Method and system for automatically monitoring operation state of electromechanical equipment in new energy station

Abstract

The invention relates to the technical field of intelligent monitoring, and provides a new energy station electromechanical equipment operation state automatic monitoring method and system, and the method comprises the steps: constructing a dynamic baseline model, so as to predict the theoretical normal operation index of equipment under a specific environment and working condition; then, by calculating the residual error between the predicted value and the actual value, the complex monitoring problem under the influence of multiple variables is converted into analysis of single residual error flow. According to the method, the size of a single residual value is not independently judged any more, and deep analysis is further performed on the time sequence mode of the residual flow, so that the monitoring dimension is improved from discrete data points to a continuous behavior mode level. By means of the mode, interference of dynamic factors such as the environment can be effectively stripped, early-stage fault symptoms of which the numerical value does not exceed the limit but the behavior mode is abnormal can be accurately captured, and finally high-sensitivity and high-accuracy state monitoring and beforehand early warning of the electromechanical equipment of the new energy field station are achieved.

Description

TECHNICAL FIELD

[0001] The present application relates to the technical field of intelligent monitoring, in particular to a new energy station electromechanical equipment operation state automatic monitoring method and system. BACKGROUND

[0002] With the rapid development of the new energy industry, the scale of new energy stations is expanding, and the stable and efficient operation of electromechanical equipment is crucial to ensuring the reliability and economy of the power system. Therefore, real-time and effective state monitoring and operation and maintenance management of a large number of scattered electromechanical equipment have become an urgent need for industry development.

[0003] However, the existing automatic monitoring methods have obvious technical limitations. The traditional monitoring logic mainly relies on a static and isolated judgment method. On the one hand, it uses fixed thresholds for alarms, but the normal operation parameters of equipment are not constant but fluctuate within a dynamic range due to external environment, self-load and other factors. Therefore, in order to avoid false alarms in extreme conditions, the alarm threshold is usually set relatively wide, which will inevitably lead to the neglect of some abnormal states that should be identified under normal conditions because they do not reach the upper limit of the alarm threshold. On the other hand, the traditional method tends to judge whether the isolated data point at the current time is out of limit, while ignoring the behavior pattern of the equipment operation. However, many progressive and hidden early failures are characterized by a continuous deviation of the equipment's running behavior from its normal pattern over a period of time, rather than a dramatic numerical jump. This makes the monitoring method that only focuses on single-point values unable to effectively identify potential failures characterized by behavior pattern abnormalities. Moreover, the current technology lacks a dynamic baseline model that can accurately describe the behavior of equipment under any operating condition, and lacks effective analysis capabilities for time series patterns of equipment operation data, which severely restricts the accuracy of the monitoring system's early warning and fault recognition capabilities.

[0004] Therefore, an optimized new energy station electromechanical equipment operation state automatic monitoring method is expected. SUMMARY

[0005] The present application aims to at least solve one of the problems in the prior art, and provides a new energy station electromechanical equipment operation state automatic monitoring method and system.

[0006] In one aspect of the present application, a new energy station electromechanical equipment operation state automatic monitoring method is provided, comprising:

[0007] Collecting multi-source heterogeneous data through a sensor network to obtain original equipment data and original environment data;

[0008] Preprocessing the original equipment data and the original environment data to obtain input feature time series and true target time series;

[0009] inputting the input feature time sequence into the trained dynamic baseline model to obtain a predicted target time sequence;

[0010] performing residual error calculation on the predicted target time sequence and the real target time sequence to obtain a residual error stream;

[0011] performing residual error mode analysis on the residual error stream to obtain an anomaly score;

[0012] determining whether an abnormality exists in the operation state of the electromechanical equipment of the new energy station based on a comparison between the anomaly score and a preset threshold.

[0013] Optionally, the electromechanical equipment of the new energy station is a string inverter.

[0014] Optionally, the original equipment data includes direct current voltage, current, alternating current power, and internal temperature data, and the original environment data includes irradiance and environmental temperature data, the input feature time sequence includes time sequences corresponding to the direct current voltage, current, internal temperature data, irradiance, and environmental temperature data, and the real target time sequence includes a time sequence of the alternating current power.

[0015] Optionally, the dynamic baseline model is an XGBoost regression model.

[0016] Optionally, the residual error mode analysis on the residual error stream to obtain an anomaly score includes:

[0017] performing sequence segmentation on the residual error stream based on a sliding window to obtain a time sequence of residual error window data;

[0018] extracting residual local time sequence mode features from each residual error window data in the time sequence of residual error window data to obtain a time sequence of residual local time sequence mode feature encoding vectors;

[0019] performing residual error mode substantial information transfer coding on the time sequence of residual local time sequence mode feature encoding vectors to obtain a residual full time domain mode feature encoding vector;

[0020] performing feature decoding regression on the residual full time domain mode feature encoding vector to obtain the anomaly score.

