operational risk distribution. The operational risk distribution is characterized by a set of quantiles and a set of risk quantile values, capable of depicting the distribution pattern of operational risks at different risk levels. This operational risk distribution then undergoes edge-cloud functional decoupling processing within the modeling framework, forming a cloud-based global evolutionary risk distribution prior and an edge-side individual operational risk residual distribution representation. The cloud-based global evolutionary risk distribution prior describes the long-term statistical evolution characteristics of operational risks under different operating conditions, while the edge-side individual operational risk residual distribution representation depicts the current unit's offset from the group risk level.

[0121] On the edge side, the system performs operational risk assessment based on the recombined time-series representation and the edge-side individual operational risk residual distribution representation. The operational risk assessment process combines risk distribution information from multiple time steps within a forward time window to cumulatively evaluate the risk, thereby generating operational risk detection results. This approach can suppress the impact of abnormal fluctuations in a single time step on the risk assessment results, making the risk detection results more stable and forward-looking. Simultaneously, after outputting the operational risk detection results in each monitoring cycle, the system calculates the distribution distance index between the edge-side individual operational risk residual distribution representation and the cloud-based global evolutionary risk distribution prior. The distribution distance index reflects the degree of deviation of the current operational risk distribution relative to the long-term evolutionary risk distribution. When the distribution distance index continuously exceeds a preset threshold for multiple consecutive monitoring cycles, the system generates a risk distribution offset trigger signal. Upon receiving the risk distribution offset trigger signal, the cloud performs restricted recalibration processing on the cloud-based global evolutionary risk distribution prior and sends the updated distribution parameters back to the edge side for synchronously updating the edge-side individual operational risk residual distribution representation, thus forming a closed loop for edge-cloud collaborative operational risk detection.

[0122] In this implementation scenario, three methods were compared and used to evaluate the performance of wind turbine operation risk detection: the traditional threshold determination method, the risk detection method based on a fixed time series model, and the operation risk detection method based on edge-cloud collaboration proposed in this invention. The evaluation focused on indicators such as operation risk detection accuracy, false alarm rate, false negative rate, and risk response stability. The relevant experimental statistical results are summarized in the table below.

[0123] Table 1. Performance Comparison Results of Different Operational Risk Detection Methods

[0124] Method Name Traditional threshold determination method Fixed Time Series Model Method Method of the present invention Risk detection accuracy (%) 82.4 86.9 90.3 False alarm rate (%) 11.6 8.4 6.1 Missed report rate (%) 6.0 4.7 3.6 Risk response volatility Larger medium smaller

[0125] As can be seen from the data in Table 1, traditional threshold determination methods, due to their failure to consider the impact of changes in operating conditions on risk distribution, exhibit a high false alarm rate and significant fluctuations in risk response under complex operating conditions. Fixed-time-series model methods reduce some false alarms by introducing time-series modeling, but because they still rely on physical time continuity modeling, the stability of risk assessment is still affected when operating conditions frequently change. In contrast, the method of this invention achieves a relatively stable improvement in risk detection accuracy, further reducing both false alarm and false negative rates, while significantly reducing the fluctuations in risk response.

[0126] The aforementioned performance improvement mainly stems from the synergistic effect formed by the present invention in the time-series representation reorganization structure driven by operating conditions, the modeling of operating risk distribution, and the edge-cloud function decoupling and recalibration mechanism. Through operating condition consistency reorganization processing, the risk modeling process can continuously characterize the risk evolution features under the same operating conditions; through the distributed modeling method, the risk detection results no longer rely on a single numerical judgment; through the edge-cloud collaborative distributed prior recalibration mechanism, the system can maintain its adaptive ability to changes in risk distribution during the long-term evolution of operating risks.

