A real-time power generation load estimation method and system for photovoltaic power generation

Abstract

The application discloses a real-time power generation load estimation method and system for photovoltaic power generation, relates to the technical field of photovoltaic power generation load estimation, and aims at the state determination delay and model mismatch estimation error caused by the rapid movement of short-time cloud shadows, adopts the following end-to-end adaptive process: power, current, voltage, irradiance and temperature signals are collected through a plurality of array sensors, high-frequency characteristic signals and low-frequency characteristic signals are separated by applying high-pass / low-pass filtering, and a static baseline multi-dimensional feature vector is generated; dynamic characteristics are extracted based on power spectrum entropy adaptive adjustment of a time window; static and dynamic characteristics are fused to generate a discrimination input; a logic regression model is used for state determination, and extended Kalman filtering is started for online correction when an energy residual error exceeds a threshold; model drift is detected based on CUSUM and Shewhart algorithms, and an incremental training set is automatically constructed to update model parameters. Real-time power generation load estimation and operation and maintenance support with high precision and low time delay under the condition of short-time cloud shadows are realized.

Description

Technical Field

[0001] This invention relates to the field of photovoltaic power generation load estimation technology, and more specifically, to a method and system for real-time photovoltaic power generation load estimation. Background Technology

[0002] Photovoltaic power generation, as a clean and renewable energy source, plays an increasingly important role in grid connection and microgrid operation. However, due to the highly dynamic nature of sunlight conditions, especially the drastic fluctuations in irradiance caused by the rapid movement of cloud shadows in short periods, traditional power generation load estimation methods based on fixed model parameters often struggle to capture the instantaneous changes in the array's operating status in a timely manner. This leads to delays or misjudgments in status determination, further causing problems such as inaccurate model matching and increased errors in output power generation estimation. Existing technologies generally use static characteristics or a single time window for estimation when the photovoltaic array is in a stable output state or completely shaded, which cannot take into account both high-frequency jitter and low-frequency trends. Furthermore, the lack of an online calibration mechanism for model drift and environmental changes makes it difficult to meet the real-time power generation load estimation accuracy under short-term cloud shadow conditions for operation and maintenance and dispatching needs.

[0003] To address the above problems, this invention proposes a solution. Summary of the Invention

[0004] The purpose of this invention is to provide a real-time power generation load estimation method and system for photovoltaic power generation, which can solve the misestimation caused by the mismatch between state determination and modeling under short-term cloud shadow.

[0005] The objective of this invention can be achieved through the following technical solutions:

[0006] A real-time photovoltaic power generation load estimation method includes the following steps: real-time acquisition of key operating data signals and environmental data signals of the photovoltaic power station through sensors deployed in arrays; multi-frequency feature analysis is performed on the original sampled data signal set of each array to obtain high-frequency feature signals and low-frequency feature signals, forming a static baseline multi-dimensional feature vector;

[0007] A dynamic window adjustment algorithm based on power spectral entropy is used to automatically adjust the feature extraction time window length. The algorithm includes: performing a fast Fourier transform on each power signal in the static baseline multidimensional feature vector to obtain the power spectral density; calculating the power spectral entropy and comparing it with a preset spectral entropy threshold; and re-extracting enhanced multidimensional features from the original sampled signal using the corresponding feature extraction time window length based on the comparison result to obtain a dynamic multidimensional feature vector.

[0008] Based on the dynamic multidimensional feature vector and the static baseline multidimensional feature vector obtained above, the final discrimination input feature is obtained by fusing them through a signal type-aware weighted mapping gating mechanism; and input into the trained photovoltaic array operation status judgment model to output the corresponding status alarm code.

[0009] As a further aspect of the present invention: performing multi-frequency feature analysis on the original sampled data signal set of each array to generate high-frequency feature signals and low-frequency feature signals, forming a static baseline multi-dimensional feature vector, including: synchronously sampling the inverter output AC power signal, string DC current and voltage signals, array irradiance signal and module temperature signal for each photovoltaic array within a fixed time window t;

[0010] For the above-mentioned original sampled data signal set High-frequency and low-frequency characteristic signals are obtained by using high-pass and low-pass filters, respectively.

[0011] High-frequency and low-frequency feature signals are aligned and concatenated according to a time window to generate a static baseline multidimensional feature vector. in, and These represent the original sampled data signals S, respectively. i (t) represents the high-frequency and low-frequency characteristic signals within a fixed time window t, where i is the index of the original sampled data signal set, i = (1,2,...,N), and τ(t) is the time window length dynamically adjusted by the power spectral entropy algorithm.

[0012] As a further aspect of the present invention: the high-pass filter and the low-pass filter are defined according to the following formula:

[0013] High-frequency channel filter An output value of 1 indicates that the selected frequency f ≥ f c,high The original sampled data signal, where 0 indicates that other frequencies of the original sampled data signal have been suppressed;

[0014] Low-frequency channel filter An output value of 1 indicates that the selected frequency f ≤ f c,low The original sampled data signal, where 0 indicates that other frequencies of the original sampled data signal have been suppressed;

[0015] Where f is the frequency variable, f c f is the preset cutoff frequency parameter. c,high with f c,low These are the preset high-pass cutoff frequency parameters and low-pass cutoff frequency parameters corresponding to each signal category.

