Method and system for detecting concept drift of new energy power prediction model

Abstract

The invention discloses a new energy power prediction model concept drift detection method and system, and belongs to the technical field of new energy power prediction and data flow mining, and the method comprises the steps: collecting high-availability basic operation data of a new energy station in real time, and carrying out the time alignment processing; a preset three-channel collaborative drift detection module is adopted for detection; the three-channel collaborative drift detection module comprises three monitoring channels which run in parallel, and alarm signals corresponding to the channels are output through prediction residual behavior monitoring, actual power distribution drift detection and physical consistency deviation detection. And calculating the confidence coefficient of each channel, carrying out weighted average on the confidence coefficients of the channels to obtain a total drift score, and judging whether the new energy power prediction model generates concept drift according to the total drift score and the corresponding duration. The method has high availability and strong robustness, can realize high-precision and low-false-alarm detection of concept drift, and provides technical support for optimization of a power prediction model.

Description

TECHNICAL FIELD

[0001] The application belongs to the technical field of new energy power prediction, and particularly relates to a detection method and system for concept drift of a new energy power prediction model. BACKGROUND

[0002] In long-term operation, the mapping relationship between input (meteorology) and output (power) of a new energy power prediction model often deviates due to equipment aging, environmental changes or control strategy adjustment, i.e. concept drift. If the model cannot be adjusted in time, the prediction accuracy will continue to decline, affecting reliability.

[0003] Power prediction refers to a process of modeling and estimating active power output of a new energy power station in a specific period in the future based on numerical weather prediction (NWP), historical operation data and physical characteristics of the station. The power prediction result is used for power grid dispatching, power market declaration and "two rules" assessment, and is a core supporting technology for new energy grid-connected operation. Concept drift refers to a phenomenon that, as time goes by, due to factors such as equipment aging, environmental changes, control strategy adjustment or external interference, the mapping relationship between meteorological prediction and output power deviates significantly and continuously, so that the original prediction model is no longer applicable, and the prediction error systematically increases or the output behavior pattern changes At present, the existing detection methods for concept drift are mostly based on general data stream algorithms (such as ADWIN, KS test), which have the problems of not considering the physical constraints of new energy output, easily misjudging normal weather fluctuations as drift, relying on fault labels or external events (such as power limiting, cleaning), and the sensitivity and robustness of a single detection dimension being difficult to balance, and the accuracy and false alarm rate of detection needing to be further improved. SUMMARY

[0004] The application provides a detection method and system for concept drift of a new energy power prediction model, aiming to solve the problems in the existing detection methods for concept drift that the physical constraints of new energy output are not considered, normal weather fluctuations are easily misjudged as drift, reliance on fault labels or external events, the sensitivity and robustness of a single detection dimension being difficult to balance, and the accuracy and false alarm rate of detection needing to be further improved.

[0005] To achieve the above purpose, the application adopts the following technical solutions: The application provides a detection method for concept drift of a new energy power prediction model, comprising the following steps: S1, collecting high-availability basic operation data of a new energy power station in real time, and performing time alignment processing on the high-availability basic operation data; S2, based on the time-aligned high-availability basic operation data, a preset three-channel cooperative drift detection module is used for detection; The three-channel cooperative drift detection module includes three monitoring channels running in parallel, which respectively output alarm signals corresponding to each channel through prediction residual behavior monitoring, actual power distribution drift detection, and physical consistency deviation detection; Among them, the three monitoring channels are prediction residual behavior monitoring channel, actual power distribution drift detection channel and physical consistency deviation detection channel, and the detection logic of each monitoring channel is preset based on the output characteristics of new energy station, meteorological condition constraints and physical power generation principle; S3, based on the alarm signals of each channel output by the three monitoring channels, the confidence of each channel is calculated, the confidence of all channels is weighted and averaged to obtain a total drift score, and according to the total drift score and the corresponding duration of the total drift score, it is determined whether the new energy power prediction model has concept drift.

[0006] In some embodiments, in S1, the high-availability basic operation data includes the actual power generation reported by the SCADA system, the measured meteorological data measured by the wind tower and the irradiance instrument of the station, the power prediction data and the meteorological prediction data of the station power prediction system, the installed capacity of the station and the corresponding timestamp information; The measured meteorological data includes wind speed, irradiance and temperature.