[0127] 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 method for detecting operational risks of wind turbine units based on edge-cloud collaboration, characterized in that, Includes the following steps: Collect the set of operating parameters of the wind turbine, perform time alignment processing, parse the operating condition information, and generate structured runtime sequence data; Based on operational condition information, a time-series representation reorganization structure is constructed, and operational condition consistency reorganization processing is performed on the structured runtime time-series data to obtain the reorganized time-series representation. A runtime risk distribution modeling framework is constructed based on the Moirai model, and distributed risk modeling processing is performed based on the recombined time series representation to obtain the runtime risk distribution; In the operational risk distribution modeling framework, the risk distribution construction process of decoupling edge and cloud functions is performed, and the operational risk distribution is constructed as a priori representation of the global evolutionary risk distribution in the cloud and a representation of the residual distribution of individual operational risks on the edge. Based on the recombined time series representation and the residual distribution representation of individual operational risks at the edge, operational risk assessment processing is performed to generate operational risk detection results; After outputting the operational risk detection results in each monitoring cycle, the distribution distance index between the residual distribution of the end-side individual operational risk in the current monitoring cycle and the prior global evolutionary risk distribution in the cloud is calculated, and a risk distribution offset trigger signal is generated. Based on the risk distribution offset trigger signal, the prior of the global evolution risk distribution in the cloud is recalibrated and transmitted back to the edge to update the residual distribution representation of individual operational risks on the edge.

2. The wind turbine operation risk detection method based on edge-cloud collaboration according to claim 1, characterized in that, The generation of the structured runtime sequence data includes: The raw operating parameters generated during the operation of the wind turbine are collected to form an operating parameter set, which consists of wind condition parameters, power parameters, structural state parameters, and control state parameters. Perform time alignment processing on all types of parameters in the set of operating parameters under a unified sampling period to generate a synchronously sampled data sequence, and obtain the time-aligned set of operating parameters. Based on the control state parameters in the time-aligned set of operating parameters, the operating condition information is analyzed to determine the control condition type of the wind turbine. Then, the control condition type is corrected by combining the power change mode corresponding to the power parameters to obtain the operating condition information. The time-aligned set of runtime parameters is bound to runtime condition information to generate structured runtime sequence data.

3. The wind turbine operation risk detection method based on edge-cloud collaboration according to claim 1, characterized in that, The generation of the recombined time-series representation includes: Based on the operating condition information, the operating condition identifier corresponding to the structured runtime sequence data is read, and the operating condition dimension is divided. Time sampling points with the same operating condition information are divided into a set of time sequence segments corresponding to the same operating condition. For each set of time segments corresponding to each operating condition, splicing is performed according to the order of time occurrence to form a recombined subsequence; During the process of splicing recombinant subsequences, isolation constraints are applied to the time segments corresponding to different operating conditions. The recombined subsequences are organized according to the operating condition information to construct an operating condition-driven time-series representation recombination structure. The internal representation order of the structured runtime sequence data is subjected to operating condition consistency recombination processing, and the recombined time-series representation is output.

4. The wind turbine operation risk detection method based on edge-cloud collaboration according to claim 1, characterized in that, The generation of the operational risk distribution includes: Based on the time series fundamental modeling mechanism of the Moirai model, a risk distribution modeling framework is constructed. The recombined time series representation is input into the operational risk distribution modeling framework, and distributed feature modeling processing is performed on the recombined time series representation to extract distributed modeling features that characterize the evolution of operational risk. Based on the distribution modeling characteristics, parametric modeling processing is performed to construct a set of quantiles and generate a set of risk quantile values, which are then combined to generate the operational risk distribution; Based on the operational condition information corresponding to the recombined time series representation, cross-operational condition constraint processing is performed on the operational risk distribution, operational condition isolation constraints are applied to the risk quantile value set, and consistency correction processing is performed.