[0021] Optionally, the residual error mode substantial information transfer coding on the time sequence of residual local time sequence mode feature encoding vectors to obtain a residual full time domain mode feature encoding vector includes:

[0022] Residual local time sequence mode state uncertainty quantification is performed on each residual local time sequence mode feature coding vector in the time sequence of the residual local time sequence mode feature coding vector, to obtain a sequence distribution of residual local time sequence entropy gain;

[0023] Based on the sequence distribution of the residual local time sequence entropy gain, a sequence distribution of residual local time sequence modulation weight is determined.

[0024] Based on the sequence distribution of the residual local time sequence modulation weight, weighted modulation is performed on the time sequence of the residual local time sequence mode feature coding vector, to obtain a time sequence of modulated residual local time sequence mode feature coding vector.

[0025] The time sequence of the modulated residual local time sequence mode feature coding vector is input into a forward LSTM-based sequence encoder, to obtain the residual full time domain mode feature coding vector.

[0026] Another aspect of the application provides a new energy station electromechanical equipment operation state automatic monitoring system, comprising:

[0027] A multi-source heterogeneous data acquisition module is configured to acquire multi-source heterogeneous data through a sensor network to obtain original equipment data and original environment data.

[0028] A data preprocessing module is configured to preprocess the original equipment data and the original environment data to obtain an input feature time sequence and a real target time sequence.

[0029] A target time sequence prediction module is configured to input the input feature time sequence into a trained dynamic baseline model to obtain a predicted target time sequence.

[0030] A residual calculation module is configured to perform residual calculation on the predicted target time sequence and the real target time sequence to obtain a residual flow.

[0031] A residual mode analysis module is configured to perform residual mode analysis on the residual flow to obtain an anomaly score.

[0032] An operation state monitoring module is configured to determine whether the operation state of the new energy station electromechanical equipment is abnormal based on a comparison between the anomaly score and a preset threshold.

[0033] Another aspect of the application provides an electronic device, comprising:

[0034] At least one processor; and,

[0035] A memory in communication connection with the at least one processor; wherein,

[0036] The memory stores instructions executable by the at least one processor, and the instructions are executed by the at least one processor to enable the at least one processor to perform the automatic monitoring method of the operating state of the electromechanical equipment of the new energy station.

[0037] Another aspect of the present application provides a computer-readable storage medium storing a computer program, and the computer program is executed by a processor to implement the automatic monitoring method of the operating state of the electromechanical equipment of the new energy station.

[0038] Another aspect of the present application provides a computer program product comprising a computer program, and the computer program is executed by a processor to implement the automatic monitoring method of the operating state of the electromechanical equipment of the new energy station.

[0039] Compared with the prior art, the automatic monitoring method and system of the operating state of the electromechanical equipment of the new energy station provided by the present application can predict the theoretical normal operating index of the equipment under a specific environment and working condition by constructing a dynamic baseline model, and then convert the complex monitoring problem under the influence of multiple variables into the analysis of a single residual flow by calculating the residual between the predicted value and the actual value. Instead of judging the size of a single residual value, the present application further analyzes the time sequence pattern of the residual flow in depth, so as to improve the monitoring dimension from discrete data points to continuous behavior pattern level. In this way, the interference of dynamic factors such as environment can be effectively stripped, and early signs of failure that are not yet over the limit in numerical value but have already been abnormal in behavior pattern can be accurately captured, so as to finally realize high-sensitivity and high-accuracy state monitoring and early warning of the electromechanical equipment of the new energy station. BRIEF DESCRIPTION OF DRAWINGS

[0040] One or more embodiments are illustrated by way of example in the figures that form a part of this patent document, these example are not to be construed as limiting the embodiments, elements having the same reference numerals in different figures represent like elements, unless otherwise indicated, the figures in the drawings are not to scale.

[0041] Figure 1 A flow chart of an automatic monitoring method of the operating state of electromechanical equipment of a new energy station provided by an embodiment of the present application;

[0042] Figure 2 A data flow schematic diagram of an automatic monitoring method of the operating state of electromechanical equipment of a new energy station provided by another embodiment of the present application;

[0043] Figure 3 A block diagram of an automatic monitoring system of the operating state of electromechanical equipment of a new energy station provided by another embodiment of the present application. DETAILED DESCRIPTION

[0044] For the purposes of making the objectives, technical solutions, and advantages of the embodiments of the present application clearer, the various embodiments of the present application will be described in detail below with reference to the drawings. However, those of ordinary skill in the art can understand that, in the various embodiments of the present application, many technical details are presented in order to make the present application better understood by the readers. However, the technical solutions claimed by the present application can be implemented even without these technical details and based on various changes and modifications of the following embodiments. The division of the following various embodiments is for the convenience of description, and should not constitute any limitation on the specific implementation of the present application, and the various embodiments can be combined with each other and referenced to each other without contradiction.

[0045] One embodiment of the present application relates to a new energy station electromechanical equipment operation state automatic monitoring method, the flowchart of which is shown in Figure 1 The steps S1 to S6 will be described in detail below. Figure 2 The steps S1 to S6 will be described in detail below.

[0046] S1, multi-source heterogeneous data acquisition through a sensor network to obtain original equipment data and original environment data.

