New energy power prediction automatic diagnosis and alarm attribution method based on ai closed loop

Abstract

The application belongs to the technical field of new energy power generation operation and maintenance, and relates to a new energy power prediction automatic diagnosis and alarm attribution method based on an AI closed loop. The method comprises the following steps: obtaining historical operation data and real-time operation data of each monitoring point of a target station, and respectively performing standardization processing to obtain a historical time series data set and a real-time time series data set; establishing a numerical fluctuation range of each monitoring point according to the data of each monitoring point in the normal working condition in the historical time series data set, as a dynamic judgment benchmark for judging power prediction abnormalities; comparing the data of each monitoring point in the real-time time series data set with the numerical fluctuation range of the corresponding monitoring point in the dynamic judgment benchmark, and if the power prediction is abnormal, extracting the abnormal features corresponding to the data; obtaining the cause type corresponding to the abnormal features according to the abnormal features; obtaining the treatment mode corresponding to the cause type according to the cause type, and executing the treatment mode. The application significantly improves the abnormal response speed and attribution accuracy.

Description

Technical Field

[0001] This invention belongs to the field of new energy power generation operation and maintenance technology, and relates to a new energy power prediction automatic diagnosis and alarm attribution method based on AI closed loop. Background Technology

[0002] Power forecasting at renewable energy power plants is a core basis for grid dispatching and power trading, and its accuracy directly affects grid safety and the economic benefits of the power plants. Currently, renewable energy power plants generally adopt an open-loop power forecasting model that combines a fixed statistical model with manual parameter adjustment. The basic process of this model is as follows: when the forecast deviation exceeds the limit, an alarm is triggered, maintenance personnel manually investigate the cause, then manually adjust the parameters, and repeat the above process after the next anomaly occurs. This model heavily relies on the personal experience of senior engineers, resulting in high maintenance costs and slow response times.

[0003] Furthermore, existing power prediction systems can only report deviations exceeding limits, unable to distinguish whether the deviation stems from numerical weather prediction (NWP) bias, data latency, equipment malfunction, or acquisition failure. Maintenance personnel typically require 0.5 to 3 hours to troubleshoot each issue. Adjustments to time-series compensation parameters, multi-source data fusion weights, and characteristic parameters rely on engineers manually based on experience, resulting in long training cycles. New personnel often struggle to achieve the same level of optimization within a short period after staff turnover, leading to significant fluctuations in prediction accuracy. Traditional prediction models retain fixed parameters once set, failing to adapt to dynamic changes such as NWP data arrival delays, sensor refresh rates, and equipment response times. This inaccuracy is particularly pronounced under extreme weather conditions. Summary of the Invention

[0004] To address the problems in existing technologies, this invention provides an AI-based closed-loop method for automatic diagnosis and alarm attribution of new energy power prediction, which significantly improves the anomaly response speed and attribution accuracy.

[0005] To achieve the above objectives, the present invention employs the following technical solution: In a first aspect, the present invention provides an automatic diagnosis and alarm attribution method for new energy power prediction based on AI closed-loop, comprising the following steps: S1. Obtain historical and real-time operational data from each monitoring point of the target station, and standardize the historical and real-time operational data respectively to obtain historical time-series datasets and real-time time-series datasets. S2. Based on the data of each monitoring point in the historical time series dataset under normal operating conditions, establish the numerical fluctuation range of each monitoring point as a dynamic judgment benchmark for determining power prediction anomalies. S3. Compare the data of each monitoring point in the real-time time series dataset with the numerical fluctuation range of the corresponding monitoring point in the dynamic judgment benchmark. If there is data in the real-time time series dataset that exceeds the numerical fluctuation range of the corresponding monitoring point, it is determined that the power prediction is abnormal, and the abnormal features corresponding to the data are extracted. S4. Obtain the corresponding cause type based on the abnormal characteristics; S5. Obtain the corresponding treatment method according to the cause type, and execute the treatment method.

[0006] Preferably, the monitoring points of the target site include one or more of the following: numerical weather forecast data monitoring points, wind tower measured data monitoring points, irradiance meter measured data monitoring points, local micro-meteorological monitoring points, wind turbine operation status monitoring points, inverter operation status monitoring points, and SCADA communication status monitoring points; the historical operation data and real-time operation data include meteorological data, equipment operation data, acquisition status data, and scheduling rule data.

[0007] Preferably, the standardization process performed on the historical running data and the real-time running data to obtain historical time-series datasets and real-time time-series datasets includes: The historical and real-time operational data are processed with a unified time stamp to align data from different sources to the same time base. Outlier removal is performed on the data after standardization of time scale. Abnormal records are identified and deleted based on the preset threshold ranges for each physical quantity. After removing outliers, missing values ​​were filled in using piecewise linear interpolation to fill in the missing time periods. The completed data is normalized to eliminate differences between different units. The processed data is arranged in chronological order to form the historical time-series dataset and the real-time time-series dataset.

[0008] Preferably, the step of establishing the numerical fluctuation range of each monitoring point based on the data of each monitoring point in the historical time series dataset under normal operating conditions, as a dynamic judgment benchmark for determining power prediction anomalies, includes: Data samples of each monitoring point under normal operating conditions are extracted from the historical time series dataset. Normal operating conditions refer to the operating conditions under which the station equipment is running normally, there are no fault alarms, unlimited power commands, and communication links are unobstructed. Statistical analysis was performed on the normal operating condition data samples of each monitoring point to calculate the upper confidence limit and lower confidence limit of that monitoring point; The upper and lower confidence limits of each monitoring point constitute the numerical fluctuation range of that monitoring point, and the numerical fluctuation ranges of all monitoring points together constitute the dynamic judgment benchmark.

