New energy station equipment operation monitoring alarm method and system

Abstract

This invention discloses a method and system for monitoring and alarming the operation of new energy power station equipment, relating to the field of equipment operation monitoring technology. The method includes: collecting multi-dimensional telemetry and teleindication data from multiple devices within the power station to construct a time-series data stream; training a deep probabilistic reconstruction model based on historical normal operation data, calculating reconstruction errors, and triggering potential fault warnings; extracting multi-dimensional time-series data from the time-series data stream, initiating spatiotemporal causal tracing analysis, and generating a fault propagation chain; and performing directed graph visualization to generate an explanatory warning report containing the root cause device, fault propagation path, and handling suggestions. This invention solves the technical problems in existing technologies, such as threshold-based alarms leading to delayed warnings, difficulty in automatically identifying fault root causes from multi-device associated alarms, and a lack of spatiotemporal causal reasoning and visualization capabilities. It achieves the technical effects of improving early potential fault perception capabilities, fault root cause location accuracy, alarm interpretability, and emergency response efficiency.

Description

Technical Field

[0001] This invention relates to the technical field of equipment operation monitoring, specifically to a method and system for monitoring and alarming the operation of new energy power station equipment. Background Technology

[0002] New energy power plants, represented by wind power and photovoltaic power generation (such as wind farms and photovoltaic power stations), are typically located in remote and harsh environments compared to traditional thermal or hydropower plants. They also involve a large number and variety of equipment, including wind turbines, inverters, transformer substations, step-up substations, and AGC / AVC systems. Their safe and stable operation directly affects the safety of the power grid and the efficiency of new energy absorption. Monitoring (i.e., centralized monitoring of equipment status by operators) is a core means of detecting anomalies and preventing accidents. Current mainstream monitoring is based on threshold alarm rules, such as alarms for temperatures exceeding 80℃ or voltages falling below 90% of the rated value, combined with simple logical judgments. However, this approach has many shortcomings. It only triggers when the measured value clearly exceeds the limit; early faults such as slight bearing wear or slow decline in inverter efficiency do not immediately lead to exceeding the threshold. It is difficult to adapt to dynamic characteristics such as equipment aging and environmental changes. Furthermore, there are strong coupling relationships between equipment within new energy power plants, making fault identification difficult under complex coupling relationships. Moreover, it cannot automatically generate propagation paths from the root cause equipment to subsequent alarms, lacking spatiotemporal causal reasoning capabilities.

[0003] Therefore, current technologies suffer from several technical problems, including reliance on threshold alarms leading to delayed warnings, difficulty in automatically identifying the root cause of faults from alarms associated with multiple devices, and a lack of spatiotemporal causal reasoning and visualization capabilities. Summary of the Invention

[0004] This application provides a monitoring and alarm method and system for the operation of new energy power station equipment, which solves the technical problems in the prior art, such as the reliance on threshold alarms leading to delayed early warnings, the difficulty in automatically identifying the root cause of faults from alarms associated with multiple devices, and the lack of spatiotemporal causal reasoning and visualization interpretation capabilities. It achieves the technical effects of improving the ability to perceive early potential faults, the accuracy of fault root cause location, the interpretability of alarms, and the efficiency of emergency response.

[0005] This application provides a method for monitoring and alarming the operation of new energy power station equipment. The method includes: collecting multi-dimensional telemetry and teleindication data from multiple devices within the power station, constructing a time-series data stream containing timestamps, device identifiers, measurement point types, and numerical values, and storing it in a real-time database; training a deep probabilistic reconstruction model based on historical normal operation data, wherein the deep probabilistic reconstruction model learns the joint probability distribution between measurement points of each device; during real-time monitoring, extracting data vectors of all relevant measurement points at the current moment from the time-series data stream and inputting them into the model to calculate the reconstruction error; triggering a potential fault warning based on the reconstruction error and recording the current warning trigger timestamp; extracting multi-dimensional time-series data within the time window before and after the warning moment from the time-series data stream according to the potential fault warning, initiating spatiotemporal causal tracing analysis, and generating a fault propagation chain; performing directed graph visualization based on the fault propagation chain, and generating an explanatory warning report containing the root cause device, fault propagation path, and handling suggestions based on the causal relationship.

[0006] In one possible implementation, a deep probabilistic reconstruction model is trained based on historical normal operation data, including: acquiring historical data covering multiple operating conditions and time periods during normal operation, and constructing a training dataset; training the deep probabilistic reconstruction model using the training dataset, wherein the input is multi-device, multi-measurement point data within a sliding time window and the global operating condition parameter sequence within the corresponding window, and the output is the probability distribution parameters of each measurement point within the corresponding window, to learn the joint probability distribution between measurement points under a given operating condition; after training, the reconstruction error is calculated for each training sample, and a quantile regression model is trained with the reconstruction error as the dependent variable and the operating condition parameters of the corresponding sample as the independent variable, to dynamically generate a reconstruction error threshold associated with real-time operating conditions.

[0007] In a possible implementation, the deep probabilistic reconstruction model adopts an encoder-decoder architecture. The encoder consists of a first long short-term memory network layer and a first fully connected layer, used to map the input window sequence to the mean and log-variance of latent variables. The decoder consists of a second fully connected layer and a second long short-term memory network layer, used to reconstruct the probability distribution of each measurement point from the latent variables. By injecting global operating condition parameters into the encoder and decoder step-by-step, the deep probabilistic reconstruction model learns the probability distribution of the measurement point data conditioned by the operating condition parameters, so that the benchmark dynamic adjustment of the same measurement point data being judged as normal or abnormal under different operating conditions.

[0008] In a possible implementation, the encoder maps the input window sequence to the mean and log-variance of latent variables, including: the first long short-term memory network layer receives a multi-device, multi-measurement point data matrix within a sliding time window, and at each time step, it concatenates the current global operating condition parameters as conditional features with the measurement point data before inputting them; the first fully connected layer maps the output of the first long short-term memory network layer to the mean vector and log-variance vector of the latent variables; the decoder reconstructs the probability distribution of each measurement point from the latent variables, including: the second fully connected layer maps the latent variable sample values ​​back to the initial reconstructed features, and the second long short-term memory network layer concatenates the current global operating condition parameters as conditional features with the reconstructed features at each time step, outputting the probability distribution parameters of each measurement point at this time step step by step.

[0009] In possible implementations, a training dataset is constructed, including: extracting time-period data from historical databases that have been confirmed by operations and maintenance to have no fault alarms and cover a variety of typical operating conditions, including wind speed, irradiance, ambient temperature, and grid connection point frequency; after performing missing value imputation, time-aligned resampling, and normalization preprocessing on the extracted normal operation data, training samples are generated using a sliding window approach, with each sample containing a multi-device, multi-measurement point data matrix within a continuous time window and a global operating condition parameter sequence within the corresponding window.

[0010] In a possible implementation, triggering a potential fault warning based on reconstruction error includes: extracting multi-device, multi-measurement point data and corresponding global operating condition parameters from the time-series data stream within the current sliding time window, inputting them into the deep probabilistic reconstruction model, and calculating the reconstruction error of the current window; inputting the operating condition parameters of the current window into the quantile regression model to obtain a dynamic threshold under the current operating condition; if the reconstruction error exceeds the dynamic threshold, triggering the potential fault warning and recording the warning trigger time.