[0016] As a further aspect of the present invention: the logic for automatically adjusting the feature extraction window length is as follows: the calculated power spectral entropy H is compared with a preset spectral entropy threshold H. th Compare and define the window adjustment function T win (t);

[0017] Among them, Tmin and T max These are the preset shortest and longest window lengths, respectively; when H(t) > H th When using a preset minimum length window T min Extract multidimensional feature vectors; when H(t) ≤ H th When using a preset maximum length window T max Extract multidimensional feature vectors.

[0018] As a further aspect of the present invention: the final discriminative input features are obtained by fusing signals through a signal type-aware weighted mapping gating mechanism. The fusion mechanism includes: for each original sampled data signal i, assuming its feature weight in the static baseline multidimensional feature vector is ρ. i The feature weights of the dynamic multidimensional feature vector are σ. i The fusion gating factor is calculated using the Sigmoid function, and the final discriminative input features are obtained through fusion.

[0019] As a further aspect of the present invention: the photovoltaic array operation status judgment model is a trained three-class logistic regression model, which takes the multi-dimensional feature vector as input, multiplies it with the preset influence weights of various data signals and adds a bias to obtain three raw scores; the three raw scores are converted into probability outputs;

[0020] If the stable category has the highest probability, then output the status alarm code 0;

[0021] If the probability of the fluctuation category is the highest, then output the status alarm code 1;

[0022] If the probability of the shutdown category is the highest, then output status alarm code 2.

[0023] As a further aspect of the present invention: all time periods determined to be in the stable category are collected, and all time windows determined to be in the stable category are summarized to form a set of stable time periods; the output power is estimated based on the preset DC power estimation model and the inverter efficiency model respectively, and the energy residual is calculated with the real-time measurement results; if the energy residual of the current unit window exceeds the residual threshold, it is determined that the preset inverter efficiency model and the preset DC power estimation model have a deviation; the extended Kalman filter inverse inference is started for online calibration.

[0024] Conversely, the preset inverter efficiency model and preset DC power estimation model are normal.

[0025] As a further aspect of the present invention, the following steps are also included: defining the length of the main monitoring sampling window and the length of the sub-window; accumulating the actual and estimated power generation for each sub-window and recording the energy residual sequence; setting warning thresholds and alarm thresholds based on the historical normal segment residual mean and standard deviation respectively; triggering a model drift event if the sub-window residual exceeds the alarm threshold, and recording the trigger time, duration and residual statistics.

[0026] The specific logic for triggering the model drift event is as follows: when the residual of any sub-window exceeds the warning threshold but does not exceed the alarm threshold, the system enters the early warning mode, generates a warning signal, records the warning start time and residual statistics, and does not perform online incremental learning.

[0027] When the residual of any sub-window exceeds the alarm threshold, the system immediately triggers model drift detection and starts online incremental learning, while generating an alarm signal.

[0028] If the mean residual of k consecutive sub-windows falls below the warning threshold, the system exits the warning mode and resumes normal monitoring, where k is the preset number of safe windows.

[0029] As a further aspect of this invention: The residual sequence is subjected to joint detection using a cumulative sum and the Shewhart algorithm. First, a forward cumulative sum C is defined based on the cumulative sum algorithm. + (t) and negative cumulative sum C - (t);

[0030] Based on the positive accumulation of residuals C + (t) and negative cumulative sum C - (t) are updated separately;

[0031] When C + (t) or C - (t) When the preset decision threshold is exceeded, a drift detection signal is output; the drift detection signal specifically includes: when the residual of any sub-window exceeds the alarm threshold, a single-point over-limit alarm signal is output;

[0032] When the residuals of L consecutive sub-windows all exceed the alarm threshold, a continuous over-limit alarm signal is output, where L is the preset number of safety windows.

[0033] If the duration of any of the above drift detection signals exceeds the preset safe duration, a model drift event is identified.

[0034] A real-time photovoltaic power generation load estimation system includes: a data acquisition module for real-time acquisition of raw signals such as inverter output AC power, string DC current / voltage, array irradiance and module temperature through multiple types of sensors deployed in arrays, and synchronously aligning the acquired multi-source data according to timestamps to output a set of raw sampled data signals.

[0035] The data acquisition module is used to collect raw signals such as inverter output AC power, string DC current / voltage, array irradiance and component temperature in real time through multiple types of sensors deployed in arrays, and synchronize and align the collected multi-source data according to timestamps to output a set of raw sampled data signals.

[0036] The feature extraction and analysis module receives the set of raw sampled data signals output by the data acquisition module, applies high-pass / low-pass filtering to separate high-frequency and low-frequency feature signals within a fixed time window, generates a static baseline multidimensional feature vector, and performs fast Fourier transform on each power signal in the static baseline multidimensional feature vector to generate power spectral entropy.