[0047] Specifically, the performance of the equipment depends not only on its own health condition, but also closely related to the dynamic changing environment, therefore, in order to comprehensively and accurately capture all relevant variables affecting the operation state of the new energy station electromechanical equipment, step S1 acquires multi-source heterogeneous data through a sensor network to obtain original equipment data and original environment data, so as to provide real and complete data support for subsequent establishment of accurate dynamic baseline model and implementation of high-sensitivity anomaly detection.

[0048] For example, the original equipment data includes direct current voltage, current, alternating current power, internal temperature data. The original environment data includes irradiance and environmental temperature data. The input features include direct current voltage, current, internal temperature data, irradiance and environmental temperature data, and the true target includes alternating current power. These data together constitute the most original and most comprehensive information set, which lays a solid foundation for subsequent data preprocessing, feature extraction and model construction.

[0049] The input feature time series includes time series corresponding to direct current voltage, current, internal temperature data, irradiance and environmental temperature data respectively, and the true target time series includes time series of alternating current power.

[0050] In implementation, step S1 can be completed by a sensor network deployed in the new energy station. The sensor network is a system composed of a large number of distributed sensor nodes, which can measure the physical parameters or environmental conditions of the target object and transmit the collected information to the data processing center through the communication link. Specifically, the data collection process is as follows: first, deploy corresponding sensors in the target electromechanical equipment and its surrounding environment; then, the sensors measure in real time or at a preset frequency, and convert the physical quantity into an electrical signal; finally, through the communication system (such as the SCADA system) in the station, the signals are collected and transmitted to the central data server to form a raw data set with a time stamp.

[0051] S2, pre-processing the original equipment data and the original environment data to obtain input feature time series and real target time series.

[0052] Specifically, the raw data collected directly from the sensor network is often rough and unorganized, may contain noise, outliers or missing values, and it is not clear which data is an influencing factor and which is a result indicator. Therefore, step S2 preprocesses the data to clean, organize and structure it, so as to convert the raw information collected from the physical world, which may contain noise, different formats and is not organized, into a structured, standardized mathematical form that can be understood and processed by the subsequent dynamic baseline model.

[0053] Among them, the input feature time series refers to a data sequence arranged in chronological order, where each time point corresponds to a vector containing multiple input variables (i.e. features), which collectively describe the working condition of the equipment at that time. The real target time series is a single variable time series, which represents the actual observation value of the core performance indicator that we hope the model can learn and predict. When the input features include direct current voltage, current, internal temperature data, irradiance and environmental temperature data, and the real target includes alternating current power, the input feature time series includes the time series corresponding to the direct current voltage, current, internal temperature data, irradiance and environmental temperature data respectively, and the real target time series includes the time series of alternating current power.

[0054] In implementation, data preprocessing is a systematic process. First, it needs to be cleaned up. The original sensor data, including raw device data and raw environmental data, often has missing values due to equipment failure or communication interruption, and abnormal values beyond the physical credible range. The cleaning process will use methods such as linear interpolation, forward and backward value filling to complete the missing data, and eliminate obviously incorrect record points to ensure the integrity and basic quality of the data. Next, data alignment and integration. Since device data and environmental data may come from different collection systems, their sampling frequencies and timestamps may not be completely consistent. This link synchronizes heterogeneous data streams from different sources to a unified time axis through timestamp matching and resampling techniques, forming a wide table data set containing all related variables. Finally, based on the understanding of the device running mechanism, the integrated data set is logically split into two parts: one part is the cause variable set that affects the device running state, and the other part is the result variable that represents the device running performance, to obtain the input feature time series and the real target time series.

[0055] S3, input the input feature time series into the trained dynamic baseline model to obtain the predicted target time series.

[0056] Specifically, through the trained dynamic baseline model, a precise and dynamically changing theoretical benchmark for the actual running performance of new energy station mechanical and electrical equipment is established. This completely abandons the static and broad threshold in traditional methods, aiming to generate a reference system that can reflect the ideal running state of the device under the current specific working condition in real time. This reference system is not a fixed value, but a time series that changes synchronously with the input conditions, which lays the foundation for accurately stripping out the abnormal deviation caused by the change of the device health state.

[0057] Among them, the dynamic baseline model is a mathematical model that can learn and simulate the complex nonlinear relationship between input features and targets. In the training stage, the model adjusts its internal parameters continuously, so that its prediction output of the input feature time series can maximize the approximation of the corresponding real target time series. In a specific example, the dynamic baseline model is an XGBoost(eXtreme Gradient Boosting) regression model. XGBoost is a high-efficiency, flexible and scalable gradient boosting decision tree algorithm, which integrates multiple weak learners(decision trees) to build a powerful prediction model, and is very suitable for processing the table type data prediction problem with multiple input features involved in the present application.