[0011] In a possible implementation, based on the potential fault warning, multi-dimensional time-series data within a time window before and after the warning time is extracted from the time-series data stream, and spatiotemporal causal tracing analysis is initiated to generate a fault propagation chain. This includes: dividing the measurement points into multiple causal analysis regions according to the electrical topology connection relationship of the station, and performing spatiotemporal causal tracing analysis independently in each causal analysis region; calculating the time delay correlation between changes in each measurement point in each causal analysis region based on the multi-dimensional time-series data, and constructing a directed acyclic causal graph with measurement points as nodes and causal influence relationships as edges using the transfer entropy method; identifying nodes with an in-degree of zero as root measurement points based on the directed acyclic causal graph, and generating a fault propagation chain according to the propagation direction of the directed edges.

[0012] In a possible implementation, a directed acyclic causal graph with measurement points as nodes and causal relationships as edges is constructed using the transfer entropy method. This includes: based on multidimensional time-series data within each causal analysis region, converting continuous telemetry values ​​into symbolic sequences and preserving the original symbols for discrete telemetry values, thus establishing a symbolic time series suitable for transfer entropy calculation; setting a maximum time delay step L_max; for any two measurement points M and N within the causal analysis region, calculating the transfer entropy value from M to N with a delay of τ (τ=1, 2, ..., L_max) and from N to M respectively, based on the symbolic time series; for each ordered pair (M, N), determining the optimal delay and the corresponding maximum transfer entropy value; comparing the maximum transfer entropy value with a preset causal threshold to determine directed edges, wherein if both bidirectional transfer entropies exceed the preset causal threshold, comparing their magnitudes and retaining only the direction corresponding to the larger transfer entropy as a directed edge; Using the measurement points as nodes, an initial directed graph is constructed with directed edges. Directed cycles in the initial directed graph are detected. For each detected cycle, the propagation entropy values ​​of each edge in the cycle are compared. Only the edge with the largest propagation entropy is retained, and the remaining edges in the cycle are deleted until there are no directed cycles in the graph, thus obtaining a directed acyclic causal graph.

[0013] In a possible implementation, the operation monitoring and alarm method for new energy power station equipment further includes: collecting normal operation data within a preset time window before the current warning trigger time, and complete multi-dimensional time series data within a preset time window after the warning trigger time; obtaining the final fault cause marked by operation and maintenance, and associating and binding the final fault cause with the root cause measurement point in the directed acyclic causal graph; using the normal operation data within the preset time window as new normal samples, incrementally learning and fine-tuning the deep probabilistic reconstruction model, and updating the model parameters to adapt to the evolution of operating conditions; and incrementally training the quantile regression model based on the collected complete multi-dimensional time series data and the final fault cause marked by operation and maintenance, and adjusting the fitting surface of the dynamic threshold.

[0014] This application also provides an operation monitoring and alarm system for new energy power station equipment. The system includes: a data acquisition module for collecting multi-dimensional telemetry and teleindication data from multiple devices within the power station, constructing a time-series data stream containing timestamps, device identifiers, measurement point types, and numerical values, and storing it in a real-time database; a fault early warning module for training a deep probabilistic reconstruction model based on historical normal operation data. The deep probabilistic reconstruction model learns the joint probability distribution between measurement points of the device. During real-time monitoring, it extracts data vectors of all relevant measurement points at the current moment from the time-series data stream and inputs them into the model to calculate the reconstruction error. Based on the reconstruction error, it triggers a potential fault early warning and records the current early warning trigger timestamp; a fault propagation chain generation module for extracting multi-dimensional time-series data within a time window before and after the early warning time from the time-series data stream based on the potential fault early warning, initiating spatiotemporal causal tracing analysis, and generating a fault propagation chain; and an early warning report generation module for performing directed graph visualization based on the fault propagation chain and generating an explanatory early warning report containing the root cause device, fault propagation path, and handling suggestions based on causal relationships.

[0015] This application proposes a method and system for monitoring and alarming the operation of new energy power station equipment. This involves collecting multi-dimensional telemetry and teleindication data from multiple devices within the power station to construct a time-series data stream. Based on historical normal operation data, a deep probabilistic reconstruction model is trained to calculate reconstruction errors and trigger potential fault warnings. Multi-dimensional time-series data is extracted from the time-series data stream, and spatiotemporal causal tracing analysis is initiated to generate fault propagation chains. A directed graph visualization is then performed to generate an explanatory warning report containing the root cause device, fault propagation path, and handling suggestions. This addresses the technical problems in existing technologies, such as threshold-based alarms leading to delayed warnings, difficulty in automatically identifying fault root causes from multi-device associated alarms, and a lack of spatiotemporal causal reasoning and visualization capabilities. The goal is to improve early potential fault detection capabilities, fault root cause location accuracy, alarm interpretability, and emergency response efficiency. Attached Figure Description

[0016] To more clearly illustrate the technical solutions of the embodiments of this disclosure, the accompanying drawings of the embodiments of this disclosure will be briefly described below. Flowcharts are used in this application to illustrate the operations performed by the system according to the embodiments of this application. It should be understood that the preceding or following operations are not necessarily performed precisely in sequence. Instead, various steps can be processed in reverse order or simultaneously as needed. Furthermore, other operations can be added to these processes, or one or more steps can be removed from these processes.

[0017] Figure 1 A schematic diagram of the operation monitoring and alarm method for new energy power station equipment provided in this application embodiment.

[0018] Figure 2A schematic diagram of the operation monitoring and alarm system for new energy power station equipment provided in this application embodiment.

[0019] Figure labeling: Data acquisition module 10, fault early warning module 20, fault propagation chain generation module 30, early warning report generation module 40. Detailed Implementation

[0020] To further illustrate the technical means and effects adopted by the present invention in order to achieve the intended purpose, the following detailed description is provided in conjunction with the accompanying drawings and preferred embodiments, based on the specific implementation methods, structures, features and effects of the present invention.

[0021] This application provides an embodiment of a method for monitoring and alarming the operation of new energy power station equipment, such as... Figure 1 As shown, the method includes: Step S100: Collect multi-dimensional telemetry and teleindication data from multiple devices within the site, construct a time-series data stream containing timestamps, device identifiers, measurement point types, and numerical values, and store it in a real-time database.

[0022] Preferably, multi-dimensional telemetry data is collected from multiple devices within the power station via wind turbine SCADA, photovoltaic inverters, and weather stations. This includes continuously changing analog measurements of the new energy power station equipment, specifically electrical quantities such as voltage, current, active power, reactive power, frequency, and power factor; non-electrical quantities such as wind speed, wind direction, ambient temperature, nacelle temperature, gearbox temperature, generator speed, pitch angle, photovoltaic irradiance, module backsheet temperature, liquid level, and pressure; and discrete status signals of the new energy power station equipment, specifically circuit breaker opening and closing status, disconnector switch position, protection device operation indicators, and alarm indicators. The system collects and processes data, including operation / shutdown / fault identifiers, manual / automatic mode identifiers, and equipment fault codes. These are then arranged chronologically to form a time-series data stream. Each record corresponds to a sampled value from a measuring point on a specific device at a given moment. Each data record includes a timestamp, device identifier, measuring point type, and numerical value, representing the absolute time of the data sampling, a unique code for the device within the site (e.g., wind turbine number, inverter number, transformer number), the physical quantity or state quantity to which the data belongs (e.g., nacelle wind speed, active power, circuit breaker position), and the actual measured value or state value of the measuring point at the corresponding moment. Finally, the time-series data stream is stored in a real-time database for high-speed read / write operations, such as an industrial real-time database or an open-source time-series database, enabling rapid writing by timestamp and querying by time range.