[0037] The dynamic window and fusion module is used to adaptively determine the dynamic window length based on the power spectral entropy output by the feature extraction and analysis module and the preset spectral entropy threshold. The determined dynamic window length is used to re-extract dynamic features, and static and dynamic features are fused through a weighted mapping gating mechanism to output the final discriminative input features.

[0038] The model inference and alarm module is used to receive the final discrimination input features output by the dynamic window and fusion module, input the trained three-class logistic regression model and generate a status alarm code; based on the energy residual exceeding the threshold, it triggers online correction of the extended Kalman filter and outputs alarm information and updated model parameters;

[0039] The storage and visualization module receives the status alarm codes, calibration signals and historical feature vectors output by the model inference and alarm module, stores them in the time series database, and provides real-time visualization on the user interface.

[0040] The beneficial effects of this invention are:

[0041] (1) This invention effectively takes into account both the high-frequency jitter and long-term trend caused by cloud shadows through multi-frequency separation, high-low frequency fusion and dynamic window adjustment based on power spectrum entropy, and significantly improves the accuracy and robustness of short-term state determination.

[0042] (2) This invention, by integrating extended Kalman filter inverse inference and residual drift detection mechanism, corrects the preset power and efficiency model parameters in real time, and triggers online incremental learning when necessary, so as to achieve adaptive compensation for non-stationary effects such as component contamination and inverter drift, which greatly reduces estimation error and extends the model validity period. Attached Figure Description

[0043] The invention will now be further described with reference to the accompanying drawings.

[0044] Figure 1 This is a schematic diagram of the overall process of a real-time power generation load estimation method for photovoltaic power generation according to the present invention.

[0045] Figure 2 This is a flowchart illustrating the process of obtaining high-frequency and low-frequency characteristic signals in step one of this invention.

[0046] Figure 3 This is a flowchart of the drift detection decision process in step four of this invention;

[0047] Figure 4 This is a schematic diagram of the structural framework of a real-time power generation load estimation system for photovoltaic power generation according to the present invention. Detailed Implementation

[0048] The technical solutions of the embodiments of the present invention will be clearly and completely described below with reference to the accompanying drawings. Obviously, the described embodiments are only some embodiments of the present invention, and not all embodiments. Based on the embodiments of the present invention, all other embodiments obtained by those skilled in the art without creative effort are within the scope of protection of the present invention.

[0049] The photovoltaic power station described in this invention adopts a distributed array deployment architecture, comprising at least two independent photovoltaic arrays; each array includes: a photovoltaic string: formed by multiple photovoltaic modules connected in series;

[0050] Inverter: Connected to the string output, performs DC / AC conversion and measures AC output power; Environmental sensor: Deployed inside the array to monitor irradiance, component temperature and ambient temperature.

[0051] Example 1

[0052] Please see Figure 1 As shown, this invention provides a real-time photovoltaic power generation load estimation method, comprising the following steps:

[0053] Step 1: Real-time acquisition of key operational and environmental data signals from the photovoltaic power station using multiple types of sensors; multi-frequency feature analysis of the acquired data signals to obtain high-frequency and low-frequency feature signals, forming a static baseline multi-dimensional feature vector describing the operating status of the photovoltaic array; specifically:

[0054] For each photovoltaic array, synchronous data signals are sampled within a fixed time window of t. Based on the sampled key operational data signals and environmental data signals of the photovoltaic power station, an original set of sampled data signals is formed.

[0055] The key operating data signals include, but are not limited to: the AC power signal P output by the inverter. AC (t); String DC current signal I DC (t) and voltage signal U DC (t);

[0056] The environmental data signals include, but are not limited to: the irradiance signal G(t) measured by the irradiance sensor deployed locally on the array; and the module temperature signal T measured by the temperature sensor attached to the backsheet of the photovoltaic module. cell (t);

[0057] Please see Figure 2 As shown, for various raw power signals acquired in real time, a multi-channel filtering method is used to simultaneously extract high-frequency and low-frequency features from the data. The specific algorithm logic flow is as follows:

[0058] Based on the original sampled data signal set: Where N represents the number of original sampled data signals and N is a positive integer; for example, in this embodiment, it can be S1(t) = P AC (t), S2(t)=I DC (t) etc.;

[0059] For any signal Where X(t) represents the original sampled value of the signal at time t, and the high-frequency channel filter H is defined. high (f) and low-frequency channel filter H low (f):

[0060] An output value of 1 indicates that the selected frequency f ≥ f c,high The original sampled data signal, where 0 indicates that other frequencies of the original sampled data signal have been suppressed;

[0061] An output value of 1 indicates that the selected frequency f ≤ f c,low The original sampled data signal, where 0 indicates that other frequencies of the original sampled data signal have been suppressed;

[0062] Where f is the frequency variable, f c f is the preset cutoff frequency parameter. c,high with f c,low These are the preset high-pass cutoff frequency parameters and low-pass cutoff frequency parameters corresponding to each signal category;

[0063] The high-frequency channel filter H high (f) and low-frequency channel filter H low (f) As a frequency domain mask, the signal decomposition formula is embedded through frequency domain multiplication operations to directly generate high-frequency feature data signal X. high (t) and low-frequency characteristic data signal X low (t);