[0058] In implementation, the execution process of step S3 is divided into two stages: offline model training stage and online prediction stage. In the offline model training stage, the XGBoost regression model is fitted using the input feature time series and the real target time series collected and preprocessed during the normal operation period until the prediction error of the model converges to an acceptable range, so as to obtain the trained XGBoost regression model. After entering the online prediction stage, for each new time point, the input feature vector corresponding to the corresponding time (i.e., the vector containing the current direct current, current, temperature, etc.) is sent to the trained XGBoost regression model, and the XGBoost regression model calculates and outputs a predicted value according to the internal law it has learned, which represents the performance index that the device should achieve under the current environment. The predicted values obtained at all time points are arranged in chronological order to obtain the predicted target time series.

[0059] For example, when the monitoring object is a string inverter, the trained XGBoost regression model has learned the AC power output mode of a healthy string inverter under different combinations of illumination, temperature and electrical parameters. In actual operation, the input feature time series obtained by real-time collection and preprocessing, i.e., the continuous data stream composed of [direct current, direct current, internal temperature, irradiance, ambient temperature], is input into the XGBoost regression model time point by time point, and the XGBoost regression model outputs a predicted AC power value sequence, i.e., the predicted target time series, which accurately depicts the AC power trajectory that the inverter should theoretically output under the current variable working conditions.

[0060] S4, residual error calculation is performed on the predicted target time series and the real target time series to obtain a residual error stream.

[0061] Specifically, the dynamic baseline model has quantified and reflected the normal influence of external factors such as environment and load on device performance in the predicted target time series. By calculating the difference between the predicted value and the real observed value, these external influences can be effectively removed, so that the final residual error value theoretically only contains the performance deviation caused by the change of the health state of the device itself and a small amount of random noise, providing an extremely pure and sensitive signal for subsequent anomaly detection, greatly improving the accuracy of fault identification.

[0062] In implementation, for each corresponding time point t in the time series, the residual error value is obtained by arithmetic subtraction, and the specific calculation process is represented by the formula:

[0063] R(t) = Y actual (t) - Y predicted (t);

[0064] where t represents any time point in the time series, R(t) represents the residual value calculated at time point t, Y actual (t) represents the value of the true target time series at time point t, i.e. the actual measurement value of the device, Y predicted (t) represents the value of the predicted target time series at time point t, i.e. the theoretical prediction value output by the dynamic baseline model. By continuously performing this operation for all time points t within the entire monitoring time range, a residual stream consisting of consecutive residual values R(t) can be generated.

[0065] S5, performing residual pattern analysis on the residual stream to obtain an anomaly score.

[0066] Specifically, traditional monitoring methods often only focus on whether the residual value exceeds a certain fixed threshold, while ignoring the dynamic behavior of the residual over time. However, many progressive and subtle faults, whose early signs are precisely reflected in the specific sequential patterns of the residual stream, such as a small but continuous drift, a change in fluctuation frequency, or a subtle morphological difference from the normal pattern. To this end, step S5 performs residual pattern analysis on the residual stream to ultimately quantify and condense these complex pattern information into an intuitive and continuous anomaly score, so as to capture early fault signals that are difficult to discover by single-point values.

[0067] In specific implementation, S5 includes the following steps S51 to S54.

[0068] S51, performing sequence segmentation on the residual stream based on a sliding window to obtain a time series of residual window data.

[0069] That is, by performing sequence segmentation on the residual stream based on a sliding window, the continuous and one-dimensional residual stream data is transformed into a series of overlapping and equal-length residual window data segments, so as to focus the analysis perspective from an infinite time line to individual local time segments.

[0070] S52, extracting residual local time series pattern features from each residual window data in the time series of residual window data to obtain a time series of residual local time series pattern feature encoding vectors.

[0071] In this process, for each residual data segment segmented from the sliding window, a set of statistical or time series features that can describe its morphology and distribution are calculated, such as the mean, variance, skewness, kurtosis, trend slope of the residual in the window. These features are combined into a multi-dimensional vector, i.e. the residual local time series pattern feature encoding vector, which constitutes a compact mathematical description of the residual behavior within the time segment. This process converts the original residual data segment into a sequence of feature vectors with more information content.

[0072] S53, residual mode substantial information transmission coding is performed on the time sequence of the residual local time sequence mode feature coding vector to obtain a residual full-time domain mode feature coding vector.

[0073] Since the AC power output of devices such as string inverters and the like fluctuates continuously with external factors such as irradiance, ambient temperature and the like, the residual sequence contains not only random noise under normal working conditions, but also systematic deviation caused by early failure. Such deviation often manifests as a persistent deviation of a specific time sequence pattern rather than a single-point value mutation, and the traditional static threshold monitoring method cannot capture the gradual abnormal behavior caused by environmental dynamics and load fluctuations during device operation. Therefore, the embodiment quantifies the information entropy change intensity of different time segments in the residual sequence, establishes a time sequence modulation weight distribution, thereby liberating the traditional LSTM encoder from the burden of uniform time modeling. This architecture that combines information theory preprocessing and deep learning coding finally outputs a feature vector that condenses the key abnormal patterns in the full-time domain, providing high-value information representation that contains both microscopic local abnormal features and reflects macroscopic evolution laws for subsequent abnormal scoring.

[0074] Specifically, S53 includes the following steps S531 to S534.