[0007] In some embodiments, in S2, the detection process of the prediction residual behavior monitoring channel includes: according to the type of new energy station, dividing typical period according to output characteristics; for each typical period, based on the power prediction data and the actual power generation in the historical high-availability basic operation data, the historical error of the target period is calculated, and the dynamic threshold of the target period is maintained; based on the dynamic threshold, the error of the typical period to which the current time belongs is monitored and smoothed in real time, the observable behavior characteristics are extracted, and the first channel alarm signal is output according to the observable behavior characteristics; Among them, the type of new energy station includes photovoltaic power station and wind power station, the typical period of photovoltaic power station is divided into sunrise period, noon period and sunset period, the typical period of wind power station is divided into daytime period and night period, and the specific time range of the typical period can be adjusted according to the actual output distribution of the station.

[0008] Further, in S2, let the current time be , the model predicted power be , the actual power be , the current time , the typical period to which the current time belongs be , and the instantaneous absolute error be ; The first historical statistical period within the period The instantaneous absolute errors corresponding to the proportion of the first quantile of all instantaneous absolute errors: ; in, For time period The instantaneous absolute error corresponding to the first quantile proportion The first quantile represents the period within the first historical statistical period. The error sequence is used to calculate the linear regression slope, with each consecutive preset time window as a unit. : ; in, The slope of the linear regression. Each data point is assigned a number, and N is the total number of data points. The average number of the numbers. yes Instantaneous absolute error at a given point in time. The average error is used; the linear regression slope corresponding to the second quantile of the linear regression slope is taken as the maximum allowable upward slope of the error sequence for the target time period. ; in, It is the second quantile. The maximum permissible rising slope of the error sequence; dynamic thresholds include and Maintenance is performed according to the preset cycle.

[0009] Furthermore, in S2, observable behavioral characteristics include drift amplitude, drift velocity, and persistence. The output conditions for the first channel alarm signal are: ; in, This is the smoothed error sequence after being smoothed by an exponentially weighted moving average. ; Drift speed , among which, among which For the most recent The slope of the linear regression fit of the smoothed error sequence at each point: ; Persistence ,in To continuously satisfy Number of time points, This is the first duration threshold; Simultaneously satisfy , , When three conditions are met, the first channel alarm signal is output.

[0010] In some implementation methods, during the detection process of the actual power distribution drift detection channel in S2: if it is a photovoltaic power station, the key meteorological parameter is the predicted irradiance. Calculate clear-sky irradiance: ; Samples within the preset current data statistics window duration Extract the satisfying The power samples constitute the current sample set Find the high availability basic operation data that meets the requirements from the first historical statistical period. The power samples constitute the historical sample set ; in, The first meteorological screening threshold is set. This is the second meteorological screening threshold; right and Construct the empirical cumulative distribution function respectively and The KS test is used to judge. and The formula for the KS statistic to determine whether two things are from the same source is as follows: ; like The corresponding p-value is lower than the preset value, and the current sample set If the value is not less than the preset value, it is determined that the actual power distribution has drifted significantly, triggering a channel alarm. in, For KS statistics, For the KS test value, This is a typical threshold.

[0011] Furthermore, in S2, during the detection process of the actual power distribution drift detection channel, if it is a wind farm, the key meteorological parameter is the predicted wind speed. Obtain the rated wind speed of the unit From the samples within the current data statistics window duration Extract the satisfying The power samples constitute the current sample set Find the high availability basic operation data that meets the requirements from the first historical statistical period. The power samples constitute the historical sample set ; in, The third meteorological screening threshold, a fourth weather screening threshold value; based on the formula of KS statistics; If , the corresponding p-value is lower than a preset value, and the current sample set is not less than a preset value, it is determined that the actual power distribution has a significant drift, triggering a channel alarm.

[0012] In some embodiments, in S2, the detection process of the physical consistency deviation detection channel includes: calculating the theoretical power of the new energy power station based on the weather forecast data in the high-availability basis operation data and the installed capacity of the station; for a photovoltaic power station, a simplified model is used, and the theoretical power calculation model is: ; wherein, is the theoretical power, is the predicted irradiance, is the installed capacity of the station, is the typical temperature coefficient, is the predicted ambient temperature; For a wind power station, a piecewise cubic function is used to fit the wind power curve: ; wherein, is the predicted wind speed, , , are the cut-in wind speed, rated wind speed, and cut-out wind speed, respectively.