[0009] Preferably, the step of comparing the data of each monitoring point in the real-time time series dataset with the numerical fluctuation range of the corresponding monitoring point in the dynamic judgment benchmark, and determining that the power prediction is abnormal if there is data in the real-time time series dataset that exceeds the numerical fluctuation range of the corresponding monitoring point, and extracting the abnormal features corresponding to the data, includes: The real-time data of each monitoring point in the real-time time series dataset is compared with the upper and lower confidence limits of the corresponding monitoring points in the dynamic judgment benchmark. If the real-time data of a certain monitoring point exceeds the upper confidence limit or falls below the lower confidence limit of that monitoring point, a preliminary anomaly alarm will be triggered. After triggering the initial anomaly alarm, a dual verification mechanism is activated: the real-time data is verified using a preset static threshold. If the real-time data exceeds both the static threshold and the dynamic judgment benchmark, the power prediction is ultimately determined to be a real anomaly. Data features of monitoring points where real anomalies occur are extracted as the anomaly features, which include at least the time of anomaly occurrence, anomaly monitoring point identifier, anomaly value, and degree of deviation.

[0010] Preferably, obtaining the corresponding cause type based on the abnormal features includes: A root cause knowledge graph of deviations is constructed, wherein the nodes of the knowledge graph include at least one or more of the following causal nodes: NWP source deviation node, local meteorological change node, equipment derating node, data acquisition interruption node, communication delay node, and limited issuance instruction node; the edges of the knowledge graph represent the association relationships between the causal nodes. Parallel retrieval is performed in the knowledge graph using the abnormal monitoring point identifiers and abnormal values ​​in the abnormal features as retrieval conditions. Match the corresponding causal nodes based on the search results, and output the causal type and corresponding confidence level.

[0011] Preferably, after obtaining the corresponding treatment method based on the cause type and executing the treatment method, the method further includes: After the above handling method is completed, check whether the real-time data of the abnormal monitoring point returns to the value fluctuation range of the dynamic judgment benchmark. If regression occurs, the loop closure is confirmed to be eliminated, and the entire chain processing information from the detection of the anomaly to the elimination of the loop is used as feedback data and input into the dynamic judgment benchmark and / or the knowledge graph to update the fluctuation range of the dynamic judgment benchmark and / or the association rules and confidence parameters of the knowledge graph. If no regression occurs, the result of this action will be treated as a newly added abnormal feature, and the attribution analysis will be re-executed in step S4. Then, step S5 will be executed again based on the newly acquired cause type until the abnormality is eliminated or the preset iteration termination condition is met.

[0012] Secondly, this invention provides an AI-based closed-loop system for automatic diagnosis and alarm attribution of new energy power prediction, comprising: Data processing module: used to acquire historical and real-time operational data of each monitoring point of the target site, and to standardize the historical and real-time operational data to obtain historical time-series datasets and real-time time-series datasets respectively; Dynamic benchmark establishment module: used to establish the numerical fluctuation range of each monitoring point based on the data of each monitoring point in the historical time series dataset under normal operating conditions, as a dynamic judgment benchmark for judging power prediction anomalies; Anomaly detection module: used to compare the data of each monitoring point in the real-time time series dataset with the numerical fluctuation range of the corresponding monitoring point in the dynamic judgment benchmark. If there is data in the real-time time series dataset that exceeds the numerical fluctuation range of the corresponding monitoring point, it is determined that the power prediction is abnormal and the abnormal features corresponding to the data are extracted. Attribution diagnosis module: used to obtain the corresponding cause type based on the abnormal features; Hierarchical closed-loop treatment module: used to obtain the corresponding treatment method according to the cause type, and execute the treatment method.

[0013] Thirdly, the present invention provides a computer device, including a memory, a processor, and a computer program stored in the memory and executable on the processor, wherein the processor executes the computer program to implement the steps of an AI-based closed-loop method for automatic diagnosis and alarm attribution of new energy power prediction.

[0014] Fourthly, the present invention provides a computer-readable storage medium storing a computer program, wherein when the computer program is executed by a processor, it implements the steps of an AI-based closed-loop method for automatic diagnosis and alarm attribution of new energy power prediction.

[0015] Compared with the prior art, the present invention has the following beneficial effects: Standardized processing effectively solves the problems of inconsistent quality and time scales of multi-source heterogeneous data; an adaptive dynamic judgment benchmark with upper and lower confidence limits is established based on historical statistics under normal operating conditions, replacing the traditional fixed threshold, so that anomaly judgment can adapt to individual differences and seasonal time-varying patterns of different stations; the combination of dynamic benchmark and preset static threshold constitutes a dual verification mechanism, which significantly reduces the false alarm rate caused by noise interference while ensuring the sensitivity of anomaly detection; with the help of knowledge graph, parallel retrieval of anomaly features and rapid root cause matching are realized, and quantitative confidence scores are output; differentiated treatment strategies are automatically matched according to the cause type, and the closed-loop elimination is verified by effect back-checking after execution, and the end-to-end processing information is fed back to dynamic benchmark and knowledge graph to achieve self-learning evolution. Attached Figure Description

[0016] To more clearly illustrate the technical solutions of the embodiments of the present invention, the accompanying drawings used in the embodiments will be briefly introduced below. It should be understood that the following drawings only show some embodiments of the present invention and should not be regarded as a limitation on the scope. For those skilled in the art, other related drawings can be obtained based on these drawings without creative effort.

[0017] Figure 1 This is a flowchart of the method of the present invention. Detailed Implementation

[0018] The present invention will now be described in detail with reference to the accompanying drawings and embodiments. It should be noted that, unless otherwise specified, the embodiments and features described in this application can be combined with each other.

[0019] The following detailed description is exemplary and intended to provide further detailed explanation of the invention. Unless otherwise specified, all technical terms used in this invention have the same meaning as commonly understood by one of ordinary skill in the art to which this application pertains. The terminology used in this invention is for the purpose of describing particular embodiments only and is not intended to limit the scope of exemplary embodiments according to the invention.