[0075] S531, residual local time sequence mode state uncertainty quantification is performed on each residual local time sequence mode feature coding vector in the time sequence of the residual local time sequence mode feature coding vector to obtain a sequence distribution of residual local time sequence entropy gain.

[0076] Specifically, the operation residual of devices such as inverters and the like under complex environments has significant non-stationary characteristics. When the traditional method directly processes the residual sequence uniformly, it is difficult to distinguish between random noise caused by normal environmental fluctuations and systematic deviation that truly predicts device failure. Therefore, the embodiment introduces entropy gain calculation based on information theory, and the system can penetrate the surface numerical fluctuation to identify abnormalities from the certainty degree of the residual evolution pattern, a deeper feature. This information dynamic-based measurement method enables the system to overcome the limitations of traditional threshold methods and capture fault signals from the inherent ordered changes in device operation patterns.

[0077] More specifically, the specific steps of performing residual local time sequence mode state uncertainty quantification on each residual local time sequence mode feature coding vector are as follows:

[0078] First, the residual local time sequence mode feature coding vector is cross-normalized to obtain a density matrix, which is represented by the formula as follows:

[0079]

[0080] wherein v t is the residual local temporal pattern feature encoding vector at the t-th time instant, v t T is the transpose of v t , and p t is the density matrix at the t-th time instant.

[0081] Next, the Shannon entropy of the density matrix is calculated to quantify the state uncertainty, and in one specific example, the greater the entropy, the higher the uncertainty; the calculation process of the Shannon entropy of the density matrix is represented by the formula:

[0082]

[0083] wherein H(p t ) is the Shannon entropy of p t , log represents a logarithm operation with 2 as the base, l i is the eigenvalue of the i-th position in the density matrix, and d is the total number of elements included in the density matrix.

[0084] Further, based on the Shannon entropy of the density matrix, the residual local temporal entropy gain of each residual local temporal pattern feature encoding vector is calculated, and this process is represented by the formula:

[0085] E t = H(p t ) - H(p t-1 ).

[0086] wherein E t is the residual local temporal entropy gain, and H(p t-1 ) is the Shannon entropy of the density matrix p t-1 at the t-1-th time instant.

[0087] S532, based on the sequence distribution of the residual local temporal entropy gain, determine the sequence distribution of the residual local temporal modulation weight.

[0088] Specifically, device abnormal behavior often manifests as non-uniform changes in operating mode uncertainty, and traditional uniform weighting processing methods can blur these key change points. When a photovoltaic inverter DC side has poor contact, the residual sequence of its power conversion efficiency will suddenly present a chaotic state at a specific operating point, but this key abnormality is easily covered by continuous environmental noise. Therefore, the present embodiment enhances the saliency of abnormal signals through information theory guided feature selection, and specifically, based on the nonlinear weight function of entropy gain, it can automatically amplify these abnormal mode fragments with diagnostic value while weakening the random fluctuations under normal working conditions. This dynamic weight allocation mechanism enables the system to automatically focus on the key time segments that truly reflect the health state transition of the device, rather than mechanically processing all data points.

[0089] More specifically, in one specific example, based on the sequence distribution of the residual local timing entropy gain, the sequence distribution of the residual local timing modulation weight is determined as follows:

[0090]

[0091] wherein μ t and σ t are the mean and standard deviation of the timing entropy gain, β represents a scaling factor of the normalized entropy gain, GELU(·) represents a GELU function, γ represents a decay intensity factor of the historical cumulative entropy, is a normalized activation term, is a historical entropy decay term, ω t is a residual local timing modulation weight.

[0092] S533, based on the sequence distribution of the residual local timing modulation weight, the time sequence of the residual local timing pattern feature encoding vector is weighted and modulated to obtain the time sequence of the modulated residual local timing pattern feature encoding vector.

[0093] Specifically, the change of the equipment operating state often presents a non-uniform time distribution characteristic. When the early degradation of the power module inside the photovoltaic inverter occurs, the abnormal features are not uniformly distributed at all time points, but are concentrated in certain specific working state intervals. The traditional uniform processing method equates these key fault signals with environmental noise, making it difficult to capture early faults in a timely manner. Therefore, the present embodiment can highlight the real fault features in the complex environmental noise background by introducing a dynamic modulation mechanism based on entropy gain, so that the system can intelligently adjust the observation intensity according to the actual operating state of the equipment, and highlight the feature segments containing potential fault information on the premise of maintaining the integrity of the overall data.

[0094] In one specific example, the time sequence of the residual local timing pattern feature encoding vector is weighted and modulated as follows to obtain the time sequence of the modulated residual local timing pattern feature encoding vector:

[0095]

[0096] wherein h t is the modulated residual local timing pattern feature encoding vector at the t-th moment, H is the time sequence of the modulated residual local timing pattern feature encoding vector, h1,h2,...,h n are the 1st, 2nd, …, n-th modulated residual local timing pattern feature encoding vectors in the time sequence of the modulated residual local timing pattern feature encoding vector, respectively.

[0097] S534, input the time series of the modulated residual local time sequence pattern feature encoding vector into a forward LSTM-based sequence encoder to obtain a residual full time domain pattern feature encoding vector.