[0023] Step S200: Based on historical normal operation data, a deep probabilistic reconstruction model is trained. The deep probabilistic reconstruction model learns the joint probability distribution between each measuring point of the device. During real-time monitoring, the data vectors of all relevant measuring points at the current moment are extracted from the time-series data stream and input into the model to calculate the reconstruction error. Based on the reconstruction error, a potential fault warning is triggered, and the current warning trigger timestamp is recorded.

[0024] Step S200 further includes acquiring historical data covering multiple operating conditions and time periods during normal operation, and constructing a training dataset; using the training dataset to train the deep probabilistic reconstruction model, wherein the input is multi-device, multi-measurement point data within a sliding time window and the global operating condition parameter sequence within the corresponding window, and the output is the probability distribution parameters of each measurement point within the corresponding window, to learn the joint probability distribution between measurement points under a given operating condition; after training, the reconstruction error is calculated for each training sample, and a quantile regression model is trained with the reconstruction error as the dependent variable and the operating condition parameters of the corresponding sample as the independent variable, to dynamically generate a reconstruction error threshold associated with the real-time operating condition.

[0025] Preferably, historical time-series data with a duration of no less than a certain period (e.g., 30 or 90 days) is extracted from an existing real-time database. During this period, no equipment failures, protection actions, or abnormal alarm events occurred at the site. Multiple operating conditions refer to external operating conditions affecting the equipment's operational status, specifically including low wind speed ranges (e.g., 0~4 m / s), medium wind speed ranges (4~12 m / s), high wind speed ranges (above 12 m / s), and nighttime (0 W / m). 2The historical data selected includes continuous operation records for each range: low irradiance, medium irradiance, high irradiance, low temperature, normal temperature, and high temperature. Then, the historical data undergoes missing value imputation, time-aligned resampling, and normalization preprocessing. Training samples are generated using a sliding window method, with a fixed window length of T time steps (e.g., T=100) and a sliding step size of S time steps. Next, multi-device, multi-point data within the sliding time window is used as input, forming a T×D matrix, where T is the number of time steps within the window, and D is the total number of all measuring points for all monitored devices. Simultaneously, the global operating condition parameter sequence within the corresponding window is used as input, forming a T×K matrix, where K is the number of operating condition parameters (e.g., wind speed, etc.). The model outputs four parameters: irradiance, ambient temperature, and grid connection frequency. Each time step corresponds to a set of operating condition parameters. The model reconstructs the probability distribution parameters of each measuring point within the corresponding window. For each measuring point, the model outputs the parameters of the probability distribution that the measuring point value follows. For measuring points that follow a Gaussian distribution, the model outputs the mean and variance, i.e., the output shape is T×D×P, where P is the number of probability distribution parameters for each measuring point. The model learns the conditional probability distribution P(measuring point data | operating condition parameters), which represents the joint probability distribution of the measuring point values ​​under given operating conditions. During training, the model minimizes the negative log-likelihood loss function, so that under normal operating conditions, the real measuring point data has a high probability of appearing in the probability distribution output by the model.

[0026] Preferably, after training, the reconstruction error is calculated for each training sample in the training dataset, which is the negative log-likelihood or root mean square error between the actual values ​​of all measuring points within the window at all time steps and the probability distribution output by the model. Each training sample corresponds to a vector of operating condition parameters (such as average wind speed, average irradiance, average ambient temperature, and average grid connection frequency within the window) and a reconstruction error value. Then, with the operating condition parameters as independent variables and the reconstruction error as the dependent variable, a quantile regression model is trained to learn the conditional quantile function Q_τ (reconstruction error | operating condition parameters). The quantile τ values ​​include 0.95, 0.98, and 0.99. After training, when any new set of operating condition parameters is input, the quantile regression model outputs a specific quantile threshold for the reconstruction error under that operating condition. Under different operating conditions, the normal fluctuation range of the equipment is different, and the distribution of the normal reconstruction error is also different. For example, under high wind speed conditions, the normal range of reconstruction error is larger, and the threshold is correspondingly higher; under low wind speed conditions, the fluctuation is smaller, and the threshold is correspondingly lower.

[0027] Furthermore, step S200 also includes that the deep probabilistic reconstruction model adopts an encoder-decoder architecture. The encoder consists of a first long short-term memory network layer and a first fully connected layer, used to map the input window sequence to the mean and log-variance of latent variables. The decoder consists of a second fully connected layer and a second long short-term memory network layer, used to reconstruct the probability distribution of each measurement point from the latent variables. In this way, by injecting global operating condition parameters into the encoder and decoder step by step, the deep probabilistic reconstruction model learns the probability distribution of the measurement point data conditioned by the operating condition parameters, so that the same measurement point data is judged as normal or abnormal under different operating conditions by the benchmark dynamic adjustment.

[0028] Step S200 further includes: an encoder mapping the input window sequence to the mean and log-variance of latent variables; the first long short-term memory network layer receiving a multi-device, multi-measurement point data matrix within a sliding time window, and inputting the current global operating condition parameters as conditional features concatenated with the measurement point data at each time step; the first fully connected layer mapping the output of the first long short-term memory network layer to the mean vector and log-variance vector of latent variables; the decoder reconstructing the probability distribution of each measurement point from the latent variables; the second fully connected layer mapping the latent variable sample values ​​back to the initial reconstructed features; the second long short-term memory network layer concatenating the current global operating condition parameters as conditional features with the reconstructed features at each time step, and outputting the probability distribution parameters of each measurement point at this time step step by step.

[0029] Preferably, the deep probabilistic reconstruction model adopts an encoder-decoder architecture. The encoder compresses the input sequence into latent variables, and the decoder restores the latent variables into output. Specifically, the encoder consists of a first long short-term memory network layer and a first fully connected layer. The first long short-term memory network layer receives the input window sequence, with an input shape of T×(D+K), where T is the number of time steps within the window, D is the total number of multiple devices and multiple measurement points, and K is the number of global operating parameters. At each time step, all measurement point data at the current time are concatenated with the global operating parameters at the current time and input into the layer. The input is processed step by step in chronological order, and a hidden state vector is output at each time step. The hidden state vector output at the last time step contains the temporal dependency information of the entire window sequence. The first fully connected layer receives the hidden state vector output at the last time step and maps the hidden state vector into two output vectors through linear transformation and activation function, including the mean and log-variance of the latent variables. The latent variables are a compressed representation of the input window sequence, used to capture the correlation and temporal dependency between all measurement points within the window. Their dimension is a preset hyperparameter, smaller than the dimension of the input data.