[0020] The primary objective of this invention is to provide an AI-based closed-loop method for automatic diagnosis and alarm attribution of new energy power prediction, such as... Figure 1 As shown, it includes the following steps: S1. Obtain historical and real-time operational data from each monitoring point of the target station, and standardize the historical and real-time operational data to obtain historical time-series datasets and real-time time-series datasets.

[0021] Acquire historical and real-time operational data from each monitoring point at the target site, perform unified time-stamping on the historical and real-time operational data, and align data from different sources to the same time reference. Outlier removal is performed on the data after standardization of time scale. Abnormal records are identified and deleted based on the preset threshold ranges for each physical quantity. After removing outliers, missing values ​​were filled in using piecewise linear interpolation to fill in the missing time periods. The completed data is normalized to eliminate differences between different units. The processed data is arranged in chronological order to form the historical time-series dataset and the real-time time-series dataset.

[0022] Specifically, it includes: The system acquires historical and real-time operational data from various monitoring points at the target site. These monitoring points broadly cover one or more of the following: numerical weather prediction data monitoring points, wind tower measured data monitoring points, irradiance measured data monitoring points, local micrometeorological monitoring points at the site, wind turbine operation status monitoring points, inverter operation status monitoring points, and SCADA communication status monitoring points. Correspondingly, the collected historical and real-time operational data encompasses meteorological data, equipment operation data, data acquisition status data, and dispatch market data, specifically including: NWP numerical weather prediction wind speed, irradiance, temperature, and air quality data. Pressure; measured wind speed from the meteorological tower and measured irradiance from the photovoltaic irradiance meter; real-time data from the micro-meteorological sensors at the site; wind turbine output, inverter active power, unit power limiting instructions, equipment derating factor, fault codes, and start / stop status; SCADA sampling interval, data transmission delay, data packet loss rate, and communication link interruption markers; deviation assessment thresholds for the "two detailed rules" of the power grid, time-of-use spot electricity prices, assessment penalty calculation rules, and the maximum allowable prediction deviation range, thereby ensuring that no information is omitted from the entire chain of information from the meteorological source to the equipment terminal and then to the dispatch boundary, providing a complete data foundation for subsequent diagnosis.

[0023] When synchronously accessing the above four types of time-series data, the unified timestamp granularity supports adjustable values ​​of 5s, 1min, 15min, and 1h to adapt to the data acquisition capabilities and business needs of different sites.

[0024] Based on this, the above-mentioned multi-source data undergoes systematic standardization processing, specifically as follows: (1) Unified Time Stamping: Data from different sources and with different sampling frequencies are synchronized to the same time base through time interpolation or timestamp alignment to eliminate timing misalignment caused by data acquisition delay and clock deviation. When data packet loss or transmission delay exists, linear interpolation is used to complete the timing at the same moment. The interpolation formula is:

[0025] in, For a moment The interpolated data after completion; To standardize and unify time scales; and for The timestamps of the two most recent valid data entries; , They are respectively and The original measured data at each moment.

[0026] (2) Outlier removal: Based on the physical meaning of each physical quantity and the reasonable threshold range preset by engineering experience, identify and delete abnormal records that obviously exceed the equipment range or physical limits to prevent extreme noise from contaminating the subsequent statistical benchmark.

[0027] (3) Missing value completion: For missing time periods after removing outliers, piecewise linear interpolation is used to fill them. The linear relationship between valid data points before and after the missing time period is used to reasonably estimate the filling value. This method can follow the local change trend of data well while ensuring computational efficiency, and avoid introducing excessive filling deviation.

[0028] (4) Normalization: The original data have extremely different dimensions (wind speed m / s, power MW, electricity price yuan / MWh), so it is necessary to eliminate the interference of these dimensional differences on the weights of subsequent model calculations and statistical criteria. A linear normalization formula is used to map all physical quantities to the [0,1] interval:

[0029] in, These are the time-series values ​​of the original physical quantities (such as wind speed in m / s, power in MW, and electricity price in yuan / MWh). and These represent the maximum and minimum values ​​of the physical quantity within the sliding window. These are dimensionless values ​​after normalization. The sliding window is set to update dynamically every 720 hours (30 days) to ensure that the baseline dynamically adapts to seasonal weather changes.

[0030] Finally, all the processed data is arranged in chronological order to form a well-structured, time-series-continuous, and scale-consistent historical time-series dataset and a real-time time-series dataset.

[0031] S2. Based on the data of each monitoring point in the historical time series dataset under normal operating conditions, establish the numerical fluctuation range of each monitoring point as a dynamic judgment benchmark for determining power prediction anomalies.

[0032] Data samples of each monitoring point under normal operating conditions are extracted from the historical time series dataset. Normal operating conditions refer to the operating conditions under which the station equipment is running normally, there are no fault alarms, unlimited power commands, and communication links are unobstructed. Statistical analysis was performed on the normal operating condition data samples of each monitoring point to calculate the upper confidence limit and lower confidence limit of that monitoring point; The upper and lower confidence limits of each monitoring point constitute the numerical fluctuation range of that monitoring point, and the numerical fluctuation ranges of all monitoring points together constitute the dynamic judgment benchmark.

[0033] Specifically, it includes: First, data samples of each monitoring point under normal operating conditions are extracted from the historical time series dataset. Normal operating conditions are strictly defined as a stable operating state in which the station equipment is running normally, there are no fault alarms, unlimited power commands, and the communication link is unobstructed. This ensures that the selected sample data can truly reflect the inherent fluctuation characteristics of the station under conditions without external interference and internal anomalies.

[0034] Subsequently, a systematic statistical analysis was conducted on the normal operating data samples of each monitoring point. It was assumed that the residual output of the power station under normal conditions followed a normal distribution. Taking the 95% confidence interval as the upper and lower limits of the dynamic baseline, the upper and lower confidence limits for this monitoring point are calculated using the following formulas:

[0035]

[0036] in, Let t be the upper confidence limit of the dynamic baseline. Let t be the lower confidence limit of the dynamic baseline at time t. Let be the mean of the predicted residuals within the sliding window at time t. 1.96 represents the standard deviation of the predicted residuals within the sliding window at time t; 1.96 corresponds to the two-sided confidence coefficient at the 95% confidence level.