5. The wind turbine operation risk detection method based on edge-cloud collaboration according to claim 1, characterized in that, The construction of the prior representation of the cloud-based global evolution risk distribution and the residual representation of the edge-side individual operational risk includes: In the operational risk distribution modeling framework, the operational risk distribution is parsed by the distribution parameter type, and the parameters of the operational risk distribution are divided into distribution structure parameters that characterize the long-term evolution characteristics across operating conditions and distribution offset parameters that characterize the current individual operational offset characteristics of the wind turbine. Perform time-dimensional statistical aggregation and evolutionary smoothing updates on the risk quantile value set corresponding to the distribution structure parameters to construct a cloud-based global evolutionary risk distribution prior; Based on the distribution offset parameter, the distribution offset calculation process is performed on the risk quantile value set corresponding to the current monitoring period and the prior global evolution risk distribution in the cloud, the distribution offset parameter is updated, and the residual distribution representation of the individual operation risk on the edge is constructed. The prior representation of the global evolution risk distribution in the cloud and the residual distribution representation of the individual operational risk on the edge are respectively incorporated into the cloud functional structure and edge functional structure of the operational risk distribution modeling framework.

6. The wind turbine operation risk detection method based on edge-cloud collaboration according to claim 1, characterized in that, The generation of the operational risk detection results includes: In the end-side functional structure of the operational risk distribution modeling framework, a risk assessment input sequence within the forward time window is constructed based on the recombined time series representation and the end-side individual operational risk residual distribution representation. Based on the risk assessment input sequence, the cumulative processing of the residual distribution of individual operational risks at the end side is performed across time steps within the forward time window to obtain the cumulative risk distribution representation within the forward time window; Based on the cumulative risk distribution representation, the cumulative risk quantity corresponding to the forward time window is extracted, risk status mapping processing is performed, and the operational risk status value is generated as the operational risk detection result.

7. The wind turbine operation risk detection method based on edge-cloud collaboration according to claim 1, characterized in that, The generation of the risk distribution offset trigger signal includes: After outputting the operational risk detection results in each monitoring cycle, the residual distribution of the operational risk of the individual on the edge corresponding to the current monitoring cycle is obtained and aligned with the prior global evolutionary risk distribution in the cloud, and an aligned risk distribution pair is generated. Based on the aligned risk distribution pairs, the difference between the risk quantile values ​​at the corresponding quantile positions of the residual distribution of individual operational risks on the end side and the prior risk distribution of global evolution in the cloud is calculated one by one in the quantile set, and a quantile difference sequence is constructed. Weighted summation processing is performed based on the quantile difference sequence to obtain the distribution distance index; The distribution distance indicator is subjected to time continuity detection processing within multiple consecutive monitoring periods. When the distribution distance indicator continuously exceeds the preset offset judgment threshold within the preset continuous monitoring period range, and the operation risk status value of the corresponding monitoring period meets the preset risk triggering condition, a risk distribution offset trigger signal is generated.

8. The wind turbine operation risk detection method based on edge-cloud collaboration according to claim 1, characterized in that, The update of the end-side individual operational risk residual distribution representation includes: In the cloud-based functional structure that runs the risk distribution modeling framework, the risk distribution offset trigger signal generated and uploaded by the edge is received as a control condition to trigger the cloud-based global evolution risk distribution prior recalibration process. After the triggering conditions are met, the distribution structure parameters in the prior of the global evolution risk distribution in the cloud are recalibrated and prepared, and the main structure parameters of the Moirai model in the running risk distribution modeling framework are locked. Based on the historical trend of the distribution distance index corresponding to the risk distribution offset trigger signal, parameter update processing is performed on the distribution structure parameters to obtain the recalibrated distribution structure parameters. Based on the recalibrated distribution structure parameters, the cloud global evolution risk distribution prior is updated to form the recalibrated cloud global evolution risk distribution prior; The updated distribution parameters corresponding to the prior of the recalibrated cloud-based global evolution risk distribution are sent back to the edge, and the residual distribution representation of the individual operational risk on the edge is synchronously updated.

Images