[0064] It should be noted that the signal decomposition formula is as follows:

[0065] Among them, X high (t) is a high-frequency characteristic signal, X low (t) represents the low-frequency characteristic signal. To convert the time-domain signal X(t) into a frequency-domain representation, To inversely transform the filtered frequency domain signal back to the time domain and generate analyzable feature signals;

[0066] Based on the generated high-frequency feature data signal X high (t) and low-frequency characteristic data signal X low (t) Aligned by time window, extract the static baseline multidimensional feature vector describing the operating status of the photovoltaic arrays for all arrays. in, and These represent the original sampled data signals S, respectively. i (t) High-frequency data signal characteristics and low-frequency data signal characteristics within a fixed time window t, where i is the index of the original sampled data signal set, i = (1,2,...,N), and τ(t) is the time window length dynamically adjusted by the power spectral entropy algorithm; the static baseline multidimensional feature vector includes, but is not limited to, the following feature vectors: power slope characteristics, power variance characteristics, high-frequency power variance characteristics, DC current change rate characteristics, component temperature change characteristics, inverter temperature change characteristics, irradiance abrupt change characteristics, high-frequency energy ratio, and low-frequency mean;

[0067] This invention addresses the issue of micro-environmental differences caused by spatial distribution through independent acquisition using a segmented array; it separates high-frequency disturbances from low-frequency trends using multi-channel parallel filtering to avoid signal aliasing; and it constructs a static baseline multi-dimensional feature vector that can quantify the operating status through spatiotemporal alignment and fusion, providing a high-dimensional basis for load estimation.

[0068] Step 2: Employ a dynamic window adjustment algorithm based on power spectral entropy to automatically adjust the feature extraction window length according to the high and low frequency characteristics of various data signals. Shorten the window during high-frequency jitter and lengthen it during stable periods. Use the obtained dynamic multidimensional feature vector as input to construct a photovoltaic array operation status judgment model based on the trained model and generate status alarm codes. The status alarm codes include: stable, fluctuating, or shutdown status codes.

[0069] For the static baseline multidimensional feature vector V array For each power signal in (t), perform a Fast Fourier Transform (FFT) to obtain the power spectral density distribution; after normalizing the power spectral density distribution, calculate the power spectral entropy H(t); then compare the calculated power spectral entropy H with a preset spectral entropy threshold H. th Compare and define the window adjustment function T win (t);

[0070] Among them, T min and T max These are the preset shortest and longest window lengths, respectively; when H(t) > H th When using a preset minimum length window T min Extract multidimensional feature vectors; when H(t) ≤ H th When using a preset maximum length window T max Extract multidimensional feature vectors;

[0071] Based on the dynamic window adjustment results described above, the system performs enhanced multidimensional feature vector extraction on the data within each window; this mainly includes: within the current time window T win Within (t), the absolute value of the deviation between the signal and its mean value within the time window is averaged to obtain the time-domain volatility; the local mean value and standard deviation are continuously calculated with small step sizes to obtain the sliding statistic; the frequency band with the highest energy density is determined in the high-frequency components and used as the dominant frequency feature to obtain the spectral dominant frequency; the difference between the current spectral entropy evaluation value and the previous evaluation value is recorded to obtain the spectral entropy change.

[0072] The static baseline multidimensional feature vector and the enhanced multidimensional feature vector are concatenated using a signal type-aware weighted mapping to obtain the dynamic multidimensional feature vector V extracted under the dynamic window. dyn (t, γ) where γ marks the time attribute dimension of the dynamic window; the specific process of the stitching is as follows: for each original sampled data signal i, let its feature weight in the static baseline multidimensional feature vector be ρ. i The feature weights of the dynamic multidimensional feature vector are σ. i The fusion gating factor is calculated using the Sigmoid function.

[0073] While retaining the enhanced multidimensional feature vector, a one-to-one semantic correspondence is achieved by projecting static and dynamic features onto the same principal component subspace, without the need for an externally predefined order. Furthermore, an implicit window length attribute dimension γ is added to the final fusion vector and encoded together with the fusion gating factor of each channel to form a multidimensional input with temporal traceability, so that the subsequent model can simultaneously perceive what the feature is and the dual semantics of the feature being extracted from the time window of which it is.

[0074] Based on the channel ordering information of the static baseline, the consistency of the channel dimension is maintained throughout the fusion process, but self-supervised learning is allowed to continuously fine-tune the influence weight of the fusion gating factor on various data signals in the order, so as to dynamically optimize the contribution of each channel feature to the overall discrimination model without destroying the original order.

[0075] The multidimensional feature vector V extracted under the dynamic window dyn(t, γ) are used as inputs to a trained three-class logistic regression model, with an output dimension of 3, corresponding to three states: stable, fluctuating, and stopped. Specifically: the input multidimensional feature vector V... dyn (t,γ) is multiplied by the preset influence weights of various data signals and a bias is added to obtain three raw scores; the three raw scores are then converted into probability outputs.