[0030] Preferably, the decoder consists of a second fully connected layer and a second long short-term memory network layer. The second fully connected layer receives the latent variable vector sampled from the latent variable distribution and maps the latent variable vector to an initial reconstructed feature vector through a linear transformation. Its dimension is the same as the dimension D of the measurement point in a single time step in the input sequence. The second long short-term memory network layer receives the initial reconstructed features output by the second fully connected layer as the initial hidden state. At each time step, it concatenates the reconstructed features of the current time step with the global operating parameters of the current time step and inputs them. Then, it outputs the output vector of each time step in chronological order. After the output vector of each time step is transformed by an additional fully connected layer, the probability distribution parameters of each measurement point at that time step are output. The probability distribution parameters include the mean and logarithmic variance of each measurement point.

[0031] Preferably, in the encoder, at each time step, the global operating condition parameter vector at the current moment is concatenated with the measurement point data vector at the current moment to form a longer vector as the input for that time step. In the decoder, at each time step, the global operating condition parameter vector at the current moment is concatenated with the reconstructed feature vector at the current time step as the input for that time step. The conditional probability distribution P(measurement point data | operating condition parameters) learned by the deep probabilistic reconstruction model represents the joint probability distribution of the measurement point values ​​under the condition of known operating conditions. The vertical line represents the conditional dependency relationship. The right side of the equal sign is the given condition, and the left side is the variable to be predicted. This dynamically adjusts the baseline for judging the same measurement point data as normal or abnormal under different operating conditions. For example, the same measurement point data value may belong to the high probability area of ​​normal distribution under high wind speed conditions; the same measurement point data value may belong to the low probability area of ​​normal distribution under low wind speed conditions, and thus be judged as abnormal.

[0032] Furthermore, step S200 also includes extracting time period data from the historical database that has been confirmed by operation and maintenance to have no fault alarms and covers a variety of typical operating conditions, including wind speed, irradiance, ambient temperature and grid connection point frequency; after performing missing value imputation, time alignment resampling and normalization preprocessing on the extracted normal operation data, training samples are generated using a sliding window method. Each sample contains a multi-device multi-measurement point data matrix within a continuous time window and a global operating condition parameter sequence within the corresponding window.

[0033] Preferably, time-period data covering multiple typical operating conditions is extracted from historical databases. The extracted data periods must be verified and confirmed by maintenance personnel, ensuring that no equipment failures, protection trips, or abnormal alarm events occurred at the site during these periods. For example, there should be no alarm entries in the event log, no fault repair records in the maintenance log, and all equipment status indicators should show normal operation. Typical operating conditions include wind speed, irradiance, ambient temperature, and grid connection frequency. The data periods must contain continuous records under different operating conditions, including at least different wind speed ranges: low wind speed (e.g., 0–4 m / s), medium wind speed (4–12 m / s), and high wind speed (above 12 m / s), and different irradiance ranges, including nighttime (0 W / m²). 2 The data was analyzed across various irradiance levels (low, medium, and high), ambient temperature ranges (low, normal, and high), and grid connection frequency ranges (below 49.5 Hz, 49.5–50.5 Hz, and above 50.5 Hz). The extracted normal operation data was examined to identify missing values ​​due to sensor malfunctions, communication interruptions, etc. The arithmetic mean of the values ​​before and after the missing points was used to fill in the missing values. Since different data sources have different sampling periods, resampling was performed to unify all measurement points to the same time interval, for example, converting all data to one record every 5 seconds. High-frequency data was downsampled to the target frequency, and low-frequency data was upsampled to the target frequency through interpolation. Finally, range normalization was used to convert measurement point data with different dimensions and numerical ranges to a unified numerical range between 0 and 1. Training samples are generated using a sliding window method. The window length W is a pre-defined fixed number of time steps, such as 100, meaning each window contains 100 consecutive time step data records. The sliding step size is the number of time steps the window moves backward each time, such as 1 or 10 time steps each time. The window slides from the start time of the historical data to the end time to obtain training samples. The order of time steps within the window is consistent with the original time series order. Each training sample contains a multi-device, multi-measurement point data matrix within the consecutive time window and a global operating condition parameter sequence within the corresponding window.

[0034] Furthermore, step S200 also includes extracting multi-device, multi-measurement point data and corresponding global operating condition parameters within the current sliding time window from the time-series data stream, inputting them into the deep probability reconstruction model, calculating the reconstruction error of the current window; inputting the operating condition parameters of the current window into the quantile regression model to obtain the dynamic threshold under the current operating condition; if the reconstruction error exceeds the dynamic threshold, triggering the potential fault warning, and recording the warning trigger time.

[0035] Preferably, the current sliding time window refers to the time interval covered by a fixed length W time steps backward from the current moment. Multi-device, multi-measurement point data and corresponding global operating condition parameters are extracted from the time-series data stream. This includes a W×D multi-device, multi-measurement point data matrix, where W is the number of time steps within the window, D is the total number of measurement points for all devices, and each element in the matrix represents the value of the corresponding measurement point at the corresponding time step; and a W×K sequence of corresponding global operating condition parameters, where W is the number of time steps within the window, K is the number of operating condition parameter types, and each element in the matrix represents the value of the corresponding operating condition parameter at the corresponding time step. The extracted W×D multi-device, multi-measurement point data matrix and... The W×K global load condition parameter sequence is simultaneously input into the deep probabilistic reconstruction model and processed using an encoder-decoder approach. Each time step injects the corresponding load condition parameters, outputting the probability distribution parameters for each measurement point at each time step within the current window, including the mean and variance of each measurement point at each time step. Then, the output probability distribution parameters are compared with the actual measurement point values ​​within the window. For each measurement point at each time step, the probability density of the actual value appearing in the model's output probability distribution is calculated. The negative logarithm is then summed or averaged to determine the reconstruction error. A larger value indicates that the model has a greater difficulty in reconstructing the current window's data, meaning the current data deviates more from the normal pattern. The load condition features of the current window are extracted from the global load condition parameter sequence, i.e., the average value of each load condition parameter across multiple time steps within the window. This load condition parameter vector is then input into the quantile regression model, outputting the quantile threshold for the reconstruction error under that load condition, corresponding to the quantile set during training. For example, setting the quantile to 0.99 outputs the 99th percentile of the reconstruction error under that load condition in normal historical data.

[0036] Preferably, the reconstruction error calculated for the current window is compared with the dynamic threshold output by the quantile regression model. If the reconstruction error is less than or equal to the dynamic threshold, the current window is considered normal and no warning is triggered. If the reconstruction error is greater than the dynamic threshold, a warning is triggered, indicating that the multi-device, multi-measurement point data pattern in the current window deviates significantly from the data pattern under historical normal conditions. This deviation does not reach the limit of the traditional threshold alarm, but it has deviated from the main area of ​​the normal joint probability distribution under operating conditions. Then, the current time of triggering the warning is recorded as a timestamp as the reference time for extracting the time windows before and after, and is simultaneously recorded in the real-time database or event log. The operating condition parameters are used as inputs to the quantile regression model to achieve dynamic adjustment of the threshold.

[0037] Step S300: Based on the potential fault warning, extract multi-dimensional time-series data within the time window before and after the warning time from the time-series data stream, initiate spatiotemporal causal tracing analysis, and generate a fault propagation chain.