[0037] The formula for calculating the predicted residual is:

[0038] in, Let be the power prediction residual at time t. This represents the actual power generation capacity of the power station. Output predicted power for the original model.

[0039] The upper and lower confidence limits of each monitoring point constitute the numerical fluctuation range of that monitoring point. The numerical fluctuation ranges of all monitoring points are summarized and integrated to form a unified dynamic judgment benchmark covering multi-source monitoring indicators of the entire station.

[0040] S3. Compare the data of each monitoring point in the real-time time series dataset with the numerical fluctuation range of the corresponding monitoring point in the dynamic judgment benchmark. If there is data in the real-time time series dataset that exceeds the numerical fluctuation range of the corresponding monitoring point, it is determined that the power prediction is abnormal, and the abnormal features corresponding to the data are extracted.

[0041] The real-time data of each monitoring point in the real-time time series dataset is compared with the upper and lower confidence limits of the corresponding monitoring points in the dynamic judgment benchmark. If the real-time data of a certain monitoring point exceeds the upper confidence limit or falls below the lower confidence limit of that monitoring point, a preliminary anomaly alarm will be triggered. After triggering the initial anomaly alarm, a dual verification mechanism is activated: the real-time data is verified using a preset static threshold. If the real-time data exceeds both the static threshold and the dynamic judgment benchmark, the power prediction is ultimately determined to be a real anomaly. Data features of monitoring points where real anomalies occur are extracted as the anomaly features, which include at least the time of anomaly occurrence, anomaly monitoring point identifier, anomaly value, and degree of deviation.

[0042] Specifically, it includes: The real-time data from each monitoring point in the real-time time series dataset is compared one by one with the upper and lower confidence limits of the corresponding monitoring point in the dynamic determination benchmark. If the real-time value of a certain monitoring point exceeds the upper confidence limit or falls below the lower confidence limit, a preliminary anomaly alarm is triggered, indicating that the current value of the monitoring point has deviated from its historical fluctuation pattern under normal operating conditions in a statistical sense.

[0043] To avoid false alarms caused by insufficient coverage of normal operating condition samples or transient spikes due to extreme weather, a dual verification mechanism is activated immediately after an initial anomaly alarm is triggered. This involves introducing static thresholds pre-set based on equipment physical limits and scheduling procedures. As a second criterion, parallel verification is performed on the same real-time data. The complete judgment condition is:

[0044] in, This is a preset static threshold or hard threshold. When the absolute value of the residual exceeds the static threshold and the residual also exceeds the confidence interval of the dynamic baseline, the power prediction is ultimately determined to be a true anomaly; if only one of these conditions is met, it is determined to be normal fluctuation or occasional noise, and the subsequent attribution process is not triggered.

[0045] After confirming the occurrence of a real anomaly, the data features corresponding to the monitoring point where the anomaly occurred are automatically extracted as anomaly features. These anomaly features include at least the time of the anomaly occurrence, the anomaly monitoring point identifier, the anomaly value, and the degree of deviation (such as the difference or percentage exceeding the confidence boundary). They are then encapsulated in a structured data format for use in subsequent attribution and diagnosis processes.

[0046] By employing the dual verification strategy of dynamic benchmark and static threshold, the advantages of dynamic judgment benchmark in accurately adapting to the individual characteristics and seasonal time-varying patterns of the site are fully utilized. On the other hand, the physical rigidity constraint of static threshold effectively eliminates the risk of misjudgment introduced by the dynamic boundary being occasionally too wide or too narrow due to sample fluctuations. The two complement each other and work together to significantly improve the anomaly detection rate while greatly reducing the false alarm rate.

[0047] S4. Obtain the corresponding cause type based on the abnormal characteristics.

[0048] A root cause knowledge graph of deviations is constructed, wherein the nodes of the knowledge graph include at least one or more of the following causal nodes: NWP source deviation node, local meteorological change node, equipment derating node, data acquisition interruption node, communication delay node, and limited issuance instruction node; the edges of the knowledge graph represent the association relationships between the causal nodes. Parallel retrieval is performed in the knowledge graph using the abnormal monitoring point identifiers and abnormal values ​​in the abnormal features as retrieval conditions. Match the corresponding causal nodes based on the search results, and output the causal type and corresponding confidence level.

[0049] Specifically, it includes: A knowledge graph of deviation root causes is pre-constructed. The nodes of this knowledge graph cover at least one or more of the following: NWP source deviation nodes, local meteorological change nodes, equipment derating nodes, data acquisition interruption nodes, communication delay nodes, and power restriction command nodes. The edges in the graph are used to represent the causal transmission, temporal correlation, or spatial coupling relationship between the causal nodes. For example, the NWP source deviation node and the local meteorological change node can be associated through the spatiotemporal transmission path of numerical deviation, and the equipment derating node and the power restriction command node can be connected through the causal chain of power regulation. This forms a knowledge system of deviation causes that covers the entire process, including meteorological sources, data acquisition links, communication channels, equipment execution, and scheduling commands. Based on this, the abnormal monitoring point identifiers extracted from the abnormal features are used as the basis for spatial positioning, and the abnormal values ​​and their deviations are used as the basis for quantitative matching. These two types of information are combined to form search conditions, and parallel search is performed in the knowledge graph. That is, the causal nodes and their associated paths that have a mapping relationship with the current abnormal features are quickly traversed on multiple related dimensions at the same time. Finally, the most likely cause type and its corresponding confidence level are comprehensively ranked based on the matching degree and path confidence level of the hit nodes in the search results. The confidence level can be calculated based on the frequency of occurrence of the same or similar abnormal patterns under the cause type in historical cases and the similarity between the current matching features and the typical patterns.