[0076] If the stable category has the highest probability, then output the status alarm code 0;

[0077] If the probability of the fluctuation category is the highest, then output the status alarm code 1;

[0078] If the probability of the shutdown category is the highest, then output status alarm code 2;

[0079] When the status alarm code is 1 or 2, the system directly sends a real-time alarm to the operation and maintenance platform.

[0080] Step 3: Collect data for all time periods classified as stable. Estimate the DC power based on the preset DC power estimation model, and then estimate the AC power based on the preset inverter efficiency model. Accumulate the actual estimated power generation within each unit time window and calculate the energy residual. When the energy residual exceeds the preset residual threshold, trigger the extended Kalman filter for inverse inference. Use the power generation and theoretical efficiency to infer the pollution level and equivalent irradiance of the photovoltaic array components, and then correct the key component parameters of the preset ambient temperature and photovoltaic array efficiency model in real time.

[0081] Collect all time periods that are classified as stable, and summarize all time windows classified as stable to form a set of stable time periods; this set corresponds to the period when the photovoltaic array operates smoothly without sudden abnormal shocks.

[0082] For each moment within the stable category period, the system reads the following measured data: equivalent irradiance obtained from a local irradiance sensor after temperature and tilt compensation; component temperature measured by a surface-mount temperature sensor; ambient temperature; actual DC power and AC power;

[0083] The system performs linear interpolation to align the sampling timestamps of different sensors, removes missing or out-of-threshold points, and removes high-frequency noise interference through hardware and software filtering to ensure the reliability of the input data for subsequent model estimation.

[0084] Based on the preset DC power estimation model, the DC power estimate is obtained by taking the current equivalent irradiance and component temperature as inputs.

[0085] Based on the current ambient temperature and the initially calibrated temperature compensation and aging factor, the preset inverter efficiency model is called, and the inverter efficiency result obtained from the preset inverter efficiency model is multiplied by the estimated output DC power to obtain the AC power estimate.

[0086] The cumulative actual power generation and the cumulative estimated power generation are obtained by summing the actual measured AC power and the inverter's estimated AC power within each stable period; the difference between the two is the energy residual.

[0087] By fitting the distribution using historical residual data, residual thresholds are automatically generated.

[0088] If the current unit window energy residual exceeds the residual threshold, it is determined that there is a deviation between the preset inverter efficiency model and the preset DC power estimation model; the extended Kalman filter is started for reverse inference.

[0089] Conversely, the preset inverter efficiency model and preset DC power estimation model are normal.

[0090] Specifically, the inverse inference of the extended Kalman filter is as follows: using the energy residual as the observation, the key parameters such as the contamination factor, temperature compensation coefficient, and efficiency decay factor of each array component are estimated in reverse through the Kalman filter equation to achieve online adaptive calibration.

[0091] The Kalman filter state equation is set as follows: the pollution attenuation factor, inverter aging attenuation factor and environmental equivalent irradiance correction term on each array component are taken as the state to be estimated.

[0092] Using energy residuals as the observation input, a nonlinear function relating them to state variables is defined and constructed based on the current preset inverter efficiency model and preset DC power estimation model.

[0093] After initializing the state covariance and measurement noise covariance, the prediction and update steps are repeatedly executed to output the state estimate.

[0094] The environmental equivalent irradiance correction term and pollution attenuation factor obtained by filtering are written back into the preset DC power estimation model; the inverter aging attenuation factor is written back into the preset inverter efficiency model, so that the subsequent estimation is closer to the actual situation on site.

[0095] The updated model parameters will automatically take effect in the next round of stable window estimation, and calibration records will be recorded at the same time for operation and maintenance personnel to audit or analyze offline.

[0096] It should be noted that the preset DC power estimation model and preset inverter efficiency model are constructed using IEC61853 series standards or offline calibration data provided by inverter manufacturers, which are existing technologies and will not be elaborated upon here. The preset DC power estimation model is based on the DC power estimate of the array equivalent irradiance and the component surface temperature; the preset inverter efficiency model takes this estimate and the ambient temperature, wind speed, etc., as inputs to output the AC power estimate of the inverter. The above models are used to generate initial DC power estimates and AC power estimates for subsequent Kalman filter correction and online incremental learning, including residual calculation and parameter updates. In the implementation of this invention, all estimates obtained based on this initial model are considered preset estimation parameters. Subsequent extended Kalman filter inverse inference, residual drift detection, incremental learning, etc., are all performed on the basis of this preset model to achieve real-time compensation and adaptive optimization for on-site pollution levels, component aging, cloud shadow fluctuations, and inverter performance drift.

[0097] Step 4: Continuously monitor the residual distribution of sampled power generation energy, detect long-term offset trends through adaptive thresholds, and distinguish between short-term seasonal fluctuations and true model drift; when significant drift occurs, automatically trigger online incremental learning to optimize and update model parameters; output the power generation load estimation results of each photovoltaic array and the entire station in real time, and dynamically generate alarm and drift signals.