[0038] Step S300 further includes dividing the measurement points into multiple causal analysis regions according to the electrical topology connection relationship of the station, and performing spatiotemporal causal tracing analysis independently in each causal analysis region; based on the multidimensional time series data, calculating the time delay correlation between changes in each measurement point in each causal analysis region, and constructing a directed acyclic causal graph with measurement points as nodes and causal influence relationships as edges using the transfer entropy method; based on the directed acyclic causal graph, identifying nodes with an in-degree of zero as root source measurement points, and generating a fault propagation chain according to the propagation direction of the directed edges.

[0039] Preferably, the electrical topology connections of the power station are obtained, including the wiring method of the collector lines, such as multiple wind turbines or inverters connected to the same collector line via feeders; transformer connections, such as the high-voltage and low-voltage side connections of box-type transformers and step-up transformers; busbar segmentation structure, such as the bus tie switch connections between different busbar segments; and grid connection points, such as the connection point between the power station outlet and the power grid system. Then, based on the electrical topology connections, the measurement points are divided into multiple causal analysis regions. Among them, all equipment and their associated measurement points on the same collector line are assigned to the same causal analysis region; the high-voltage and low-voltage side measurement points of the same transformer are assigned to the same causal analysis region; and all incoming and outgoing line bay measurement points on the same busbar segment are assigned to the same causal analysis region. This further divides the complete set of measurement points into multiple smaller subsets, with strong electrical coupling relationships between the measurement points in each subset.

[0040] Preferably, the change of a measurement point refers to the amount or direction of change of the measurement point value over time. For telemetry data, the change in adjacent time steps is calculated, and for teleindication data, state change events are recorded. The changes of each measurement point are determined, and then the time delay correlation between the changes of each measurement point in each causal analysis area is analyzed, i.e., a fixed time offset. For example, measurement point A changes first, and measurement point B changes after τ time steps. The specific value of the time delay τ is defined according to the sampling interval, and τ is an integer multiple of the time steps. Then, the transfer entropy method is used to quantify the predictive contribution of one time series X to another time series Y, and the time delay is considered in the calculation process. A directed acyclic causal graph is constructed with measurement points as nodes and causal influence relationships as edges. The node identifier includes the device identifier and measurement point type, and the direction of the edge is determined by the asymmetry of the transfer entropy calculation result. Then, based on the directed acyclic causal graph, nodes with an in-degree of zero are identified as root source test points. Fault propagation chains are generated according to the propagation direction of directed edges. Specifically, the in-degree of a node is the number of directed edges pointing to that node. A node with an in-degree of zero indicates that no other test point has a causal impact on that node. In the directed acyclic causal graph, all nodes with an in-degree of zero are identified as potential root source test points. When a potential fault warning is triggered, the node that changes earliest before the warning time is marked as the final root source test point. Starting from the root source test point, the entire directed acyclic graph is traversed along the direction of the directed edges. Each path from the root source test point to the terminal node constitutes a fault propagation chain. Each directed edge is marked with a time delay τ, which represents the lag time steps of the causal impact. Finally, multiple fault propagation chains are generated, each chain containing a test point sequence and a corresponding time delay sequence.

[0041] Furthermore, step S300 also includes: based on the multidimensional time-series data within each causal analysis region, converting continuous telemetry values ​​into symbol sequences and preserving the original symbols for discrete telemetry values, establishing a symbol time series suitable for calculating transfer entropy; setting a maximum time delay step L_max, and for any two measurement points M and N within the causal analysis region, calculating the transfer entropy value from M to N with a delay of τ (τ=1, 2, ..., L_max) and the transfer entropy value from N to M based on the symbol time series; for each ordered pair (M, N), determining the optimal delay. The maximum transfer entropy value is obtained; the maximum transfer entropy value is compared with a preset causal threshold to determine the directed edges. If both bidirectional transfer entropies exceed the preset causal threshold, the two values ​​are compared, and only the direction corresponding to the larger transfer entropy is retained as the directed edge. An initial directed graph is constructed with the measurement point as the node and the directed edges, and directed loops in the initial directed graph are detected. For each detected loop, the transfer entropy values ​​of each edge in the loop are compared, and only the edge with the largest transfer entropy is retained. The remaining edges in the loop are deleted until there are no directed loops in the graph, thus obtaining a directed acyclic causal graph.

[0042] Preferably, the telemetry continuous value sequence of each measuring point in each causal analysis region is converted into a discrete symbol sequence. This includes calculating the median of the measuring point sequence, marking values ​​greater than the median as the symbol "H" (high bit), values ​​less than the median as the symbol "L" (low bit), and values ​​equal to the median as the symbol "E" (equal). After conversion, the original floating-point time series becomes a time series composed of three symbols: H, L, and E. Remote signaling discrete values ​​refer to quantities with discrete states, such as switch states and protection action signals. Their original state symbols are kept unchanged. For example, the closed state is marked as "1" and the open state is marked as "0". After conversion, the time series of the measuring point is a time series composed of two symbols: 0 and 1. Thus, a symbol time series suitable for transfer entropy calculation is established.

[0043] Preferably, the maximum time delay step L_max is a pre-set positive integer, representing the maximum time lag step to consider causal effects. For example, if the sampling interval is 5 seconds and L_max = 10, then a maximum time delay of 50 seconds is considered. The specific value is determined according to the device's dynamic response time. The ordered pair (M, N) represents the direction from measurement point M to measurement point N, where M is a candidate cause and N is a candidate result. For all measurement points in the causal analysis area, pairwise combinations are used to calculate the transfer entropy in both directions. The transfer entropy from M to N with a delay of τ is calculated as the amount of additional information about the future symbol of N that M can provide from the symbol τ time steps ago, given the current symbol of N. Similarly, the transfer entropy values ​​under each delay from N to M are calculated. For the ordered pair (M, N), the L_max transfer entropy values ​​from τ = 1 to L_max are compared, and the τ with the largest transfer entropy value is selected as the optimal delay. The transfer entropy value corresponding to the optimal delay is recorded as the maximum transfer entropy value, representing the causal influence strength of M on N under the optimal delay.

[0044] Preferably, a preset causal threshold is used to determine whether a causal relationship exists. A transit entropy value greater than or equal to the threshold indicates the existence of a causal influence. The maximum transit entropy value is compared with the preset causal threshold to determine directed edges. If both bidirectional transit entropies exceed the preset causal threshold, their magnitudes are compared, and only the direction corresponding to the larger transit entropy is retained as a directed edge. Then, with the measurement point as the node, an initial directed graph is constructed with directed edges, and a depth-first traversal algorithm is used to mark the access status of each node. When a node already in the current search path is encountered during the traversal, a cycle is detected, and the directed cycle in the initial directed graph is determined. For each detected cycle, the transit entropy values ​​of each edge in the cycle are compared, and only the edge with the largest transit entropy is retained. The remaining edges in the cycle are deleted, and the cycle detection and elimination steps are repeated until there are no directed cycles in the graph, resulting in a directed acyclic causal graph. In this graph, starting from any node and moving along the direction of the edge, it is impossible to return to that node.