[0013] In some embodiments, in S3, the calculation process of the confidence level of each channel includes: the first channel confidence level ; wherein is the drift speed, is the drift amplitude, is the drift amplitude threshold value; the second channel confidence level ; wherein, is the KS test corresponding value, when the second channel alarm condition is triggered ; when the alarm condition is not triggered, ; the third channel confidence level ; When , ; wherein, is the KS statistics, is the KS statistics threshold value, is the current sample set, is a 7-day rolling mean of PR, is a PR value corresponding to a third quantile proportion of a historical PR distribution, is a deviation degree threshold value, is a third quantile proportion; The total drift score calculation uses the following formula: wherein, is a total drift score; if is greater than a preset value and the state is continuously maintained for more than a preset time, it is determined that concept drift occurs; otherwise, it is determined that the model state is normal.

[0014] The application also provides a new energy power prediction model concept drift detection system to implement the new energy power prediction model concept drift detection method described above, which includes a data acquisition and time alignment module, a three-channel drift detection module, and a concept drift determination output module. The data acquisition and time alignment module is used to acquire high-availability basic operation data of a new energy station in real time and perform time alignment processing on the high-availability basic operation data. The three-channel drift detection module is used to detect the time-aligned high-availability basic operation data using a preset three-channel collaborative drift detection module. The three-channel collaborative drift detection module includes three monitoring channels that run in parallel and output corresponding alarm signals for each channel through prediction residual behavior monitoring, actual power distribution drift detection, and physical consistency deviation detection. The three monitoring channels are a prediction residual behavior monitoring channel, an actual power distribution drift detection channel, and a physical consistency deviation detection channel, and the detection logic of each monitoring channel is preset based on the output characteristics of the new energy station, meteorological condition constraints, and physical power generation principles. The concept drift determination output module is used to calculate the confidence of each channel based on the alarm signals output by the three monitoring channels, perform weighted averaging on the confidence of all channels to obtain a total drift score, and determine whether the new energy power prediction model has concept drift based on the total drift score and the corresponding duration of the total drift score.

[0015] Compared with the prior art, the new energy power prediction model concept drift detection method and system has the following beneficial effects: ​The application discloses a detection method for concept drift of a new energy power prediction model. The application discloses a detection method for concept drift of a new energy power prediction model. BRIEF DESCRIPTION OF DRAWINGS

[0016] The accompanying drawings are included to provide a further understanding of the application, and are incorporated in and constitute a part of this specification, illustrate embodiments of the application and explain them, and do not limit the application.

[0017] Figure 1 The application discloses a detection method for concept drift of a new energy power prediction model. DETAILED DESCRIPTION

[0018] In order to make the objects, technical solutions and advantages of the embodiments of the present application clearer, the following will be combined with the accompanying drawings to clearly and completely describe the technical solutions in the embodiments of the present application. Obviously, the described embodiments are a part of the embodiments of the present application, rather than all the embodiments. The components of the embodiments of the present application described and shown in the accompanying drawings can be arranged and designed in various different configurations.

[0019] Therefore, the following detailed description of the embodiments of the application provided in the accompanying drawings is not intended to limit the scope of the claimed application, but only represents selected embodiments of the application. All other embodiments obtained by a person of ordinary skill in the art based on the embodiments in the present application without creative labor are within the scope of protection of the present application.

[0020] It should be noted that the terms "comprising", "containing", or any other similar term as used herein are intended to encompass a non-exclusive inclusion, such that a process, method, article, or apparatus that comprises a list of elements does not include only those elements in the list, but can also include other elements not expressly listed or inherent to such process, method, article, or apparatus. Without further limitation, an element defined by the phrase "comprising a" does not exclude the presence of additional identical elements in the process, method, article, or apparatus that includes the element.