[0098] Define the monitoring sampling window: for example, T main =1h is the main window length, and the sliding step size is ΔT main =15min; T per main window main The internal sub-window is further divided into 6 sub-windows, each with a duration of 10 minutes. sub Used for refined analysis;

[0099] For each sub-window, sum up the actual measured AC output power;

[0100] Similarly, by applying the updated preset inverter efficiency model and preset DC power estimation model from step three, the cumulative sampled actual power generation and the cumulative sampled estimated power generation are obtained.

[0101] The energy residual is calculated based on the cumulative sampled actual power generation and the cumulative sampled estimated power generation. This residual is recorded as a sub-window residual sequence and then stored in the database. Each main window outputs statistical values ​​such as the average residual, maximum residual, and minimum residual for subsequent feature extraction.

[0102] Based on the residual mean and standard deviation of the historical normal operation period (past 30 days), the 95th percentile and 99th percentile are taken as the initial warning threshold and initial alarm threshold, respectively.

[0103] The historical average residual μ is obtained by applying an exponentially weighted moving average to the daily main window residuals;

[0104] Similarly, update the influence weights of various data signals in step three, and adjust the initial warning threshold and initial alarm threshold accordingly to ensure that the initial warning threshold and initial alarm threshold can drift smoothly with seasonal and environmental changes.

[0105] When the residual of any sub-window exceeds the warning threshold but does not exceed the alarm threshold, the system enters the early warning mode, generates a warning signal, and records the warning start time and residual statistics, without performing online incremental learning.

[0106] When the residual of any sub-window exceeds the alarm threshold, the system immediately triggers model drift detection and starts online incremental learning, while generating an alarm signal.

[0107] If the mean residual of k consecutive sub-windows falls below the warning threshold, the system exits the warning mode and resumes normal monitoring, where k is the preset number of safe windows;

[0108] After collecting and initially warning about the energy residuals, the cumulative sum algorithm and the Shewhart algorithm are used to detect drift in the residual sequence, including:

[0109] Define positive cumulative sum and negative cumulative sum:

[0110] C + (t)=max{0,C + (t-1)+(△E(t)-μ)-P};

[0111] C - (t)=max(0,C - (t-1)+(△E(t)-μ)-P); where △E(t) is the current sub-window residual, μ is the historical average residual, and P is the preset allowable offset constant;

[0112] Based on the positive accumulation of residuals C + (t) and negative cumulative sum C - (t) are updated separately;

[0113] When C + (t) or C - (t) When the preset decision threshold is exceeded, a drift detection signal is output; the drift detection signal specifically includes: when the residual of any sub-window exceeds the alarm threshold, a single-point over-limit alarm signal is output;

[0114] When the residuals of L consecutive sub-windows all exceed the alarm threshold, a continuous over-limit alarm signal is output, where L is the preset number of safety windows.

[0115] If the duration of any of the above drift detection signals exceeds the preset safe duration, a model drift event is identified.

[0116] Drift events are recorded periodically, including trigger time, duration, triggering algorithm type, residual statistics, etc., for subsequent analysis.

[0117] After detecting a model drift event, the system automatically pauses online inference for new samples, but continues to collect raw data and state labels.

[0118] Synchronously record drift event metadata, including trigger threshold, trigger signal type, associated main window number, etc.

[0119] Select all sub-windows that were determined to be in a stable state within 24 hours before the drift event occurred, along with their corresponding dynamic input vectors and actual output power;

[0120] Construct an incremental training set, with sample formats including input vector, actual power, and residuals before and after drift; perform LOF outlier detection on the incremental samples to remove abnormal samples.

[0121] Incremental samples are named with the date and drift event number, written to the time series database and archived; each new sample generated by a drift event forms a version, and supports backtracking, comparison and rollback.

[0122] To ensure the continuous high accuracy and robustness of the estimation model, after completing the online calibration of model parameters in step three, this invention further monitors the residual power generation during system operation over a long period. When model drift or environmental changes cause continuous error accumulation, an online incremental learning mechanism is automatically triggered to adaptively optimize and update the model. This step mainly includes: data acquisition and preprocessing, residual distribution feature extraction, adaptive threshold dynamic adjustment, drift detection algorithm, drift event triggering and incremental sample management, online incremental learning strategy, model validation and smooth switching, and system integration and closed-loop operation and maintenance.

[0123] like Figure 2 As shown, the present invention also discloses a real-time photovoltaic power generation load estimation system for implementing the above method, including a data acquisition module, a feature extraction and analysis module, a dynamic window and fusion module, a model reasoning and alarm module, and a storage and visualization module, with signal connections between the modules;

[0124] The data acquisition module is used to collect raw signals such as inverter output AC power, string DC current / voltage, array irradiance and component temperature in real time through multiple types of sensors deployed in arrays, and synchronize and align the collected multi-source data according to timestamps to output a set of raw sampled data signals.

[0125] The feature extraction and analysis module receives the set of raw sampled data signals output by the data acquisition module, applies high-pass / low-pass filtering to separate high-frequency and low-frequency feature signals within a fixed time window, generates a static baseline multidimensional feature vector, and performs fast Fourier transform on each power signal in the static baseline multidimensional feature vector to generate power spectral entropy.