[0045] Furthermore, step S300 also includes collecting normal operation data within a preset time window before the current warning trigger time, and complete multi-dimensional time series data within a preset time window after the warning trigger time; obtaining the final fault cause marked by operation and maintenance, and associating and binding the final fault cause with the root cause measurement point in the directed acyclic causal graph; using the normal operation data within the preset time window as new normal samples, incrementally learning and fine-tuning the deep probability reconstruction model, and updating the model parameters to adapt to the evolution of operating conditions; and based on the collected complete multi-dimensional time series data and the final fault cause marked by operation and maintenance, incrementally training the quantile regression model and adjusting the fitting surface of the dynamic threshold.

[0046] Preferably, normal operation data within a preset time window prior to the warning trigger time is collected, and all multi-device, multi-measurement point data and global operating condition parameter sequences within that window are extracted as new normal samples. Then, all process data from the warning trigger time to the fault development and maintenance handling completion within that window are obtained, including the complete process of the fault from its initial occurrence to its evolution after the warning. The final fault cause marked by maintenance is obtained, including the faulty device identifier, fault type, fault occurrence time, fault handling record, etc. A one-to-one correspondence is established between the final fault cause marked by maintenance and the root cause measurement point. For example, a related record is recorded in the database or knowledge base, including the root cause measurement point identifier, the final fault cause description, the warning trigger time, and the fault confirmation time.

[0047] Preferably, the normal operation data within a preset time window undergoes missing value imputation, time-aligned resampling, and normalization preprocessing. Then, a sliding window is used to generate multiple training samples for incremental learning fine-tuning of the deep probabilistic reconstruction model. This includes retaining the current model parameters as initial values, constructing mini-batch training data using newly added normal samples, setting a small learning rate, and performing a limited number of training rounds to minimize the negative log-likelihood loss function, thereby updating the model parameters to adapt to evolving operating conditions. Based on the collected complete multi-dimensional time-series data and the final fault causes labeled by operations and maintenance, the quantile regression model is further refined. In the incremental training, specifically, the data from the current warning and subsequent faults are used as new training samples. The reconstruction error value is calculated and determined. Based on the original quantile regression model parameters, incremental updates are performed to learn the quantile surface of the reconstruction error as the operating condition parameters change. The fitting surface of the dynamic threshold is adjusted. For example, if the current warning is confirmed as a real fault by the operation and maintenance personnel, the reconstruction error of this sample is above the current threshold surface. After incremental training, the threshold surface is slightly raised near this operating condition. If the current warning is confirmed as a false alarm by the operation and maintenance personnel, the reconstruction error of this sample may be higher due to special reasons. After incremental training, the threshold surface is slightly lower near this operating condition.

[0048] Step S400: Visualize the directed graph based on the fault propagation chain, and generate an explanatory warning report containing the root cause device, fault propagation path, and handling suggestions based on the causal relationship.

[0049] Preferably, the fault propagation chain is visualized as a directed graph. Each node represents a measurement point of the device, and the device identifier and measurement point type are marked on the node. Directed edges connect two nodes with arrowed line segments, and the direction of the arrow indicates the direction of causal influence, i.e., from the cause measurement point to the result measurement point. The edge attributes include time delay values. The graph is displayed graphically on the human-machine interface of the monitoring system. The node corresponding to the root device is identified by a color or shape that is different from other nodes, and they are arranged from left to right or from top to bottom according to the time sequence of propagation. The graph also supports interactive operations such as zooming, panning, and node highlighting. When there are multiple fault propagation chains, multiple directed graphs can be displayed or they can be displayed as subgraphs of different colors in the same directed graph. An explanatory early warning report is generated based on causal relationships, including the root cause device, the fault propagation path, and handling suggestions. The root cause device is the device information corresponding to the node with zero in-degree in the fault propagation chain, clearly defining the root cause device identifier, specific measurement point type, and deviation. The fault propagation path is the devices and measurement points listed sequentially according to the direction of the edges in the directed graph, with each step marked with the time delay of the causal impact and describing the specific changes in the intermediate links. The handling suggestions are corresponding disposal measures matched from a preset knowledge base based on the type and fault characteristics of the root cause device, including at least the name of the equipment component to be checked, the suggested operation actions, such as "shutdown inspection", "remote reset", "power-limited operation", and the urgency level classification, such as "immediate handling", "handling within 2 hours", and "planned maintenance".

[0050] In the above text, refer to Figure 1 The method for monitoring and alarming the operation of new energy power station equipment according to embodiments of the present invention is described in detail. Next, we will refer to... Figure 2 This invention describes an operation monitoring and alarm system for new energy power station equipment according to an embodiment of the present invention.

[0051] The operation monitoring and alarm system for new energy power station equipment according to embodiments of the present invention addresses the technical problems in existing technologies, such as threshold-based alarms leading to delayed early warnings, difficulty in automatically identifying fault root causes from multi-device associated alarms, and a lack of spatiotemporal causal reasoning and visual interpretation capabilities. It achieves the technical effects of improving early potential fault perception capabilities, fault root cause location accuracy, alarm interpretability, and emergency response efficiency. Figure 2 As shown, the operation monitoring and alarm system for new energy power station equipment includes: a data acquisition module 10, a fault early warning module 20, a fault propagation chain generation module 30, and an early warning report generation module 40.

[0052] The data acquisition module 10 is used to collect multi-dimensional telemetry and teleindication data from multiple devices within the site, construct a time-series data stream containing timestamps, device identifiers, measurement point types, and numerical values, and store it in a real-time database. The fault early warning module 20 is used to train a deep probabilistic reconstruction model based on historical normal operation data. The deep probabilistic reconstruction model learns the joint probability distribution between measurement points of each device. During real-time monitoring, it extracts the data vectors of all relevant measurement points at the current moment from the time-series data stream and inputs them into the model to calculate the reconstruction error. Based on the reconstruction error, it triggers a potential fault early warning and records the current early warning trigger timestamp. The fault propagation chain generation module 30 is used to extract multi-dimensional time-series data within the time window before and after the early warning moment from the time-series data stream according to the potential fault early warning, initiates spatiotemporal causal tracing analysis, and generates a fault propagation chain. The early warning report generation module 40 is used to perform directed graph visualization based on the fault propagation chain and generate an explanatory early warning report containing the root cause device, fault propagation path, and handling suggestions based on causal relationships.

[0053] The specific configuration of the fault early warning module 20 will be described in detail below. The fault early warning module 20 further includes: acquiring historical data covering multiple operating conditions during normal operation to construct a training dataset; training the deep probabilistic reconstruction model using the training dataset, wherein the input is multi-device, multi-measurement point data within a sliding time window and the global operating condition parameter sequence within the corresponding window, and the output is the probability distribution parameters of each measurement point within the corresponding window, learning the joint probability distribution between measurement points under given operating conditions; after training, calculating the reconstruction error for each training sample, and training a quantile regression model using the reconstruction error as the dependent variable and the corresponding sample's operating condition parameters as the independent variable, to dynamically generate a reconstruction error threshold associated with real-time operating conditions.

