A clear sky residual event prototype migration-based cross-site photovoltaic power prediction method and system

Abstract

This invention relates to the field of photovoltaic power prediction technology, and in particular to a cross-site photovoltaic power prediction method and system based on clear-sky residual event prototype migration. The method includes constructing a multi-timescale historical window at each prediction time based on a constructed clear-sky disturbance characterization sequence, generating a multi-timescale sample sequence; identifying photovoltaic operation events based on the generated multi-timescale sample sequence; establishing a cross-site event prototype library based on the identified photovoltaic operation events; using the cross-site event prototype library to perform source domain pre-training and target domain few-sample calibration of the event prototype contrast loss; and dynamically fusing the source domain pre-training and target domain few-sample calibration results to reconstruct the future power and prediction range of the target power plant. This invention can reduce the negative migration caused by installed capacity, geographical location, solar irradiance phase, and local weather differences, and improve the prediction accuracy and stability of newly built photovoltaic power plants under small-sample conditions.

Description

Technical Field

[0001] This invention relates to the field of photovoltaic power prediction technology, and in particular to a cross-site photovoltaic power prediction method and system based on clear-sky residual event prototype migration. Background Technology

[0002] With a high proportion of new energy sources being integrated into the grid, photovoltaic power generation forecasting has become a crucial foundation for power system operation control, reserve capacity configuration, energy storage dispatch, and day-ahead planning. Unlike traditional controllable power sources such as thermal power and hydropower, photovoltaic power output is directly affected by a combination of factors including solar irradiance, temperature, cloud cover, humidity, wind speed, air transparency, and local meteorological processes. It is characterized by rapid fluctuations, strong nonlinearity, high randomness, and significant multi-timescale coupling.

[0003] From a physical perspective, minute-level or hourly power changes are often affected by cloud movement and local shading, manifesting as short-term spikes, sudden drops, and rapid recovery. On an intra-day scale, changes in solar altitude angle, sunrise and sunset times, and the evolution of weather processes determine the basic shape of the power curve. On a trans-day scale, frontal systems, seasonal changes, regional climate conditions, and continuous shifts between sunny and cloudy weather can cause systematic shifts in the statistical distribution of power over several days.

[0004] In cross-site photovoltaic power prediction tasks, if the original output power or capacity-normalized power is directly used as the transfer target, the model's learned representation often contains three types of differences: the first type is the difference in theoretically generateable power caused by changes in latitude and longitude, season, and intraday solar altitude angle; the second type is the inherent capacity difference of the power station caused by installed capacity, module efficiency, inverter efficiency, and system losses; and the third type is the difference caused by weather disturbances or operational disturbances such as cloud cover, cloudy / sunny weather transitions, power curtailment, and equipment malfunctions. These three types of differences are intertwined in the original power sequence, making it difficult for the model to determine whether the difference in the power curves of two power stations originates from solar geometry and installed conditions, or from transferable weather disturbance patterns. This problem leads to inaccurate construction of positive and negative samples in cross-site projects and further triggers negative transfer.

[0005] While existing technologies employ photovoltaic power prediction methods based on physical models, residual correction, transfer learning, and comparative learning, these methods typically suffer from the following shortcomings: First, residuals are often used as post-processing corrections for prediction errors, rather than being defined as physical deviations from theoretical clear-sky power. Second, comparative learning sample pairs are often constructed based on temporal adjacency, weather labels, or overall curve similarity, lacking a sample construction mechanism that uses physical operational events such as sudden cloud cover, cloud recovery, continuous cloudy days with low output, and peak shaving as transfer units. Third, cross-domain alignment often involves statistical alignment of the entire station's characteristic distribution or the entire time series distribution, making it difficult to ensure that similar weather disturbance events are correctly aligned between different stations. Fourth, when fine-tuning a small number of samples at the target station, only model parameters are usually adjusted, lacking joint calibration of the target station's clear-sky baseline, residual distribution, and event prototype matching weights.

[0006] Therefore, for cross-site photovoltaic power prediction schemes to have stable migration capabilities, they cannot rely solely on capacity-normalized power, whole-time distribution alignment, or model parameter reuse. Instead, they should first separate deterministic differences such as solar geometry, installed capacity, and module installation methods from random disturbance differences such as cloud cover, cloudy / sunny weather transitions, power curtailment and peak shaving, and equipment anomalies. Furthermore, they should use operational events with clear physical meaning as cross-site alignment units and calibrate the clear-sky benchmark, residual distribution, and event prototype matching relationship under limited sample conditions in the target domain. This will prevent negative migration from general comparative learning or overall distribution migration in cross-site photovoltaic prediction scenarios. Summary of the Invention

[0007] To address the problem that the coupling of different power station installed capacities, latitudes and longitudes, solar irradiance phases, equipment efficiency and local weather disturbances in cross-site photovoltaic power prediction makes it difficult to stably migrate the historical operating modes of source domain power stations to small sample prediction scenarios of target domain power stations, this invention provides a cross-site photovoltaic power prediction method and system based on clear-sky residual event prototype migration.

[0008] In a first aspect, the present invention provides a cross-site photovoltaic power prediction method based on clear-sky residual event prototype migration, which adopts the following technical solution:

[0009] A cross-site photovoltaic power prediction method based on clear-sky residual event prototype transfer includes:

[0010] Acquire cross-site multi-source data from photovoltaic power plants in both the source and target domains;

[0011] Perform time alignment, missing anomaly marking, and capacity normalization on cross-site multi-source data;

[0012] Theoretical clear-sky power is constructed based on normalized data;

[0013] Construct a clear-sky perturbation characterization sequence based on theoretical clear-sky power;

[0014] Based on the constructed clear-sky disturbance characterization sequence, a multi-timescale historical window is constructed at each prediction time to generate a multi-timescale sample sequence.

[0015] Identifying photovoltaic operation events based on generated multi-timescale sample sequences;

[0016] A cross-site event prototype library was established based on the identified photovoltaic operation events.

[0017] The event prototype contrastive loss is pre-trained in the source domain and calibrated with few samples in the target domain using a cross-site event prototype library.

[0018] Dynamic fusion of source domain pre-training and target domain few-sample calibration results is performed to reconstruct the future power and prediction range of the target power plant.

[0019] Secondly, a cross-site photovoltaic power prediction system based on clear-sky residual event prototype transfer includes:

[0020] The data acquisition module is configured to acquire cross-site multi-source data from source domain photovoltaic power plants and target domain photovoltaic power plants;

[0021] The preprocessing module is configured to perform time alignment, missing anomaly marking, and capacity normalization on cross-site multi-source data;

[0022] The clear sky module is configured to construct theoretical clear sky power based on normalized data.

[0023] The perturbation module is configured to construct a clear-sky perturbation characterization sequence based on theoretical clear-sky power;

[0024] The window module is configured to construct a multi-timescale historical window at each prediction time based on the constructed clear-sky disturbance characterization sequence, and generate a multi-timescale sample sequence.

[0025] The photovoltaic module is configured to identify photovoltaic operating events based on generated multi-timescale sample sequences;

[0026] The prototype library module is configured to build a cross-site event prototype library based on the identified photovoltaic operation events.

[0027] The calibration module is configured to perform source domain pre-training and target domain few-sample calibration of the event prototype contrastive loss using a cross-site event prototype library.

[0028] The prediction module is configured to dynamically fuse source domain pre-training and target domain few-sample calibration results to reconstruct the future power and prediction range of the target power plant.

[0029] Thirdly, the present invention provides a computer-readable storage medium storing a plurality of instructions adapted for loading and execution by a processor of a terminal device of the cross-site photovoltaic power prediction method based on clear-sky residual event prototype migration.

[0030] Fourthly, the present invention provides a terminal device, including a processor and a computer-readable storage medium, wherein the processor is used to implement various instructions; the computer-readable storage medium is used to store multiple instructions, the instructions being adapted to be loaded and executed by the processor to provide a cross-site photovoltaic power prediction method based on clear-sky residual event prototype migration.

[0031] In summary, the present invention has the following beneficial technical effects:

[0032] First, this invention decomposes the original photovoltaic power into a deterministic clear-sky benchmark and a random disturbance residual by using theoretical clear-sky power. This transforms the cross-site migration target from the original power curve into a clear-sky attenuation coefficient and a clear-sky residual, reducing the interference of differences in installed capacity, latitude and longitude, solar irradiance phase, and theoretically achievable power on transfer learning. At the same time, during the prediction stage, the final power value is reconstructed based on the future theoretical clear-sky power, so that the prediction result has a clear physical meaning and meets the boundary requirements of non-negativity and not exceeding the installed capacity.

[0033] Second, this invention uses photovoltaic operation events such as sudden cloud cover, cloud recovery, continuous cloudy days with low output, peak shaving and abnormal low power as migration units, and establishes a cross-site event prototype library to aggregate similar events in multiple source domain photovoltaic power plants into reusable disturbance prototypes, thereby reducing the mismatch of positive and negative samples caused by similar original power but different disturbance mechanisms, or different original power but the same disturbance mechanism.

[0034] Third, this invention improves the prediction stability and migration efficiency of newly built power plants under small sample conditions by constraining similar photovoltaic operation events to cluster in the characterization space and separating events with different physical mechanisms in the characterization space through event prototype comparison loss constraints, and by jointly calibrating the theoretical clear-sky power benchmark, clear-sky attenuation coefficient distribution, clear-sky residual distribution and event prototype matching weights during the target domain migration stage.