[0050] For example, a root cause knowledge graph of deviations can be pre-constructed, and the mathematical definition of this knowledge graph is:

[0051] Among them, the node set These correspond to NWP source deviation nodes, local meteorological change nodes, equipment derating nodes, data acquisition interruption nodes, communication delay nodes, and limited issuance command nodes, respectively.

[0052] The correlation between each causal node and its corresponding abnormal features is shown in Table 1: Table 1. Correlation between each causal node and its corresponding abnormal features

[0053] In the graph, edges E represent causal transmission, temporal correlation, or spatial coupling relationships between various causal nodes. For example, the NWP source deviation node and the local meteorological mutation node can be linked through the spatiotemporal transmission path of numerical deviation, while the equipment derating node and the power restriction command node can be connected through the causal chain of power regulation. This forms a knowledge system of deviation causes covering the entire process from meteorological sources, data acquisition links, communication channels, equipment execution, and scheduling commands. The edge weight matrix W represents the historical probability weight of each root cause leading to the deviation.

[0054] Based on this, the abnormal monitoring point identifiers extracted from the abnormal features are used as the basis for spatial positioning, and the abnormal values ​​and their deviations are used as the basis for quantitative matching. These two types of information are combined to form search conditions, and parallel search is performed in the knowledge graph. That is, the causal nodes and their associated paths that have a mapping relationship with the current abnormal features are quickly traversed on multiple related dimensions at the same time.

[0055] Given anomaly observation features (i.e., the abnormal features extracted in step S3), single root cause node confidence level Calculation using Bayesian probabilistic inference:

[0056] in, root cause Prior probability is obtained from historical failure statistics; root cause Prior probability; root cause When it occurs, the current abnormal observation characteristics appear. The item; The total number of root cause nodes in the graph; the denominator is the normalization term, which ensures that the sum of the confidence scores of all root causes is 1.

[0057] Automatic Filtering The root cause is used as the diagnostic conclusion, and the standardized treatment strategy bound to the matching map is synchronized to automatically generate a complete diagnostic report.

[0058] S5. Obtain the corresponding treatment method according to the cause type, and execute the treatment method.

[0059] Based on the cause type and its confidence level output by the cause diagnosis process, the system matches and retrieves the corresponding handling method from a pre-set handling strategy library. This library pre-configures differentiated handling logic for different cause types. For example, if the cause type is determined to be NWP source deviation, it triggers dynamic adjustment of the multi-source meteorological data fusion weights or adaptive update of the NWP correction coefficient; if it's equipment throttling, it issues an equipment status verification command and initiates a backup power allocation plan; if it's data acquisition interruption or communication delay, it performs data replenishment or communication link switching operations; if it's a limited-issue command, it verifies the timeliness and execution deviation of the scheduling command; if it's a sudden change in local weather, it triggers a recalculation process for ultra-short-term rolling forecasts. The confidence level is used as the basis for classifying the handling intensity and execution priority. High-confidence attribution results directly trigger automated handling processes, while low-confidence results are executed after manual confirmation. After the handling method is determined, specific control commands are issued through the automatic execution unit at the station or the scheduling side interface to complete the handling action.

[0060] For example, based on the cause type and its confidence level output by the attribution diagnosis process, the corresponding treatment method is matched and obtained from a pre-set treatment strategy library. Treatment priority is primarily based on the quantified value of the assessment loss. Make a judgment:

[0061] in, The spot electricity price for period t; This is the duration of the current time series window; This is the deviation penalty coefficient for the "two detailed rules" of the power grid.

[0062] loss The larger the value, the higher the priority the system will execute automatic processing.

[0063] The logical mathematical determination for the third-level processing is as follows: (1) Level 1, no human intervention required ( (Low loss and slight deviation): Calculate the NWP data transmission delay, sensor refresh delay, and unit output response delay respectively.

[0064] The above three types of real-time calculation models for timing delay are:

[0065] in, Due to the time delay of NWP weather forecasts; For the time it takes for the site to receive NWP data; For NWP forecast time; For sensor data acquisition refresh latency; This refers to the local sampling time of the sensor. To standardize the time scale for the system; For the unit's output response delay; This refers to the actual power output time of the unit; This is the time when control commands are issued.

[0066] Global timing compensation coefficient The calculation formula is:

[0067] in, , , These are three types of time delay weight coefficients, dynamically updated by offline training of the cloud model, with values ​​ranging from [0,1] and summing to 1.

[0068] Corrected model input timing: To align with the timing of real physical operating conditions.

[0069] The final predicted power is output by fusing the main model for power prediction across different scenarios:

[0070] in, The final predicted power after timing correction; Output of the basic prediction model for typical scenarios; Output for a dedicated extreme model for high winds / intense radiation and high electricity prices; , As scene weight, satisfy When an extreme weather condition determination is established Automatically increases to 0.7~0.9 during normal periods. =0.8~0.95.

[0071] The quantification metric for prediction accuracy is the Mean Absolute Percentage Error (MAPE).

[0072] in, This represents the total number of time-series samples within the statistical period. This represents the actual power generation at time k. This represents the predicted power of the model after correction at time k. Supplementary optimization metric RMSE (Root Mean Square Error, used to help assess significant deviations): .

[0073] (2) Level 2, manual confirmation and handling ( (Equipment / data acquisition failure): Automatically generate maintenance work orders, push them to site personnel, and issue maintenance instructions after confirmation.

[0074] (3) Level 3, dispatching and reporting appeals ( (Major systemic deviation): Automatically integrate NWP raw data, device logs, and diagnostic reports to generate standardized appeal materials and push them to the centralized control and scheduling interface.

[0075] in, and The two-level loss threshold can be preset and configured according to the station scheduling rules.

[0076] S6. Self-learning for feedback and closed-loop feedback on treatment effectiveness.