[0021] It should be noted that the devices and methods disclosed in the embodiments herein can also be implemented in other ways. The device embodiments described above are merely illustrative, for example, the flowcharts and block diagrams in the drawings show possible architectures, functions and operations of devices, methods and computer program products according to the embodiments herein. In this regard, each block in the flowcharts or block diagrams can represent a module, a program segment or a portion of code that contains one or more executable instructions for implementing the specified logical functions. It should also be noted that in some alternative implementations, the functions noted in the blocks can occur in a different order than that shown in the drawings. For example, two consecutive blocks can actually be executed substantially in parallel, and they can also be executed in reverse order, depending on the functions involved. It should also be noted that each block in the block diagrams and / or flowcharts, and the combination of blocks in the block diagrams and / or flowcharts, can be implemented by a dedicated hardware-based system for performing the specified functions or actions, or can be implemented by a combination of dedicated hardware and computer instructions.

[0022] In addition, the functional modules in each of the embodiments herein can be integrated together to form an independent part, or each module can exist independently, or two or more modules can be integrated to form an independent part.

[0023] How to provide a special drift detection method that fuses multi-source observable signals, introduces physical priors, and only relies on high-availability data.

[0024] As shown in Figure 1 The detection method for conceptual drift of a new energy power prediction model according to the present application includes the following steps: S1, real-time acquisition of high-availability basic operation data of a new energy station, time alignment processing of the high-availability basic operation data; S2, based on the time-aligned high-availability basic operation data, detection is performed using a preset three-channel collaborative drift detection module; The three-channel cooperative drift detection module includes three monitoring channels running in parallel, which respectively output corresponding alarm signals of each channel through predicted residual behavior monitoring, actual power distribution drift detection and physical consistency deviation detection; The three monitoring channels are respectively a predicted residual behavior monitoring channel, an actual power distribution drift detection channel and a physical consistency deviation detection channel, and the detection logic of each monitoring channel is preset based on the output characteristics of the new energy station, meteorological condition constraints and physical power generation principles; S3, based on the channel alarm signals output by the three monitoring channels, the confidence of each channel is calculated, the confidence of all channels is weighted and averaged to obtain a total drift score, and according to the total drift score and the duration corresponding to the total drift score, it is determined whether the new energy power prediction model has concept drift.

[0025] The detection method of the application is based on the first channel of dynamic normalized predicted residual of typical output period, the second channel of distribution comparison based on meteorological similar condition constraint, and the third channel of performance ratio calculated based on online estimation of physical parameters, and drift detection is realized through cooperation of the three. A confidence weighted soft fusion mechanism is used for drift detection determination. According to the continuous confidence score and duration calculated by each channel monitoring index, it is determined whether concept drift occurs; the data used only includes SCADA power, historical meteorological data, installed capacity, station power prediction data, meteorological prediction data and time stamp, and does not depend on equipment fault diagnosis or operation and maintenance work order. The application is suitable for detecting concept drift of power prediction model of new energy station such as wind farm and photovoltaic power station, and is especially suitable for realizing high sensitivity and low false alarm rate automatic detection of performance degradation of prediction model under the condition of only relying on SCADA power data, historical meteorological data and time stamp and other high availability operation data, and providing direction guidance for power prediction model optimization.

[0026] In some embodiments, the application is a detection method for concept drift of a new energy power prediction model, which constructs a three-channel cooperative detection architecture, extracts observable features from three dimensions of predicted residual behavior, actual output distribution and physical consistency, and uses a multi-channel fusion confidence determination mechanism to realize high-precision and low-false detection of concept drift. This method does not require external fault diagnosis information and can be run only using dispatching mandatory reporting data, and has strong engineering landing performance.

[0027] Step 1: Obtain high-availability basic data; Real-time collection of high-availability basic operation data of new energy station, including: actual power generation reported by SCADA system, measured meteorological data (such as wind speed, irradiance, temperature, etc.) measured by station wind tower and radiometer, power prediction data and meteorological prediction data of station power prediction system, station installed capacity and corresponding timestamp information, all data time alignment, time resolution is 15 minutes.

[0028] Step 2: Construct a multi-channel drift detection module; A multi-channel drift detection module is constructed, which contains three monitoring channels running in parallel, which cooperatively judge the conceptual drift of the power prediction model from three dimensions of prediction residual behavior, actual output distribution and physical consistency.

[0029] 1) The first channel is the prediction residual behavior monitoring channel. This channel identifies abnormal changes in the statistical characteristics of the error sequence of the current main prediction model by analyzing the error sequence.