[0035] Fourth, this invention uses short-scale, medium-scale, and long-scale historical windows to characterize short-term cloud cover disturbances, intraday weather events, and cross-day residual drift, respectively, and integrates them through a multi-timescale prediction network. This can take into account both ultra-short-term and short-term photovoltaic power prediction needs and is applicable to engineering scenarios such as grid dispatching, energy storage control, spot trading, and power plant operation and maintenance. Attached Figure Description

[0036] Figure 1 This is a schematic diagram of the overall process of the photovoltaic power prediction method based on the prototype migration of clear-sky residual events of the present invention.

[0037] Figure 2 This is a schematic diagram illustrating the extraction of photovoltaic operation events and the construction of a cross-site event prototype library according to the present invention.

[0038] Figure 3 This is a schematic diagram of the event prototype comparison pre-training architecture of the present invention.

[0039] Figure 4 This is a schematic diagram of the target domain few-sample calibration and future power reconstruction of the present invention.

[0040] Figure 5 This is a schematic diagram of the photovoltaic power prediction system of the present invention.

[0041] Figure 6 This is a comparison chart of the prediction performance of different methods on the target domain test set.

[0042] Figure 7 A graph showing the variation in prediction accuracy under different target domain sample sizes is provided. Detailed Implementation

[0043] The present invention will be further described in detail below with reference to the accompanying drawings.

[0044] Example 1

[0045] Reference Figure 1 This embodiment presents a cross-site photovoltaic power prediction method based on clear-sky residual event prototype transfer. It utilizes historical operating data from source-domain photovoltaic power plants and a limited amount of labeled data from target-domain photovoltaic power plants. The method follows a main line: "multi-source cross-site data preprocessing—theoretical clear-sky power benchmark construction—clear-sky attenuation coefficient and clear-sky residual characterization generation—multi-scale operating event identification—cross-site event prototype transfer learning—target-domain limited-sample calibration—future photovoltaic power reconstruction output." This establishes a cross-site photovoltaic power prediction framework that maintains high prediction accuracy even with insufficient data at the target site. In this embodiment, the clear-sky residual event prototype refers to a cross-site event prototype obtained by encoding clear-sky attenuation coefficient, normalized clear-sky residual, and their dynamic change characteristics using photovoltaic operating event segments as basic samples. The transfer object is not the original power curve itself, but rather the perturbation pattern relative to the theoretical clear-sky power benchmark.

[0046] (1) Multi-site data acquisition and preprocessing

[0047] The core objective of this step is to integrate heterogeneous raw data from mature multi-source photovoltaic power plants and newly built target power plants, unify the sampling interval, time reference, feature units, installed capacity scale, and operating condition labels of each site, and output standardized time-series samples. This solves the problems of inconsistent data sources, varying sampling frequencies, missing and abnormal data, large differences in installed capacity, and inconsistent equipment operating condition recording standards in cross-site scenarios, providing basic input for subsequent clear-sky power decoupling, residual characterization, event recognition, and cross-domain migration modeling.

[0048] The original input data fields required by the method of this invention include at least the historical output power of the photovoltaic power plant, multi-dimensional irradiance data (horizontal and inclined irradiance), ambient temperature, photovoltaic module temperature, wind speed and direction, atmospheric pressure, relative humidity, time series labels, and static attribute data of the power plant. The aforementioned multi-source raw data can be obtained from photovoltaic power prediction supporting data acquisition terminals and business platforms such as the photovoltaic power plant SCADA operation monitoring system, ground meteorological observation stations, numerical weather prediction platforms, power plant ledger configuration systems, and power grid dispatching and power limitation record management systems.

[0049] Let the set of photovoltaic power stations in the source region be:

[0050] ,

[0051] in Indicates the number of power plants in the source region. Indicates the first Historical datasets of power plants in the source region.

[0052] The few-sample dataset of newly built photovoltaic power plants in the target domain is represented as follows:

[0053] ,

[0054] in, This represents a small sample dataset of newly built photovoltaic power plants in the target domain. This represents the total number of continuously labeled samples in the few-sample dataset of the target domain; Indicates the first Group target domain labeled samples, Each sample group contains a continuous time-series observation. Indicates the first The dynamic input features, station static attributes, and operating condition markers corresponding to the group samples. This represents the measured output power of the field station at the corresponding time point for this sample set. This small number of labeled samples are used for target domain calibration of subsequent theoretical clear-sky power benchmarks, clear-sky attenuation coefficient distributions, and clear-sky residual distributions.

[0055] For any power station The collected data is divided into three categories. The first category is time-series dynamic observation characteristics, which change with time, including real-time SCADA output power. Horizontal irradiance Radiance of inclined surface Ambient temperature Photovoltaic module temperature Atmospheric pressure Wind speed ,wind direction relative humidity Cloud cover The first category includes real-time weather types and time series labels. The second category is time series operating condition labels, which generate masks over time, including scheduling and limited-release records and operation and maintenance records. Limited-release masks are generated during the preprocessing stage. and maintenance mask The third category is the static attributes of the station. Its static attributes are fixed and do not change over time. These static attributes of the stations are uniformly encapsulated as static feature vectors. It should include at least the installed capacity, latitude and longitude, altitude, module tilt angle, module azimuth angle, inverter efficiency, system loss coefficient, temperature power coefficient and module nominal operating temperature.

[0056] First, the data from each power station underwent time unification and resampling. Data from different systems was unified to the same time zone and fixed sampling interval. ,in The timeframe can be 5 minutes, 10 minutes, 15 minutes, or 30 minutes. This applies to the original observation sequence. The time-aligned sequence is represented as:

[0057] ,

[0058] in, This indicates the operation of resampling and time alignment at fixed sampling intervals. This indicates a unified time index.

[0059] Then, the missing data is processed by category. Let the missing mask be:

[0060] ,

[0061] For short-term missing data, linear interpolation, spline interpolation, or the mean of adjacent time periods are used for imputation; for long-term missing data, it is not directly used as a valid training sample, but the missing label is retained. The imputed features are represented as follows:

[0062] ,

[0063] in, This represents the interpolation function based on the neighborhood window. Indicates the length of the interpolation reference window.

[0064] Next, abnormal samples are identified and labeled. Abnormal samples include negative power, power exceeding installed capacity, sensor abrupt changes, abnormal irradiance spikes, non-zero power at night, maintenance shutdowns, and dispatch-limited power generation. The anomaly detection is represented as follows:

[0065] ,

[0066] in, For the first The installed capacity of each power station To allow for error margin, This is the threshold for sensor switching. Samples under power limiting, maintenance, or abnormal operating conditions are not removed; instead, their corresponding operating condition labels are retained for subsequent identification of power limiting / peak shaving events and abnormal low-power events.

[0067] Power data is normalized to reduce cross-site migration bias caused by differences in installed capacity between different power plants.

[0068] ,

[0069] in, To prevent extremely small constants with a denominator of zero.

[0070] Robust standardization is applied to meteorological linear features:

[0071] ,

[0072] In the formula, This represents a linear meteorological feature sequence after missing data imputation, including horizontal irradiance, inclined irradiance, ambient temperature, component temperature, atmospheric pressure, wind speed, relative humidity, and cloud cover. This represents the median. This represents the interquartile range. Standardized meteorological features include... , , , , , and .

[0073] Wind direction characteristics Perform periodic encoding to obtain the wind direction encoding vector:

[0074] ,

[0075] Simultaneously, time features are periodically encoded to preserve daily and annual cycle patterns:

[0076] ,

[0077] in, Indicates the hour of the current time. This indicates the current date's position within the year.

[0078] Furthermore, a sample quality score is constructed to screen high-confidence samples and adjust the weights of various samples during the training of the main prediction model; the event recognition stage retains the complete working condition labeling, which is not filtered by the sample quality score.

[0079] ,

[0080] in, Indicates the power effective mask. Indicates the effective irradiance mask. Indicates an anomaly marker. Indicates a limited issuance flag. Indicates a maintenance / repair sign. to , where is the weighting coefficient. For samples with limited power generation and those undergoing maintenance, their sample weights are reduced during the training of the master prediction model, while their operating condition labels are retained during the event identification phase.

[0081] Finally, a unified feature sequence is constructed for each power station:

[0082] ,

[0083] in: ,

[0084] It is a multi-dimensional sample mask vector, which contains four types of operating condition labels: missing, abnormal, limited, and maintenance.

[0085] This step outputs a cross-site standardized feature sequence with a unified sampling interval, capacity-normalized power, robust standardized meteorological features, wind direction periodic coding, time periodic coding, station static attributes, sample quality score, and masks for various operating conditions such as missing, abnormal, limited power generation, and maintenance. This provides a reliable input with quality labels for subsequent construction of theoretical clear-sky power benchmarks, generation of clear-sky attenuation coefficient and clear-sky residual characterization, photovoltaic operation event identification, and cross-site event prototype migration.

[0086] (2) Calculate the theoretical power under clear skies

[0087] This step is used to use photovoltaic astrophysical models to isolate the inherent power output differences of different power plants caused by factors such as geographical location, installed capacity, component installation angle, equipment efficiency and operating temperature, so as to provide a physical benchmark for the subsequent generation of clear sky attenuation coefficient and clear sky residual characterization.

[0088] Based on the cross-site standardized feature sequence output in step one Sample mask and station static attribute vector A theoretical clear-sky power benchmark is constructed for each photovoltaic power station at the corresponding time. The calculation process of theoretical clear-sky power uses the original dimensional observation data after time alignment, missing data imputation, and anomaly labeling in step one, and does not directly use the robustly normalized dimensionless feature sequence. For variables involved in physical modeling calculations, such as irradiance, ambient temperature, module temperature, and atmospheric pressure, the imputed original dimensional sequence is used. The calculation is performed, while the sample mask generated in step one is retained to determine the reliability of the theoretical clear-sky baseline result at the corresponding time. (The superscript is not included in this translation.) This represents the original dimensional observation sequence after missing data filling in step one.