[0126] The dynamic window and fusion module is used to adaptively determine the dynamic window length based on the power spectral entropy output by the feature extraction and analysis module and the preset spectral entropy threshold. The determined dynamic window length is used to re-extract dynamic features, and static and dynamic features are fused through a weighted mapping gating mechanism to output the final discriminative input features.

[0127] The model inference and alarm module is used to receive the final discrimination input features output by the dynamic window and fusion module, input the trained three-class logistic regression model and generate a status alarm code; based on the energy residual exceeding the threshold, it triggers online correction of the extended Kalman filter and outputs alarm information and updated model parameters;

[0128] The storage and visualization module receives the status alarm codes, calibration signals and historical feature vectors output by the model inference and alarm module, stores them in the time series database, and provides real-time visualization on the user interface.

[0129] The above formulas are all dimensionless calculations. The formulas are derived from software simulations based on a large amount of collected data to obtain the most recent real-world results. The preset parameters in the formulas are set by those skilled in the art according to the actual situation.

[0130] The above embodiments can be implemented, in whole or in part, by software, hardware, firmware, or any other combination thereof. When implemented using software, the above embodiments can be implemented, in whole or in part, in the form of a computer program product.

[0131] Those skilled in the art will recognize that the modules and algorithm steps of the various examples described in conjunction with the embodiments disclosed herein can be implemented in electronic hardware, or a combination of computer software and electronic hardware. Whether these functions are implemented in hardware or software depends on the specific application of the technical solution and the constraints of the invention;

[0132] Skilled professionals may use different methods to implement the described functions for each specific application, but such implementation should not be considered beyond the scope of this application.

[0133] In addition, the functional modules in the various embodiments of this application can be integrated into one processing module, or each module can exist physically separately, or two or more modules can be integrated into one module.

[0134] The above description is merely a specific embodiment of this application, but the scope of protection of this application is not limited thereto. Any variations or substitutions that can be easily conceived by those skilled in the art within the scope of the technology disclosed in this application should be included within the scope of protection of this application. Therefore, the scope of protection of this application should be determined by the scope of the claims.

[0135] In conclusion, the above description is only a preferred embodiment of the present invention and is not intended to limit the present invention. Any modifications, equivalent substitutions, improvements, etc., made within the spirit and principles of the present invention should be included within the protection scope of the present invention.

Claims

1. A method for estimating real-time photovoltaic power generation load, characterized in that, Includes the following steps: The key operating data signals and environmental data signals of the photovoltaic power station are collected in real time by sensors deployed in arrays; multi-frequency feature analysis is performed on the raw sampled data signal set of each photovoltaic array to obtain high-frequency feature signals and low-frequency feature signals, forming a static baseline multi-dimensional feature vector; among them, the key operating data signals include the AC power signal output by the inverter, the string DC current signal and voltage signal; A dynamic window adjustment algorithm based on power spectral entropy is used to automatically adjust the feature extraction time window length. The algorithm includes: performing a fast Fourier transform on each power signal in the static baseline multidimensional feature vector to obtain the power spectral density; then calculating the power spectral entropy and comparing it with a preset spectral entropy threshold; and, based on the comparison result, re-extracting enhanced multidimensional features from the original sampled signal using the corresponding feature extraction time window length to obtain a dynamic multidimensional feature vector. Based on the dynamic multidimensional feature vector and the static baseline multidimensional feature vector obtained above, the final discrimination input feature is obtained by fusing them through a signal type-aware weighted mapping gating mechanism; and input into the trained photovoltaic array operation status judgment model, outputting the corresponding status alarm code; wherein, the photovoltaic array operation status judgment model is a trained three-class logistic regression model, which multiplies the multidimensional feature vector as input with the preset influence weights of various data signals and adds a preset bias to obtain three raw scores; the three raw scores are then converted into probability output; If the stable category has the highest probability, then output the status alarm code 0; If the probability of the fluctuation category is the highest, then output the status alarm code 1; If the probability of the shutdown category is the highest, then output status alarm code 2; Collect all time periods classified as stable, summarize all time windows classified as stable to form a set of stable time periods; estimate the output power based on the preset DC power estimation model and the inverter efficiency model respectively, and calculate the energy residual with the real-time measurement results; if the energy residual of the current unit window exceeds the preset residual threshold, it is determined that there is a deviation between the preset inverter efficiency model and the preset DC power estimation model; start the extended Kalman filter inverse inference for online calibration. Conversely, the preset inverter efficiency model and preset DC power estimation model are normal. The system defines the length of the main monitoring sampling window and the length of the sub-windows; it accumulates the actual and estimated power generation for each sub-window and records the energy residual sequence; it sets warning thresholds and alarm thresholds based on the historical normal segment residual mean and standard deviation; if the sub-window residual exceeds the alarm threshold, a model drift event is triggered, and the trigger time, duration, and residual statistics are recorded.