[0030] Firstly, time period division is performed. For photovoltaic power stations, three typical time periods are divided according to output characteristics, namely sunrise period 6:00-10:00, noon period 10:00-16:00, sunset period 16:00-20:00, and night period 20:00-6:00 of the next day, the power of which is basically 0 and does not participate in the monitoring of prediction residual behavior. For wind farms, the output characteristics of daytime and nighttime are generally different, and two typical time periods are divided according to output characteristics, namely daytime period 8:00-20:00 and nighttime period 20:00-8:00 of the next day. It should be noted that for wind farms and photovoltaic power stations in different geographical locations, the above division of typical time periods may differ, and can be adjusted according to the actual output distribution.

[0031] Then, dynamic threshold maintenance is performed. For each time period k, the errors of all time points belonging to time period k in the past 30 days are calculated. Let the current time be t, the time granularity be 15 minutes, the model predicted power be , the actual power be , and the instantaneous absolute error be: ; Calculate the 95% quantile of the errors of this period in the past 30 days: ; Calculate the maximum allowed rising slope of the error sequence of this period in the past 30 days. For each consecutive 3-hour (i.e. 12 time points) window in the error sequence of this period, calculate the linear regression slope: ; Wherein, is the linear regression slope, Each data point is numbered 1-12, and N is the total number of data points, which is 12. The average number of the numbers. yes The instantaneous absolute error at a given moment.

[0032] in This represents the average error over 12 points. The 99th percentile of the historical linear regression slope is taken as the maximum permissible upward slope of the error series for that period. .

[0033] Dynamic threshold and Offline maintenance, daily updates.

[0034] Finally, based on the dynamic threshold, the error for each time period is monitored and alarmed in real time. That is: determine the time period k to which the current time t belongs, and calculate the instantaneous absolute error at each point. To suppress random fluctuations, an exponentially weighted moving average (EWMA) is used to smooth the error series, resulting in a smoothed error series.

[0035] ; in, This is a smoothing coefficient, typically ranging from 0.2 to 0.3.

[0036] Based on the smoothed error sequence, the following observable behavioral characteristics are calculated: a) Drift amplitude Calculate the relative deviation of the current smoothing error from the historical error distribution: ; like (Typical threshold) If the error is too large, then the error range is determined to be too large.

[0037] b) Drift speed At the most recent N points (e.g., N=12, i.e., the most recent 3 hours), Perform linear regression and fit the slope. : ; like If this happens, the judgment error shows a significant upward trend.

[0038] c) Persistence Statistical continuous satisfaction Number of time points ,like (like If the error persists for more than 1 hour, then the error is considered to be too large.

[0039] In summary, the condition for the first channel to output an alarm signal is that all of the following three conditions are met , , The first channel alarm signal is output , otherwise .

[0040] 2) The second channel is an actual power distribution drift detection channel. This channel does not rely on a prediction model, but directly analyzes whether the statistical distribution of the actual power time series under similar weather conditions has shifted. It is particularly suitable for detecting systematic output capacity decline caused by equipment or environmental changes.

[0041] The second monitoring channel of the present application improves the KS (Kolmogorov-Smirnov) test. The power data distribution under current weather and historical similar weather conditions is used for KS test, and the prediction deviation and test instability that may occur at low power are eliminated, enhancing the accuracy of the test. The existing traditional KS test method uses all current data and historical data for testing.

[0042] For a photovoltaic power station, the predicted irradiance data in the historical weather forecast data is obtained , . For each sample in the recent 7-day window , the power samples that meet are extracted from the current window to form . The purpose of this approach is to eliminate the impact of overcast days on power distribution. At the same time, the weather similar samples in the historical 30-day data are searched, i.e.: ; The empirical cumulative distribution functions and are constructed for and respectively. The Kolmogorov-Smirnov (KS) test is used to determine whether the two distributions are homologous. The KS statistic is defined as: ; If (typical threshold 0.25) and the corresponding p-value is less than 0.05, and the sample size , then it is determined that the actual power distribution has a significant shift, triggering the channel alarm and outputting the first channel alarm signal , otherwise .