[0089] For the A photovoltaic power station, its static attribute vector At least latitude ,longitude ,altitude Installed capacity Component tilt angle Component azimuth angle Inverter efficiency System loss coefficient Temperature power coefficient and the component's nominal operating temperature .

[0090] For a unified time index According to its corresponding date sequence number Hourly Information And the station's latitude and longitude are used to calculate the sun's geometric position. The sun's declination angle is expressed as:

[0091] ,

[0092] True solar time is obtained by correcting for station longitude and standard time zone longitude. The solar hour angle is expressed as:

[0093] ,

[0094] in, Indicates the first Each power station at any time The corresponding true solar time.

[0095] Solar altitude angle It can be represented as:

[0096] ,

[0097] when At 0:00, it indicates that the time is nighttime or the sun is below the horizon, and the corresponding theoretical clear-sky power is set to 0. Clear-sky horizontal irradiance is expressed as:

[0098] ,

[0099] in, Indicates solar irradiance outside the atmosphere. This represents the clear sky transmittance after taking into account the effects of atmospheric transparency, altitude, and air pressure.

[0100] Solar irradiance outside the atmosphere is expressed as:

[0101] ,

[0102] in, Let be the solar constant. Based on the atmospheric pressure, relative humidity, and altitude obtained in step one, clear-sky transmittance can be expressed as:

[0103] ,

[0104] in, To fill the back atmospheric pressure, Indicates standard atmospheric pressure. This indicates the relative humidity after filling. Indicates the elevation of the power station. , , These are empirical correction coefficients obtained based on pre-calibration of regional meteorological conditions.

[0105] Considering the influence of module tilt angle and azimuth angle on received irradiance, the clear-sky irradiance on the horizontal plane is converted to clear-sky irradiance on the tilted plane of the module. Module incident angle Measured by solar altitude angle, solar azimuth angle, and component tilt angle. and component azimuth angle The calculated effective irradiance model for a tilted surface under clear sky is expressed as follows:

[0106] ,

[0107] in, This represents the clear-sky irradiance of the inclined surface calculated by the model. Indicates the incident angle of the component. This represents a minimal constant to prevent the denominator from being zero.

[0108] Based on the measured irradiance of the tilted surface after filling, as output in step one. Weighted corrections are applied to the model's calculated values:

[0109] ,

[0110] in, This represents the corrected clear-sky irradiance of the tilted surface. This represents the fusion coefficient between the model-calculated values ​​and the measured infill values, and .

[0111] Subsequently, the theoretical output power is corrected based on the component temperature, and the component temperature is selected based on the validity of the corresponding component temperature dimension in the missing mask from step one. When the component temperature is valid, the measured component temperature after filling in the missing information is used. When the module temperature is invalid, the module temperature is estimated based on the filled ambient temperature and the corrected clear-sky irradiance of the tilted surface. The estimation formula is as follows:

[0112] ,

[0113] in, This indicates an estimated component temperature. This represents the ambient temperature after filling, obtained in step one. Indicates the first The nominal operating temperature of the components in each power station. This represents the corrected clear-sky irradiance of the tilted surface.

[0114] Then, the component temperature used for theoretical clear-sky power calculations is uniformly expressed as:

[0115] ,

[0116] in, This indicates the component temperature ultimately used for theoretical clear-sky power calculations.

[0117] The theoretical power under clear skies is expressed as:

[0118] ,

[0119] in, Indicates the first Each power station at any time The theoretical clear-sky power, Indicates installed capacity. Indicates standard test irradiance. Indicates inverter efficiency. Represents the system loss coefficient.

[0120] Indicates the temperature power coefficient. Indicates the standard test temperature.

[0121] To ensure that the theoretical clear-sky power conforms to the physical boundary of photovoltaic output, constraints are imposed on it:

[0122] ,

[0123] in, This represents the theoretical clear-sky power after boundary constraints, and its value is not less than 0 and does not exceed the installed capacity. .

[0124] Final output theoretical clear-sky power sequence:

[0125] ,

[0126] in, Indicates the first The theoretical clear-sky power sequence of each power station, Indicates the first The total number of valid time-series samples for each photovoltaic power station. This theoretical clear-sky power series serves as the physical benchmark for subsequent calculations of the clear-sky attenuation coefficient and clear-sky residual. It is used to isolate the inherent output differences of each photovoltaic power station caused by factors such as installed capacity, geographical location, component installation angle, inherent equipment efficiency, and operating temperature. This allows subsequent models to focus on dynamic disturbance patterns that can migrate across power stations, such as cloud cover, overcast weather disturbances, dispatch-limited power generation, and equipment anomalies.

[0127] (3) Construct the clear sky attenuation coefficient and clear sky residual.

[0128] This step, based on the theoretical clear-sky power benchmark obtained in step two, transforms the measured power of each photovoltaic power station into a disturbance representation relative to the clear-sky benchmark. This reduces the impact of differences in installed capacity, geographical location, component installation method, and inherent efficiency among different power stations on cross-site migration modeling. Unlike directly using the original power sequence for migration, this step constructs a representation of "how much the actual power deviates from the theoretical clear-sky power," enabling subsequent models to focus on learning common cross-site operational change patterns such as cloud cover, cloudy day attenuation, peak shaving, equipment anomalies, and local weather disturbances.

[0129] Based on the already obtained number Measured power sequence of a photovoltaic power station after time alignment, missing data filling, and anomaly marking. Installed capacity Sample quality score Multidimensional sample mask Limited distribution mask Inspection mask Exception markers Theoretical clear-sky power after boundary constraints This step further constructs the clear-sky attenuation coefficient, clear-sky residual, normalized clear-sky residual, and their dynamic variation characteristics.

[0130] First, to avoid distortion in proportional calculations due to excessively low theoretical clear-sky power at night or low solar altitude angles, an effective threshold for theoretical clear-sky power is set. This is used to filter out times when theoretical power is close to zero, such as early morning, evening, and night, and to construct effective markers for clear skies:

[0131] ,

[0132] in, Indicates the first Each power station at any time Is it within a suitable period for calculating the proportion of clear skies?

[0133] Define the clear-sky attenuation coefficient during the effective clear-sky period:

[0134] ,

[0135] in, Indicates the first Each power station at any time The clear-sky attenuation coefficient, To fill the gap in the measured power, Theoretical clear-sky power, To prevent extremely small constants with a denominator of zero.

[0136] To avoid unreasonable extreme values ​​in the attenuation coefficient due to sensor errors or deviations in the theoretical clear-sky model, boundary constraints are applied to the clear-sky attenuation coefficient:

[0137] ,

[0138] in, This represents the clear-sky attenuation coefficient after constraints. This represents the upper limit of the attenuation coefficient. Further definition of clear-sky residual:

[0139] ,

[0140] in, Indicates the first Each power station at any time Clear-sky residuals describe the absolute deviation between measured power and theoretical clear-sky power.

[0141] Considering the different installed capacities of different power plants, directly comparing absolute residuals would still be affected by the capacity scale. Therefore, a normalized clear-sky residual is constructed:

[0142] ,

[0143] in, This represents the normalized clear sky residual. Indicates the first The installed capacity of each power station.

[0144] To describe the rate of change of the disturbance, the change in the clear-sky attenuation coefficient is constructed:

[0145]

[0146] in, It represents the change in the clear sky attenuation coefficient between adjacent moments, in order to depict the sudden drop caused by the rapid obstruction of clouds, and the recovery process after the clouds move away.

[0147] Simultaneously, the normalized clear-sky residual variation is constructed:

[0148] ,

[0149] in, It represents the change in normalized clear-sky residuals between adjacent time points, describing dynamic characteristics such as short-term upslopes, rapid drops, and recovery slopes.

[0150] To describe the intensity of the disturbance over a period of time, a length of... Calculate the intensity of local residual fluctuations within a historical window:

[0151] ,

[0152] in, Indicates the first Each power station at any time forward Normalized clear-sky residual fluctuation intensity within each sampling point; The preset window length; This represents the function for calculating standard deviation.

[0153] Further construct the local residual mean:

[0154] ,

[0155] in, Indicates length is The normalized clear-sky residual mean within a historical window is used to represent the degree of continuous deviation of actual output from the clear-sky benchmark over a period of time.

[0156] Finally, the clear-sky perturbation representation vector is constructed:

[0157]

[0158] in, Indicates the first Each power station at any time The clear-sky disturbance characterization vector; This represents the clear-sky attenuation coefficient after constraints. Indicates residuals in clear skies; This represents the normalized clear sky residual; and Indicates the characteristics of disturbance changes; and This indicates the intensity of residual fluctuations and the degree of average deviation within a local window; Indicates a valid clear sky marker; and These are the sample quality score and the multidimensional sample mask, respectively.

[0159] for At low solar altitude angles or nighttime hours, the proportional clear-sky attenuation coefficient is not used as an effective migration characteristic. Set to 0 or the default value, and mark as clear sky. This proportional feature is marked as invalid to avoid distortion of proportional features due to excessively low theoretical clear-sky power. For samples marked as limited power generation, maintenance, or anomalies, they are not directly deleted but are instead added to the clear-sky disturbance characterization vector. The corresponding sample quality score is retained. and multidimensional sample mask This enables subsequent operational event identification to distinguish between output reduction caused by weather disturbances and output reduction caused by non-weather reasons.