[0098] Specifically, the evolution of equipment failure often presents complex time-dependent characteristics. When the power module of a photovoltaic inverter appears to be aging, its performance degradation does not occur instantaneously, but rather as a result of the continuous accumulation and propagation of abnormal patterns in specific operating states. The residual sequence after pre-modulation processing highlights the key abnormal time points, but the implicit fault evolution logic still needs to be fully explored through deep time sequence modeling. Therefore, in this embodiment, the LSTM encoder can adaptively select to remember or forget historical information, thereby accurately tracking the propagation path of the fault feature in a dynamically changing operating environment. By establishing this time sequence level fault awareness, the system can distinguish between temporary disturbances and real equipment degradation, and make more accurate predictions of the development trend of the fault.

[0099] In one specific example, the time series of the modulated residual local time sequence pattern feature encoding vector is input into a forward LSTM-based sequence encoder to obtain a residual full time domain pattern feature encoding vector according to the following formula:

[0100] v f = LSTM(H);

[0101] where LSTM(·) represents the LSTM sequence encoding, and v f is the residual full time domain pattern feature encoding vector.

[0102] S54, feature decoding regression is performed on the residual full time domain pattern feature encoding vector to obtain an abnormal score.

[0103] Specifically, although the residual full time domain pattern feature encoding vector contains rich pattern information, it cannot be directly used for alarm judgment. Therefore, the residual full time domain pattern feature encoding vector is feature decoded in this embodiment to convert the high-dimensional, abstract and non-intuitive feature vector output in the previous stage into a single, continuous and explicitly physically meaningful scalar, i.e., an abnormal score.

[0104] In specific implementation, the residual full time domain pattern feature encoding vector is input into a pre-trained feature decoding regression model. The feature decoding regression model is one or more fully connected neural network layers (also known as dense layers or linear layers), the structure of which is designed to map a high-dimensional input vector to a one-dimensional output scalar. This process is a nonlinear function fitting, and the model learns how to regress and predict the abnormal degree corresponding to the complex input feature pattern. The core process of this calculation can be represented by the following formula:

[0105] S anomaly = σ(W · v f + b);

[0106] wherein σ(·) denotes a Sigmoid activation function, v f is a residual full-time mode feature encoding vector, W denotes a weight matrix, b is a bias term, and S anomaly denotes an anomaly score. In this process, the anomaly score is the final product of the process and is a normalized continuous scalar. This score comprehensively reflects all analysis results from the local time series pattern of the residual flow to the global evolution law. The higher the score, the more significant the detected abnormal pattern, and the higher the likelihood of potential failure of the device. Conversely, the lower the score, the more consistent the device operating state is with the health baseline model.

[0107] S6, based on a comparison between the anomaly score and a preset threshold, determining whether the operating state of the electromechanical equipment of the new energy station is abnormal.

[0108] Specifically, the continuous and quantitative anomaly score condensed from all previous complex analyses is converted into a discrete and explicit business guidance judgment conclusion to provide a direct and clear operating state monitoring result for the operation and maintenance personnel, thereby triggering subsequent alarm, repair work order generation or expert review and other operation and maintenance management actions, which is the last link connecting intelligent analysis and actual station operation and maintenance. The preset threshold is a key parameter for realizing effective alarm, and its setting is a process that needs to be carefully weighed to achieve the best balance between alarm sensitivity and accuracy. This preset threshold is usually determined based on a large amount of historical data through statistical analysis of the anomaly score distribution of normal devices and known faulty devices at different stages, and can also be adjusted in combination with the experience of operation and maintenance experts. For example, an optimal cut-off point can be found by analyzing the receiver operating characteristic (ROC) curve to maximize the differentiation between normal and abnormal states. In specific implementation, the operating state of the electromechanical equipment of the new energy station is determined to be abnormal or normal by comparing the anomaly score with the preset threshold. In a specific example, in response to the anomaly score being greater than or equal to the preset threshold, the operating state of the electromechanical equipment of the new energy station is determined to be abnormal; otherwise, it is not. Illustratively, the electromechanical equipment of the new energy station is a string inverter.

[0109] The new energy station electromechanical equipment operation state automatic monitoring method provided by the embodiment of the application, compared with the prior art, a dynamic baseline model is constructed to predict the theoretical normal operation index of the equipment under specific environment and working conditions; subsequently, the residual error between the predicted value and the actual value is calculated, and the complex monitoring problem under the influence of multiple variables is converted into the analysis of a single residual flow. Instead of judging the size of a single residual value in isolation, the time sequence pattern of the residual flow is further analyzed in depth, so that the monitoring dimension is improved from discrete data points to continuous behavior pattern level. In this way, the interference of dynamic factors such as environment can be effectively stripped, and early signs of failure that are not yet over-limit in numerical value but already abnormal in behavior pattern can be accurately captured, and finally high-sensitivity and high-accuracy state monitoring and early warning of new energy station electromechanical equipment are realized.