[0077] Check whether the real-time data of the abnormal monitoring points has returned to the value fluctuation range of the dynamic judgment benchmark; If regression occurs, the loop closure is confirmed to be eliminated, and the entire chain processing information from the detection of the anomaly to the elimination of the loop is used as feedback data and input into the dynamic judgment benchmark and / or the knowledge graph to update the fluctuation range of the dynamic judgment benchmark and / or the association rules and confidence parameters of the knowledge graph. If no regression occurs, the result of this action will be treated as a newly added abnormal feature, and the attribution analysis will be re-executed in step S4. Then, step S5 will be executed again based on the newly acquired cause type until the abnormality is eliminated or the preset iteration termination condition is met.

[0078] Specifically, it includes: Check whether the real-time data of the abnormal monitoring point has returned to the corresponding value fluctuation range in the dynamic judgment benchmark, and use this as a quantitative criterion for the effectiveness of the handling; if the real-time data has fallen back to the normal range defined by the upper and lower confidence limits, it is confirmed that the abnormal event has been eliminated in a closed loop. At this time, the entire chain processing information of the abnormal event, from abnormal detection triggering, abnormal feature extraction, attribution diagnosis, handling method matching to the final closed loop elimination, is structured and organized, and input as high-value feedback data into the dynamic judgment benchmark and / or knowledge graph.

[0079] Specifically, when inputting to the dynamic judgment benchmark, the original confidence boundary can be weighted and slid-corrected based on the actual fluctuation amplitude of the abnormal monitoring point in this abnormal event, so as to gradually optimize the characterization accuracy of the statistical distribution of normal working conditions. When inputting to the knowledge graph, the weight coefficient of the association rule and the cause confidence parameter can be dynamically adjusted according to the matching degree between the current attribution result and the actual treatment effect, thereby realizing the positive evolution of the diagnostic knowledge base. If the real-time data has not yet returned to the normal fluctuation range, the current treatment result is used as a new abnormal feature, that is, the current monitoring point value, the treatment action already performed and its effect deviation are recorded and returned to step S4 to re-execute the attribution analysis, and step S5 is executed again according to the cause type obtained by the secondary attribution, until the abnormality is eliminated or the preset iteration termination condition is reached (such as the maximum number of iterations or the longest treatment time limit). If the abnormality is still not eliminated after termination, it is upgraded to the manual intervention mode.

[0080] For example, the actual fluctuation amplitude of the abnormal monitoring points in this abnormal event is used as a new sample, and the original confidence boundary is corrected using a weighted moving average method:

[0081]

[0082] in, The first learning rate controls the rate at which new samples update the baseline. and These are the mean and standard deviation of the actual fluctuation data segment at this monitoring point during this abnormal event, respectively. This is the new average value after weighting and updating based on the feedback from this anomaly event; This represents the historical average of the normal operating condition data sample at this monitoring point before the update. This is the new standard deviation after weighting and updating based on the feedback from this anomaly event; This represents the historical standard deviation of the normal operating condition data sample at this monitoring point before the update.

[0083] The knowledge graph update strategy is as follows: based on the degree of matching between the attribution result and the actual treatment effect, the edge weights of the hit path are positively strengthened or negatively weakened.

[0084] in, For the updated causal nodes To the deviation result node Edge weights; The second learning rate; The effectiveness of this treatment will be scored (whether the abnormality was eliminated after the treatment; 1 for elimination and 0 for non-elimination). This assigns a historical average performance score to the path. When the action is successful, the corresponding path's weight increases; conversely, its weight decreases, thus achieving positive evolution of the diagnostic knowledge base.

[0085] If the real-time data has not yet returned to the normal fluctuation range, the result of this handling will be treated as a new abnormal feature (record the current monitoring point value, the handling action that has been performed and its effect deviation), return to step S4 to re-execute the attribution analysis, and execute step S5 again according to the cause type obtained by the secondary attribution, until the abnormality is eliminated or the preset iteration termination condition is reached (such as the maximum number of iterations or the longest handling time limit). If the abnormality is still not eliminated after termination, it will be upgraded to manual intervention mode.

[0086] The second objective of this invention is to provide an AI-based closed-loop system for automatic diagnosis and alarm attribution of renewable energy power prediction, comprising: Data processing module: used to acquire historical and real-time operational data of each monitoring point of the target site, and to standardize the historical and real-time operational data to obtain historical time-series datasets and real-time time-series datasets respectively; Dynamic benchmark establishment module: used to establish the numerical fluctuation range of each monitoring point based on the data of each monitoring point in the historical time series dataset under normal operating conditions, as a dynamic judgment benchmark for judging power prediction anomalies; Anomaly detection module: used to compare the data of each monitoring point in the real-time time series dataset with the numerical fluctuation range of the corresponding monitoring point in the dynamic judgment benchmark. If there is data in the real-time time series dataset that exceeds the numerical fluctuation range of the corresponding monitoring point, it is determined that the power prediction is abnormal and the abnormal features corresponding to the data are extracted. Attribution diagnosis module: used to obtain the corresponding cause type based on the abnormal features; Hierarchical closed-loop treatment module: used to obtain the corresponding treatment method according to the cause type, and execute the treatment method.

[0087] The system standardizes and manages multi-source heterogeneous data through a data processing module, transforming historical and real-time operational data—which are scattered, time-scaled, and dimensionless—into high-quality time-series datasets with regular structure and uniform scale, laying a solid data foundation for subsequent diagnostic processes. The dynamic benchmark establishment module establishes confidence boundaries for each monitoring point based on historical statistical characteristics under normal operating conditions, forming an adaptive and evolvable dynamic judgment benchmark. This overcomes the inherent limitations of traditional fixed thresholds, which cannot adapt to individual differences and seasonal variations among different stations. The anomaly detection module works in conjunction with dynamic benchmarks and static thresholds... The dual verification mechanism effectively suppresses the risk of false alarms while ensuring high sensitivity, and completes multi-dimensional structured feature extraction of real abnormal events, including time, space and magnitude. The cause diagnosis module uses a deviation root cause knowledge graph covering the entire process from meteorological source to dispatch instructions to quickly map isolated abnormal features to specific cause types and output quantitative confidence scores, realizing intelligent and efficient root cause location. The hierarchical closed-loop handling module automatically matches and executes differentiated handling strategies according to cause type and confidence score, forming a fully automated closed loop from anomaly detection, cause diagnosis, handling execution to effect feedback.