[0054] The specific configuration of the fault early warning module 20 will be described in detail below. The fault early warning module 20 further includes: the deep probabilistic reconstruction model adopts an encoder-decoder architecture. The encoder consists of a first long short-term memory network layer and a first fully connected layer, used to map the input window sequence to the mean and logarithmic variance of latent variables; the decoder consists of a second fully connected layer and a second long short-term memory network layer, used to reconstruct the probability distribution of each measurement point from the latent variables; wherein, by injecting global operating condition parameters into the encoder and decoder step-by-step, the deep probabilistic reconstruction model learns the probability distribution of the measurement point data conditioned by the operating condition parameters, enabling dynamic adjustment of the baseline for judging the same measurement point data as normal or abnormal under different operating conditions.

[0055] The specific configuration of the fault warning module 20 will be described in detail below. The fault warning module 20 further includes: an encoder for mapping the input window sequence to the mean and log-variance of latent variables, comprising: the first long short-term memory network layer receiving a multi-device, multi-measurement point data matrix within a sliding time window, and inputting the current global operating condition parameters as conditional features concatenated with the measurement point data at each time step; the first fully connected layer mapping the output of the first long short-term memory network layer to the mean vector and log-variance vector of the latent variables; and a decoder for reconstructing the probability distribution of each measurement point from the latent variables, comprising: the second fully connected layer mapping the latent variable sample values ​​back to the initial reconstructed features; the second long short-term memory network layer concatenating the current global operating condition parameters as conditional features with the reconstructed features at each time step, and outputting the probability distribution parameters of each measurement point at this time step step by step.

[0056] The specific configuration of the fault early warning module 20 will be described in detail below. The fault early warning module 20 further includes: extracting time period data from the historical database that has been confirmed by operation and maintenance to have no fault alarms and covers a variety of typical operating conditions, including wind speed, irradiance, ambient temperature and grid connection point frequency; after performing missing value imputation, time alignment resampling and normalization preprocessing on the extracted normal operation data, training samples are generated using a sliding window method. Each sample contains a multi-device multi-measurement point data matrix within a continuous time window and a global operating condition parameter sequence within the corresponding window.

[0057] The specific configuration of the fault warning module 20 will be described in detail below. The fault warning module 20 further includes: extracting multi-device, multi-measurement point data and corresponding global operating condition parameters within the current sliding time window from the time-series data stream, inputting them into the deep probability reconstruction model, and calculating the reconstruction error of the current window; inputting the operating condition parameters of the current window into the quantile regression model to obtain a dynamic threshold under the current operating condition; if the reconstruction error exceeds the dynamic threshold, triggering the potential fault warning and recording the warning trigger time.

[0058] The specific configuration of the fault propagation chain generation module 30 will be described in detail below. The fault propagation chain generation module 30 further includes: dividing the measurement points into multiple causal analysis regions based on the electrical topology of the station; performing spatiotemporal causal tracing analysis independently in each causal analysis region; calculating the time delay correlation between changes in each measurement point in each causal analysis region based on the multidimensional time-series data; constructing a directed acyclic causal graph with measurement points as nodes and causal influence relationships as edges using the transfer entropy method; identifying nodes with an in-degree of zero as root measurement points based on the directed acyclic causal graph; and generating a fault propagation chain based on the propagation direction of the directed edges.

[0059] The specific configuration of the fault propagation chain generation module 30 will be described in detail below. The fault propagation chain generation module 30 further includes: based on multi-dimensional time-series data within each causal analysis region, converting continuous telemetry values ​​into symbol sequences while preserving the original symbols for discrete telemetry values, and establishing a symbol time series suitable for propagation entropy calculation; setting a maximum time delay step L_max, and for any two measurement points M and N within the causal analysis region, calculating the propagation entropy value from M to N with a delay of τ (τ=1, 2, ..., L_max), and the propagation entropy value from N to M, respectively, based on the symbol time series; for each ordered pair (M, N), determining the optimal delay... The maximum transfer entropy value is determined by comparing the maximum transfer entropy value with a preset causal threshold to identify directed edges. If both bidirectional transfer entropies exceed the preset causal threshold, the two values ​​are compared, and only the direction corresponding to the larger transfer entropy is retained as a directed edge. An initial directed graph is constructed using the measured point as a node and the directed edges, and directed loops in the initial directed graph are detected. For each detected loop, the transfer entropy values ​​of each edge in the loop are compared, and only the edge with the largest transfer entropy is retained. The remaining edges in the loop are deleted until there are no directed loops in the graph, thus obtaining a directed acyclic causal graph.

[0060] The specific configuration of the fault propagation chain generation module 30 will be described in detail below. The fault propagation chain generation module 30 further includes: collecting normal operation data within a preset time window before the current warning trigger time, and complete multi-dimensional time-series data within a preset time window after the warning trigger time; obtaining the final fault cause labeled by operations and maintenance, and associating the final fault cause with the root cause measurement point in the directed acyclic causal graph; using the normal operation data within the preset time window as new normal samples, incrementally learning and fine-tuning the deep probabilistic reconstruction model, updating the model parameters to adapt to the evolution of operating conditions; and based on the collected complete multi-dimensional time-series data and the final fault cause labeled by operations and maintenance, incrementally training the quantile regression model, adjusting the fitting surface of the dynamic threshold.

[0061] The operation monitoring and alarm system for new energy power station equipment provided in this embodiment of the invention can execute the operation monitoring and alarm method for new energy power station equipment provided in any embodiment of the invention, and has the corresponding functional modules and beneficial effects of the execution method.

[0062] The above description is merely a preferred embodiment of the present invention and is not intended to limit the present invention in any way. Although the present invention has been disclosed above with reference to preferred embodiments, it is not intended to limit the present invention. Any person skilled in the art can make some modifications or alterations to the above-disclosed technical content to create equivalent embodiments without departing from the scope of the present invention. Any modifications, equivalent changes, and alterations made to the above embodiments based on the technical essence of the present invention without departing from the scope of the present invention shall still fall within the scope of the present invention.

Claims

1. A method for monitoring and alarming the operation of new energy power station equipment, characterized in that, include: Collect multi-dimensional telemetry and teleindication data from multiple devices within the site, construct a time-series data stream containing timestamps, device identifiers, measurement point types, and numerical values, and store it in a real-time database; Based on historical normal operation data, a deep probabilistic reconstruction model is trained. The deep probabilistic reconstruction model learns the joint probability distribution between each measuring point of the device. During real-time monitoring, the data vectors of all relevant measuring points at the current moment are extracted from the time-series data stream and input into the model to calculate the reconstruction error. Based on the reconstruction error, a potential fault warning is triggered, and the current warning trigger timestamp is recorded. Based on the potential fault warning, multi-dimensional time-series data within the time window before and after the warning time is extracted from the time-series data stream, and spatiotemporal causal tracing analysis is initiated to generate a fault propagation chain. The fault propagation chain is visualized using a directed graph, and an explanatory warning report containing the root cause device, fault propagation path, and handling suggestions is generated based on the causal relationship.

2. The method for monitoring and alarming the operation of new energy power station equipment according to claim 1, characterized in that, Based on historical normal operation data, a deep probabilistic reconstruction model is trained, including: Acquire historical data that is in normal operation and covers multiple working conditions during different time periods, and construct a training dataset; The deep probabilistic reconstruction model is trained using the training dataset, wherein the input is multi-device, multi-measurement point data within a sliding time window and the global operating condition parameter sequence within the corresponding window, and the output is the probability distribution parameters of each measurement point within the corresponding window, thereby learning the joint probability distribution among each measurement point under a given operating condition. After training is completed, the reconstruction error is calculated for each training sample, and a quantile regression model is trained with the reconstruction error as the dependent variable and the corresponding sample's operating condition parameters as the independent variables. This model is used to dynamically generate a reconstruction error threshold associated with real-time operating conditions.