[0160] The output clear-sky disturbance characterization sequence is denoted as:

[0161] ,

[0162] in, Indicates the first A clear-sky disturbance characterization sequence for a photovoltaic power station under a unified time index. This represents the total number of valid time-series samples for the power plant. This sequence transforms the original power prediction problem into a degradation and residual modeling problem under a clear-sky baseline, providing a unified and comparable input basis for subsequent construction of multi-timescale historical windows, identification of photovoltaic operation events, and cross-site event prototype transfer.

[0163] (4) Constructing a multi-timescale historical window

[0164] Based on the clear-sky disturbance characterization sequence generated in step three, this step constructs short-scale, meso-scale, and long-scale historical windows for each prediction time, enabling subsequent models to simultaneously acquire minute-level rapid disturbances, hour-level weather evolution, and cross-day residual distribution change information.

[0165] Let the predicted time be The predicted step size is The future moment to be predicted is .in, The preset prediction step size is used. To simultaneously describe the operational status across different time ranges, the short-scale window length is set to [value]. The length of the mesoscale window is The length of the long-scale window is And satisfy:

[0166] ,

[0167] in, It mainly covers the most recent sampling points. Reflecting the continuous changes within the day, This reflects the evolution of trans-day disturbances and changes in residual distribution. The lengths of all three windows are based on the sampling interval unified in step one. It is the basic unit.

[0168] For the prediction time Constructing short-scale historical windows:

[0169] ,

[0170] in, Indicates the first Each power station at the predicted reference time The corresponding short-scale historical window. This window mainly retains the changes in clear-sky attenuation coefficient, residual abrupt changes, local fluctuations, and operating condition markers over a recent period. It is suitable for describing rapid changes such as rapid cloud cover, cloud recovery, short-term sensor anomalies, and power ramp-ups.

[0171] Constructing a mesoscale historical window:

[0172] ,

[0173] in, Indicates the first Each power station at the predicted reference time The corresponding mesoscale historical window. This window covers a longer range of intraday variation than the short-scale window, and is used to describe weather evolution, changes in overall attenuation levels, peak reduction patterns, and continuous low-output segments during periods such as morning to noon and afternoon to evening.

[0174] Constructing long-scale historical windows:

[0175] ,

[0176] in, Indicates the first Each power station at the predicted reference time The corresponding long-scale historical window. This window is used to represent the residual distribution shift caused by cross-day weather processes, continuous rain, equipment aging, or long-term operation and maintenance, enabling subsequent models to identify whether the overall operating baseline of the target power plant has changed recently.

[0177] To enable multi-scale windows to be incorporated into a unified modeling process, statistical summaries are further constructed for each of the three windows. The decay rate of the short-scale window can be expressed as:

[0178] ,

[0179] in, It represents the range of variation of the clear-sky attenuation coefficient within a short-scale window, and is used to characterize the intensity of short-time attenuation fluctuations.

[0180] The average normalized residual of the mesoscale window can be expressed as:

[0181] ,

[0182] in, This represents the mean of the normalized clear-sky residuals within a mesoscale window, used to describe the degree of overall deviation from the clear-sky baseline during the day.

[0183] The residual distribution drift over a long-scale window can be expressed as:

[0184] ,

[0185] in, It indicates the degree of deviation of the short-term residual mean from the long-term residual mean, and determines whether the current output status has changed relative to the long-term operating baseline.

[0186] By combining statistical summaries from three scales, a multi-scale summary vector is constructed:

[0187] ,

[0188] in, Indicates the first Each power station at the predicted reference time The multi-scale statistical summary provides more compact auxiliary features for subsequent event recognition.

[0189] Furthermore, the static attribute vector of the site output in step one is... Time period coding Wind direction cycle coding Sample quality score and multidimensional sample mask Combined with three historical windows, this yields multi-scale input samples for the prediction time:

[0190] ,

[0191] in, Indicates the first Each power station at the predicted reference time Complete multi-timescale input samples.

[0192] For starting periods with insufficient historical length, i.e. The window length is padded using forward padding, zero padding, and effective fragment masking, and the padding markers are retained in the multidimensional sample mask to prevent padded samples from being mistaken for true observations. Let the effective window proportion be:

[0193] ,

[0194] in, This represents the overall effective proportion within a long-scale window. The clear sky markers obtained in step three are valid. The sample quality score obtained in step one. When... Below the preset window quality threshold At this time, the window is not used as a high-confidence training sample, but its condition label can still be retained for event recognition analysis.

[0195] The final output is a multi-timescale sample sequence:

[0196] ,

[0197] in, Let represent the multi-timescale input sample sequence of the i-th photovoltaic power station. For the first Total number of valid time-series samples for each power station To predict the step size, this sequence extends the single-point clear-sky disturbance characterization to include historical context encompassing short-term variations, intraday evolution, and cross-day background, providing structured input for subsequent photovoltaic operation event identification, cross-site event prototype library construction, and target domain few-sample migration prediction.

[0198] (5) Identify photovoltaic operation events

[0199] This step, based on the clear-sky disturbance characterization sequence generated in step three and the multi-timescale historical window constructed in step four, identifies photovoltaic operation events with clear physical meaning and merges moments that continuously meet the same event conditions into event segments. The preceding steps have already converted the original power sequence into attenuation, residual, and dynamic change characteristics relative to the theoretical clear-sky power. This step further identifies typical operation events such as sudden cloud cover drop, cloud recovery, continuous cloudy low output, peak shaving, and abnormally low power based on these characteristics. This ensures that subsequent inter-site migration no longer uses the entire original power curve as the object, but rather uses event segments with physical meaning as migration units.

[0200] This step is based on the clear-sky perturbation representation vector and multi-timescale input samples. Perform event identification.

[0201] First, identify clear-sky stable events. A clear-sky stable event indicates that the actual output of a photovoltaic power station is close to the theoretical output under clear skies, and the degradation coefficient and residual fluctuations are small over a short period. The criteria for this event can be expressed as:

[0202] ,

[0203] Among them, table Show the first Each power station at any time Whether it was identified as a clear-sky stable event; This represents the threshold value for the clear-sky attenuation coefficient. This represents the threshold value for the change in the clear-sky attenuation coefficient. This represents the threshold for the intensity of local residual fluctuations. (This is achieved through constraints.) This is used to ensure that the change in the attenuation coefficient is small at that moment.

[0204] Secondly, identify cloud cover abrupt events. Cloud cover abrupt events typically manifest as a rapid decrease in actual power relative to a clear-sky baseline, even though the theoretical clear-sky power remains within the effective power generation period, and the clear-sky attenuation coefficient decreases significantly within a short period. Therefore, a comparison interval for the abrupt event is set as follows: The event of sudden cloud cover is then classified as follows:

[0205] ,

[0206] in, Indicates a sudden cloud cover event. This indicates the threshold for the decrease in the clear-sky attenuation coefficient. This represents the threshold for the normalized residual to decrease. This indicates that the system was not under power restrictions at that time. This indicates that the device is not under maintenance at that time.

[0207] Next, identify post-cloud recovery events. Post-cloud recovery typically occurs after a sudden cloud cover event, characterized by a rapid recovery of actual power relative to theoretical clear-sky power, and a significant increase in the clear-sky attenuation coefficient within a short period. Let the recovery comparison interval be... Its discrimination condition can be expressed as:

[0208] ,

[0209] in, Indicates the event marker for cloud recovery. Indicates the interval between recovery magnitudes. This indicates the recovery threshold of the clear sky attenuation coefficient. This represents the normalized residual recovery threshold. This event describes the process by which the power output approaches the clear-sky baseline again after the clouds have moved away.

[0210] Further identification of consecutive cloudy days with low power output events is needed. Consecutive cloudy days with low power output are not single-point abrupt changes, but rather are characterized by a persistently low clear-sky attenuation coefficient, negative residual mean, and low volatility within a mesoscale or long-scale window. Therefore, the mesoscale window length is set as... The average clear-sky attenuation coefficient within the window is expressed as:

[0211] ,

[0212] in, Indicates the first The average clear-sky attenuation coefficient of a power station within a mesoscale window. The criteria for identifying consecutive cloudy days and low power output events are:

[0213] ,

[0214] in, This indicates a continuous period of low power output due to overcast skies. This represents the threshold for the mesoscale average attenuation coefficient. This represents the threshold value of the local residual mean. This represents the threshold for the intensity of local residual fluctuations. This event is used to distinguish weather processes that are characterized by sustained low output but relatively gradual changes.

[0215] Next, identify peak shaving events. Peak shaving typically manifests as a theoretically high clear-sky power, but the measured power is suppressed to near a certain plateau, and the peak shaving mask in step one is in a valid state. Its criterion can be expressed as:

[0216] ,

[0217] in, This indicates a flag for a peak-shaving event. This indicates the measured power after filling. This represents the theoretical clear-sky power. This indicates the threshold for the difference between the power reduction and peak load factoring. This represents the threshold value for the decay change near the firing limit platform. This event depends on the firing limit mask. This is used to distinguish between dispatch restrictions and natural attenuation caused by weather.

[0218] Finally, identify abnormally low power events. Abnormally low power typically manifests as theoretical clear-sky power being effective, but actual power being significantly lower than the theoretical clear-sky power, without any indication of power limitation or maintenance. This can be further assessed by combining the abnormality indicators from step one. The criteria for this assessment can be expressed as:

[0219] ,

[0220] in, Indicates an abnormally low power event flag. This represents the threshold value for the clear-sky attenuation coefficient corresponding to abnormally low power. This indicates the obtained anomaly flag. This event is used to describe abnormally low output caused by equipment malfunction, obstruction, sensor malfunction, or inverter malfunction under non-limited power generation and non-maintenance conditions.