[0088] In one embodiment of the present invention, a computer device is provided, comprising a processor and a memory. The memory stores a computer program, which includes program instructions. The processor executes the program instructions stored in the computer storage medium. The processor may be a Central Processing Unit (CPU), or other general-purpose processors, digital signal processors (DSPs), application-specific integrated circuits (ASICs), field-programmable gate arrays (FPGAs), or other programmable logic devices, discrete gate or transistor logic devices, discrete hardware components, etc. It is the computing and control core of the terminal, suitable for implementing one or more instructions, specifically suitable for loading and executing one or more instructions from the computer storage medium to achieve a corresponding method flow or function. The processor described in this embodiment of the present invention can be used for the operation of an AI-based closed-loop method for automatic diagnosis and alarm attribution of new energy power prediction.

[0089] This invention also provides a storage medium, specifically a computer-readable storage medium (Memory), which is a memory device in a computer device used to store programs and data. It is understood that the computer-readable storage medium here can include both the built-in storage medium in the computer device and extended storage media supported by the computer device. The computer-readable storage medium provides storage space that stores the terminal's operating system. Furthermore, this storage space also stores one or more instructions suitable for loading and execution by a processor. These instructions can be one or more computer programs (including program code). It should be noted that the computer-readable storage medium here can be high-speed RAM or non-volatile memory, such as at least one disk storage device. The processor can load and execute one or more instructions stored in the computer-readable storage medium to implement the corresponding steps of the AI-based closed-loop automatic diagnosis and alarm attribution method for new energy power prediction in the above embodiments.

[0090] Those skilled in the art will understand that embodiments of the present invention can be provided as methods, systems, or computer program products. Therefore, the present invention can take the form of a completely hardware embodiment, a completely software embodiment, or an embodiment combining software and hardware aspects. Furthermore, the present invention can take the form of a computer program product embodied on one or more computer-usable storage media (including, but not limited to, disk storage, CD-ROM, optical storage, etc.) containing computer-usable program code.

[0091] This invention is described with reference to flowchart illustrations and / or block diagrams of methods, apparatus (systems), and computer program products according to embodiments of the invention. It will be understood that each block of the flowchart illustrations and / or block diagrams, and combinations of blocks in the flowchart illustrations and / or block diagrams, can be implemented by computer program instructions. These computer program instructions can be provided to a processor of a general-purpose computer, special-purpose computer, embedded processor, or other programmable data processing apparatus to produce a machine, such that the instructions, which execute via the processor of the computer or other programmable data processing apparatus, generate instructions for implementing the flowchart illustrations and / or block diagrams. Figure 1 One or more processes and / or boxes Figure 1 A device that provides the functions specified in one or more boxes.

[0092] These computer program instructions may also be stored in a computer-readable storage medium that can direct a computer or other programmable data processing device to function in a particular manner, such that the instructions stored in the computer-readable storage medium produce an article of manufacture including instruction means, which are implemented in a process Figure 1 One or more processes and / or boxes Figure 1 The function specified in one or more boxes.

[0093] These computer program instructions may also be loaded onto a computer or other programmable data processing equipment to cause a series of operational steps to be performed on the computer or other programmable equipment to produce a computer-implemented process, thereby providing instructions that execute on the computer or other programmable equipment for implementing the process. Figure 1 One or more processes and / or boxes Figure 1 The steps of the function specified in one or more boxes.

[0094] Finally, it should be noted that the above embodiments are only used to illustrate the technical solutions of the present invention and not to limit it. Although the present invention has been described in detail with reference to the above embodiments, those skilled in the art should understand that modifications or equivalent substitutions can still be made to the specific implementation of the present invention. Any modifications or equivalent substitutions that do not depart from the spirit and scope of the present invention should be covered within the scope of protection of the claims of the present invention.

Claims

1. A method for automatic diagnosis and alarm attribution of new energy power prediction based on AI closed-loop, characterized in that, Includes the following steps: S1. Obtain historical and real-time operational data from each monitoring point of the target station, and standardize the historical and real-time operational data respectively to obtain historical time-series datasets and real-time time-series datasets. S2. Based on the data of each monitoring point in the historical time series dataset under normal operating conditions, establish the numerical fluctuation range of each monitoring point as a dynamic judgment benchmark for determining power prediction anomalies. S3. Compare the data of each monitoring point in the real-time time series dataset with the numerical fluctuation range of the corresponding monitoring point in the dynamic judgment benchmark. If there is data in the real-time time series dataset that exceeds the numerical fluctuation range of the corresponding monitoring point, it is determined that the power prediction is abnormal, and the abnormal features corresponding to the data are extracted. S4. Obtain the corresponding cause type based on the abnormal characteristics; S5. Obtain the corresponding treatment method according to the cause type, and execute the treatment method.

2. The method for automatic diagnosis and alarm attribution of new energy power prediction based on AI closed loop as described in claim 1, characterized in that, The target site's monitoring points include one or more of the following: numerical weather forecast data monitoring points, wind tower measured data monitoring points, irradiance meter measured data monitoring points, local micro-meteorological monitoring points, wind turbine operation status monitoring points, inverter operation status monitoring points, and SCADA communication status monitoring points; the historical and real-time operation data include meteorological data, equipment operation data, data acquisition status data, and scheduling rule data.