[0110] Another embodiment of the application relates to a new energy station electromechanical equipment operation state automatic monitoring system 300, as shown in the figure, comprising a multi-source heterogeneous data acquisition module 310, a data preprocessing module 320, a target time sequence prediction module 330, a residual error calculation module 340, a residual error pattern analysis module 350, and an operation state monitoring module 360. Figure 3

[0111] The multi-source heterogeneous data acquisition module 310 is used to acquire multi-source heterogeneous data through a sensor network to obtain original equipment data and original environment data.

[0112] The data preprocessing module 320 is used to preprocess the original equipment data and the original environment data to obtain input feature time sequences and real target time sequences.

[0113] The target time sequence prediction module 330 is used to input the input feature time sequences into the trained dynamic baseline model to obtain the predicted target time sequences.

[0114] The residual error calculation module 340 is used to calculate the residual error of the predicted target time sequences and the real target time sequences to obtain the residual flow.

[0115] The residual error pattern analysis module 350 is used to analyze the residual error pattern of the residual flow to obtain an abnormal score.

[0116] The operation state monitoring module 360 is used to determine whether the operation state of the new energy station electromechanical equipment is abnormal based on the comparison between the abnormal score and the preset threshold.

[0117] The specific implementation method of the new energy station electromechanical equipment operation state automatic monitoring system provided by the embodiment of the application can be referred to the new energy station electromechanical equipment operation state automatic monitoring method provided by the embodiment of the application, which will not be described here.

[0118] ​The new energy station electromechanical equipment operation state automatic monitoring system provided by the embodiment of the application can predict the theoretical normal operation index of the equipment under specific environment and working conditions by constructing a dynamic baseline model, and then convert the complex monitoring problem under multivariate influence into the analysis of a single residual flow by calculating the residual between the predicted value and the actual value. The system no longer judges the size of a single residual value in isolation, but further analyzes the time sequence pattern of the residual flow, thereby improving the monitoring dimension from discrete data points to continuous behavior pattern level. In this way, the interference of dynamic factors such as environment can be effectively stripped, and early fault signs that are not over-limit in numerical value but have already been abnormal in behavior pattern can be accurately captured, and finally high-sensitivity and high-accuracy state monitoring and early warning of the new energy station electromechanical equipment can be realized.

[0119] The new energy station electromechanical equipment operation state automatic monitoring system 300 can be implemented in various wireless terminals, such as a server with a new energy station electromechanical equipment operation state automatic monitoring algorithm. In a possible implementation manner, the new energy station electromechanical equipment operation state automatic monitoring system 300 can be integrated into a wireless terminal as a software module and / or a hardware module. For example, the new energy station electromechanical equipment operation state automatic monitoring system 300 can be a software module in the operating system of the wireless terminal, or can be an application program developed for the wireless terminal; of course, the new energy station electromechanical equipment operation state automatic monitoring system 300 can also be one of the many hardware modules of the wireless terminal.

[0120] The new energy station electromechanical equipment operation state automatic monitoring system 300 and the wireless terminal can also be separate devices, and the new energy station electromechanical equipment operation state automatic monitoring system 300 can be connected to the wireless terminal through a wired and / or wireless network, and transmit interactive information in a conventional data format.

[0121] Another embodiment of the application relates to an electronic device, including: at least one processor; and a memory connected with the at least one processor in communication; wherein the memory stores instructions executable by the at least one processor, and the instructions are executed by the at least one processor to enable the at least one processor to perform the new energy station electromechanical equipment operation state automatic monitoring method described in the above embodiment.

[0122] The memory and the processor are connected by a bus in a bus manner, the bus can include any number of interconnected buses and bridges, the bus connects various circuits of one or more processors and memories together. The bus can also connect various other circuits such as peripheral devices, voltage stabilizers and power management circuits together, which are well known in the art, therefore, they will not be further described herein. The bus interface provides an interface between the bus and the transceiver. The transceiver can be one element or multiple elements such as multiple receivers and transmitters, which provide units for communicating with various other devices on the transmission medium. The data processed by the processor is transmitted on the wireless medium through the antenna, further, the antenna also receives data and transmits the data to the processor.

[0123] The processor is responsible for managing the bus and general processing, and can also provide various functions, including timing, peripheral interface, voltage regulation, power management and other control functions. And the memory can be used to store the data used by the processor in the execution operation.

[0124] Another embodiment of the present application relates to a computer readable storage medium, which stores a computer program, and the computer program is executed by a processor to realize the new energy station electromechanical equipment operation state automatic monitoring method in the above-mentioned embodiment.

[0125] That is, those skilled in the art can understand that all or part of the steps in the method described in the above-mentioned embodiment can be completed by programs instructing related hardware, the programs are stored in a storage medium, and the storage medium includes a plurality of instructions for making a device (which can be a single-chip microcomputer, a chip, etc.) or a processor execute all or part of the steps of the method described in each embodiment of the present application. The foregoing storage medium includes a U disk, a mobile hard disk, a read-only memory (ROM), a random access memory (RAM), a magnetic disk or an optical disk and various storage program codes.