[0221] After completing the single-moment event discrimination, consecutive moments that meet the discrimination criteria under the same event type are merged into an event segment. Let the set of event types be:

[0222] ,

[0223] in, This represents the set of event types identified in this step, which includes, in order: clear and stable skies, sudden cloud cover, cloud cover recovery, continuous overcast skies with low power output, power restriction and peak shaving, and abnormally low power. For any event type... If within the time interval Internal continuity satisfies:

[0224] ,

[0225] And the duration of the event satisfies:

[0226] ,

[0227] Then the continuous intervals will be merged into a single event segment:

[0228] ,

[0229] in, Indicates the first The first of the photovoltaic power stations The types are runtime event fragments, Indicates event type The minimum duration threshold is used to filter out occasional noise points. to This is the sequence of clear-sky disturbance characterization vectors obtained in step three.

[0230] Furthermore, an event description vector is generated for each event fragment:

[0231] ,

[0232] in, Indicates the event type, and These represent the start and end times of the event, respectively. This represents the average clear-sky attenuation coefficient within an event segment. This represents the average normalized clear-sky residual within the event segment. This indicates the magnitude of change in the clear-sky attenuation coefficient within the event segment. This indicates the magnitude of the change in the normalized residual within the event segment. This represents the average sample quality score within an event segment.

[0233] The above statistics are defined as follows:

[0234] ,

[0235] in, Represents event fragments In the start and end intervals The average clear-sky attenuation coefficient within the area, Indicates the time of the i-th power station The clear sky attenuation coefficient after constraint.

[0236] ,

[0237] in, Represents event fragments Within the start and end intervals Average normalized clear-sky residuals Indicates the time of the i-th power station The normalized clear sky residual.

[0238] ,

[0239] in, This indicates the variation range of the clear-sky attenuation coefficient within an event segment, used to describe the range of attenuation of actual power output relative to theoretical clear-sky power output during the event.

[0240] ,

[0241] in, It represents the variation range of the normalized clear-sky residual within an event segment, used to describe the fluctuation range of power deviation from the clear-sky reference during the event.

[0242] ,

[0243] in, This represents the average sample quality score within an event segment. Indicates the obtained first Each power station at any time The sample quality score.

[0244] The final output of this step is number [number]. A collection of event fragments from individual power stations:

[0245] ,

[0246] in, Indicates the first The set of photovoltaic operation events identified by each power station Indicates the first Among the power stations, the type is The number of event fragments. This event set breaks down continuous power sequences into running events with clear physical meaning, providing a sample basis for subsequent cross-site event prototype library construction, event-level comparative pre-training, and target domain few-shot transfer.

[0247] Unlike methods that directly construct training samples based on fixed time slices or daily curves, this step first identifies events based on attenuation, residuals, and operating condition masks under a theoretical clear-sky power benchmark, and then uses event segments as transfer samples. Therefore, the same type of cloud cover abrupt drops, cloud recovery, or power rationing events can be compared within a unified clear-sky residual space, even if they occur at power plants with different latitudes and longitudes and different installed capacities. Furthermore, segments with similar original power curve shapes but different physical causes will be distinguished based on event type and operating condition mask.

[0248] (6) Establish a cross-site event prototype library

[0249] This step is used to build a cross-site event prototype library based on the set of photovoltaic operation event fragments output in step five, and to call the event fragments. and its event description vector This method maps similar events from different source power plants to a unified representation space and aggregates them in a weighted manner into event prototypes that can be transferred across stations.

[0250] In this embodiment, the event encoder Local event feature extraction function It consists of an attention convergence function and a local event feature extraction function. The local event feature extraction function is used to extract the clear sky disturbance features at each moment in the event segment, and the attention convergence function is used to assign higher weights to key moments within the event segment, ultimately obtaining an event embedding vector of uniform dimension. Thus, the event encoding process in step six has a clear computational chain: first, it is composed of... , and The local event features are obtained, and then the local event features are weighted and aggregated to obtain the event embedding.

[0251] Joint encoding of event fragments, event description vectors, and station static attributes:

[0252] ,

[0253] For event fragments At any time Clear-sky disturbance characterization First, construct local event features:

[0254] ,

[0255] Then, attention weights are assigned to local features at different times within the event segment:

[0256] ,

[0257] in, Indicates the time within an event segment Attention weights This represents the attention scoring function, used to map local event features to importance scores. This weight is used to highlight critical moments in an event segment, such as power drops, recovery transitions, peak-shaving plateaus, or unusually low output.

[0258] The event embedding vector is represented as:

[0259] ,

[0260] in, It is a unified dimensional representation obtained by weighted aggregation of local event features within an event fragment.

[0261] For any event type Collect event embedding vectors belonging to this event type from all source domain power plants to form a set of event embeddings of the same type:

[0262] ,

[0263] in, Indicates event type The corresponding source domain event embedding collection, This indicates the number of power plants in the source region.

[0264] To calculate the aggregate weight of similar events, the event type is first obtained. Initial center:

[0265] ,

[0266] in, Indicates event type The initial event center, Represents a set The total number of event embedding vectors.

[0267] Then, the aggregation weight is calculated based on the event fragment quality score and the event embedding similarity:

[0268] ,

[0269] in, Represents event fragments In event type Weights in prototype aggregation The event description vector in step five The average sample quality score already included. This represents the weighting coefficient for the quality score. This represents the distance penalty weight coefficient. This represents the squared Euclidean distance.

[0270] Finally, by weighted aggregation of event embeddings under the same event type, the event prototype is obtained:

[0271] ,

[0272] Repeat the above process for all event types to obtain the cross-site event prototype library:

[0273] ,

[0274] in, This indicates the cross-site event prototype library. This represents the set of event types defined in step five.

[0275] Therefore, by compressing similar operational events from multiple source power plants into a unified prototype representation, subsequent event-level comparative pre-training can focus on the goal of "similar events being close together and different types of events being separated" to carry out representation learning, thus completing the transformation from "event fragments" to "cross-station transferable event prototypes".

[0276] (7) Event-level contrastive pre-training and target domain fine-tuning

[0277] This step is based on the cross-site event prototype library. This process involves pre-training event-level comparisons of event fragments from multiple sites in the source domain, followed by calibration on a small sample dataset of newly built photovoltaic power plants in the target domain. This allows the target domain model to reuse the typical clear-sky disturbance event patterns already learned from the source domain power plants. This step comprises two phases: source domain event-level pre-training and target domain small sample calibration. The source domain phase focuses on learning event representations, bringing similar events closer together and separating different types of events. The target domain phase adjusts the matching relationship between the target site and the source domain event prototypes, and calibrates the clear-sky attenuation distribution and residual distribution of the target site.

[0278] First, define the similarity function between event embeddings and event prototypes:

[0279] ,

[0280] in, Represents cosine similarity. Indicates event type The prototype of cross-site incidents.

[0281] Based on this similarity, the event prototype contrast loss is defined as:

[0282] ,

[0283] in, Indicates the loss compared to the event prototype. This represents the temperature coefficient for comparative learning. The loss factor is used for event fragments. It is closer to the prototype corresponding to its real event type. and away from prototypes of other event types. .

[0284] Simultaneously, a multi-scale prediction network is constructed. Input samples at multiple time scales Using event prototype matching information and station static attributes as input, predict the clear-sky attenuation coefficient, normalized clear-sky residual, power value, and event category probability for future moments:

[0285] ,

[0286] in, Indicates the future Predicted clear-sky attenuation coefficient at any given time. This represents the predicted normalized clear-sky residual. Indicates predicted power. Indicates the predicted probability of event categories. This indicates the predicted network parameters.

[0287] Predictive loss is defined as:

[0288] ,

[0289] in, , , This is the loss weighting coefficient.

[0290] To constrain the event category prediction results, an event classification loss is constructed:

[0291] ,

[0292] in, This represents the real event category label obtained in step five. Indicates the predicted probability of event categories. To prevent extremely small constants where the logarithmic term is zero.

[0293] Further, a physical consistency loss is constructed. Based on the predicted clear-sky attenuation path and the residual path, two power reconstruction results can be obtained:

[0294] ,

[0295] ,

[0296] in, This represents the predicted power obtained from the clear-sky attenuation path. This represents the predicted power obtained from the normalized clear sky residual path.

[0297] Physical consistency loss is defined as:

[0298] ,

[0299] The total loss for source domain pre-training is:

[0300] ,

[0301] in, , , These represent the weight coefficients of the event prototype comparison loss, event classification loss, and physical consistency loss, respectively.

[0302] In the target domain few-sample calibration stage, the target domain few-sample dataset obtained in step one is used as a basis. Following the processing steps two through six, the target domain few-sample event embedding set is obtained:

[0303] ,

[0304] in, Indicates the target domain. The types are Event embedding, This represents the set of event types that occur in a small number of samples within the target domain. Indicates that the target domain has a type of The number of event fragments.

[0305] Calculate the mean representation of events in the target domain:

[0306] ,

[0307] in, This represents the mean of the few-sample event embeddings in the target domain. This represents the total number of few-sample event fragments in the target domain. The matching weight of the target domain to the event prototype of the source domain is defined as:

[0308] ,

[0309] in, Indicates the target domain and event type Source domain prototype matching weights, This represents the target domain prototype matching temperature coefficient.