[0043] For wind farm, the determination method is similar. Obtain the predicted wind speed data in historical meteorological forecast data , obtain the rated wind speed of the unit , usually 10 m / s or 12 m / s. For each sample in the recent 7-day window , extract the power sample from the current window that satisfies , which is , the purpose of this approach is to eliminate the influence of unstable small wind on power distribution. Similarly, search for similar samples in the historical 30-day data , that is: ; Similarly, construct the empirical cumulative distribution functions and , respectively and , and use the Kolmogorov-Smirnov (KS) test to determine whether the two distributions are homologous. The KS statistic is defined as: ; If (typical threshold 0.10) and the corresponding p-value is less than 0.05, and the sample size , then it is determined that the actual power distribution has a significant drift, triggering a channel alarm, outputting a first channel alarm signal , otherwise .

[0044] 3) The third channel is a physical consistency deviation detection channel, which introduces the physical constraints of new energy power generation to determine whether the actual power continuously deviates from the reasonable range.

[0045] The third monitoring channel of the application, for the calculation of the theoretical power of photovoltaic, adopts a self-learning method to obtain a typical temperature coefficient and is updated regularly.

[0046] a) Theoretical power calculation For a photovoltaic power station, a simplified model is used, and the theoretical power calculation model is: ; Where is the predicted irradiance, is the installed capacity of the station, is the typical temperature coefficient, is the predicted ambient temperature.

[0047] For the typical temperature coefficient , a self-learning method is used to determine. Define a high-irradiance window on a sunny day , and solve the least squares problem in to obtain i.e. ; The analytical solution of , is updated once a month, and the latest value is used to calculate the theoretical power of the photovoltaic power station.

[0048] For a wind farm, a piecewise cubic function is used to fit the wind power curve: ; wherein is the predicted wind speed, , , are the cut-in wind speed, rated wind speed, and cut-out wind speed, respectively; are the typical values of the cut-in wind speed, rated wind speed, and cut-out wind speed, respectively.

[0049] b) Performance ratio (PR) calculation and monitoring; The rolling performance ratio is calculated as: ; The 7-day rolling mean of PR is calculated and compared with the 10% quantile of the historical PR distribution : ; If and lasts for more than 24 hours, it is determined that there is a non-meteorological factor causing performance degradation, and the actual power deviates too much from the theoretical power, triggering a channel alarm, outputting a third channel alarm signal , otherwise . The typical value of

[0050] Step 3: Concept drift cooperative judgment module The cooperative judgment module of the present application uses continuous confidence fusion for judgment, which not only reflects whether it is abnormal, but also quantifies the severity of the abnormality, effectively balancing sensitivity and robustness.

[0051] The above three channels simultaneously and continuously detect the running data. In order to reduce the false positive rate, the present application uses a soft fusion mechanism based on confidence weighting. For the first channel, the confidence is calculated using the Sigmoid function: ; For the second channel, the confidence is calculated using the p-value. When and , the confidence of the second channel is: ; Otherwise

[0052] For the third channel, a 7-day rolling mean is adopted The confidence is calculated as the degree of deviation from the historical 10% quantile, when The third channel confidence is:

[0053] Otherwise .

[0054] The total drift score is calculated as: ; If and the state is continuously maintained for more than 1 hour, it is determined that concept drift occurs; otherwise, it is determined that the model state is normal.

[0055] The application is applied to the case of a certain photovoltaic power station: A certain photovoltaic power station causes efficiency to decrease due to dust accumulation, thereby reducing photovoltaic power prediction accuracy. The system detects that PR is continuously lower than the historical lower limit through the third channel, while the first channel shows that the error slowly rises, and the second channel shows that the output distribution is overall lowered. After the fusion of the three channels, the total drift score exceeds 0.6 and the duration is more than 1 hour, the system determines that concept drift occurs, outputs the detection result, and provides the upper system with decision-making. If it is not determined that concept drift occurs, the specific value of the total drift score can also quantitatively reflect the trend and degree of model change, finely identify the possible cause of the decrease in power prediction accuracy, thereby providing technical support for prediction accuracy improvement, and has better practical significance.

[0056] Finally, it should be noted that: the above described, only for the preferred embodiments of the present application, not to the present application in any form of limitation; for the ordinary skill in the art can be shown in the specification and the above described and smoothly implement the present application, the above disclosed technical content of a few changes, modifications and evolution of equivalent changes, are equivalent embodiments of the present application; at the same time, any equivalent changes, changes, modifications and evolution of the above embodiments according to the spirit of the present application, all still belong to the scope of protection of the technical solutions of the present application.