2. The method for real-time generation load estimation of photovoltaic power generation according to claim 1, characterized in that: Multi-frequency feature analysis is performed on the raw sampled data signal set of each array to generate high-frequency feature signals and low-frequency feature signals, forming a static baseline multi-dimensional feature vector, including: for each photovoltaic array, synchronous sampling of inverter output AC power signal, string DC current and voltage signal, array irradiance signal and module temperature signal within a fixed time window t; For the original sampled data signal set High-frequency and low-frequency characteristic signals are obtained by using high-pass and low-pass filters, respectively. High-frequency and low-frequency feature signals are aligned and concatenated according to a time window to generate a static baseline multidimensional feature vector. ,in, and These represent the original sampled data signals respectively. (t) represents the high-frequency and low-frequency characteristic signals within a fixed time window t, where i is the index of the original sampled data signal set. .

3. The method for real-time generation load estimation of photovoltaic power generation according to claim 2, characterized in that: The high-pass filter and low-pass filter are defined according to the following formulas: High-frequency channel filter ; An output value of 1 indicates the selected frequency. The original sampled data signal, where 0 indicates that other frequencies of the original sampled data signal have been suppressed; Low-frequency channel filter ; An output value of 1 indicates the selected frequency. The original sampled data signal, where 0 indicates that other frequencies of the original sampled data signal have been suppressed; Where f is the frequency variable, For the preset cutoff frequency parameter, and These are the preset high-pass cutoff frequency parameters and low-pass cutoff frequency parameters corresponding to each signal category.

4. The method for real-time generation load estimation of photovoltaic power generation according to claim 1, characterized in that, The logic for automatically adjusting the feature extraction window length is as follows: the calculated power spectral entropy H is compared with the preset spectral entropy threshold. Compare and define window adjustment functions ; ;in, and These are the preset shortest and longest window lengths, respectively; when When using a preset minimum window length Extract multidimensional feature vectors; when When using a window with a preset maximum length, Extract multidimensional feature vectors.

5. The method for real-time generation load estimation of photovoltaic power generation according to claim 1, characterized in that, The final discriminative input features are obtained through a signal type-aware weighted mapping gating mechanism, including: for each original sampled data signal i, let its feature weight in the static baseline multidimensional feature vector be... The feature weights of the dynamic multidimensional feature vector are The fusion gating factor is calculated using the Sigmoid function, and the final discriminative input features are obtained through fusion.

6. The method for real-time generation load estimation of photovoltaic power generation according to claim 1, characterized in that, The triggering logic for model drift events is as follows: when the residual of any sub-window exceeds the warning threshold but does not exceed the alarm threshold, the system enters the early warning mode, generates a warning signal, records the warning start time and residual statistics, and does not perform online incremental learning. When the residual of any sub-window exceeds the alarm threshold, the system immediately triggers model drift detection and starts online incremental learning, while generating an alarm signal. If the mean residual of k consecutive sub-windows falls below the warning threshold, the system exits the warning mode and resumes normal monitoring, where k is the preset number of safe windows.

7. The method for real-time generation load estimation of photovoltaic power generation according to claim 6, characterized in that, The residual sequence is subjected to joint detection using cumulative sum and Shewhart algorithm. First, a forward cumulative sum is defined based on the cumulative sum algorithm. Sum and negative cumulative sum ; Based on the positive accumulation of residuals Sum and negative cumulative sum Update them separately; when or When the preset decision threshold is exceeded, a drift detection signal is output; the drift detection signal specifically includes: when the residual of any sub-window exceeds the alarm threshold, a single-point over-limit alarm signal is output; When the residuals of L consecutive sub-windows all exceed the alarm threshold, a continuous over-limit alarm signal is output, where L is the preset number of safe windows. If the duration of any of the above drift detection signals exceeds the preset safe duration, a model drift event is identified.

8. A real-time photovoltaic power generation load estimation system, used to implement the real-time photovoltaic power generation load estimation method according to any one of claims 1-7, characterized in that, include: The data acquisition module is used to collect the inverter's output AC power signal, string DC current signal and voltage signal, array irradiance signal and component temperature signal in real time through multiple types of sensors deployed in arrays, and synchronize and align the collected multi-source data according to timestamps to output the raw sampled data signal set; The feature extraction and analysis module receives the raw sampled data signal set output by the data acquisition module, applies a high-pass / low-pass filter to it within a fixed time window to separate high-frequency and low-frequency feature signals, and generates a static baseline multidimensional feature vector; performs a fast Fourier transform on each power signal in the static baseline multidimensional feature vector to obtain the power spectral density, and calculates the power spectral entropy after normalizing the power spectral density; The dynamic window and fusion module is used to adaptively determine the dynamic window length based on the power spectral entropy output by the feature extraction and analysis module and the preset spectral entropy threshold. The determined dynamic window length is used to re-extract dynamic features, and static and dynamic features are fused through a weighted mapping gating mechanism to output the final discriminative input features. The model inference and alarm module is used to receive the final discrimination input features output by the dynamic window and the fusion module, input the trained three-class logistic regression model, and generate status alarm codes. When the energy residual exceeds the preset residual threshold, the extended Kalman filter online correction is triggered, and alarm information and updated model parameters are output. The storage and visualization module receives the status alarm codes, calibration signals and historical feature vectors output by the model inference and alarm module, stores them in the time series database, and provides real-time visualization on the user interface.

Images