[0310] Further construct the target domain calibration prototype:

[0311] ,

[0312] in, This represents a hybrid prototype representation of the target domain based on the source domain event prototype library. During target domain fine-tuning, the main parameters of the event encoder are frozen, and only the target domain site embedding parameters, normalization parameters, prototype matching weights, and prediction output layer parameters are updated. The target domain calibration loss is defined as:

[0313] ,

[0314] in, This indicates the prediction loss for a small number of samples in the target domain. This represents the physical consistency loss of the target domain. This represents the prediction network parameters after fine-tuning in the target domain. Represents the source domain pre-trained parameters. , , The loss weights are calibrated for the target domain. This step adjusts the target station's unique clear-sky attenuation distribution, residual distribution, and event prototype matching relationship using a small number of target domain samples, reducing the risk of negative migration caused by direct migration.

[0315] Furthermore, a lightweight Adapter module is set up during the target domain fine-tuning stage. This Adapter module is located between the hidden representation layer and the prediction output layer of the multi-scale temporal encoder. Its inputs are the target domain historical window encoding vector, the target domain site embedding vector, and the target domain hybrid prototype representation; its output is the target domain correction representation. During target domain fine-tuning, the parameters of the source domain backbone encoder and the cross-site event prototype library remain unchanged. Only the Adapter module parameters, target domain normalization parameters, site embedding parameters, prototype matching weights, and prediction output layer parameters are updated. This ensures that a small number of target domain samples are primarily used to correct site-specific biases without disrupting the event prototype structure already learned in the source domain.

[0316] The target domain correction representation can be expressed as hT'=hT+gA([hT,eT,cT];θA), where hT represents the target domain multi-scale temporal coding vector, eT represents the target domain site embedding vector, cT represents the target domain hybrid prototype representation, gA represents the adapter mapping function, and θA represents the adapter parameters. Through this residual adaptation structure, the target domain model can complete local parameter correction with a small number of labeled samples, while maintaining the cross-site transferability of the source domain event representation.

[0317] (8) Output the attenuation coefficient or residual and reconstruct the power prediction value.

[0318] This step, based on the target domain prediction network, outputs the predicted power value, clear-sky attenuation coefficient, normalized clear-sky residual, event category, and prototype matching weights for the target domain photovoltaic power plant at future times. Furthermore, it combines the clear-sky disturbance prediction output by the model with the theoretical future clear-sky power benchmark to reconstruct the final photovoltaic power prediction result.

[0319] To ensure the reliability of both the clear-sky attenuation path and the residual path under different operating conditions, this step dynamically generates fusion weights based on the predicted event type at future times, the target domain prototype matching weights, and the prediction uncertainty. Let the clear-sky attenuation path weights be... The residual path weight is Both conditions are met:

[0320] ,

[0321] in, and These are used to measure the reliability of the clear-sky attenuation path and the residual path under the current prediction conditions, respectively. The clear-sky attenuation path is more suitable for scenarios with stable clear skies and weak disturbances; the residual path is more suitable for scenarios with significant deviations from the clear-sky baseline, such as cloud cover, limited power generation, and abnormally low power. Therefore, this step defines the following path reliability score based on event category probability, target domain prototype matching weight, and prediction uncertainty:

[0322] ,

[0323] ,

[0324] in, This indicates the reliability score of the clear-sky attenuation path. This indicates the residual path reliability score; This indicates the probability that a clear, stable event will occur at a future time. This represents the matching weight between the target domain and the clear-sky stable prototype; This represents a set of disturbance events, including sudden cloud cover, cloud recovery, continuous cloudy days with low power output, power rationing and peak reduction, and abnormally low power output. Indicates the target domain and event type Prototype matching weights; Indicates the score of prediction uncertainty. These are the weighting coefficients.

[0325] The two reliability scores are further normalized into a fusion weight:

[0326] ,

[0327] ,

[0328] Among them, the larger the fusion weight, the more reliable the corresponding path is under the current prediction conditions. When the future time is closer to a clear-sky stable event, Increased density leads to a higher proportion of clear-sky attenuation paths; when future times are closer to disturbances such as cloud cover, limited power generation, or anomalies. As the number of residual paths increases, the proportion of residual paths also increases.

[0329] Based on the above fusion weights, the two prediction results are fused:

[0330] ,

[0331] To ensure that the output results meet the physical boundaries of photovoltaic power output, constraints are imposed on the fused prediction values:

[0332] ,

[0333] in, Indicates the target domain power station at a future time. The final predicted power.

[0334] Furthermore, the prediction interval is constructed. First, the interval width coefficient is generated based on the prediction uncertainty score, the event disturbance intensity, and the target domain prototype matching dispersion:

[0335] ,

[0336] in, Indicates future time The interval width coefficient, This indicates that future moments belong to the event type. The predicted probability, This represents the maximum matching weight of the target domain to the event prototype of the source domain. This is the interval width adjustment coefficient. The higher the prediction uncertainty, the higher the probability of the disturbance event, and the more dispersed the prototype matching, the larger the interval width.

[0337] The upper and lower bounds of the prediction interval are defined as follows:

[0338] ,

[0339] ,

[0340] Ultimately, the future will be output. Power prediction sequence at each prediction time:

[0341] ,

[0342] Output the predicted interval sequence:

[0343] ,

[0344] And output the prediction explanation information:

[0345] ,

[0346] in, This includes predicting the clear sky attenuation coefficient, predicting the normalized clear sky residual, predicting the event category probability, target domain prototype matching weight, two-path fusion weight, interval width coefficient, and prediction uncertainty score.

[0347] The focus of this step is not on redefining the clear-sky attenuation path or residual path, but rather on dynamically fusing the existing two prediction results based on event type, prototype matching relationship, and prediction uncertainty, and outputting the final power value, prediction range, and interpretation information that satisfy the physical boundaries. This processing enables the target domain prediction results to inherit the perturbation experience from the source domain event prototype, and to perform interpretable power reconstruction by combining the target station's own theoretical clear-sky power and installed capacity boundaries.

[0348] Therefore, this embodiment forms a closed loop in the prediction phase: "event prototype matching—two-path disturbance prediction—physical boundary reconstruction—interval and interpretation information output." This closed loop ensures that the model output includes not only the future power value of the target domain, but also the event category probability, prototype matching weight, and path fusion weight that led to the prediction result. This facilitates grid dispatching or power plant operation and maintenance personnel in determining whether the prediction error mainly originates from weather disturbances, power generation restriction conditions, or the target station's baseline deviation.

[0349] Experimental verification

[0350] To ensure that the accompanying drawings correspond to the technical process, system structure, and experimental results of this invention, the actual content in the drawings is now explained in conjunction with the drawings. Figures 1 to 7 The following explanation is provided.

[0351] Figure 1 This is an overall flowchart of the method of the present invention. Figure 2This diagram illustrates the construction of the source station event fragment and cross-site event prototype library. In the diagram, the event fragment sets of source station A, source station B, and source station C first enter the event rule filtering stage. The filtering types include typical scenarios such as sudden cloud cover, post-cloud recovery, continuous cloudy days, peak shaving and power reduction, and abnormally low power. Figure 3 This diagram illustrates the pre-training process for event prototype comparison. The event fragment feature vector is input to a transferable encoder, which extracts shared temporal representations and event semantic representations, generating event embedding vectors. Figure 4 This diagram illustrates the target domain calibration and fine-tuning process. The pre-training parameters, the cross-site event prototype library, and the target station's few-sample labeled data are all input into the target domain calibration unit. The calibration unit includes clear-sky baseline calibration, residual distribution calibration, event prototype matching, and boundary constraint calibration. Figure 5 This is a schematic diagram of the photovoltaic power prediction system of the present invention. The diagram is divided into three layers according to the system hierarchy: data access layer, model service layer, and prediction application layer. The data access layer includes meteorological services, SCADA historical power, power plant configuration, and power grid dispatching and power rationing records. The model service layer includes data cleaning and time alignment, clear sky benchmark decoupling module, residual characterization construction module, event identification and prototype library, and target domain calibration and fine-tuning. Figure 6 A bar chart comparing the errors of different prediction methods on the target domain test set is provided. The horizontal axis lists six methods: ARIMA, LSTM, single-scale T, multi-scale T, supervised pre-training, and the present invention. The vertical axis represents the error value, and the bars show MAE and RMSE. In the figure, the MAE of the present invention is 0.65 and the RMSE is 0.91, both lower than the other comparison methods, indicating that the prediction error of the target domain is reduced after introducing clear sky residuals, event prototype transfer, and target domain calibration. Figure 7 Line graphs showing the change in prediction error under different target domain labeled sample sizes are provided. The horizontal axis represents the target domain labeled data size, specifically 1 week, 2 weeks, and 4 weeks, while the vertical axis represents the error value. The lines represent MAE and RMSE, respectively. As the target domain labeled sample size increases from 1 week to 4 weeks, MAE decreases from 0.758 to 0.613, and RMSE decreases from 1.046 to 0.868. This indicates that the present invention can complete cold-start prediction under small sample conditions, and the prediction accuracy will further improve as the target domain labeled data increases.

[0352] To verify the effectiveness of the method of this invention, experiments were conducted using publicly available photovoltaic (PV) power plant operation data and a target domain partitioning method with a small sample size. Two PV power plants with sufficient data and significant climate differences were selected as the source domain, and a newly commissioned power plant was selected as the target domain. The source domain data spanned from January 2023 to December 2025, with a sampling interval of 15 minutes; the target domain only provided labeled data for the first two weeks for fine-tuning, and the test set was the subsequent month. The evaluation metrics used were Mean Absolute Error (MAE), Root Mean Square Error (RMSE), Mean Absolute Percentage Error (MAPE), and Normalized Root Mean Square Error (nRMSE). All comparison methods used the same target domain labeled samples, the same test interval, the same prediction step size, and the same data preprocessing rules. Except for differences in model structure and migration strategy, the input meteorological characteristics, power label processing methods, and evaluation metrics were kept consistent.