Claims

1. A method for detecting concept drift in a new energy power prediction model, characterized in that, Includes the following steps: S1. Real-time collection of high-availability basic operation data of new energy power stations, and time alignment processing of the high-availability basic operation data; S2. Based on the time-aligned high-availability basic operation data, a preset three-channel collaborative drift detection module is used for detection; The three-channel collaborative drift detection module includes three parallel monitoring channels, which output alarm signals corresponding to each channel through predictive residual behavior monitoring, actual power distribution drift detection, and physical consistency deviation detection, respectively. Among them, the three monitoring channels are the prediction residual behavior monitoring channel, the actual power distribution drift detection channel, and the physical consistency deviation detection channel. The detection logic of each monitoring channel is based on the output characteristics of the new energy power station, meteorological conditions, and the physical power generation principle. S3. Based on the alarm signals output from the three monitoring channels, calculate the confidence level of each channel, and calculate the total drift score by weighted averaging of the confidence levels of all channels. Based on the total drift score and the duration corresponding to the total drift score, determine whether the new energy power prediction model has experienced concept drift.

2. The method for detecting concept drift in the new energy power prediction model according to claim 1, characterized in that, In S1, the high availability basic operation data includes the actual power generation reported by the SCADA system, the measured meteorological data measured by the wind measurement tower and irradiance meter at the power station, the power prediction data and meteorological prediction data of the power prediction system at the power station, the installed capacity of the power station, and the corresponding timestamp information; the measured meteorological data includes wind speed, irradiance, and temperature.

3. The method for detecting concept drift in the new energy power prediction model according to claim 1, characterized in that, In S2, the detection process of the prediction residual behavior monitoring channel includes: dividing typical time periods according to the type of new energy power station and its output characteristics; for each typical time period, calculating the historical error of the target time period based on the power prediction data and actual power generation in the historical high availability basic operation data, and maintaining the dynamic threshold of the target time period; based on the dynamic threshold, performing real-time monitoring and smoothing of the error of the typical time period to which the current moment belongs, extracting observable behavior features, and outputting the first channel alarm signal based on the observable behavior features; Among them, the types of new energy power stations include photovoltaic power stations and wind farms. The typical time periods for photovoltaic power stations are divided into sunrise, midday and sunset periods, while the typical time periods for wind farms are divided into daytime and nighttime periods. The specific time range of the typical time periods can be adjusted according to the actual power output distribution of the power station.

4. The method for detecting concept drift in the new energy power prediction model according to claim 3, characterized in that, In S2, let the current time be... The model predicts the power as follows: The actual power is At present The typical time period is The instantaneous absolute error is: ； Calculate the time period within the first historical statistical period The instantaneous absolute errors corresponding to the proportion of the first quantile of all instantaneous absolute errors: ； in, For time period The instantaneous absolute error corresponding to the first quantile proportion The first quantile represents the period within the first historical statistical period. The error sequence is used to calculate the linear regression slope, with each consecutive preset time window as a unit. : ； in, The slope of the linear regression. Each data point is assigned a number, and N is the total number of data points. The average number of the numbers. yes Instantaneous absolute error at a given point in time. The average error is used; the linear regression slope corresponding to the second quantile of the linear regression slope is taken as the maximum allowable upward slope of the error sequence for the target time period. ； in, It is the second quantile. The maximum permissible rising slope of the error sequence; dynamic thresholds include and Maintenance is performed according to the preset cycle.

5. The method for detecting concept drift in the new energy power prediction model according to claim 3, characterized in that, In S2, the observable behavioral characteristics include drift amplitude, drift velocity, and persistence. The output conditions for the first channel alarm signal are: ； in, This is the smoothed error sequence after being smoothed by an exponentially weighted moving average. ； Drift speed , among which, among which For the most recent The slope of the linear regression fit of the smoothed error sequence at each point: ； Persistence ,in To continuously satisfy Number of time points, This is the first duration threshold; Simultaneously satisfy , , When three conditions are met, the first channel alarm signal is output.