3. The method for automatic diagnosis and alarm attribution of new energy power prediction based on AI closed loop as described in claim 1, characterized in that, The standardization process is performed on the historical operational data and the real-time operational data respectively to obtain historical time-series datasets and real-time time-series datasets, including: The historical and real-time operational data are processed with a unified time stamp to align data from different sources to the same time base. Outlier removal is performed on the data after standardization of time scale. Abnormal records are identified and deleted based on the preset threshold ranges for each physical quantity. After removing outliers, missing values ​​were filled in using piecewise linear interpolation to fill in the missing time periods. The completed data is normalized to eliminate differences between different units. The processed data is arranged in chronological order to form the historical time-series dataset and the real-time time-series dataset.

4. The method for automatic diagnosis and alarm attribution of new energy power prediction based on AI closed loop as described in claim 1, characterized in that, The step of establishing the numerical fluctuation range of each monitoring point based on the data of each monitoring point in the historical time series dataset under normal operating conditions, as a dynamic judgment benchmark for determining power prediction anomalies, includes: Data samples of each monitoring point under normal operating conditions are extracted from the historical time series dataset. Normal operating conditions refer to the operating conditions under which the station equipment is running normally, there are no fault alarms, unlimited power commands, and communication links are unobstructed. Statistical analysis was performed on the normal operating condition data samples of each monitoring point to calculate the upper confidence limit and lower confidence limit of that monitoring point; The upper and lower confidence limits of each monitoring point constitute the numerical fluctuation range of that monitoring point, and the numerical fluctuation ranges of all monitoring points together constitute the dynamic judgment benchmark.

5. The method for automatic diagnosis and alarm attribution of new energy power prediction based on AI closed loop as described in claim 1, characterized in that, The step involves comparing the data from each monitoring point in the real-time time-series dataset with the numerical fluctuation range of the corresponding monitoring point in the dynamic judgment benchmark. If any data in the real-time time-series dataset exceeds the numerical fluctuation range of the corresponding monitoring point, then the power prediction is determined to be abnormal, and the abnormal features corresponding to the data are extracted, including: The real-time data of each monitoring point in the real-time time series dataset is compared with the upper and lower confidence limits of the corresponding monitoring points in the dynamic judgment benchmark. If the real-time data of a certain monitoring point exceeds the upper confidence limit or falls below the lower confidence limit of that monitoring point, a preliminary anomaly alarm will be triggered. After triggering the initial anomaly alarm, a dual verification mechanism is activated: the real-time data is verified using a preset static threshold. If the real-time data exceeds both the static threshold and the dynamic judgment benchmark, the power prediction is ultimately determined to be a real anomaly. Data features of monitoring points where real anomalies occur are extracted as the anomaly features, which include at least the time of anomaly occurrence, anomaly monitoring point identifier, anomaly value, and degree of deviation.

6. The method for automatic diagnosis and alarm attribution of new energy power prediction based on AI closed loop as described in claim 1, characterized in that, The step of obtaining the corresponding cause type based on the abnormal features includes: A root cause knowledge graph of deviations is constructed, wherein the nodes of the knowledge graph include at least one or more of the following causal nodes: NWP source deviation node, local meteorological change node, equipment derating node, data acquisition interruption node, communication delay node, and limited issuance instruction node; the edges of the knowledge graph represent the association relationships between the causal nodes. Parallel retrieval is performed in the knowledge graph using the abnormal monitoring point identifiers and abnormal values ​​in the abnormal features as retrieval conditions. Match the corresponding causal nodes based on the search results, and output the causal type and corresponding confidence level.

7. The method for automatic diagnosis and alarm attribution of new energy power prediction based on AI closed loop as described in claim 1, characterized in that, After obtaining the corresponding treatment method based on the cause type and executing the treatment method, the method further includes: After the above handling method is completed, check whether the real-time data of the abnormal monitoring point returns to the value fluctuation range of the dynamic judgment benchmark. If regression occurs, the loop closure is confirmed to be eliminated, and the entire chain processing information from the detection of the anomaly to the elimination of the loop is used as feedback data and input into the dynamic judgment benchmark and / or the knowledge graph to update the fluctuation range of the dynamic judgment benchmark and / or the association rules and confidence parameters of the knowledge graph. If no regression occurs, the result of this action will be treated as a newly added abnormal feature, and the process will return to step S4 to re-execute the attribution analysis. Then, step S5 will be executed again based on the newly acquired cause type until the abnormality is eliminated or the preset iteration termination condition is met.

8. A new energy power prediction automatic diagnosis and alarm attribution system based on AI closed loop, characterized in that, include: Data processing module: used to acquire historical and real-time operational data of each monitoring point of the target site, and to standardize the historical and real-time operational data to obtain historical time-series datasets and real-time time-series datasets respectively; Dynamic benchmark establishment module: used to establish the numerical fluctuation range of each monitoring point based on the data of each monitoring point in the historical time series dataset under normal operating conditions, as a dynamic judgment benchmark for judging power prediction anomalies; Anomaly detection module: used to compare the data of each monitoring point in the real-time time series dataset with the numerical fluctuation range of the corresponding monitoring point in the dynamic judgment benchmark. If there is data in the real-time time series dataset that exceeds the numerical fluctuation range of the corresponding monitoring point, it is determined that the power prediction is abnormal and the abnormal features corresponding to the data are extracted. Attribution diagnosis module: used to obtain the corresponding cause type based on the abnormal features; Hierarchical closed-loop treatment module: used to obtain the corresponding treatment method according to the cause type, and execute the treatment method.

9. A computer device comprising a memory, a processor, and a computer program stored in the memory and executable on the processor, characterized in that, When the processor executes the computer program, it implements the steps of the method as described in any one of claims 1-7.

10. A computer-readable storage medium storing a computer program, characterized in that, When the computer program is executed by a processor, it implements the steps of the method as described in any one of claims 1-7.

Images