[0353] Table 1. Comparison of prediction performance of different methods on the target domain test set.

[0354] As shown in Table 1, the method of this invention outperforms the comparative methods in all evaluation metrics. Compared with LSTM trained from scratch, the method of this invention reduces MAE by approximately 26.90% and RMSE by approximately 24.82%, indicating that by decoupling from the theoretical clear-sky power benchmark, using a cross-site event prototype library, and event-level self-supervised transfer pre-training, the prediction accuracy in small-sample scenarios in the target domain can be significantly improved. Compared with supervised pre-training + fine-tuning methods, the method of this invention still has significant advantages, indicating that using clear-sky residual events as transfer units is more conducive to reusing typical weather disturbance experience from the source station than directly transferring the entire power sequence or simply transferring model parameters.

[0355] The above comparative results demonstrate that the technical effectiveness of this invention does not solely stem from increasing model size or extending training time, but rather from the synergistic effect between theoretical clear-sky power decoupling, operational event prototype construction, event-level comparative pre-training, and target domain few-sample calibration. This experimental setup addresses the technical problem raised in the specification: how to avoid scale shifts and event mismatches caused by directly migrating the original power curve when there are insufficient samples of newly built power plants in the target domain.

[0356] Table 2 Performance of the method of the present invention under different target domain labeled data scales

[0357] As shown in Table 2, when the target domain has only one week of labeled data, the method of the present invention can still obtain good prediction results. As the sample size of the target domain increases from one week to four weeks, the MAE continues to decrease, indicating that the present invention is not only suitable for cold starts, but also has the ability to continuously optimize as the sample accumulates. Figure 7 This further illustrates the trend of change in a more intuitive way.

[0358] Further ablation experiments were conducted. As shown in Table 3, the complete method achieved the best results in terms of MAE, RMSE, MAPE, and nRMSE. After removing the theoretical clear-sky power decoupling, MAE increased from 0.6523 kW to 0.9146 kW, and nRMSE increased from 10.16% to 14.19%, indicating that the theoretical clear-sky power benchmark can weaken the impact of differences in installed capacity and solar irradiance phase on cross-site migration. After removing the clear-sky residual characterization or directly outputting the predicted power value, the error increased significantly, indicating that predicting the clear-sky attenuation coefficient or clear-sky residual before reconstructing the power is beneficial to reducing cross-site scale offset. After removing the event identification mechanism, cross-site event prototype library, or event prototype comparison loss, the errors of each method were higher than those of the complete method, indicating that using photovoltaic operation events as migration units can enhance the cross-site reuse capability of similar weather disturbances. After removing the target domain event prototype weight calibration, the prediction performance also decreased, indicating that the target domain few-sample calibration can improve the adaptability and migration stability of the target power station.

[0359] Each ablation term in Table 3 is based on the complete method, with only one key technical component removed. The remaining data inputs, training rounds, target domain sample size, and evaluation metrics remain the same. These are used to verify the contributions of theoretical clear-sky power decoupling, clear-sky residual characterization, event identification, cross-site event prototype library, event prototype contrast loss, and target domain event prototype weight calibration to prediction performance.

[0360] Table 3 Ablation Experiment Results of the Method of the Present Invention

[0361] Example 2

[0362] This embodiment provides a cross-site photovoltaic power prediction system based on clear-sky residual event prototype migration.

[0363] A computer-readable storage medium storing a plurality of instructions adapted for loading and execution by a processor of a terminal device of the cross-site photovoltaic power prediction method based on clear-sky residual event prototype migration.

[0364] A terminal device includes a processor and a computer-readable storage medium, the processor being configured to implement various instructions; the computer-readable storage medium being configured to store multiple instructions adapted for loading and execution by the processor of the cross-site photovoltaic power prediction method based on clear-sky residual event prototype migration.

[0365] The above are all preferred embodiments of the present invention and are not intended to limit the scope of protection of the present invention. Therefore, all equivalent changes made in accordance with the structure, shape and principle of the present invention should be covered within the scope of protection of the present invention.

Claims

1. A cross-site photovoltaic power prediction method based on clear-sky residual event prototype transfer, characterized in that, include: Acquire cross-site multi-source data from photovoltaic power plants in both the source and target domains; Perform time alignment, missing anomaly marking, and capacity normalization on cross-site multi-source data; Theoretical clear-sky power is constructed based on normalized data; Construct a clear-sky perturbation characterization sequence based on theoretical clear-sky power; Based on the constructed clear-sky disturbance characterization sequence, a multi-timescale historical window is constructed at each prediction time to generate a multi-timescale sample sequence. Identifying photovoltaic operation events based on generated multi-timescale sample sequences; A cross-site event prototype library was established based on the identified photovoltaic operation events. The event prototype contrastive loss is pre-trained in the source domain and calibrated with few samples in the target domain using a cross-site event prototype library. Dynamic fusion is performed based on source domain pre-training and target domain few-sample calibration results to reconstruct the future power and prediction range of the target power plant. The method of using a cross-site event prototype library to perform source domain pre-training and target domain few-sample calibration of the event prototype contrastive loss includes the use of a cross-site event prototype library. This study pre-trains an event-level comparison dataset on event fragments from multiple sites in the source domain and calibrates it on a small sample dataset of newly built photovoltaic power plants in the target domain. First, a similarity function between event embeddings and event prototypes is defined, and then an event prototype comparison loss is defined based on this similarity function. Simultaneously, a multi-scale prediction network is constructed. Input samples at multiple time scales Using event prototype matching information and station static attributes as input, the system predicts the clear-sky attenuation coefficient, normalized clear-sky residual, power value, and event class probability at future times, and defines the prediction loss. To constrain the event class prediction results, an event classification loss and a physical consistency loss are constructed. Based on the predicted clear-sky attenuation path and residual path, the power reconstruction result is obtained, leading to the source domain pre-training total loss. In the target domain few-sample calibration stage, the obtained target domain few-sample dataset is used... We obtain a set of embedded events from a small sample in the target domain, and then calculate the mean representation of events in the target domain and the matching weight of the target domain to the event prototypes in the source domain. Further, a target domain calibration prototype is constructed. During target domain fine-tuning, the main parameters of the event encoder are frozen, and only the target domain site embedding parameters, normalization parameters, prototype matching weights, and prediction output layer parameters are updated. The clear-sky attenuation distribution, residual distribution, and event prototype matching relationship specific to the target site are adjusted through a small number of target domain samples to reduce the risk of negative migration caused by direct migration. Finally, a lightweight Adapter module is set in the target domain fine-tuning stage, located between the hidden representation layer and the prediction output layer of the multi-scale temporal encoder. The input is the target domain historical window encoding vector, the target domain site embedding vector, and the target domain hybrid prototype representation, and the output is the target domain correction representation.

2. The cross-site photovoltaic power prediction method based on clear-sky residual event prototype migration according to claim 1, characterized in that, The process of performing time alignment, missing / anomaly marking, and capacity normalization on cross-station multi-source data includes, firstly, performing time unification and resampling processing on data from each power station, unifying data from different systems to the same time zone and fixed sampling interval. For the original observation sequence The time-aligned sequence is represented as: in, This indicates the operation of resampling and time alignment at fixed sampling intervals. A unified time index is used; then, missing data is processed by category. For short-term missing data, linear interpolation, spline interpolation, or the mean of adjacent time periods is used for imputation; for long-term missing data, the missing data marker is retained; next, outlier samples are identified and marked, and then the power data is normalized to reduce cross-station migration bias caused by differences in the installed capacity of different power plants, and robust standardization is applied to meteorological linear features; then, a sample quality score is constructed to screen high-confidence samples and adjust the weights of various samples during the training of the master prediction model; for samples with limited power generation and maintenance, their sample weights are reduced during the training of the master prediction model, and their operating condition markers are retained during the operation event identification stage. Finally, a unified feature sequence is constructed for each power plant.

3. The method for cross-site photovoltaic power prediction based on clear-sky residual event prototype migration according to claim 2, characterized in that, The theoretical clear-sky power constructed based on normalized data includes variables such as irradiance, ambient temperature, component temperature, and atmospheric pressure involved in physical modeling calculations, using the imputed original dimensional sequence. The calculation is performed, and the generated sample mask is retained to determine the reliability of the theoretical clear-sky benchmark result at the corresponding time. The superscript... This represents the original dimensional observation sequence after missing data imputation; for the first... A photovoltaic power station, static attribute vector At least latitude ,longitude ,altitude Installed capacity Component tilt angle Component azimuth angle Inverter efficiency System loss coefficient Temperature power coefficient and the component's nominal operating temperature For a unified time index According to its corresponding date sequence number Hourly Information The system calculates the sun's geometric position based on the station's latitude and longitude, and corrects for true solar time by adjusting for station longitude and standard time zone longitude to calculate the solar hour angle and solar altitude angle. Considering the influence of module tilt angle and azimuth angle on received irradiance, the horizontal clear-sky irradiance is converted to the clear-sky irradiance on the module's tilted surface, and the module incident angle is... Measured by solar altitude angle, solar azimuth angle, and component tilt angle. and component azimuth angle The effective irradiance model for the inclined surface under clear sky was then calculated, and the calculated value was obtained based on the measured irradiance of the inclined surface after filling. The model's calculated values ​​are weighted and corrected; the theoretical output power is adjusted based on the component temperature. The component temperature is selected based on the validity of the corresponding component temperature dimension in the missing mask. When the component temperature is valid, the measured component temperature after filling is used. When the component temperature is invalid, the component temperature is estimated based on the filled ambient temperature and the corrected clear sky irradiance of the tilted surface. Finally, the component temperature used for the theoretical clear sky power calculation is uniformly represented to obtain the theoretical clear sky power. At the same time, in order to ensure that the theoretical clear sky power conforms to the physical boundary of photovoltaic output, it is constrained. Finally, the theoretical clear sky power sequence is output as the physical benchmark for subsequent calculation of clear sky attenuation coefficient and clear sky residual.