[0126] Another embodiment of the present application relates to a computer program product, which includes a computer program, and the computer program is executed by a processor to realize the new energy station electromechanical equipment operation state automatic monitoring method in the above-mentioned embodiment.

[0127] Those skilled in the art can understand that the above-mentioned embodiments are specific embodiments for realizing the present application, and in actual application, various changes can be made in form and details without departing from the spirit and scope of the present application.

Claims

1. A method for automatically monitoring the operating status of electromechanical equipment in a new energy power station, characterized in that, include: Multi-source heterogeneous data acquisition is performed through sensor networks to obtain raw device data and raw environmental data; The raw equipment data and the raw environmental data are preprocessed to obtain the input feature time series and the real target time series; The input feature time series is input into the trained dynamic baseline model to obtain the predicted target time series; Residual flow is obtained by calculating the residuals of the predicted target time series and the actual target time series; Residual pattern analysis is performed on the residual flow to obtain anomaly scores; Based on the comparison between the anomaly score and the preset threshold, it is determined whether there is an anomaly in the operating status of the electromechanical equipment in the new energy power station.

2. The automatic monitoring method for the operating status of electromechanical equipment in new energy power stations according to claim 1, characterized in that, The electromechanical equipment in the new energy power station is a string inverter.

3. The automatic monitoring method for the operating status of electromechanical equipment in new energy power stations according to claim 2, characterized in that, The raw equipment data includes DC voltage, current, AC power, and internal temperature data; the raw environmental data includes irradiance and ambient temperature data; the input feature time series includes time series corresponding to DC voltage, current, internal temperature, irradiance, and ambient temperature data, respectively; and the real target time series includes the AC power time series.

4. The automatic monitoring method for the operating status of electromechanical equipment in new energy power stations according to claim 1, characterized in that, The dynamic baseline model is an XGBoost regression model.

5. The automatic monitoring method for the operating status of electromechanical equipment in new energy power stations according to claim 1, characterized in that, The residual pattern analysis of the residual flow to obtain anomaly scores includes: The residual stream is segmented into sequences based on a sliding window to obtain a time series of residual window data. Extract residual local temporal pattern features from each residual window data in the time series of the residual window data to obtain the time series of residual local temporal pattern feature encoding vectors; The residual pattern essence information transfer encoding is performed on the time series of the residual local temporal pattern feature encoding vector to obtain the residual full temporal pattern feature encoding vector. The anomaly score is obtained by performing feature decoding regression on the residual full-time-domain pattern feature encoding vector.

6. The automatic monitoring method for the operating status of electromechanical equipment in new energy power stations according to claim 5, characterized in that, The step of performing residual pattern essential information transfer encoding on the time series of the residual local temporal pattern feature encoding vector to obtain the residual full-temporal pattern feature encoding vector includes: The residual local temporal pattern state uncertainty is quantized for each residual local temporal pattern feature encoding vector in the time series of the residual local temporal pattern feature encoding vector to obtain the sequence distribution of the residual local temporal entropy gain; Based on the sequence distribution of the residual local temporal entropy gain, the sequence distribution of the residual local temporal modulation weight is determined; Based on the sequence distribution of the residual local temporal modulation weights, the time series of the residual local temporal pattern feature encoding vector is weighted and modulated to obtain the time series of the modulated residual local temporal pattern feature encoding vector. The time series of the modulated residual local temporal pattern feature encoding vector is input into a forward LSTM-based sequence encoder to obtain the residual full-temporal pattern feature encoding vector.

7. An automatic monitoring system for the operating status of electromechanical equipment in a new energy power station, characterized in that, include: The multi-source heterogeneous data acquisition module is used to acquire multi-source heterogeneous data through a sensor network to obtain raw device data and raw environmental data. The data preprocessing module is used to preprocess the raw equipment data and the raw environmental data to obtain the input feature time series and the real target time series; The target time series prediction module is used to input the input feature time series into the trained dynamic baseline model to obtain the predicted target time series; The residual calculation module is used to perform residual calculation on the predicted target time series and the real target time series to obtain a residual stream; The residual pattern analysis module is used to perform residual pattern analysis on the residual flow to obtain anomaly scores; The operation status monitoring module is used to determine whether there are any abnormalities in the operation status of the electromechanical equipment in the new energy power station based on the comparison between the abnormality score and the preset threshold.

8. An electronic device, characterized in that, include: At least one processor; as well as, A memory communicatively connected to the at least one processor; wherein, The memory stores instructions that can be executed by the at least one processor, which, when executed by the at least one processor, enables the at least one processor to perform the automatic monitoring method for the operating status of electromechanical equipment in new energy power stations as described in any one of claims 1 to 6.

9. A computer-readable storage medium storing a computer program, characterized in that, When the computer program is executed by the processor, it implements the automatic monitoring method for the operating status of electromechanical equipment in new energy power stations as described in any one of claims 1 to 6.

10. A computer program product, comprising a computer program, characterized in that, When the computer program is executed by the processor, it implements the automatic monitoring method for the operating status of electromechanical equipment in new energy power stations as described in any one of claims 1 to 6.

Images