6. The method for detecting concept drift in the new energy power prediction model according to claim 1, characterized in that, In S2, during the detection process of the actual power distribution drift detection channel: if it is a photovoltaic power station, the key meteorological parameter is the predicted irradiance. Calculate clear-sky irradiance: ； Samples within the preset current data statistics window duration Extract the satisfying The power samples constitute the current sample set Find the high availability basic operation data that meets the requirements from the first historical statistical period. The power samples constitute the historical sample set ; in, The first meteorological screening threshold is set. This is the second meteorological screening threshold; right and Construct the empirical cumulative distribution function respectively and The KS test is used to judge. and The formula for the KS statistic to determine whether two things are from the same source is as follows: ;like The corresponding p-value is lower than the preset value, and the current sample set If the value is not less than the preset value, it is determined that the actual power distribution has drifted significantly, triggering a channel alarm. in, For KS statistics, For the KS test value, This is a typical threshold.

7. The method for detecting concept drift in the new energy power prediction model according to claim 6, characterized in that, In S2, during the detection process of the actual power distribution drift detection channel, if it is a wind farm, the key meteorological parameter is the predicted wind speed. Obtain the rated wind speed of the unit From the samples within the current data statistics window duration Extract the satisfying The power samples constitute the current sample set Find the high availability basic operation data that meets the requirements from the first historical statistical period. The power samples constitute the historical sample set ; in, The third meteorological screening threshold, The fourth meteorological screening threshold; Formula based on KS statistic; like The corresponding p-value is lower than the preset value, and the current sample set If the value is not less than the preset value, it is determined that the actual power distribution has drifted significantly, triggering a channel alarm.

8. The method for detecting concept drift in the new energy power prediction model according to claim 1, characterized in that, In S2, the detection process for the physical consistency deviation detection channel includes: calculating the theoretical power of the new energy power station based on meteorological forecast data and the installed capacity of the power station in the high-availability basic operation data; for photovoltaic power stations, a simplified model is adopted, and the theoretical power calculation model is as follows: ； in, Theoretical power, To predict irradiance, For the installed capacity of the station, Typical temperature coefficient, To predict ambient temperature; For wind farms, a piecewise cubic function is used to fit the wind power curve: ； in, To predict wind speed, , , These are the cut-in wind speed, rated wind speed, and cut-out wind speed, respectively.

9. The method for detecting concept drift in the new energy power prediction model according to claim 1, characterized in that, In S3, the calculation process for the confidence level of each channel includes: First channel confidence level ; in For drifting speed, For the flowing amplitude, The drift amplitude threshold; Second channel confidence level ; in, For the KS test The value, when the second channel alarm condition is triggered. When the alarm conditions are not triggered, ; Third channel confidence level ; when hour, ; in, For KS statistics, The threshold for the KS statistic is... For the current sample set, This is the 7-day rolling average of PR. This represents the PR value corresponding to the third quantile of the historical PR distribution. The threshold for the degree of deviation. The proportion of the third quantile; The total drift score is calculated using the following formula: ； in, For the total drift score; if If the value is greater than the preset value and the state is maintained continuously for more than the preset time, then concept drift is determined to have occurred; otherwise, the model state is determined to be normal.

10. A system for detecting concept drift in a new energy power prediction model, used to implement the method for detecting concept drift in a new energy power prediction model as described in any one of claims 1-9, characterized in that, It includes a data acquisition and time alignment module, a three-channel drift detection module, and a concept drift determination and output module, wherein: Data acquisition and time alignment module: used to collect high-availability basic operation data of new energy power plants in real time and perform time alignment processing on the high-availability basic operation data; Three-channel drift detection module: used for detection based on time-aligned high-availability basic operating data, using a preset three-channel collaborative drift detection module; The three-channel collaborative drift detection module includes three parallel monitoring channels, which output alarm signals corresponding to each channel through predictive residual behavior monitoring, actual power distribution drift detection, and physical consistency deviation detection, respectively. Among them, the three monitoring channels are the prediction residual behavior monitoring channel, the actual power distribution drift detection channel, and the physical consistency deviation detection channel. The detection logic of each monitoring channel is based on the output characteristics of the new energy power station, meteorological conditions, and the physical power generation principle. Concept drift determination output module: Based on the alarm signals output from the three monitoring channels, it calculates the confidence level of each channel, performs a weighted average of the confidence levels of all channels to obtain the total drift score, and determines whether the new energy power prediction model has experienced concept drift based on the total drift score and the duration corresponding to the total drift score.

Images