4. The cross-site photovoltaic power prediction method based on clear-sky residual event prototype migration according to claim 3, characterized in that, The construction of the clear-sky disturbance characterization sequence based on theoretical clear-sky power includes setting an effective threshold for theoretical clear-sky power to avoid distortion in proportional calculations caused by excessively low theoretical clear-sky power at night or low solar altitude angles. To filter out moments when the theoretical power is close to zero and construct an effective clear sky marker; at the same time, a clear sky attenuation coefficient is defined during the effective clear sky period; To avoid unreasonable extreme values ​​in the attenuation coefficient caused by sensor errors or deviations in the theoretical clear-sky model, boundary constraints are imposed on the clear-sky attenuation coefficient, and the clear-sky residual is further defined. Since the installed capacity of different power plants is different, direct comparison of absolute residuals is affected by the capacity scale. Therefore, a normalized clear-sky residual is constructed, and the change in the clear-sky attenuation coefficient and the change in the normalized clear-sky residual are obtained based on the rate of change of the disturbance. To describe the intensity of the disturbance over a period of time, in a length of Within the historical window, the intensity of local residual fluctuations is calculated, and the mean of local residuals is further constructed. Finally, the clear-sky disturbance characterization vector is constructed, and the clear-sky disturbance characterization sequence is output. This is used to transform the original power prediction problem into an attenuation and residual modeling problem under a clear-sky baseline.

5. The cross-site photovoltaic power prediction method based on clear-sky residual event prototype migration according to claim 4, characterized in that, The constructed clear-sky disturbance characterization sequence generates a multi-timescale historical window at each prediction time. This includes constructing short-scale, mesoscale, and long-scale historical windows for each prediction time based on the generated clear-sky disturbance characterization sequence, with a prediction step size of [missing information]. The future moment to be predicted is ,in To preset the prediction step size and to simultaneously describe the operational status within different time ranges, the short-scale window length is set to be [value missing]. The length of the mesoscale window is The length of the long-scale window is The lengths of all three windows are based on the sampling interval. As the basic unit; for the predicted time When constructing a short-scale historical window, the changes in clear sky attenuation coefficient, residual mutations, local fluctuations, and operating condition markers in the recent period are retained, which is suitable for describing the rapid changes in cloud cover, cloud recovery, short-term sensor anomalies, and power ramp-up. When constructing a mesoscale historical window, it covers a longer intraday variation range than the short-scale window, and is used to describe the weather evolution, overall attenuation level changes, peak reduction patterns and continuous low-output segments from morning to noon and afternoon to evening. When constructing long-scale historical windows, residual distribution shifts caused by cross-day weather processes, continuous overcast and rainy weather, equipment aging, or long-term operation and maintenance are represented. To integrate multi-scale windows into a unified modeling process, statistical summaries are constructed for each of the three windows, and a multi-scale summary vector is constructed by combining the statistical summaries of the three scales. The output station static attribute vector is then used to... Time period coding Wind direction cycle coding Sample quality score and multidimensional sample mask Combined with three historical windows, a multi-scale input sample for the predicted time is obtained. For the initial period with insufficient historical length, forward padding, zero padding and effective fragment masking are used to fill the window length, and the padding mark is retained in the multi-dimensional sample mask to avoid the padding sample being mistaken for the real observation. Finally, a multi-timescale sample sequence is output.

6. The cross-site photovoltaic power prediction method based on clear-sky residual event prototype migration according to claim 5, characterized in that, The method for identifying photovoltaic operation events based on generated multi-timescale sample sequences includes using clear-sky disturbance characterization vectors and multi-timescale input samples. To identify events, first identify the clear-sky stable event, which indicates that the actual output of the photovoltaic power station is close to the theoretical clear-sky output, and the attenuation coefficient and residual fluctuation are small in a short period of time. Second identify the cloud shading event, which is characterized by the theoretical clear-sky power being in the effective power generation period, but the actual power dropping rapidly relative to the clear-sky benchmark, and the clear-sky attenuation coefficient decreasing in a short period of time. Secondly, identify post-cloud recovery events. Post-cloud recovery occurs after a sudden cloud cover event, characterized by a rapid recovery of actual power relative to theoretical clear-sky power, and a short-term increase in the clear-sky attenuation coefficient. Further, identify continuous cloudy low-output events, assuming a mesoscale window length of [missing information]. The average clear-sky attenuation coefficient within the window is calculated to distinguish weather processes that are sustained at low power but change relatively slowly. Then, peak shaving events are identified. Peak shaving events are characterized by high theoretical clear-sky power, but the measured power is suppressed to near a certain plateau, and the peak shaving mask is valid. Finally, abnormally low power events are identified. Abnormally low power events are characterized by valid theoretical clear-sky power, but the actual power is significantly lower than the theoretical clear-sky power, and there are no peak shaving or maintenance markers. After completing the single-moment event discrimination, moments that consecutively meet the discrimination criteria under the same event type are merged into event segments. Finally, an event description vector is generated for each event segment, and the first event description vector is output. A collection of event fragments from a power station.

7. The cross-site photovoltaic power prediction method based on clear-sky residual event prototype migration according to claim 6, characterized in that, The establishment of a cross-site event prototype library based on the identified photovoltaic operation events includes constructing the cross-site event prototype library based on the output set of photovoltaic operation event fragments, and calling event fragments. and event description vector This method maps similar events from different source power plants to a unified representation space and aggregates them in a weighted manner into event prototypes that can be transferred across power plants. The event encoder is one such prototype. Local event feature extraction function Composed of an attention convergence function and a local event feature extraction function, the local event feature extraction function is used to extract the clear sky disturbance features at each moment in the event segment, and the attention convergence function is used to assign higher weights to key moments within the event segment, ultimately obtaining an event embedding vector of uniform dimension. Thus, local event features are obtained based on the computational link of the event encoding process, and then the local event features are weighted and converged to obtain the event embedding; finally, for any event type... The event embedding vectors belonging to the event type are collected from all source domain power stations to form a set of similar event embeddings. Then, the event embeddings under the same event type are weighted and aggregated based on the aggregation weight of similar events to obtain the event prototype, and thus obtain the cross-site event prototype library.

8. The cross-site photovoltaic power prediction method based on clear-sky residual event prototype migration according to claim 7, characterized in that, The process involves dynamically fusing source domain pre-training and target domain few-sample calibration results to reconstruct the future power and prediction range of the target power plant. This includes outputting the predicted power value, clear-sky attenuation coefficient, normalized clear-sky residual, event category, and prototype matching weight of the target domain photovoltaic power plant at future times based on the target domain prediction network. The output clear-sky disturbance prediction is then combined with the future theoretical clear-sky power benchmark to reconstruct the final photovoltaic power prediction result. To balance the reliability of the clear-sky attenuation path and residual path under different operating conditions, fusion weights are dynamically generated based on the predicted event type, target domain prototype matching weight, and prediction uncertainty at future times. A path reliability score is defined based on the event category probability, target domain prototype matching weight, and prediction uncertainty. The two... The reliability score of the path is normalized into a fusion weight. The larger the fusion weight, the more reliable the corresponding path is under the current prediction conditions. When the future time is close to a clear sky stable event, the proportion of clear sky attenuation paths increases; when the future time is close to cloud cover, power limitation, or abnormal disturbance events, the proportion of residual paths increases. Then, based on the fusion weight, the two prediction results are fused. To ensure that the output result meets the physical boundary of photovoltaic power output, the fused prediction value is constrained. To construct the prediction interval, the interval width coefficient is first generated based on the prediction uncertainty score, the event disturbance intensity, and the target domain prototype matching dispersion. The higher the prediction uncertainty, the higher the probability of the disturbance event, and the more dispersed the prototype matching, the larger the interval width. By defining the upper and lower bounds of the prediction interval, the future prediction is finally output. The power prediction sequence and prediction interval sequence at each prediction time, as well as prediction interpretation information, including the predicted clear sky attenuation coefficient, the predicted normalized clear sky residual, the predicted event category probability, the target domain prototype matching weight, the two path fusion weight, the interval width coefficient, and the prediction uncertainty score.

9. A cross-site photovoltaic power prediction system based on clear-sky residual event prototype transfer, executing the cross-site photovoltaic power prediction method based on clear-sky residual event prototype transfer as described in claim 1, characterized in that, include: The data acquisition module is configured to acquire cross-site multi-source data from source domain photovoltaic power plants and target domain photovoltaic power plants; The preprocessing module is configured to perform time alignment, missing anomaly marking, and capacity normalization on cross-site multi-source data; The clear sky module is configured to construct theoretical clear sky power based on normalized data. The perturbation module is configured to construct a clear-sky perturbation characterization sequence based on theoretical clear-sky power; The window module is configured to construct a multi-timescale historical window at each prediction time based on the constructed clear-sky disturbance characterization sequence, and generate a multi-timescale sample sequence. The photovoltaic module is configured to identify photovoltaic operating events based on generated multi-timescale sample sequences; The prototype library module is configured to build a cross-site event prototype library based on the identified photovoltaic operation events. The calibration module is configured to perform source domain pre-training and target domain few-sample calibration of the event prototype contrastive loss using a cross-site event prototype library. The prediction module is configured to dynamically fuse source domain pre-training and target domain few-sample calibration results to reconstruct the future power and prediction range of the target power plant.

Images