3. The method for monitoring and alarming the operation of new energy power station equipment according to claim 2, characterized in that, The deep probabilistic reconstruction model adopts an encoder-decoder architecture. The encoder consists of a first long short-term memory network layer and a first fully connected layer, which is used to map the input window sequence to the mean and log-variance of the latent variables. The decoder consists of a second fully connected layer and a second long short-term memory network layer, which is used to reconstruct the probability distribution of each measurement point from the latent variables. Specifically, by injecting global operating condition parameters into the encoder and decoder at each time step, the deep probabilistic reconstruction model learns the probability distribution of measurement point data conditioned on the operating condition parameters, thereby dynamically adjusting the benchmark to determine whether the same measurement point data is normal or abnormal under different operating conditions.

4. The operation monitoring and alarm method for new energy power station equipment according to claim 3, characterized in that, The encoder is used to map the input window sequence to the mean and log-variance of latent variables, including: the first long short-term memory network layer receives a multi-device, multi-measurement point data matrix within a sliding time window, and at each time step, it concatenates the global operating parameters at the current time as conditional features with the measurement point data and inputs them; the first fully connected layer maps the output of the first long short-term memory network layer to the mean vector and log-variance vector of latent variables; The decoder is used to reconstruct the probability distribution of each measurement point from the latent variables, including: the second fully connected layer maps the latent variable sample values ​​back to the initial reconstruction features, and the second long short-term memory network layer concatenates the global operating condition parameters at the current time step as conditional features with the reconstruction features at each time step, and outputs the probability distribution parameters of each measurement point at the current time step by time step.

5. The method for monitoring and alarming the operation of new energy power station equipment according to claim 2, characterized in that, Construct the training dataset, including: Extract time period data from the historical database that has been confirmed by operation and maintenance to have no fault alarms and covers a variety of typical operating conditions, including wind speed, irradiance, ambient temperature and grid connection frequency. After performing missing value imputation, time-aligned resampling, and normalization preprocessing on the extracted normal operation data, training samples are generated using a sliding window approach. Each sample contains a multi-device, multi-measurement point data matrix within a continuous time window and a global operating condition parameter sequence within the corresponding window.

6. The method for monitoring and alarming the operation of new energy power station equipment according to claim 2, characterized in that, Early warnings of potential faults triggered by reconstruction errors include: Extract multi-device, multi-measurement point data and corresponding global operating condition parameters within the current sliding time window from the time-series data stream, input them into the depth probability reconstruction model, and calculate the reconstruction error of the current window; Input the operating parameters of the current window into the quantile regression model to obtain the dynamic threshold under the current operating conditions; If the reconstruction error exceeds the dynamic threshold, the potential fault warning is triggered, and the warning trigger time is recorded.

7. The method for monitoring and alarming the operation of new energy power station equipment according to claim 2, characterized in that, Based on the potential fault warning, multi-dimensional time-series data within the time window before and after the warning time are extracted from the time-series data stream. Spatiotemporal causal tracing analysis is initiated to generate a fault propagation chain, including: Based on the electrical topology of the station, the measurement points are divided into multiple causal analysis regions, and spatiotemporal causal tracing analysis is performed independently in each causal analysis region; Based on the multidimensional time series data, the time delay correlation between the changes of each measurement point in each causal analysis region is calculated respectively. The transfer entropy method is used to construct a directed acyclic causal graph with measurement points as nodes and causal influence relationships as edges. Based on the directed acyclic causal graph, nodes with an in-degree of zero are identified as root source test points, and a fault propagation chain is generated according to the propagation direction of the directed edges.

8. The method for monitoring and alarming the operation of new energy power station equipment according to claim 7, characterized in that, A directed acyclic causal graph is constructed using the transitive entropy method, with measurement points as nodes and causal relationships as edges, including: Based on multidimensional time series data within each causal analysis region, telemetry continuous values ​​are converted into symbol sequences, while remote signaling discrete values ​​retain their original symbols, thus establishing a symbol time series suitable for transfer entropy calculation. Set the maximum time delay step L_max. For any two measurement points M and N in the causal analysis region, calculate the transfer entropy value from M to N with a delay of τ (τ=1,2,L_max) and the transfer entropy value from N to M based on the symbol time series. For each ordered pair (M,N), determine the optimal delay and the corresponding maximum propagation entropy value; The maximum transit entropy value is compared with a preset causal threshold to determine the directed edges. If both the bidirectional transit entropy values ​​exceed the preset causal threshold, the two values ​​are compared, and only the direction corresponding to the larger transit entropy is retained as the directed edge. Using the measurement points as nodes, an initial directed graph is constructed with directed edges. Directed cycles in the initial directed graph are detected. For each detected cycle, the propagation entropy values ​​of each edge in the cycle are compared. Only the edge with the largest propagation entropy is retained, and the remaining edges in the cycle are deleted until there are no directed cycles in the graph, thus obtaining a directed acyclic causal graph.

9. The method for monitoring and alarming the operation of new energy power station equipment according to claim 7, characterized in that, Also includes: Collect normal operation data within a preset time window before the warning trigger time, and complete multi-dimensional time series data within a preset time window after the warning trigger time; Obtain the final fault cause marked by the operation and maintenance, and associate and bind the final fault cause with the root source measurement point in the directed acyclic causal graph; The normal operation data within the preset time window is used as new normal samples to perform incremental learning and fine-tuning on the deep probability reconstruction model, and the model parameters are updated to adapt to the evolution of working conditions. Based on the collected complete multidimensional time-series data and the final fault causes marked by operation and maintenance, the quantile regression model is incrementally trained to adjust the fitting surface of the dynamic threshold.

10. An operation monitoring and alarm system for new energy power station equipment, characterized in that, The system is used to implement the operation monitoring and alarm method for new energy power station equipment according to any one of claims 1 to 9, and the system includes: The data acquisition module is used to collect multi-dimensional telemetry and teleindication data from multiple devices within the site, construct a time-series data stream containing timestamps, device identifiers, measurement point types, and numerical values, and store it in a real-time database. The fault early warning module is used to train a deep probabilistic reconstruction model based on historical normal operation data. The deep probabilistic reconstruction model learns the joint probability distribution between various measurement points of the device. During real-time monitoring, the data vectors of all relevant measurement points at the current moment are extracted from the time-series data stream and input into the model to calculate the reconstruction error. Based on the reconstruction error, a potential fault early warning is triggered, and the current early warning trigger timestamp is recorded. The fault propagation chain generation module is used to extract multi-dimensional time-series data within the time window before and after the warning time from the time-series data stream based on the potential fault warning, initiate spatiotemporal causal tracing analysis, and generate a fault propagation chain. The early warning report generation module is used to visualize the fault propagation chain in a directed graph and generate an explanatory early warning report based on causal relationships, including the root cause device, the fault propagation path, and handling suggestions.

Images